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 <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/page_ext.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
67 #include <asm/sections.h>
68 #include <asm/tlbflush.h>
69 #include <asm/div64.h>
72 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
73 static DEFINE_MUTEX(pcp_batch_high_lock);
74 #define MIN_PERCPU_PAGELIST_FRACTION (8)
76 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
77 DEFINE_PER_CPU(int, numa_node);
78 EXPORT_PER_CPU_SYMBOL(numa_node);
81 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
83 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
84 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
85 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
86 * defined in <linux/topology.h>.
88 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
89 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
90 int _node_numa_mem_[MAX_NUMNODES];
94 * Array of node states.
96 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
97 [N_POSSIBLE] = NODE_MASK_ALL,
98 [N_ONLINE] = { { [0] = 1UL } },
100 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
101 #ifdef CONFIG_HIGHMEM
102 [N_HIGH_MEMORY] = { { [0] = 1UL } },
104 #ifdef CONFIG_MOVABLE_NODE
105 [N_MEMORY] = { { [0] = 1UL } },
107 [N_CPU] = { { [0] = 1UL } },
110 EXPORT_SYMBOL(node_states);
112 /* Protect totalram_pages and zone->managed_pages */
113 static DEFINE_SPINLOCK(managed_page_count_lock);
115 unsigned long totalram_pages __read_mostly;
116 unsigned long totalreserve_pages __read_mostly;
117 unsigned long totalcma_pages __read_mostly;
119 int percpu_pagelist_fraction;
120 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
123 * A cached value of the page's pageblock's migratetype, used when the page is
124 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
125 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
126 * Also the migratetype set in the page does not necessarily match the pcplist
127 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
128 * other index - this ensures that it will be put on the correct CMA freelist.
130 static inline int get_pcppage_migratetype(struct page *page)
135 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
137 page->index = migratetype;
140 #ifdef CONFIG_PM_SLEEP
142 * The following functions are used by the suspend/hibernate code to temporarily
143 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
144 * while devices are suspended. To avoid races with the suspend/hibernate code,
145 * they should always be called with pm_mutex held (gfp_allowed_mask also should
146 * only be modified with pm_mutex held, unless the suspend/hibernate code is
147 * guaranteed not to run in parallel with that modification).
150 static gfp_t saved_gfp_mask;
152 void pm_restore_gfp_mask(void)
154 WARN_ON(!mutex_is_locked(&pm_mutex));
155 if (saved_gfp_mask) {
156 gfp_allowed_mask = saved_gfp_mask;
161 void pm_restrict_gfp_mask(void)
163 WARN_ON(!mutex_is_locked(&pm_mutex));
164 WARN_ON(saved_gfp_mask);
165 saved_gfp_mask = gfp_allowed_mask;
166 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
169 bool pm_suspended_storage(void)
171 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
175 #endif /* CONFIG_PM_SLEEP */
177 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
178 unsigned int pageblock_order __read_mostly;
181 static void __free_pages_ok(struct page *page, unsigned int order);
184 * results with 256, 32 in the lowmem_reserve sysctl:
185 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
186 * 1G machine -> (16M dma, 784M normal, 224M high)
187 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
188 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
189 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
191 * TBD: should special case ZONE_DMA32 machines here - in those we normally
192 * don't need any ZONE_NORMAL reservation
194 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
195 #ifdef CONFIG_ZONE_DMA
198 #ifdef CONFIG_ZONE_DMA32
201 #ifdef CONFIG_HIGHMEM
207 EXPORT_SYMBOL(totalram_pages);
209 static char * const zone_names[MAX_NR_ZONES] = {
210 #ifdef CONFIG_ZONE_DMA
213 #ifdef CONFIG_ZONE_DMA32
217 #ifdef CONFIG_HIGHMEM
221 #ifdef CONFIG_ZONE_DEVICE
226 char * const migratetype_names[MIGRATE_TYPES] = {
234 #ifdef CONFIG_MEMORY_ISOLATION
239 compound_page_dtor * const compound_page_dtors[] = {
242 #ifdef CONFIG_HUGETLB_PAGE
245 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
250 int min_free_kbytes = 1024;
251 int user_min_free_kbytes = -1;
253 static unsigned long __meminitdata nr_kernel_pages;
254 static unsigned long __meminitdata nr_all_pages;
255 static unsigned long __meminitdata dma_reserve;
257 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
258 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
259 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
260 static unsigned long __initdata required_kernelcore;
261 static unsigned long __initdata required_movablecore;
262 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
263 static bool mirrored_kernelcore;
265 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
267 EXPORT_SYMBOL(movable_zone);
268 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
271 int nr_node_ids __read_mostly = MAX_NUMNODES;
272 int nr_online_nodes __read_mostly = 1;
273 EXPORT_SYMBOL(nr_node_ids);
274 EXPORT_SYMBOL(nr_online_nodes);
277 int page_group_by_mobility_disabled __read_mostly;
279 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
280 static inline void reset_deferred_meminit(pg_data_t *pgdat)
282 pgdat->first_deferred_pfn = ULONG_MAX;
285 /* Returns true if the struct page for the pfn is uninitialised */
286 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
288 if (pfn >= NODE_DATA(early_pfn_to_nid(pfn))->first_deferred_pfn)
294 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
296 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
303 * Returns false when the remaining initialisation should be deferred until
304 * later in the boot cycle when it can be parallelised.
306 static inline bool update_defer_init(pg_data_t *pgdat,
307 unsigned long pfn, unsigned long zone_end,
308 unsigned long *nr_initialised)
310 /* Always populate low zones for address-contrained allocations */
311 if (zone_end < pgdat_end_pfn(pgdat))
314 /* Initialise at least 2G of the highest zone */
316 if (*nr_initialised > (2UL << (30 - PAGE_SHIFT)) &&
317 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
318 pgdat->first_deferred_pfn = pfn;
325 static inline void reset_deferred_meminit(pg_data_t *pgdat)
329 static inline bool early_page_uninitialised(unsigned long pfn)
334 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
339 static inline bool update_defer_init(pg_data_t *pgdat,
340 unsigned long pfn, unsigned long zone_end,
341 unsigned long *nr_initialised)
348 void set_pageblock_migratetype(struct page *page, int migratetype)
350 if (unlikely(page_group_by_mobility_disabled &&
351 migratetype < MIGRATE_PCPTYPES))
352 migratetype = MIGRATE_UNMOVABLE;
354 set_pageblock_flags_group(page, (unsigned long)migratetype,
355 PB_migrate, PB_migrate_end);
358 #ifdef CONFIG_DEBUG_VM
359 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
363 unsigned long pfn = page_to_pfn(page);
364 unsigned long sp, start_pfn;
367 seq = zone_span_seqbegin(zone);
368 start_pfn = zone->zone_start_pfn;
369 sp = zone->spanned_pages;
370 if (!zone_spans_pfn(zone, pfn))
372 } while (zone_span_seqretry(zone, seq));
375 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
376 pfn, zone_to_nid(zone), zone->name,
377 start_pfn, start_pfn + sp);
382 static int page_is_consistent(struct zone *zone, struct page *page)
384 if (!pfn_valid_within(page_to_pfn(page)))
386 if (zone != page_zone(page))
392 * Temporary debugging check for pages not lying within a given zone.
394 static int bad_range(struct zone *zone, struct page *page)
396 if (page_outside_zone_boundaries(zone, page))
398 if (!page_is_consistent(zone, page))
404 static inline int bad_range(struct zone *zone, struct page *page)
410 static void bad_page(struct page *page, const char *reason,
411 unsigned long bad_flags)
413 static unsigned long resume;
414 static unsigned long nr_shown;
415 static unsigned long nr_unshown;
417 /* Don't complain about poisoned pages */
418 if (PageHWPoison(page)) {
419 page_mapcount_reset(page); /* remove PageBuddy */
424 * Allow a burst of 60 reports, then keep quiet for that minute;
425 * or allow a steady drip of one report per second.
427 if (nr_shown == 60) {
428 if (time_before(jiffies, resume)) {
434 "BUG: Bad page state: %lu messages suppressed\n",
441 resume = jiffies + 60 * HZ;
443 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
444 current->comm, page_to_pfn(page));
445 __dump_page(page, reason);
446 bad_flags &= page->flags;
448 pr_alert("bad because of flags: %#lx(%pGp)\n",
449 bad_flags, &bad_flags);
450 dump_page_owner(page);
455 /* Leave bad fields for debug, except PageBuddy could make trouble */
456 page_mapcount_reset(page); /* remove PageBuddy */
457 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
461 * Higher-order pages are called "compound pages". They are structured thusly:
463 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
465 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
466 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
468 * The first tail page's ->compound_dtor holds the offset in array of compound
469 * page destructors. See compound_page_dtors.
471 * The first tail page's ->compound_order holds the order of allocation.
472 * This usage means that zero-order pages may not be compound.
475 void free_compound_page(struct page *page)
477 __free_pages_ok(page, compound_order(page));
480 void prep_compound_page(struct page *page, unsigned int order)
483 int nr_pages = 1 << order;
485 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
486 set_compound_order(page, order);
488 for (i = 1; i < nr_pages; i++) {
489 struct page *p = page + i;
490 set_page_count(p, 0);
491 p->mapping = TAIL_MAPPING;
492 set_compound_head(p, page);
494 atomic_set(compound_mapcount_ptr(page), -1);
497 #ifdef CONFIG_DEBUG_PAGEALLOC
498 unsigned int _debug_guardpage_minorder;
499 bool _debug_pagealloc_enabled __read_mostly
500 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
501 bool _debug_guardpage_enabled __read_mostly;
503 static int __init early_debug_pagealloc(char *buf)
508 if (strcmp(buf, "on") == 0)
509 _debug_pagealloc_enabled = true;
511 if (strcmp(buf, "off") == 0)
512 _debug_pagealloc_enabled = false;
516 early_param("debug_pagealloc", early_debug_pagealloc);
518 static bool need_debug_guardpage(void)
520 /* If we don't use debug_pagealloc, we don't need guard page */
521 if (!debug_pagealloc_enabled())
527 static void init_debug_guardpage(void)
529 if (!debug_pagealloc_enabled())
532 _debug_guardpage_enabled = true;
535 struct page_ext_operations debug_guardpage_ops = {
536 .need = need_debug_guardpage,
537 .init = init_debug_guardpage,
540 static int __init debug_guardpage_minorder_setup(char *buf)
544 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
545 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
548 _debug_guardpage_minorder = res;
549 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
552 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
554 static inline void set_page_guard(struct zone *zone, struct page *page,
555 unsigned int order, int migratetype)
557 struct page_ext *page_ext;
559 if (!debug_guardpage_enabled())
562 page_ext = lookup_page_ext(page);
563 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
565 INIT_LIST_HEAD(&page->lru);
566 set_page_private(page, order);
567 /* Guard pages are not available for any usage */
568 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
571 static inline void clear_page_guard(struct zone *zone, struct page *page,
572 unsigned int order, int migratetype)
574 struct page_ext *page_ext;
576 if (!debug_guardpage_enabled())
579 page_ext = lookup_page_ext(page);
580 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
582 set_page_private(page, 0);
583 if (!is_migrate_isolate(migratetype))
584 __mod_zone_freepage_state(zone, (1 << order), migratetype);
587 struct page_ext_operations debug_guardpage_ops = { NULL, };
588 static inline void set_page_guard(struct zone *zone, struct page *page,
589 unsigned int order, int migratetype) {}
590 static inline void clear_page_guard(struct zone *zone, struct page *page,
591 unsigned int order, int migratetype) {}
594 static inline void set_page_order(struct page *page, unsigned int order)
596 set_page_private(page, order);
597 __SetPageBuddy(page);
600 static inline void rmv_page_order(struct page *page)
602 __ClearPageBuddy(page);
603 set_page_private(page, 0);
607 * This function checks whether a page is free && is the buddy
608 * we can do coalesce a page and its buddy if
609 * (a) the buddy is not in a hole &&
610 * (b) the buddy is in the buddy system &&
611 * (c) a page and its buddy have the same order &&
612 * (d) a page and its buddy are in the same zone.
614 * For recording whether a page is in the buddy system, we set ->_mapcount
615 * PAGE_BUDDY_MAPCOUNT_VALUE.
616 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
617 * serialized by zone->lock.
619 * For recording page's order, we use page_private(page).
621 static inline int page_is_buddy(struct page *page, struct page *buddy,
624 if (!pfn_valid_within(page_to_pfn(buddy)))
627 if (page_is_guard(buddy) && page_order(buddy) == order) {
628 if (page_zone_id(page) != page_zone_id(buddy))
631 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
636 if (PageBuddy(buddy) && page_order(buddy) == order) {
638 * zone check is done late to avoid uselessly
639 * calculating zone/node ids for pages that could
642 if (page_zone_id(page) != page_zone_id(buddy))
645 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
653 * Freeing function for a buddy system allocator.
655 * The concept of a buddy system is to maintain direct-mapped table
656 * (containing bit values) for memory blocks of various "orders".
657 * The bottom level table contains the map for the smallest allocatable
658 * units of memory (here, pages), and each level above it describes
659 * pairs of units from the levels below, hence, "buddies".
660 * At a high level, all that happens here is marking the table entry
661 * at the bottom level available, and propagating the changes upward
662 * as necessary, plus some accounting needed to play nicely with other
663 * parts of the VM system.
664 * At each level, we keep a list of pages, which are heads of continuous
665 * free pages of length of (1 << order) and marked with _mapcount
666 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
668 * So when we are allocating or freeing one, we can derive the state of the
669 * other. That is, if we allocate a small block, and both were
670 * free, the remainder of the region must be split into blocks.
671 * If a block is freed, and its buddy is also free, then this
672 * triggers coalescing into a block of larger size.
677 static inline void __free_one_page(struct page *page,
679 struct zone *zone, unsigned int order,
682 unsigned long page_idx;
683 unsigned long combined_idx;
684 unsigned long uninitialized_var(buddy_idx);
686 unsigned int max_order = MAX_ORDER;
688 VM_BUG_ON(!zone_is_initialized(zone));
689 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
691 VM_BUG_ON(migratetype == -1);
692 if (is_migrate_isolate(migratetype)) {
694 * We restrict max order of merging to prevent merge
695 * between freepages on isolate pageblock and normal
696 * pageblock. Without this, pageblock isolation
697 * could cause incorrect freepage accounting.
699 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
701 __mod_zone_freepage_state(zone, 1 << order, migratetype);
704 page_idx = pfn & ((1 << max_order) - 1);
706 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
707 VM_BUG_ON_PAGE(bad_range(zone, page), page);
709 while (order < max_order - 1) {
710 buddy_idx = __find_buddy_index(page_idx, order);
711 buddy = page + (buddy_idx - page_idx);
712 if (!page_is_buddy(page, buddy, order))
715 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
716 * merge with it and move up one order.
718 if (page_is_guard(buddy)) {
719 clear_page_guard(zone, buddy, order, migratetype);
721 list_del(&buddy->lru);
722 zone->free_area[order].nr_free--;
723 rmv_page_order(buddy);
725 combined_idx = buddy_idx & page_idx;
726 page = page + (combined_idx - page_idx);
727 page_idx = combined_idx;
730 set_page_order(page, order);
733 * If this is not the largest possible page, check if the buddy
734 * of the next-highest order is free. If it is, it's possible
735 * that pages are being freed that will coalesce soon. In case,
736 * that is happening, add the free page to the tail of the list
737 * so it's less likely to be used soon and more likely to be merged
738 * as a higher order page
740 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
741 struct page *higher_page, *higher_buddy;
742 combined_idx = buddy_idx & page_idx;
743 higher_page = page + (combined_idx - page_idx);
744 buddy_idx = __find_buddy_index(combined_idx, order + 1);
745 higher_buddy = higher_page + (buddy_idx - combined_idx);
746 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
747 list_add_tail(&page->lru,
748 &zone->free_area[order].free_list[migratetype]);
753 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
755 zone->free_area[order].nr_free++;
758 static inline int free_pages_check(struct page *page)
760 const char *bad_reason = NULL;
761 unsigned long bad_flags = 0;
763 if (unlikely(atomic_read(&page->_mapcount) != -1))
764 bad_reason = "nonzero mapcount";
765 if (unlikely(page->mapping != NULL))
766 bad_reason = "non-NULL mapping";
767 if (unlikely(atomic_read(&page->_count) != 0))
768 bad_reason = "nonzero _count";
769 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
770 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
771 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
774 if (unlikely(page->mem_cgroup))
775 bad_reason = "page still charged to cgroup";
777 if (unlikely(bad_reason)) {
778 bad_page(page, bad_reason, bad_flags);
781 page_cpupid_reset_last(page);
782 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
783 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
788 * Frees a number of pages from the PCP lists
789 * Assumes all pages on list are in same zone, and of same order.
790 * count is the number of pages to free.
792 * If the zone was previously in an "all pages pinned" state then look to
793 * see if this freeing clears that state.
795 * And clear the zone's pages_scanned counter, to hold off the "all pages are
796 * pinned" detection logic.
798 static void free_pcppages_bulk(struct zone *zone, int count,
799 struct per_cpu_pages *pcp)
804 unsigned long nr_scanned;
806 spin_lock(&zone->lock);
807 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
809 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
813 struct list_head *list;
816 * Remove pages from lists in a round-robin fashion. A
817 * batch_free count is maintained that is incremented when an
818 * empty list is encountered. This is so more pages are freed
819 * off fuller lists instead of spinning excessively around empty
824 if (++migratetype == MIGRATE_PCPTYPES)
826 list = &pcp->lists[migratetype];
827 } while (list_empty(list));
829 /* This is the only non-empty list. Free them all. */
830 if (batch_free == MIGRATE_PCPTYPES)
831 batch_free = to_free;
834 int mt; /* migratetype of the to-be-freed page */
836 page = list_last_entry(list, struct page, lru);
837 /* must delete as __free_one_page list manipulates */
838 list_del(&page->lru);
840 mt = get_pcppage_migratetype(page);
841 /* MIGRATE_ISOLATE page should not go to pcplists */
842 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
843 /* Pageblock could have been isolated meanwhile */
844 if (unlikely(has_isolate_pageblock(zone)))
845 mt = get_pageblock_migratetype(page);
847 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
848 trace_mm_page_pcpu_drain(page, 0, mt);
849 } while (--to_free && --batch_free && !list_empty(list));
851 spin_unlock(&zone->lock);
854 static void free_one_page(struct zone *zone,
855 struct page *page, unsigned long pfn,
859 unsigned long nr_scanned;
860 spin_lock(&zone->lock);
861 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
863 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
865 if (unlikely(has_isolate_pageblock(zone) ||
866 is_migrate_isolate(migratetype))) {
867 migratetype = get_pfnblock_migratetype(page, pfn);
869 __free_one_page(page, pfn, zone, order, migratetype);
870 spin_unlock(&zone->lock);
873 static int free_tail_pages_check(struct page *head_page, struct page *page)
878 * We rely page->lru.next never has bit 0 set, unless the page
879 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
881 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
883 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
887 switch (page - head_page) {
889 /* the first tail page: ->mapping is compound_mapcount() */
890 if (unlikely(compound_mapcount(page))) {
891 bad_page(page, "nonzero compound_mapcount", 0);
897 * the second tail page: ->mapping is
898 * page_deferred_list().next -- ignore value.
902 if (page->mapping != TAIL_MAPPING) {
903 bad_page(page, "corrupted mapping in tail page", 0);
908 if (unlikely(!PageTail(page))) {
909 bad_page(page, "PageTail not set", 0);
912 if (unlikely(compound_head(page) != head_page)) {
913 bad_page(page, "compound_head not consistent", 0);
918 page->mapping = NULL;
919 clear_compound_head(page);
923 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
924 unsigned long zone, int nid)
926 set_page_links(page, zone, nid, pfn);
927 init_page_count(page);
928 page_mapcount_reset(page);
929 page_cpupid_reset_last(page);
931 INIT_LIST_HEAD(&page->lru);
932 #ifdef WANT_PAGE_VIRTUAL
933 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
934 if (!is_highmem_idx(zone))
935 set_page_address(page, __va(pfn << PAGE_SHIFT));
939 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
942 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
945 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
946 static void init_reserved_page(unsigned long pfn)
951 if (!early_page_uninitialised(pfn))
954 nid = early_pfn_to_nid(pfn);
955 pgdat = NODE_DATA(nid);
957 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
958 struct zone *zone = &pgdat->node_zones[zid];
960 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
963 __init_single_pfn(pfn, zid, nid);
966 static inline void init_reserved_page(unsigned long pfn)
969 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
972 * Initialised pages do not have PageReserved set. This function is
973 * called for each range allocated by the bootmem allocator and
974 * marks the pages PageReserved. The remaining valid pages are later
975 * sent to the buddy page allocator.
977 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
979 unsigned long start_pfn = PFN_DOWN(start);
980 unsigned long end_pfn = PFN_UP(end);
982 for (; start_pfn < end_pfn; start_pfn++) {
983 if (pfn_valid(start_pfn)) {
984 struct page *page = pfn_to_page(start_pfn);
986 init_reserved_page(start_pfn);
988 /* Avoid false-positive PageTail() */
989 INIT_LIST_HEAD(&page->lru);
991 SetPageReserved(page);
996 static bool free_pages_prepare(struct page *page, unsigned int order)
998 bool compound = PageCompound(page);
1001 VM_BUG_ON_PAGE(PageTail(page), page);
1002 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1004 trace_mm_page_free(page, order);
1005 kmemcheck_free_shadow(page, order);
1006 kasan_free_pages(page, order);
1009 page->mapping = NULL;
1010 bad += free_pages_check(page);
1011 for (i = 1; i < (1 << order); i++) {
1013 bad += free_tail_pages_check(page, page + i);
1014 bad += free_pages_check(page + i);
1019 reset_page_owner(page, order);
1021 if (!PageHighMem(page)) {
1022 debug_check_no_locks_freed(page_address(page),
1023 PAGE_SIZE << order);
1024 debug_check_no_obj_freed(page_address(page),
1025 PAGE_SIZE << order);
1027 arch_free_page(page, order);
1028 kernel_poison_pages(page, 1 << order, 0);
1029 kernel_map_pages(page, 1 << order, 0);
1034 static void __free_pages_ok(struct page *page, unsigned int order)
1036 unsigned long flags;
1038 unsigned long pfn = page_to_pfn(page);
1040 if (!free_pages_prepare(page, order))
1043 migratetype = get_pfnblock_migratetype(page, pfn);
1044 local_irq_save(flags);
1045 __count_vm_events(PGFREE, 1 << order);
1046 free_one_page(page_zone(page), page, pfn, order, migratetype);
1047 local_irq_restore(flags);
1050 static void __init __free_pages_boot_core(struct page *page,
1051 unsigned long pfn, unsigned int order)
1053 unsigned int nr_pages = 1 << order;
1054 struct page *p = page;
1058 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1060 __ClearPageReserved(p);
1061 set_page_count(p, 0);
1063 __ClearPageReserved(p);
1064 set_page_count(p, 0);
1066 page_zone(page)->managed_pages += nr_pages;
1067 set_page_refcounted(page);
1068 __free_pages(page, order);
1071 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1072 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1074 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1076 int __meminit early_pfn_to_nid(unsigned long pfn)
1078 static DEFINE_SPINLOCK(early_pfn_lock);
1081 spin_lock(&early_pfn_lock);
1082 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1085 spin_unlock(&early_pfn_lock);
1091 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1092 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1093 struct mminit_pfnnid_cache *state)
1097 nid = __early_pfn_to_nid(pfn, state);
1098 if (nid >= 0 && nid != node)
1103 /* Only safe to use early in boot when initialisation is single-threaded */
1104 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1106 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1111 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1115 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1116 struct mminit_pfnnid_cache *state)
1123 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1126 if (early_page_uninitialised(pfn))
1128 return __free_pages_boot_core(page, pfn, order);
1131 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1132 static void __init deferred_free_range(struct page *page,
1133 unsigned long pfn, int nr_pages)
1140 /* Free a large naturally-aligned chunk if possible */
1141 if (nr_pages == MAX_ORDER_NR_PAGES &&
1142 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1143 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1144 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1148 for (i = 0; i < nr_pages; i++, page++, pfn++)
1149 __free_pages_boot_core(page, pfn, 0);
1152 /* Completion tracking for deferred_init_memmap() threads */
1153 static atomic_t pgdat_init_n_undone __initdata;
1154 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1156 static inline void __init pgdat_init_report_one_done(void)
1158 if (atomic_dec_and_test(&pgdat_init_n_undone))
1159 complete(&pgdat_init_all_done_comp);
1162 /* Initialise remaining memory on a node */
1163 static int __init deferred_init_memmap(void *data)
1165 pg_data_t *pgdat = data;
1166 int nid = pgdat->node_id;
1167 struct mminit_pfnnid_cache nid_init_state = { };
1168 unsigned long start = jiffies;
1169 unsigned long nr_pages = 0;
1170 unsigned long walk_start, walk_end;
1173 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1174 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1176 if (first_init_pfn == ULONG_MAX) {
1177 pgdat_init_report_one_done();
1181 /* Bind memory initialisation thread to a local node if possible */
1182 if (!cpumask_empty(cpumask))
1183 set_cpus_allowed_ptr(current, cpumask);
1185 /* Sanity check boundaries */
1186 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1187 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1188 pgdat->first_deferred_pfn = ULONG_MAX;
1190 /* Only the highest zone is deferred so find it */
1191 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1192 zone = pgdat->node_zones + zid;
1193 if (first_init_pfn < zone_end_pfn(zone))
1197 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1198 unsigned long pfn, end_pfn;
1199 struct page *page = NULL;
1200 struct page *free_base_page = NULL;
1201 unsigned long free_base_pfn = 0;
1204 end_pfn = min(walk_end, zone_end_pfn(zone));
1205 pfn = first_init_pfn;
1206 if (pfn < walk_start)
1208 if (pfn < zone->zone_start_pfn)
1209 pfn = zone->zone_start_pfn;
1211 for (; pfn < end_pfn; pfn++) {
1212 if (!pfn_valid_within(pfn))
1216 * Ensure pfn_valid is checked every
1217 * MAX_ORDER_NR_PAGES for memory holes
1219 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1220 if (!pfn_valid(pfn)) {
1226 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1231 /* Minimise pfn page lookups and scheduler checks */
1232 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1235 nr_pages += nr_to_free;
1236 deferred_free_range(free_base_page,
1237 free_base_pfn, nr_to_free);
1238 free_base_page = NULL;
1239 free_base_pfn = nr_to_free = 0;
1241 page = pfn_to_page(pfn);
1246 VM_BUG_ON(page_zone(page) != zone);
1250 __init_single_page(page, pfn, zid, nid);
1251 if (!free_base_page) {
1252 free_base_page = page;
1253 free_base_pfn = pfn;
1258 /* Where possible, batch up pages for a single free */
1261 /* Free the current block of pages to allocator */
1262 nr_pages += nr_to_free;
1263 deferred_free_range(free_base_page, free_base_pfn,
1265 free_base_page = NULL;
1266 free_base_pfn = nr_to_free = 0;
1269 first_init_pfn = max(end_pfn, first_init_pfn);
1272 /* Sanity check that the next zone really is unpopulated */
1273 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1275 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1276 jiffies_to_msecs(jiffies - start));
1278 pgdat_init_report_one_done();
1282 void __init page_alloc_init_late(void)
1286 /* There will be num_node_state(N_MEMORY) threads */
1287 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1288 for_each_node_state(nid, N_MEMORY) {
1289 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1292 /* Block until all are initialised */
1293 wait_for_completion(&pgdat_init_all_done_comp);
1295 /* Reinit limits that are based on free pages after the kernel is up */
1296 files_maxfiles_init();
1298 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1301 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1302 void __init init_cma_reserved_pageblock(struct page *page)
1304 unsigned i = pageblock_nr_pages;
1305 struct page *p = page;
1308 __ClearPageReserved(p);
1309 set_page_count(p, 0);
1312 set_pageblock_migratetype(page, MIGRATE_CMA);
1314 if (pageblock_order >= MAX_ORDER) {
1315 i = pageblock_nr_pages;
1318 set_page_refcounted(p);
1319 __free_pages(p, MAX_ORDER - 1);
1320 p += MAX_ORDER_NR_PAGES;
1321 } while (i -= MAX_ORDER_NR_PAGES);
1323 set_page_refcounted(page);
1324 __free_pages(page, pageblock_order);
1327 adjust_managed_page_count(page, pageblock_nr_pages);
1332 * The order of subdivision here is critical for the IO subsystem.
1333 * Please do not alter this order without good reasons and regression
1334 * testing. Specifically, as large blocks of memory are subdivided,
1335 * the order in which smaller blocks are delivered depends on the order
1336 * they're subdivided in this function. This is the primary factor
1337 * influencing the order in which pages are delivered to the IO
1338 * subsystem according to empirical testing, and this is also justified
1339 * by considering the behavior of a buddy system containing a single
1340 * large block of memory acted on by a series of small allocations.
1341 * This behavior is a critical factor in sglist merging's success.
1345 static inline void expand(struct zone *zone, struct page *page,
1346 int low, int high, struct free_area *area,
1349 unsigned long size = 1 << high;
1351 while (high > low) {
1355 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1357 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1358 debug_guardpage_enabled() &&
1359 high < debug_guardpage_minorder()) {
1361 * Mark as guard pages (or page), that will allow to
1362 * merge back to allocator when buddy will be freed.
1363 * Corresponding page table entries will not be touched,
1364 * pages will stay not present in virtual address space
1366 set_page_guard(zone, &page[size], high, migratetype);
1369 list_add(&page[size].lru, &area->free_list[migratetype]);
1371 set_page_order(&page[size], high);
1376 * This page is about to be returned from the page allocator
1378 static inline int check_new_page(struct page *page)
1380 const char *bad_reason = NULL;
1381 unsigned long bad_flags = 0;
1383 if (unlikely(atomic_read(&page->_mapcount) != -1))
1384 bad_reason = "nonzero mapcount";
1385 if (unlikely(page->mapping != NULL))
1386 bad_reason = "non-NULL mapping";
1387 if (unlikely(atomic_read(&page->_count) != 0))
1388 bad_reason = "nonzero _count";
1389 if (unlikely(page->flags & __PG_HWPOISON)) {
1390 bad_reason = "HWPoisoned (hardware-corrupted)";
1391 bad_flags = __PG_HWPOISON;
1393 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1394 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1395 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1398 if (unlikely(page->mem_cgroup))
1399 bad_reason = "page still charged to cgroup";
1401 if (unlikely(bad_reason)) {
1402 bad_page(page, bad_reason, bad_flags);
1408 static inline bool free_pages_prezeroed(bool poisoned)
1410 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1411 page_poisoning_enabled() && poisoned;
1414 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1418 bool poisoned = true;
1420 for (i = 0; i < (1 << order); i++) {
1421 struct page *p = page + i;
1422 if (unlikely(check_new_page(p)))
1425 poisoned &= page_is_poisoned(p);
1428 set_page_private(page, 0);
1429 set_page_refcounted(page);
1431 arch_alloc_page(page, order);
1432 kernel_map_pages(page, 1 << order, 1);
1433 kernel_poison_pages(page, 1 << order, 1);
1434 kasan_alloc_pages(page, order);
1436 if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO))
1437 for (i = 0; i < (1 << order); i++)
1438 clear_highpage(page + i);
1440 if (order && (gfp_flags & __GFP_COMP))
1441 prep_compound_page(page, order);
1443 set_page_owner(page, order, gfp_flags);
1446 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1447 * allocate the page. The expectation is that the caller is taking
1448 * steps that will free more memory. The caller should avoid the page
1449 * being used for !PFMEMALLOC purposes.
1451 if (alloc_flags & ALLOC_NO_WATERMARKS)
1452 set_page_pfmemalloc(page);
1454 clear_page_pfmemalloc(page);
1460 * Go through the free lists for the given migratetype and remove
1461 * the smallest available page from the freelists
1464 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1467 unsigned int current_order;
1468 struct free_area *area;
1471 /* Find a page of the appropriate size in the preferred list */
1472 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1473 area = &(zone->free_area[current_order]);
1474 page = list_first_entry_or_null(&area->free_list[migratetype],
1478 list_del(&page->lru);
1479 rmv_page_order(page);
1481 expand(zone, page, order, current_order, area, migratetype);
1482 set_pcppage_migratetype(page, migratetype);
1491 * This array describes the order lists are fallen back to when
1492 * the free lists for the desirable migrate type are depleted
1494 static int fallbacks[MIGRATE_TYPES][4] = {
1495 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1496 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1497 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1499 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1501 #ifdef CONFIG_MEMORY_ISOLATION
1502 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1507 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1510 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1513 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1514 unsigned int order) { return NULL; }
1518 * Move the free pages in a range to the free lists of the requested type.
1519 * Note that start_page and end_pages are not aligned on a pageblock
1520 * boundary. If alignment is required, use move_freepages_block()
1522 int move_freepages(struct zone *zone,
1523 struct page *start_page, struct page *end_page,
1528 int pages_moved = 0;
1530 #ifndef CONFIG_HOLES_IN_ZONE
1532 * page_zone is not safe to call in this context when
1533 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1534 * anyway as we check zone boundaries in move_freepages_block().
1535 * Remove at a later date when no bug reports exist related to
1536 * grouping pages by mobility
1538 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1541 for (page = start_page; page <= end_page;) {
1542 /* Make sure we are not inadvertently changing nodes */
1543 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1545 if (!pfn_valid_within(page_to_pfn(page))) {
1550 if (!PageBuddy(page)) {
1555 order = page_order(page);
1556 list_move(&page->lru,
1557 &zone->free_area[order].free_list[migratetype]);
1559 pages_moved += 1 << order;
1565 int move_freepages_block(struct zone *zone, struct page *page,
1568 unsigned long start_pfn, end_pfn;
1569 struct page *start_page, *end_page;
1571 start_pfn = page_to_pfn(page);
1572 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1573 start_page = pfn_to_page(start_pfn);
1574 end_page = start_page + pageblock_nr_pages - 1;
1575 end_pfn = start_pfn + pageblock_nr_pages - 1;
1577 /* Do not cross zone boundaries */
1578 if (!zone_spans_pfn(zone, start_pfn))
1580 if (!zone_spans_pfn(zone, end_pfn))
1583 return move_freepages(zone, start_page, end_page, migratetype);
1586 static void change_pageblock_range(struct page *pageblock_page,
1587 int start_order, int migratetype)
1589 int nr_pageblocks = 1 << (start_order - pageblock_order);
1591 while (nr_pageblocks--) {
1592 set_pageblock_migratetype(pageblock_page, migratetype);
1593 pageblock_page += pageblock_nr_pages;
1598 * When we are falling back to another migratetype during allocation, try to
1599 * steal extra free pages from the same pageblocks to satisfy further
1600 * allocations, instead of polluting multiple pageblocks.
1602 * If we are stealing a relatively large buddy page, it is likely there will
1603 * be more free pages in the pageblock, so try to steal them all. For
1604 * reclaimable and unmovable allocations, we steal regardless of page size,
1605 * as fragmentation caused by those allocations polluting movable pageblocks
1606 * is worse than movable allocations stealing from unmovable and reclaimable
1609 static bool can_steal_fallback(unsigned int order, int start_mt)
1612 * Leaving this order check is intended, although there is
1613 * relaxed order check in next check. The reason is that
1614 * we can actually steal whole pageblock if this condition met,
1615 * but, below check doesn't guarantee it and that is just heuristic
1616 * so could be changed anytime.
1618 if (order >= pageblock_order)
1621 if (order >= pageblock_order / 2 ||
1622 start_mt == MIGRATE_RECLAIMABLE ||
1623 start_mt == MIGRATE_UNMOVABLE ||
1624 page_group_by_mobility_disabled)
1631 * This function implements actual steal behaviour. If order is large enough,
1632 * we can steal whole pageblock. If not, we first move freepages in this
1633 * pageblock and check whether half of pages are moved or not. If half of
1634 * pages are moved, we can change migratetype of pageblock and permanently
1635 * use it's pages as requested migratetype in the future.
1637 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1640 unsigned int current_order = page_order(page);
1643 /* Take ownership for orders >= pageblock_order */
1644 if (current_order >= pageblock_order) {
1645 change_pageblock_range(page, current_order, start_type);
1649 pages = move_freepages_block(zone, page, start_type);
1651 /* Claim the whole block if over half of it is free */
1652 if (pages >= (1 << (pageblock_order-1)) ||
1653 page_group_by_mobility_disabled)
1654 set_pageblock_migratetype(page, start_type);
1658 * Check whether there is a suitable fallback freepage with requested order.
1659 * If only_stealable is true, this function returns fallback_mt only if
1660 * we can steal other freepages all together. This would help to reduce
1661 * fragmentation due to mixed migratetype pages in one pageblock.
1663 int find_suitable_fallback(struct free_area *area, unsigned int order,
1664 int migratetype, bool only_stealable, bool *can_steal)
1669 if (area->nr_free == 0)
1674 fallback_mt = fallbacks[migratetype][i];
1675 if (fallback_mt == MIGRATE_TYPES)
1678 if (list_empty(&area->free_list[fallback_mt]))
1681 if (can_steal_fallback(order, migratetype))
1684 if (!only_stealable)
1695 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1696 * there are no empty page blocks that contain a page with a suitable order
1698 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1699 unsigned int alloc_order)
1702 unsigned long max_managed, flags;
1705 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1706 * Check is race-prone but harmless.
1708 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1709 if (zone->nr_reserved_highatomic >= max_managed)
1712 spin_lock_irqsave(&zone->lock, flags);
1714 /* Recheck the nr_reserved_highatomic limit under the lock */
1715 if (zone->nr_reserved_highatomic >= max_managed)
1719 mt = get_pageblock_migratetype(page);
1720 if (mt != MIGRATE_HIGHATOMIC &&
1721 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1722 zone->nr_reserved_highatomic += pageblock_nr_pages;
1723 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1724 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1728 spin_unlock_irqrestore(&zone->lock, flags);
1732 * Used when an allocation is about to fail under memory pressure. This
1733 * potentially hurts the reliability of high-order allocations when under
1734 * intense memory pressure but failed atomic allocations should be easier
1735 * to recover from than an OOM.
1737 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1739 struct zonelist *zonelist = ac->zonelist;
1740 unsigned long flags;
1746 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1748 /* Preserve at least one pageblock */
1749 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1752 spin_lock_irqsave(&zone->lock, flags);
1753 for (order = 0; order < MAX_ORDER; order++) {
1754 struct free_area *area = &(zone->free_area[order]);
1756 page = list_first_entry_or_null(
1757 &area->free_list[MIGRATE_HIGHATOMIC],
1763 * It should never happen but changes to locking could
1764 * inadvertently allow a per-cpu drain to add pages
1765 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1766 * and watch for underflows.
1768 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1769 zone->nr_reserved_highatomic);
1772 * Convert to ac->migratetype and avoid the normal
1773 * pageblock stealing heuristics. Minimally, the caller
1774 * is doing the work and needs the pages. More
1775 * importantly, if the block was always converted to
1776 * MIGRATE_UNMOVABLE or another type then the number
1777 * of pageblocks that cannot be completely freed
1780 set_pageblock_migratetype(page, ac->migratetype);
1781 move_freepages_block(zone, page, ac->migratetype);
1782 spin_unlock_irqrestore(&zone->lock, flags);
1785 spin_unlock_irqrestore(&zone->lock, flags);
1789 /* Remove an element from the buddy allocator from the fallback list */
1790 static inline struct page *
1791 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1793 struct free_area *area;
1794 unsigned int current_order;
1799 /* Find the largest possible block of pages in the other list */
1800 for (current_order = MAX_ORDER-1;
1801 current_order >= order && current_order <= MAX_ORDER-1;
1803 area = &(zone->free_area[current_order]);
1804 fallback_mt = find_suitable_fallback(area, current_order,
1805 start_migratetype, false, &can_steal);
1806 if (fallback_mt == -1)
1809 page = list_first_entry(&area->free_list[fallback_mt],
1812 steal_suitable_fallback(zone, page, start_migratetype);
1814 /* Remove the page from the freelists */
1816 list_del(&page->lru);
1817 rmv_page_order(page);
1819 expand(zone, page, order, current_order, area,
1822 * The pcppage_migratetype may differ from pageblock's
1823 * migratetype depending on the decisions in
1824 * find_suitable_fallback(). This is OK as long as it does not
1825 * differ for MIGRATE_CMA pageblocks. Those can be used as
1826 * fallback only via special __rmqueue_cma_fallback() function
1828 set_pcppage_migratetype(page, start_migratetype);
1830 trace_mm_page_alloc_extfrag(page, order, current_order,
1831 start_migratetype, fallback_mt);
1840 * Do the hard work of removing an element from the buddy allocator.
1841 * Call me with the zone->lock already held.
1843 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1848 page = __rmqueue_smallest(zone, order, migratetype);
1849 if (unlikely(!page)) {
1850 if (migratetype == MIGRATE_MOVABLE)
1851 page = __rmqueue_cma_fallback(zone, order);
1854 page = __rmqueue_fallback(zone, order, migratetype);
1857 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1862 * Obtain a specified number of elements from the buddy allocator, all under
1863 * a single hold of the lock, for efficiency. Add them to the supplied list.
1864 * Returns the number of new pages which were placed at *list.
1866 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1867 unsigned long count, struct list_head *list,
1868 int migratetype, bool cold)
1872 spin_lock(&zone->lock);
1873 for (i = 0; i < count; ++i) {
1874 struct page *page = __rmqueue(zone, order, migratetype);
1875 if (unlikely(page == NULL))
1879 * Split buddy pages returned by expand() are received here
1880 * in physical page order. The page is added to the callers and
1881 * list and the list head then moves forward. From the callers
1882 * perspective, the linked list is ordered by page number in
1883 * some conditions. This is useful for IO devices that can
1884 * merge IO requests if the physical pages are ordered
1888 list_add(&page->lru, list);
1890 list_add_tail(&page->lru, list);
1892 if (is_migrate_cma(get_pcppage_migratetype(page)))
1893 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1896 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1897 spin_unlock(&zone->lock);
1903 * Called from the vmstat counter updater to drain pagesets of this
1904 * currently executing processor on remote nodes after they have
1907 * Note that this function must be called with the thread pinned to
1908 * a single processor.
1910 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1912 unsigned long flags;
1913 int to_drain, batch;
1915 local_irq_save(flags);
1916 batch = READ_ONCE(pcp->batch);
1917 to_drain = min(pcp->count, batch);
1919 free_pcppages_bulk(zone, to_drain, pcp);
1920 pcp->count -= to_drain;
1922 local_irq_restore(flags);
1927 * Drain pcplists of the indicated processor and zone.
1929 * The processor must either be the current processor and the
1930 * thread pinned to the current processor or a processor that
1933 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1935 unsigned long flags;
1936 struct per_cpu_pageset *pset;
1937 struct per_cpu_pages *pcp;
1939 local_irq_save(flags);
1940 pset = per_cpu_ptr(zone->pageset, cpu);
1944 free_pcppages_bulk(zone, pcp->count, pcp);
1947 local_irq_restore(flags);
1951 * Drain pcplists of all zones on the indicated processor.
1953 * The processor must either be the current processor and the
1954 * thread pinned to the current processor or a processor that
1957 static void drain_pages(unsigned int cpu)
1961 for_each_populated_zone(zone) {
1962 drain_pages_zone(cpu, zone);
1967 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1969 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1970 * the single zone's pages.
1972 void drain_local_pages(struct zone *zone)
1974 int cpu = smp_processor_id();
1977 drain_pages_zone(cpu, zone);
1983 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1985 * When zone parameter is non-NULL, spill just the single zone's pages.
1987 * Note that this code is protected against sending an IPI to an offline
1988 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1989 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1990 * nothing keeps CPUs from showing up after we populated the cpumask and
1991 * before the call to on_each_cpu_mask().
1993 void drain_all_pages(struct zone *zone)
1998 * Allocate in the BSS so we wont require allocation in
1999 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2001 static cpumask_t cpus_with_pcps;
2004 * We don't care about racing with CPU hotplug event
2005 * as offline notification will cause the notified
2006 * cpu to drain that CPU pcps and on_each_cpu_mask
2007 * disables preemption as part of its processing
2009 for_each_online_cpu(cpu) {
2010 struct per_cpu_pageset *pcp;
2012 bool has_pcps = false;
2015 pcp = per_cpu_ptr(zone->pageset, cpu);
2019 for_each_populated_zone(z) {
2020 pcp = per_cpu_ptr(z->pageset, cpu);
2021 if (pcp->pcp.count) {
2029 cpumask_set_cpu(cpu, &cpus_with_pcps);
2031 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2033 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2037 #ifdef CONFIG_HIBERNATION
2039 void mark_free_pages(struct zone *zone)
2041 unsigned long pfn, max_zone_pfn;
2042 unsigned long flags;
2043 unsigned int order, t;
2046 if (zone_is_empty(zone))
2049 spin_lock_irqsave(&zone->lock, flags);
2051 max_zone_pfn = zone_end_pfn(zone);
2052 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2053 if (pfn_valid(pfn)) {
2054 page = pfn_to_page(pfn);
2055 if (!swsusp_page_is_forbidden(page))
2056 swsusp_unset_page_free(page);
2059 for_each_migratetype_order(order, t) {
2060 list_for_each_entry(page,
2061 &zone->free_area[order].free_list[t], lru) {
2064 pfn = page_to_pfn(page);
2065 for (i = 0; i < (1UL << order); i++)
2066 swsusp_set_page_free(pfn_to_page(pfn + i));
2069 spin_unlock_irqrestore(&zone->lock, flags);
2071 #endif /* CONFIG_PM */
2074 * Free a 0-order page
2075 * cold == true ? free a cold page : free a hot page
2077 void free_hot_cold_page(struct page *page, bool cold)
2079 struct zone *zone = page_zone(page);
2080 struct per_cpu_pages *pcp;
2081 unsigned long flags;
2082 unsigned long pfn = page_to_pfn(page);
2085 if (!free_pages_prepare(page, 0))
2088 migratetype = get_pfnblock_migratetype(page, pfn);
2089 set_pcppage_migratetype(page, migratetype);
2090 local_irq_save(flags);
2091 __count_vm_event(PGFREE);
2094 * We only track unmovable, reclaimable and movable on pcp lists.
2095 * Free ISOLATE pages back to the allocator because they are being
2096 * offlined but treat RESERVE as movable pages so we can get those
2097 * areas back if necessary. Otherwise, we may have to free
2098 * excessively into the page allocator
2100 if (migratetype >= MIGRATE_PCPTYPES) {
2101 if (unlikely(is_migrate_isolate(migratetype))) {
2102 free_one_page(zone, page, pfn, 0, migratetype);
2105 migratetype = MIGRATE_MOVABLE;
2108 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2110 list_add(&page->lru, &pcp->lists[migratetype]);
2112 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2114 if (pcp->count >= pcp->high) {
2115 unsigned long batch = READ_ONCE(pcp->batch);
2116 free_pcppages_bulk(zone, batch, pcp);
2117 pcp->count -= batch;
2121 local_irq_restore(flags);
2125 * Free a list of 0-order pages
2127 void free_hot_cold_page_list(struct list_head *list, bool cold)
2129 struct page *page, *next;
2131 list_for_each_entry_safe(page, next, list, lru) {
2132 trace_mm_page_free_batched(page, cold);
2133 free_hot_cold_page(page, cold);
2138 * split_page takes a non-compound higher-order page, and splits it into
2139 * n (1<<order) sub-pages: page[0..n]
2140 * Each sub-page must be freed individually.
2142 * Note: this is probably too low level an operation for use in drivers.
2143 * Please consult with lkml before using this in your driver.
2145 void split_page(struct page *page, unsigned int order)
2150 VM_BUG_ON_PAGE(PageCompound(page), page);
2151 VM_BUG_ON_PAGE(!page_count(page), page);
2153 #ifdef CONFIG_KMEMCHECK
2155 * Split shadow pages too, because free(page[0]) would
2156 * otherwise free the whole shadow.
2158 if (kmemcheck_page_is_tracked(page))
2159 split_page(virt_to_page(page[0].shadow), order);
2162 gfp_mask = get_page_owner_gfp(page);
2163 set_page_owner(page, 0, gfp_mask);
2164 for (i = 1; i < (1 << order); i++) {
2165 set_page_refcounted(page + i);
2166 set_page_owner(page + i, 0, gfp_mask);
2169 EXPORT_SYMBOL_GPL(split_page);
2171 int __isolate_free_page(struct page *page, unsigned int order)
2173 unsigned long watermark;
2177 BUG_ON(!PageBuddy(page));
2179 zone = page_zone(page);
2180 mt = get_pageblock_migratetype(page);
2182 if (!is_migrate_isolate(mt)) {
2183 /* Obey watermarks as if the page was being allocated */
2184 watermark = low_wmark_pages(zone) + (1 << order);
2185 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2188 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2191 /* Remove page from free list */
2192 list_del(&page->lru);
2193 zone->free_area[order].nr_free--;
2194 rmv_page_order(page);
2196 set_page_owner(page, order, __GFP_MOVABLE);
2198 /* Set the pageblock if the isolated page is at least a pageblock */
2199 if (order >= pageblock_order - 1) {
2200 struct page *endpage = page + (1 << order) - 1;
2201 for (; page < endpage; page += pageblock_nr_pages) {
2202 int mt = get_pageblock_migratetype(page);
2203 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2204 set_pageblock_migratetype(page,
2210 return 1UL << order;
2214 * Similar to split_page except the page is already free. As this is only
2215 * being used for migration, the migratetype of the block also changes.
2216 * As this is called with interrupts disabled, the caller is responsible
2217 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2220 * Note: this is probably too low level an operation for use in drivers.
2221 * Please consult with lkml before using this in your driver.
2223 int split_free_page(struct page *page)
2228 order = page_order(page);
2230 nr_pages = __isolate_free_page(page, order);
2234 /* Split into individual pages */
2235 set_page_refcounted(page);
2236 split_page(page, order);
2241 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2244 struct page *buffered_rmqueue(struct zone *preferred_zone,
2245 struct zone *zone, unsigned int order,
2246 gfp_t gfp_flags, int alloc_flags, int migratetype)
2248 unsigned long flags;
2250 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2252 if (likely(order == 0)) {
2253 struct per_cpu_pages *pcp;
2254 struct list_head *list;
2256 local_irq_save(flags);
2257 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2258 list = &pcp->lists[migratetype];
2259 if (list_empty(list)) {
2260 pcp->count += rmqueue_bulk(zone, 0,
2263 if (unlikely(list_empty(list)))
2268 page = list_last_entry(list, struct page, lru);
2270 page = list_first_entry(list, struct page, lru);
2272 list_del(&page->lru);
2275 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2277 * __GFP_NOFAIL is not to be used in new code.
2279 * All __GFP_NOFAIL callers should be fixed so that they
2280 * properly detect and handle allocation failures.
2282 * We most definitely don't want callers attempting to
2283 * allocate greater than order-1 page units with
2286 WARN_ON_ONCE(order > 1);
2288 spin_lock_irqsave(&zone->lock, flags);
2291 if (alloc_flags & ALLOC_HARDER) {
2292 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2294 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2297 page = __rmqueue(zone, order, migratetype);
2298 spin_unlock(&zone->lock);
2301 __mod_zone_freepage_state(zone, -(1 << order),
2302 get_pcppage_migratetype(page));
2305 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2306 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2307 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2308 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2310 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2311 zone_statistics(preferred_zone, zone, gfp_flags);
2312 local_irq_restore(flags);
2314 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2318 local_irq_restore(flags);
2322 #ifdef CONFIG_FAIL_PAGE_ALLOC
2325 struct fault_attr attr;
2327 bool ignore_gfp_highmem;
2328 bool ignore_gfp_reclaim;
2330 } fail_page_alloc = {
2331 .attr = FAULT_ATTR_INITIALIZER,
2332 .ignore_gfp_reclaim = true,
2333 .ignore_gfp_highmem = true,
2337 static int __init setup_fail_page_alloc(char *str)
2339 return setup_fault_attr(&fail_page_alloc.attr, str);
2341 __setup("fail_page_alloc=", setup_fail_page_alloc);
2343 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2345 if (order < fail_page_alloc.min_order)
2347 if (gfp_mask & __GFP_NOFAIL)
2349 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2351 if (fail_page_alloc.ignore_gfp_reclaim &&
2352 (gfp_mask & __GFP_DIRECT_RECLAIM))
2355 return should_fail(&fail_page_alloc.attr, 1 << order);
2358 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2360 static int __init fail_page_alloc_debugfs(void)
2362 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2365 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2366 &fail_page_alloc.attr);
2368 return PTR_ERR(dir);
2370 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2371 &fail_page_alloc.ignore_gfp_reclaim))
2373 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2374 &fail_page_alloc.ignore_gfp_highmem))
2376 if (!debugfs_create_u32("min-order", mode, dir,
2377 &fail_page_alloc.min_order))
2382 debugfs_remove_recursive(dir);
2387 late_initcall(fail_page_alloc_debugfs);
2389 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2391 #else /* CONFIG_FAIL_PAGE_ALLOC */
2393 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2398 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2401 * Return true if free base pages are above 'mark'. For high-order checks it
2402 * will return true of the order-0 watermark is reached and there is at least
2403 * one free page of a suitable size. Checking now avoids taking the zone lock
2404 * to check in the allocation paths if no pages are free.
2406 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2407 unsigned long mark, int classzone_idx, int alloc_flags,
2412 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2414 /* free_pages may go negative - that's OK */
2415 free_pages -= (1 << order) - 1;
2417 if (alloc_flags & ALLOC_HIGH)
2421 * If the caller does not have rights to ALLOC_HARDER then subtract
2422 * the high-atomic reserves. This will over-estimate the size of the
2423 * atomic reserve but it avoids a search.
2425 if (likely(!alloc_harder))
2426 free_pages -= z->nr_reserved_highatomic;
2431 /* If allocation can't use CMA areas don't use free CMA pages */
2432 if (!(alloc_flags & ALLOC_CMA))
2433 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2437 * Check watermarks for an order-0 allocation request. If these
2438 * are not met, then a high-order request also cannot go ahead
2439 * even if a suitable page happened to be free.
2441 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2444 /* If this is an order-0 request then the watermark is fine */
2448 /* For a high-order request, check at least one suitable page is free */
2449 for (o = order; o < MAX_ORDER; o++) {
2450 struct free_area *area = &z->free_area[o];
2459 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2460 if (!list_empty(&area->free_list[mt]))
2465 if ((alloc_flags & ALLOC_CMA) &&
2466 !list_empty(&area->free_list[MIGRATE_CMA])) {
2474 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2475 int classzone_idx, int alloc_flags)
2477 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2478 zone_page_state(z, NR_FREE_PAGES));
2481 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2482 unsigned long mark, int classzone_idx)
2484 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2486 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2487 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2489 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2494 static bool zone_local(struct zone *local_zone, struct zone *zone)
2496 return local_zone->node == zone->node;
2499 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2501 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2504 #else /* CONFIG_NUMA */
2505 static bool zone_local(struct zone *local_zone, struct zone *zone)
2510 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2514 #endif /* CONFIG_NUMA */
2516 static void reset_alloc_batches(struct zone *preferred_zone)
2518 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2521 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2522 high_wmark_pages(zone) - low_wmark_pages(zone) -
2523 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2524 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2525 } while (zone++ != preferred_zone);
2529 * get_page_from_freelist goes through the zonelist trying to allocate
2532 static struct page *
2533 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2534 const struct alloc_context *ac)
2536 struct zonelist *zonelist = ac->zonelist;
2538 struct page *page = NULL;
2540 int nr_fair_skipped = 0;
2541 bool zonelist_rescan;
2544 zonelist_rescan = false;
2547 * Scan zonelist, looking for a zone with enough free.
2548 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2550 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2554 if (cpusets_enabled() &&
2555 (alloc_flags & ALLOC_CPUSET) &&
2556 !cpuset_zone_allowed(zone, gfp_mask))
2559 * Distribute pages in proportion to the individual
2560 * zone size to ensure fair page aging. The zone a
2561 * page was allocated in should have no effect on the
2562 * time the page has in memory before being reclaimed.
2564 if (alloc_flags & ALLOC_FAIR) {
2565 if (!zone_local(ac->preferred_zone, zone))
2567 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2573 * When allocating a page cache page for writing, we
2574 * want to get it from a zone that is within its dirty
2575 * limit, such that no single zone holds more than its
2576 * proportional share of globally allowed dirty pages.
2577 * The dirty limits take into account the zone's
2578 * lowmem reserves and high watermark so that kswapd
2579 * should be able to balance it without having to
2580 * write pages from its LRU list.
2582 * This may look like it could increase pressure on
2583 * lower zones by failing allocations in higher zones
2584 * before they are full. But the pages that do spill
2585 * over are limited as the lower zones are protected
2586 * by this very same mechanism. It should not become
2587 * a practical burden to them.
2589 * XXX: For now, allow allocations to potentially
2590 * exceed the per-zone dirty limit in the slowpath
2591 * (spread_dirty_pages unset) before going into reclaim,
2592 * which is important when on a NUMA setup the allowed
2593 * zones are together not big enough to reach the
2594 * global limit. The proper fix for these situations
2595 * will require awareness of zones in the
2596 * dirty-throttling and the flusher threads.
2598 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2601 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2602 if (!zone_watermark_ok(zone, order, mark,
2603 ac->classzone_idx, alloc_flags)) {
2606 /* Checked here to keep the fast path fast */
2607 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2608 if (alloc_flags & ALLOC_NO_WATERMARKS)
2611 if (zone_reclaim_mode == 0 ||
2612 !zone_allows_reclaim(ac->preferred_zone, zone))
2615 ret = zone_reclaim(zone, gfp_mask, order);
2617 case ZONE_RECLAIM_NOSCAN:
2620 case ZONE_RECLAIM_FULL:
2621 /* scanned but unreclaimable */
2624 /* did we reclaim enough */
2625 if (zone_watermark_ok(zone, order, mark,
2626 ac->classzone_idx, alloc_flags))
2634 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2635 gfp_mask, alloc_flags, ac->migratetype);
2637 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2641 * If this is a high-order atomic allocation then check
2642 * if the pageblock should be reserved for the future
2644 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2645 reserve_highatomic_pageblock(page, zone, order);
2652 * The first pass makes sure allocations are spread fairly within the
2653 * local node. However, the local node might have free pages left
2654 * after the fairness batches are exhausted, and remote zones haven't
2655 * even been considered yet. Try once more without fairness, and
2656 * include remote zones now, before entering the slowpath and waking
2657 * kswapd: prefer spilling to a remote zone over swapping locally.
2659 if (alloc_flags & ALLOC_FAIR) {
2660 alloc_flags &= ~ALLOC_FAIR;
2661 if (nr_fair_skipped) {
2662 zonelist_rescan = true;
2663 reset_alloc_batches(ac->preferred_zone);
2665 if (nr_online_nodes > 1)
2666 zonelist_rescan = true;
2669 if (zonelist_rescan)
2676 * Large machines with many possible nodes should not always dump per-node
2677 * meminfo in irq context.
2679 static inline bool should_suppress_show_mem(void)
2684 ret = in_interrupt();
2689 static DEFINE_RATELIMIT_STATE(nopage_rs,
2690 DEFAULT_RATELIMIT_INTERVAL,
2691 DEFAULT_RATELIMIT_BURST);
2693 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2695 unsigned int filter = SHOW_MEM_FILTER_NODES;
2697 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2698 debug_guardpage_minorder() > 0)
2702 * This documents exceptions given to allocations in certain
2703 * contexts that are allowed to allocate outside current's set
2706 if (!(gfp_mask & __GFP_NOMEMALLOC))
2707 if (test_thread_flag(TIF_MEMDIE) ||
2708 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2709 filter &= ~SHOW_MEM_FILTER_NODES;
2710 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2711 filter &= ~SHOW_MEM_FILTER_NODES;
2714 struct va_format vaf;
2717 va_start(args, fmt);
2722 pr_warn("%pV", &vaf);
2727 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
2728 current->comm, order, gfp_mask, &gfp_mask);
2730 if (!should_suppress_show_mem())
2734 static inline struct page *
2735 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2736 const struct alloc_context *ac, unsigned long *did_some_progress)
2738 struct oom_control oc = {
2739 .zonelist = ac->zonelist,
2740 .nodemask = ac->nodemask,
2741 .gfp_mask = gfp_mask,
2746 *did_some_progress = 0;
2749 * Acquire the oom lock. If that fails, somebody else is
2750 * making progress for us.
2752 if (!mutex_trylock(&oom_lock)) {
2753 *did_some_progress = 1;
2754 schedule_timeout_uninterruptible(1);
2759 * Go through the zonelist yet one more time, keep very high watermark
2760 * here, this is only to catch a parallel oom killing, we must fail if
2761 * we're still under heavy pressure.
2763 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2764 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2768 if (!(gfp_mask & __GFP_NOFAIL)) {
2769 /* Coredumps can quickly deplete all memory reserves */
2770 if (current->flags & PF_DUMPCORE)
2772 /* The OOM killer will not help higher order allocs */
2773 if (order > PAGE_ALLOC_COSTLY_ORDER)
2775 /* The OOM killer does not needlessly kill tasks for lowmem */
2776 if (ac->high_zoneidx < ZONE_NORMAL)
2778 /* The OOM killer does not compensate for IO-less reclaim */
2779 if (!(gfp_mask & __GFP_FS)) {
2781 * XXX: Page reclaim didn't yield anything,
2782 * and the OOM killer can't be invoked, but
2783 * keep looping as per tradition.
2785 *did_some_progress = 1;
2788 if (pm_suspended_storage())
2790 /* The OOM killer may not free memory on a specific node */
2791 if (gfp_mask & __GFP_THISNODE)
2794 /* Exhausted what can be done so it's blamo time */
2795 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
2796 *did_some_progress = 1;
2798 if (gfp_mask & __GFP_NOFAIL) {
2799 page = get_page_from_freelist(gfp_mask, order,
2800 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
2802 * fallback to ignore cpuset restriction if our nodes
2806 page = get_page_from_freelist(gfp_mask, order,
2807 ALLOC_NO_WATERMARKS, ac);
2811 mutex_unlock(&oom_lock);
2815 #ifdef CONFIG_COMPACTION
2816 /* Try memory compaction for high-order allocations before reclaim */
2817 static struct page *
2818 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2819 int alloc_flags, const struct alloc_context *ac,
2820 enum migrate_mode mode, int *contended_compaction,
2821 bool *deferred_compaction)
2823 unsigned long compact_result;
2829 current->flags |= PF_MEMALLOC;
2830 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2831 mode, contended_compaction);
2832 current->flags &= ~PF_MEMALLOC;
2834 switch (compact_result) {
2835 case COMPACT_DEFERRED:
2836 *deferred_compaction = true;
2838 case COMPACT_SKIPPED:
2845 * At least in one zone compaction wasn't deferred or skipped, so let's
2846 * count a compaction stall
2848 count_vm_event(COMPACTSTALL);
2850 page = get_page_from_freelist(gfp_mask, order,
2851 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2854 struct zone *zone = page_zone(page);
2856 zone->compact_blockskip_flush = false;
2857 compaction_defer_reset(zone, order, true);
2858 count_vm_event(COMPACTSUCCESS);
2863 * It's bad if compaction run occurs and fails. The most likely reason
2864 * is that pages exist, but not enough to satisfy watermarks.
2866 count_vm_event(COMPACTFAIL);
2873 static inline struct page *
2874 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2875 int alloc_flags, const struct alloc_context *ac,
2876 enum migrate_mode mode, int *contended_compaction,
2877 bool *deferred_compaction)
2881 #endif /* CONFIG_COMPACTION */
2883 /* Perform direct synchronous page reclaim */
2885 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2886 const struct alloc_context *ac)
2888 struct reclaim_state reclaim_state;
2893 /* We now go into synchronous reclaim */
2894 cpuset_memory_pressure_bump();
2895 current->flags |= PF_MEMALLOC;
2896 lockdep_set_current_reclaim_state(gfp_mask);
2897 reclaim_state.reclaimed_slab = 0;
2898 current->reclaim_state = &reclaim_state;
2900 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2903 current->reclaim_state = NULL;
2904 lockdep_clear_current_reclaim_state();
2905 current->flags &= ~PF_MEMALLOC;
2912 /* The really slow allocator path where we enter direct reclaim */
2913 static inline struct page *
2914 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2915 int alloc_flags, const struct alloc_context *ac,
2916 unsigned long *did_some_progress)
2918 struct page *page = NULL;
2919 bool drained = false;
2921 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2922 if (unlikely(!(*did_some_progress)))
2926 page = get_page_from_freelist(gfp_mask, order,
2927 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2930 * If an allocation failed after direct reclaim, it could be because
2931 * pages are pinned on the per-cpu lists or in high alloc reserves.
2932 * Shrink them them and try again
2934 if (!page && !drained) {
2935 unreserve_highatomic_pageblock(ac);
2936 drain_all_pages(NULL);
2944 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2949 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2950 ac->high_zoneidx, ac->nodemask)
2951 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2955 gfp_to_alloc_flags(gfp_t gfp_mask)
2957 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2959 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2960 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2963 * The caller may dip into page reserves a bit more if the caller
2964 * cannot run direct reclaim, or if the caller has realtime scheduling
2965 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2966 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
2968 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2970 if (gfp_mask & __GFP_ATOMIC) {
2972 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2973 * if it can't schedule.
2975 if (!(gfp_mask & __GFP_NOMEMALLOC))
2976 alloc_flags |= ALLOC_HARDER;
2978 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2979 * comment for __cpuset_node_allowed().
2981 alloc_flags &= ~ALLOC_CPUSET;
2982 } else if (unlikely(rt_task(current)) && !in_interrupt())
2983 alloc_flags |= ALLOC_HARDER;
2985 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2986 if (gfp_mask & __GFP_MEMALLOC)
2987 alloc_flags |= ALLOC_NO_WATERMARKS;
2988 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2989 alloc_flags |= ALLOC_NO_WATERMARKS;
2990 else if (!in_interrupt() &&
2991 ((current->flags & PF_MEMALLOC) ||
2992 unlikely(test_thread_flag(TIF_MEMDIE))))
2993 alloc_flags |= ALLOC_NO_WATERMARKS;
2996 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2997 alloc_flags |= ALLOC_CMA;
3002 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3004 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
3007 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
3009 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
3012 static inline struct page *
3013 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3014 struct alloc_context *ac)
3016 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3017 struct page *page = NULL;
3019 unsigned long pages_reclaimed = 0;
3020 unsigned long did_some_progress;
3021 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3022 bool deferred_compaction = false;
3023 int contended_compaction = COMPACT_CONTENDED_NONE;
3026 * In the slowpath, we sanity check order to avoid ever trying to
3027 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3028 * be using allocators in order of preference for an area that is
3031 if (order >= MAX_ORDER) {
3032 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3037 * We also sanity check to catch abuse of atomic reserves being used by
3038 * callers that are not in atomic context.
3040 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3041 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3042 gfp_mask &= ~__GFP_ATOMIC;
3045 * If this allocation cannot block and it is for a specific node, then
3046 * fail early. There's no need to wakeup kswapd or retry for a
3047 * speculative node-specific allocation.
3049 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
3053 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3054 wake_all_kswapds(order, ac);
3057 * OK, we're below the kswapd watermark and have kicked background
3058 * reclaim. Now things get more complex, so set up alloc_flags according
3059 * to how we want to proceed.
3061 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3064 * Find the true preferred zone if the allocation is unconstrained by
3067 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3068 struct zoneref *preferred_zoneref;
3069 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3070 ac->high_zoneidx, NULL, &ac->preferred_zone);
3071 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3074 /* This is the last chance, in general, before the goto nopage. */
3075 page = get_page_from_freelist(gfp_mask, order,
3076 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3080 /* Allocate without watermarks if the context allows */
3081 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3083 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3084 * the allocation is high priority and these type of
3085 * allocations are system rather than user orientated
3087 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3088 page = get_page_from_freelist(gfp_mask, order,
3089 ALLOC_NO_WATERMARKS, ac);
3094 /* Caller is not willing to reclaim, we can't balance anything */
3095 if (!can_direct_reclaim) {
3097 * All existing users of the __GFP_NOFAIL are blockable, so warn
3098 * of any new users that actually allow this type of allocation
3101 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3105 /* Avoid recursion of direct reclaim */
3106 if (current->flags & PF_MEMALLOC) {
3108 * __GFP_NOFAIL request from this context is rather bizarre
3109 * because we cannot reclaim anything and only can loop waiting
3110 * for somebody to do a work for us.
3112 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3119 /* Avoid allocations with no watermarks from looping endlessly */
3120 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3124 * Try direct compaction. The first pass is asynchronous. Subsequent
3125 * attempts after direct reclaim are synchronous
3127 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3129 &contended_compaction,
3130 &deferred_compaction);
3134 /* Checks for THP-specific high-order allocations */
3135 if (is_thp_gfp_mask(gfp_mask)) {
3137 * If compaction is deferred for high-order allocations, it is
3138 * because sync compaction recently failed. If this is the case
3139 * and the caller requested a THP allocation, we do not want
3140 * to heavily disrupt the system, so we fail the allocation
3141 * instead of entering direct reclaim.
3143 if (deferred_compaction)
3147 * In all zones where compaction was attempted (and not
3148 * deferred or skipped), lock contention has been detected.
3149 * For THP allocation we do not want to disrupt the others
3150 * so we fallback to base pages instead.
3152 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3156 * If compaction was aborted due to need_resched(), we do not
3157 * want to further increase allocation latency, unless it is
3158 * khugepaged trying to collapse.
3160 if (contended_compaction == COMPACT_CONTENDED_SCHED
3161 && !(current->flags & PF_KTHREAD))
3166 * It can become very expensive to allocate transparent hugepages at
3167 * fault, so use asynchronous memory compaction for THP unless it is
3168 * khugepaged trying to collapse.
3170 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3171 migration_mode = MIGRATE_SYNC_LIGHT;
3173 /* Try direct reclaim and then allocating */
3174 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3175 &did_some_progress);
3179 /* Do not loop if specifically requested */
3180 if (gfp_mask & __GFP_NORETRY)
3183 /* Keep reclaiming pages as long as there is reasonable progress */
3184 pages_reclaimed += did_some_progress;
3185 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3186 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3187 /* Wait for some write requests to complete then retry */
3188 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3192 /* Reclaim has failed us, start killing things */
3193 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3197 /* Retry as long as the OOM killer is making progress */
3198 if (did_some_progress)
3203 * High-order allocations do not necessarily loop after
3204 * direct reclaim and reclaim/compaction depends on compaction
3205 * being called after reclaim so call directly if necessary
3207 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3209 &contended_compaction,
3210 &deferred_compaction);
3214 warn_alloc_failed(gfp_mask, order, NULL);
3220 * This is the 'heart' of the zoned buddy allocator.
3223 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3224 struct zonelist *zonelist, nodemask_t *nodemask)
3226 struct zoneref *preferred_zoneref;
3227 struct page *page = NULL;
3228 unsigned int cpuset_mems_cookie;
3229 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3230 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3231 struct alloc_context ac = {
3232 .high_zoneidx = gfp_zone(gfp_mask),
3233 .nodemask = nodemask,
3234 .migratetype = gfpflags_to_migratetype(gfp_mask),
3237 gfp_mask &= gfp_allowed_mask;
3239 lockdep_trace_alloc(gfp_mask);
3241 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3243 if (should_fail_alloc_page(gfp_mask, order))
3247 * Check the zones suitable for the gfp_mask contain at least one
3248 * valid zone. It's possible to have an empty zonelist as a result
3249 * of __GFP_THISNODE and a memoryless node
3251 if (unlikely(!zonelist->_zonerefs->zone))
3254 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3255 alloc_flags |= ALLOC_CMA;
3258 cpuset_mems_cookie = read_mems_allowed_begin();
3260 /* We set it here, as __alloc_pages_slowpath might have changed it */
3261 ac.zonelist = zonelist;
3263 /* Dirty zone balancing only done in the fast path */
3264 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3266 /* The preferred zone is used for statistics later */
3267 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3268 ac.nodemask ? : &cpuset_current_mems_allowed,
3269 &ac.preferred_zone);
3270 if (!ac.preferred_zone)
3272 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3274 /* First allocation attempt */
3275 alloc_mask = gfp_mask|__GFP_HARDWALL;
3276 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3277 if (unlikely(!page)) {
3279 * Runtime PM, block IO and its error handling path
3280 * can deadlock because I/O on the device might not
3283 alloc_mask = memalloc_noio_flags(gfp_mask);
3284 ac.spread_dirty_pages = false;
3286 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3289 if (kmemcheck_enabled && page)
3290 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3292 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3296 * When updating a task's mems_allowed, it is possible to race with
3297 * parallel threads in such a way that an allocation can fail while
3298 * the mask is being updated. If a page allocation is about to fail,
3299 * check if the cpuset changed during allocation and if so, retry.
3301 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3306 EXPORT_SYMBOL(__alloc_pages_nodemask);
3309 * Common helper functions.
3311 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3316 * __get_free_pages() returns a 32-bit address, which cannot represent
3319 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3321 page = alloc_pages(gfp_mask, order);
3324 return (unsigned long) page_address(page);
3326 EXPORT_SYMBOL(__get_free_pages);
3328 unsigned long get_zeroed_page(gfp_t gfp_mask)
3330 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3332 EXPORT_SYMBOL(get_zeroed_page);
3334 void __free_pages(struct page *page, unsigned int order)
3336 if (put_page_testzero(page)) {
3338 free_hot_cold_page(page, false);
3340 __free_pages_ok(page, order);
3344 EXPORT_SYMBOL(__free_pages);
3346 void free_pages(unsigned long addr, unsigned int order)
3349 VM_BUG_ON(!virt_addr_valid((void *)addr));
3350 __free_pages(virt_to_page((void *)addr), order);
3354 EXPORT_SYMBOL(free_pages);
3358 * An arbitrary-length arbitrary-offset area of memory which resides
3359 * within a 0 or higher order page. Multiple fragments within that page
3360 * are individually refcounted, in the page's reference counter.
3362 * The page_frag functions below provide a simple allocation framework for
3363 * page fragments. This is used by the network stack and network device
3364 * drivers to provide a backing region of memory for use as either an
3365 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3367 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3370 struct page *page = NULL;
3371 gfp_t gfp = gfp_mask;
3373 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3374 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3376 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3377 PAGE_FRAG_CACHE_MAX_ORDER);
3378 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3380 if (unlikely(!page))
3381 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3383 nc->va = page ? page_address(page) : NULL;
3388 void *__alloc_page_frag(struct page_frag_cache *nc,
3389 unsigned int fragsz, gfp_t gfp_mask)
3391 unsigned int size = PAGE_SIZE;
3395 if (unlikely(!nc->va)) {
3397 page = __page_frag_refill(nc, gfp_mask);
3401 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3402 /* if size can vary use size else just use PAGE_SIZE */
3405 /* Even if we own the page, we do not use atomic_set().
3406 * This would break get_page_unless_zero() users.
3408 atomic_add(size - 1, &page->_count);
3410 /* reset page count bias and offset to start of new frag */
3411 nc->pfmemalloc = page_is_pfmemalloc(page);
3412 nc->pagecnt_bias = size;
3416 offset = nc->offset - fragsz;
3417 if (unlikely(offset < 0)) {
3418 page = virt_to_page(nc->va);
3420 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3423 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3424 /* if size can vary use size else just use PAGE_SIZE */
3427 /* OK, page count is 0, we can safely set it */
3428 atomic_set(&page->_count, size);
3430 /* reset page count bias and offset to start of new frag */
3431 nc->pagecnt_bias = size;
3432 offset = size - fragsz;
3436 nc->offset = offset;
3438 return nc->va + offset;
3440 EXPORT_SYMBOL(__alloc_page_frag);
3443 * Frees a page fragment allocated out of either a compound or order 0 page.
3445 void __free_page_frag(void *addr)
3447 struct page *page = virt_to_head_page(addr);
3449 if (unlikely(put_page_testzero(page)))
3450 __free_pages_ok(page, compound_order(page));
3452 EXPORT_SYMBOL(__free_page_frag);
3455 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3456 * of the current memory cgroup if __GFP_ACCOUNT is set, other than that it is
3457 * equivalent to alloc_pages.
3459 * It should be used when the caller would like to use kmalloc, but since the
3460 * allocation is large, it has to fall back to the page allocator.
3462 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3466 page = alloc_pages(gfp_mask, order);
3467 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3468 __free_pages(page, order);
3474 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3478 page = alloc_pages_node(nid, gfp_mask, order);
3479 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3480 __free_pages(page, order);
3487 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3490 void __free_kmem_pages(struct page *page, unsigned int order)
3492 memcg_kmem_uncharge(page, order);
3493 __free_pages(page, order);
3496 void free_kmem_pages(unsigned long addr, unsigned int order)
3499 VM_BUG_ON(!virt_addr_valid((void *)addr));
3500 __free_kmem_pages(virt_to_page((void *)addr), order);
3504 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3508 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3509 unsigned long used = addr + PAGE_ALIGN(size);
3511 split_page(virt_to_page((void *)addr), order);
3512 while (used < alloc_end) {
3517 return (void *)addr;
3521 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3522 * @size: the number of bytes to allocate
3523 * @gfp_mask: GFP flags for the allocation
3525 * This function is similar to alloc_pages(), except that it allocates the
3526 * minimum number of pages to satisfy the request. alloc_pages() can only
3527 * allocate memory in power-of-two pages.
3529 * This function is also limited by MAX_ORDER.
3531 * Memory allocated by this function must be released by free_pages_exact().
3533 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3535 unsigned int order = get_order(size);
3538 addr = __get_free_pages(gfp_mask, order);
3539 return make_alloc_exact(addr, order, size);
3541 EXPORT_SYMBOL(alloc_pages_exact);
3544 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3546 * @nid: the preferred node ID where memory should be allocated
3547 * @size: the number of bytes to allocate
3548 * @gfp_mask: GFP flags for the allocation
3550 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3553 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3555 unsigned int order = get_order(size);
3556 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3559 return make_alloc_exact((unsigned long)page_address(p), order, size);
3563 * free_pages_exact - release memory allocated via alloc_pages_exact()
3564 * @virt: the value returned by alloc_pages_exact.
3565 * @size: size of allocation, same value as passed to alloc_pages_exact().
3567 * Release the memory allocated by a previous call to alloc_pages_exact.
3569 void free_pages_exact(void *virt, size_t size)
3571 unsigned long addr = (unsigned long)virt;
3572 unsigned long end = addr + PAGE_ALIGN(size);
3574 while (addr < end) {
3579 EXPORT_SYMBOL(free_pages_exact);
3582 * nr_free_zone_pages - count number of pages beyond high watermark
3583 * @offset: The zone index of the highest zone
3585 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3586 * high watermark within all zones at or below a given zone index. For each
3587 * zone, the number of pages is calculated as:
3588 * managed_pages - high_pages
3590 static unsigned long nr_free_zone_pages(int offset)
3595 /* Just pick one node, since fallback list is circular */
3596 unsigned long sum = 0;
3598 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3600 for_each_zone_zonelist(zone, z, zonelist, offset) {
3601 unsigned long size = zone->managed_pages;
3602 unsigned long high = high_wmark_pages(zone);
3611 * nr_free_buffer_pages - count number of pages beyond high watermark
3613 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3614 * watermark within ZONE_DMA and ZONE_NORMAL.
3616 unsigned long nr_free_buffer_pages(void)
3618 return nr_free_zone_pages(gfp_zone(GFP_USER));
3620 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3623 * nr_free_pagecache_pages - count number of pages beyond high watermark
3625 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3626 * high watermark within all zones.
3628 unsigned long nr_free_pagecache_pages(void)
3630 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3633 static inline void show_node(struct zone *zone)
3635 if (IS_ENABLED(CONFIG_NUMA))
3636 printk("Node %d ", zone_to_nid(zone));
3639 void si_meminfo(struct sysinfo *val)
3641 val->totalram = totalram_pages;
3642 val->sharedram = global_page_state(NR_SHMEM);
3643 val->freeram = global_page_state(NR_FREE_PAGES);
3644 val->bufferram = nr_blockdev_pages();
3645 val->totalhigh = totalhigh_pages;
3646 val->freehigh = nr_free_highpages();
3647 val->mem_unit = PAGE_SIZE;
3650 EXPORT_SYMBOL(si_meminfo);
3653 void si_meminfo_node(struct sysinfo *val, int nid)
3655 int zone_type; /* needs to be signed */
3656 unsigned long managed_pages = 0;
3657 pg_data_t *pgdat = NODE_DATA(nid);
3659 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3660 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3661 val->totalram = managed_pages;
3662 val->sharedram = node_page_state(nid, NR_SHMEM);
3663 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3664 #ifdef CONFIG_HIGHMEM
3665 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3666 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3672 val->mem_unit = PAGE_SIZE;
3677 * Determine whether the node should be displayed or not, depending on whether
3678 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3680 bool skip_free_areas_node(unsigned int flags, int nid)
3683 unsigned int cpuset_mems_cookie;
3685 if (!(flags & SHOW_MEM_FILTER_NODES))
3689 cpuset_mems_cookie = read_mems_allowed_begin();
3690 ret = !node_isset(nid, cpuset_current_mems_allowed);
3691 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3696 #define K(x) ((x) << (PAGE_SHIFT-10))
3698 static void show_migration_types(unsigned char type)
3700 static const char types[MIGRATE_TYPES] = {
3701 [MIGRATE_UNMOVABLE] = 'U',
3702 [MIGRATE_MOVABLE] = 'M',
3703 [MIGRATE_RECLAIMABLE] = 'E',
3704 [MIGRATE_HIGHATOMIC] = 'H',
3706 [MIGRATE_CMA] = 'C',
3708 #ifdef CONFIG_MEMORY_ISOLATION
3709 [MIGRATE_ISOLATE] = 'I',
3712 char tmp[MIGRATE_TYPES + 1];
3716 for (i = 0; i < MIGRATE_TYPES; i++) {
3717 if (type & (1 << i))
3722 printk("(%s) ", tmp);
3726 * Show free area list (used inside shift_scroll-lock stuff)
3727 * We also calculate the percentage fragmentation. We do this by counting the
3728 * memory on each free list with the exception of the first item on the list.
3731 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3734 void show_free_areas(unsigned int filter)
3736 unsigned long free_pcp = 0;
3740 for_each_populated_zone(zone) {
3741 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3744 for_each_online_cpu(cpu)
3745 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3748 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3749 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3750 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3751 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3752 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3753 " free:%lu free_pcp:%lu free_cma:%lu\n",
3754 global_page_state(NR_ACTIVE_ANON),
3755 global_page_state(NR_INACTIVE_ANON),
3756 global_page_state(NR_ISOLATED_ANON),
3757 global_page_state(NR_ACTIVE_FILE),
3758 global_page_state(NR_INACTIVE_FILE),
3759 global_page_state(NR_ISOLATED_FILE),
3760 global_page_state(NR_UNEVICTABLE),
3761 global_page_state(NR_FILE_DIRTY),
3762 global_page_state(NR_WRITEBACK),
3763 global_page_state(NR_UNSTABLE_NFS),
3764 global_page_state(NR_SLAB_RECLAIMABLE),
3765 global_page_state(NR_SLAB_UNRECLAIMABLE),
3766 global_page_state(NR_FILE_MAPPED),
3767 global_page_state(NR_SHMEM),
3768 global_page_state(NR_PAGETABLE),
3769 global_page_state(NR_BOUNCE),
3770 global_page_state(NR_FREE_PAGES),
3772 global_page_state(NR_FREE_CMA_PAGES));
3774 for_each_populated_zone(zone) {
3777 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3781 for_each_online_cpu(cpu)
3782 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3790 " active_anon:%lukB"
3791 " inactive_anon:%lukB"
3792 " active_file:%lukB"
3793 " inactive_file:%lukB"
3794 " unevictable:%lukB"
3795 " isolated(anon):%lukB"
3796 " isolated(file):%lukB"
3804 " slab_reclaimable:%lukB"
3805 " slab_unreclaimable:%lukB"
3806 " kernel_stack:%lukB"
3813 " writeback_tmp:%lukB"
3814 " pages_scanned:%lu"
3815 " all_unreclaimable? %s"
3818 K(zone_page_state(zone, NR_FREE_PAGES)),
3819 K(min_wmark_pages(zone)),
3820 K(low_wmark_pages(zone)),
3821 K(high_wmark_pages(zone)),
3822 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3823 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3824 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3825 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3826 K(zone_page_state(zone, NR_UNEVICTABLE)),
3827 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3828 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3829 K(zone->present_pages),
3830 K(zone->managed_pages),
3831 K(zone_page_state(zone, NR_MLOCK)),
3832 K(zone_page_state(zone, NR_FILE_DIRTY)),
3833 K(zone_page_state(zone, NR_WRITEBACK)),
3834 K(zone_page_state(zone, NR_FILE_MAPPED)),
3835 K(zone_page_state(zone, NR_SHMEM)),
3836 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3837 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3838 zone_page_state(zone, NR_KERNEL_STACK) *
3840 K(zone_page_state(zone, NR_PAGETABLE)),
3841 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3842 K(zone_page_state(zone, NR_BOUNCE)),
3844 K(this_cpu_read(zone->pageset->pcp.count)),
3845 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3846 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3847 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3848 (!zone_reclaimable(zone) ? "yes" : "no")
3850 printk("lowmem_reserve[]:");
3851 for (i = 0; i < MAX_NR_ZONES; i++)
3852 printk(" %ld", zone->lowmem_reserve[i]);
3856 for_each_populated_zone(zone) {
3858 unsigned long nr[MAX_ORDER], flags, total = 0;
3859 unsigned char types[MAX_ORDER];
3861 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3864 printk("%s: ", zone->name);
3866 spin_lock_irqsave(&zone->lock, flags);
3867 for (order = 0; order < MAX_ORDER; order++) {
3868 struct free_area *area = &zone->free_area[order];
3871 nr[order] = area->nr_free;
3872 total += nr[order] << order;
3875 for (type = 0; type < MIGRATE_TYPES; type++) {
3876 if (!list_empty(&area->free_list[type]))
3877 types[order] |= 1 << type;
3880 spin_unlock_irqrestore(&zone->lock, flags);
3881 for (order = 0; order < MAX_ORDER; order++) {
3882 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3884 show_migration_types(types[order]);
3886 printk("= %lukB\n", K(total));
3889 hugetlb_show_meminfo();
3891 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3893 show_swap_cache_info();
3896 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3898 zoneref->zone = zone;
3899 zoneref->zone_idx = zone_idx(zone);
3903 * Builds allocation fallback zone lists.
3905 * Add all populated zones of a node to the zonelist.
3907 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3911 enum zone_type zone_type = MAX_NR_ZONES;
3915 zone = pgdat->node_zones + zone_type;
3916 if (populated_zone(zone)) {
3917 zoneref_set_zone(zone,
3918 &zonelist->_zonerefs[nr_zones++]);
3919 check_highest_zone(zone_type);
3921 } while (zone_type);
3929 * 0 = automatic detection of better ordering.
3930 * 1 = order by ([node] distance, -zonetype)
3931 * 2 = order by (-zonetype, [node] distance)
3933 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3934 * the same zonelist. So only NUMA can configure this param.
3936 #define ZONELIST_ORDER_DEFAULT 0
3937 #define ZONELIST_ORDER_NODE 1
3938 #define ZONELIST_ORDER_ZONE 2
3940 /* zonelist order in the kernel.
3941 * set_zonelist_order() will set this to NODE or ZONE.
3943 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3944 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3948 /* The value user specified ....changed by config */
3949 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3950 /* string for sysctl */
3951 #define NUMA_ZONELIST_ORDER_LEN 16
3952 char numa_zonelist_order[16] = "default";
3955 * interface for configure zonelist ordering.
3956 * command line option "numa_zonelist_order"
3957 * = "[dD]efault - default, automatic configuration.
3958 * = "[nN]ode - order by node locality, then by zone within node
3959 * = "[zZ]one - order by zone, then by locality within zone
3962 static int __parse_numa_zonelist_order(char *s)
3964 if (*s == 'd' || *s == 'D') {
3965 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3966 } else if (*s == 'n' || *s == 'N') {
3967 user_zonelist_order = ZONELIST_ORDER_NODE;
3968 } else if (*s == 'z' || *s == 'Z') {
3969 user_zonelist_order = ZONELIST_ORDER_ZONE;
3972 "Ignoring invalid numa_zonelist_order value: "
3979 static __init int setup_numa_zonelist_order(char *s)
3986 ret = __parse_numa_zonelist_order(s);
3988 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3992 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3995 * sysctl handler for numa_zonelist_order
3997 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3998 void __user *buffer, size_t *length,
4001 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4003 static DEFINE_MUTEX(zl_order_mutex);
4005 mutex_lock(&zl_order_mutex);
4007 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4011 strcpy(saved_string, (char *)table->data);
4013 ret = proc_dostring(table, write, buffer, length, ppos);
4017 int oldval = user_zonelist_order;
4019 ret = __parse_numa_zonelist_order((char *)table->data);
4022 * bogus value. restore saved string
4024 strncpy((char *)table->data, saved_string,
4025 NUMA_ZONELIST_ORDER_LEN);
4026 user_zonelist_order = oldval;
4027 } else if (oldval != user_zonelist_order) {
4028 mutex_lock(&zonelists_mutex);
4029 build_all_zonelists(NULL, NULL);
4030 mutex_unlock(&zonelists_mutex);
4034 mutex_unlock(&zl_order_mutex);
4039 #define MAX_NODE_LOAD (nr_online_nodes)
4040 static int node_load[MAX_NUMNODES];
4043 * find_next_best_node - find the next node that should appear in a given node's fallback list
4044 * @node: node whose fallback list we're appending
4045 * @used_node_mask: nodemask_t of already used nodes
4047 * We use a number of factors to determine which is the next node that should
4048 * appear on a given node's fallback list. The node should not have appeared
4049 * already in @node's fallback list, and it should be the next closest node
4050 * according to the distance array (which contains arbitrary distance values
4051 * from each node to each node in the system), and should also prefer nodes
4052 * with no CPUs, since presumably they'll have very little allocation pressure
4053 * on them otherwise.
4054 * It returns -1 if no node is found.
4056 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4059 int min_val = INT_MAX;
4060 int best_node = NUMA_NO_NODE;
4061 const struct cpumask *tmp = cpumask_of_node(0);
4063 /* Use the local node if we haven't already */
4064 if (!node_isset(node, *used_node_mask)) {
4065 node_set(node, *used_node_mask);
4069 for_each_node_state(n, N_MEMORY) {
4071 /* Don't want a node to appear more than once */
4072 if (node_isset(n, *used_node_mask))
4075 /* Use the distance array to find the distance */
4076 val = node_distance(node, n);
4078 /* Penalize nodes under us ("prefer the next node") */
4081 /* Give preference to headless and unused nodes */
4082 tmp = cpumask_of_node(n);
4083 if (!cpumask_empty(tmp))
4084 val += PENALTY_FOR_NODE_WITH_CPUS;
4086 /* Slight preference for less loaded node */
4087 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4088 val += node_load[n];
4090 if (val < min_val) {
4097 node_set(best_node, *used_node_mask);
4104 * Build zonelists ordered by node and zones within node.
4105 * This results in maximum locality--normal zone overflows into local
4106 * DMA zone, if any--but risks exhausting DMA zone.
4108 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4111 struct zonelist *zonelist;
4113 zonelist = &pgdat->node_zonelists[0];
4114 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4116 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4117 zonelist->_zonerefs[j].zone = NULL;
4118 zonelist->_zonerefs[j].zone_idx = 0;
4122 * Build gfp_thisnode zonelists
4124 static void build_thisnode_zonelists(pg_data_t *pgdat)
4127 struct zonelist *zonelist;
4129 zonelist = &pgdat->node_zonelists[1];
4130 j = build_zonelists_node(pgdat, zonelist, 0);
4131 zonelist->_zonerefs[j].zone = NULL;
4132 zonelist->_zonerefs[j].zone_idx = 0;
4136 * Build zonelists ordered by zone and nodes within zones.
4137 * This results in conserving DMA zone[s] until all Normal memory is
4138 * exhausted, but results in overflowing to remote node while memory
4139 * may still exist in local DMA zone.
4141 static int node_order[MAX_NUMNODES];
4143 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4146 int zone_type; /* needs to be signed */
4148 struct zonelist *zonelist;
4150 zonelist = &pgdat->node_zonelists[0];
4152 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4153 for (j = 0; j < nr_nodes; j++) {
4154 node = node_order[j];
4155 z = &NODE_DATA(node)->node_zones[zone_type];
4156 if (populated_zone(z)) {
4158 &zonelist->_zonerefs[pos++]);
4159 check_highest_zone(zone_type);
4163 zonelist->_zonerefs[pos].zone = NULL;
4164 zonelist->_zonerefs[pos].zone_idx = 0;
4167 #if defined(CONFIG_64BIT)
4169 * Devices that require DMA32/DMA are relatively rare and do not justify a
4170 * penalty to every machine in case the specialised case applies. Default
4171 * to Node-ordering on 64-bit NUMA machines
4173 static int default_zonelist_order(void)
4175 return ZONELIST_ORDER_NODE;
4179 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4180 * by the kernel. If processes running on node 0 deplete the low memory zone
4181 * then reclaim will occur more frequency increasing stalls and potentially
4182 * be easier to OOM if a large percentage of the zone is under writeback or
4183 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4184 * Hence, default to zone ordering on 32-bit.
4186 static int default_zonelist_order(void)
4188 return ZONELIST_ORDER_ZONE;
4190 #endif /* CONFIG_64BIT */
4192 static void set_zonelist_order(void)
4194 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4195 current_zonelist_order = default_zonelist_order();
4197 current_zonelist_order = user_zonelist_order;
4200 static void build_zonelists(pg_data_t *pgdat)
4203 nodemask_t used_mask;
4204 int local_node, prev_node;
4205 struct zonelist *zonelist;
4206 unsigned int order = current_zonelist_order;
4208 /* initialize zonelists */
4209 for (i = 0; i < MAX_ZONELISTS; i++) {
4210 zonelist = pgdat->node_zonelists + i;
4211 zonelist->_zonerefs[0].zone = NULL;
4212 zonelist->_zonerefs[0].zone_idx = 0;
4215 /* NUMA-aware ordering of nodes */
4216 local_node = pgdat->node_id;
4217 load = nr_online_nodes;
4218 prev_node = local_node;
4219 nodes_clear(used_mask);
4221 memset(node_order, 0, sizeof(node_order));
4224 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4226 * We don't want to pressure a particular node.
4227 * So adding penalty to the first node in same
4228 * distance group to make it round-robin.
4230 if (node_distance(local_node, node) !=
4231 node_distance(local_node, prev_node))
4232 node_load[node] = load;
4236 if (order == ZONELIST_ORDER_NODE)
4237 build_zonelists_in_node_order(pgdat, node);
4239 node_order[i++] = node; /* remember order */
4242 if (order == ZONELIST_ORDER_ZONE) {
4243 /* calculate node order -- i.e., DMA last! */
4244 build_zonelists_in_zone_order(pgdat, i);
4247 build_thisnode_zonelists(pgdat);
4250 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4252 * Return node id of node used for "local" allocations.
4253 * I.e., first node id of first zone in arg node's generic zonelist.
4254 * Used for initializing percpu 'numa_mem', which is used primarily
4255 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4257 int local_memory_node(int node)
4261 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4262 gfp_zone(GFP_KERNEL),
4269 #else /* CONFIG_NUMA */
4271 static void set_zonelist_order(void)
4273 current_zonelist_order = ZONELIST_ORDER_ZONE;
4276 static void build_zonelists(pg_data_t *pgdat)
4278 int node, local_node;
4280 struct zonelist *zonelist;
4282 local_node = pgdat->node_id;
4284 zonelist = &pgdat->node_zonelists[0];
4285 j = build_zonelists_node(pgdat, zonelist, 0);
4288 * Now we build the zonelist so that it contains the zones
4289 * of all the other nodes.
4290 * We don't want to pressure a particular node, so when
4291 * building the zones for node N, we make sure that the
4292 * zones coming right after the local ones are those from
4293 * node N+1 (modulo N)
4295 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4296 if (!node_online(node))
4298 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4300 for (node = 0; node < local_node; node++) {
4301 if (!node_online(node))
4303 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4306 zonelist->_zonerefs[j].zone = NULL;
4307 zonelist->_zonerefs[j].zone_idx = 0;
4310 #endif /* CONFIG_NUMA */
4313 * Boot pageset table. One per cpu which is going to be used for all
4314 * zones and all nodes. The parameters will be set in such a way
4315 * that an item put on a list will immediately be handed over to
4316 * the buddy list. This is safe since pageset manipulation is done
4317 * with interrupts disabled.
4319 * The boot_pagesets must be kept even after bootup is complete for
4320 * unused processors and/or zones. They do play a role for bootstrapping
4321 * hotplugged processors.
4323 * zoneinfo_show() and maybe other functions do
4324 * not check if the processor is online before following the pageset pointer.
4325 * Other parts of the kernel may not check if the zone is available.
4327 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4328 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4329 static void setup_zone_pageset(struct zone *zone);
4332 * Global mutex to protect against size modification of zonelists
4333 * as well as to serialize pageset setup for the new populated zone.
4335 DEFINE_MUTEX(zonelists_mutex);
4337 /* return values int ....just for stop_machine() */
4338 static int __build_all_zonelists(void *data)
4342 pg_data_t *self = data;
4345 memset(node_load, 0, sizeof(node_load));
4348 if (self && !node_online(self->node_id)) {
4349 build_zonelists(self);
4352 for_each_online_node(nid) {
4353 pg_data_t *pgdat = NODE_DATA(nid);
4355 build_zonelists(pgdat);
4359 * Initialize the boot_pagesets that are going to be used
4360 * for bootstrapping processors. The real pagesets for
4361 * each zone will be allocated later when the per cpu
4362 * allocator is available.
4364 * boot_pagesets are used also for bootstrapping offline
4365 * cpus if the system is already booted because the pagesets
4366 * are needed to initialize allocators on a specific cpu too.
4367 * F.e. the percpu allocator needs the page allocator which
4368 * needs the percpu allocator in order to allocate its pagesets
4369 * (a chicken-egg dilemma).
4371 for_each_possible_cpu(cpu) {
4372 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4374 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4376 * We now know the "local memory node" for each node--
4377 * i.e., the node of the first zone in the generic zonelist.
4378 * Set up numa_mem percpu variable for on-line cpus. During
4379 * boot, only the boot cpu should be on-line; we'll init the
4380 * secondary cpus' numa_mem as they come on-line. During
4381 * node/memory hotplug, we'll fixup all on-line cpus.
4383 if (cpu_online(cpu))
4384 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4391 static noinline void __init
4392 build_all_zonelists_init(void)
4394 __build_all_zonelists(NULL);
4395 mminit_verify_zonelist();
4396 cpuset_init_current_mems_allowed();
4400 * Called with zonelists_mutex held always
4401 * unless system_state == SYSTEM_BOOTING.
4403 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4404 * [we're only called with non-NULL zone through __meminit paths] and
4405 * (2) call of __init annotated helper build_all_zonelists_init
4406 * [protected by SYSTEM_BOOTING].
4408 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4410 set_zonelist_order();
4412 if (system_state == SYSTEM_BOOTING) {
4413 build_all_zonelists_init();
4415 #ifdef CONFIG_MEMORY_HOTPLUG
4417 setup_zone_pageset(zone);
4419 /* we have to stop all cpus to guarantee there is no user
4421 stop_machine(__build_all_zonelists, pgdat, NULL);
4422 /* cpuset refresh routine should be here */
4424 vm_total_pages = nr_free_pagecache_pages();
4426 * Disable grouping by mobility if the number of pages in the
4427 * system is too low to allow the mechanism to work. It would be
4428 * more accurate, but expensive to check per-zone. This check is
4429 * made on memory-hotadd so a system can start with mobility
4430 * disabled and enable it later
4432 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4433 page_group_by_mobility_disabled = 1;
4435 page_group_by_mobility_disabled = 0;
4437 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4438 "Total pages: %ld\n",
4440 zonelist_order_name[current_zonelist_order],
4441 page_group_by_mobility_disabled ? "off" : "on",
4444 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4449 * Helper functions to size the waitqueue hash table.
4450 * Essentially these want to choose hash table sizes sufficiently
4451 * large so that collisions trying to wait on pages are rare.
4452 * But in fact, the number of active page waitqueues on typical
4453 * systems is ridiculously low, less than 200. So this is even
4454 * conservative, even though it seems large.
4456 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4457 * waitqueues, i.e. the size of the waitq table given the number of pages.
4459 #define PAGES_PER_WAITQUEUE 256
4461 #ifndef CONFIG_MEMORY_HOTPLUG
4462 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4464 unsigned long size = 1;
4466 pages /= PAGES_PER_WAITQUEUE;
4468 while (size < pages)
4472 * Once we have dozens or even hundreds of threads sleeping
4473 * on IO we've got bigger problems than wait queue collision.
4474 * Limit the size of the wait table to a reasonable size.
4476 size = min(size, 4096UL);
4478 return max(size, 4UL);
4482 * A zone's size might be changed by hot-add, so it is not possible to determine
4483 * a suitable size for its wait_table. So we use the maximum size now.
4485 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4487 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4488 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4489 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4491 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4492 * or more by the traditional way. (See above). It equals:
4494 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4495 * ia64(16K page size) : = ( 8G + 4M)byte.
4496 * powerpc (64K page size) : = (32G +16M)byte.
4498 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4505 * This is an integer logarithm so that shifts can be used later
4506 * to extract the more random high bits from the multiplicative
4507 * hash function before the remainder is taken.
4509 static inline unsigned long wait_table_bits(unsigned long size)
4515 * Initially all pages are reserved - free ones are freed
4516 * up by free_all_bootmem() once the early boot process is
4517 * done. Non-atomic initialization, single-pass.
4519 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4520 unsigned long start_pfn, enum memmap_context context)
4522 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
4523 unsigned long end_pfn = start_pfn + size;
4524 pg_data_t *pgdat = NODE_DATA(nid);
4526 unsigned long nr_initialised = 0;
4527 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4528 struct memblock_region *r = NULL, *tmp;
4531 if (highest_memmap_pfn < end_pfn - 1)
4532 highest_memmap_pfn = end_pfn - 1;
4535 * Honor reservation requested by the driver for this ZONE_DEVICE
4538 if (altmap && start_pfn == altmap->base_pfn)
4539 start_pfn += altmap->reserve;
4541 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4543 * There can be holes in boot-time mem_map[]s handed to this
4544 * function. They do not exist on hotplugged memory.
4546 if (context != MEMMAP_EARLY)
4549 if (!early_pfn_valid(pfn))
4551 if (!early_pfn_in_nid(pfn, nid))
4553 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
4556 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4558 * If not mirrored_kernelcore and ZONE_MOVABLE exists, range
4559 * from zone_movable_pfn[nid] to end of each node should be
4560 * ZONE_MOVABLE not ZONE_NORMAL. skip it.
4562 if (!mirrored_kernelcore && zone_movable_pfn[nid])
4563 if (zone == ZONE_NORMAL && pfn >= zone_movable_pfn[nid])
4567 * Check given memblock attribute by firmware which can affect
4568 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
4569 * mirrored, it's an overlapped memmap init. skip it.
4571 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
4572 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
4573 for_each_memblock(memory, tmp)
4574 if (pfn < memblock_region_memory_end_pfn(tmp))
4578 if (pfn >= memblock_region_memory_base_pfn(r) &&
4579 memblock_is_mirror(r)) {
4580 /* already initialized as NORMAL */
4581 pfn = memblock_region_memory_end_pfn(r);
4589 * Mark the block movable so that blocks are reserved for
4590 * movable at startup. This will force kernel allocations
4591 * to reserve their blocks rather than leaking throughout
4592 * the address space during boot when many long-lived
4593 * kernel allocations are made.
4595 * bitmap is created for zone's valid pfn range. but memmap
4596 * can be created for invalid pages (for alignment)
4597 * check here not to call set_pageblock_migratetype() against
4600 if (!(pfn & (pageblock_nr_pages - 1))) {
4601 struct page *page = pfn_to_page(pfn);
4603 __init_single_page(page, pfn, zone, nid);
4604 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4606 __init_single_pfn(pfn, zone, nid);
4611 static void __meminit zone_init_free_lists(struct zone *zone)
4613 unsigned int order, t;
4614 for_each_migratetype_order(order, t) {
4615 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4616 zone->free_area[order].nr_free = 0;
4620 #ifndef __HAVE_ARCH_MEMMAP_INIT
4621 #define memmap_init(size, nid, zone, start_pfn) \
4622 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4625 static int zone_batchsize(struct zone *zone)
4631 * The per-cpu-pages pools are set to around 1000th of the
4632 * size of the zone. But no more than 1/2 of a meg.
4634 * OK, so we don't know how big the cache is. So guess.
4636 batch = zone->managed_pages / 1024;
4637 if (batch * PAGE_SIZE > 512 * 1024)
4638 batch = (512 * 1024) / PAGE_SIZE;
4639 batch /= 4; /* We effectively *= 4 below */
4644 * Clamp the batch to a 2^n - 1 value. Having a power
4645 * of 2 value was found to be more likely to have
4646 * suboptimal cache aliasing properties in some cases.
4648 * For example if 2 tasks are alternately allocating
4649 * batches of pages, one task can end up with a lot
4650 * of pages of one half of the possible page colors
4651 * and the other with pages of the other colors.
4653 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4658 /* The deferral and batching of frees should be suppressed under NOMMU
4661 * The problem is that NOMMU needs to be able to allocate large chunks
4662 * of contiguous memory as there's no hardware page translation to
4663 * assemble apparent contiguous memory from discontiguous pages.
4665 * Queueing large contiguous runs of pages for batching, however,
4666 * causes the pages to actually be freed in smaller chunks. As there
4667 * can be a significant delay between the individual batches being
4668 * recycled, this leads to the once large chunks of space being
4669 * fragmented and becoming unavailable for high-order allocations.
4676 * pcp->high and pcp->batch values are related and dependent on one another:
4677 * ->batch must never be higher then ->high.
4678 * The following function updates them in a safe manner without read side
4681 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4682 * those fields changing asynchronously (acording the the above rule).
4684 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4685 * outside of boot time (or some other assurance that no concurrent updaters
4688 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4689 unsigned long batch)
4691 /* start with a fail safe value for batch */
4695 /* Update high, then batch, in order */
4702 /* a companion to pageset_set_high() */
4703 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4705 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4708 static void pageset_init(struct per_cpu_pageset *p)
4710 struct per_cpu_pages *pcp;
4713 memset(p, 0, sizeof(*p));
4717 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4718 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4721 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4724 pageset_set_batch(p, batch);
4728 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4729 * to the value high for the pageset p.
4731 static void pageset_set_high(struct per_cpu_pageset *p,
4734 unsigned long batch = max(1UL, high / 4);
4735 if ((high / 4) > (PAGE_SHIFT * 8))
4736 batch = PAGE_SHIFT * 8;
4738 pageset_update(&p->pcp, high, batch);
4741 static void pageset_set_high_and_batch(struct zone *zone,
4742 struct per_cpu_pageset *pcp)
4744 if (percpu_pagelist_fraction)
4745 pageset_set_high(pcp,
4746 (zone->managed_pages /
4747 percpu_pagelist_fraction));
4749 pageset_set_batch(pcp, zone_batchsize(zone));
4752 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4754 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4757 pageset_set_high_and_batch(zone, pcp);
4760 static void __meminit setup_zone_pageset(struct zone *zone)
4763 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4764 for_each_possible_cpu(cpu)
4765 zone_pageset_init(zone, cpu);
4769 * Allocate per cpu pagesets and initialize them.
4770 * Before this call only boot pagesets were available.
4772 void __init setup_per_cpu_pageset(void)
4776 for_each_populated_zone(zone)
4777 setup_zone_pageset(zone);
4780 static noinline __init_refok
4781 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4787 * The per-page waitqueue mechanism uses hashed waitqueues
4790 zone->wait_table_hash_nr_entries =
4791 wait_table_hash_nr_entries(zone_size_pages);
4792 zone->wait_table_bits =
4793 wait_table_bits(zone->wait_table_hash_nr_entries);
4794 alloc_size = zone->wait_table_hash_nr_entries
4795 * sizeof(wait_queue_head_t);
4797 if (!slab_is_available()) {
4798 zone->wait_table = (wait_queue_head_t *)
4799 memblock_virt_alloc_node_nopanic(
4800 alloc_size, zone->zone_pgdat->node_id);
4803 * This case means that a zone whose size was 0 gets new memory
4804 * via memory hot-add.
4805 * But it may be the case that a new node was hot-added. In
4806 * this case vmalloc() will not be able to use this new node's
4807 * memory - this wait_table must be initialized to use this new
4808 * node itself as well.
4809 * To use this new node's memory, further consideration will be
4812 zone->wait_table = vmalloc(alloc_size);
4814 if (!zone->wait_table)
4817 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4818 init_waitqueue_head(zone->wait_table + i);
4823 static __meminit void zone_pcp_init(struct zone *zone)
4826 * per cpu subsystem is not up at this point. The following code
4827 * relies on the ability of the linker to provide the
4828 * offset of a (static) per cpu variable into the per cpu area.
4830 zone->pageset = &boot_pageset;
4832 if (populated_zone(zone))
4833 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4834 zone->name, zone->present_pages,
4835 zone_batchsize(zone));
4838 int __meminit init_currently_empty_zone(struct zone *zone,
4839 unsigned long zone_start_pfn,
4842 struct pglist_data *pgdat = zone->zone_pgdat;
4844 ret = zone_wait_table_init(zone, size);
4847 pgdat->nr_zones = zone_idx(zone) + 1;
4849 zone->zone_start_pfn = zone_start_pfn;
4851 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4852 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4854 (unsigned long)zone_idx(zone),
4855 zone_start_pfn, (zone_start_pfn + size));
4857 zone_init_free_lists(zone);
4862 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4863 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4866 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4868 int __meminit __early_pfn_to_nid(unsigned long pfn,
4869 struct mminit_pfnnid_cache *state)
4871 unsigned long start_pfn, end_pfn;
4874 if (state->last_start <= pfn && pfn < state->last_end)
4875 return state->last_nid;
4877 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4879 state->last_start = start_pfn;
4880 state->last_end = end_pfn;
4881 state->last_nid = nid;
4886 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4889 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4890 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4891 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4893 * If an architecture guarantees that all ranges registered contain no holes
4894 * and may be freed, this this function may be used instead of calling
4895 * memblock_free_early_nid() manually.
4897 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4899 unsigned long start_pfn, end_pfn;
4902 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4903 start_pfn = min(start_pfn, max_low_pfn);
4904 end_pfn = min(end_pfn, max_low_pfn);
4906 if (start_pfn < end_pfn)
4907 memblock_free_early_nid(PFN_PHYS(start_pfn),
4908 (end_pfn - start_pfn) << PAGE_SHIFT,
4914 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4915 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4917 * If an architecture guarantees that all ranges registered contain no holes and may
4918 * be freed, this function may be used instead of calling memory_present() manually.
4920 void __init sparse_memory_present_with_active_regions(int nid)
4922 unsigned long start_pfn, end_pfn;
4925 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4926 memory_present(this_nid, start_pfn, end_pfn);
4930 * get_pfn_range_for_nid - Return the start and end page frames for a node
4931 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4932 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4933 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4935 * It returns the start and end page frame of a node based on information
4936 * provided by memblock_set_node(). If called for a node
4937 * with no available memory, a warning is printed and the start and end
4940 void __meminit get_pfn_range_for_nid(unsigned int nid,
4941 unsigned long *start_pfn, unsigned long *end_pfn)
4943 unsigned long this_start_pfn, this_end_pfn;
4949 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4950 *start_pfn = min(*start_pfn, this_start_pfn);
4951 *end_pfn = max(*end_pfn, this_end_pfn);
4954 if (*start_pfn == -1UL)
4959 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4960 * assumption is made that zones within a node are ordered in monotonic
4961 * increasing memory addresses so that the "highest" populated zone is used
4963 static void __init find_usable_zone_for_movable(void)
4966 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4967 if (zone_index == ZONE_MOVABLE)
4970 if (arch_zone_highest_possible_pfn[zone_index] >
4971 arch_zone_lowest_possible_pfn[zone_index])
4975 VM_BUG_ON(zone_index == -1);
4976 movable_zone = zone_index;
4980 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4981 * because it is sized independent of architecture. Unlike the other zones,
4982 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4983 * in each node depending on the size of each node and how evenly kernelcore
4984 * is distributed. This helper function adjusts the zone ranges
4985 * provided by the architecture for a given node by using the end of the
4986 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4987 * zones within a node are in order of monotonic increases memory addresses
4989 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4990 unsigned long zone_type,
4991 unsigned long node_start_pfn,
4992 unsigned long node_end_pfn,
4993 unsigned long *zone_start_pfn,
4994 unsigned long *zone_end_pfn)
4996 /* Only adjust if ZONE_MOVABLE is on this node */
4997 if (zone_movable_pfn[nid]) {
4998 /* Size ZONE_MOVABLE */
4999 if (zone_type == ZONE_MOVABLE) {
5000 *zone_start_pfn = zone_movable_pfn[nid];
5001 *zone_end_pfn = min(node_end_pfn,
5002 arch_zone_highest_possible_pfn[movable_zone]);
5004 /* Check if this whole range is within ZONE_MOVABLE */
5005 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5006 *zone_start_pfn = *zone_end_pfn;
5011 * Return the number of pages a zone spans in a node, including holes
5012 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5014 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5015 unsigned long zone_type,
5016 unsigned long node_start_pfn,
5017 unsigned long node_end_pfn,
5018 unsigned long *zone_start_pfn,
5019 unsigned long *zone_end_pfn,
5020 unsigned long *ignored)
5022 /* When hotadd a new node from cpu_up(), the node should be empty */
5023 if (!node_start_pfn && !node_end_pfn)
5026 /* Get the start and end of the zone */
5027 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5028 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5029 adjust_zone_range_for_zone_movable(nid, zone_type,
5030 node_start_pfn, node_end_pfn,
5031 zone_start_pfn, zone_end_pfn);
5033 /* Check that this node has pages within the zone's required range */
5034 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5037 /* Move the zone boundaries inside the node if necessary */
5038 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5039 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5041 /* Return the spanned pages */
5042 return *zone_end_pfn - *zone_start_pfn;
5046 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5047 * then all holes in the requested range will be accounted for.
5049 unsigned long __meminit __absent_pages_in_range(int nid,
5050 unsigned long range_start_pfn,
5051 unsigned long range_end_pfn)
5053 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5054 unsigned long start_pfn, end_pfn;
5057 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5058 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5059 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5060 nr_absent -= end_pfn - start_pfn;
5066 * absent_pages_in_range - Return number of page frames in holes within a range
5067 * @start_pfn: The start PFN to start searching for holes
5068 * @end_pfn: The end PFN to stop searching for holes
5070 * It returns the number of pages frames in memory holes within a range.
5072 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5073 unsigned long end_pfn)
5075 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5078 /* Return the number of page frames in holes in a zone on a node */
5079 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5080 unsigned long zone_type,
5081 unsigned long node_start_pfn,
5082 unsigned long node_end_pfn,
5083 unsigned long *ignored)
5085 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5086 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5087 unsigned long zone_start_pfn, zone_end_pfn;
5088 unsigned long nr_absent;
5090 /* When hotadd a new node from cpu_up(), the node should be empty */
5091 if (!node_start_pfn && !node_end_pfn)
5094 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5095 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5097 adjust_zone_range_for_zone_movable(nid, zone_type,
5098 node_start_pfn, node_end_pfn,
5099 &zone_start_pfn, &zone_end_pfn);
5100 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5103 * ZONE_MOVABLE handling.
5104 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5107 if (zone_movable_pfn[nid]) {
5108 if (mirrored_kernelcore) {
5109 unsigned long start_pfn, end_pfn;
5110 struct memblock_region *r;
5112 for_each_memblock(memory, r) {
5113 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5114 zone_start_pfn, zone_end_pfn);
5115 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5116 zone_start_pfn, zone_end_pfn);
5118 if (zone_type == ZONE_MOVABLE &&
5119 memblock_is_mirror(r))
5120 nr_absent += end_pfn - start_pfn;
5122 if (zone_type == ZONE_NORMAL &&
5123 !memblock_is_mirror(r))
5124 nr_absent += end_pfn - start_pfn;
5127 if (zone_type == ZONE_NORMAL)
5128 nr_absent += node_end_pfn - zone_movable_pfn[nid];
5135 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5136 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5137 unsigned long zone_type,
5138 unsigned long node_start_pfn,
5139 unsigned long node_end_pfn,
5140 unsigned long *zone_start_pfn,
5141 unsigned long *zone_end_pfn,
5142 unsigned long *zones_size)
5146 *zone_start_pfn = node_start_pfn;
5147 for (zone = 0; zone < zone_type; zone++)
5148 *zone_start_pfn += zones_size[zone];
5150 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5152 return zones_size[zone_type];
5155 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5156 unsigned long zone_type,
5157 unsigned long node_start_pfn,
5158 unsigned long node_end_pfn,
5159 unsigned long *zholes_size)
5164 return zholes_size[zone_type];
5167 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5169 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5170 unsigned long node_start_pfn,
5171 unsigned long node_end_pfn,
5172 unsigned long *zones_size,
5173 unsigned long *zholes_size)
5175 unsigned long realtotalpages = 0, totalpages = 0;
5178 for (i = 0; i < MAX_NR_ZONES; i++) {
5179 struct zone *zone = pgdat->node_zones + i;
5180 unsigned long zone_start_pfn, zone_end_pfn;
5181 unsigned long size, real_size;
5183 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5189 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5190 node_start_pfn, node_end_pfn,
5193 zone->zone_start_pfn = zone_start_pfn;
5195 zone->zone_start_pfn = 0;
5196 zone->spanned_pages = size;
5197 zone->present_pages = real_size;
5200 realtotalpages += real_size;
5203 pgdat->node_spanned_pages = totalpages;
5204 pgdat->node_present_pages = realtotalpages;
5205 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5209 #ifndef CONFIG_SPARSEMEM
5211 * Calculate the size of the zone->blockflags rounded to an unsigned long
5212 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5213 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5214 * round what is now in bits to nearest long in bits, then return it in
5217 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5219 unsigned long usemapsize;
5221 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5222 usemapsize = roundup(zonesize, pageblock_nr_pages);
5223 usemapsize = usemapsize >> pageblock_order;
5224 usemapsize *= NR_PAGEBLOCK_BITS;
5225 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5227 return usemapsize / 8;
5230 static void __init setup_usemap(struct pglist_data *pgdat,
5232 unsigned long zone_start_pfn,
5233 unsigned long zonesize)
5235 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5236 zone->pageblock_flags = NULL;
5238 zone->pageblock_flags =
5239 memblock_virt_alloc_node_nopanic(usemapsize,
5243 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5244 unsigned long zone_start_pfn, unsigned long zonesize) {}
5245 #endif /* CONFIG_SPARSEMEM */
5247 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5249 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5250 void __paginginit set_pageblock_order(void)
5254 /* Check that pageblock_nr_pages has not already been setup */
5255 if (pageblock_order)
5258 if (HPAGE_SHIFT > PAGE_SHIFT)
5259 order = HUGETLB_PAGE_ORDER;
5261 order = MAX_ORDER - 1;
5264 * Assume the largest contiguous order of interest is a huge page.
5265 * This value may be variable depending on boot parameters on IA64 and
5268 pageblock_order = order;
5270 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5273 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5274 * is unused as pageblock_order is set at compile-time. See
5275 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5278 void __paginginit set_pageblock_order(void)
5282 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5284 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5285 unsigned long present_pages)
5287 unsigned long pages = spanned_pages;
5290 * Provide a more accurate estimation if there are holes within
5291 * the zone and SPARSEMEM is in use. If there are holes within the
5292 * zone, each populated memory region may cost us one or two extra
5293 * memmap pages due to alignment because memmap pages for each
5294 * populated regions may not naturally algined on page boundary.
5295 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5297 if (spanned_pages > present_pages + (present_pages >> 4) &&
5298 IS_ENABLED(CONFIG_SPARSEMEM))
5299 pages = present_pages;
5301 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5305 * Set up the zone data structures:
5306 * - mark all pages reserved
5307 * - mark all memory queues empty
5308 * - clear the memory bitmaps
5310 * NOTE: pgdat should get zeroed by caller.
5312 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5315 int nid = pgdat->node_id;
5318 pgdat_resize_init(pgdat);
5319 #ifdef CONFIG_NUMA_BALANCING
5320 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5321 pgdat->numabalancing_migrate_nr_pages = 0;
5322 pgdat->numabalancing_migrate_next_window = jiffies;
5324 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5325 spin_lock_init(&pgdat->split_queue_lock);
5326 INIT_LIST_HEAD(&pgdat->split_queue);
5327 pgdat->split_queue_len = 0;
5329 init_waitqueue_head(&pgdat->kswapd_wait);
5330 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5331 pgdat_page_ext_init(pgdat);
5333 for (j = 0; j < MAX_NR_ZONES; j++) {
5334 struct zone *zone = pgdat->node_zones + j;
5335 unsigned long size, realsize, freesize, memmap_pages;
5336 unsigned long zone_start_pfn = zone->zone_start_pfn;
5338 size = zone->spanned_pages;
5339 realsize = freesize = zone->present_pages;
5342 * Adjust freesize so that it accounts for how much memory
5343 * is used by this zone for memmap. This affects the watermark
5344 * and per-cpu initialisations
5346 memmap_pages = calc_memmap_size(size, realsize);
5347 if (!is_highmem_idx(j)) {
5348 if (freesize >= memmap_pages) {
5349 freesize -= memmap_pages;
5352 " %s zone: %lu pages used for memmap\n",
5353 zone_names[j], memmap_pages);
5356 " %s zone: %lu pages exceeds freesize %lu\n",
5357 zone_names[j], memmap_pages, freesize);
5360 /* Account for reserved pages */
5361 if (j == 0 && freesize > dma_reserve) {
5362 freesize -= dma_reserve;
5363 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5364 zone_names[0], dma_reserve);
5367 if (!is_highmem_idx(j))
5368 nr_kernel_pages += freesize;
5369 /* Charge for highmem memmap if there are enough kernel pages */
5370 else if (nr_kernel_pages > memmap_pages * 2)
5371 nr_kernel_pages -= memmap_pages;
5372 nr_all_pages += freesize;
5375 * Set an approximate value for lowmem here, it will be adjusted
5376 * when the bootmem allocator frees pages into the buddy system.
5377 * And all highmem pages will be managed by the buddy system.
5379 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5382 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5384 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5386 zone->name = zone_names[j];
5387 spin_lock_init(&zone->lock);
5388 spin_lock_init(&zone->lru_lock);
5389 zone_seqlock_init(zone);
5390 zone->zone_pgdat = pgdat;
5391 zone_pcp_init(zone);
5393 /* For bootup, initialized properly in watermark setup */
5394 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5396 lruvec_init(&zone->lruvec);
5400 set_pageblock_order();
5401 setup_usemap(pgdat, zone, zone_start_pfn, size);
5402 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5404 memmap_init(size, nid, j, zone_start_pfn);
5408 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5410 unsigned long __maybe_unused start = 0;
5411 unsigned long __maybe_unused offset = 0;
5413 /* Skip empty nodes */
5414 if (!pgdat->node_spanned_pages)
5417 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5418 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5419 offset = pgdat->node_start_pfn - start;
5420 /* ia64 gets its own node_mem_map, before this, without bootmem */
5421 if (!pgdat->node_mem_map) {
5422 unsigned long size, end;
5426 * The zone's endpoints aren't required to be MAX_ORDER
5427 * aligned but the node_mem_map endpoints must be in order
5428 * for the buddy allocator to function correctly.
5430 end = pgdat_end_pfn(pgdat);
5431 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5432 size = (end - start) * sizeof(struct page);
5433 map = alloc_remap(pgdat->node_id, size);
5435 map = memblock_virt_alloc_node_nopanic(size,
5437 pgdat->node_mem_map = map + offset;
5439 #ifndef CONFIG_NEED_MULTIPLE_NODES
5441 * With no DISCONTIG, the global mem_map is just set as node 0's
5443 if (pgdat == NODE_DATA(0)) {
5444 mem_map = NODE_DATA(0)->node_mem_map;
5445 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5446 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5448 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5451 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5454 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5455 unsigned long node_start_pfn, unsigned long *zholes_size)
5457 pg_data_t *pgdat = NODE_DATA(nid);
5458 unsigned long start_pfn = 0;
5459 unsigned long end_pfn = 0;
5461 /* pg_data_t should be reset to zero when it's allocated */
5462 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5464 reset_deferred_meminit(pgdat);
5465 pgdat->node_id = nid;
5466 pgdat->node_start_pfn = node_start_pfn;
5467 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5468 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5469 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5470 (u64)start_pfn << PAGE_SHIFT,
5471 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5473 start_pfn = node_start_pfn;
5475 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5476 zones_size, zholes_size);
5478 alloc_node_mem_map(pgdat);
5479 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5480 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5481 nid, (unsigned long)pgdat,
5482 (unsigned long)pgdat->node_mem_map);
5485 free_area_init_core(pgdat);
5488 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5490 #if MAX_NUMNODES > 1
5492 * Figure out the number of possible node ids.
5494 void __init setup_nr_node_ids(void)
5496 unsigned int highest;
5498 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5499 nr_node_ids = highest + 1;
5504 * node_map_pfn_alignment - determine the maximum internode alignment
5506 * This function should be called after node map is populated and sorted.
5507 * It calculates the maximum power of two alignment which can distinguish
5510 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5511 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5512 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5513 * shifted, 1GiB is enough and this function will indicate so.
5515 * This is used to test whether pfn -> nid mapping of the chosen memory
5516 * model has fine enough granularity to avoid incorrect mapping for the
5517 * populated node map.
5519 * Returns the determined alignment in pfn's. 0 if there is no alignment
5520 * requirement (single node).
5522 unsigned long __init node_map_pfn_alignment(void)
5524 unsigned long accl_mask = 0, last_end = 0;
5525 unsigned long start, end, mask;
5529 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5530 if (!start || last_nid < 0 || last_nid == nid) {
5537 * Start with a mask granular enough to pin-point to the
5538 * start pfn and tick off bits one-by-one until it becomes
5539 * too coarse to separate the current node from the last.
5541 mask = ~((1 << __ffs(start)) - 1);
5542 while (mask && last_end <= (start & (mask << 1)))
5545 /* accumulate all internode masks */
5549 /* convert mask to number of pages */
5550 return ~accl_mask + 1;
5553 /* Find the lowest pfn for a node */
5554 static unsigned long __init find_min_pfn_for_node(int nid)
5556 unsigned long min_pfn = ULONG_MAX;
5557 unsigned long start_pfn;
5560 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5561 min_pfn = min(min_pfn, start_pfn);
5563 if (min_pfn == ULONG_MAX) {
5565 "Could not find start_pfn for node %d\n", nid);
5573 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5575 * It returns the minimum PFN based on information provided via
5576 * memblock_set_node().
5578 unsigned long __init find_min_pfn_with_active_regions(void)
5580 return find_min_pfn_for_node(MAX_NUMNODES);
5584 * early_calculate_totalpages()
5585 * Sum pages in active regions for movable zone.
5586 * Populate N_MEMORY for calculating usable_nodes.
5588 static unsigned long __init early_calculate_totalpages(void)
5590 unsigned long totalpages = 0;
5591 unsigned long start_pfn, end_pfn;
5594 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5595 unsigned long pages = end_pfn - start_pfn;
5597 totalpages += pages;
5599 node_set_state(nid, N_MEMORY);
5605 * Find the PFN the Movable zone begins in each node. Kernel memory
5606 * is spread evenly between nodes as long as the nodes have enough
5607 * memory. When they don't, some nodes will have more kernelcore than
5610 static void __init find_zone_movable_pfns_for_nodes(void)
5613 unsigned long usable_startpfn;
5614 unsigned long kernelcore_node, kernelcore_remaining;
5615 /* save the state before borrow the nodemask */
5616 nodemask_t saved_node_state = node_states[N_MEMORY];
5617 unsigned long totalpages = early_calculate_totalpages();
5618 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5619 struct memblock_region *r;
5621 /* Need to find movable_zone earlier when movable_node is specified. */
5622 find_usable_zone_for_movable();
5625 * If movable_node is specified, ignore kernelcore and movablecore
5628 if (movable_node_is_enabled()) {
5629 for_each_memblock(memory, r) {
5630 if (!memblock_is_hotpluggable(r))
5635 usable_startpfn = PFN_DOWN(r->base);
5636 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5637 min(usable_startpfn, zone_movable_pfn[nid]) :
5645 * If kernelcore=mirror is specified, ignore movablecore option
5647 if (mirrored_kernelcore) {
5648 bool mem_below_4gb_not_mirrored = false;
5650 for_each_memblock(memory, r) {
5651 if (memblock_is_mirror(r))
5656 usable_startpfn = memblock_region_memory_base_pfn(r);
5658 if (usable_startpfn < 0x100000) {
5659 mem_below_4gb_not_mirrored = true;
5663 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5664 min(usable_startpfn, zone_movable_pfn[nid]) :
5668 if (mem_below_4gb_not_mirrored)
5669 pr_warn("This configuration results in unmirrored kernel memory.");
5675 * If movablecore=nn[KMG] was specified, calculate what size of
5676 * kernelcore that corresponds so that memory usable for
5677 * any allocation type is evenly spread. If both kernelcore
5678 * and movablecore are specified, then the value of kernelcore
5679 * will be used for required_kernelcore if it's greater than
5680 * what movablecore would have allowed.
5682 if (required_movablecore) {
5683 unsigned long corepages;
5686 * Round-up so that ZONE_MOVABLE is at least as large as what
5687 * was requested by the user
5689 required_movablecore =
5690 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5691 required_movablecore = min(totalpages, required_movablecore);
5692 corepages = totalpages - required_movablecore;
5694 required_kernelcore = max(required_kernelcore, corepages);
5698 * If kernelcore was not specified or kernelcore size is larger
5699 * than totalpages, there is no ZONE_MOVABLE.
5701 if (!required_kernelcore || required_kernelcore >= totalpages)
5704 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5705 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5708 /* Spread kernelcore memory as evenly as possible throughout nodes */
5709 kernelcore_node = required_kernelcore / usable_nodes;
5710 for_each_node_state(nid, N_MEMORY) {
5711 unsigned long start_pfn, end_pfn;
5714 * Recalculate kernelcore_node if the division per node
5715 * now exceeds what is necessary to satisfy the requested
5716 * amount of memory for the kernel
5718 if (required_kernelcore < kernelcore_node)
5719 kernelcore_node = required_kernelcore / usable_nodes;
5722 * As the map is walked, we track how much memory is usable
5723 * by the kernel using kernelcore_remaining. When it is
5724 * 0, the rest of the node is usable by ZONE_MOVABLE
5726 kernelcore_remaining = kernelcore_node;
5728 /* Go through each range of PFNs within this node */
5729 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5730 unsigned long size_pages;
5732 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5733 if (start_pfn >= end_pfn)
5736 /* Account for what is only usable for kernelcore */
5737 if (start_pfn < usable_startpfn) {
5738 unsigned long kernel_pages;
5739 kernel_pages = min(end_pfn, usable_startpfn)
5742 kernelcore_remaining -= min(kernel_pages,
5743 kernelcore_remaining);
5744 required_kernelcore -= min(kernel_pages,
5745 required_kernelcore);
5747 /* Continue if range is now fully accounted */
5748 if (end_pfn <= usable_startpfn) {
5751 * Push zone_movable_pfn to the end so
5752 * that if we have to rebalance
5753 * kernelcore across nodes, we will
5754 * not double account here
5756 zone_movable_pfn[nid] = end_pfn;
5759 start_pfn = usable_startpfn;
5763 * The usable PFN range for ZONE_MOVABLE is from
5764 * start_pfn->end_pfn. Calculate size_pages as the
5765 * number of pages used as kernelcore
5767 size_pages = end_pfn - start_pfn;
5768 if (size_pages > kernelcore_remaining)
5769 size_pages = kernelcore_remaining;
5770 zone_movable_pfn[nid] = start_pfn + size_pages;
5773 * Some kernelcore has been met, update counts and
5774 * break if the kernelcore for this node has been
5777 required_kernelcore -= min(required_kernelcore,
5779 kernelcore_remaining -= size_pages;
5780 if (!kernelcore_remaining)
5786 * If there is still required_kernelcore, we do another pass with one
5787 * less node in the count. This will push zone_movable_pfn[nid] further
5788 * along on the nodes that still have memory until kernelcore is
5792 if (usable_nodes && required_kernelcore > usable_nodes)
5796 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5797 for (nid = 0; nid < MAX_NUMNODES; nid++)
5798 zone_movable_pfn[nid] =
5799 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5802 /* restore the node_state */
5803 node_states[N_MEMORY] = saved_node_state;
5806 /* Any regular or high memory on that node ? */
5807 static void check_for_memory(pg_data_t *pgdat, int nid)
5809 enum zone_type zone_type;
5811 if (N_MEMORY == N_NORMAL_MEMORY)
5814 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5815 struct zone *zone = &pgdat->node_zones[zone_type];
5816 if (populated_zone(zone)) {
5817 node_set_state(nid, N_HIGH_MEMORY);
5818 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5819 zone_type <= ZONE_NORMAL)
5820 node_set_state(nid, N_NORMAL_MEMORY);
5827 * free_area_init_nodes - Initialise all pg_data_t and zone data
5828 * @max_zone_pfn: an array of max PFNs for each zone
5830 * This will call free_area_init_node() for each active node in the system.
5831 * Using the page ranges provided by memblock_set_node(), the size of each
5832 * zone in each node and their holes is calculated. If the maximum PFN
5833 * between two adjacent zones match, it is assumed that the zone is empty.
5834 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5835 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5836 * starts where the previous one ended. For example, ZONE_DMA32 starts
5837 * at arch_max_dma_pfn.
5839 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5841 unsigned long start_pfn, end_pfn;
5844 /* Record where the zone boundaries are */
5845 memset(arch_zone_lowest_possible_pfn, 0,
5846 sizeof(arch_zone_lowest_possible_pfn));
5847 memset(arch_zone_highest_possible_pfn, 0,
5848 sizeof(arch_zone_highest_possible_pfn));
5849 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5850 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5851 for (i = 1; i < MAX_NR_ZONES; i++) {
5852 if (i == ZONE_MOVABLE)
5854 arch_zone_lowest_possible_pfn[i] =
5855 arch_zone_highest_possible_pfn[i-1];
5856 arch_zone_highest_possible_pfn[i] =
5857 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5859 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5860 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5862 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5863 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5864 find_zone_movable_pfns_for_nodes();
5866 /* Print out the zone ranges */
5867 pr_info("Zone ranges:\n");
5868 for (i = 0; i < MAX_NR_ZONES; i++) {
5869 if (i == ZONE_MOVABLE)
5871 pr_info(" %-8s ", zone_names[i]);
5872 if (arch_zone_lowest_possible_pfn[i] ==
5873 arch_zone_highest_possible_pfn[i])
5876 pr_cont("[mem %#018Lx-%#018Lx]\n",
5877 (u64)arch_zone_lowest_possible_pfn[i]
5879 ((u64)arch_zone_highest_possible_pfn[i]
5880 << PAGE_SHIFT) - 1);
5883 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5884 pr_info("Movable zone start for each node\n");
5885 for (i = 0; i < MAX_NUMNODES; i++) {
5886 if (zone_movable_pfn[i])
5887 pr_info(" Node %d: %#018Lx\n", i,
5888 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5891 /* Print out the early node map */
5892 pr_info("Early memory node ranges\n");
5893 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5894 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5895 (u64)start_pfn << PAGE_SHIFT,
5896 ((u64)end_pfn << PAGE_SHIFT) - 1);
5898 /* Initialise every node */
5899 mminit_verify_pageflags_layout();
5900 setup_nr_node_ids();
5901 for_each_online_node(nid) {
5902 pg_data_t *pgdat = NODE_DATA(nid);
5903 free_area_init_node(nid, NULL,
5904 find_min_pfn_for_node(nid), NULL);
5906 /* Any memory on that node */
5907 if (pgdat->node_present_pages)
5908 node_set_state(nid, N_MEMORY);
5909 check_for_memory(pgdat, nid);
5913 static int __init cmdline_parse_core(char *p, unsigned long *core)
5915 unsigned long long coremem;
5919 coremem = memparse(p, &p);
5920 *core = coremem >> PAGE_SHIFT;
5922 /* Paranoid check that UL is enough for the coremem value */
5923 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5929 * kernelcore=size sets the amount of memory for use for allocations that
5930 * cannot be reclaimed or migrated.
5932 static int __init cmdline_parse_kernelcore(char *p)
5934 /* parse kernelcore=mirror */
5935 if (parse_option_str(p, "mirror")) {
5936 mirrored_kernelcore = true;
5940 return cmdline_parse_core(p, &required_kernelcore);
5944 * movablecore=size sets the amount of memory for use for allocations that
5945 * can be reclaimed or migrated.
5947 static int __init cmdline_parse_movablecore(char *p)
5949 return cmdline_parse_core(p, &required_movablecore);
5952 early_param("kernelcore", cmdline_parse_kernelcore);
5953 early_param("movablecore", cmdline_parse_movablecore);
5955 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5957 void adjust_managed_page_count(struct page *page, long count)
5959 spin_lock(&managed_page_count_lock);
5960 page_zone(page)->managed_pages += count;
5961 totalram_pages += count;
5962 #ifdef CONFIG_HIGHMEM
5963 if (PageHighMem(page))
5964 totalhigh_pages += count;
5966 spin_unlock(&managed_page_count_lock);
5968 EXPORT_SYMBOL(adjust_managed_page_count);
5970 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5973 unsigned long pages = 0;
5975 start = (void *)PAGE_ALIGN((unsigned long)start);
5976 end = (void *)((unsigned long)end & PAGE_MASK);
5977 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5978 if ((unsigned int)poison <= 0xFF)
5979 memset(pos, poison, PAGE_SIZE);
5980 free_reserved_page(virt_to_page(pos));
5984 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5985 s, pages << (PAGE_SHIFT - 10), start, end);
5989 EXPORT_SYMBOL(free_reserved_area);
5991 #ifdef CONFIG_HIGHMEM
5992 void free_highmem_page(struct page *page)
5994 __free_reserved_page(page);
5996 page_zone(page)->managed_pages++;
6002 void __init mem_init_print_info(const char *str)
6004 unsigned long physpages, codesize, datasize, rosize, bss_size;
6005 unsigned long init_code_size, init_data_size;
6007 physpages = get_num_physpages();
6008 codesize = _etext - _stext;
6009 datasize = _edata - _sdata;
6010 rosize = __end_rodata - __start_rodata;
6011 bss_size = __bss_stop - __bss_start;
6012 init_data_size = __init_end - __init_begin;
6013 init_code_size = _einittext - _sinittext;
6016 * Detect special cases and adjust section sizes accordingly:
6017 * 1) .init.* may be embedded into .data sections
6018 * 2) .init.text.* may be out of [__init_begin, __init_end],
6019 * please refer to arch/tile/kernel/vmlinux.lds.S.
6020 * 3) .rodata.* may be embedded into .text or .data sections.
6022 #define adj_init_size(start, end, size, pos, adj) \
6024 if (start <= pos && pos < end && size > adj) \
6028 adj_init_size(__init_begin, __init_end, init_data_size,
6029 _sinittext, init_code_size);
6030 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6031 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6032 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6033 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6035 #undef adj_init_size
6037 pr_info("Memory: %luK/%luK available "
6038 "(%luK kernel code, %luK rwdata, %luK rodata, "
6039 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
6040 #ifdef CONFIG_HIGHMEM
6044 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
6045 codesize >> 10, datasize >> 10, rosize >> 10,
6046 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6047 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
6048 totalcma_pages << (PAGE_SHIFT-10),
6049 #ifdef CONFIG_HIGHMEM
6050 totalhigh_pages << (PAGE_SHIFT-10),
6052 str ? ", " : "", str ? str : "");
6056 * set_dma_reserve - set the specified number of pages reserved in the first zone
6057 * @new_dma_reserve: The number of pages to mark reserved
6059 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6060 * In the DMA zone, a significant percentage may be consumed by kernel image
6061 * and other unfreeable allocations which can skew the watermarks badly. This
6062 * function may optionally be used to account for unfreeable pages in the
6063 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6064 * smaller per-cpu batchsize.
6066 void __init set_dma_reserve(unsigned long new_dma_reserve)
6068 dma_reserve = new_dma_reserve;
6071 void __init free_area_init(unsigned long *zones_size)
6073 free_area_init_node(0, zones_size,
6074 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6077 static int page_alloc_cpu_notify(struct notifier_block *self,
6078 unsigned long action, void *hcpu)
6080 int cpu = (unsigned long)hcpu;
6082 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6083 lru_add_drain_cpu(cpu);
6087 * Spill the event counters of the dead processor
6088 * into the current processors event counters.
6089 * This artificially elevates the count of the current
6092 vm_events_fold_cpu(cpu);
6095 * Zero the differential counters of the dead processor
6096 * so that the vm statistics are consistent.
6098 * This is only okay since the processor is dead and cannot
6099 * race with what we are doing.
6101 cpu_vm_stats_fold(cpu);
6106 void __init page_alloc_init(void)
6108 hotcpu_notifier(page_alloc_cpu_notify, 0);
6112 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6113 * or min_free_kbytes changes.
6115 static void calculate_totalreserve_pages(void)
6117 struct pglist_data *pgdat;
6118 unsigned long reserve_pages = 0;
6119 enum zone_type i, j;
6121 for_each_online_pgdat(pgdat) {
6122 for (i = 0; i < MAX_NR_ZONES; i++) {
6123 struct zone *zone = pgdat->node_zones + i;
6126 /* Find valid and maximum lowmem_reserve in the zone */
6127 for (j = i; j < MAX_NR_ZONES; j++) {
6128 if (zone->lowmem_reserve[j] > max)
6129 max = zone->lowmem_reserve[j];
6132 /* we treat the high watermark as reserved pages. */
6133 max += high_wmark_pages(zone);
6135 if (max > zone->managed_pages)
6136 max = zone->managed_pages;
6138 zone->totalreserve_pages = max;
6140 reserve_pages += max;
6143 totalreserve_pages = reserve_pages;
6147 * setup_per_zone_lowmem_reserve - called whenever
6148 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6149 * has a correct pages reserved value, so an adequate number of
6150 * pages are left in the zone after a successful __alloc_pages().
6152 static void setup_per_zone_lowmem_reserve(void)
6154 struct pglist_data *pgdat;
6155 enum zone_type j, idx;
6157 for_each_online_pgdat(pgdat) {
6158 for (j = 0; j < MAX_NR_ZONES; j++) {
6159 struct zone *zone = pgdat->node_zones + j;
6160 unsigned long managed_pages = zone->managed_pages;
6162 zone->lowmem_reserve[j] = 0;
6166 struct zone *lower_zone;
6170 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6171 sysctl_lowmem_reserve_ratio[idx] = 1;
6173 lower_zone = pgdat->node_zones + idx;
6174 lower_zone->lowmem_reserve[j] = managed_pages /
6175 sysctl_lowmem_reserve_ratio[idx];
6176 managed_pages += lower_zone->managed_pages;
6181 /* update totalreserve_pages */
6182 calculate_totalreserve_pages();
6185 static void __setup_per_zone_wmarks(void)
6187 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6188 unsigned long lowmem_pages = 0;
6190 unsigned long flags;
6192 /* Calculate total number of !ZONE_HIGHMEM pages */
6193 for_each_zone(zone) {
6194 if (!is_highmem(zone))
6195 lowmem_pages += zone->managed_pages;
6198 for_each_zone(zone) {
6201 spin_lock_irqsave(&zone->lock, flags);
6202 tmp = (u64)pages_min * zone->managed_pages;
6203 do_div(tmp, lowmem_pages);
6204 if (is_highmem(zone)) {
6206 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6207 * need highmem pages, so cap pages_min to a small
6210 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6211 * deltas control asynch page reclaim, and so should
6212 * not be capped for highmem.
6214 unsigned long min_pages;
6216 min_pages = zone->managed_pages / 1024;
6217 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6218 zone->watermark[WMARK_MIN] = min_pages;
6221 * If it's a lowmem zone, reserve a number of pages
6222 * proportionate to the zone's size.
6224 zone->watermark[WMARK_MIN] = tmp;
6227 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
6228 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6230 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6231 high_wmark_pages(zone) - low_wmark_pages(zone) -
6232 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6234 spin_unlock_irqrestore(&zone->lock, flags);
6237 /* update totalreserve_pages */
6238 calculate_totalreserve_pages();
6242 * setup_per_zone_wmarks - called when min_free_kbytes changes
6243 * or when memory is hot-{added|removed}
6245 * Ensures that the watermark[min,low,high] values for each zone are set
6246 * correctly with respect to min_free_kbytes.
6248 void setup_per_zone_wmarks(void)
6250 mutex_lock(&zonelists_mutex);
6251 __setup_per_zone_wmarks();
6252 mutex_unlock(&zonelists_mutex);
6256 * The inactive anon list should be small enough that the VM never has to
6257 * do too much work, but large enough that each inactive page has a chance
6258 * to be referenced again before it is swapped out.
6260 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6261 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6262 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6263 * the anonymous pages are kept on the inactive list.
6266 * memory ratio inactive anon
6267 * -------------------------------------
6276 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6278 unsigned int gb, ratio;
6280 /* Zone size in gigabytes */
6281 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6283 ratio = int_sqrt(10 * gb);
6287 zone->inactive_ratio = ratio;
6290 static void __meminit setup_per_zone_inactive_ratio(void)
6295 calculate_zone_inactive_ratio(zone);
6299 * Initialise min_free_kbytes.
6301 * For small machines we want it small (128k min). For large machines
6302 * we want it large (64MB max). But it is not linear, because network
6303 * bandwidth does not increase linearly with machine size. We use
6305 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6306 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6322 int __meminit init_per_zone_wmark_min(void)
6324 unsigned long lowmem_kbytes;
6325 int new_min_free_kbytes;
6327 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6328 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6330 if (new_min_free_kbytes > user_min_free_kbytes) {
6331 min_free_kbytes = new_min_free_kbytes;
6332 if (min_free_kbytes < 128)
6333 min_free_kbytes = 128;
6334 if (min_free_kbytes > 65536)
6335 min_free_kbytes = 65536;
6337 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6338 new_min_free_kbytes, user_min_free_kbytes);
6340 setup_per_zone_wmarks();
6341 refresh_zone_stat_thresholds();
6342 setup_per_zone_lowmem_reserve();
6343 setup_per_zone_inactive_ratio();
6346 module_init(init_per_zone_wmark_min)
6349 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6350 * that we can call two helper functions whenever min_free_kbytes
6353 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6354 void __user *buffer, size_t *length, loff_t *ppos)
6358 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6363 user_min_free_kbytes = min_free_kbytes;
6364 setup_per_zone_wmarks();
6370 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6371 void __user *buffer, size_t *length, loff_t *ppos)
6376 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6381 zone->min_unmapped_pages = (zone->managed_pages *
6382 sysctl_min_unmapped_ratio) / 100;
6386 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6387 void __user *buffer, size_t *length, loff_t *ppos)
6392 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6397 zone->min_slab_pages = (zone->managed_pages *
6398 sysctl_min_slab_ratio) / 100;
6404 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6405 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6406 * whenever sysctl_lowmem_reserve_ratio changes.
6408 * The reserve ratio obviously has absolutely no relation with the
6409 * minimum watermarks. The lowmem reserve ratio can only make sense
6410 * if in function of the boot time zone sizes.
6412 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6413 void __user *buffer, size_t *length, loff_t *ppos)
6415 proc_dointvec_minmax(table, write, buffer, length, ppos);
6416 setup_per_zone_lowmem_reserve();
6421 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6422 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6423 * pagelist can have before it gets flushed back to buddy allocator.
6425 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6426 void __user *buffer, size_t *length, loff_t *ppos)
6429 int old_percpu_pagelist_fraction;
6432 mutex_lock(&pcp_batch_high_lock);
6433 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6435 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6436 if (!write || ret < 0)
6439 /* Sanity checking to avoid pcp imbalance */
6440 if (percpu_pagelist_fraction &&
6441 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6442 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6448 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6451 for_each_populated_zone(zone) {
6454 for_each_possible_cpu(cpu)
6455 pageset_set_high_and_batch(zone,
6456 per_cpu_ptr(zone->pageset, cpu));
6459 mutex_unlock(&pcp_batch_high_lock);
6464 int hashdist = HASHDIST_DEFAULT;
6466 static int __init set_hashdist(char *str)
6470 hashdist = simple_strtoul(str, &str, 0);
6473 __setup("hashdist=", set_hashdist);
6477 * allocate a large system hash table from bootmem
6478 * - it is assumed that the hash table must contain an exact power-of-2
6479 * quantity of entries
6480 * - limit is the number of hash buckets, not the total allocation size
6482 void *__init alloc_large_system_hash(const char *tablename,
6483 unsigned long bucketsize,
6484 unsigned long numentries,
6487 unsigned int *_hash_shift,
6488 unsigned int *_hash_mask,
6489 unsigned long low_limit,
6490 unsigned long high_limit)
6492 unsigned long long max = high_limit;
6493 unsigned long log2qty, size;
6496 /* allow the kernel cmdline to have a say */
6498 /* round applicable memory size up to nearest megabyte */
6499 numentries = nr_kernel_pages;
6501 /* It isn't necessary when PAGE_SIZE >= 1MB */
6502 if (PAGE_SHIFT < 20)
6503 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6505 /* limit to 1 bucket per 2^scale bytes of low memory */
6506 if (scale > PAGE_SHIFT)
6507 numentries >>= (scale - PAGE_SHIFT);
6509 numentries <<= (PAGE_SHIFT - scale);
6511 /* Make sure we've got at least a 0-order allocation.. */
6512 if (unlikely(flags & HASH_SMALL)) {
6513 /* Makes no sense without HASH_EARLY */
6514 WARN_ON(!(flags & HASH_EARLY));
6515 if (!(numentries >> *_hash_shift)) {
6516 numentries = 1UL << *_hash_shift;
6517 BUG_ON(!numentries);
6519 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6520 numentries = PAGE_SIZE / bucketsize;
6522 numentries = roundup_pow_of_two(numentries);
6524 /* limit allocation size to 1/16 total memory by default */
6526 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6527 do_div(max, bucketsize);
6529 max = min(max, 0x80000000ULL);
6531 if (numentries < low_limit)
6532 numentries = low_limit;
6533 if (numentries > max)
6536 log2qty = ilog2(numentries);
6539 size = bucketsize << log2qty;
6540 if (flags & HASH_EARLY)
6541 table = memblock_virt_alloc_nopanic(size, 0);
6543 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6546 * If bucketsize is not a power-of-two, we may free
6547 * some pages at the end of hash table which
6548 * alloc_pages_exact() automatically does
6550 if (get_order(size) < MAX_ORDER) {
6551 table = alloc_pages_exact(size, GFP_ATOMIC);
6552 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6555 } while (!table && size > PAGE_SIZE && --log2qty);
6558 panic("Failed to allocate %s hash table\n", tablename);
6560 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6563 ilog2(size) - PAGE_SHIFT,
6567 *_hash_shift = log2qty;
6569 *_hash_mask = (1 << log2qty) - 1;
6574 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6575 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6578 #ifdef CONFIG_SPARSEMEM
6579 return __pfn_to_section(pfn)->pageblock_flags;
6581 return zone->pageblock_flags;
6582 #endif /* CONFIG_SPARSEMEM */
6585 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6587 #ifdef CONFIG_SPARSEMEM
6588 pfn &= (PAGES_PER_SECTION-1);
6589 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6591 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6592 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6593 #endif /* CONFIG_SPARSEMEM */
6597 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6598 * @page: The page within the block of interest
6599 * @pfn: The target page frame number
6600 * @end_bitidx: The last bit of interest to retrieve
6601 * @mask: mask of bits that the caller is interested in
6603 * Return: pageblock_bits flags
6605 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6606 unsigned long end_bitidx,
6610 unsigned long *bitmap;
6611 unsigned long bitidx, word_bitidx;
6614 zone = page_zone(page);
6615 bitmap = get_pageblock_bitmap(zone, pfn);
6616 bitidx = pfn_to_bitidx(zone, pfn);
6617 word_bitidx = bitidx / BITS_PER_LONG;
6618 bitidx &= (BITS_PER_LONG-1);
6620 word = bitmap[word_bitidx];
6621 bitidx += end_bitidx;
6622 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6626 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6627 * @page: The page within the block of interest
6628 * @flags: The flags to set
6629 * @pfn: The target page frame number
6630 * @end_bitidx: The last bit of interest
6631 * @mask: mask of bits that the caller is interested in
6633 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6635 unsigned long end_bitidx,
6639 unsigned long *bitmap;
6640 unsigned long bitidx, word_bitidx;
6641 unsigned long old_word, word;
6643 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6645 zone = page_zone(page);
6646 bitmap = get_pageblock_bitmap(zone, pfn);
6647 bitidx = pfn_to_bitidx(zone, pfn);
6648 word_bitidx = bitidx / BITS_PER_LONG;
6649 bitidx &= (BITS_PER_LONG-1);
6651 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6653 bitidx += end_bitidx;
6654 mask <<= (BITS_PER_LONG - bitidx - 1);
6655 flags <<= (BITS_PER_LONG - bitidx - 1);
6657 word = READ_ONCE(bitmap[word_bitidx]);
6659 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6660 if (word == old_word)
6667 * This function checks whether pageblock includes unmovable pages or not.
6668 * If @count is not zero, it is okay to include less @count unmovable pages
6670 * PageLRU check without isolation or lru_lock could race so that
6671 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6672 * expect this function should be exact.
6674 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6675 bool skip_hwpoisoned_pages)
6677 unsigned long pfn, iter, found;
6681 * For avoiding noise data, lru_add_drain_all() should be called
6682 * If ZONE_MOVABLE, the zone never contains unmovable pages
6684 if (zone_idx(zone) == ZONE_MOVABLE)
6686 mt = get_pageblock_migratetype(page);
6687 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6690 pfn = page_to_pfn(page);
6691 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6692 unsigned long check = pfn + iter;
6694 if (!pfn_valid_within(check))
6697 page = pfn_to_page(check);
6700 * Hugepages are not in LRU lists, but they're movable.
6701 * We need not scan over tail pages bacause we don't
6702 * handle each tail page individually in migration.
6704 if (PageHuge(page)) {
6705 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6710 * We can't use page_count without pin a page
6711 * because another CPU can free compound page.
6712 * This check already skips compound tails of THP
6713 * because their page->_count is zero at all time.
6715 if (!atomic_read(&page->_count)) {
6716 if (PageBuddy(page))
6717 iter += (1 << page_order(page)) - 1;
6722 * The HWPoisoned page may be not in buddy system, and
6723 * page_count() is not 0.
6725 if (skip_hwpoisoned_pages && PageHWPoison(page))
6731 * If there are RECLAIMABLE pages, we need to check
6732 * it. But now, memory offline itself doesn't call
6733 * shrink_node_slabs() and it still to be fixed.
6736 * If the page is not RAM, page_count()should be 0.
6737 * we don't need more check. This is an _used_ not-movable page.
6739 * The problematic thing here is PG_reserved pages. PG_reserved
6740 * is set to both of a memory hole page and a _used_ kernel
6749 bool is_pageblock_removable_nolock(struct page *page)
6755 * We have to be careful here because we are iterating over memory
6756 * sections which are not zone aware so we might end up outside of
6757 * the zone but still within the section.
6758 * We have to take care about the node as well. If the node is offline
6759 * its NODE_DATA will be NULL - see page_zone.
6761 if (!node_online(page_to_nid(page)))
6764 zone = page_zone(page);
6765 pfn = page_to_pfn(page);
6766 if (!zone_spans_pfn(zone, pfn))
6769 return !has_unmovable_pages(zone, page, 0, true);
6772 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
6774 static unsigned long pfn_max_align_down(unsigned long pfn)
6776 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6777 pageblock_nr_pages) - 1);
6780 static unsigned long pfn_max_align_up(unsigned long pfn)
6782 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6783 pageblock_nr_pages));
6786 /* [start, end) must belong to a single zone. */
6787 static int __alloc_contig_migrate_range(struct compact_control *cc,
6788 unsigned long start, unsigned long end)
6790 /* This function is based on compact_zone() from compaction.c. */
6791 unsigned long nr_reclaimed;
6792 unsigned long pfn = start;
6793 unsigned int tries = 0;
6798 while (pfn < end || !list_empty(&cc->migratepages)) {
6799 if (fatal_signal_pending(current)) {
6804 if (list_empty(&cc->migratepages)) {
6805 cc->nr_migratepages = 0;
6806 pfn = isolate_migratepages_range(cc, pfn, end);
6812 } else if (++tries == 5) {
6813 ret = ret < 0 ? ret : -EBUSY;
6817 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6819 cc->nr_migratepages -= nr_reclaimed;
6821 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6822 NULL, 0, cc->mode, MR_CMA);
6825 putback_movable_pages(&cc->migratepages);
6832 * alloc_contig_range() -- tries to allocate given range of pages
6833 * @start: start PFN to allocate
6834 * @end: one-past-the-last PFN to allocate
6835 * @migratetype: migratetype of the underlaying pageblocks (either
6836 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6837 * in range must have the same migratetype and it must
6838 * be either of the two.
6840 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6841 * aligned, however it's the caller's responsibility to guarantee that
6842 * we are the only thread that changes migrate type of pageblocks the
6845 * The PFN range must belong to a single zone.
6847 * Returns zero on success or negative error code. On success all
6848 * pages which PFN is in [start, end) are allocated for the caller and
6849 * need to be freed with free_contig_range().
6851 int alloc_contig_range(unsigned long start, unsigned long end,
6852 unsigned migratetype)
6854 unsigned long outer_start, outer_end;
6858 struct compact_control cc = {
6859 .nr_migratepages = 0,
6861 .zone = page_zone(pfn_to_page(start)),
6862 .mode = MIGRATE_SYNC,
6863 .ignore_skip_hint = true,
6865 INIT_LIST_HEAD(&cc.migratepages);
6868 * What we do here is we mark all pageblocks in range as
6869 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6870 * have different sizes, and due to the way page allocator
6871 * work, we align the range to biggest of the two pages so
6872 * that page allocator won't try to merge buddies from
6873 * different pageblocks and change MIGRATE_ISOLATE to some
6874 * other migration type.
6876 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6877 * migrate the pages from an unaligned range (ie. pages that
6878 * we are interested in). This will put all the pages in
6879 * range back to page allocator as MIGRATE_ISOLATE.
6881 * When this is done, we take the pages in range from page
6882 * allocator removing them from the buddy system. This way
6883 * page allocator will never consider using them.
6885 * This lets us mark the pageblocks back as
6886 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6887 * aligned range but not in the unaligned, original range are
6888 * put back to page allocator so that buddy can use them.
6891 ret = start_isolate_page_range(pfn_max_align_down(start),
6892 pfn_max_align_up(end), migratetype,
6898 * In case of -EBUSY, we'd like to know which page causes problem.
6899 * So, just fall through. We will check it in test_pages_isolated().
6901 ret = __alloc_contig_migrate_range(&cc, start, end);
6902 if (ret && ret != -EBUSY)
6906 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6907 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6908 * more, all pages in [start, end) are free in page allocator.
6909 * What we are going to do is to allocate all pages from
6910 * [start, end) (that is remove them from page allocator).
6912 * The only problem is that pages at the beginning and at the
6913 * end of interesting range may be not aligned with pages that
6914 * page allocator holds, ie. they can be part of higher order
6915 * pages. Because of this, we reserve the bigger range and
6916 * once this is done free the pages we are not interested in.
6918 * We don't have to hold zone->lock here because the pages are
6919 * isolated thus they won't get removed from buddy.
6922 lru_add_drain_all();
6923 drain_all_pages(cc.zone);
6926 outer_start = start;
6927 while (!PageBuddy(pfn_to_page(outer_start))) {
6928 if (++order >= MAX_ORDER) {
6929 outer_start = start;
6932 outer_start &= ~0UL << order;
6935 if (outer_start != start) {
6936 order = page_order(pfn_to_page(outer_start));
6939 * outer_start page could be small order buddy page and
6940 * it doesn't include start page. Adjust outer_start
6941 * in this case to report failed page properly
6942 * on tracepoint in test_pages_isolated()
6944 if (outer_start + (1UL << order) <= start)
6945 outer_start = start;
6948 /* Make sure the range is really isolated. */
6949 if (test_pages_isolated(outer_start, end, false)) {
6950 pr_info("%s: [%lx, %lx) PFNs busy\n",
6951 __func__, outer_start, end);
6956 /* Grab isolated pages from freelists. */
6957 outer_end = isolate_freepages_range(&cc, outer_start, end);
6963 /* Free head and tail (if any) */
6964 if (start != outer_start)
6965 free_contig_range(outer_start, start - outer_start);
6966 if (end != outer_end)
6967 free_contig_range(end, outer_end - end);
6970 undo_isolate_page_range(pfn_max_align_down(start),
6971 pfn_max_align_up(end), migratetype);
6975 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6977 unsigned int count = 0;
6979 for (; nr_pages--; pfn++) {
6980 struct page *page = pfn_to_page(pfn);
6982 count += page_count(page) != 1;
6985 WARN(count != 0, "%d pages are still in use!\n", count);
6989 #ifdef CONFIG_MEMORY_HOTPLUG
6991 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6992 * page high values need to be recalulated.
6994 void __meminit zone_pcp_update(struct zone *zone)
6997 mutex_lock(&pcp_batch_high_lock);
6998 for_each_possible_cpu(cpu)
6999 pageset_set_high_and_batch(zone,
7000 per_cpu_ptr(zone->pageset, cpu));
7001 mutex_unlock(&pcp_batch_high_lock);
7005 void zone_pcp_reset(struct zone *zone)
7007 unsigned long flags;
7009 struct per_cpu_pageset *pset;
7011 /* avoid races with drain_pages() */
7012 local_irq_save(flags);
7013 if (zone->pageset != &boot_pageset) {
7014 for_each_online_cpu(cpu) {
7015 pset = per_cpu_ptr(zone->pageset, cpu);
7016 drain_zonestat(zone, pset);
7018 free_percpu(zone->pageset);
7019 zone->pageset = &boot_pageset;
7021 local_irq_restore(flags);
7024 #ifdef CONFIG_MEMORY_HOTREMOVE
7026 * All pages in the range must be isolated before calling this.
7029 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7033 unsigned int order, i;
7035 unsigned long flags;
7036 /* find the first valid pfn */
7037 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7042 zone = page_zone(pfn_to_page(pfn));
7043 spin_lock_irqsave(&zone->lock, flags);
7045 while (pfn < end_pfn) {
7046 if (!pfn_valid(pfn)) {
7050 page = pfn_to_page(pfn);
7052 * The HWPoisoned page may be not in buddy system, and
7053 * page_count() is not 0.
7055 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7057 SetPageReserved(page);
7061 BUG_ON(page_count(page));
7062 BUG_ON(!PageBuddy(page));
7063 order = page_order(page);
7064 #ifdef CONFIG_DEBUG_VM
7065 printk(KERN_INFO "remove from free list %lx %d %lx\n",
7066 pfn, 1 << order, end_pfn);
7068 list_del(&page->lru);
7069 rmv_page_order(page);
7070 zone->free_area[order].nr_free--;
7071 for (i = 0; i < (1 << order); i++)
7072 SetPageReserved((page+i));
7073 pfn += (1 << order);
7075 spin_unlock_irqrestore(&zone->lock, flags);
7079 #ifdef CONFIG_MEMORY_FAILURE
7080 bool is_free_buddy_page(struct page *page)
7082 struct zone *zone = page_zone(page);
7083 unsigned long pfn = page_to_pfn(page);
7084 unsigned long flags;
7087 spin_lock_irqsave(&zone->lock, flags);
7088 for (order = 0; order < MAX_ORDER; order++) {
7089 struct page *page_head = page - (pfn & ((1 << order) - 1));
7091 if (PageBuddy(page_head) && page_order(page_head) >= order)
7094 spin_unlock_irqrestore(&zone->lock, flags);
7096 return order < MAX_ORDER;