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_map_pages(page, 1 << order, 0);
1033 static void __free_pages_ok(struct page *page, unsigned int order)
1035 unsigned long flags;
1037 unsigned long pfn = page_to_pfn(page);
1039 if (!free_pages_prepare(page, order))
1042 migratetype = get_pfnblock_migratetype(page, pfn);
1043 local_irq_save(flags);
1044 __count_vm_events(PGFREE, 1 << order);
1045 free_one_page(page_zone(page), page, pfn, order, migratetype);
1046 local_irq_restore(flags);
1049 static void __init __free_pages_boot_core(struct page *page,
1050 unsigned long pfn, unsigned int order)
1052 unsigned int nr_pages = 1 << order;
1053 struct page *p = page;
1057 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1059 __ClearPageReserved(p);
1060 set_page_count(p, 0);
1062 __ClearPageReserved(p);
1063 set_page_count(p, 0);
1065 page_zone(page)->managed_pages += nr_pages;
1066 set_page_refcounted(page);
1067 __free_pages(page, order);
1070 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1071 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1073 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1075 int __meminit early_pfn_to_nid(unsigned long pfn)
1077 static DEFINE_SPINLOCK(early_pfn_lock);
1080 spin_lock(&early_pfn_lock);
1081 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1084 spin_unlock(&early_pfn_lock);
1090 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1091 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1092 struct mminit_pfnnid_cache *state)
1096 nid = __early_pfn_to_nid(pfn, state);
1097 if (nid >= 0 && nid != node)
1102 /* Only safe to use early in boot when initialisation is single-threaded */
1103 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1105 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1110 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1114 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1115 struct mminit_pfnnid_cache *state)
1122 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1125 if (early_page_uninitialised(pfn))
1127 return __free_pages_boot_core(page, pfn, order);
1130 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1131 static void __init deferred_free_range(struct page *page,
1132 unsigned long pfn, int nr_pages)
1139 /* Free a large naturally-aligned chunk if possible */
1140 if (nr_pages == MAX_ORDER_NR_PAGES &&
1141 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1142 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1143 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1147 for (i = 0; i < nr_pages; i++, page++, pfn++)
1148 __free_pages_boot_core(page, pfn, 0);
1151 /* Completion tracking for deferred_init_memmap() threads */
1152 static atomic_t pgdat_init_n_undone __initdata;
1153 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1155 static inline void __init pgdat_init_report_one_done(void)
1157 if (atomic_dec_and_test(&pgdat_init_n_undone))
1158 complete(&pgdat_init_all_done_comp);
1161 /* Initialise remaining memory on a node */
1162 static int __init deferred_init_memmap(void *data)
1164 pg_data_t *pgdat = data;
1165 int nid = pgdat->node_id;
1166 struct mminit_pfnnid_cache nid_init_state = { };
1167 unsigned long start = jiffies;
1168 unsigned long nr_pages = 0;
1169 unsigned long walk_start, walk_end;
1172 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1173 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1175 if (first_init_pfn == ULONG_MAX) {
1176 pgdat_init_report_one_done();
1180 /* Bind memory initialisation thread to a local node if possible */
1181 if (!cpumask_empty(cpumask))
1182 set_cpus_allowed_ptr(current, cpumask);
1184 /* Sanity check boundaries */
1185 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1186 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1187 pgdat->first_deferred_pfn = ULONG_MAX;
1189 /* Only the highest zone is deferred so find it */
1190 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1191 zone = pgdat->node_zones + zid;
1192 if (first_init_pfn < zone_end_pfn(zone))
1196 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1197 unsigned long pfn, end_pfn;
1198 struct page *page = NULL;
1199 struct page *free_base_page = NULL;
1200 unsigned long free_base_pfn = 0;
1203 end_pfn = min(walk_end, zone_end_pfn(zone));
1204 pfn = first_init_pfn;
1205 if (pfn < walk_start)
1207 if (pfn < zone->zone_start_pfn)
1208 pfn = zone->zone_start_pfn;
1210 for (; pfn < end_pfn; pfn++) {
1211 if (!pfn_valid_within(pfn))
1215 * Ensure pfn_valid is checked every
1216 * MAX_ORDER_NR_PAGES for memory holes
1218 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1219 if (!pfn_valid(pfn)) {
1225 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1230 /* Minimise pfn page lookups and scheduler checks */
1231 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1234 nr_pages += nr_to_free;
1235 deferred_free_range(free_base_page,
1236 free_base_pfn, nr_to_free);
1237 free_base_page = NULL;
1238 free_base_pfn = nr_to_free = 0;
1240 page = pfn_to_page(pfn);
1245 VM_BUG_ON(page_zone(page) != zone);
1249 __init_single_page(page, pfn, zid, nid);
1250 if (!free_base_page) {
1251 free_base_page = page;
1252 free_base_pfn = pfn;
1257 /* Where possible, batch up pages for a single free */
1260 /* Free the current block of pages to allocator */
1261 nr_pages += nr_to_free;
1262 deferred_free_range(free_base_page, free_base_pfn,
1264 free_base_page = NULL;
1265 free_base_pfn = nr_to_free = 0;
1268 first_init_pfn = max(end_pfn, first_init_pfn);
1271 /* Sanity check that the next zone really is unpopulated */
1272 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1274 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1275 jiffies_to_msecs(jiffies - start));
1277 pgdat_init_report_one_done();
1281 void __init page_alloc_init_late(void)
1285 /* There will be num_node_state(N_MEMORY) threads */
1286 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1287 for_each_node_state(nid, N_MEMORY) {
1288 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1291 /* Block until all are initialised */
1292 wait_for_completion(&pgdat_init_all_done_comp);
1294 /* Reinit limits that are based on free pages after the kernel is up */
1295 files_maxfiles_init();
1297 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1300 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1301 void __init init_cma_reserved_pageblock(struct page *page)
1303 unsigned i = pageblock_nr_pages;
1304 struct page *p = page;
1307 __ClearPageReserved(p);
1308 set_page_count(p, 0);
1311 set_pageblock_migratetype(page, MIGRATE_CMA);
1313 if (pageblock_order >= MAX_ORDER) {
1314 i = pageblock_nr_pages;
1317 set_page_refcounted(p);
1318 __free_pages(p, MAX_ORDER - 1);
1319 p += MAX_ORDER_NR_PAGES;
1320 } while (i -= MAX_ORDER_NR_PAGES);
1322 set_page_refcounted(page);
1323 __free_pages(page, pageblock_order);
1326 adjust_managed_page_count(page, pageblock_nr_pages);
1331 * The order of subdivision here is critical for the IO subsystem.
1332 * Please do not alter this order without good reasons and regression
1333 * testing. Specifically, as large blocks of memory are subdivided,
1334 * the order in which smaller blocks are delivered depends on the order
1335 * they're subdivided in this function. This is the primary factor
1336 * influencing the order in which pages are delivered to the IO
1337 * subsystem according to empirical testing, and this is also justified
1338 * by considering the behavior of a buddy system containing a single
1339 * large block of memory acted on by a series of small allocations.
1340 * This behavior is a critical factor in sglist merging's success.
1344 static inline void expand(struct zone *zone, struct page *page,
1345 int low, int high, struct free_area *area,
1348 unsigned long size = 1 << high;
1350 while (high > low) {
1354 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1356 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1357 debug_guardpage_enabled() &&
1358 high < debug_guardpage_minorder()) {
1360 * Mark as guard pages (or page), that will allow to
1361 * merge back to allocator when buddy will be freed.
1362 * Corresponding page table entries will not be touched,
1363 * pages will stay not present in virtual address space
1365 set_page_guard(zone, &page[size], high, migratetype);
1368 list_add(&page[size].lru, &area->free_list[migratetype]);
1370 set_page_order(&page[size], high);
1375 * This page is about to be returned from the page allocator
1377 static inline int check_new_page(struct page *page)
1379 const char *bad_reason = NULL;
1380 unsigned long bad_flags = 0;
1382 if (unlikely(atomic_read(&page->_mapcount) != -1))
1383 bad_reason = "nonzero mapcount";
1384 if (unlikely(page->mapping != NULL))
1385 bad_reason = "non-NULL mapping";
1386 if (unlikely(atomic_read(&page->_count) != 0))
1387 bad_reason = "nonzero _count";
1388 if (unlikely(page->flags & __PG_HWPOISON)) {
1389 bad_reason = "HWPoisoned (hardware-corrupted)";
1390 bad_flags = __PG_HWPOISON;
1392 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1393 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1394 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1397 if (unlikely(page->mem_cgroup))
1398 bad_reason = "page still charged to cgroup";
1400 if (unlikely(bad_reason)) {
1401 bad_page(page, bad_reason, bad_flags);
1407 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1412 for (i = 0; i < (1 << order); i++) {
1413 struct page *p = page + i;
1414 if (unlikely(check_new_page(p)))
1418 set_page_private(page, 0);
1419 set_page_refcounted(page);
1421 arch_alloc_page(page, order);
1422 kernel_map_pages(page, 1 << order, 1);
1423 kasan_alloc_pages(page, order);
1425 if (gfp_flags & __GFP_ZERO)
1426 for (i = 0; i < (1 << order); i++)
1427 clear_highpage(page + i);
1429 if (order && (gfp_flags & __GFP_COMP))
1430 prep_compound_page(page, order);
1432 set_page_owner(page, order, gfp_flags);
1435 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1436 * allocate the page. The expectation is that the caller is taking
1437 * steps that will free more memory. The caller should avoid the page
1438 * being used for !PFMEMALLOC purposes.
1440 if (alloc_flags & ALLOC_NO_WATERMARKS)
1441 set_page_pfmemalloc(page);
1443 clear_page_pfmemalloc(page);
1449 * Go through the free lists for the given migratetype and remove
1450 * the smallest available page from the freelists
1453 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1456 unsigned int current_order;
1457 struct free_area *area;
1460 /* Find a page of the appropriate size in the preferred list */
1461 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1462 area = &(zone->free_area[current_order]);
1463 page = list_first_entry_or_null(&area->free_list[migratetype],
1467 list_del(&page->lru);
1468 rmv_page_order(page);
1470 expand(zone, page, order, current_order, area, migratetype);
1471 set_pcppage_migratetype(page, migratetype);
1480 * This array describes the order lists are fallen back to when
1481 * the free lists for the desirable migrate type are depleted
1483 static int fallbacks[MIGRATE_TYPES][4] = {
1484 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1485 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1486 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1488 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1490 #ifdef CONFIG_MEMORY_ISOLATION
1491 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1496 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1499 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1502 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1503 unsigned int order) { return NULL; }
1507 * Move the free pages in a range to the free lists of the requested type.
1508 * Note that start_page and end_pages are not aligned on a pageblock
1509 * boundary. If alignment is required, use move_freepages_block()
1511 int move_freepages(struct zone *zone,
1512 struct page *start_page, struct page *end_page,
1517 int pages_moved = 0;
1519 #ifndef CONFIG_HOLES_IN_ZONE
1521 * page_zone is not safe to call in this context when
1522 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1523 * anyway as we check zone boundaries in move_freepages_block().
1524 * Remove at a later date when no bug reports exist related to
1525 * grouping pages by mobility
1527 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1530 for (page = start_page; page <= end_page;) {
1531 /* Make sure we are not inadvertently changing nodes */
1532 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1534 if (!pfn_valid_within(page_to_pfn(page))) {
1539 if (!PageBuddy(page)) {
1544 order = page_order(page);
1545 list_move(&page->lru,
1546 &zone->free_area[order].free_list[migratetype]);
1548 pages_moved += 1 << order;
1554 int move_freepages_block(struct zone *zone, struct page *page,
1557 unsigned long start_pfn, end_pfn;
1558 struct page *start_page, *end_page;
1560 start_pfn = page_to_pfn(page);
1561 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1562 start_page = pfn_to_page(start_pfn);
1563 end_page = start_page + pageblock_nr_pages - 1;
1564 end_pfn = start_pfn + pageblock_nr_pages - 1;
1566 /* Do not cross zone boundaries */
1567 if (!zone_spans_pfn(zone, start_pfn))
1569 if (!zone_spans_pfn(zone, end_pfn))
1572 return move_freepages(zone, start_page, end_page, migratetype);
1575 static void change_pageblock_range(struct page *pageblock_page,
1576 int start_order, int migratetype)
1578 int nr_pageblocks = 1 << (start_order - pageblock_order);
1580 while (nr_pageblocks--) {
1581 set_pageblock_migratetype(pageblock_page, migratetype);
1582 pageblock_page += pageblock_nr_pages;
1587 * When we are falling back to another migratetype during allocation, try to
1588 * steal extra free pages from the same pageblocks to satisfy further
1589 * allocations, instead of polluting multiple pageblocks.
1591 * If we are stealing a relatively large buddy page, it is likely there will
1592 * be more free pages in the pageblock, so try to steal them all. For
1593 * reclaimable and unmovable allocations, we steal regardless of page size,
1594 * as fragmentation caused by those allocations polluting movable pageblocks
1595 * is worse than movable allocations stealing from unmovable and reclaimable
1598 static bool can_steal_fallback(unsigned int order, int start_mt)
1601 * Leaving this order check is intended, although there is
1602 * relaxed order check in next check. The reason is that
1603 * we can actually steal whole pageblock if this condition met,
1604 * but, below check doesn't guarantee it and that is just heuristic
1605 * so could be changed anytime.
1607 if (order >= pageblock_order)
1610 if (order >= pageblock_order / 2 ||
1611 start_mt == MIGRATE_RECLAIMABLE ||
1612 start_mt == MIGRATE_UNMOVABLE ||
1613 page_group_by_mobility_disabled)
1620 * This function implements actual steal behaviour. If order is large enough,
1621 * we can steal whole pageblock. If not, we first move freepages in this
1622 * pageblock and check whether half of pages are moved or not. If half of
1623 * pages are moved, we can change migratetype of pageblock and permanently
1624 * use it's pages as requested migratetype in the future.
1626 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1629 unsigned int current_order = page_order(page);
1632 /* Take ownership for orders >= pageblock_order */
1633 if (current_order >= pageblock_order) {
1634 change_pageblock_range(page, current_order, start_type);
1638 pages = move_freepages_block(zone, page, start_type);
1640 /* Claim the whole block if over half of it is free */
1641 if (pages >= (1 << (pageblock_order-1)) ||
1642 page_group_by_mobility_disabled)
1643 set_pageblock_migratetype(page, start_type);
1647 * Check whether there is a suitable fallback freepage with requested order.
1648 * If only_stealable is true, this function returns fallback_mt only if
1649 * we can steal other freepages all together. This would help to reduce
1650 * fragmentation due to mixed migratetype pages in one pageblock.
1652 int find_suitable_fallback(struct free_area *area, unsigned int order,
1653 int migratetype, bool only_stealable, bool *can_steal)
1658 if (area->nr_free == 0)
1663 fallback_mt = fallbacks[migratetype][i];
1664 if (fallback_mt == MIGRATE_TYPES)
1667 if (list_empty(&area->free_list[fallback_mt]))
1670 if (can_steal_fallback(order, migratetype))
1673 if (!only_stealable)
1684 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1685 * there are no empty page blocks that contain a page with a suitable order
1687 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1688 unsigned int alloc_order)
1691 unsigned long max_managed, flags;
1694 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1695 * Check is race-prone but harmless.
1697 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1698 if (zone->nr_reserved_highatomic >= max_managed)
1701 spin_lock_irqsave(&zone->lock, flags);
1703 /* Recheck the nr_reserved_highatomic limit under the lock */
1704 if (zone->nr_reserved_highatomic >= max_managed)
1708 mt = get_pageblock_migratetype(page);
1709 if (mt != MIGRATE_HIGHATOMIC &&
1710 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1711 zone->nr_reserved_highatomic += pageblock_nr_pages;
1712 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1713 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1717 spin_unlock_irqrestore(&zone->lock, flags);
1721 * Used when an allocation is about to fail under memory pressure. This
1722 * potentially hurts the reliability of high-order allocations when under
1723 * intense memory pressure but failed atomic allocations should be easier
1724 * to recover from than an OOM.
1726 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1728 struct zonelist *zonelist = ac->zonelist;
1729 unsigned long flags;
1735 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1737 /* Preserve at least one pageblock */
1738 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1741 spin_lock_irqsave(&zone->lock, flags);
1742 for (order = 0; order < MAX_ORDER; order++) {
1743 struct free_area *area = &(zone->free_area[order]);
1745 page = list_first_entry_or_null(
1746 &area->free_list[MIGRATE_HIGHATOMIC],
1752 * It should never happen but changes to locking could
1753 * inadvertently allow a per-cpu drain to add pages
1754 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1755 * and watch for underflows.
1757 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1758 zone->nr_reserved_highatomic);
1761 * Convert to ac->migratetype and avoid the normal
1762 * pageblock stealing heuristics. Minimally, the caller
1763 * is doing the work and needs the pages. More
1764 * importantly, if the block was always converted to
1765 * MIGRATE_UNMOVABLE or another type then the number
1766 * of pageblocks that cannot be completely freed
1769 set_pageblock_migratetype(page, ac->migratetype);
1770 move_freepages_block(zone, page, ac->migratetype);
1771 spin_unlock_irqrestore(&zone->lock, flags);
1774 spin_unlock_irqrestore(&zone->lock, flags);
1778 /* Remove an element from the buddy allocator from the fallback list */
1779 static inline struct page *
1780 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1782 struct free_area *area;
1783 unsigned int current_order;
1788 /* Find the largest possible block of pages in the other list */
1789 for (current_order = MAX_ORDER-1;
1790 current_order >= order && current_order <= MAX_ORDER-1;
1792 area = &(zone->free_area[current_order]);
1793 fallback_mt = find_suitable_fallback(area, current_order,
1794 start_migratetype, false, &can_steal);
1795 if (fallback_mt == -1)
1798 page = list_first_entry(&area->free_list[fallback_mt],
1801 steal_suitable_fallback(zone, page, start_migratetype);
1803 /* Remove the page from the freelists */
1805 list_del(&page->lru);
1806 rmv_page_order(page);
1808 expand(zone, page, order, current_order, area,
1811 * The pcppage_migratetype may differ from pageblock's
1812 * migratetype depending on the decisions in
1813 * find_suitable_fallback(). This is OK as long as it does not
1814 * differ for MIGRATE_CMA pageblocks. Those can be used as
1815 * fallback only via special __rmqueue_cma_fallback() function
1817 set_pcppage_migratetype(page, start_migratetype);
1819 trace_mm_page_alloc_extfrag(page, order, current_order,
1820 start_migratetype, fallback_mt);
1829 * Do the hard work of removing an element from the buddy allocator.
1830 * Call me with the zone->lock already held.
1832 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1837 page = __rmqueue_smallest(zone, order, migratetype);
1838 if (unlikely(!page)) {
1839 if (migratetype == MIGRATE_MOVABLE)
1840 page = __rmqueue_cma_fallback(zone, order);
1843 page = __rmqueue_fallback(zone, order, migratetype);
1846 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1851 * Obtain a specified number of elements from the buddy allocator, all under
1852 * a single hold of the lock, for efficiency. Add them to the supplied list.
1853 * Returns the number of new pages which were placed at *list.
1855 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1856 unsigned long count, struct list_head *list,
1857 int migratetype, bool cold)
1861 spin_lock(&zone->lock);
1862 for (i = 0; i < count; ++i) {
1863 struct page *page = __rmqueue(zone, order, migratetype);
1864 if (unlikely(page == NULL))
1868 * Split buddy pages returned by expand() are received here
1869 * in physical page order. The page is added to the callers and
1870 * list and the list head then moves forward. From the callers
1871 * perspective, the linked list is ordered by page number in
1872 * some conditions. This is useful for IO devices that can
1873 * merge IO requests if the physical pages are ordered
1877 list_add(&page->lru, list);
1879 list_add_tail(&page->lru, list);
1881 if (is_migrate_cma(get_pcppage_migratetype(page)))
1882 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1885 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1886 spin_unlock(&zone->lock);
1892 * Called from the vmstat counter updater to drain pagesets of this
1893 * currently executing processor on remote nodes after they have
1896 * Note that this function must be called with the thread pinned to
1897 * a single processor.
1899 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1901 unsigned long flags;
1902 int to_drain, batch;
1904 local_irq_save(flags);
1905 batch = READ_ONCE(pcp->batch);
1906 to_drain = min(pcp->count, batch);
1908 free_pcppages_bulk(zone, to_drain, pcp);
1909 pcp->count -= to_drain;
1911 local_irq_restore(flags);
1916 * Drain pcplists of the indicated processor and zone.
1918 * The processor must either be the current processor and the
1919 * thread pinned to the current processor or a processor that
1922 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1924 unsigned long flags;
1925 struct per_cpu_pageset *pset;
1926 struct per_cpu_pages *pcp;
1928 local_irq_save(flags);
1929 pset = per_cpu_ptr(zone->pageset, cpu);
1933 free_pcppages_bulk(zone, pcp->count, pcp);
1936 local_irq_restore(flags);
1940 * Drain pcplists of all zones on the indicated processor.
1942 * The processor must either be the current processor and the
1943 * thread pinned to the current processor or a processor that
1946 static void drain_pages(unsigned int cpu)
1950 for_each_populated_zone(zone) {
1951 drain_pages_zone(cpu, zone);
1956 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1958 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1959 * the single zone's pages.
1961 void drain_local_pages(struct zone *zone)
1963 int cpu = smp_processor_id();
1966 drain_pages_zone(cpu, zone);
1972 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1974 * When zone parameter is non-NULL, spill just the single zone's pages.
1976 * Note that this code is protected against sending an IPI to an offline
1977 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1978 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1979 * nothing keeps CPUs from showing up after we populated the cpumask and
1980 * before the call to on_each_cpu_mask().
1982 void drain_all_pages(struct zone *zone)
1987 * Allocate in the BSS so we wont require allocation in
1988 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1990 static cpumask_t cpus_with_pcps;
1993 * We don't care about racing with CPU hotplug event
1994 * as offline notification will cause the notified
1995 * cpu to drain that CPU pcps and on_each_cpu_mask
1996 * disables preemption as part of its processing
1998 for_each_online_cpu(cpu) {
1999 struct per_cpu_pageset *pcp;
2001 bool has_pcps = false;
2004 pcp = per_cpu_ptr(zone->pageset, cpu);
2008 for_each_populated_zone(z) {
2009 pcp = per_cpu_ptr(z->pageset, cpu);
2010 if (pcp->pcp.count) {
2018 cpumask_set_cpu(cpu, &cpus_with_pcps);
2020 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2022 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2026 #ifdef CONFIG_HIBERNATION
2028 void mark_free_pages(struct zone *zone)
2030 unsigned long pfn, max_zone_pfn;
2031 unsigned long flags;
2032 unsigned int order, t;
2035 if (zone_is_empty(zone))
2038 spin_lock_irqsave(&zone->lock, flags);
2040 max_zone_pfn = zone_end_pfn(zone);
2041 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2042 if (pfn_valid(pfn)) {
2043 page = pfn_to_page(pfn);
2044 if (!swsusp_page_is_forbidden(page))
2045 swsusp_unset_page_free(page);
2048 for_each_migratetype_order(order, t) {
2049 list_for_each_entry(page,
2050 &zone->free_area[order].free_list[t], lru) {
2053 pfn = page_to_pfn(page);
2054 for (i = 0; i < (1UL << order); i++)
2055 swsusp_set_page_free(pfn_to_page(pfn + i));
2058 spin_unlock_irqrestore(&zone->lock, flags);
2060 #endif /* CONFIG_PM */
2063 * Free a 0-order page
2064 * cold == true ? free a cold page : free a hot page
2066 void free_hot_cold_page(struct page *page, bool cold)
2068 struct zone *zone = page_zone(page);
2069 struct per_cpu_pages *pcp;
2070 unsigned long flags;
2071 unsigned long pfn = page_to_pfn(page);
2074 if (!free_pages_prepare(page, 0))
2077 migratetype = get_pfnblock_migratetype(page, pfn);
2078 set_pcppage_migratetype(page, migratetype);
2079 local_irq_save(flags);
2080 __count_vm_event(PGFREE);
2083 * We only track unmovable, reclaimable and movable on pcp lists.
2084 * Free ISOLATE pages back to the allocator because they are being
2085 * offlined but treat RESERVE as movable pages so we can get those
2086 * areas back if necessary. Otherwise, we may have to free
2087 * excessively into the page allocator
2089 if (migratetype >= MIGRATE_PCPTYPES) {
2090 if (unlikely(is_migrate_isolate(migratetype))) {
2091 free_one_page(zone, page, pfn, 0, migratetype);
2094 migratetype = MIGRATE_MOVABLE;
2097 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2099 list_add(&page->lru, &pcp->lists[migratetype]);
2101 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2103 if (pcp->count >= pcp->high) {
2104 unsigned long batch = READ_ONCE(pcp->batch);
2105 free_pcppages_bulk(zone, batch, pcp);
2106 pcp->count -= batch;
2110 local_irq_restore(flags);
2114 * Free a list of 0-order pages
2116 void free_hot_cold_page_list(struct list_head *list, bool cold)
2118 struct page *page, *next;
2120 list_for_each_entry_safe(page, next, list, lru) {
2121 trace_mm_page_free_batched(page, cold);
2122 free_hot_cold_page(page, cold);
2127 * split_page takes a non-compound higher-order page, and splits it into
2128 * n (1<<order) sub-pages: page[0..n]
2129 * Each sub-page must be freed individually.
2131 * Note: this is probably too low level an operation for use in drivers.
2132 * Please consult with lkml before using this in your driver.
2134 void split_page(struct page *page, unsigned int order)
2139 VM_BUG_ON_PAGE(PageCompound(page), page);
2140 VM_BUG_ON_PAGE(!page_count(page), page);
2142 #ifdef CONFIG_KMEMCHECK
2144 * Split shadow pages too, because free(page[0]) would
2145 * otherwise free the whole shadow.
2147 if (kmemcheck_page_is_tracked(page))
2148 split_page(virt_to_page(page[0].shadow), order);
2151 gfp_mask = get_page_owner_gfp(page);
2152 set_page_owner(page, 0, gfp_mask);
2153 for (i = 1; i < (1 << order); i++) {
2154 set_page_refcounted(page + i);
2155 set_page_owner(page + i, 0, gfp_mask);
2158 EXPORT_SYMBOL_GPL(split_page);
2160 int __isolate_free_page(struct page *page, unsigned int order)
2162 unsigned long watermark;
2166 BUG_ON(!PageBuddy(page));
2168 zone = page_zone(page);
2169 mt = get_pageblock_migratetype(page);
2171 if (!is_migrate_isolate(mt)) {
2172 /* Obey watermarks as if the page was being allocated */
2173 watermark = low_wmark_pages(zone) + (1 << order);
2174 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2177 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2180 /* Remove page from free list */
2181 list_del(&page->lru);
2182 zone->free_area[order].nr_free--;
2183 rmv_page_order(page);
2185 set_page_owner(page, order, __GFP_MOVABLE);
2187 /* Set the pageblock if the isolated page is at least a pageblock */
2188 if (order >= pageblock_order - 1) {
2189 struct page *endpage = page + (1 << order) - 1;
2190 for (; page < endpage; page += pageblock_nr_pages) {
2191 int mt = get_pageblock_migratetype(page);
2192 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2193 set_pageblock_migratetype(page,
2199 return 1UL << order;
2203 * Similar to split_page except the page is already free. As this is only
2204 * being used for migration, the migratetype of the block also changes.
2205 * As this is called with interrupts disabled, the caller is responsible
2206 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2209 * Note: this is probably too low level an operation for use in drivers.
2210 * Please consult with lkml before using this in your driver.
2212 int split_free_page(struct page *page)
2217 order = page_order(page);
2219 nr_pages = __isolate_free_page(page, order);
2223 /* Split into individual pages */
2224 set_page_refcounted(page);
2225 split_page(page, order);
2230 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2233 struct page *buffered_rmqueue(struct zone *preferred_zone,
2234 struct zone *zone, unsigned int order,
2235 gfp_t gfp_flags, int alloc_flags, int migratetype)
2237 unsigned long flags;
2239 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2241 if (likely(order == 0)) {
2242 struct per_cpu_pages *pcp;
2243 struct list_head *list;
2245 local_irq_save(flags);
2246 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2247 list = &pcp->lists[migratetype];
2248 if (list_empty(list)) {
2249 pcp->count += rmqueue_bulk(zone, 0,
2252 if (unlikely(list_empty(list)))
2257 page = list_last_entry(list, struct page, lru);
2259 page = list_first_entry(list, struct page, lru);
2261 list_del(&page->lru);
2264 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2266 * __GFP_NOFAIL is not to be used in new code.
2268 * All __GFP_NOFAIL callers should be fixed so that they
2269 * properly detect and handle allocation failures.
2271 * We most definitely don't want callers attempting to
2272 * allocate greater than order-1 page units with
2275 WARN_ON_ONCE(order > 1);
2277 spin_lock_irqsave(&zone->lock, flags);
2280 if (alloc_flags & ALLOC_HARDER) {
2281 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2283 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2286 page = __rmqueue(zone, order, migratetype);
2287 spin_unlock(&zone->lock);
2290 __mod_zone_freepage_state(zone, -(1 << order),
2291 get_pcppage_migratetype(page));
2294 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2295 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2296 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2297 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2299 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2300 zone_statistics(preferred_zone, zone, gfp_flags);
2301 local_irq_restore(flags);
2303 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2307 local_irq_restore(flags);
2311 #ifdef CONFIG_FAIL_PAGE_ALLOC
2314 struct fault_attr attr;
2316 bool ignore_gfp_highmem;
2317 bool ignore_gfp_reclaim;
2319 } fail_page_alloc = {
2320 .attr = FAULT_ATTR_INITIALIZER,
2321 .ignore_gfp_reclaim = true,
2322 .ignore_gfp_highmem = true,
2326 static int __init setup_fail_page_alloc(char *str)
2328 return setup_fault_attr(&fail_page_alloc.attr, str);
2330 __setup("fail_page_alloc=", setup_fail_page_alloc);
2332 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2334 if (order < fail_page_alloc.min_order)
2336 if (gfp_mask & __GFP_NOFAIL)
2338 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2340 if (fail_page_alloc.ignore_gfp_reclaim &&
2341 (gfp_mask & __GFP_DIRECT_RECLAIM))
2344 return should_fail(&fail_page_alloc.attr, 1 << order);
2347 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2349 static int __init fail_page_alloc_debugfs(void)
2351 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2354 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2355 &fail_page_alloc.attr);
2357 return PTR_ERR(dir);
2359 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2360 &fail_page_alloc.ignore_gfp_reclaim))
2362 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2363 &fail_page_alloc.ignore_gfp_highmem))
2365 if (!debugfs_create_u32("min-order", mode, dir,
2366 &fail_page_alloc.min_order))
2371 debugfs_remove_recursive(dir);
2376 late_initcall(fail_page_alloc_debugfs);
2378 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2380 #else /* CONFIG_FAIL_PAGE_ALLOC */
2382 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2387 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2390 * Return true if free base pages are above 'mark'. For high-order checks it
2391 * will return true of the order-0 watermark is reached and there is at least
2392 * one free page of a suitable size. Checking now avoids taking the zone lock
2393 * to check in the allocation paths if no pages are free.
2395 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2396 unsigned long mark, int classzone_idx, int alloc_flags,
2401 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2403 /* free_pages may go negative - that's OK */
2404 free_pages -= (1 << order) - 1;
2406 if (alloc_flags & ALLOC_HIGH)
2410 * If the caller does not have rights to ALLOC_HARDER then subtract
2411 * the high-atomic reserves. This will over-estimate the size of the
2412 * atomic reserve but it avoids a search.
2414 if (likely(!alloc_harder))
2415 free_pages -= z->nr_reserved_highatomic;
2420 /* If allocation can't use CMA areas don't use free CMA pages */
2421 if (!(alloc_flags & ALLOC_CMA))
2422 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2426 * Check watermarks for an order-0 allocation request. If these
2427 * are not met, then a high-order request also cannot go ahead
2428 * even if a suitable page happened to be free.
2430 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2433 /* If this is an order-0 request then the watermark is fine */
2437 /* For a high-order request, check at least one suitable page is free */
2438 for (o = order; o < MAX_ORDER; o++) {
2439 struct free_area *area = &z->free_area[o];
2448 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2449 if (!list_empty(&area->free_list[mt]))
2454 if ((alloc_flags & ALLOC_CMA) &&
2455 !list_empty(&area->free_list[MIGRATE_CMA])) {
2463 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2464 int classzone_idx, int alloc_flags)
2466 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2467 zone_page_state(z, NR_FREE_PAGES));
2470 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2471 unsigned long mark, int classzone_idx)
2473 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2475 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2476 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2478 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2483 static bool zone_local(struct zone *local_zone, struct zone *zone)
2485 return local_zone->node == zone->node;
2488 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2490 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2493 #else /* CONFIG_NUMA */
2494 static bool zone_local(struct zone *local_zone, struct zone *zone)
2499 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2503 #endif /* CONFIG_NUMA */
2505 static void reset_alloc_batches(struct zone *preferred_zone)
2507 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2510 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2511 high_wmark_pages(zone) - low_wmark_pages(zone) -
2512 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2513 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2514 } while (zone++ != preferred_zone);
2518 * get_page_from_freelist goes through the zonelist trying to allocate
2521 static struct page *
2522 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2523 const struct alloc_context *ac)
2525 struct zonelist *zonelist = ac->zonelist;
2527 struct page *page = NULL;
2529 int nr_fair_skipped = 0;
2530 bool zonelist_rescan;
2533 zonelist_rescan = false;
2536 * Scan zonelist, looking for a zone with enough free.
2537 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2539 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2543 if (cpusets_enabled() &&
2544 (alloc_flags & ALLOC_CPUSET) &&
2545 !cpuset_zone_allowed(zone, gfp_mask))
2548 * Distribute pages in proportion to the individual
2549 * zone size to ensure fair page aging. The zone a
2550 * page was allocated in should have no effect on the
2551 * time the page has in memory before being reclaimed.
2553 if (alloc_flags & ALLOC_FAIR) {
2554 if (!zone_local(ac->preferred_zone, zone))
2556 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2562 * When allocating a page cache page for writing, we
2563 * want to get it from a zone that is within its dirty
2564 * limit, such that no single zone holds more than its
2565 * proportional share of globally allowed dirty pages.
2566 * The dirty limits take into account the zone's
2567 * lowmem reserves and high watermark so that kswapd
2568 * should be able to balance it without having to
2569 * write pages from its LRU list.
2571 * This may look like it could increase pressure on
2572 * lower zones by failing allocations in higher zones
2573 * before they are full. But the pages that do spill
2574 * over are limited as the lower zones are protected
2575 * by this very same mechanism. It should not become
2576 * a practical burden to them.
2578 * XXX: For now, allow allocations to potentially
2579 * exceed the per-zone dirty limit in the slowpath
2580 * (spread_dirty_pages unset) before going into reclaim,
2581 * which is important when on a NUMA setup the allowed
2582 * zones are together not big enough to reach the
2583 * global limit. The proper fix for these situations
2584 * will require awareness of zones in the
2585 * dirty-throttling and the flusher threads.
2587 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2590 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2591 if (!zone_watermark_ok(zone, order, mark,
2592 ac->classzone_idx, alloc_flags)) {
2595 /* Checked here to keep the fast path fast */
2596 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2597 if (alloc_flags & ALLOC_NO_WATERMARKS)
2600 if (zone_reclaim_mode == 0 ||
2601 !zone_allows_reclaim(ac->preferred_zone, zone))
2604 ret = zone_reclaim(zone, gfp_mask, order);
2606 case ZONE_RECLAIM_NOSCAN:
2609 case ZONE_RECLAIM_FULL:
2610 /* scanned but unreclaimable */
2613 /* did we reclaim enough */
2614 if (zone_watermark_ok(zone, order, mark,
2615 ac->classzone_idx, alloc_flags))
2623 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2624 gfp_mask, alloc_flags, ac->migratetype);
2626 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2630 * If this is a high-order atomic allocation then check
2631 * if the pageblock should be reserved for the future
2633 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2634 reserve_highatomic_pageblock(page, zone, order);
2641 * The first pass makes sure allocations are spread fairly within the
2642 * local node. However, the local node might have free pages left
2643 * after the fairness batches are exhausted, and remote zones haven't
2644 * even been considered yet. Try once more without fairness, and
2645 * include remote zones now, before entering the slowpath and waking
2646 * kswapd: prefer spilling to a remote zone over swapping locally.
2648 if (alloc_flags & ALLOC_FAIR) {
2649 alloc_flags &= ~ALLOC_FAIR;
2650 if (nr_fair_skipped) {
2651 zonelist_rescan = true;
2652 reset_alloc_batches(ac->preferred_zone);
2654 if (nr_online_nodes > 1)
2655 zonelist_rescan = true;
2658 if (zonelist_rescan)
2665 * Large machines with many possible nodes should not always dump per-node
2666 * meminfo in irq context.
2668 static inline bool should_suppress_show_mem(void)
2673 ret = in_interrupt();
2678 static DEFINE_RATELIMIT_STATE(nopage_rs,
2679 DEFAULT_RATELIMIT_INTERVAL,
2680 DEFAULT_RATELIMIT_BURST);
2682 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2684 unsigned int filter = SHOW_MEM_FILTER_NODES;
2686 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2687 debug_guardpage_minorder() > 0)
2691 * This documents exceptions given to allocations in certain
2692 * contexts that are allowed to allocate outside current's set
2695 if (!(gfp_mask & __GFP_NOMEMALLOC))
2696 if (test_thread_flag(TIF_MEMDIE) ||
2697 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2698 filter &= ~SHOW_MEM_FILTER_NODES;
2699 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2700 filter &= ~SHOW_MEM_FILTER_NODES;
2703 struct va_format vaf;
2706 va_start(args, fmt);
2711 pr_warn("%pV", &vaf);
2716 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
2717 current->comm, order, gfp_mask, &gfp_mask);
2719 if (!should_suppress_show_mem())
2723 static inline struct page *
2724 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2725 const struct alloc_context *ac, unsigned long *did_some_progress)
2727 struct oom_control oc = {
2728 .zonelist = ac->zonelist,
2729 .nodemask = ac->nodemask,
2730 .gfp_mask = gfp_mask,
2735 *did_some_progress = 0;
2738 * Acquire the oom lock. If that fails, somebody else is
2739 * making progress for us.
2741 if (!mutex_trylock(&oom_lock)) {
2742 *did_some_progress = 1;
2743 schedule_timeout_uninterruptible(1);
2748 * Go through the zonelist yet one more time, keep very high watermark
2749 * here, this is only to catch a parallel oom killing, we must fail if
2750 * we're still under heavy pressure.
2752 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2753 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2757 if (!(gfp_mask & __GFP_NOFAIL)) {
2758 /* Coredumps can quickly deplete all memory reserves */
2759 if (current->flags & PF_DUMPCORE)
2761 /* The OOM killer will not help higher order allocs */
2762 if (order > PAGE_ALLOC_COSTLY_ORDER)
2764 /* The OOM killer does not needlessly kill tasks for lowmem */
2765 if (ac->high_zoneidx < ZONE_NORMAL)
2767 /* The OOM killer does not compensate for IO-less reclaim */
2768 if (!(gfp_mask & __GFP_FS)) {
2770 * XXX: Page reclaim didn't yield anything,
2771 * and the OOM killer can't be invoked, but
2772 * keep looping as per tradition.
2774 *did_some_progress = 1;
2777 if (pm_suspended_storage())
2779 /* The OOM killer may not free memory on a specific node */
2780 if (gfp_mask & __GFP_THISNODE)
2783 /* Exhausted what can be done so it's blamo time */
2784 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
2785 *did_some_progress = 1;
2787 if (gfp_mask & __GFP_NOFAIL) {
2788 page = get_page_from_freelist(gfp_mask, order,
2789 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
2791 * fallback to ignore cpuset restriction if our nodes
2795 page = get_page_from_freelist(gfp_mask, order,
2796 ALLOC_NO_WATERMARKS, ac);
2800 mutex_unlock(&oom_lock);
2804 #ifdef CONFIG_COMPACTION
2805 /* Try memory compaction for high-order allocations before reclaim */
2806 static struct page *
2807 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2808 int alloc_flags, const struct alloc_context *ac,
2809 enum migrate_mode mode, int *contended_compaction,
2810 bool *deferred_compaction)
2812 unsigned long compact_result;
2818 current->flags |= PF_MEMALLOC;
2819 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2820 mode, contended_compaction);
2821 current->flags &= ~PF_MEMALLOC;
2823 switch (compact_result) {
2824 case COMPACT_DEFERRED:
2825 *deferred_compaction = true;
2827 case COMPACT_SKIPPED:
2834 * At least in one zone compaction wasn't deferred or skipped, so let's
2835 * count a compaction stall
2837 count_vm_event(COMPACTSTALL);
2839 page = get_page_from_freelist(gfp_mask, order,
2840 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2843 struct zone *zone = page_zone(page);
2845 zone->compact_blockskip_flush = false;
2846 compaction_defer_reset(zone, order, true);
2847 count_vm_event(COMPACTSUCCESS);
2852 * It's bad if compaction run occurs and fails. The most likely reason
2853 * is that pages exist, but not enough to satisfy watermarks.
2855 count_vm_event(COMPACTFAIL);
2862 static inline struct page *
2863 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2864 int alloc_flags, const struct alloc_context *ac,
2865 enum migrate_mode mode, int *contended_compaction,
2866 bool *deferred_compaction)
2870 #endif /* CONFIG_COMPACTION */
2872 /* Perform direct synchronous page reclaim */
2874 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2875 const struct alloc_context *ac)
2877 struct reclaim_state reclaim_state;
2882 /* We now go into synchronous reclaim */
2883 cpuset_memory_pressure_bump();
2884 current->flags |= PF_MEMALLOC;
2885 lockdep_set_current_reclaim_state(gfp_mask);
2886 reclaim_state.reclaimed_slab = 0;
2887 current->reclaim_state = &reclaim_state;
2889 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2892 current->reclaim_state = NULL;
2893 lockdep_clear_current_reclaim_state();
2894 current->flags &= ~PF_MEMALLOC;
2901 /* The really slow allocator path where we enter direct reclaim */
2902 static inline struct page *
2903 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2904 int alloc_flags, const struct alloc_context *ac,
2905 unsigned long *did_some_progress)
2907 struct page *page = NULL;
2908 bool drained = false;
2910 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2911 if (unlikely(!(*did_some_progress)))
2915 page = get_page_from_freelist(gfp_mask, order,
2916 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2919 * If an allocation failed after direct reclaim, it could be because
2920 * pages are pinned on the per-cpu lists or in high alloc reserves.
2921 * Shrink them them and try again
2923 if (!page && !drained) {
2924 unreserve_highatomic_pageblock(ac);
2925 drain_all_pages(NULL);
2933 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2938 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2939 ac->high_zoneidx, ac->nodemask)
2940 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2944 gfp_to_alloc_flags(gfp_t gfp_mask)
2946 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2948 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2949 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2952 * The caller may dip into page reserves a bit more if the caller
2953 * cannot run direct reclaim, or if the caller has realtime scheduling
2954 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2955 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
2957 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2959 if (gfp_mask & __GFP_ATOMIC) {
2961 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2962 * if it can't schedule.
2964 if (!(gfp_mask & __GFP_NOMEMALLOC))
2965 alloc_flags |= ALLOC_HARDER;
2967 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2968 * comment for __cpuset_node_allowed().
2970 alloc_flags &= ~ALLOC_CPUSET;
2971 } else if (unlikely(rt_task(current)) && !in_interrupt())
2972 alloc_flags |= ALLOC_HARDER;
2974 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2975 if (gfp_mask & __GFP_MEMALLOC)
2976 alloc_flags |= ALLOC_NO_WATERMARKS;
2977 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2978 alloc_flags |= ALLOC_NO_WATERMARKS;
2979 else if (!in_interrupt() &&
2980 ((current->flags & PF_MEMALLOC) ||
2981 unlikely(test_thread_flag(TIF_MEMDIE))))
2982 alloc_flags |= ALLOC_NO_WATERMARKS;
2985 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2986 alloc_flags |= ALLOC_CMA;
2991 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2993 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2996 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
2998 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
3001 static inline struct page *
3002 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3003 struct alloc_context *ac)
3005 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3006 struct page *page = NULL;
3008 unsigned long pages_reclaimed = 0;
3009 unsigned long did_some_progress;
3010 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3011 bool deferred_compaction = false;
3012 int contended_compaction = COMPACT_CONTENDED_NONE;
3015 * In the slowpath, we sanity check order to avoid ever trying to
3016 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3017 * be using allocators in order of preference for an area that is
3020 if (order >= MAX_ORDER) {
3021 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3026 * We also sanity check to catch abuse of atomic reserves being used by
3027 * callers that are not in atomic context.
3029 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3030 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3031 gfp_mask &= ~__GFP_ATOMIC;
3034 * If this allocation cannot block and it is for a specific node, then
3035 * fail early. There's no need to wakeup kswapd or retry for a
3036 * speculative node-specific allocation.
3038 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
3042 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3043 wake_all_kswapds(order, ac);
3046 * OK, we're below the kswapd watermark and have kicked background
3047 * reclaim. Now things get more complex, so set up alloc_flags according
3048 * to how we want to proceed.
3050 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3053 * Find the true preferred zone if the allocation is unconstrained by
3056 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3057 struct zoneref *preferred_zoneref;
3058 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3059 ac->high_zoneidx, NULL, &ac->preferred_zone);
3060 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3063 /* This is the last chance, in general, before the goto nopage. */
3064 page = get_page_from_freelist(gfp_mask, order,
3065 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3069 /* Allocate without watermarks if the context allows */
3070 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3072 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3073 * the allocation is high priority and these type of
3074 * allocations are system rather than user orientated
3076 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3077 page = get_page_from_freelist(gfp_mask, order,
3078 ALLOC_NO_WATERMARKS, ac);
3083 /* Caller is not willing to reclaim, we can't balance anything */
3084 if (!can_direct_reclaim) {
3086 * All existing users of the __GFP_NOFAIL are blockable, so warn
3087 * of any new users that actually allow this type of allocation
3090 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3094 /* Avoid recursion of direct reclaim */
3095 if (current->flags & PF_MEMALLOC) {
3097 * __GFP_NOFAIL request from this context is rather bizarre
3098 * because we cannot reclaim anything and only can loop waiting
3099 * for somebody to do a work for us.
3101 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3108 /* Avoid allocations with no watermarks from looping endlessly */
3109 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3113 * Try direct compaction. The first pass is asynchronous. Subsequent
3114 * attempts after direct reclaim are synchronous
3116 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3118 &contended_compaction,
3119 &deferred_compaction);
3123 /* Checks for THP-specific high-order allocations */
3124 if (is_thp_gfp_mask(gfp_mask)) {
3126 * If compaction is deferred for high-order allocations, it is
3127 * because sync compaction recently failed. If this is the case
3128 * and the caller requested a THP allocation, we do not want
3129 * to heavily disrupt the system, so we fail the allocation
3130 * instead of entering direct reclaim.
3132 if (deferred_compaction)
3136 * In all zones where compaction was attempted (and not
3137 * deferred or skipped), lock contention has been detected.
3138 * For THP allocation we do not want to disrupt the others
3139 * so we fallback to base pages instead.
3141 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3145 * If compaction was aborted due to need_resched(), we do not
3146 * want to further increase allocation latency, unless it is
3147 * khugepaged trying to collapse.
3149 if (contended_compaction == COMPACT_CONTENDED_SCHED
3150 && !(current->flags & PF_KTHREAD))
3155 * It can become very expensive to allocate transparent hugepages at
3156 * fault, so use asynchronous memory compaction for THP unless it is
3157 * khugepaged trying to collapse.
3159 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3160 migration_mode = MIGRATE_SYNC_LIGHT;
3162 /* Try direct reclaim and then allocating */
3163 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3164 &did_some_progress);
3168 /* Do not loop if specifically requested */
3169 if (gfp_mask & __GFP_NORETRY)
3172 /* Keep reclaiming pages as long as there is reasonable progress */
3173 pages_reclaimed += did_some_progress;
3174 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3175 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3176 /* Wait for some write requests to complete then retry */
3177 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3181 /* Reclaim has failed us, start killing things */
3182 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3186 /* Retry as long as the OOM killer is making progress */
3187 if (did_some_progress)
3192 * High-order allocations do not necessarily loop after
3193 * direct reclaim and reclaim/compaction depends on compaction
3194 * being called after reclaim so call directly if necessary
3196 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3198 &contended_compaction,
3199 &deferred_compaction);
3203 warn_alloc_failed(gfp_mask, order, NULL);
3209 * This is the 'heart' of the zoned buddy allocator.
3212 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3213 struct zonelist *zonelist, nodemask_t *nodemask)
3215 struct zoneref *preferred_zoneref;
3216 struct page *page = NULL;
3217 unsigned int cpuset_mems_cookie;
3218 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3219 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3220 struct alloc_context ac = {
3221 .high_zoneidx = gfp_zone(gfp_mask),
3222 .nodemask = nodemask,
3223 .migratetype = gfpflags_to_migratetype(gfp_mask),
3226 gfp_mask &= gfp_allowed_mask;
3228 lockdep_trace_alloc(gfp_mask);
3230 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3232 if (should_fail_alloc_page(gfp_mask, order))
3236 * Check the zones suitable for the gfp_mask contain at least one
3237 * valid zone. It's possible to have an empty zonelist as a result
3238 * of __GFP_THISNODE and a memoryless node
3240 if (unlikely(!zonelist->_zonerefs->zone))
3243 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3244 alloc_flags |= ALLOC_CMA;
3247 cpuset_mems_cookie = read_mems_allowed_begin();
3249 /* We set it here, as __alloc_pages_slowpath might have changed it */
3250 ac.zonelist = zonelist;
3252 /* Dirty zone balancing only done in the fast path */
3253 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3255 /* The preferred zone is used for statistics later */
3256 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3257 ac.nodemask ? : &cpuset_current_mems_allowed,
3258 &ac.preferred_zone);
3259 if (!ac.preferred_zone)
3261 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3263 /* First allocation attempt */
3264 alloc_mask = gfp_mask|__GFP_HARDWALL;
3265 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3266 if (unlikely(!page)) {
3268 * Runtime PM, block IO and its error handling path
3269 * can deadlock because I/O on the device might not
3272 alloc_mask = memalloc_noio_flags(gfp_mask);
3273 ac.spread_dirty_pages = false;
3275 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3278 if (kmemcheck_enabled && page)
3279 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3281 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3285 * When updating a task's mems_allowed, it is possible to race with
3286 * parallel threads in such a way that an allocation can fail while
3287 * the mask is being updated. If a page allocation is about to fail,
3288 * check if the cpuset changed during allocation and if so, retry.
3290 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3295 EXPORT_SYMBOL(__alloc_pages_nodemask);
3298 * Common helper functions.
3300 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3305 * __get_free_pages() returns a 32-bit address, which cannot represent
3308 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3310 page = alloc_pages(gfp_mask, order);
3313 return (unsigned long) page_address(page);
3315 EXPORT_SYMBOL(__get_free_pages);
3317 unsigned long get_zeroed_page(gfp_t gfp_mask)
3319 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3321 EXPORT_SYMBOL(get_zeroed_page);
3323 void __free_pages(struct page *page, unsigned int order)
3325 if (put_page_testzero(page)) {
3327 free_hot_cold_page(page, false);
3329 __free_pages_ok(page, order);
3333 EXPORT_SYMBOL(__free_pages);
3335 void free_pages(unsigned long addr, unsigned int order)
3338 VM_BUG_ON(!virt_addr_valid((void *)addr));
3339 __free_pages(virt_to_page((void *)addr), order);
3343 EXPORT_SYMBOL(free_pages);
3347 * An arbitrary-length arbitrary-offset area of memory which resides
3348 * within a 0 or higher order page. Multiple fragments within that page
3349 * are individually refcounted, in the page's reference counter.
3351 * The page_frag functions below provide a simple allocation framework for
3352 * page fragments. This is used by the network stack and network device
3353 * drivers to provide a backing region of memory for use as either an
3354 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3356 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3359 struct page *page = NULL;
3360 gfp_t gfp = gfp_mask;
3362 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3363 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3365 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3366 PAGE_FRAG_CACHE_MAX_ORDER);
3367 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3369 if (unlikely(!page))
3370 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3372 nc->va = page ? page_address(page) : NULL;
3377 void *__alloc_page_frag(struct page_frag_cache *nc,
3378 unsigned int fragsz, gfp_t gfp_mask)
3380 unsigned int size = PAGE_SIZE;
3384 if (unlikely(!nc->va)) {
3386 page = __page_frag_refill(nc, gfp_mask);
3390 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3391 /* if size can vary use size else just use PAGE_SIZE */
3394 /* Even if we own the page, we do not use atomic_set().
3395 * This would break get_page_unless_zero() users.
3397 atomic_add(size - 1, &page->_count);
3399 /* reset page count bias and offset to start of new frag */
3400 nc->pfmemalloc = page_is_pfmemalloc(page);
3401 nc->pagecnt_bias = size;
3405 offset = nc->offset - fragsz;
3406 if (unlikely(offset < 0)) {
3407 page = virt_to_page(nc->va);
3409 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3412 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3413 /* if size can vary use size else just use PAGE_SIZE */
3416 /* OK, page count is 0, we can safely set it */
3417 atomic_set(&page->_count, size);
3419 /* reset page count bias and offset to start of new frag */
3420 nc->pagecnt_bias = size;
3421 offset = size - fragsz;
3425 nc->offset = offset;
3427 return nc->va + offset;
3429 EXPORT_SYMBOL(__alloc_page_frag);
3432 * Frees a page fragment allocated out of either a compound or order 0 page.
3434 void __free_page_frag(void *addr)
3436 struct page *page = virt_to_head_page(addr);
3438 if (unlikely(put_page_testzero(page)))
3439 __free_pages_ok(page, compound_order(page));
3441 EXPORT_SYMBOL(__free_page_frag);
3444 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3445 * of the current memory cgroup if __GFP_ACCOUNT is set, other than that it is
3446 * equivalent to alloc_pages.
3448 * It should be used when the caller would like to use kmalloc, but since the
3449 * allocation is large, it has to fall back to the page allocator.
3451 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3455 page = alloc_pages(gfp_mask, order);
3456 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3457 __free_pages(page, order);
3463 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3467 page = alloc_pages_node(nid, gfp_mask, order);
3468 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3469 __free_pages(page, order);
3476 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3479 void __free_kmem_pages(struct page *page, unsigned int order)
3481 memcg_kmem_uncharge(page, order);
3482 __free_pages(page, order);
3485 void free_kmem_pages(unsigned long addr, unsigned int order)
3488 VM_BUG_ON(!virt_addr_valid((void *)addr));
3489 __free_kmem_pages(virt_to_page((void *)addr), order);
3493 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3497 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3498 unsigned long used = addr + PAGE_ALIGN(size);
3500 split_page(virt_to_page((void *)addr), order);
3501 while (used < alloc_end) {
3506 return (void *)addr;
3510 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3511 * @size: the number of bytes to allocate
3512 * @gfp_mask: GFP flags for the allocation
3514 * This function is similar to alloc_pages(), except that it allocates the
3515 * minimum number of pages to satisfy the request. alloc_pages() can only
3516 * allocate memory in power-of-two pages.
3518 * This function is also limited by MAX_ORDER.
3520 * Memory allocated by this function must be released by free_pages_exact().
3522 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3524 unsigned int order = get_order(size);
3527 addr = __get_free_pages(gfp_mask, order);
3528 return make_alloc_exact(addr, order, size);
3530 EXPORT_SYMBOL(alloc_pages_exact);
3533 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3535 * @nid: the preferred node ID where memory should be allocated
3536 * @size: the number of bytes to allocate
3537 * @gfp_mask: GFP flags for the allocation
3539 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3542 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3544 unsigned int order = get_order(size);
3545 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3548 return make_alloc_exact((unsigned long)page_address(p), order, size);
3552 * free_pages_exact - release memory allocated via alloc_pages_exact()
3553 * @virt: the value returned by alloc_pages_exact.
3554 * @size: size of allocation, same value as passed to alloc_pages_exact().
3556 * Release the memory allocated by a previous call to alloc_pages_exact.
3558 void free_pages_exact(void *virt, size_t size)
3560 unsigned long addr = (unsigned long)virt;
3561 unsigned long end = addr + PAGE_ALIGN(size);
3563 while (addr < end) {
3568 EXPORT_SYMBOL(free_pages_exact);
3571 * nr_free_zone_pages - count number of pages beyond high watermark
3572 * @offset: The zone index of the highest zone
3574 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3575 * high watermark within all zones at or below a given zone index. For each
3576 * zone, the number of pages is calculated as:
3577 * managed_pages - high_pages
3579 static unsigned long nr_free_zone_pages(int offset)
3584 /* Just pick one node, since fallback list is circular */
3585 unsigned long sum = 0;
3587 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3589 for_each_zone_zonelist(zone, z, zonelist, offset) {
3590 unsigned long size = zone->managed_pages;
3591 unsigned long high = high_wmark_pages(zone);
3600 * nr_free_buffer_pages - count number of pages beyond high watermark
3602 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3603 * watermark within ZONE_DMA and ZONE_NORMAL.
3605 unsigned long nr_free_buffer_pages(void)
3607 return nr_free_zone_pages(gfp_zone(GFP_USER));
3609 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3612 * nr_free_pagecache_pages - count number of pages beyond high watermark
3614 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3615 * high watermark within all zones.
3617 unsigned long nr_free_pagecache_pages(void)
3619 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3622 static inline void show_node(struct zone *zone)
3624 if (IS_ENABLED(CONFIG_NUMA))
3625 printk("Node %d ", zone_to_nid(zone));
3628 void si_meminfo(struct sysinfo *val)
3630 val->totalram = totalram_pages;
3631 val->sharedram = global_page_state(NR_SHMEM);
3632 val->freeram = global_page_state(NR_FREE_PAGES);
3633 val->bufferram = nr_blockdev_pages();
3634 val->totalhigh = totalhigh_pages;
3635 val->freehigh = nr_free_highpages();
3636 val->mem_unit = PAGE_SIZE;
3639 EXPORT_SYMBOL(si_meminfo);
3642 void si_meminfo_node(struct sysinfo *val, int nid)
3644 int zone_type; /* needs to be signed */
3645 unsigned long managed_pages = 0;
3646 pg_data_t *pgdat = NODE_DATA(nid);
3648 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3649 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3650 val->totalram = managed_pages;
3651 val->sharedram = node_page_state(nid, NR_SHMEM);
3652 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3653 #ifdef CONFIG_HIGHMEM
3654 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3655 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3661 val->mem_unit = PAGE_SIZE;
3666 * Determine whether the node should be displayed or not, depending on whether
3667 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3669 bool skip_free_areas_node(unsigned int flags, int nid)
3672 unsigned int cpuset_mems_cookie;
3674 if (!(flags & SHOW_MEM_FILTER_NODES))
3678 cpuset_mems_cookie = read_mems_allowed_begin();
3679 ret = !node_isset(nid, cpuset_current_mems_allowed);
3680 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3685 #define K(x) ((x) << (PAGE_SHIFT-10))
3687 static void show_migration_types(unsigned char type)
3689 static const char types[MIGRATE_TYPES] = {
3690 [MIGRATE_UNMOVABLE] = 'U',
3691 [MIGRATE_MOVABLE] = 'M',
3692 [MIGRATE_RECLAIMABLE] = 'E',
3693 [MIGRATE_HIGHATOMIC] = 'H',
3695 [MIGRATE_CMA] = 'C',
3697 #ifdef CONFIG_MEMORY_ISOLATION
3698 [MIGRATE_ISOLATE] = 'I',
3701 char tmp[MIGRATE_TYPES + 1];
3705 for (i = 0; i < MIGRATE_TYPES; i++) {
3706 if (type & (1 << i))
3711 printk("(%s) ", tmp);
3715 * Show free area list (used inside shift_scroll-lock stuff)
3716 * We also calculate the percentage fragmentation. We do this by counting the
3717 * memory on each free list with the exception of the first item on the list.
3720 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3723 void show_free_areas(unsigned int filter)
3725 unsigned long free_pcp = 0;
3729 for_each_populated_zone(zone) {
3730 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3733 for_each_online_cpu(cpu)
3734 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3737 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3738 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3739 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3740 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3741 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3742 " free:%lu free_pcp:%lu free_cma:%lu\n",
3743 global_page_state(NR_ACTIVE_ANON),
3744 global_page_state(NR_INACTIVE_ANON),
3745 global_page_state(NR_ISOLATED_ANON),
3746 global_page_state(NR_ACTIVE_FILE),
3747 global_page_state(NR_INACTIVE_FILE),
3748 global_page_state(NR_ISOLATED_FILE),
3749 global_page_state(NR_UNEVICTABLE),
3750 global_page_state(NR_FILE_DIRTY),
3751 global_page_state(NR_WRITEBACK),
3752 global_page_state(NR_UNSTABLE_NFS),
3753 global_page_state(NR_SLAB_RECLAIMABLE),
3754 global_page_state(NR_SLAB_UNRECLAIMABLE),
3755 global_page_state(NR_FILE_MAPPED),
3756 global_page_state(NR_SHMEM),
3757 global_page_state(NR_PAGETABLE),
3758 global_page_state(NR_BOUNCE),
3759 global_page_state(NR_FREE_PAGES),
3761 global_page_state(NR_FREE_CMA_PAGES));
3763 for_each_populated_zone(zone) {
3766 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3770 for_each_online_cpu(cpu)
3771 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3779 " active_anon:%lukB"
3780 " inactive_anon:%lukB"
3781 " active_file:%lukB"
3782 " inactive_file:%lukB"
3783 " unevictable:%lukB"
3784 " isolated(anon):%lukB"
3785 " isolated(file):%lukB"
3793 " slab_reclaimable:%lukB"
3794 " slab_unreclaimable:%lukB"
3795 " kernel_stack:%lukB"
3802 " writeback_tmp:%lukB"
3803 " pages_scanned:%lu"
3804 " all_unreclaimable? %s"
3807 K(zone_page_state(zone, NR_FREE_PAGES)),
3808 K(min_wmark_pages(zone)),
3809 K(low_wmark_pages(zone)),
3810 K(high_wmark_pages(zone)),
3811 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3812 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3813 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3814 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3815 K(zone_page_state(zone, NR_UNEVICTABLE)),
3816 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3817 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3818 K(zone->present_pages),
3819 K(zone->managed_pages),
3820 K(zone_page_state(zone, NR_MLOCK)),
3821 K(zone_page_state(zone, NR_FILE_DIRTY)),
3822 K(zone_page_state(zone, NR_WRITEBACK)),
3823 K(zone_page_state(zone, NR_FILE_MAPPED)),
3824 K(zone_page_state(zone, NR_SHMEM)),
3825 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3826 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3827 zone_page_state(zone, NR_KERNEL_STACK) *
3829 K(zone_page_state(zone, NR_PAGETABLE)),
3830 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3831 K(zone_page_state(zone, NR_BOUNCE)),
3833 K(this_cpu_read(zone->pageset->pcp.count)),
3834 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3835 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3836 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3837 (!zone_reclaimable(zone) ? "yes" : "no")
3839 printk("lowmem_reserve[]:");
3840 for (i = 0; i < MAX_NR_ZONES; i++)
3841 printk(" %ld", zone->lowmem_reserve[i]);
3845 for_each_populated_zone(zone) {
3847 unsigned long nr[MAX_ORDER], flags, total = 0;
3848 unsigned char types[MAX_ORDER];
3850 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3853 printk("%s: ", zone->name);
3855 spin_lock_irqsave(&zone->lock, flags);
3856 for (order = 0; order < MAX_ORDER; order++) {
3857 struct free_area *area = &zone->free_area[order];
3860 nr[order] = area->nr_free;
3861 total += nr[order] << order;
3864 for (type = 0; type < MIGRATE_TYPES; type++) {
3865 if (!list_empty(&area->free_list[type]))
3866 types[order] |= 1 << type;
3869 spin_unlock_irqrestore(&zone->lock, flags);
3870 for (order = 0; order < MAX_ORDER; order++) {
3871 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3873 show_migration_types(types[order]);
3875 printk("= %lukB\n", K(total));
3878 hugetlb_show_meminfo();
3880 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3882 show_swap_cache_info();
3885 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3887 zoneref->zone = zone;
3888 zoneref->zone_idx = zone_idx(zone);
3892 * Builds allocation fallback zone lists.
3894 * Add all populated zones of a node to the zonelist.
3896 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3900 enum zone_type zone_type = MAX_NR_ZONES;
3904 zone = pgdat->node_zones + zone_type;
3905 if (populated_zone(zone)) {
3906 zoneref_set_zone(zone,
3907 &zonelist->_zonerefs[nr_zones++]);
3908 check_highest_zone(zone_type);
3910 } while (zone_type);
3918 * 0 = automatic detection of better ordering.
3919 * 1 = order by ([node] distance, -zonetype)
3920 * 2 = order by (-zonetype, [node] distance)
3922 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3923 * the same zonelist. So only NUMA can configure this param.
3925 #define ZONELIST_ORDER_DEFAULT 0
3926 #define ZONELIST_ORDER_NODE 1
3927 #define ZONELIST_ORDER_ZONE 2
3929 /* zonelist order in the kernel.
3930 * set_zonelist_order() will set this to NODE or ZONE.
3932 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3933 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3937 /* The value user specified ....changed by config */
3938 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3939 /* string for sysctl */
3940 #define NUMA_ZONELIST_ORDER_LEN 16
3941 char numa_zonelist_order[16] = "default";
3944 * interface for configure zonelist ordering.
3945 * command line option "numa_zonelist_order"
3946 * = "[dD]efault - default, automatic configuration.
3947 * = "[nN]ode - order by node locality, then by zone within node
3948 * = "[zZ]one - order by zone, then by locality within zone
3951 static int __parse_numa_zonelist_order(char *s)
3953 if (*s == 'd' || *s == 'D') {
3954 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3955 } else if (*s == 'n' || *s == 'N') {
3956 user_zonelist_order = ZONELIST_ORDER_NODE;
3957 } else if (*s == 'z' || *s == 'Z') {
3958 user_zonelist_order = ZONELIST_ORDER_ZONE;
3961 "Ignoring invalid numa_zonelist_order value: "
3968 static __init int setup_numa_zonelist_order(char *s)
3975 ret = __parse_numa_zonelist_order(s);
3977 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3981 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3984 * sysctl handler for numa_zonelist_order
3986 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3987 void __user *buffer, size_t *length,
3990 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3992 static DEFINE_MUTEX(zl_order_mutex);
3994 mutex_lock(&zl_order_mutex);
3996 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4000 strcpy(saved_string, (char *)table->data);
4002 ret = proc_dostring(table, write, buffer, length, ppos);
4006 int oldval = user_zonelist_order;
4008 ret = __parse_numa_zonelist_order((char *)table->data);
4011 * bogus value. restore saved string
4013 strncpy((char *)table->data, saved_string,
4014 NUMA_ZONELIST_ORDER_LEN);
4015 user_zonelist_order = oldval;
4016 } else if (oldval != user_zonelist_order) {
4017 mutex_lock(&zonelists_mutex);
4018 build_all_zonelists(NULL, NULL);
4019 mutex_unlock(&zonelists_mutex);
4023 mutex_unlock(&zl_order_mutex);
4028 #define MAX_NODE_LOAD (nr_online_nodes)
4029 static int node_load[MAX_NUMNODES];
4032 * find_next_best_node - find the next node that should appear in a given node's fallback list
4033 * @node: node whose fallback list we're appending
4034 * @used_node_mask: nodemask_t of already used nodes
4036 * We use a number of factors to determine which is the next node that should
4037 * appear on a given node's fallback list. The node should not have appeared
4038 * already in @node's fallback list, and it should be the next closest node
4039 * according to the distance array (which contains arbitrary distance values
4040 * from each node to each node in the system), and should also prefer nodes
4041 * with no CPUs, since presumably they'll have very little allocation pressure
4042 * on them otherwise.
4043 * It returns -1 if no node is found.
4045 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4048 int min_val = INT_MAX;
4049 int best_node = NUMA_NO_NODE;
4050 const struct cpumask *tmp = cpumask_of_node(0);
4052 /* Use the local node if we haven't already */
4053 if (!node_isset(node, *used_node_mask)) {
4054 node_set(node, *used_node_mask);
4058 for_each_node_state(n, N_MEMORY) {
4060 /* Don't want a node to appear more than once */
4061 if (node_isset(n, *used_node_mask))
4064 /* Use the distance array to find the distance */
4065 val = node_distance(node, n);
4067 /* Penalize nodes under us ("prefer the next node") */
4070 /* Give preference to headless and unused nodes */
4071 tmp = cpumask_of_node(n);
4072 if (!cpumask_empty(tmp))
4073 val += PENALTY_FOR_NODE_WITH_CPUS;
4075 /* Slight preference for less loaded node */
4076 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4077 val += node_load[n];
4079 if (val < min_val) {
4086 node_set(best_node, *used_node_mask);
4093 * Build zonelists ordered by node and zones within node.
4094 * This results in maximum locality--normal zone overflows into local
4095 * DMA zone, if any--but risks exhausting DMA zone.
4097 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4100 struct zonelist *zonelist;
4102 zonelist = &pgdat->node_zonelists[0];
4103 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4105 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4106 zonelist->_zonerefs[j].zone = NULL;
4107 zonelist->_zonerefs[j].zone_idx = 0;
4111 * Build gfp_thisnode zonelists
4113 static void build_thisnode_zonelists(pg_data_t *pgdat)
4116 struct zonelist *zonelist;
4118 zonelist = &pgdat->node_zonelists[1];
4119 j = build_zonelists_node(pgdat, zonelist, 0);
4120 zonelist->_zonerefs[j].zone = NULL;
4121 zonelist->_zonerefs[j].zone_idx = 0;
4125 * Build zonelists ordered by zone and nodes within zones.
4126 * This results in conserving DMA zone[s] until all Normal memory is
4127 * exhausted, but results in overflowing to remote node while memory
4128 * may still exist in local DMA zone.
4130 static int node_order[MAX_NUMNODES];
4132 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4135 int zone_type; /* needs to be signed */
4137 struct zonelist *zonelist;
4139 zonelist = &pgdat->node_zonelists[0];
4141 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4142 for (j = 0; j < nr_nodes; j++) {
4143 node = node_order[j];
4144 z = &NODE_DATA(node)->node_zones[zone_type];
4145 if (populated_zone(z)) {
4147 &zonelist->_zonerefs[pos++]);
4148 check_highest_zone(zone_type);
4152 zonelist->_zonerefs[pos].zone = NULL;
4153 zonelist->_zonerefs[pos].zone_idx = 0;
4156 #if defined(CONFIG_64BIT)
4158 * Devices that require DMA32/DMA are relatively rare and do not justify a
4159 * penalty to every machine in case the specialised case applies. Default
4160 * to Node-ordering on 64-bit NUMA machines
4162 static int default_zonelist_order(void)
4164 return ZONELIST_ORDER_NODE;
4168 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4169 * by the kernel. If processes running on node 0 deplete the low memory zone
4170 * then reclaim will occur more frequency increasing stalls and potentially
4171 * be easier to OOM if a large percentage of the zone is under writeback or
4172 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4173 * Hence, default to zone ordering on 32-bit.
4175 static int default_zonelist_order(void)
4177 return ZONELIST_ORDER_ZONE;
4179 #endif /* CONFIG_64BIT */
4181 static void set_zonelist_order(void)
4183 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4184 current_zonelist_order = default_zonelist_order();
4186 current_zonelist_order = user_zonelist_order;
4189 static void build_zonelists(pg_data_t *pgdat)
4192 nodemask_t used_mask;
4193 int local_node, prev_node;
4194 struct zonelist *zonelist;
4195 unsigned int order = current_zonelist_order;
4197 /* initialize zonelists */
4198 for (i = 0; i < MAX_ZONELISTS; i++) {
4199 zonelist = pgdat->node_zonelists + i;
4200 zonelist->_zonerefs[0].zone = NULL;
4201 zonelist->_zonerefs[0].zone_idx = 0;
4204 /* NUMA-aware ordering of nodes */
4205 local_node = pgdat->node_id;
4206 load = nr_online_nodes;
4207 prev_node = local_node;
4208 nodes_clear(used_mask);
4210 memset(node_order, 0, sizeof(node_order));
4213 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4215 * We don't want to pressure a particular node.
4216 * So adding penalty to the first node in same
4217 * distance group to make it round-robin.
4219 if (node_distance(local_node, node) !=
4220 node_distance(local_node, prev_node))
4221 node_load[node] = load;
4225 if (order == ZONELIST_ORDER_NODE)
4226 build_zonelists_in_node_order(pgdat, node);
4228 node_order[i++] = node; /* remember order */
4231 if (order == ZONELIST_ORDER_ZONE) {
4232 /* calculate node order -- i.e., DMA last! */
4233 build_zonelists_in_zone_order(pgdat, i);
4236 build_thisnode_zonelists(pgdat);
4239 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4241 * Return node id of node used for "local" allocations.
4242 * I.e., first node id of first zone in arg node's generic zonelist.
4243 * Used for initializing percpu 'numa_mem', which is used primarily
4244 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4246 int local_memory_node(int node)
4250 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4251 gfp_zone(GFP_KERNEL),
4258 #else /* CONFIG_NUMA */
4260 static void set_zonelist_order(void)
4262 current_zonelist_order = ZONELIST_ORDER_ZONE;
4265 static void build_zonelists(pg_data_t *pgdat)
4267 int node, local_node;
4269 struct zonelist *zonelist;
4271 local_node = pgdat->node_id;
4273 zonelist = &pgdat->node_zonelists[0];
4274 j = build_zonelists_node(pgdat, zonelist, 0);
4277 * Now we build the zonelist so that it contains the zones
4278 * of all the other nodes.
4279 * We don't want to pressure a particular node, so when
4280 * building the zones for node N, we make sure that the
4281 * zones coming right after the local ones are those from
4282 * node N+1 (modulo N)
4284 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4285 if (!node_online(node))
4287 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4289 for (node = 0; node < local_node; node++) {
4290 if (!node_online(node))
4292 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4295 zonelist->_zonerefs[j].zone = NULL;
4296 zonelist->_zonerefs[j].zone_idx = 0;
4299 #endif /* CONFIG_NUMA */
4302 * Boot pageset table. One per cpu which is going to be used for all
4303 * zones and all nodes. The parameters will be set in such a way
4304 * that an item put on a list will immediately be handed over to
4305 * the buddy list. This is safe since pageset manipulation is done
4306 * with interrupts disabled.
4308 * The boot_pagesets must be kept even after bootup is complete for
4309 * unused processors and/or zones. They do play a role for bootstrapping
4310 * hotplugged processors.
4312 * zoneinfo_show() and maybe other functions do
4313 * not check if the processor is online before following the pageset pointer.
4314 * Other parts of the kernel may not check if the zone is available.
4316 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4317 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4318 static void setup_zone_pageset(struct zone *zone);
4321 * Global mutex to protect against size modification of zonelists
4322 * as well as to serialize pageset setup for the new populated zone.
4324 DEFINE_MUTEX(zonelists_mutex);
4326 /* return values int ....just for stop_machine() */
4327 static int __build_all_zonelists(void *data)
4331 pg_data_t *self = data;
4334 memset(node_load, 0, sizeof(node_load));
4337 if (self && !node_online(self->node_id)) {
4338 build_zonelists(self);
4341 for_each_online_node(nid) {
4342 pg_data_t *pgdat = NODE_DATA(nid);
4344 build_zonelists(pgdat);
4348 * Initialize the boot_pagesets that are going to be used
4349 * for bootstrapping processors. The real pagesets for
4350 * each zone will be allocated later when the per cpu
4351 * allocator is available.
4353 * boot_pagesets are used also for bootstrapping offline
4354 * cpus if the system is already booted because the pagesets
4355 * are needed to initialize allocators on a specific cpu too.
4356 * F.e. the percpu allocator needs the page allocator which
4357 * needs the percpu allocator in order to allocate its pagesets
4358 * (a chicken-egg dilemma).
4360 for_each_possible_cpu(cpu) {
4361 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4363 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4365 * We now know the "local memory node" for each node--
4366 * i.e., the node of the first zone in the generic zonelist.
4367 * Set up numa_mem percpu variable for on-line cpus. During
4368 * boot, only the boot cpu should be on-line; we'll init the
4369 * secondary cpus' numa_mem as they come on-line. During
4370 * node/memory hotplug, we'll fixup all on-line cpus.
4372 if (cpu_online(cpu))
4373 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4380 static noinline void __init
4381 build_all_zonelists_init(void)
4383 __build_all_zonelists(NULL);
4384 mminit_verify_zonelist();
4385 cpuset_init_current_mems_allowed();
4389 * Called with zonelists_mutex held always
4390 * unless system_state == SYSTEM_BOOTING.
4392 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4393 * [we're only called with non-NULL zone through __meminit paths] and
4394 * (2) call of __init annotated helper build_all_zonelists_init
4395 * [protected by SYSTEM_BOOTING].
4397 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4399 set_zonelist_order();
4401 if (system_state == SYSTEM_BOOTING) {
4402 build_all_zonelists_init();
4404 #ifdef CONFIG_MEMORY_HOTPLUG
4406 setup_zone_pageset(zone);
4408 /* we have to stop all cpus to guarantee there is no user
4410 stop_machine(__build_all_zonelists, pgdat, NULL);
4411 /* cpuset refresh routine should be here */
4413 vm_total_pages = nr_free_pagecache_pages();
4415 * Disable grouping by mobility if the number of pages in the
4416 * system is too low to allow the mechanism to work. It would be
4417 * more accurate, but expensive to check per-zone. This check is
4418 * made on memory-hotadd so a system can start with mobility
4419 * disabled and enable it later
4421 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4422 page_group_by_mobility_disabled = 1;
4424 page_group_by_mobility_disabled = 0;
4426 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4427 "Total pages: %ld\n",
4429 zonelist_order_name[current_zonelist_order],
4430 page_group_by_mobility_disabled ? "off" : "on",
4433 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4438 * Helper functions to size the waitqueue hash table.
4439 * Essentially these want to choose hash table sizes sufficiently
4440 * large so that collisions trying to wait on pages are rare.
4441 * But in fact, the number of active page waitqueues on typical
4442 * systems is ridiculously low, less than 200. So this is even
4443 * conservative, even though it seems large.
4445 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4446 * waitqueues, i.e. the size of the waitq table given the number of pages.
4448 #define PAGES_PER_WAITQUEUE 256
4450 #ifndef CONFIG_MEMORY_HOTPLUG
4451 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4453 unsigned long size = 1;
4455 pages /= PAGES_PER_WAITQUEUE;
4457 while (size < pages)
4461 * Once we have dozens or even hundreds of threads sleeping
4462 * on IO we've got bigger problems than wait queue collision.
4463 * Limit the size of the wait table to a reasonable size.
4465 size = min(size, 4096UL);
4467 return max(size, 4UL);
4471 * A zone's size might be changed by hot-add, so it is not possible to determine
4472 * a suitable size for its wait_table. So we use the maximum size now.
4474 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4476 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4477 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4478 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4480 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4481 * or more by the traditional way. (See above). It equals:
4483 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4484 * ia64(16K page size) : = ( 8G + 4M)byte.
4485 * powerpc (64K page size) : = (32G +16M)byte.
4487 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4494 * This is an integer logarithm so that shifts can be used later
4495 * to extract the more random high bits from the multiplicative
4496 * hash function before the remainder is taken.
4498 static inline unsigned long wait_table_bits(unsigned long size)
4504 * Initially all pages are reserved - free ones are freed
4505 * up by free_all_bootmem() once the early boot process is
4506 * done. Non-atomic initialization, single-pass.
4508 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4509 unsigned long start_pfn, enum memmap_context context)
4511 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
4512 unsigned long end_pfn = start_pfn + size;
4513 pg_data_t *pgdat = NODE_DATA(nid);
4515 unsigned long nr_initialised = 0;
4516 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4517 struct memblock_region *r = NULL, *tmp;
4520 if (highest_memmap_pfn < end_pfn - 1)
4521 highest_memmap_pfn = end_pfn - 1;
4524 * Honor reservation requested by the driver for this ZONE_DEVICE
4527 if (altmap && start_pfn == altmap->base_pfn)
4528 start_pfn += altmap->reserve;
4530 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4532 * There can be holes in boot-time mem_map[]s handed to this
4533 * function. They do not exist on hotplugged memory.
4535 if (context != MEMMAP_EARLY)
4538 if (!early_pfn_valid(pfn))
4540 if (!early_pfn_in_nid(pfn, nid))
4542 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
4545 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4547 * If not mirrored_kernelcore and ZONE_MOVABLE exists, range
4548 * from zone_movable_pfn[nid] to end of each node should be
4549 * ZONE_MOVABLE not ZONE_NORMAL. skip it.
4551 if (!mirrored_kernelcore && zone_movable_pfn[nid])
4552 if (zone == ZONE_NORMAL && pfn >= zone_movable_pfn[nid])
4556 * Check given memblock attribute by firmware which can affect
4557 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
4558 * mirrored, it's an overlapped memmap init. skip it.
4560 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
4561 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
4562 for_each_memblock(memory, tmp)
4563 if (pfn < memblock_region_memory_end_pfn(tmp))
4567 if (pfn >= memblock_region_memory_base_pfn(r) &&
4568 memblock_is_mirror(r)) {
4569 /* already initialized as NORMAL */
4570 pfn = memblock_region_memory_end_pfn(r);
4578 * Mark the block movable so that blocks are reserved for
4579 * movable at startup. This will force kernel allocations
4580 * to reserve their blocks rather than leaking throughout
4581 * the address space during boot when many long-lived
4582 * kernel allocations are made.
4584 * bitmap is created for zone's valid pfn range. but memmap
4585 * can be created for invalid pages (for alignment)
4586 * check here not to call set_pageblock_migratetype() against
4589 if (!(pfn & (pageblock_nr_pages - 1))) {
4590 struct page *page = pfn_to_page(pfn);
4592 __init_single_page(page, pfn, zone, nid);
4593 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4595 __init_single_pfn(pfn, zone, nid);
4600 static void __meminit zone_init_free_lists(struct zone *zone)
4602 unsigned int order, t;
4603 for_each_migratetype_order(order, t) {
4604 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4605 zone->free_area[order].nr_free = 0;
4609 #ifndef __HAVE_ARCH_MEMMAP_INIT
4610 #define memmap_init(size, nid, zone, start_pfn) \
4611 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4614 static int zone_batchsize(struct zone *zone)
4620 * The per-cpu-pages pools are set to around 1000th of the
4621 * size of the zone. But no more than 1/2 of a meg.
4623 * OK, so we don't know how big the cache is. So guess.
4625 batch = zone->managed_pages / 1024;
4626 if (batch * PAGE_SIZE > 512 * 1024)
4627 batch = (512 * 1024) / PAGE_SIZE;
4628 batch /= 4; /* We effectively *= 4 below */
4633 * Clamp the batch to a 2^n - 1 value. Having a power
4634 * of 2 value was found to be more likely to have
4635 * suboptimal cache aliasing properties in some cases.
4637 * For example if 2 tasks are alternately allocating
4638 * batches of pages, one task can end up with a lot
4639 * of pages of one half of the possible page colors
4640 * and the other with pages of the other colors.
4642 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4647 /* The deferral and batching of frees should be suppressed under NOMMU
4650 * The problem is that NOMMU needs to be able to allocate large chunks
4651 * of contiguous memory as there's no hardware page translation to
4652 * assemble apparent contiguous memory from discontiguous pages.
4654 * Queueing large contiguous runs of pages for batching, however,
4655 * causes the pages to actually be freed in smaller chunks. As there
4656 * can be a significant delay between the individual batches being
4657 * recycled, this leads to the once large chunks of space being
4658 * fragmented and becoming unavailable for high-order allocations.
4665 * pcp->high and pcp->batch values are related and dependent on one another:
4666 * ->batch must never be higher then ->high.
4667 * The following function updates them in a safe manner without read side
4670 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4671 * those fields changing asynchronously (acording the the above rule).
4673 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4674 * outside of boot time (or some other assurance that no concurrent updaters
4677 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4678 unsigned long batch)
4680 /* start with a fail safe value for batch */
4684 /* Update high, then batch, in order */
4691 /* a companion to pageset_set_high() */
4692 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4694 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4697 static void pageset_init(struct per_cpu_pageset *p)
4699 struct per_cpu_pages *pcp;
4702 memset(p, 0, sizeof(*p));
4706 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4707 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4710 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4713 pageset_set_batch(p, batch);
4717 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4718 * to the value high for the pageset p.
4720 static void pageset_set_high(struct per_cpu_pageset *p,
4723 unsigned long batch = max(1UL, high / 4);
4724 if ((high / 4) > (PAGE_SHIFT * 8))
4725 batch = PAGE_SHIFT * 8;
4727 pageset_update(&p->pcp, high, batch);
4730 static void pageset_set_high_and_batch(struct zone *zone,
4731 struct per_cpu_pageset *pcp)
4733 if (percpu_pagelist_fraction)
4734 pageset_set_high(pcp,
4735 (zone->managed_pages /
4736 percpu_pagelist_fraction));
4738 pageset_set_batch(pcp, zone_batchsize(zone));
4741 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4743 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4746 pageset_set_high_and_batch(zone, pcp);
4749 static void __meminit setup_zone_pageset(struct zone *zone)
4752 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4753 for_each_possible_cpu(cpu)
4754 zone_pageset_init(zone, cpu);
4758 * Allocate per cpu pagesets and initialize them.
4759 * Before this call only boot pagesets were available.
4761 void __init setup_per_cpu_pageset(void)
4765 for_each_populated_zone(zone)
4766 setup_zone_pageset(zone);
4769 static noinline __init_refok
4770 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4776 * The per-page waitqueue mechanism uses hashed waitqueues
4779 zone->wait_table_hash_nr_entries =
4780 wait_table_hash_nr_entries(zone_size_pages);
4781 zone->wait_table_bits =
4782 wait_table_bits(zone->wait_table_hash_nr_entries);
4783 alloc_size = zone->wait_table_hash_nr_entries
4784 * sizeof(wait_queue_head_t);
4786 if (!slab_is_available()) {
4787 zone->wait_table = (wait_queue_head_t *)
4788 memblock_virt_alloc_node_nopanic(
4789 alloc_size, zone->zone_pgdat->node_id);
4792 * This case means that a zone whose size was 0 gets new memory
4793 * via memory hot-add.
4794 * But it may be the case that a new node was hot-added. In
4795 * this case vmalloc() will not be able to use this new node's
4796 * memory - this wait_table must be initialized to use this new
4797 * node itself as well.
4798 * To use this new node's memory, further consideration will be
4801 zone->wait_table = vmalloc(alloc_size);
4803 if (!zone->wait_table)
4806 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4807 init_waitqueue_head(zone->wait_table + i);
4812 static __meminit void zone_pcp_init(struct zone *zone)
4815 * per cpu subsystem is not up at this point. The following code
4816 * relies on the ability of the linker to provide the
4817 * offset of a (static) per cpu variable into the per cpu area.
4819 zone->pageset = &boot_pageset;
4821 if (populated_zone(zone))
4822 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4823 zone->name, zone->present_pages,
4824 zone_batchsize(zone));
4827 int __meminit init_currently_empty_zone(struct zone *zone,
4828 unsigned long zone_start_pfn,
4831 struct pglist_data *pgdat = zone->zone_pgdat;
4833 ret = zone_wait_table_init(zone, size);
4836 pgdat->nr_zones = zone_idx(zone) + 1;
4838 zone->zone_start_pfn = zone_start_pfn;
4840 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4841 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4843 (unsigned long)zone_idx(zone),
4844 zone_start_pfn, (zone_start_pfn + size));
4846 zone_init_free_lists(zone);
4851 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4852 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4855 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4857 int __meminit __early_pfn_to_nid(unsigned long pfn,
4858 struct mminit_pfnnid_cache *state)
4860 unsigned long start_pfn, end_pfn;
4863 if (state->last_start <= pfn && pfn < state->last_end)
4864 return state->last_nid;
4866 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4868 state->last_start = start_pfn;
4869 state->last_end = end_pfn;
4870 state->last_nid = nid;
4875 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4878 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4879 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4880 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4882 * If an architecture guarantees that all ranges registered contain no holes
4883 * and may be freed, this this function may be used instead of calling
4884 * memblock_free_early_nid() manually.
4886 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4888 unsigned long start_pfn, end_pfn;
4891 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4892 start_pfn = min(start_pfn, max_low_pfn);
4893 end_pfn = min(end_pfn, max_low_pfn);
4895 if (start_pfn < end_pfn)
4896 memblock_free_early_nid(PFN_PHYS(start_pfn),
4897 (end_pfn - start_pfn) << PAGE_SHIFT,
4903 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4904 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4906 * If an architecture guarantees that all ranges registered contain no holes and may
4907 * be freed, this function may be used instead of calling memory_present() manually.
4909 void __init sparse_memory_present_with_active_regions(int nid)
4911 unsigned long start_pfn, end_pfn;
4914 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4915 memory_present(this_nid, start_pfn, end_pfn);
4919 * get_pfn_range_for_nid - Return the start and end page frames for a node
4920 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4921 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4922 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4924 * It returns the start and end page frame of a node based on information
4925 * provided by memblock_set_node(). If called for a node
4926 * with no available memory, a warning is printed and the start and end
4929 void __meminit get_pfn_range_for_nid(unsigned int nid,
4930 unsigned long *start_pfn, unsigned long *end_pfn)
4932 unsigned long this_start_pfn, this_end_pfn;
4938 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4939 *start_pfn = min(*start_pfn, this_start_pfn);
4940 *end_pfn = max(*end_pfn, this_end_pfn);
4943 if (*start_pfn == -1UL)
4948 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4949 * assumption is made that zones within a node are ordered in monotonic
4950 * increasing memory addresses so that the "highest" populated zone is used
4952 static void __init find_usable_zone_for_movable(void)
4955 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4956 if (zone_index == ZONE_MOVABLE)
4959 if (arch_zone_highest_possible_pfn[zone_index] >
4960 arch_zone_lowest_possible_pfn[zone_index])
4964 VM_BUG_ON(zone_index == -1);
4965 movable_zone = zone_index;
4969 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4970 * because it is sized independent of architecture. Unlike the other zones,
4971 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4972 * in each node depending on the size of each node and how evenly kernelcore
4973 * is distributed. This helper function adjusts the zone ranges
4974 * provided by the architecture for a given node by using the end of the
4975 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4976 * zones within a node are in order of monotonic increases memory addresses
4978 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4979 unsigned long zone_type,
4980 unsigned long node_start_pfn,
4981 unsigned long node_end_pfn,
4982 unsigned long *zone_start_pfn,
4983 unsigned long *zone_end_pfn)
4985 /* Only adjust if ZONE_MOVABLE is on this node */
4986 if (zone_movable_pfn[nid]) {
4987 /* Size ZONE_MOVABLE */
4988 if (zone_type == ZONE_MOVABLE) {
4989 *zone_start_pfn = zone_movable_pfn[nid];
4990 *zone_end_pfn = min(node_end_pfn,
4991 arch_zone_highest_possible_pfn[movable_zone]);
4993 /* Check if this whole range is within ZONE_MOVABLE */
4994 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4995 *zone_start_pfn = *zone_end_pfn;
5000 * Return the number of pages a zone spans in a node, including holes
5001 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5003 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5004 unsigned long zone_type,
5005 unsigned long node_start_pfn,
5006 unsigned long node_end_pfn,
5007 unsigned long *zone_start_pfn,
5008 unsigned long *zone_end_pfn,
5009 unsigned long *ignored)
5011 /* When hotadd a new node from cpu_up(), the node should be empty */
5012 if (!node_start_pfn && !node_end_pfn)
5015 /* Get the start and end of the zone */
5016 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5017 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5018 adjust_zone_range_for_zone_movable(nid, zone_type,
5019 node_start_pfn, node_end_pfn,
5020 zone_start_pfn, zone_end_pfn);
5022 /* Check that this node has pages within the zone's required range */
5023 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5026 /* Move the zone boundaries inside the node if necessary */
5027 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5028 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5030 /* Return the spanned pages */
5031 return *zone_end_pfn - *zone_start_pfn;
5035 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5036 * then all holes in the requested range will be accounted for.
5038 unsigned long __meminit __absent_pages_in_range(int nid,
5039 unsigned long range_start_pfn,
5040 unsigned long range_end_pfn)
5042 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5043 unsigned long start_pfn, end_pfn;
5046 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5047 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5048 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5049 nr_absent -= end_pfn - start_pfn;
5055 * absent_pages_in_range - Return number of page frames in holes within a range
5056 * @start_pfn: The start PFN to start searching for holes
5057 * @end_pfn: The end PFN to stop searching for holes
5059 * It returns the number of pages frames in memory holes within a range.
5061 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5062 unsigned long end_pfn)
5064 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5067 /* Return the number of page frames in holes in a zone on a node */
5068 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5069 unsigned long zone_type,
5070 unsigned long node_start_pfn,
5071 unsigned long node_end_pfn,
5072 unsigned long *ignored)
5074 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5075 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5076 unsigned long zone_start_pfn, zone_end_pfn;
5077 unsigned long nr_absent;
5079 /* When hotadd a new node from cpu_up(), the node should be empty */
5080 if (!node_start_pfn && !node_end_pfn)
5083 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5084 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5086 adjust_zone_range_for_zone_movable(nid, zone_type,
5087 node_start_pfn, node_end_pfn,
5088 &zone_start_pfn, &zone_end_pfn);
5089 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5092 * ZONE_MOVABLE handling.
5093 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5096 if (zone_movable_pfn[nid]) {
5097 if (mirrored_kernelcore) {
5098 unsigned long start_pfn, end_pfn;
5099 struct memblock_region *r;
5101 for_each_memblock(memory, r) {
5102 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5103 zone_start_pfn, zone_end_pfn);
5104 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5105 zone_start_pfn, zone_end_pfn);
5107 if (zone_type == ZONE_MOVABLE &&
5108 memblock_is_mirror(r))
5109 nr_absent += end_pfn - start_pfn;
5111 if (zone_type == ZONE_NORMAL &&
5112 !memblock_is_mirror(r))
5113 nr_absent += end_pfn - start_pfn;
5116 if (zone_type == ZONE_NORMAL)
5117 nr_absent += node_end_pfn - zone_movable_pfn[nid];
5124 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5125 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5126 unsigned long zone_type,
5127 unsigned long node_start_pfn,
5128 unsigned long node_end_pfn,
5129 unsigned long *zone_start_pfn,
5130 unsigned long *zone_end_pfn,
5131 unsigned long *zones_size)
5135 *zone_start_pfn = node_start_pfn;
5136 for (zone = 0; zone < zone_type; zone++)
5137 *zone_start_pfn += zones_size[zone];
5139 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5141 return zones_size[zone_type];
5144 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5145 unsigned long zone_type,
5146 unsigned long node_start_pfn,
5147 unsigned long node_end_pfn,
5148 unsigned long *zholes_size)
5153 return zholes_size[zone_type];
5156 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5158 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5159 unsigned long node_start_pfn,
5160 unsigned long node_end_pfn,
5161 unsigned long *zones_size,
5162 unsigned long *zholes_size)
5164 unsigned long realtotalpages = 0, totalpages = 0;
5167 for (i = 0; i < MAX_NR_ZONES; i++) {
5168 struct zone *zone = pgdat->node_zones + i;
5169 unsigned long zone_start_pfn, zone_end_pfn;
5170 unsigned long size, real_size;
5172 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5178 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5179 node_start_pfn, node_end_pfn,
5182 zone->zone_start_pfn = zone_start_pfn;
5184 zone->zone_start_pfn = 0;
5185 zone->spanned_pages = size;
5186 zone->present_pages = real_size;
5189 realtotalpages += real_size;
5192 pgdat->node_spanned_pages = totalpages;
5193 pgdat->node_present_pages = realtotalpages;
5194 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5198 #ifndef CONFIG_SPARSEMEM
5200 * Calculate the size of the zone->blockflags rounded to an unsigned long
5201 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5202 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5203 * round what is now in bits to nearest long in bits, then return it in
5206 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5208 unsigned long usemapsize;
5210 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5211 usemapsize = roundup(zonesize, pageblock_nr_pages);
5212 usemapsize = usemapsize >> pageblock_order;
5213 usemapsize *= NR_PAGEBLOCK_BITS;
5214 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5216 return usemapsize / 8;
5219 static void __init setup_usemap(struct pglist_data *pgdat,
5221 unsigned long zone_start_pfn,
5222 unsigned long zonesize)
5224 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5225 zone->pageblock_flags = NULL;
5227 zone->pageblock_flags =
5228 memblock_virt_alloc_node_nopanic(usemapsize,
5232 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5233 unsigned long zone_start_pfn, unsigned long zonesize) {}
5234 #endif /* CONFIG_SPARSEMEM */
5236 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5238 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5239 void __paginginit set_pageblock_order(void)
5243 /* Check that pageblock_nr_pages has not already been setup */
5244 if (pageblock_order)
5247 if (HPAGE_SHIFT > PAGE_SHIFT)
5248 order = HUGETLB_PAGE_ORDER;
5250 order = MAX_ORDER - 1;
5253 * Assume the largest contiguous order of interest is a huge page.
5254 * This value may be variable depending on boot parameters on IA64 and
5257 pageblock_order = order;
5259 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5262 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5263 * is unused as pageblock_order is set at compile-time. See
5264 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5267 void __paginginit set_pageblock_order(void)
5271 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5273 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5274 unsigned long present_pages)
5276 unsigned long pages = spanned_pages;
5279 * Provide a more accurate estimation if there are holes within
5280 * the zone and SPARSEMEM is in use. If there are holes within the
5281 * zone, each populated memory region may cost us one or two extra
5282 * memmap pages due to alignment because memmap pages for each
5283 * populated regions may not naturally algined on page boundary.
5284 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5286 if (spanned_pages > present_pages + (present_pages >> 4) &&
5287 IS_ENABLED(CONFIG_SPARSEMEM))
5288 pages = present_pages;
5290 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5294 * Set up the zone data structures:
5295 * - mark all pages reserved
5296 * - mark all memory queues empty
5297 * - clear the memory bitmaps
5299 * NOTE: pgdat should get zeroed by caller.
5301 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5304 int nid = pgdat->node_id;
5307 pgdat_resize_init(pgdat);
5308 #ifdef CONFIG_NUMA_BALANCING
5309 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5310 pgdat->numabalancing_migrate_nr_pages = 0;
5311 pgdat->numabalancing_migrate_next_window = jiffies;
5313 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5314 spin_lock_init(&pgdat->split_queue_lock);
5315 INIT_LIST_HEAD(&pgdat->split_queue);
5316 pgdat->split_queue_len = 0;
5318 init_waitqueue_head(&pgdat->kswapd_wait);
5319 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5320 pgdat_page_ext_init(pgdat);
5322 for (j = 0; j < MAX_NR_ZONES; j++) {
5323 struct zone *zone = pgdat->node_zones + j;
5324 unsigned long size, realsize, freesize, memmap_pages;
5325 unsigned long zone_start_pfn = zone->zone_start_pfn;
5327 size = zone->spanned_pages;
5328 realsize = freesize = zone->present_pages;
5331 * Adjust freesize so that it accounts for how much memory
5332 * is used by this zone for memmap. This affects the watermark
5333 * and per-cpu initialisations
5335 memmap_pages = calc_memmap_size(size, realsize);
5336 if (!is_highmem_idx(j)) {
5337 if (freesize >= memmap_pages) {
5338 freesize -= memmap_pages;
5341 " %s zone: %lu pages used for memmap\n",
5342 zone_names[j], memmap_pages);
5345 " %s zone: %lu pages exceeds freesize %lu\n",
5346 zone_names[j], memmap_pages, freesize);
5349 /* Account for reserved pages */
5350 if (j == 0 && freesize > dma_reserve) {
5351 freesize -= dma_reserve;
5352 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5353 zone_names[0], dma_reserve);
5356 if (!is_highmem_idx(j))
5357 nr_kernel_pages += freesize;
5358 /* Charge for highmem memmap if there are enough kernel pages */
5359 else if (nr_kernel_pages > memmap_pages * 2)
5360 nr_kernel_pages -= memmap_pages;
5361 nr_all_pages += freesize;
5364 * Set an approximate value for lowmem here, it will be adjusted
5365 * when the bootmem allocator frees pages into the buddy system.
5366 * And all highmem pages will be managed by the buddy system.
5368 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5371 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5373 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5375 zone->name = zone_names[j];
5376 spin_lock_init(&zone->lock);
5377 spin_lock_init(&zone->lru_lock);
5378 zone_seqlock_init(zone);
5379 zone->zone_pgdat = pgdat;
5380 zone_pcp_init(zone);
5382 /* For bootup, initialized properly in watermark setup */
5383 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5385 lruvec_init(&zone->lruvec);
5389 set_pageblock_order();
5390 setup_usemap(pgdat, zone, zone_start_pfn, size);
5391 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5393 memmap_init(size, nid, j, zone_start_pfn);
5397 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5399 unsigned long __maybe_unused start = 0;
5400 unsigned long __maybe_unused offset = 0;
5402 /* Skip empty nodes */
5403 if (!pgdat->node_spanned_pages)
5406 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5407 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5408 offset = pgdat->node_start_pfn - start;
5409 /* ia64 gets its own node_mem_map, before this, without bootmem */
5410 if (!pgdat->node_mem_map) {
5411 unsigned long size, end;
5415 * The zone's endpoints aren't required to be MAX_ORDER
5416 * aligned but the node_mem_map endpoints must be in order
5417 * for the buddy allocator to function correctly.
5419 end = pgdat_end_pfn(pgdat);
5420 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5421 size = (end - start) * sizeof(struct page);
5422 map = alloc_remap(pgdat->node_id, size);
5424 map = memblock_virt_alloc_node_nopanic(size,
5426 pgdat->node_mem_map = map + offset;
5428 #ifndef CONFIG_NEED_MULTIPLE_NODES
5430 * With no DISCONTIG, the global mem_map is just set as node 0's
5432 if (pgdat == NODE_DATA(0)) {
5433 mem_map = NODE_DATA(0)->node_mem_map;
5434 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5435 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5437 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5440 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5443 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5444 unsigned long node_start_pfn, unsigned long *zholes_size)
5446 pg_data_t *pgdat = NODE_DATA(nid);
5447 unsigned long start_pfn = 0;
5448 unsigned long end_pfn = 0;
5450 /* pg_data_t should be reset to zero when it's allocated */
5451 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5453 reset_deferred_meminit(pgdat);
5454 pgdat->node_id = nid;
5455 pgdat->node_start_pfn = node_start_pfn;
5456 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5457 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5458 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5459 (u64)start_pfn << PAGE_SHIFT,
5460 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5462 start_pfn = node_start_pfn;
5464 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5465 zones_size, zholes_size);
5467 alloc_node_mem_map(pgdat);
5468 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5469 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5470 nid, (unsigned long)pgdat,
5471 (unsigned long)pgdat->node_mem_map);
5474 free_area_init_core(pgdat);
5477 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5479 #if MAX_NUMNODES > 1
5481 * Figure out the number of possible node ids.
5483 void __init setup_nr_node_ids(void)
5485 unsigned int highest;
5487 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5488 nr_node_ids = highest + 1;
5493 * node_map_pfn_alignment - determine the maximum internode alignment
5495 * This function should be called after node map is populated and sorted.
5496 * It calculates the maximum power of two alignment which can distinguish
5499 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5500 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5501 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5502 * shifted, 1GiB is enough and this function will indicate so.
5504 * This is used to test whether pfn -> nid mapping of the chosen memory
5505 * model has fine enough granularity to avoid incorrect mapping for the
5506 * populated node map.
5508 * Returns the determined alignment in pfn's. 0 if there is no alignment
5509 * requirement (single node).
5511 unsigned long __init node_map_pfn_alignment(void)
5513 unsigned long accl_mask = 0, last_end = 0;
5514 unsigned long start, end, mask;
5518 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5519 if (!start || last_nid < 0 || last_nid == nid) {
5526 * Start with a mask granular enough to pin-point to the
5527 * start pfn and tick off bits one-by-one until it becomes
5528 * too coarse to separate the current node from the last.
5530 mask = ~((1 << __ffs(start)) - 1);
5531 while (mask && last_end <= (start & (mask << 1)))
5534 /* accumulate all internode masks */
5538 /* convert mask to number of pages */
5539 return ~accl_mask + 1;
5542 /* Find the lowest pfn for a node */
5543 static unsigned long __init find_min_pfn_for_node(int nid)
5545 unsigned long min_pfn = ULONG_MAX;
5546 unsigned long start_pfn;
5549 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5550 min_pfn = min(min_pfn, start_pfn);
5552 if (min_pfn == ULONG_MAX) {
5554 "Could not find start_pfn for node %d\n", nid);
5562 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5564 * It returns the minimum PFN based on information provided via
5565 * memblock_set_node().
5567 unsigned long __init find_min_pfn_with_active_regions(void)
5569 return find_min_pfn_for_node(MAX_NUMNODES);
5573 * early_calculate_totalpages()
5574 * Sum pages in active regions for movable zone.
5575 * Populate N_MEMORY for calculating usable_nodes.
5577 static unsigned long __init early_calculate_totalpages(void)
5579 unsigned long totalpages = 0;
5580 unsigned long start_pfn, end_pfn;
5583 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5584 unsigned long pages = end_pfn - start_pfn;
5586 totalpages += pages;
5588 node_set_state(nid, N_MEMORY);
5594 * Find the PFN the Movable zone begins in each node. Kernel memory
5595 * is spread evenly between nodes as long as the nodes have enough
5596 * memory. When they don't, some nodes will have more kernelcore than
5599 static void __init find_zone_movable_pfns_for_nodes(void)
5602 unsigned long usable_startpfn;
5603 unsigned long kernelcore_node, kernelcore_remaining;
5604 /* save the state before borrow the nodemask */
5605 nodemask_t saved_node_state = node_states[N_MEMORY];
5606 unsigned long totalpages = early_calculate_totalpages();
5607 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5608 struct memblock_region *r;
5610 /* Need to find movable_zone earlier when movable_node is specified. */
5611 find_usable_zone_for_movable();
5614 * If movable_node is specified, ignore kernelcore and movablecore
5617 if (movable_node_is_enabled()) {
5618 for_each_memblock(memory, r) {
5619 if (!memblock_is_hotpluggable(r))
5624 usable_startpfn = PFN_DOWN(r->base);
5625 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5626 min(usable_startpfn, zone_movable_pfn[nid]) :
5634 * If kernelcore=mirror is specified, ignore movablecore option
5636 if (mirrored_kernelcore) {
5637 bool mem_below_4gb_not_mirrored = false;
5639 for_each_memblock(memory, r) {
5640 if (memblock_is_mirror(r))
5645 usable_startpfn = memblock_region_memory_base_pfn(r);
5647 if (usable_startpfn < 0x100000) {
5648 mem_below_4gb_not_mirrored = true;
5652 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5653 min(usable_startpfn, zone_movable_pfn[nid]) :
5657 if (mem_below_4gb_not_mirrored)
5658 pr_warn("This configuration results in unmirrored kernel memory.");
5664 * If movablecore=nn[KMG] was specified, calculate what size of
5665 * kernelcore that corresponds so that memory usable for
5666 * any allocation type is evenly spread. If both kernelcore
5667 * and movablecore are specified, then the value of kernelcore
5668 * will be used for required_kernelcore if it's greater than
5669 * what movablecore would have allowed.
5671 if (required_movablecore) {
5672 unsigned long corepages;
5675 * Round-up so that ZONE_MOVABLE is at least as large as what
5676 * was requested by the user
5678 required_movablecore =
5679 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5680 required_movablecore = min(totalpages, required_movablecore);
5681 corepages = totalpages - required_movablecore;
5683 required_kernelcore = max(required_kernelcore, corepages);
5687 * If kernelcore was not specified or kernelcore size is larger
5688 * than totalpages, there is no ZONE_MOVABLE.
5690 if (!required_kernelcore || required_kernelcore >= totalpages)
5693 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5694 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5697 /* Spread kernelcore memory as evenly as possible throughout nodes */
5698 kernelcore_node = required_kernelcore / usable_nodes;
5699 for_each_node_state(nid, N_MEMORY) {
5700 unsigned long start_pfn, end_pfn;
5703 * Recalculate kernelcore_node if the division per node
5704 * now exceeds what is necessary to satisfy the requested
5705 * amount of memory for the kernel
5707 if (required_kernelcore < kernelcore_node)
5708 kernelcore_node = required_kernelcore / usable_nodes;
5711 * As the map is walked, we track how much memory is usable
5712 * by the kernel using kernelcore_remaining. When it is
5713 * 0, the rest of the node is usable by ZONE_MOVABLE
5715 kernelcore_remaining = kernelcore_node;
5717 /* Go through each range of PFNs within this node */
5718 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5719 unsigned long size_pages;
5721 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5722 if (start_pfn >= end_pfn)
5725 /* Account for what is only usable for kernelcore */
5726 if (start_pfn < usable_startpfn) {
5727 unsigned long kernel_pages;
5728 kernel_pages = min(end_pfn, usable_startpfn)
5731 kernelcore_remaining -= min(kernel_pages,
5732 kernelcore_remaining);
5733 required_kernelcore -= min(kernel_pages,
5734 required_kernelcore);
5736 /* Continue if range is now fully accounted */
5737 if (end_pfn <= usable_startpfn) {
5740 * Push zone_movable_pfn to the end so
5741 * that if we have to rebalance
5742 * kernelcore across nodes, we will
5743 * not double account here
5745 zone_movable_pfn[nid] = end_pfn;
5748 start_pfn = usable_startpfn;
5752 * The usable PFN range for ZONE_MOVABLE is from
5753 * start_pfn->end_pfn. Calculate size_pages as the
5754 * number of pages used as kernelcore
5756 size_pages = end_pfn - start_pfn;
5757 if (size_pages > kernelcore_remaining)
5758 size_pages = kernelcore_remaining;
5759 zone_movable_pfn[nid] = start_pfn + size_pages;
5762 * Some kernelcore has been met, update counts and
5763 * break if the kernelcore for this node has been
5766 required_kernelcore -= min(required_kernelcore,
5768 kernelcore_remaining -= size_pages;
5769 if (!kernelcore_remaining)
5775 * If there is still required_kernelcore, we do another pass with one
5776 * less node in the count. This will push zone_movable_pfn[nid] further
5777 * along on the nodes that still have memory until kernelcore is
5781 if (usable_nodes && required_kernelcore > usable_nodes)
5785 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5786 for (nid = 0; nid < MAX_NUMNODES; nid++)
5787 zone_movable_pfn[nid] =
5788 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5791 /* restore the node_state */
5792 node_states[N_MEMORY] = saved_node_state;
5795 /* Any regular or high memory on that node ? */
5796 static void check_for_memory(pg_data_t *pgdat, int nid)
5798 enum zone_type zone_type;
5800 if (N_MEMORY == N_NORMAL_MEMORY)
5803 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5804 struct zone *zone = &pgdat->node_zones[zone_type];
5805 if (populated_zone(zone)) {
5806 node_set_state(nid, N_HIGH_MEMORY);
5807 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5808 zone_type <= ZONE_NORMAL)
5809 node_set_state(nid, N_NORMAL_MEMORY);
5816 * free_area_init_nodes - Initialise all pg_data_t and zone data
5817 * @max_zone_pfn: an array of max PFNs for each zone
5819 * This will call free_area_init_node() for each active node in the system.
5820 * Using the page ranges provided by memblock_set_node(), the size of each
5821 * zone in each node and their holes is calculated. If the maximum PFN
5822 * between two adjacent zones match, it is assumed that the zone is empty.
5823 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5824 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5825 * starts where the previous one ended. For example, ZONE_DMA32 starts
5826 * at arch_max_dma_pfn.
5828 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5830 unsigned long start_pfn, end_pfn;
5833 /* Record where the zone boundaries are */
5834 memset(arch_zone_lowest_possible_pfn, 0,
5835 sizeof(arch_zone_lowest_possible_pfn));
5836 memset(arch_zone_highest_possible_pfn, 0,
5837 sizeof(arch_zone_highest_possible_pfn));
5838 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5839 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5840 for (i = 1; i < MAX_NR_ZONES; i++) {
5841 if (i == ZONE_MOVABLE)
5843 arch_zone_lowest_possible_pfn[i] =
5844 arch_zone_highest_possible_pfn[i-1];
5845 arch_zone_highest_possible_pfn[i] =
5846 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5848 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5849 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5851 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5852 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5853 find_zone_movable_pfns_for_nodes();
5855 /* Print out the zone ranges */
5856 pr_info("Zone ranges:\n");
5857 for (i = 0; i < MAX_NR_ZONES; i++) {
5858 if (i == ZONE_MOVABLE)
5860 pr_info(" %-8s ", zone_names[i]);
5861 if (arch_zone_lowest_possible_pfn[i] ==
5862 arch_zone_highest_possible_pfn[i])
5865 pr_cont("[mem %#018Lx-%#018Lx]\n",
5866 (u64)arch_zone_lowest_possible_pfn[i]
5868 ((u64)arch_zone_highest_possible_pfn[i]
5869 << PAGE_SHIFT) - 1);
5872 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5873 pr_info("Movable zone start for each node\n");
5874 for (i = 0; i < MAX_NUMNODES; i++) {
5875 if (zone_movable_pfn[i])
5876 pr_info(" Node %d: %#018Lx\n", i,
5877 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5880 /* Print out the early node map */
5881 pr_info("Early memory node ranges\n");
5882 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5883 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5884 (u64)start_pfn << PAGE_SHIFT,
5885 ((u64)end_pfn << PAGE_SHIFT) - 1);
5887 /* Initialise every node */
5888 mminit_verify_pageflags_layout();
5889 setup_nr_node_ids();
5890 for_each_online_node(nid) {
5891 pg_data_t *pgdat = NODE_DATA(nid);
5892 free_area_init_node(nid, NULL,
5893 find_min_pfn_for_node(nid), NULL);
5895 /* Any memory on that node */
5896 if (pgdat->node_present_pages)
5897 node_set_state(nid, N_MEMORY);
5898 check_for_memory(pgdat, nid);
5902 static int __init cmdline_parse_core(char *p, unsigned long *core)
5904 unsigned long long coremem;
5908 coremem = memparse(p, &p);
5909 *core = coremem >> PAGE_SHIFT;
5911 /* Paranoid check that UL is enough for the coremem value */
5912 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5918 * kernelcore=size sets the amount of memory for use for allocations that
5919 * cannot be reclaimed or migrated.
5921 static int __init cmdline_parse_kernelcore(char *p)
5923 /* parse kernelcore=mirror */
5924 if (parse_option_str(p, "mirror")) {
5925 mirrored_kernelcore = true;
5929 return cmdline_parse_core(p, &required_kernelcore);
5933 * movablecore=size sets the amount of memory for use for allocations that
5934 * can be reclaimed or migrated.
5936 static int __init cmdline_parse_movablecore(char *p)
5938 return cmdline_parse_core(p, &required_movablecore);
5941 early_param("kernelcore", cmdline_parse_kernelcore);
5942 early_param("movablecore", cmdline_parse_movablecore);
5944 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5946 void adjust_managed_page_count(struct page *page, long count)
5948 spin_lock(&managed_page_count_lock);
5949 page_zone(page)->managed_pages += count;
5950 totalram_pages += count;
5951 #ifdef CONFIG_HIGHMEM
5952 if (PageHighMem(page))
5953 totalhigh_pages += count;
5955 spin_unlock(&managed_page_count_lock);
5957 EXPORT_SYMBOL(adjust_managed_page_count);
5959 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5962 unsigned long pages = 0;
5964 start = (void *)PAGE_ALIGN((unsigned long)start);
5965 end = (void *)((unsigned long)end & PAGE_MASK);
5966 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5967 if ((unsigned int)poison <= 0xFF)
5968 memset(pos, poison, PAGE_SIZE);
5969 free_reserved_page(virt_to_page(pos));
5973 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5974 s, pages << (PAGE_SHIFT - 10), start, end);
5978 EXPORT_SYMBOL(free_reserved_area);
5980 #ifdef CONFIG_HIGHMEM
5981 void free_highmem_page(struct page *page)
5983 __free_reserved_page(page);
5985 page_zone(page)->managed_pages++;
5991 void __init mem_init_print_info(const char *str)
5993 unsigned long physpages, codesize, datasize, rosize, bss_size;
5994 unsigned long init_code_size, init_data_size;
5996 physpages = get_num_physpages();
5997 codesize = _etext - _stext;
5998 datasize = _edata - _sdata;
5999 rosize = __end_rodata - __start_rodata;
6000 bss_size = __bss_stop - __bss_start;
6001 init_data_size = __init_end - __init_begin;
6002 init_code_size = _einittext - _sinittext;
6005 * Detect special cases and adjust section sizes accordingly:
6006 * 1) .init.* may be embedded into .data sections
6007 * 2) .init.text.* may be out of [__init_begin, __init_end],
6008 * please refer to arch/tile/kernel/vmlinux.lds.S.
6009 * 3) .rodata.* may be embedded into .text or .data sections.
6011 #define adj_init_size(start, end, size, pos, adj) \
6013 if (start <= pos && pos < end && size > adj) \
6017 adj_init_size(__init_begin, __init_end, init_data_size,
6018 _sinittext, init_code_size);
6019 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6020 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6021 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6022 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6024 #undef adj_init_size
6026 pr_info("Memory: %luK/%luK available "
6027 "(%luK kernel code, %luK rwdata, %luK rodata, "
6028 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
6029 #ifdef CONFIG_HIGHMEM
6033 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
6034 codesize >> 10, datasize >> 10, rosize >> 10,
6035 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6036 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
6037 totalcma_pages << (PAGE_SHIFT-10),
6038 #ifdef CONFIG_HIGHMEM
6039 totalhigh_pages << (PAGE_SHIFT-10),
6041 str ? ", " : "", str ? str : "");
6045 * set_dma_reserve - set the specified number of pages reserved in the first zone
6046 * @new_dma_reserve: The number of pages to mark reserved
6048 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6049 * In the DMA zone, a significant percentage may be consumed by kernel image
6050 * and other unfreeable allocations which can skew the watermarks badly. This
6051 * function may optionally be used to account for unfreeable pages in the
6052 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6053 * smaller per-cpu batchsize.
6055 void __init set_dma_reserve(unsigned long new_dma_reserve)
6057 dma_reserve = new_dma_reserve;
6060 void __init free_area_init(unsigned long *zones_size)
6062 free_area_init_node(0, zones_size,
6063 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6066 static int page_alloc_cpu_notify(struct notifier_block *self,
6067 unsigned long action, void *hcpu)
6069 int cpu = (unsigned long)hcpu;
6071 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6072 lru_add_drain_cpu(cpu);
6076 * Spill the event counters of the dead processor
6077 * into the current processors event counters.
6078 * This artificially elevates the count of the current
6081 vm_events_fold_cpu(cpu);
6084 * Zero the differential counters of the dead processor
6085 * so that the vm statistics are consistent.
6087 * This is only okay since the processor is dead and cannot
6088 * race with what we are doing.
6090 cpu_vm_stats_fold(cpu);
6095 void __init page_alloc_init(void)
6097 hotcpu_notifier(page_alloc_cpu_notify, 0);
6101 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6102 * or min_free_kbytes changes.
6104 static void calculate_totalreserve_pages(void)
6106 struct pglist_data *pgdat;
6107 unsigned long reserve_pages = 0;
6108 enum zone_type i, j;
6110 for_each_online_pgdat(pgdat) {
6111 for (i = 0; i < MAX_NR_ZONES; i++) {
6112 struct zone *zone = pgdat->node_zones + i;
6115 /* Find valid and maximum lowmem_reserve in the zone */
6116 for (j = i; j < MAX_NR_ZONES; j++) {
6117 if (zone->lowmem_reserve[j] > max)
6118 max = zone->lowmem_reserve[j];
6121 /* we treat the high watermark as reserved pages. */
6122 max += high_wmark_pages(zone);
6124 if (max > zone->managed_pages)
6125 max = zone->managed_pages;
6127 zone->totalreserve_pages = max;
6129 reserve_pages += max;
6132 totalreserve_pages = reserve_pages;
6136 * setup_per_zone_lowmem_reserve - called whenever
6137 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6138 * has a correct pages reserved value, so an adequate number of
6139 * pages are left in the zone after a successful __alloc_pages().
6141 static void setup_per_zone_lowmem_reserve(void)
6143 struct pglist_data *pgdat;
6144 enum zone_type j, idx;
6146 for_each_online_pgdat(pgdat) {
6147 for (j = 0; j < MAX_NR_ZONES; j++) {
6148 struct zone *zone = pgdat->node_zones + j;
6149 unsigned long managed_pages = zone->managed_pages;
6151 zone->lowmem_reserve[j] = 0;
6155 struct zone *lower_zone;
6159 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6160 sysctl_lowmem_reserve_ratio[idx] = 1;
6162 lower_zone = pgdat->node_zones + idx;
6163 lower_zone->lowmem_reserve[j] = managed_pages /
6164 sysctl_lowmem_reserve_ratio[idx];
6165 managed_pages += lower_zone->managed_pages;
6170 /* update totalreserve_pages */
6171 calculate_totalreserve_pages();
6174 static void __setup_per_zone_wmarks(void)
6176 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6177 unsigned long lowmem_pages = 0;
6179 unsigned long flags;
6181 /* Calculate total number of !ZONE_HIGHMEM pages */
6182 for_each_zone(zone) {
6183 if (!is_highmem(zone))
6184 lowmem_pages += zone->managed_pages;
6187 for_each_zone(zone) {
6190 spin_lock_irqsave(&zone->lock, flags);
6191 tmp = (u64)pages_min * zone->managed_pages;
6192 do_div(tmp, lowmem_pages);
6193 if (is_highmem(zone)) {
6195 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6196 * need highmem pages, so cap pages_min to a small
6199 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6200 * deltas control asynch page reclaim, and so should
6201 * not be capped for highmem.
6203 unsigned long min_pages;
6205 min_pages = zone->managed_pages / 1024;
6206 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6207 zone->watermark[WMARK_MIN] = min_pages;
6210 * If it's a lowmem zone, reserve a number of pages
6211 * proportionate to the zone's size.
6213 zone->watermark[WMARK_MIN] = tmp;
6216 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
6217 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6219 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6220 high_wmark_pages(zone) - low_wmark_pages(zone) -
6221 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6223 spin_unlock_irqrestore(&zone->lock, flags);
6226 /* update totalreserve_pages */
6227 calculate_totalreserve_pages();
6231 * setup_per_zone_wmarks - called when min_free_kbytes changes
6232 * or when memory is hot-{added|removed}
6234 * Ensures that the watermark[min,low,high] values for each zone are set
6235 * correctly with respect to min_free_kbytes.
6237 void setup_per_zone_wmarks(void)
6239 mutex_lock(&zonelists_mutex);
6240 __setup_per_zone_wmarks();
6241 mutex_unlock(&zonelists_mutex);
6245 * The inactive anon list should be small enough that the VM never has to
6246 * do too much work, but large enough that each inactive page has a chance
6247 * to be referenced again before it is swapped out.
6249 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6250 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6251 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6252 * the anonymous pages are kept on the inactive list.
6255 * memory ratio inactive anon
6256 * -------------------------------------
6265 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6267 unsigned int gb, ratio;
6269 /* Zone size in gigabytes */
6270 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6272 ratio = int_sqrt(10 * gb);
6276 zone->inactive_ratio = ratio;
6279 static void __meminit setup_per_zone_inactive_ratio(void)
6284 calculate_zone_inactive_ratio(zone);
6288 * Initialise min_free_kbytes.
6290 * For small machines we want it small (128k min). For large machines
6291 * we want it large (64MB max). But it is not linear, because network
6292 * bandwidth does not increase linearly with machine size. We use
6294 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6295 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6311 int __meminit init_per_zone_wmark_min(void)
6313 unsigned long lowmem_kbytes;
6314 int new_min_free_kbytes;
6316 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6317 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6319 if (new_min_free_kbytes > user_min_free_kbytes) {
6320 min_free_kbytes = new_min_free_kbytes;
6321 if (min_free_kbytes < 128)
6322 min_free_kbytes = 128;
6323 if (min_free_kbytes > 65536)
6324 min_free_kbytes = 65536;
6326 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6327 new_min_free_kbytes, user_min_free_kbytes);
6329 setup_per_zone_wmarks();
6330 refresh_zone_stat_thresholds();
6331 setup_per_zone_lowmem_reserve();
6332 setup_per_zone_inactive_ratio();
6335 module_init(init_per_zone_wmark_min)
6338 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6339 * that we can call two helper functions whenever min_free_kbytes
6342 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6343 void __user *buffer, size_t *length, loff_t *ppos)
6347 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6352 user_min_free_kbytes = min_free_kbytes;
6353 setup_per_zone_wmarks();
6359 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6360 void __user *buffer, size_t *length, loff_t *ppos)
6365 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6370 zone->min_unmapped_pages = (zone->managed_pages *
6371 sysctl_min_unmapped_ratio) / 100;
6375 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6376 void __user *buffer, size_t *length, loff_t *ppos)
6381 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6386 zone->min_slab_pages = (zone->managed_pages *
6387 sysctl_min_slab_ratio) / 100;
6393 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6394 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6395 * whenever sysctl_lowmem_reserve_ratio changes.
6397 * The reserve ratio obviously has absolutely no relation with the
6398 * minimum watermarks. The lowmem reserve ratio can only make sense
6399 * if in function of the boot time zone sizes.
6401 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6402 void __user *buffer, size_t *length, loff_t *ppos)
6404 proc_dointvec_minmax(table, write, buffer, length, ppos);
6405 setup_per_zone_lowmem_reserve();
6410 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6411 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6412 * pagelist can have before it gets flushed back to buddy allocator.
6414 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6415 void __user *buffer, size_t *length, loff_t *ppos)
6418 int old_percpu_pagelist_fraction;
6421 mutex_lock(&pcp_batch_high_lock);
6422 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6424 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6425 if (!write || ret < 0)
6428 /* Sanity checking to avoid pcp imbalance */
6429 if (percpu_pagelist_fraction &&
6430 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6431 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6437 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6440 for_each_populated_zone(zone) {
6443 for_each_possible_cpu(cpu)
6444 pageset_set_high_and_batch(zone,
6445 per_cpu_ptr(zone->pageset, cpu));
6448 mutex_unlock(&pcp_batch_high_lock);
6453 int hashdist = HASHDIST_DEFAULT;
6455 static int __init set_hashdist(char *str)
6459 hashdist = simple_strtoul(str, &str, 0);
6462 __setup("hashdist=", set_hashdist);
6466 * allocate a large system hash table from bootmem
6467 * - it is assumed that the hash table must contain an exact power-of-2
6468 * quantity of entries
6469 * - limit is the number of hash buckets, not the total allocation size
6471 void *__init alloc_large_system_hash(const char *tablename,
6472 unsigned long bucketsize,
6473 unsigned long numentries,
6476 unsigned int *_hash_shift,
6477 unsigned int *_hash_mask,
6478 unsigned long low_limit,
6479 unsigned long high_limit)
6481 unsigned long long max = high_limit;
6482 unsigned long log2qty, size;
6485 /* allow the kernel cmdline to have a say */
6487 /* round applicable memory size up to nearest megabyte */
6488 numentries = nr_kernel_pages;
6490 /* It isn't necessary when PAGE_SIZE >= 1MB */
6491 if (PAGE_SHIFT < 20)
6492 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6494 /* limit to 1 bucket per 2^scale bytes of low memory */
6495 if (scale > PAGE_SHIFT)
6496 numentries >>= (scale - PAGE_SHIFT);
6498 numentries <<= (PAGE_SHIFT - scale);
6500 /* Make sure we've got at least a 0-order allocation.. */
6501 if (unlikely(flags & HASH_SMALL)) {
6502 /* Makes no sense without HASH_EARLY */
6503 WARN_ON(!(flags & HASH_EARLY));
6504 if (!(numentries >> *_hash_shift)) {
6505 numentries = 1UL << *_hash_shift;
6506 BUG_ON(!numentries);
6508 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6509 numentries = PAGE_SIZE / bucketsize;
6511 numentries = roundup_pow_of_two(numentries);
6513 /* limit allocation size to 1/16 total memory by default */
6515 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6516 do_div(max, bucketsize);
6518 max = min(max, 0x80000000ULL);
6520 if (numentries < low_limit)
6521 numentries = low_limit;
6522 if (numentries > max)
6525 log2qty = ilog2(numentries);
6528 size = bucketsize << log2qty;
6529 if (flags & HASH_EARLY)
6530 table = memblock_virt_alloc_nopanic(size, 0);
6532 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6535 * If bucketsize is not a power-of-two, we may free
6536 * some pages at the end of hash table which
6537 * alloc_pages_exact() automatically does
6539 if (get_order(size) < MAX_ORDER) {
6540 table = alloc_pages_exact(size, GFP_ATOMIC);
6541 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6544 } while (!table && size > PAGE_SIZE && --log2qty);
6547 panic("Failed to allocate %s hash table\n", tablename);
6549 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6552 ilog2(size) - PAGE_SHIFT,
6556 *_hash_shift = log2qty;
6558 *_hash_mask = (1 << log2qty) - 1;
6563 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6564 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6567 #ifdef CONFIG_SPARSEMEM
6568 return __pfn_to_section(pfn)->pageblock_flags;
6570 return zone->pageblock_flags;
6571 #endif /* CONFIG_SPARSEMEM */
6574 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6576 #ifdef CONFIG_SPARSEMEM
6577 pfn &= (PAGES_PER_SECTION-1);
6578 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6580 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6581 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6582 #endif /* CONFIG_SPARSEMEM */
6586 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6587 * @page: The page within the block of interest
6588 * @pfn: The target page frame number
6589 * @end_bitidx: The last bit of interest to retrieve
6590 * @mask: mask of bits that the caller is interested in
6592 * Return: pageblock_bits flags
6594 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6595 unsigned long end_bitidx,
6599 unsigned long *bitmap;
6600 unsigned long bitidx, word_bitidx;
6603 zone = page_zone(page);
6604 bitmap = get_pageblock_bitmap(zone, pfn);
6605 bitidx = pfn_to_bitidx(zone, pfn);
6606 word_bitidx = bitidx / BITS_PER_LONG;
6607 bitidx &= (BITS_PER_LONG-1);
6609 word = bitmap[word_bitidx];
6610 bitidx += end_bitidx;
6611 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6615 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6616 * @page: The page within the block of interest
6617 * @flags: The flags to set
6618 * @pfn: The target page frame number
6619 * @end_bitidx: The last bit of interest
6620 * @mask: mask of bits that the caller is interested in
6622 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6624 unsigned long end_bitidx,
6628 unsigned long *bitmap;
6629 unsigned long bitidx, word_bitidx;
6630 unsigned long old_word, word;
6632 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6634 zone = page_zone(page);
6635 bitmap = get_pageblock_bitmap(zone, pfn);
6636 bitidx = pfn_to_bitidx(zone, pfn);
6637 word_bitidx = bitidx / BITS_PER_LONG;
6638 bitidx &= (BITS_PER_LONG-1);
6640 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6642 bitidx += end_bitidx;
6643 mask <<= (BITS_PER_LONG - bitidx - 1);
6644 flags <<= (BITS_PER_LONG - bitidx - 1);
6646 word = READ_ONCE(bitmap[word_bitidx]);
6648 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6649 if (word == old_word)
6656 * This function checks whether pageblock includes unmovable pages or not.
6657 * If @count is not zero, it is okay to include less @count unmovable pages
6659 * PageLRU check without isolation or lru_lock could race so that
6660 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6661 * expect this function should be exact.
6663 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6664 bool skip_hwpoisoned_pages)
6666 unsigned long pfn, iter, found;
6670 * For avoiding noise data, lru_add_drain_all() should be called
6671 * If ZONE_MOVABLE, the zone never contains unmovable pages
6673 if (zone_idx(zone) == ZONE_MOVABLE)
6675 mt = get_pageblock_migratetype(page);
6676 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6679 pfn = page_to_pfn(page);
6680 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6681 unsigned long check = pfn + iter;
6683 if (!pfn_valid_within(check))
6686 page = pfn_to_page(check);
6689 * Hugepages are not in LRU lists, but they're movable.
6690 * We need not scan over tail pages bacause we don't
6691 * handle each tail page individually in migration.
6693 if (PageHuge(page)) {
6694 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6699 * We can't use page_count without pin a page
6700 * because another CPU can free compound page.
6701 * This check already skips compound tails of THP
6702 * because their page->_count is zero at all time.
6704 if (!atomic_read(&page->_count)) {
6705 if (PageBuddy(page))
6706 iter += (1 << page_order(page)) - 1;
6711 * The HWPoisoned page may be not in buddy system, and
6712 * page_count() is not 0.
6714 if (skip_hwpoisoned_pages && PageHWPoison(page))
6720 * If there are RECLAIMABLE pages, we need to check
6721 * it. But now, memory offline itself doesn't call
6722 * shrink_node_slabs() and it still to be fixed.
6725 * If the page is not RAM, page_count()should be 0.
6726 * we don't need more check. This is an _used_ not-movable page.
6728 * The problematic thing here is PG_reserved pages. PG_reserved
6729 * is set to both of a memory hole page and a _used_ kernel
6738 bool is_pageblock_removable_nolock(struct page *page)
6744 * We have to be careful here because we are iterating over memory
6745 * sections which are not zone aware so we might end up outside of
6746 * the zone but still within the section.
6747 * We have to take care about the node as well. If the node is offline
6748 * its NODE_DATA will be NULL - see page_zone.
6750 if (!node_online(page_to_nid(page)))
6753 zone = page_zone(page);
6754 pfn = page_to_pfn(page);
6755 if (!zone_spans_pfn(zone, pfn))
6758 return !has_unmovable_pages(zone, page, 0, true);
6761 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
6763 static unsigned long pfn_max_align_down(unsigned long pfn)
6765 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6766 pageblock_nr_pages) - 1);
6769 static unsigned long pfn_max_align_up(unsigned long pfn)
6771 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6772 pageblock_nr_pages));
6775 /* [start, end) must belong to a single zone. */
6776 static int __alloc_contig_migrate_range(struct compact_control *cc,
6777 unsigned long start, unsigned long end)
6779 /* This function is based on compact_zone() from compaction.c. */
6780 unsigned long nr_reclaimed;
6781 unsigned long pfn = start;
6782 unsigned int tries = 0;
6787 while (pfn < end || !list_empty(&cc->migratepages)) {
6788 if (fatal_signal_pending(current)) {
6793 if (list_empty(&cc->migratepages)) {
6794 cc->nr_migratepages = 0;
6795 pfn = isolate_migratepages_range(cc, pfn, end);
6801 } else if (++tries == 5) {
6802 ret = ret < 0 ? ret : -EBUSY;
6806 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6808 cc->nr_migratepages -= nr_reclaimed;
6810 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6811 NULL, 0, cc->mode, MR_CMA);
6814 putback_movable_pages(&cc->migratepages);
6821 * alloc_contig_range() -- tries to allocate given range of pages
6822 * @start: start PFN to allocate
6823 * @end: one-past-the-last PFN to allocate
6824 * @migratetype: migratetype of the underlaying pageblocks (either
6825 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6826 * in range must have the same migratetype and it must
6827 * be either of the two.
6829 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6830 * aligned, however it's the caller's responsibility to guarantee that
6831 * we are the only thread that changes migrate type of pageblocks the
6834 * The PFN range must belong to a single zone.
6836 * Returns zero on success or negative error code. On success all
6837 * pages which PFN is in [start, end) are allocated for the caller and
6838 * need to be freed with free_contig_range().
6840 int alloc_contig_range(unsigned long start, unsigned long end,
6841 unsigned migratetype)
6843 unsigned long outer_start, outer_end;
6847 struct compact_control cc = {
6848 .nr_migratepages = 0,
6850 .zone = page_zone(pfn_to_page(start)),
6851 .mode = MIGRATE_SYNC,
6852 .ignore_skip_hint = true,
6854 INIT_LIST_HEAD(&cc.migratepages);
6857 * What we do here is we mark all pageblocks in range as
6858 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6859 * have different sizes, and due to the way page allocator
6860 * work, we align the range to biggest of the two pages so
6861 * that page allocator won't try to merge buddies from
6862 * different pageblocks and change MIGRATE_ISOLATE to some
6863 * other migration type.
6865 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6866 * migrate the pages from an unaligned range (ie. pages that
6867 * we are interested in). This will put all the pages in
6868 * range back to page allocator as MIGRATE_ISOLATE.
6870 * When this is done, we take the pages in range from page
6871 * allocator removing them from the buddy system. This way
6872 * page allocator will never consider using them.
6874 * This lets us mark the pageblocks back as
6875 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6876 * aligned range but not in the unaligned, original range are
6877 * put back to page allocator so that buddy can use them.
6880 ret = start_isolate_page_range(pfn_max_align_down(start),
6881 pfn_max_align_up(end), migratetype,
6887 * In case of -EBUSY, we'd like to know which page causes problem.
6888 * So, just fall through. We will check it in test_pages_isolated().
6890 ret = __alloc_contig_migrate_range(&cc, start, end);
6891 if (ret && ret != -EBUSY)
6895 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6896 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6897 * more, all pages in [start, end) are free in page allocator.
6898 * What we are going to do is to allocate all pages from
6899 * [start, end) (that is remove them from page allocator).
6901 * The only problem is that pages at the beginning and at the
6902 * end of interesting range may be not aligned with pages that
6903 * page allocator holds, ie. they can be part of higher order
6904 * pages. Because of this, we reserve the bigger range and
6905 * once this is done free the pages we are not interested in.
6907 * We don't have to hold zone->lock here because the pages are
6908 * isolated thus they won't get removed from buddy.
6911 lru_add_drain_all();
6912 drain_all_pages(cc.zone);
6915 outer_start = start;
6916 while (!PageBuddy(pfn_to_page(outer_start))) {
6917 if (++order >= MAX_ORDER) {
6918 outer_start = start;
6921 outer_start &= ~0UL << order;
6924 if (outer_start != start) {
6925 order = page_order(pfn_to_page(outer_start));
6928 * outer_start page could be small order buddy page and
6929 * it doesn't include start page. Adjust outer_start
6930 * in this case to report failed page properly
6931 * on tracepoint in test_pages_isolated()
6933 if (outer_start + (1UL << order) <= start)
6934 outer_start = start;
6937 /* Make sure the range is really isolated. */
6938 if (test_pages_isolated(outer_start, end, false)) {
6939 pr_info("%s: [%lx, %lx) PFNs busy\n",
6940 __func__, outer_start, end);
6945 /* Grab isolated pages from freelists. */
6946 outer_end = isolate_freepages_range(&cc, outer_start, end);
6952 /* Free head and tail (if any) */
6953 if (start != outer_start)
6954 free_contig_range(outer_start, start - outer_start);
6955 if (end != outer_end)
6956 free_contig_range(end, outer_end - end);
6959 undo_isolate_page_range(pfn_max_align_down(start),
6960 pfn_max_align_up(end), migratetype);
6964 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6966 unsigned int count = 0;
6968 for (; nr_pages--; pfn++) {
6969 struct page *page = pfn_to_page(pfn);
6971 count += page_count(page) != 1;
6974 WARN(count != 0, "%d pages are still in use!\n", count);
6978 #ifdef CONFIG_MEMORY_HOTPLUG
6980 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6981 * page high values need to be recalulated.
6983 void __meminit zone_pcp_update(struct zone *zone)
6986 mutex_lock(&pcp_batch_high_lock);
6987 for_each_possible_cpu(cpu)
6988 pageset_set_high_and_batch(zone,
6989 per_cpu_ptr(zone->pageset, cpu));
6990 mutex_unlock(&pcp_batch_high_lock);
6994 void zone_pcp_reset(struct zone *zone)
6996 unsigned long flags;
6998 struct per_cpu_pageset *pset;
7000 /* avoid races with drain_pages() */
7001 local_irq_save(flags);
7002 if (zone->pageset != &boot_pageset) {
7003 for_each_online_cpu(cpu) {
7004 pset = per_cpu_ptr(zone->pageset, cpu);
7005 drain_zonestat(zone, pset);
7007 free_percpu(zone->pageset);
7008 zone->pageset = &boot_pageset;
7010 local_irq_restore(flags);
7013 #ifdef CONFIG_MEMORY_HOTREMOVE
7015 * All pages in the range must be isolated before calling this.
7018 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7022 unsigned int order, i;
7024 unsigned long flags;
7025 /* find the first valid pfn */
7026 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7031 zone = page_zone(pfn_to_page(pfn));
7032 spin_lock_irqsave(&zone->lock, flags);
7034 while (pfn < end_pfn) {
7035 if (!pfn_valid(pfn)) {
7039 page = pfn_to_page(pfn);
7041 * The HWPoisoned page may be not in buddy system, and
7042 * page_count() is not 0.
7044 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7046 SetPageReserved(page);
7050 BUG_ON(page_count(page));
7051 BUG_ON(!PageBuddy(page));
7052 order = page_order(page);
7053 #ifdef CONFIG_DEBUG_VM
7054 printk(KERN_INFO "remove from free list %lx %d %lx\n",
7055 pfn, 1 << order, end_pfn);
7057 list_del(&page->lru);
7058 rmv_page_order(page);
7059 zone->free_area[order].nr_free--;
7060 for (i = 0; i < (1 << order); i++)
7061 SetPageReserved((page+i));
7062 pfn += (1 << order);
7064 spin_unlock_irqrestore(&zone->lock, flags);
7068 #ifdef CONFIG_MEMORY_FAILURE
7069 bool is_free_buddy_page(struct page *page)
7071 struct zone *zone = page_zone(page);
7072 unsigned long pfn = page_to_pfn(page);
7073 unsigned long flags;
7076 spin_lock_irqsave(&zone->lock, flags);
7077 for (order = 0; order < MAX_ORDER; order++) {
7078 struct page *page_head = page - (pfn & ((1 << order) - 1));
7080 if (PageBuddy(page_head) && page_order(page_head) >= order)
7083 spin_unlock_irqrestore(&zone->lock, flags);
7085 return order < MAX_ORDER;