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;
252 int watermark_scale_factor = 10;
254 static unsigned long __meminitdata nr_kernel_pages;
255 static unsigned long __meminitdata nr_all_pages;
256 static unsigned long __meminitdata dma_reserve;
258 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
259 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
260 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
261 static unsigned long __initdata required_kernelcore;
262 static unsigned long __initdata required_movablecore;
263 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
264 static bool mirrored_kernelcore;
266 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
268 EXPORT_SYMBOL(movable_zone);
269 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
272 int nr_node_ids __read_mostly = MAX_NUMNODES;
273 int nr_online_nodes __read_mostly = 1;
274 EXPORT_SYMBOL(nr_node_ids);
275 EXPORT_SYMBOL(nr_online_nodes);
278 int page_group_by_mobility_disabled __read_mostly;
280 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
281 static inline void reset_deferred_meminit(pg_data_t *pgdat)
283 pgdat->first_deferred_pfn = ULONG_MAX;
286 /* Returns true if the struct page for the pfn is uninitialised */
287 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
289 if (pfn >= NODE_DATA(early_pfn_to_nid(pfn))->first_deferred_pfn)
295 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
297 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
304 * Returns false when the remaining initialisation should be deferred until
305 * later in the boot cycle when it can be parallelised.
307 static inline bool update_defer_init(pg_data_t *pgdat,
308 unsigned long pfn, unsigned long zone_end,
309 unsigned long *nr_initialised)
311 unsigned long max_initialise;
313 /* Always populate low zones for address-contrained allocations */
314 if (zone_end < pgdat_end_pfn(pgdat))
317 * Initialise at least 2G of a node but also take into account that
318 * two large system hashes that can take up 1GB for 0.25TB/node.
320 max_initialise = max(2UL << (30 - PAGE_SHIFT),
321 (pgdat->node_spanned_pages >> 8));
324 if ((*nr_initialised > max_initialise) &&
325 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
326 pgdat->first_deferred_pfn = pfn;
333 static inline void reset_deferred_meminit(pg_data_t *pgdat)
337 static inline bool early_page_uninitialised(unsigned long pfn)
342 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
347 static inline bool update_defer_init(pg_data_t *pgdat,
348 unsigned long pfn, unsigned long zone_end,
349 unsigned long *nr_initialised)
356 void set_pageblock_migratetype(struct page *page, int migratetype)
358 if (unlikely(page_group_by_mobility_disabled &&
359 migratetype < MIGRATE_PCPTYPES))
360 migratetype = MIGRATE_UNMOVABLE;
362 set_pageblock_flags_group(page, (unsigned long)migratetype,
363 PB_migrate, PB_migrate_end);
366 #ifdef CONFIG_DEBUG_VM
367 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
371 unsigned long pfn = page_to_pfn(page);
372 unsigned long sp, start_pfn;
375 seq = zone_span_seqbegin(zone);
376 start_pfn = zone->zone_start_pfn;
377 sp = zone->spanned_pages;
378 if (!zone_spans_pfn(zone, pfn))
380 } while (zone_span_seqretry(zone, seq));
383 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
384 pfn, zone_to_nid(zone), zone->name,
385 start_pfn, start_pfn + sp);
390 static int page_is_consistent(struct zone *zone, struct page *page)
392 if (!pfn_valid_within(page_to_pfn(page)))
394 if (zone != page_zone(page))
400 * Temporary debugging check for pages not lying within a given zone.
402 static int bad_range(struct zone *zone, struct page *page)
404 if (page_outside_zone_boundaries(zone, page))
406 if (!page_is_consistent(zone, page))
412 static inline int bad_range(struct zone *zone, struct page *page)
418 static void bad_page(struct page *page, const char *reason,
419 unsigned long bad_flags)
421 static unsigned long resume;
422 static unsigned long nr_shown;
423 static unsigned long nr_unshown;
425 /* Don't complain about poisoned pages */
426 if (PageHWPoison(page)) {
427 page_mapcount_reset(page); /* remove PageBuddy */
432 * Allow a burst of 60 reports, then keep quiet for that minute;
433 * or allow a steady drip of one report per second.
435 if (nr_shown == 60) {
436 if (time_before(jiffies, resume)) {
442 "BUG: Bad page state: %lu messages suppressed\n",
449 resume = jiffies + 60 * HZ;
451 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
452 current->comm, page_to_pfn(page));
453 __dump_page(page, reason);
454 bad_flags &= page->flags;
456 pr_alert("bad because of flags: %#lx(%pGp)\n",
457 bad_flags, &bad_flags);
458 dump_page_owner(page);
463 /* Leave bad fields for debug, except PageBuddy could make trouble */
464 page_mapcount_reset(page); /* remove PageBuddy */
465 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
469 * Higher-order pages are called "compound pages". They are structured thusly:
471 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
473 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
474 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
476 * The first tail page's ->compound_dtor holds the offset in array of compound
477 * page destructors. See compound_page_dtors.
479 * The first tail page's ->compound_order holds the order of allocation.
480 * This usage means that zero-order pages may not be compound.
483 void free_compound_page(struct page *page)
485 __free_pages_ok(page, compound_order(page));
488 void prep_compound_page(struct page *page, unsigned int order)
491 int nr_pages = 1 << order;
493 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
494 set_compound_order(page, order);
496 for (i = 1; i < nr_pages; i++) {
497 struct page *p = page + i;
498 set_page_count(p, 0);
499 p->mapping = TAIL_MAPPING;
500 set_compound_head(p, page);
502 atomic_set(compound_mapcount_ptr(page), -1);
505 #ifdef CONFIG_DEBUG_PAGEALLOC
506 unsigned int _debug_guardpage_minorder;
507 bool _debug_pagealloc_enabled __read_mostly
508 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
509 EXPORT_SYMBOL(_debug_pagealloc_enabled);
510 bool _debug_guardpage_enabled __read_mostly;
512 static int __init early_debug_pagealloc(char *buf)
517 if (strcmp(buf, "on") == 0)
518 _debug_pagealloc_enabled = true;
520 if (strcmp(buf, "off") == 0)
521 _debug_pagealloc_enabled = false;
525 early_param("debug_pagealloc", early_debug_pagealloc);
527 static bool need_debug_guardpage(void)
529 /* If we don't use debug_pagealloc, we don't need guard page */
530 if (!debug_pagealloc_enabled())
536 static void init_debug_guardpage(void)
538 if (!debug_pagealloc_enabled())
541 _debug_guardpage_enabled = true;
544 struct page_ext_operations debug_guardpage_ops = {
545 .need = need_debug_guardpage,
546 .init = init_debug_guardpage,
549 static int __init debug_guardpage_minorder_setup(char *buf)
553 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
554 pr_err("Bad debug_guardpage_minorder value\n");
557 _debug_guardpage_minorder = res;
558 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
561 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
563 static inline void set_page_guard(struct zone *zone, struct page *page,
564 unsigned int order, int migratetype)
566 struct page_ext *page_ext;
568 if (!debug_guardpage_enabled())
571 page_ext = lookup_page_ext(page);
572 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
574 INIT_LIST_HEAD(&page->lru);
575 set_page_private(page, order);
576 /* Guard pages are not available for any usage */
577 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
580 static inline void clear_page_guard(struct zone *zone, struct page *page,
581 unsigned int order, int migratetype)
583 struct page_ext *page_ext;
585 if (!debug_guardpage_enabled())
588 page_ext = lookup_page_ext(page);
589 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
591 set_page_private(page, 0);
592 if (!is_migrate_isolate(migratetype))
593 __mod_zone_freepage_state(zone, (1 << order), migratetype);
596 struct page_ext_operations debug_guardpage_ops = { NULL, };
597 static inline void set_page_guard(struct zone *zone, struct page *page,
598 unsigned int order, int migratetype) {}
599 static inline void clear_page_guard(struct zone *zone, struct page *page,
600 unsigned int order, int migratetype) {}
603 static inline void set_page_order(struct page *page, unsigned int order)
605 set_page_private(page, order);
606 __SetPageBuddy(page);
609 static inline void rmv_page_order(struct page *page)
611 __ClearPageBuddy(page);
612 set_page_private(page, 0);
616 * This function checks whether a page is free && is the buddy
617 * we can do coalesce a page and its buddy if
618 * (a) the buddy is not in a hole &&
619 * (b) the buddy is in the buddy system &&
620 * (c) a page and its buddy have the same order &&
621 * (d) a page and its buddy are in the same zone.
623 * For recording whether a page is in the buddy system, we set ->_mapcount
624 * PAGE_BUDDY_MAPCOUNT_VALUE.
625 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
626 * serialized by zone->lock.
628 * For recording page's order, we use page_private(page).
630 static inline int page_is_buddy(struct page *page, struct page *buddy,
633 if (!pfn_valid_within(page_to_pfn(buddy)))
636 if (page_is_guard(buddy) && page_order(buddy) == order) {
637 if (page_zone_id(page) != page_zone_id(buddy))
640 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
645 if (PageBuddy(buddy) && page_order(buddy) == order) {
647 * zone check is done late to avoid uselessly
648 * calculating zone/node ids for pages that could
651 if (page_zone_id(page) != page_zone_id(buddy))
654 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
662 * Freeing function for a buddy system allocator.
664 * The concept of a buddy system is to maintain direct-mapped table
665 * (containing bit values) for memory blocks of various "orders".
666 * The bottom level table contains the map for the smallest allocatable
667 * units of memory (here, pages), and each level above it describes
668 * pairs of units from the levels below, hence, "buddies".
669 * At a high level, all that happens here is marking the table entry
670 * at the bottom level available, and propagating the changes upward
671 * as necessary, plus some accounting needed to play nicely with other
672 * parts of the VM system.
673 * At each level, we keep a list of pages, which are heads of continuous
674 * free pages of length of (1 << order) and marked with _mapcount
675 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
677 * So when we are allocating or freeing one, we can derive the state of the
678 * other. That is, if we allocate a small block, and both were
679 * free, the remainder of the region must be split into blocks.
680 * If a block is freed, and its buddy is also free, then this
681 * triggers coalescing into a block of larger size.
686 static inline void __free_one_page(struct page *page,
688 struct zone *zone, unsigned int order,
691 unsigned long page_idx;
692 unsigned long combined_idx;
693 unsigned long uninitialized_var(buddy_idx);
695 unsigned int max_order;
697 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
699 VM_BUG_ON(!zone_is_initialized(zone));
700 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
702 VM_BUG_ON(migratetype == -1);
703 if (likely(!is_migrate_isolate(migratetype)))
704 __mod_zone_freepage_state(zone, 1 << order, migratetype);
706 page_idx = pfn & ((1 << MAX_ORDER) - 1);
708 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
709 VM_BUG_ON_PAGE(bad_range(zone, page), page);
712 while (order < max_order - 1) {
713 buddy_idx = __find_buddy_index(page_idx, order);
714 buddy = page + (buddy_idx - page_idx);
715 if (!page_is_buddy(page, buddy, order))
718 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
719 * merge with it and move up one order.
721 if (page_is_guard(buddy)) {
722 clear_page_guard(zone, buddy, order, migratetype);
724 list_del(&buddy->lru);
725 zone->free_area[order].nr_free--;
726 rmv_page_order(buddy);
728 combined_idx = buddy_idx & page_idx;
729 page = page + (combined_idx - page_idx);
730 page_idx = combined_idx;
733 if (max_order < MAX_ORDER) {
734 /* If we are here, it means order is >= pageblock_order.
735 * We want to prevent merge between freepages on isolate
736 * pageblock and normal pageblock. Without this, pageblock
737 * isolation could cause incorrect freepage or CMA accounting.
739 * We don't want to hit this code for the more frequent
742 if (unlikely(has_isolate_pageblock(zone))) {
745 buddy_idx = __find_buddy_index(page_idx, order);
746 buddy = page + (buddy_idx - page_idx);
747 buddy_mt = get_pageblock_migratetype(buddy);
749 if (migratetype != buddy_mt
750 && (is_migrate_isolate(migratetype) ||
751 is_migrate_isolate(buddy_mt)))
755 goto continue_merging;
759 set_page_order(page, order);
762 * If this is not the largest possible page, check if the buddy
763 * of the next-highest order is free. If it is, it's possible
764 * that pages are being freed that will coalesce soon. In case,
765 * that is happening, add the free page to the tail of the list
766 * so it's less likely to be used soon and more likely to be merged
767 * as a higher order page
769 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
770 struct page *higher_page, *higher_buddy;
771 combined_idx = buddy_idx & page_idx;
772 higher_page = page + (combined_idx - page_idx);
773 buddy_idx = __find_buddy_index(combined_idx, order + 1);
774 higher_buddy = higher_page + (buddy_idx - combined_idx);
775 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
776 list_add_tail(&page->lru,
777 &zone->free_area[order].free_list[migratetype]);
782 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
784 zone->free_area[order].nr_free++;
787 static inline int free_pages_check(struct page *page)
789 const char *bad_reason = NULL;
790 unsigned long bad_flags = 0;
792 if (unlikely(atomic_read(&page->_mapcount) != -1))
793 bad_reason = "nonzero mapcount";
794 if (unlikely(page->mapping != NULL))
795 bad_reason = "non-NULL mapping";
796 if (unlikely(page_ref_count(page) != 0))
797 bad_reason = "nonzero _refcount";
798 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
799 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
800 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
803 if (unlikely(page->mem_cgroup))
804 bad_reason = "page still charged to cgroup";
806 if (unlikely(bad_reason)) {
807 bad_page(page, bad_reason, bad_flags);
810 page_cpupid_reset_last(page);
811 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
812 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
817 * Frees a number of pages from the PCP lists
818 * Assumes all pages on list are in same zone, and of same order.
819 * count is the number of pages to free.
821 * If the zone was previously in an "all pages pinned" state then look to
822 * see if this freeing clears that state.
824 * And clear the zone's pages_scanned counter, to hold off the "all pages are
825 * pinned" detection logic.
827 static void free_pcppages_bulk(struct zone *zone, int count,
828 struct per_cpu_pages *pcp)
833 unsigned long nr_scanned;
834 bool isolated_pageblocks;
836 spin_lock(&zone->lock);
837 isolated_pageblocks = has_isolate_pageblock(zone);
838 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
840 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
844 struct list_head *list;
847 * Remove pages from lists in a round-robin fashion. A
848 * batch_free count is maintained that is incremented when an
849 * empty list is encountered. This is so more pages are freed
850 * off fuller lists instead of spinning excessively around empty
855 if (++migratetype == MIGRATE_PCPTYPES)
857 list = &pcp->lists[migratetype];
858 } while (list_empty(list));
860 /* This is the only non-empty list. Free them all. */
861 if (batch_free == MIGRATE_PCPTYPES)
862 batch_free = to_free;
865 int mt; /* migratetype of the to-be-freed page */
867 page = list_last_entry(list, struct page, lru);
868 /* must delete as __free_one_page list manipulates */
869 list_del(&page->lru);
871 mt = get_pcppage_migratetype(page);
872 /* MIGRATE_ISOLATE page should not go to pcplists */
873 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
874 /* Pageblock could have been isolated meanwhile */
875 if (unlikely(isolated_pageblocks))
876 mt = get_pageblock_migratetype(page);
878 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
879 trace_mm_page_pcpu_drain(page, 0, mt);
880 } while (--to_free && --batch_free && !list_empty(list));
882 spin_unlock(&zone->lock);
885 static void free_one_page(struct zone *zone,
886 struct page *page, unsigned long pfn,
890 unsigned long nr_scanned;
891 spin_lock(&zone->lock);
892 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
894 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
896 if (unlikely(has_isolate_pageblock(zone) ||
897 is_migrate_isolate(migratetype))) {
898 migratetype = get_pfnblock_migratetype(page, pfn);
900 __free_one_page(page, pfn, zone, order, migratetype);
901 spin_unlock(&zone->lock);
904 static int free_tail_pages_check(struct page *head_page, struct page *page)
909 * We rely page->lru.next never has bit 0 set, unless the page
910 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
912 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
914 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
918 switch (page - head_page) {
920 /* the first tail page: ->mapping is compound_mapcount() */
921 if (unlikely(compound_mapcount(page))) {
922 bad_page(page, "nonzero compound_mapcount", 0);
928 * the second tail page: ->mapping is
929 * page_deferred_list().next -- ignore value.
933 if (page->mapping != TAIL_MAPPING) {
934 bad_page(page, "corrupted mapping in tail page", 0);
939 if (unlikely(!PageTail(page))) {
940 bad_page(page, "PageTail not set", 0);
943 if (unlikely(compound_head(page) != head_page)) {
944 bad_page(page, "compound_head not consistent", 0);
949 page->mapping = NULL;
950 clear_compound_head(page);
954 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
955 unsigned long zone, int nid)
957 set_page_links(page, zone, nid, pfn);
958 init_page_count(page);
959 page_mapcount_reset(page);
960 page_cpupid_reset_last(page);
962 INIT_LIST_HEAD(&page->lru);
963 #ifdef WANT_PAGE_VIRTUAL
964 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
965 if (!is_highmem_idx(zone))
966 set_page_address(page, __va(pfn << PAGE_SHIFT));
970 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
973 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
976 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
977 static void init_reserved_page(unsigned long pfn)
982 if (!early_page_uninitialised(pfn))
985 nid = early_pfn_to_nid(pfn);
986 pgdat = NODE_DATA(nid);
988 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
989 struct zone *zone = &pgdat->node_zones[zid];
991 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
994 __init_single_pfn(pfn, zid, nid);
997 static inline void init_reserved_page(unsigned long pfn)
1000 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1003 * Initialised pages do not have PageReserved set. This function is
1004 * called for each range allocated by the bootmem allocator and
1005 * marks the pages PageReserved. The remaining valid pages are later
1006 * sent to the buddy page allocator.
1008 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
1010 unsigned long start_pfn = PFN_DOWN(start);
1011 unsigned long end_pfn = PFN_UP(end);
1013 for (; start_pfn < end_pfn; start_pfn++) {
1014 if (pfn_valid(start_pfn)) {
1015 struct page *page = pfn_to_page(start_pfn);
1017 init_reserved_page(start_pfn);
1019 /* Avoid false-positive PageTail() */
1020 INIT_LIST_HEAD(&page->lru);
1022 SetPageReserved(page);
1027 static bool free_pages_prepare(struct page *page, unsigned int order)
1031 VM_BUG_ON_PAGE(PageTail(page), page);
1033 trace_mm_page_free(page, order);
1034 kmemcheck_free_shadow(page, order);
1035 kasan_free_pages(page, order);
1038 * Check tail pages before head page information is cleared to
1039 * avoid checking PageCompound for order-0 pages.
1041 if (unlikely(order)) {
1042 bool compound = PageCompound(page);
1045 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1047 for (i = 1; i < (1 << order); i++) {
1049 bad += free_tail_pages_check(page, page + i);
1050 bad += free_pages_check(page + i);
1053 if (PageAnonHead(page))
1054 page->mapping = NULL;
1055 bad += free_pages_check(page);
1059 reset_page_owner(page, order);
1061 if (!PageHighMem(page)) {
1062 debug_check_no_locks_freed(page_address(page),
1063 PAGE_SIZE << order);
1064 debug_check_no_obj_freed(page_address(page),
1065 PAGE_SIZE << order);
1067 arch_free_page(page, order);
1068 kernel_poison_pages(page, 1 << order, 0);
1069 kernel_map_pages(page, 1 << order, 0);
1074 static void __free_pages_ok(struct page *page, unsigned int order)
1076 unsigned long flags;
1078 unsigned long pfn = page_to_pfn(page);
1080 if (!free_pages_prepare(page, order))
1083 migratetype = get_pfnblock_migratetype(page, pfn);
1084 local_irq_save(flags);
1085 __count_vm_events(PGFREE, 1 << order);
1086 free_one_page(page_zone(page), page, pfn, order, migratetype);
1087 local_irq_restore(flags);
1090 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1092 unsigned int nr_pages = 1 << order;
1093 struct page *p = page;
1097 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1099 __ClearPageReserved(p);
1100 set_page_count(p, 0);
1102 __ClearPageReserved(p);
1103 set_page_count(p, 0);
1105 page_zone(page)->managed_pages += nr_pages;
1106 set_page_refcounted(page);
1107 __free_pages(page, order);
1110 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1111 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1113 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1115 int __meminit early_pfn_to_nid(unsigned long pfn)
1117 static DEFINE_SPINLOCK(early_pfn_lock);
1120 spin_lock(&early_pfn_lock);
1121 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1124 spin_unlock(&early_pfn_lock);
1130 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1131 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1132 struct mminit_pfnnid_cache *state)
1136 nid = __early_pfn_to_nid(pfn, state);
1137 if (nid >= 0 && nid != node)
1142 /* Only safe to use early in boot when initialisation is single-threaded */
1143 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1145 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1150 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1154 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1155 struct mminit_pfnnid_cache *state)
1162 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1165 if (early_page_uninitialised(pfn))
1167 return __free_pages_boot_core(page, order);
1171 * Check that the whole (or subset of) a pageblock given by the interval of
1172 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1173 * with the migration of free compaction scanner. The scanners then need to
1174 * use only pfn_valid_within() check for arches that allow holes within
1177 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1179 * It's possible on some configurations to have a setup like node0 node1 node0
1180 * i.e. it's possible that all pages within a zones range of pages do not
1181 * belong to a single zone. We assume that a border between node0 and node1
1182 * can occur within a single pageblock, but not a node0 node1 node0
1183 * interleaving within a single pageblock. It is therefore sufficient to check
1184 * the first and last page of a pageblock and avoid checking each individual
1185 * page in a pageblock.
1187 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1188 unsigned long end_pfn, struct zone *zone)
1190 struct page *start_page;
1191 struct page *end_page;
1193 /* end_pfn is one past the range we are checking */
1196 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1199 start_page = pfn_to_page(start_pfn);
1201 if (page_zone(start_page) != zone)
1204 end_page = pfn_to_page(end_pfn);
1206 /* This gives a shorter code than deriving page_zone(end_page) */
1207 if (page_zone_id(start_page) != page_zone_id(end_page))
1213 void set_zone_contiguous(struct zone *zone)
1215 unsigned long block_start_pfn = zone->zone_start_pfn;
1216 unsigned long block_end_pfn;
1218 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1219 for (; block_start_pfn < zone_end_pfn(zone);
1220 block_start_pfn = block_end_pfn,
1221 block_end_pfn += pageblock_nr_pages) {
1223 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1225 if (!__pageblock_pfn_to_page(block_start_pfn,
1226 block_end_pfn, zone))
1230 /* We confirm that there is no hole */
1231 zone->contiguous = true;
1234 void clear_zone_contiguous(struct zone *zone)
1236 zone->contiguous = false;
1239 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1240 static void __init deferred_free_range(struct page *page,
1241 unsigned long pfn, int nr_pages)
1248 /* Free a large naturally-aligned chunk if possible */
1249 if (nr_pages == MAX_ORDER_NR_PAGES &&
1250 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1251 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1252 __free_pages_boot_core(page, MAX_ORDER-1);
1256 for (i = 0; i < nr_pages; i++, page++)
1257 __free_pages_boot_core(page, 0);
1260 /* Completion tracking for deferred_init_memmap() threads */
1261 static atomic_t pgdat_init_n_undone __initdata;
1262 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1264 static inline void __init pgdat_init_report_one_done(void)
1266 if (atomic_dec_and_test(&pgdat_init_n_undone))
1267 complete(&pgdat_init_all_done_comp);
1270 /* Initialise remaining memory on a node */
1271 static int __init deferred_init_memmap(void *data)
1273 pg_data_t *pgdat = data;
1274 int nid = pgdat->node_id;
1275 struct mminit_pfnnid_cache nid_init_state = { };
1276 unsigned long start = jiffies;
1277 unsigned long nr_pages = 0;
1278 unsigned long walk_start, walk_end;
1281 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1282 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1284 if (first_init_pfn == ULONG_MAX) {
1285 pgdat_init_report_one_done();
1289 /* Bind memory initialisation thread to a local node if possible */
1290 if (!cpumask_empty(cpumask))
1291 set_cpus_allowed_ptr(current, cpumask);
1293 /* Sanity check boundaries */
1294 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1295 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1296 pgdat->first_deferred_pfn = ULONG_MAX;
1298 /* Only the highest zone is deferred so find it */
1299 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1300 zone = pgdat->node_zones + zid;
1301 if (first_init_pfn < zone_end_pfn(zone))
1305 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1306 unsigned long pfn, end_pfn;
1307 struct page *page = NULL;
1308 struct page *free_base_page = NULL;
1309 unsigned long free_base_pfn = 0;
1312 end_pfn = min(walk_end, zone_end_pfn(zone));
1313 pfn = first_init_pfn;
1314 if (pfn < walk_start)
1316 if (pfn < zone->zone_start_pfn)
1317 pfn = zone->zone_start_pfn;
1319 for (; pfn < end_pfn; pfn++) {
1320 if (!pfn_valid_within(pfn))
1324 * Ensure pfn_valid is checked every
1325 * MAX_ORDER_NR_PAGES for memory holes
1327 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1328 if (!pfn_valid(pfn)) {
1334 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1339 /* Minimise pfn page lookups and scheduler checks */
1340 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1343 nr_pages += nr_to_free;
1344 deferred_free_range(free_base_page,
1345 free_base_pfn, nr_to_free);
1346 free_base_page = NULL;
1347 free_base_pfn = nr_to_free = 0;
1349 page = pfn_to_page(pfn);
1354 VM_BUG_ON(page_zone(page) != zone);
1358 __init_single_page(page, pfn, zid, nid);
1359 if (!free_base_page) {
1360 free_base_page = page;
1361 free_base_pfn = pfn;
1366 /* Where possible, batch up pages for a single free */
1369 /* Free the current block of pages to allocator */
1370 nr_pages += nr_to_free;
1371 deferred_free_range(free_base_page, free_base_pfn,
1373 free_base_page = NULL;
1374 free_base_pfn = nr_to_free = 0;
1377 first_init_pfn = max(end_pfn, first_init_pfn);
1380 /* Sanity check that the next zone really is unpopulated */
1381 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1383 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1384 jiffies_to_msecs(jiffies - start));
1386 pgdat_init_report_one_done();
1389 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1391 void __init page_alloc_init_late(void)
1395 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1398 /* There will be num_node_state(N_MEMORY) threads */
1399 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1400 for_each_node_state(nid, N_MEMORY) {
1401 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1404 /* Block until all are initialised */
1405 wait_for_completion(&pgdat_init_all_done_comp);
1407 /* Reinit limits that are based on free pages after the kernel is up */
1408 files_maxfiles_init();
1411 for_each_populated_zone(zone)
1412 set_zone_contiguous(zone);
1416 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1417 void __init init_cma_reserved_pageblock(struct page *page)
1419 unsigned i = pageblock_nr_pages;
1420 struct page *p = page;
1423 __ClearPageReserved(p);
1424 set_page_count(p, 0);
1427 set_pageblock_migratetype(page, MIGRATE_CMA);
1429 if (pageblock_order >= MAX_ORDER) {
1430 i = pageblock_nr_pages;
1433 set_page_refcounted(p);
1434 __free_pages(p, MAX_ORDER - 1);
1435 p += MAX_ORDER_NR_PAGES;
1436 } while (i -= MAX_ORDER_NR_PAGES);
1438 set_page_refcounted(page);
1439 __free_pages(page, pageblock_order);
1442 adjust_managed_page_count(page, pageblock_nr_pages);
1447 * The order of subdivision here is critical for the IO subsystem.
1448 * Please do not alter this order without good reasons and regression
1449 * testing. Specifically, as large blocks of memory are subdivided,
1450 * the order in which smaller blocks are delivered depends on the order
1451 * they're subdivided in this function. This is the primary factor
1452 * influencing the order in which pages are delivered to the IO
1453 * subsystem according to empirical testing, and this is also justified
1454 * by considering the behavior of a buddy system containing a single
1455 * large block of memory acted on by a series of small allocations.
1456 * This behavior is a critical factor in sglist merging's success.
1460 static inline void expand(struct zone *zone, struct page *page,
1461 int low, int high, struct free_area *area,
1464 unsigned long size = 1 << high;
1466 while (high > low) {
1470 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1472 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1473 debug_guardpage_enabled() &&
1474 high < debug_guardpage_minorder()) {
1476 * Mark as guard pages (or page), that will allow to
1477 * merge back to allocator when buddy will be freed.
1478 * Corresponding page table entries will not be touched,
1479 * pages will stay not present in virtual address space
1481 set_page_guard(zone, &page[size], high, migratetype);
1484 list_add(&page[size].lru, &area->free_list[migratetype]);
1486 set_page_order(&page[size], high);
1491 * This page is about to be returned from the page allocator
1493 static inline int check_new_page(struct page *page)
1495 const char *bad_reason = NULL;
1496 unsigned long bad_flags = 0;
1498 if (unlikely(atomic_read(&page->_mapcount) != -1))
1499 bad_reason = "nonzero mapcount";
1500 if (unlikely(page->mapping != NULL))
1501 bad_reason = "non-NULL mapping";
1502 if (unlikely(page_ref_count(page) != 0))
1503 bad_reason = "nonzero _count";
1504 if (unlikely(page->flags & __PG_HWPOISON)) {
1505 bad_reason = "HWPoisoned (hardware-corrupted)";
1506 bad_flags = __PG_HWPOISON;
1508 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1509 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1510 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1513 if (unlikely(page->mem_cgroup))
1514 bad_reason = "page still charged to cgroup";
1516 if (unlikely(bad_reason)) {
1517 bad_page(page, bad_reason, bad_flags);
1523 static inline bool free_pages_prezeroed(bool poisoned)
1525 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1526 page_poisoning_enabled() && poisoned;
1529 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1530 unsigned int alloc_flags)
1533 bool poisoned = true;
1535 for (i = 0; i < (1 << order); i++) {
1536 struct page *p = page + i;
1537 if (unlikely(check_new_page(p)))
1540 poisoned &= page_is_poisoned(p);
1543 set_page_private(page, 0);
1544 set_page_refcounted(page);
1546 arch_alloc_page(page, order);
1547 kernel_map_pages(page, 1 << order, 1);
1548 kernel_poison_pages(page, 1 << order, 1);
1549 kasan_alloc_pages(page, order);
1551 if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO))
1552 for (i = 0; i < (1 << order); i++)
1553 clear_highpage(page + i);
1555 if (order && (gfp_flags & __GFP_COMP))
1556 prep_compound_page(page, order);
1558 set_page_owner(page, order, gfp_flags);
1561 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1562 * allocate the page. The expectation is that the caller is taking
1563 * steps that will free more memory. The caller should avoid the page
1564 * being used for !PFMEMALLOC purposes.
1566 if (alloc_flags & ALLOC_NO_WATERMARKS)
1567 set_page_pfmemalloc(page);
1569 clear_page_pfmemalloc(page);
1575 * Go through the free lists for the given migratetype and remove
1576 * the smallest available page from the freelists
1579 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1582 unsigned int current_order;
1583 struct free_area *area;
1586 /* Find a page of the appropriate size in the preferred list */
1587 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1588 area = &(zone->free_area[current_order]);
1589 page = list_first_entry_or_null(&area->free_list[migratetype],
1593 list_del(&page->lru);
1594 rmv_page_order(page);
1596 expand(zone, page, order, current_order, area, migratetype);
1597 set_pcppage_migratetype(page, migratetype);
1606 * This array describes the order lists are fallen back to when
1607 * the free lists for the desirable migrate type are depleted
1609 static int fallbacks[MIGRATE_TYPES][4] = {
1610 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1611 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1612 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1614 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1616 #ifdef CONFIG_MEMORY_ISOLATION
1617 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1622 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1625 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1628 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1629 unsigned int order) { return NULL; }
1633 * Move the free pages in a range to the free lists of the requested type.
1634 * Note that start_page and end_pages are not aligned on a pageblock
1635 * boundary. If alignment is required, use move_freepages_block()
1637 int move_freepages(struct zone *zone,
1638 struct page *start_page, struct page *end_page,
1643 int pages_moved = 0;
1645 #ifndef CONFIG_HOLES_IN_ZONE
1647 * page_zone is not safe to call in this context when
1648 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1649 * anyway as we check zone boundaries in move_freepages_block().
1650 * Remove at a later date when no bug reports exist related to
1651 * grouping pages by mobility
1653 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1656 for (page = start_page; page <= end_page;) {
1657 /* Make sure we are not inadvertently changing nodes */
1658 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1660 if (!pfn_valid_within(page_to_pfn(page))) {
1665 if (!PageBuddy(page)) {
1670 order = page_order(page);
1671 list_move(&page->lru,
1672 &zone->free_area[order].free_list[migratetype]);
1674 pages_moved += 1 << order;
1680 int move_freepages_block(struct zone *zone, struct page *page,
1683 unsigned long start_pfn, end_pfn;
1684 struct page *start_page, *end_page;
1686 start_pfn = page_to_pfn(page);
1687 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1688 start_page = pfn_to_page(start_pfn);
1689 end_page = start_page + pageblock_nr_pages - 1;
1690 end_pfn = start_pfn + pageblock_nr_pages - 1;
1692 /* Do not cross zone boundaries */
1693 if (!zone_spans_pfn(zone, start_pfn))
1695 if (!zone_spans_pfn(zone, end_pfn))
1698 return move_freepages(zone, start_page, end_page, migratetype);
1701 static void change_pageblock_range(struct page *pageblock_page,
1702 int start_order, int migratetype)
1704 int nr_pageblocks = 1 << (start_order - pageblock_order);
1706 while (nr_pageblocks--) {
1707 set_pageblock_migratetype(pageblock_page, migratetype);
1708 pageblock_page += pageblock_nr_pages;
1713 * When we are falling back to another migratetype during allocation, try to
1714 * steal extra free pages from the same pageblocks to satisfy further
1715 * allocations, instead of polluting multiple pageblocks.
1717 * If we are stealing a relatively large buddy page, it is likely there will
1718 * be more free pages in the pageblock, so try to steal them all. For
1719 * reclaimable and unmovable allocations, we steal regardless of page size,
1720 * as fragmentation caused by those allocations polluting movable pageblocks
1721 * is worse than movable allocations stealing from unmovable and reclaimable
1724 static bool can_steal_fallback(unsigned int order, int start_mt)
1727 * Leaving this order check is intended, although there is
1728 * relaxed order check in next check. The reason is that
1729 * we can actually steal whole pageblock if this condition met,
1730 * but, below check doesn't guarantee it and that is just heuristic
1731 * so could be changed anytime.
1733 if (order >= pageblock_order)
1736 if (order >= pageblock_order / 2 ||
1737 start_mt == MIGRATE_RECLAIMABLE ||
1738 start_mt == MIGRATE_UNMOVABLE ||
1739 page_group_by_mobility_disabled)
1746 * This function implements actual steal behaviour. If order is large enough,
1747 * we can steal whole pageblock. If not, we first move freepages in this
1748 * pageblock and check whether half of pages are moved or not. If half of
1749 * pages are moved, we can change migratetype of pageblock and permanently
1750 * use it's pages as requested migratetype in the future.
1752 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1755 unsigned int current_order = page_order(page);
1758 /* Take ownership for orders >= pageblock_order */
1759 if (current_order >= pageblock_order) {
1760 change_pageblock_range(page, current_order, start_type);
1764 pages = move_freepages_block(zone, page, start_type);
1766 /* Claim the whole block if over half of it is free */
1767 if (pages >= (1 << (pageblock_order-1)) ||
1768 page_group_by_mobility_disabled)
1769 set_pageblock_migratetype(page, start_type);
1773 * Check whether there is a suitable fallback freepage with requested order.
1774 * If only_stealable is true, this function returns fallback_mt only if
1775 * we can steal other freepages all together. This would help to reduce
1776 * fragmentation due to mixed migratetype pages in one pageblock.
1778 int find_suitable_fallback(struct free_area *area, unsigned int order,
1779 int migratetype, bool only_stealable, bool *can_steal)
1784 if (area->nr_free == 0)
1789 fallback_mt = fallbacks[migratetype][i];
1790 if (fallback_mt == MIGRATE_TYPES)
1793 if (list_empty(&area->free_list[fallback_mt]))
1796 if (can_steal_fallback(order, migratetype))
1799 if (!only_stealable)
1810 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1811 * there are no empty page blocks that contain a page with a suitable order
1813 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1814 unsigned int alloc_order)
1817 unsigned long max_managed, flags;
1820 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1821 * Check is race-prone but harmless.
1823 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1824 if (zone->nr_reserved_highatomic >= max_managed)
1827 spin_lock_irqsave(&zone->lock, flags);
1829 /* Recheck the nr_reserved_highatomic limit under the lock */
1830 if (zone->nr_reserved_highatomic >= max_managed)
1834 mt = get_pageblock_migratetype(page);
1835 if (mt != MIGRATE_HIGHATOMIC &&
1836 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1837 zone->nr_reserved_highatomic += pageblock_nr_pages;
1838 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1839 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1843 spin_unlock_irqrestore(&zone->lock, flags);
1847 * Used when an allocation is about to fail under memory pressure. This
1848 * potentially hurts the reliability of high-order allocations when under
1849 * intense memory pressure but failed atomic allocations should be easier
1850 * to recover from than an OOM.
1852 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1854 struct zonelist *zonelist = ac->zonelist;
1855 unsigned long flags;
1861 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1863 /* Preserve at least one pageblock */
1864 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1867 spin_lock_irqsave(&zone->lock, flags);
1868 for (order = 0; order < MAX_ORDER; order++) {
1869 struct free_area *area = &(zone->free_area[order]);
1871 page = list_first_entry_or_null(
1872 &area->free_list[MIGRATE_HIGHATOMIC],
1878 * It should never happen but changes to locking could
1879 * inadvertently allow a per-cpu drain to add pages
1880 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1881 * and watch for underflows.
1883 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1884 zone->nr_reserved_highatomic);
1887 * Convert to ac->migratetype and avoid the normal
1888 * pageblock stealing heuristics. Minimally, the caller
1889 * is doing the work and needs the pages. More
1890 * importantly, if the block was always converted to
1891 * MIGRATE_UNMOVABLE or another type then the number
1892 * of pageblocks that cannot be completely freed
1895 set_pageblock_migratetype(page, ac->migratetype);
1896 move_freepages_block(zone, page, ac->migratetype);
1897 spin_unlock_irqrestore(&zone->lock, flags);
1900 spin_unlock_irqrestore(&zone->lock, flags);
1904 /* Remove an element from the buddy allocator from the fallback list */
1905 static inline struct page *
1906 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1908 struct free_area *area;
1909 unsigned int current_order;
1914 /* Find the largest possible block of pages in the other list */
1915 for (current_order = MAX_ORDER-1;
1916 current_order >= order && current_order <= MAX_ORDER-1;
1918 area = &(zone->free_area[current_order]);
1919 fallback_mt = find_suitable_fallback(area, current_order,
1920 start_migratetype, false, &can_steal);
1921 if (fallback_mt == -1)
1924 page = list_first_entry(&area->free_list[fallback_mt],
1927 steal_suitable_fallback(zone, page, start_migratetype);
1929 /* Remove the page from the freelists */
1931 list_del(&page->lru);
1932 rmv_page_order(page);
1934 expand(zone, page, order, current_order, area,
1937 * The pcppage_migratetype may differ from pageblock's
1938 * migratetype depending on the decisions in
1939 * find_suitable_fallback(). This is OK as long as it does not
1940 * differ for MIGRATE_CMA pageblocks. Those can be used as
1941 * fallback only via special __rmqueue_cma_fallback() function
1943 set_pcppage_migratetype(page, start_migratetype);
1945 trace_mm_page_alloc_extfrag(page, order, current_order,
1946 start_migratetype, fallback_mt);
1955 * Do the hard work of removing an element from the buddy allocator.
1956 * Call me with the zone->lock already held.
1958 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1963 page = __rmqueue_smallest(zone, order, migratetype);
1964 if (unlikely(!page)) {
1965 if (migratetype == MIGRATE_MOVABLE)
1966 page = __rmqueue_cma_fallback(zone, order);
1969 page = __rmqueue_fallback(zone, order, migratetype);
1972 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1977 * Obtain a specified number of elements from the buddy allocator, all under
1978 * a single hold of the lock, for efficiency. Add them to the supplied list.
1979 * Returns the number of new pages which were placed at *list.
1981 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1982 unsigned long count, struct list_head *list,
1983 int migratetype, bool cold)
1987 spin_lock(&zone->lock);
1988 for (i = 0; i < count; ++i) {
1989 struct page *page = __rmqueue(zone, order, migratetype);
1990 if (unlikely(page == NULL))
1994 * Split buddy pages returned by expand() are received here
1995 * in physical page order. The page is added to the callers and
1996 * list and the list head then moves forward. From the callers
1997 * perspective, the linked list is ordered by page number in
1998 * some conditions. This is useful for IO devices that can
1999 * merge IO requests if the physical pages are ordered
2003 list_add(&page->lru, list);
2005 list_add_tail(&page->lru, list);
2007 if (is_migrate_cma(get_pcppage_migratetype(page)))
2008 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2011 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2012 spin_unlock(&zone->lock);
2018 * Called from the vmstat counter updater to drain pagesets of this
2019 * currently executing processor on remote nodes after they have
2022 * Note that this function must be called with the thread pinned to
2023 * a single processor.
2025 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2027 unsigned long flags;
2028 int to_drain, batch;
2030 local_irq_save(flags);
2031 batch = READ_ONCE(pcp->batch);
2032 to_drain = min(pcp->count, batch);
2034 free_pcppages_bulk(zone, to_drain, pcp);
2035 pcp->count -= to_drain;
2037 local_irq_restore(flags);
2042 * Drain pcplists of the indicated processor and zone.
2044 * The processor must either be the current processor and the
2045 * thread pinned to the current processor or a processor that
2048 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2050 unsigned long flags;
2051 struct per_cpu_pageset *pset;
2052 struct per_cpu_pages *pcp;
2054 local_irq_save(flags);
2055 pset = per_cpu_ptr(zone->pageset, cpu);
2059 free_pcppages_bulk(zone, pcp->count, pcp);
2062 local_irq_restore(flags);
2066 * Drain pcplists of all zones on the indicated processor.
2068 * The processor must either be the current processor and the
2069 * thread pinned to the current processor or a processor that
2072 static void drain_pages(unsigned int cpu)
2076 for_each_populated_zone(zone) {
2077 drain_pages_zone(cpu, zone);
2082 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2084 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2085 * the single zone's pages.
2087 void drain_local_pages(struct zone *zone)
2089 int cpu = smp_processor_id();
2092 drain_pages_zone(cpu, zone);
2098 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2100 * When zone parameter is non-NULL, spill just the single zone's pages.
2102 * Note that this code is protected against sending an IPI to an offline
2103 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2104 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2105 * nothing keeps CPUs from showing up after we populated the cpumask and
2106 * before the call to on_each_cpu_mask().
2108 void drain_all_pages(struct zone *zone)
2113 * Allocate in the BSS so we wont require allocation in
2114 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2116 static cpumask_t cpus_with_pcps;
2119 * We don't care about racing with CPU hotplug event
2120 * as offline notification will cause the notified
2121 * cpu to drain that CPU pcps and on_each_cpu_mask
2122 * disables preemption as part of its processing
2124 for_each_online_cpu(cpu) {
2125 struct per_cpu_pageset *pcp;
2127 bool has_pcps = false;
2130 pcp = per_cpu_ptr(zone->pageset, cpu);
2134 for_each_populated_zone(z) {
2135 pcp = per_cpu_ptr(z->pageset, cpu);
2136 if (pcp->pcp.count) {
2144 cpumask_set_cpu(cpu, &cpus_with_pcps);
2146 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2148 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2152 #ifdef CONFIG_HIBERNATION
2154 void mark_free_pages(struct zone *zone)
2156 unsigned long pfn, max_zone_pfn;
2157 unsigned long flags;
2158 unsigned int order, t;
2161 if (zone_is_empty(zone))
2164 spin_lock_irqsave(&zone->lock, flags);
2166 max_zone_pfn = zone_end_pfn(zone);
2167 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2168 if (pfn_valid(pfn)) {
2169 page = pfn_to_page(pfn);
2171 if (page_zone(page) != zone)
2174 if (!swsusp_page_is_forbidden(page))
2175 swsusp_unset_page_free(page);
2178 for_each_migratetype_order(order, t) {
2179 list_for_each_entry(page,
2180 &zone->free_area[order].free_list[t], lru) {
2183 pfn = page_to_pfn(page);
2184 for (i = 0; i < (1UL << order); i++)
2185 swsusp_set_page_free(pfn_to_page(pfn + i));
2188 spin_unlock_irqrestore(&zone->lock, flags);
2190 #endif /* CONFIG_PM */
2193 * Free a 0-order page
2194 * cold == true ? free a cold page : free a hot page
2196 void free_hot_cold_page(struct page *page, bool cold)
2198 struct zone *zone = page_zone(page);
2199 struct per_cpu_pages *pcp;
2200 unsigned long flags;
2201 unsigned long pfn = page_to_pfn(page);
2204 if (!free_pages_prepare(page, 0))
2207 migratetype = get_pfnblock_migratetype(page, pfn);
2208 set_pcppage_migratetype(page, migratetype);
2209 local_irq_save(flags);
2210 __count_vm_event(PGFREE);
2213 * We only track unmovable, reclaimable and movable on pcp lists.
2214 * Free ISOLATE pages back to the allocator because they are being
2215 * offlined but treat RESERVE as movable pages so we can get those
2216 * areas back if necessary. Otherwise, we may have to free
2217 * excessively into the page allocator
2219 if (migratetype >= MIGRATE_PCPTYPES) {
2220 if (unlikely(is_migrate_isolate(migratetype))) {
2221 free_one_page(zone, page, pfn, 0, migratetype);
2224 migratetype = MIGRATE_MOVABLE;
2227 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2229 list_add(&page->lru, &pcp->lists[migratetype]);
2231 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2233 if (pcp->count >= pcp->high) {
2234 unsigned long batch = READ_ONCE(pcp->batch);
2235 free_pcppages_bulk(zone, batch, pcp);
2236 pcp->count -= batch;
2240 local_irq_restore(flags);
2244 * Free a list of 0-order pages
2246 void free_hot_cold_page_list(struct list_head *list, bool cold)
2248 struct page *page, *next;
2250 list_for_each_entry_safe(page, next, list, lru) {
2251 trace_mm_page_free_batched(page, cold);
2252 free_hot_cold_page(page, cold);
2257 * split_page takes a non-compound higher-order page, and splits it into
2258 * n (1<<order) sub-pages: page[0..n]
2259 * Each sub-page must be freed individually.
2261 * Note: this is probably too low level an operation for use in drivers.
2262 * Please consult with lkml before using this in your driver.
2264 void split_page(struct page *page, unsigned int order)
2269 VM_BUG_ON_PAGE(PageCompound(page), page);
2270 VM_BUG_ON_PAGE(!page_count(page), page);
2272 #ifdef CONFIG_KMEMCHECK
2274 * Split shadow pages too, because free(page[0]) would
2275 * otherwise free the whole shadow.
2277 if (kmemcheck_page_is_tracked(page))
2278 split_page(virt_to_page(page[0].shadow), order);
2281 gfp_mask = get_page_owner_gfp(page);
2282 set_page_owner(page, 0, gfp_mask);
2283 for (i = 1; i < (1 << order); i++) {
2284 set_page_refcounted(page + i);
2285 set_page_owner(page + i, 0, gfp_mask);
2288 EXPORT_SYMBOL_GPL(split_page);
2290 int __isolate_free_page(struct page *page, unsigned int order)
2292 unsigned long watermark;
2296 BUG_ON(!PageBuddy(page));
2298 zone = page_zone(page);
2299 mt = get_pageblock_migratetype(page);
2301 if (!is_migrate_isolate(mt)) {
2302 /* Obey watermarks as if the page was being allocated */
2303 watermark = low_wmark_pages(zone) + (1 << order);
2304 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2307 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2310 /* Remove page from free list */
2311 list_del(&page->lru);
2312 zone->free_area[order].nr_free--;
2313 rmv_page_order(page);
2315 set_page_owner(page, order, __GFP_MOVABLE);
2317 /* Set the pageblock if the isolated page is at least a pageblock */
2318 if (order >= pageblock_order - 1) {
2319 struct page *endpage = page + (1 << order) - 1;
2320 for (; page < endpage; page += pageblock_nr_pages) {
2321 int mt = get_pageblock_migratetype(page);
2322 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2323 set_pageblock_migratetype(page,
2329 return 1UL << order;
2333 * Similar to split_page except the page is already free. As this is only
2334 * being used for migration, the migratetype of the block also changes.
2335 * As this is called with interrupts disabled, the caller is responsible
2336 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2339 * Note: this is probably too low level an operation for use in drivers.
2340 * Please consult with lkml before using this in your driver.
2342 int split_free_page(struct page *page)
2347 order = page_order(page);
2349 nr_pages = __isolate_free_page(page, order);
2353 /* Split into individual pages */
2354 set_page_refcounted(page);
2355 split_page(page, order);
2360 * Update NUMA hit/miss statistics
2362 * Must be called with interrupts disabled.
2364 * When __GFP_OTHER_NODE is set assume the node of the preferred
2365 * zone is the local node. This is useful for daemons who allocate
2366 * memory on behalf of other processes.
2368 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2372 int local_nid = numa_node_id();
2373 enum zone_stat_item local_stat = NUMA_LOCAL;
2375 if (unlikely(flags & __GFP_OTHER_NODE)) {
2376 local_stat = NUMA_OTHER;
2377 local_nid = preferred_zone->node;
2380 if (z->node == local_nid) {
2381 __inc_zone_state(z, NUMA_HIT);
2382 __inc_zone_state(z, local_stat);
2384 __inc_zone_state(z, NUMA_MISS);
2385 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
2391 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2394 struct page *buffered_rmqueue(struct zone *preferred_zone,
2395 struct zone *zone, unsigned int order,
2396 gfp_t gfp_flags, unsigned int alloc_flags,
2399 unsigned long flags;
2401 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2403 if (likely(order == 0)) {
2404 struct per_cpu_pages *pcp;
2405 struct list_head *list;
2407 local_irq_save(flags);
2408 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2409 list = &pcp->lists[migratetype];
2410 if (list_empty(list)) {
2411 pcp->count += rmqueue_bulk(zone, 0,
2414 if (unlikely(list_empty(list)))
2419 page = list_last_entry(list, struct page, lru);
2421 page = list_first_entry(list, struct page, lru);
2423 __dec_zone_state(zone, NR_ALLOC_BATCH);
2424 list_del(&page->lru);
2428 * We most definitely don't want callers attempting to
2429 * allocate greater than order-1 page units with __GFP_NOFAIL.
2431 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2432 spin_lock_irqsave(&zone->lock, flags);
2435 if (alloc_flags & ALLOC_HARDER) {
2436 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2438 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2441 page = __rmqueue(zone, order, migratetype);
2442 spin_unlock(&zone->lock);
2445 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2446 __mod_zone_freepage_state(zone, -(1 << order),
2447 get_pcppage_migratetype(page));
2450 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2451 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2452 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2454 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2455 zone_statistics(preferred_zone, zone, gfp_flags);
2456 local_irq_restore(flags);
2458 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2462 local_irq_restore(flags);
2466 #ifdef CONFIG_FAIL_PAGE_ALLOC
2469 struct fault_attr attr;
2471 bool ignore_gfp_highmem;
2472 bool ignore_gfp_reclaim;
2474 } fail_page_alloc = {
2475 .attr = FAULT_ATTR_INITIALIZER,
2476 .ignore_gfp_reclaim = true,
2477 .ignore_gfp_highmem = true,
2481 static int __init setup_fail_page_alloc(char *str)
2483 return setup_fault_attr(&fail_page_alloc.attr, str);
2485 __setup("fail_page_alloc=", setup_fail_page_alloc);
2487 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2489 if (order < fail_page_alloc.min_order)
2491 if (gfp_mask & __GFP_NOFAIL)
2493 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2495 if (fail_page_alloc.ignore_gfp_reclaim &&
2496 (gfp_mask & __GFP_DIRECT_RECLAIM))
2499 return should_fail(&fail_page_alloc.attr, 1 << order);
2502 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2504 static int __init fail_page_alloc_debugfs(void)
2506 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2509 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2510 &fail_page_alloc.attr);
2512 return PTR_ERR(dir);
2514 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2515 &fail_page_alloc.ignore_gfp_reclaim))
2517 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2518 &fail_page_alloc.ignore_gfp_highmem))
2520 if (!debugfs_create_u32("min-order", mode, dir,
2521 &fail_page_alloc.min_order))
2526 debugfs_remove_recursive(dir);
2531 late_initcall(fail_page_alloc_debugfs);
2533 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2535 #else /* CONFIG_FAIL_PAGE_ALLOC */
2537 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2542 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2545 * Return true if free base pages are above 'mark'. For high-order checks it
2546 * will return true of the order-0 watermark is reached and there is at least
2547 * one free page of a suitable size. Checking now avoids taking the zone lock
2548 * to check in the allocation paths if no pages are free.
2550 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2551 unsigned long mark, int classzone_idx,
2552 unsigned int alloc_flags,
2557 const bool alloc_harder = (alloc_flags & ALLOC_HARDER);
2559 /* free_pages may go negative - that's OK */
2560 free_pages -= (1 << order) - 1;
2562 if (alloc_flags & ALLOC_HIGH)
2566 * If the caller does not have rights to ALLOC_HARDER then subtract
2567 * the high-atomic reserves. This will over-estimate the size of the
2568 * atomic reserve but it avoids a search.
2570 if (likely(!alloc_harder))
2571 free_pages -= z->nr_reserved_highatomic;
2576 /* If allocation can't use CMA areas don't use free CMA pages */
2577 if (!(alloc_flags & ALLOC_CMA))
2578 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2582 * Check watermarks for an order-0 allocation request. If these
2583 * are not met, then a high-order request also cannot go ahead
2584 * even if a suitable page happened to be free.
2586 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2589 /* If this is an order-0 request then the watermark is fine */
2593 /* For a high-order request, check at least one suitable page is free */
2594 for (o = order; o < MAX_ORDER; o++) {
2595 struct free_area *area = &z->free_area[o];
2604 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2605 if (!list_empty(&area->free_list[mt]))
2610 if ((alloc_flags & ALLOC_CMA) &&
2611 !list_empty(&area->free_list[MIGRATE_CMA])) {
2619 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2620 int classzone_idx, unsigned int alloc_flags)
2622 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2623 zone_page_state(z, NR_FREE_PAGES));
2626 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
2627 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
2629 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2633 /* If allocation can't use CMA areas don't use free CMA pages */
2634 if (!(alloc_flags & ALLOC_CMA))
2635 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
2639 * Fast check for order-0 only. If this fails then the reserves
2640 * need to be calculated. There is a corner case where the check
2641 * passes but only the high-order atomic reserve are free. If
2642 * the caller is !atomic then it'll uselessly search the free
2643 * list. That corner case is then slower but it is harmless.
2645 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
2648 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2652 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2653 unsigned long mark, int classzone_idx)
2655 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2657 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2658 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2660 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2665 static bool zone_local(struct zone *local_zone, struct zone *zone)
2667 return local_zone->node == zone->node;
2670 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2672 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2675 #else /* CONFIG_NUMA */
2676 static bool zone_local(struct zone *local_zone, struct zone *zone)
2681 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2685 #endif /* CONFIG_NUMA */
2687 static void reset_alloc_batches(struct zone *preferred_zone)
2689 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2692 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2693 high_wmark_pages(zone) - low_wmark_pages(zone) -
2694 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2695 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2696 } while (zone++ != preferred_zone);
2700 * get_page_from_freelist goes through the zonelist trying to allocate
2703 static struct page *
2704 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2705 const struct alloc_context *ac)
2707 struct zoneref *z = ac->preferred_zoneref;
2709 bool fair_skipped = false;
2710 bool apply_fair = (alloc_flags & ALLOC_FAIR);
2714 * Scan zonelist, looking for a zone with enough free.
2715 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2717 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2722 if (cpusets_enabled() &&
2723 (alloc_flags & ALLOC_CPUSET) &&
2724 !cpuset_zone_allowed(zone, gfp_mask))
2727 * Distribute pages in proportion to the individual
2728 * zone size to ensure fair page aging. The zone a
2729 * page was allocated in should have no effect on the
2730 * time the page has in memory before being reclaimed.
2733 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2734 fair_skipped = true;
2737 if (!zone_local(ac->preferred_zoneref->zone, zone)) {
2744 * When allocating a page cache page for writing, we
2745 * want to get it from a zone that is within its dirty
2746 * limit, such that no single zone holds more than its
2747 * proportional share of globally allowed dirty pages.
2748 * The dirty limits take into account the zone's
2749 * lowmem reserves and high watermark so that kswapd
2750 * should be able to balance it without having to
2751 * write pages from its LRU list.
2753 * This may look like it could increase pressure on
2754 * lower zones by failing allocations in higher zones
2755 * before they are full. But the pages that do spill
2756 * over are limited as the lower zones are protected
2757 * by this very same mechanism. It should not become
2758 * a practical burden to them.
2760 * XXX: For now, allow allocations to potentially
2761 * exceed the per-zone dirty limit in the slowpath
2762 * (spread_dirty_pages unset) before going into reclaim,
2763 * which is important when on a NUMA setup the allowed
2764 * zones are together not big enough to reach the
2765 * global limit. The proper fix for these situations
2766 * will require awareness of zones in the
2767 * dirty-throttling and the flusher threads.
2769 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2772 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2773 if (!zone_watermark_fast(zone, order, mark,
2774 ac_classzone_idx(ac), alloc_flags)) {
2777 /* Checked here to keep the fast path fast */
2778 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2779 if (alloc_flags & ALLOC_NO_WATERMARKS)
2782 if (zone_reclaim_mode == 0 ||
2783 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
2786 ret = zone_reclaim(zone, gfp_mask, order);
2788 case ZONE_RECLAIM_NOSCAN:
2791 case ZONE_RECLAIM_FULL:
2792 /* scanned but unreclaimable */
2795 /* did we reclaim enough */
2796 if (zone_watermark_ok(zone, order, mark,
2797 ac_classzone_idx(ac), alloc_flags))
2805 page = buffered_rmqueue(ac->preferred_zoneref->zone, zone, order,
2806 gfp_mask, alloc_flags, ac->migratetype);
2808 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2812 * If this is a high-order atomic allocation then check
2813 * if the pageblock should be reserved for the future
2815 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2816 reserve_highatomic_pageblock(page, zone, order);
2823 * The first pass makes sure allocations are spread fairly within the
2824 * local node. However, the local node might have free pages left
2825 * after the fairness batches are exhausted, and remote zones haven't
2826 * even been considered yet. Try once more without fairness, and
2827 * include remote zones now, before entering the slowpath and waking
2828 * kswapd: prefer spilling to a remote zone over swapping locally.
2833 fair_skipped = false;
2834 reset_alloc_batches(ac->preferred_zoneref->zone);
2842 * Large machines with many possible nodes should not always dump per-node
2843 * meminfo in irq context.
2845 static inline bool should_suppress_show_mem(void)
2850 ret = in_interrupt();
2855 static DEFINE_RATELIMIT_STATE(nopage_rs,
2856 DEFAULT_RATELIMIT_INTERVAL,
2857 DEFAULT_RATELIMIT_BURST);
2859 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2861 unsigned int filter = SHOW_MEM_FILTER_NODES;
2863 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2864 debug_guardpage_minorder() > 0)
2868 * This documents exceptions given to allocations in certain
2869 * contexts that are allowed to allocate outside current's set
2872 if (!(gfp_mask & __GFP_NOMEMALLOC))
2873 if (test_thread_flag(TIF_MEMDIE) ||
2874 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2875 filter &= ~SHOW_MEM_FILTER_NODES;
2876 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2877 filter &= ~SHOW_MEM_FILTER_NODES;
2880 struct va_format vaf;
2883 va_start(args, fmt);
2888 pr_warn("%pV", &vaf);
2893 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
2894 current->comm, order, gfp_mask, &gfp_mask);
2896 if (!should_suppress_show_mem())
2900 static inline struct page *
2901 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2902 const struct alloc_context *ac, unsigned long *did_some_progress)
2904 struct oom_control oc = {
2905 .zonelist = ac->zonelist,
2906 .nodemask = ac->nodemask,
2907 .gfp_mask = gfp_mask,
2912 *did_some_progress = 0;
2915 * Acquire the oom lock. If that fails, somebody else is
2916 * making progress for us.
2918 if (!mutex_trylock(&oom_lock)) {
2919 *did_some_progress = 1;
2920 schedule_timeout_uninterruptible(1);
2925 * Go through the zonelist yet one more time, keep very high watermark
2926 * here, this is only to catch a parallel oom killing, we must fail if
2927 * we're still under heavy pressure.
2929 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2930 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2934 if (!(gfp_mask & __GFP_NOFAIL)) {
2935 /* Coredumps can quickly deplete all memory reserves */
2936 if (current->flags & PF_DUMPCORE)
2938 /* The OOM killer will not help higher order allocs */
2939 if (order > PAGE_ALLOC_COSTLY_ORDER)
2941 /* The OOM killer does not needlessly kill tasks for lowmem */
2942 if (ac->high_zoneidx < ZONE_NORMAL)
2944 if (pm_suspended_storage())
2947 * XXX: GFP_NOFS allocations should rather fail than rely on
2948 * other request to make a forward progress.
2949 * We are in an unfortunate situation where out_of_memory cannot
2950 * do much for this context but let's try it to at least get
2951 * access to memory reserved if the current task is killed (see
2952 * out_of_memory). Once filesystems are ready to handle allocation
2953 * failures more gracefully we should just bail out here.
2956 /* The OOM killer may not free memory on a specific node */
2957 if (gfp_mask & __GFP_THISNODE)
2960 /* Exhausted what can be done so it's blamo time */
2961 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
2962 *did_some_progress = 1;
2964 if (gfp_mask & __GFP_NOFAIL) {
2965 page = get_page_from_freelist(gfp_mask, order,
2966 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
2968 * fallback to ignore cpuset restriction if our nodes
2972 page = get_page_from_freelist(gfp_mask, order,
2973 ALLOC_NO_WATERMARKS, ac);
2977 mutex_unlock(&oom_lock);
2981 #ifdef CONFIG_COMPACTION
2982 /* Try memory compaction for high-order allocations before reclaim */
2983 static struct page *
2984 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2985 unsigned int alloc_flags, const struct alloc_context *ac,
2986 enum migrate_mode mode, int *contended_compaction,
2987 bool *deferred_compaction)
2989 unsigned long compact_result;
2995 current->flags |= PF_MEMALLOC;
2996 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2997 mode, contended_compaction);
2998 current->flags &= ~PF_MEMALLOC;
3000 switch (compact_result) {
3001 case COMPACT_DEFERRED:
3002 *deferred_compaction = true;
3004 case COMPACT_SKIPPED:
3011 * At least in one zone compaction wasn't deferred or skipped, so let's
3012 * count a compaction stall
3014 count_vm_event(COMPACTSTALL);
3016 page = get_page_from_freelist(gfp_mask, order,
3017 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3020 struct zone *zone = page_zone(page);
3022 zone->compact_blockskip_flush = false;
3023 compaction_defer_reset(zone, order, true);
3024 count_vm_event(COMPACTSUCCESS);
3029 * It's bad if compaction run occurs and fails. The most likely reason
3030 * is that pages exist, but not enough to satisfy watermarks.
3032 count_vm_event(COMPACTFAIL);
3039 static inline struct page *
3040 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3041 unsigned int alloc_flags, const struct alloc_context *ac,
3042 enum migrate_mode mode, int *contended_compaction,
3043 bool *deferred_compaction)
3047 #endif /* CONFIG_COMPACTION */
3049 /* Perform direct synchronous page reclaim */
3051 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3052 const struct alloc_context *ac)
3054 struct reclaim_state reclaim_state;
3059 /* We now go into synchronous reclaim */
3060 cpuset_memory_pressure_bump();
3061 current->flags |= PF_MEMALLOC;
3062 lockdep_set_current_reclaim_state(gfp_mask);
3063 reclaim_state.reclaimed_slab = 0;
3064 current->reclaim_state = &reclaim_state;
3066 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3069 current->reclaim_state = NULL;
3070 lockdep_clear_current_reclaim_state();
3071 current->flags &= ~PF_MEMALLOC;
3078 /* The really slow allocator path where we enter direct reclaim */
3079 static inline struct page *
3080 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3081 unsigned int alloc_flags, const struct alloc_context *ac,
3082 unsigned long *did_some_progress)
3084 struct page *page = NULL;
3085 bool drained = false;
3087 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3088 if (unlikely(!(*did_some_progress)))
3092 page = get_page_from_freelist(gfp_mask, order,
3093 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3096 * If an allocation failed after direct reclaim, it could be because
3097 * pages are pinned on the per-cpu lists or in high alloc reserves.
3098 * Shrink them them and try again
3100 if (!page && !drained) {
3101 unreserve_highatomic_pageblock(ac);
3102 drain_all_pages(NULL);
3110 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3115 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3116 ac->high_zoneidx, ac->nodemask)
3117 wakeup_kswapd(zone, order, ac_classzone_idx(ac));
3120 static inline unsigned int
3121 gfp_to_alloc_flags(gfp_t gfp_mask)
3123 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3125 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3126 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3129 * The caller may dip into page reserves a bit more if the caller
3130 * cannot run direct reclaim, or if the caller has realtime scheduling
3131 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3132 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3134 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3136 if (gfp_mask & __GFP_ATOMIC) {
3138 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3139 * if it can't schedule.
3141 if (!(gfp_mask & __GFP_NOMEMALLOC))
3142 alloc_flags |= ALLOC_HARDER;
3144 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3145 * comment for __cpuset_node_allowed().
3147 alloc_flags &= ~ALLOC_CPUSET;
3148 } else if (unlikely(rt_task(current)) && !in_interrupt())
3149 alloc_flags |= ALLOC_HARDER;
3151 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
3152 if (gfp_mask & __GFP_MEMALLOC)
3153 alloc_flags |= ALLOC_NO_WATERMARKS;
3154 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3155 alloc_flags |= ALLOC_NO_WATERMARKS;
3156 else if (!in_interrupt() &&
3157 ((current->flags & PF_MEMALLOC) ||
3158 unlikely(test_thread_flag(TIF_MEMDIE))))
3159 alloc_flags |= ALLOC_NO_WATERMARKS;
3162 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3163 alloc_flags |= ALLOC_CMA;
3168 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3170 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
3173 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
3175 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
3178 static inline struct page *
3179 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3180 struct alloc_context *ac)
3182 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3183 struct page *page = NULL;
3184 unsigned int alloc_flags;
3185 unsigned long pages_reclaimed = 0;
3186 unsigned long did_some_progress;
3187 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3188 bool deferred_compaction = false;
3189 int contended_compaction = COMPACT_CONTENDED_NONE;
3192 * In the slowpath, we sanity check order to avoid ever trying to
3193 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3194 * be using allocators in order of preference for an area that is
3197 if (order >= MAX_ORDER) {
3198 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3203 * We also sanity check to catch abuse of atomic reserves being used by
3204 * callers that are not in atomic context.
3206 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3207 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3208 gfp_mask &= ~__GFP_ATOMIC;
3211 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3212 wake_all_kswapds(order, ac);
3215 * OK, we're below the kswapd watermark and have kicked background
3216 * reclaim. Now things get more complex, so set up alloc_flags according
3217 * to how we want to proceed.
3219 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3221 /* This is the last chance, in general, before the goto nopage. */
3222 page = get_page_from_freelist(gfp_mask, order,
3223 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3227 /* Allocate without watermarks if the context allows */
3228 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3230 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3231 * the allocation is high priority and these type of
3232 * allocations are system rather than user orientated
3234 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3235 page = get_page_from_freelist(gfp_mask, order,
3236 ALLOC_NO_WATERMARKS, ac);
3241 /* Caller is not willing to reclaim, we can't balance anything */
3242 if (!can_direct_reclaim) {
3244 * All existing users of the __GFP_NOFAIL are blockable, so warn
3245 * of any new users that actually allow this type of allocation
3248 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3252 /* Avoid recursion of direct reclaim */
3253 if (current->flags & PF_MEMALLOC) {
3255 * __GFP_NOFAIL request from this context is rather bizarre
3256 * because we cannot reclaim anything and only can loop waiting
3257 * for somebody to do a work for us.
3259 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3266 /* Avoid allocations with no watermarks from looping endlessly */
3267 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3271 * Try direct compaction. The first pass is asynchronous. Subsequent
3272 * attempts after direct reclaim are synchronous
3274 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3276 &contended_compaction,
3277 &deferred_compaction);
3281 /* Checks for THP-specific high-order allocations */
3282 if (is_thp_gfp_mask(gfp_mask)) {
3284 * If compaction is deferred for high-order allocations, it is
3285 * because sync compaction recently failed. If this is the case
3286 * and the caller requested a THP allocation, we do not want
3287 * to heavily disrupt the system, so we fail the allocation
3288 * instead of entering direct reclaim.
3290 if (deferred_compaction)
3294 * In all zones where compaction was attempted (and not
3295 * deferred or skipped), lock contention has been detected.
3296 * For THP allocation we do not want to disrupt the others
3297 * so we fallback to base pages instead.
3299 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3303 * If compaction was aborted due to need_resched(), we do not
3304 * want to further increase allocation latency, unless it is
3305 * khugepaged trying to collapse.
3307 if (contended_compaction == COMPACT_CONTENDED_SCHED
3308 && !(current->flags & PF_KTHREAD))
3313 * It can become very expensive to allocate transparent hugepages at
3314 * fault, so use asynchronous memory compaction for THP unless it is
3315 * khugepaged trying to collapse.
3317 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3318 migration_mode = MIGRATE_SYNC_LIGHT;
3320 /* Try direct reclaim and then allocating */
3321 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3322 &did_some_progress);
3326 /* Do not loop if specifically requested */
3327 if (gfp_mask & __GFP_NORETRY)
3330 /* Keep reclaiming pages as long as there is reasonable progress */
3331 pages_reclaimed += did_some_progress;
3332 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3333 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3334 /* Wait for some write requests to complete then retry */
3335 wait_iff_congested(ac->preferred_zoneref->zone, BLK_RW_ASYNC, HZ/50);
3339 /* Reclaim has failed us, start killing things */
3340 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3344 /* Retry as long as the OOM killer is making progress */
3345 if (did_some_progress)
3350 * High-order allocations do not necessarily loop after
3351 * direct reclaim and reclaim/compaction depends on compaction
3352 * being called after reclaim so call directly if necessary
3354 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3356 &contended_compaction,
3357 &deferred_compaction);
3361 warn_alloc_failed(gfp_mask, order, NULL);
3367 * This is the 'heart' of the zoned buddy allocator.
3370 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3371 struct zonelist *zonelist, nodemask_t *nodemask)
3374 unsigned int cpuset_mems_cookie;
3375 unsigned int alloc_flags = ALLOC_WMARK_LOW|ALLOC_FAIR;
3376 gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */
3377 struct alloc_context ac = {
3378 .high_zoneidx = gfp_zone(gfp_mask),
3379 .zonelist = zonelist,
3380 .nodemask = nodemask,
3381 .migratetype = gfpflags_to_migratetype(gfp_mask),
3384 if (cpusets_enabled()) {
3385 alloc_mask |= __GFP_HARDWALL;
3386 alloc_flags |= ALLOC_CPUSET;
3388 ac.nodemask = &cpuset_current_mems_allowed;
3391 gfp_mask &= gfp_allowed_mask;
3393 lockdep_trace_alloc(gfp_mask);
3395 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3397 if (should_fail_alloc_page(gfp_mask, order))
3401 * Check the zones suitable for the gfp_mask contain at least one
3402 * valid zone. It's possible to have an empty zonelist as a result
3403 * of __GFP_THISNODE and a memoryless node
3405 if (unlikely(!zonelist->_zonerefs->zone))
3408 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3409 alloc_flags |= ALLOC_CMA;
3412 cpuset_mems_cookie = read_mems_allowed_begin();
3414 /* Dirty zone balancing only done in the fast path */
3415 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3417 /* The preferred zone is used for statistics later */
3418 ac.preferred_zoneref = first_zones_zonelist(ac.zonelist,
3419 ac.high_zoneidx, ac.nodemask);
3420 if (!ac.preferred_zoneref) {
3425 /* First allocation attempt */
3426 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3431 * Runtime PM, block IO and its error handling path can deadlock
3432 * because I/O on the device might not complete.
3434 alloc_mask = memalloc_noio_flags(gfp_mask);
3435 ac.spread_dirty_pages = false;
3437 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3441 * When updating a task's mems_allowed, it is possible to race with
3442 * parallel threads in such a way that an allocation can fail while
3443 * the mask is being updated. If a page allocation is about to fail,
3444 * check if the cpuset changed during allocation and if so, retry.
3446 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) {
3447 alloc_mask = gfp_mask;
3452 if (kmemcheck_enabled && page)
3453 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3455 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3459 EXPORT_SYMBOL(__alloc_pages_nodemask);
3462 * Common helper functions.
3464 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3469 * __get_free_pages() returns a 32-bit address, which cannot represent
3472 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3474 page = alloc_pages(gfp_mask, order);
3477 return (unsigned long) page_address(page);
3479 EXPORT_SYMBOL(__get_free_pages);
3481 unsigned long get_zeroed_page(gfp_t gfp_mask)
3483 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3485 EXPORT_SYMBOL(get_zeroed_page);
3487 void __free_pages(struct page *page, unsigned int order)
3489 if (put_page_testzero(page)) {
3491 free_hot_cold_page(page, false);
3493 __free_pages_ok(page, order);
3497 EXPORT_SYMBOL(__free_pages);
3499 void free_pages(unsigned long addr, unsigned int order)
3502 VM_BUG_ON(!virt_addr_valid((void *)addr));
3503 __free_pages(virt_to_page((void *)addr), order);
3507 EXPORT_SYMBOL(free_pages);
3511 * An arbitrary-length arbitrary-offset area of memory which resides
3512 * within a 0 or higher order page. Multiple fragments within that page
3513 * are individually refcounted, in the page's reference counter.
3515 * The page_frag functions below provide a simple allocation framework for
3516 * page fragments. This is used by the network stack and network device
3517 * drivers to provide a backing region of memory for use as either an
3518 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3520 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3523 struct page *page = NULL;
3524 gfp_t gfp = gfp_mask;
3526 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3527 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3529 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3530 PAGE_FRAG_CACHE_MAX_ORDER);
3531 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3533 if (unlikely(!page))
3534 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3536 nc->va = page ? page_address(page) : NULL;
3541 void *__alloc_page_frag(struct page_frag_cache *nc,
3542 unsigned int fragsz, gfp_t gfp_mask)
3544 unsigned int size = PAGE_SIZE;
3548 if (unlikely(!nc->va)) {
3550 page = __page_frag_refill(nc, gfp_mask);
3554 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3555 /* if size can vary use size else just use PAGE_SIZE */
3558 /* Even if we own the page, we do not use atomic_set().
3559 * This would break get_page_unless_zero() users.
3561 page_ref_add(page, size - 1);
3563 /* reset page count bias and offset to start of new frag */
3564 nc->pfmemalloc = page_is_pfmemalloc(page);
3565 nc->pagecnt_bias = size;
3569 offset = nc->offset - fragsz;
3570 if (unlikely(offset < 0)) {
3571 page = virt_to_page(nc->va);
3573 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
3576 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3577 /* if size can vary use size else just use PAGE_SIZE */
3580 /* OK, page count is 0, we can safely set it */
3581 set_page_count(page, size);
3583 /* reset page count bias and offset to start of new frag */
3584 nc->pagecnt_bias = size;
3585 offset = size - fragsz;
3589 nc->offset = offset;
3591 return nc->va + offset;
3593 EXPORT_SYMBOL(__alloc_page_frag);
3596 * Frees a page fragment allocated out of either a compound or order 0 page.
3598 void __free_page_frag(void *addr)
3600 struct page *page = virt_to_head_page(addr);
3602 if (unlikely(put_page_testzero(page)))
3603 __free_pages_ok(page, compound_order(page));
3605 EXPORT_SYMBOL(__free_page_frag);
3608 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3609 * of the current memory cgroup if __GFP_ACCOUNT is set, other than that it is
3610 * equivalent to alloc_pages.
3612 * It should be used when the caller would like to use kmalloc, but since the
3613 * allocation is large, it has to fall back to the page allocator.
3615 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3619 page = alloc_pages(gfp_mask, order);
3620 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3621 __free_pages(page, order);
3627 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3631 page = alloc_pages_node(nid, gfp_mask, order);
3632 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3633 __free_pages(page, order);
3640 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3643 void __free_kmem_pages(struct page *page, unsigned int order)
3645 memcg_kmem_uncharge(page, order);
3646 __free_pages(page, order);
3649 void free_kmem_pages(unsigned long addr, unsigned int order)
3652 VM_BUG_ON(!virt_addr_valid((void *)addr));
3653 __free_kmem_pages(virt_to_page((void *)addr), order);
3657 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3661 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3662 unsigned long used = addr + PAGE_ALIGN(size);
3664 split_page(virt_to_page((void *)addr), order);
3665 while (used < alloc_end) {
3670 return (void *)addr;
3674 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3675 * @size: the number of bytes to allocate
3676 * @gfp_mask: GFP flags for the allocation
3678 * This function is similar to alloc_pages(), except that it allocates the
3679 * minimum number of pages to satisfy the request. alloc_pages() can only
3680 * allocate memory in power-of-two pages.
3682 * This function is also limited by MAX_ORDER.
3684 * Memory allocated by this function must be released by free_pages_exact().
3686 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3688 unsigned int order = get_order(size);
3691 addr = __get_free_pages(gfp_mask, order);
3692 return make_alloc_exact(addr, order, size);
3694 EXPORT_SYMBOL(alloc_pages_exact);
3697 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3699 * @nid: the preferred node ID where memory should be allocated
3700 * @size: the number of bytes to allocate
3701 * @gfp_mask: GFP flags for the allocation
3703 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3706 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3708 unsigned int order = get_order(size);
3709 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3712 return make_alloc_exact((unsigned long)page_address(p), order, size);
3716 * free_pages_exact - release memory allocated via alloc_pages_exact()
3717 * @virt: the value returned by alloc_pages_exact.
3718 * @size: size of allocation, same value as passed to alloc_pages_exact().
3720 * Release the memory allocated by a previous call to alloc_pages_exact.
3722 void free_pages_exact(void *virt, size_t size)
3724 unsigned long addr = (unsigned long)virt;
3725 unsigned long end = addr + PAGE_ALIGN(size);
3727 while (addr < end) {
3732 EXPORT_SYMBOL(free_pages_exact);
3735 * nr_free_zone_pages - count number of pages beyond high watermark
3736 * @offset: The zone index of the highest zone
3738 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3739 * high watermark within all zones at or below a given zone index. For each
3740 * zone, the number of pages is calculated as:
3741 * managed_pages - high_pages
3743 static unsigned long nr_free_zone_pages(int offset)
3748 /* Just pick one node, since fallback list is circular */
3749 unsigned long sum = 0;
3751 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3753 for_each_zone_zonelist(zone, z, zonelist, offset) {
3754 unsigned long size = zone->managed_pages;
3755 unsigned long high = high_wmark_pages(zone);
3764 * nr_free_buffer_pages - count number of pages beyond high watermark
3766 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3767 * watermark within ZONE_DMA and ZONE_NORMAL.
3769 unsigned long nr_free_buffer_pages(void)
3771 return nr_free_zone_pages(gfp_zone(GFP_USER));
3773 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3776 * nr_free_pagecache_pages - count number of pages beyond high watermark
3778 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3779 * high watermark within all zones.
3781 unsigned long nr_free_pagecache_pages(void)
3783 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3786 static inline void show_node(struct zone *zone)
3788 if (IS_ENABLED(CONFIG_NUMA))
3789 printk("Node %d ", zone_to_nid(zone));
3792 long si_mem_available(void)
3795 unsigned long pagecache;
3796 unsigned long wmark_low = 0;
3797 unsigned long pages[NR_LRU_LISTS];
3801 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
3802 pages[lru] = global_page_state(NR_LRU_BASE + lru);
3805 wmark_low += zone->watermark[WMARK_LOW];
3808 * Estimate the amount of memory available for userspace allocations,
3809 * without causing swapping.
3811 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
3814 * Not all the page cache can be freed, otherwise the system will
3815 * start swapping. Assume at least half of the page cache, or the
3816 * low watermark worth of cache, needs to stay.
3818 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
3819 pagecache -= min(pagecache / 2, wmark_low);
3820 available += pagecache;
3823 * Part of the reclaimable slab consists of items that are in use,
3824 * and cannot be freed. Cap this estimate at the low watermark.
3826 available += global_page_state(NR_SLAB_RECLAIMABLE) -
3827 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
3833 EXPORT_SYMBOL_GPL(si_mem_available);
3835 void si_meminfo(struct sysinfo *val)
3837 val->totalram = totalram_pages;
3838 val->sharedram = global_page_state(NR_SHMEM);
3839 val->freeram = global_page_state(NR_FREE_PAGES);
3840 val->bufferram = nr_blockdev_pages();
3841 val->totalhigh = totalhigh_pages;
3842 val->freehigh = nr_free_highpages();
3843 val->mem_unit = PAGE_SIZE;
3846 EXPORT_SYMBOL(si_meminfo);
3849 void si_meminfo_node(struct sysinfo *val, int nid)
3851 int zone_type; /* needs to be signed */
3852 unsigned long managed_pages = 0;
3853 unsigned long managed_highpages = 0;
3854 unsigned long free_highpages = 0;
3855 pg_data_t *pgdat = NODE_DATA(nid);
3857 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3858 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3859 val->totalram = managed_pages;
3860 val->sharedram = node_page_state(nid, NR_SHMEM);
3861 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3862 #ifdef CONFIG_HIGHMEM
3863 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3864 struct zone *zone = &pgdat->node_zones[zone_type];
3866 if (is_highmem(zone)) {
3867 managed_highpages += zone->managed_pages;
3868 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
3871 val->totalhigh = managed_highpages;
3872 val->freehigh = free_highpages;
3874 val->totalhigh = managed_highpages;
3875 val->freehigh = free_highpages;
3877 val->mem_unit = PAGE_SIZE;
3882 * Determine whether the node should be displayed or not, depending on whether
3883 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3885 bool skip_free_areas_node(unsigned int flags, int nid)
3888 unsigned int cpuset_mems_cookie;
3890 if (!(flags & SHOW_MEM_FILTER_NODES))
3894 cpuset_mems_cookie = read_mems_allowed_begin();
3895 ret = !node_isset(nid, cpuset_current_mems_allowed);
3896 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3901 #define K(x) ((x) << (PAGE_SHIFT-10))
3903 static void show_migration_types(unsigned char type)
3905 static const char types[MIGRATE_TYPES] = {
3906 [MIGRATE_UNMOVABLE] = 'U',
3907 [MIGRATE_MOVABLE] = 'M',
3908 [MIGRATE_RECLAIMABLE] = 'E',
3909 [MIGRATE_HIGHATOMIC] = 'H',
3911 [MIGRATE_CMA] = 'C',
3913 #ifdef CONFIG_MEMORY_ISOLATION
3914 [MIGRATE_ISOLATE] = 'I',
3917 char tmp[MIGRATE_TYPES + 1];
3921 for (i = 0; i < MIGRATE_TYPES; i++) {
3922 if (type & (1 << i))
3927 printk("(%s) ", tmp);
3931 * Show free area list (used inside shift_scroll-lock stuff)
3932 * We also calculate the percentage fragmentation. We do this by counting the
3933 * memory on each free list with the exception of the first item on the list.
3936 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3939 void show_free_areas(unsigned int filter)
3941 unsigned long free_pcp = 0;
3945 for_each_populated_zone(zone) {
3946 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3949 for_each_online_cpu(cpu)
3950 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3953 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3954 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3955 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3956 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3957 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3958 " free:%lu free_pcp:%lu free_cma:%lu\n",
3959 global_page_state(NR_ACTIVE_ANON),
3960 global_page_state(NR_INACTIVE_ANON),
3961 global_page_state(NR_ISOLATED_ANON),
3962 global_page_state(NR_ACTIVE_FILE),
3963 global_page_state(NR_INACTIVE_FILE),
3964 global_page_state(NR_ISOLATED_FILE),
3965 global_page_state(NR_UNEVICTABLE),
3966 global_page_state(NR_FILE_DIRTY),
3967 global_page_state(NR_WRITEBACK),
3968 global_page_state(NR_UNSTABLE_NFS),
3969 global_page_state(NR_SLAB_RECLAIMABLE),
3970 global_page_state(NR_SLAB_UNRECLAIMABLE),
3971 global_page_state(NR_FILE_MAPPED),
3972 global_page_state(NR_SHMEM),
3973 global_page_state(NR_PAGETABLE),
3974 global_page_state(NR_BOUNCE),
3975 global_page_state(NR_FREE_PAGES),
3977 global_page_state(NR_FREE_CMA_PAGES));
3979 for_each_populated_zone(zone) {
3982 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3986 for_each_online_cpu(cpu)
3987 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3995 " active_anon:%lukB"
3996 " inactive_anon:%lukB"
3997 " active_file:%lukB"
3998 " inactive_file:%lukB"
3999 " unevictable:%lukB"
4000 " isolated(anon):%lukB"
4001 " isolated(file):%lukB"
4009 " slab_reclaimable:%lukB"
4010 " slab_unreclaimable:%lukB"
4011 " kernel_stack:%lukB"
4018 " writeback_tmp:%lukB"
4019 " pages_scanned:%lu"
4020 " all_unreclaimable? %s"
4023 K(zone_page_state(zone, NR_FREE_PAGES)),
4024 K(min_wmark_pages(zone)),
4025 K(low_wmark_pages(zone)),
4026 K(high_wmark_pages(zone)),
4027 K(zone_page_state(zone, NR_ACTIVE_ANON)),
4028 K(zone_page_state(zone, NR_INACTIVE_ANON)),
4029 K(zone_page_state(zone, NR_ACTIVE_FILE)),
4030 K(zone_page_state(zone, NR_INACTIVE_FILE)),
4031 K(zone_page_state(zone, NR_UNEVICTABLE)),
4032 K(zone_page_state(zone, NR_ISOLATED_ANON)),
4033 K(zone_page_state(zone, NR_ISOLATED_FILE)),
4034 K(zone->present_pages),
4035 K(zone->managed_pages),
4036 K(zone_page_state(zone, NR_MLOCK)),
4037 K(zone_page_state(zone, NR_FILE_DIRTY)),
4038 K(zone_page_state(zone, NR_WRITEBACK)),
4039 K(zone_page_state(zone, NR_FILE_MAPPED)),
4040 K(zone_page_state(zone, NR_SHMEM)),
4041 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
4042 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
4043 zone_page_state(zone, NR_KERNEL_STACK) *
4045 K(zone_page_state(zone, NR_PAGETABLE)),
4046 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
4047 K(zone_page_state(zone, NR_BOUNCE)),
4049 K(this_cpu_read(zone->pageset->pcp.count)),
4050 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
4051 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
4052 K(zone_page_state(zone, NR_PAGES_SCANNED)),
4053 (!zone_reclaimable(zone) ? "yes" : "no")
4055 printk("lowmem_reserve[]:");
4056 for (i = 0; i < MAX_NR_ZONES; i++)
4057 printk(" %ld", zone->lowmem_reserve[i]);
4061 for_each_populated_zone(zone) {
4063 unsigned long nr[MAX_ORDER], flags, total = 0;
4064 unsigned char types[MAX_ORDER];
4066 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4069 printk("%s: ", zone->name);
4071 spin_lock_irqsave(&zone->lock, flags);
4072 for (order = 0; order < MAX_ORDER; order++) {
4073 struct free_area *area = &zone->free_area[order];
4076 nr[order] = area->nr_free;
4077 total += nr[order] << order;
4080 for (type = 0; type < MIGRATE_TYPES; type++) {
4081 if (!list_empty(&area->free_list[type]))
4082 types[order] |= 1 << type;
4085 spin_unlock_irqrestore(&zone->lock, flags);
4086 for (order = 0; order < MAX_ORDER; order++) {
4087 printk("%lu*%lukB ", nr[order], K(1UL) << order);
4089 show_migration_types(types[order]);
4091 printk("= %lukB\n", K(total));
4094 hugetlb_show_meminfo();
4096 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
4098 show_swap_cache_info();
4101 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4103 zoneref->zone = zone;
4104 zoneref->zone_idx = zone_idx(zone);
4108 * Builds allocation fallback zone lists.
4110 * Add all populated zones of a node to the zonelist.
4112 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4116 enum zone_type zone_type = MAX_NR_ZONES;
4120 zone = pgdat->node_zones + zone_type;
4121 if (populated_zone(zone)) {
4122 zoneref_set_zone(zone,
4123 &zonelist->_zonerefs[nr_zones++]);
4124 check_highest_zone(zone_type);
4126 } while (zone_type);
4134 * 0 = automatic detection of better ordering.
4135 * 1 = order by ([node] distance, -zonetype)
4136 * 2 = order by (-zonetype, [node] distance)
4138 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4139 * the same zonelist. So only NUMA can configure this param.
4141 #define ZONELIST_ORDER_DEFAULT 0
4142 #define ZONELIST_ORDER_NODE 1
4143 #define ZONELIST_ORDER_ZONE 2
4145 /* zonelist order in the kernel.
4146 * set_zonelist_order() will set this to NODE or ZONE.
4148 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4149 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4153 /* The value user specified ....changed by config */
4154 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4155 /* string for sysctl */
4156 #define NUMA_ZONELIST_ORDER_LEN 16
4157 char numa_zonelist_order[16] = "default";
4160 * interface for configure zonelist ordering.
4161 * command line option "numa_zonelist_order"
4162 * = "[dD]efault - default, automatic configuration.
4163 * = "[nN]ode - order by node locality, then by zone within node
4164 * = "[zZ]one - order by zone, then by locality within zone
4167 static int __parse_numa_zonelist_order(char *s)
4169 if (*s == 'd' || *s == 'D') {
4170 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4171 } else if (*s == 'n' || *s == 'N') {
4172 user_zonelist_order = ZONELIST_ORDER_NODE;
4173 } else if (*s == 'z' || *s == 'Z') {
4174 user_zonelist_order = ZONELIST_ORDER_ZONE;
4176 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4182 static __init int setup_numa_zonelist_order(char *s)
4189 ret = __parse_numa_zonelist_order(s);
4191 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4195 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4198 * sysctl handler for numa_zonelist_order
4200 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4201 void __user *buffer, size_t *length,
4204 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4206 static DEFINE_MUTEX(zl_order_mutex);
4208 mutex_lock(&zl_order_mutex);
4210 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4214 strcpy(saved_string, (char *)table->data);
4216 ret = proc_dostring(table, write, buffer, length, ppos);
4220 int oldval = user_zonelist_order;
4222 ret = __parse_numa_zonelist_order((char *)table->data);
4225 * bogus value. restore saved string
4227 strncpy((char *)table->data, saved_string,
4228 NUMA_ZONELIST_ORDER_LEN);
4229 user_zonelist_order = oldval;
4230 } else if (oldval != user_zonelist_order) {
4231 mutex_lock(&zonelists_mutex);
4232 build_all_zonelists(NULL, NULL);
4233 mutex_unlock(&zonelists_mutex);
4237 mutex_unlock(&zl_order_mutex);
4242 #define MAX_NODE_LOAD (nr_online_nodes)
4243 static int node_load[MAX_NUMNODES];
4246 * find_next_best_node - find the next node that should appear in a given node's fallback list
4247 * @node: node whose fallback list we're appending
4248 * @used_node_mask: nodemask_t of already used nodes
4250 * We use a number of factors to determine which is the next node that should
4251 * appear on a given node's fallback list. The node should not have appeared
4252 * already in @node's fallback list, and it should be the next closest node
4253 * according to the distance array (which contains arbitrary distance values
4254 * from each node to each node in the system), and should also prefer nodes
4255 * with no CPUs, since presumably they'll have very little allocation pressure
4256 * on them otherwise.
4257 * It returns -1 if no node is found.
4259 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4262 int min_val = INT_MAX;
4263 int best_node = NUMA_NO_NODE;
4264 const struct cpumask *tmp = cpumask_of_node(0);
4266 /* Use the local node if we haven't already */
4267 if (!node_isset(node, *used_node_mask)) {
4268 node_set(node, *used_node_mask);
4272 for_each_node_state(n, N_MEMORY) {
4274 /* Don't want a node to appear more than once */
4275 if (node_isset(n, *used_node_mask))
4278 /* Use the distance array to find the distance */
4279 val = node_distance(node, n);
4281 /* Penalize nodes under us ("prefer the next node") */
4284 /* Give preference to headless and unused nodes */
4285 tmp = cpumask_of_node(n);
4286 if (!cpumask_empty(tmp))
4287 val += PENALTY_FOR_NODE_WITH_CPUS;
4289 /* Slight preference for less loaded node */
4290 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4291 val += node_load[n];
4293 if (val < min_val) {
4300 node_set(best_node, *used_node_mask);
4307 * Build zonelists ordered by node and zones within node.
4308 * This results in maximum locality--normal zone overflows into local
4309 * DMA zone, if any--but risks exhausting DMA zone.
4311 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4314 struct zonelist *zonelist;
4316 zonelist = &pgdat->node_zonelists[0];
4317 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4319 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4320 zonelist->_zonerefs[j].zone = NULL;
4321 zonelist->_zonerefs[j].zone_idx = 0;
4325 * Build gfp_thisnode zonelists
4327 static void build_thisnode_zonelists(pg_data_t *pgdat)
4330 struct zonelist *zonelist;
4332 zonelist = &pgdat->node_zonelists[1];
4333 j = build_zonelists_node(pgdat, zonelist, 0);
4334 zonelist->_zonerefs[j].zone = NULL;
4335 zonelist->_zonerefs[j].zone_idx = 0;
4339 * Build zonelists ordered by zone and nodes within zones.
4340 * This results in conserving DMA zone[s] until all Normal memory is
4341 * exhausted, but results in overflowing to remote node while memory
4342 * may still exist in local DMA zone.
4344 static int node_order[MAX_NUMNODES];
4346 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4349 int zone_type; /* needs to be signed */
4351 struct zonelist *zonelist;
4353 zonelist = &pgdat->node_zonelists[0];
4355 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4356 for (j = 0; j < nr_nodes; j++) {
4357 node = node_order[j];
4358 z = &NODE_DATA(node)->node_zones[zone_type];
4359 if (populated_zone(z)) {
4361 &zonelist->_zonerefs[pos++]);
4362 check_highest_zone(zone_type);
4366 zonelist->_zonerefs[pos].zone = NULL;
4367 zonelist->_zonerefs[pos].zone_idx = 0;
4370 #if defined(CONFIG_64BIT)
4372 * Devices that require DMA32/DMA are relatively rare and do not justify a
4373 * penalty to every machine in case the specialised case applies. Default
4374 * to Node-ordering on 64-bit NUMA machines
4376 static int default_zonelist_order(void)
4378 return ZONELIST_ORDER_NODE;
4382 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4383 * by the kernel. If processes running on node 0 deplete the low memory zone
4384 * then reclaim will occur more frequency increasing stalls and potentially
4385 * be easier to OOM if a large percentage of the zone is under writeback or
4386 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4387 * Hence, default to zone ordering on 32-bit.
4389 static int default_zonelist_order(void)
4391 return ZONELIST_ORDER_ZONE;
4393 #endif /* CONFIG_64BIT */
4395 static void set_zonelist_order(void)
4397 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4398 current_zonelist_order = default_zonelist_order();
4400 current_zonelist_order = user_zonelist_order;
4403 static void build_zonelists(pg_data_t *pgdat)
4406 nodemask_t used_mask;
4407 int local_node, prev_node;
4408 struct zonelist *zonelist;
4409 unsigned int order = current_zonelist_order;
4411 /* initialize zonelists */
4412 for (i = 0; i < MAX_ZONELISTS; i++) {
4413 zonelist = pgdat->node_zonelists + i;
4414 zonelist->_zonerefs[0].zone = NULL;
4415 zonelist->_zonerefs[0].zone_idx = 0;
4418 /* NUMA-aware ordering of nodes */
4419 local_node = pgdat->node_id;
4420 load = nr_online_nodes;
4421 prev_node = local_node;
4422 nodes_clear(used_mask);
4424 memset(node_order, 0, sizeof(node_order));
4427 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4429 * We don't want to pressure a particular node.
4430 * So adding penalty to the first node in same
4431 * distance group to make it round-robin.
4433 if (node_distance(local_node, node) !=
4434 node_distance(local_node, prev_node))
4435 node_load[node] = load;
4439 if (order == ZONELIST_ORDER_NODE)
4440 build_zonelists_in_node_order(pgdat, node);
4442 node_order[i++] = node; /* remember order */
4445 if (order == ZONELIST_ORDER_ZONE) {
4446 /* calculate node order -- i.e., DMA last! */
4447 build_zonelists_in_zone_order(pgdat, i);
4450 build_thisnode_zonelists(pgdat);
4453 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4455 * Return node id of node used for "local" allocations.
4456 * I.e., first node id of first zone in arg node's generic zonelist.
4457 * Used for initializing percpu 'numa_mem', which is used primarily
4458 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4460 int local_memory_node(int node)
4464 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4465 gfp_zone(GFP_KERNEL),
4467 return z->zone->node;
4471 #else /* CONFIG_NUMA */
4473 static void set_zonelist_order(void)
4475 current_zonelist_order = ZONELIST_ORDER_ZONE;
4478 static void build_zonelists(pg_data_t *pgdat)
4480 int node, local_node;
4482 struct zonelist *zonelist;
4484 local_node = pgdat->node_id;
4486 zonelist = &pgdat->node_zonelists[0];
4487 j = build_zonelists_node(pgdat, zonelist, 0);
4490 * Now we build the zonelist so that it contains the zones
4491 * of all the other nodes.
4492 * We don't want to pressure a particular node, so when
4493 * building the zones for node N, we make sure that the
4494 * zones coming right after the local ones are those from
4495 * node N+1 (modulo N)
4497 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4498 if (!node_online(node))
4500 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4502 for (node = 0; node < local_node; node++) {
4503 if (!node_online(node))
4505 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4508 zonelist->_zonerefs[j].zone = NULL;
4509 zonelist->_zonerefs[j].zone_idx = 0;
4512 #endif /* CONFIG_NUMA */
4515 * Boot pageset table. One per cpu which is going to be used for all
4516 * zones and all nodes. The parameters will be set in such a way
4517 * that an item put on a list will immediately be handed over to
4518 * the buddy list. This is safe since pageset manipulation is done
4519 * with interrupts disabled.
4521 * The boot_pagesets must be kept even after bootup is complete for
4522 * unused processors and/or zones. They do play a role for bootstrapping
4523 * hotplugged processors.
4525 * zoneinfo_show() and maybe other functions do
4526 * not check if the processor is online before following the pageset pointer.
4527 * Other parts of the kernel may not check if the zone is available.
4529 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4530 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4531 static void setup_zone_pageset(struct zone *zone);
4534 * Global mutex to protect against size modification of zonelists
4535 * as well as to serialize pageset setup for the new populated zone.
4537 DEFINE_MUTEX(zonelists_mutex);
4539 /* return values int ....just for stop_machine() */
4540 static int __build_all_zonelists(void *data)
4544 pg_data_t *self = data;
4547 memset(node_load, 0, sizeof(node_load));
4550 if (self && !node_online(self->node_id)) {
4551 build_zonelists(self);
4554 for_each_online_node(nid) {
4555 pg_data_t *pgdat = NODE_DATA(nid);
4557 build_zonelists(pgdat);
4561 * Initialize the boot_pagesets that are going to be used
4562 * for bootstrapping processors. The real pagesets for
4563 * each zone will be allocated later when the per cpu
4564 * allocator is available.
4566 * boot_pagesets are used also for bootstrapping offline
4567 * cpus if the system is already booted because the pagesets
4568 * are needed to initialize allocators on a specific cpu too.
4569 * F.e. the percpu allocator needs the page allocator which
4570 * needs the percpu allocator in order to allocate its pagesets
4571 * (a chicken-egg dilemma).
4573 for_each_possible_cpu(cpu) {
4574 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4576 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4578 * We now know the "local memory node" for each node--
4579 * i.e., the node of the first zone in the generic zonelist.
4580 * Set up numa_mem percpu variable for on-line cpus. During
4581 * boot, only the boot cpu should be on-line; we'll init the
4582 * secondary cpus' numa_mem as they come on-line. During
4583 * node/memory hotplug, we'll fixup all on-line cpus.
4585 if (cpu_online(cpu))
4586 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4593 static noinline void __init
4594 build_all_zonelists_init(void)
4596 __build_all_zonelists(NULL);
4597 mminit_verify_zonelist();
4598 cpuset_init_current_mems_allowed();
4602 * Called with zonelists_mutex held always
4603 * unless system_state == SYSTEM_BOOTING.
4605 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4606 * [we're only called with non-NULL zone through __meminit paths] and
4607 * (2) call of __init annotated helper build_all_zonelists_init
4608 * [protected by SYSTEM_BOOTING].
4610 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4612 set_zonelist_order();
4614 if (system_state == SYSTEM_BOOTING) {
4615 build_all_zonelists_init();
4617 #ifdef CONFIG_MEMORY_HOTPLUG
4619 setup_zone_pageset(zone);
4621 /* we have to stop all cpus to guarantee there is no user
4623 stop_machine(__build_all_zonelists, pgdat, NULL);
4624 /* cpuset refresh routine should be here */
4626 vm_total_pages = nr_free_pagecache_pages();
4628 * Disable grouping by mobility if the number of pages in the
4629 * system is too low to allow the mechanism to work. It would be
4630 * more accurate, but expensive to check per-zone. This check is
4631 * made on memory-hotadd so a system can start with mobility
4632 * disabled and enable it later
4634 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4635 page_group_by_mobility_disabled = 1;
4637 page_group_by_mobility_disabled = 0;
4639 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
4641 zonelist_order_name[current_zonelist_order],
4642 page_group_by_mobility_disabled ? "off" : "on",
4645 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4650 * Helper functions to size the waitqueue hash table.
4651 * Essentially these want to choose hash table sizes sufficiently
4652 * large so that collisions trying to wait on pages are rare.
4653 * But in fact, the number of active page waitqueues on typical
4654 * systems is ridiculously low, less than 200. So this is even
4655 * conservative, even though it seems large.
4657 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4658 * waitqueues, i.e. the size of the waitq table given the number of pages.
4660 #define PAGES_PER_WAITQUEUE 256
4662 #ifndef CONFIG_MEMORY_HOTPLUG
4663 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4665 unsigned long size = 1;
4667 pages /= PAGES_PER_WAITQUEUE;
4669 while (size < pages)
4673 * Once we have dozens or even hundreds of threads sleeping
4674 * on IO we've got bigger problems than wait queue collision.
4675 * Limit the size of the wait table to a reasonable size.
4677 size = min(size, 4096UL);
4679 return max(size, 4UL);
4683 * A zone's size might be changed by hot-add, so it is not possible to determine
4684 * a suitable size for its wait_table. So we use the maximum size now.
4686 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4688 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4689 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4690 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4692 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4693 * or more by the traditional way. (See above). It equals:
4695 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4696 * ia64(16K page size) : = ( 8G + 4M)byte.
4697 * powerpc (64K page size) : = (32G +16M)byte.
4699 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4706 * This is an integer logarithm so that shifts can be used later
4707 * to extract the more random high bits from the multiplicative
4708 * hash function before the remainder is taken.
4710 static inline unsigned long wait_table_bits(unsigned long size)
4716 * Initially all pages are reserved - free ones are freed
4717 * up by free_all_bootmem() once the early boot process is
4718 * done. Non-atomic initialization, single-pass.
4720 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4721 unsigned long start_pfn, enum memmap_context context)
4723 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
4724 unsigned long end_pfn = start_pfn + size;
4725 pg_data_t *pgdat = NODE_DATA(nid);
4727 unsigned long nr_initialised = 0;
4728 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4729 struct memblock_region *r = NULL, *tmp;
4732 if (highest_memmap_pfn < end_pfn - 1)
4733 highest_memmap_pfn = end_pfn - 1;
4736 * Honor reservation requested by the driver for this ZONE_DEVICE
4739 if (altmap && start_pfn == altmap->base_pfn)
4740 start_pfn += altmap->reserve;
4742 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4744 * There can be holes in boot-time mem_map[]s handed to this
4745 * function. They do not exist on hotplugged memory.
4747 if (context != MEMMAP_EARLY)
4750 if (!early_pfn_valid(pfn))
4752 if (!early_pfn_in_nid(pfn, nid))
4754 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
4757 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4759 * If not mirrored_kernelcore and ZONE_MOVABLE exists, range
4760 * from zone_movable_pfn[nid] to end of each node should be
4761 * ZONE_MOVABLE not ZONE_NORMAL. skip it.
4763 if (!mirrored_kernelcore && zone_movable_pfn[nid])
4764 if (zone == ZONE_NORMAL && pfn >= zone_movable_pfn[nid])
4768 * Check given memblock attribute by firmware which can affect
4769 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
4770 * mirrored, it's an overlapped memmap init. skip it.
4772 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
4773 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
4774 for_each_memblock(memory, tmp)
4775 if (pfn < memblock_region_memory_end_pfn(tmp))
4779 if (pfn >= memblock_region_memory_base_pfn(r) &&
4780 memblock_is_mirror(r)) {
4781 /* already initialized as NORMAL */
4782 pfn = memblock_region_memory_end_pfn(r);
4790 * Mark the block movable so that blocks are reserved for
4791 * movable at startup. This will force kernel allocations
4792 * to reserve their blocks rather than leaking throughout
4793 * the address space during boot when many long-lived
4794 * kernel allocations are made.
4796 * bitmap is created for zone's valid pfn range. but memmap
4797 * can be created for invalid pages (for alignment)
4798 * check here not to call set_pageblock_migratetype() against
4801 if (!(pfn & (pageblock_nr_pages - 1))) {
4802 struct page *page = pfn_to_page(pfn);
4804 __init_single_page(page, pfn, zone, nid);
4805 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4807 __init_single_pfn(pfn, zone, nid);
4812 static void __meminit zone_init_free_lists(struct zone *zone)
4814 unsigned int order, t;
4815 for_each_migratetype_order(order, t) {
4816 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4817 zone->free_area[order].nr_free = 0;
4821 #ifndef __HAVE_ARCH_MEMMAP_INIT
4822 #define memmap_init(size, nid, zone, start_pfn) \
4823 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4826 static int zone_batchsize(struct zone *zone)
4832 * The per-cpu-pages pools are set to around 1000th of the
4833 * size of the zone. But no more than 1/2 of a meg.
4835 * OK, so we don't know how big the cache is. So guess.
4837 batch = zone->managed_pages / 1024;
4838 if (batch * PAGE_SIZE > 512 * 1024)
4839 batch = (512 * 1024) / PAGE_SIZE;
4840 batch /= 4; /* We effectively *= 4 below */
4845 * Clamp the batch to a 2^n - 1 value. Having a power
4846 * of 2 value was found to be more likely to have
4847 * suboptimal cache aliasing properties in some cases.
4849 * For example if 2 tasks are alternately allocating
4850 * batches of pages, one task can end up with a lot
4851 * of pages of one half of the possible page colors
4852 * and the other with pages of the other colors.
4854 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4859 /* The deferral and batching of frees should be suppressed under NOMMU
4862 * The problem is that NOMMU needs to be able to allocate large chunks
4863 * of contiguous memory as there's no hardware page translation to
4864 * assemble apparent contiguous memory from discontiguous pages.
4866 * Queueing large contiguous runs of pages for batching, however,
4867 * causes the pages to actually be freed in smaller chunks. As there
4868 * can be a significant delay between the individual batches being
4869 * recycled, this leads to the once large chunks of space being
4870 * fragmented and becoming unavailable for high-order allocations.
4877 * pcp->high and pcp->batch values are related and dependent on one another:
4878 * ->batch must never be higher then ->high.
4879 * The following function updates them in a safe manner without read side
4882 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4883 * those fields changing asynchronously (acording the the above rule).
4885 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4886 * outside of boot time (or some other assurance that no concurrent updaters
4889 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4890 unsigned long batch)
4892 /* start with a fail safe value for batch */
4896 /* Update high, then batch, in order */
4903 /* a companion to pageset_set_high() */
4904 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4906 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4909 static void pageset_init(struct per_cpu_pageset *p)
4911 struct per_cpu_pages *pcp;
4914 memset(p, 0, sizeof(*p));
4918 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4919 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4922 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4925 pageset_set_batch(p, batch);
4929 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4930 * to the value high for the pageset p.
4932 static void pageset_set_high(struct per_cpu_pageset *p,
4935 unsigned long batch = max(1UL, high / 4);
4936 if ((high / 4) > (PAGE_SHIFT * 8))
4937 batch = PAGE_SHIFT * 8;
4939 pageset_update(&p->pcp, high, batch);
4942 static void pageset_set_high_and_batch(struct zone *zone,
4943 struct per_cpu_pageset *pcp)
4945 if (percpu_pagelist_fraction)
4946 pageset_set_high(pcp,
4947 (zone->managed_pages /
4948 percpu_pagelist_fraction));
4950 pageset_set_batch(pcp, zone_batchsize(zone));
4953 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4955 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4958 pageset_set_high_and_batch(zone, pcp);
4961 static void __meminit setup_zone_pageset(struct zone *zone)
4964 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4965 for_each_possible_cpu(cpu)
4966 zone_pageset_init(zone, cpu);
4970 * Allocate per cpu pagesets and initialize them.
4971 * Before this call only boot pagesets were available.
4973 void __init setup_per_cpu_pageset(void)
4977 for_each_populated_zone(zone)
4978 setup_zone_pageset(zone);
4981 static noinline __init_refok
4982 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4988 * The per-page waitqueue mechanism uses hashed waitqueues
4991 zone->wait_table_hash_nr_entries =
4992 wait_table_hash_nr_entries(zone_size_pages);
4993 zone->wait_table_bits =
4994 wait_table_bits(zone->wait_table_hash_nr_entries);
4995 alloc_size = zone->wait_table_hash_nr_entries
4996 * sizeof(wait_queue_head_t);
4998 if (!slab_is_available()) {
4999 zone->wait_table = (wait_queue_head_t *)
5000 memblock_virt_alloc_node_nopanic(
5001 alloc_size, zone->zone_pgdat->node_id);
5004 * This case means that a zone whose size was 0 gets new memory
5005 * via memory hot-add.
5006 * But it may be the case that a new node was hot-added. In
5007 * this case vmalloc() will not be able to use this new node's
5008 * memory - this wait_table must be initialized to use this new
5009 * node itself as well.
5010 * To use this new node's memory, further consideration will be
5013 zone->wait_table = vmalloc(alloc_size);
5015 if (!zone->wait_table)
5018 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
5019 init_waitqueue_head(zone->wait_table + i);
5024 static __meminit void zone_pcp_init(struct zone *zone)
5027 * per cpu subsystem is not up at this point. The following code
5028 * relies on the ability of the linker to provide the
5029 * offset of a (static) per cpu variable into the per cpu area.
5031 zone->pageset = &boot_pageset;
5033 if (populated_zone(zone))
5034 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5035 zone->name, zone->present_pages,
5036 zone_batchsize(zone));
5039 int __meminit init_currently_empty_zone(struct zone *zone,
5040 unsigned long zone_start_pfn,
5043 struct pglist_data *pgdat = zone->zone_pgdat;
5045 ret = zone_wait_table_init(zone, size);
5048 pgdat->nr_zones = zone_idx(zone) + 1;
5050 zone->zone_start_pfn = zone_start_pfn;
5052 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5053 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5055 (unsigned long)zone_idx(zone),
5056 zone_start_pfn, (zone_start_pfn + size));
5058 zone_init_free_lists(zone);
5063 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5064 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5067 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5069 int __meminit __early_pfn_to_nid(unsigned long pfn,
5070 struct mminit_pfnnid_cache *state)
5072 unsigned long start_pfn, end_pfn;
5075 if (state->last_start <= pfn && pfn < state->last_end)
5076 return state->last_nid;
5078 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5080 state->last_start = start_pfn;
5081 state->last_end = end_pfn;
5082 state->last_nid = nid;
5087 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5090 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5091 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5092 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5094 * If an architecture guarantees that all ranges registered contain no holes
5095 * and may be freed, this this function may be used instead of calling
5096 * memblock_free_early_nid() manually.
5098 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5100 unsigned long start_pfn, end_pfn;
5103 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5104 start_pfn = min(start_pfn, max_low_pfn);
5105 end_pfn = min(end_pfn, max_low_pfn);
5107 if (start_pfn < end_pfn)
5108 memblock_free_early_nid(PFN_PHYS(start_pfn),
5109 (end_pfn - start_pfn) << PAGE_SHIFT,
5115 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5116 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5118 * If an architecture guarantees that all ranges registered contain no holes and may
5119 * be freed, this function may be used instead of calling memory_present() manually.
5121 void __init sparse_memory_present_with_active_regions(int nid)
5123 unsigned long start_pfn, end_pfn;
5126 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5127 memory_present(this_nid, start_pfn, end_pfn);
5131 * get_pfn_range_for_nid - Return the start and end page frames for a node
5132 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5133 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5134 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5136 * It returns the start and end page frame of a node based on information
5137 * provided by memblock_set_node(). If called for a node
5138 * with no available memory, a warning is printed and the start and end
5141 void __meminit get_pfn_range_for_nid(unsigned int nid,
5142 unsigned long *start_pfn, unsigned long *end_pfn)
5144 unsigned long this_start_pfn, this_end_pfn;
5150 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5151 *start_pfn = min(*start_pfn, this_start_pfn);
5152 *end_pfn = max(*end_pfn, this_end_pfn);
5155 if (*start_pfn == -1UL)
5160 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5161 * assumption is made that zones within a node are ordered in monotonic
5162 * increasing memory addresses so that the "highest" populated zone is used
5164 static void __init find_usable_zone_for_movable(void)
5167 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5168 if (zone_index == ZONE_MOVABLE)
5171 if (arch_zone_highest_possible_pfn[zone_index] >
5172 arch_zone_lowest_possible_pfn[zone_index])
5176 VM_BUG_ON(zone_index == -1);
5177 movable_zone = zone_index;
5181 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5182 * because it is sized independent of architecture. Unlike the other zones,
5183 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5184 * in each node depending on the size of each node and how evenly kernelcore
5185 * is distributed. This helper function adjusts the zone ranges
5186 * provided by the architecture for a given node by using the end of the
5187 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5188 * zones within a node are in order of monotonic increases memory addresses
5190 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5191 unsigned long zone_type,
5192 unsigned long node_start_pfn,
5193 unsigned long node_end_pfn,
5194 unsigned long *zone_start_pfn,
5195 unsigned long *zone_end_pfn)
5197 /* Only adjust if ZONE_MOVABLE is on this node */
5198 if (zone_movable_pfn[nid]) {
5199 /* Size ZONE_MOVABLE */
5200 if (zone_type == ZONE_MOVABLE) {
5201 *zone_start_pfn = zone_movable_pfn[nid];
5202 *zone_end_pfn = min(node_end_pfn,
5203 arch_zone_highest_possible_pfn[movable_zone]);
5205 /* Check if this whole range is within ZONE_MOVABLE */
5206 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5207 *zone_start_pfn = *zone_end_pfn;
5212 * Return the number of pages a zone spans in a node, including holes
5213 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5215 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5216 unsigned long zone_type,
5217 unsigned long node_start_pfn,
5218 unsigned long node_end_pfn,
5219 unsigned long *zone_start_pfn,
5220 unsigned long *zone_end_pfn,
5221 unsigned long *ignored)
5223 /* When hotadd a new node from cpu_up(), the node should be empty */
5224 if (!node_start_pfn && !node_end_pfn)
5227 /* Get the start and end of the zone */
5228 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5229 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5230 adjust_zone_range_for_zone_movable(nid, zone_type,
5231 node_start_pfn, node_end_pfn,
5232 zone_start_pfn, zone_end_pfn);
5234 /* Check that this node has pages within the zone's required range */
5235 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5238 /* Move the zone boundaries inside the node if necessary */
5239 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5240 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5242 /* Return the spanned pages */
5243 return *zone_end_pfn - *zone_start_pfn;
5247 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5248 * then all holes in the requested range will be accounted for.
5250 unsigned long __meminit __absent_pages_in_range(int nid,
5251 unsigned long range_start_pfn,
5252 unsigned long range_end_pfn)
5254 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5255 unsigned long start_pfn, end_pfn;
5258 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5259 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5260 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5261 nr_absent -= end_pfn - start_pfn;
5267 * absent_pages_in_range - Return number of page frames in holes within a range
5268 * @start_pfn: The start PFN to start searching for holes
5269 * @end_pfn: The end PFN to stop searching for holes
5271 * It returns the number of pages frames in memory holes within a range.
5273 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5274 unsigned long end_pfn)
5276 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5279 /* Return the number of page frames in holes in a zone on a node */
5280 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5281 unsigned long zone_type,
5282 unsigned long node_start_pfn,
5283 unsigned long node_end_pfn,
5284 unsigned long *ignored)
5286 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5287 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5288 unsigned long zone_start_pfn, zone_end_pfn;
5289 unsigned long nr_absent;
5291 /* When hotadd a new node from cpu_up(), the node should be empty */
5292 if (!node_start_pfn && !node_end_pfn)
5295 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5296 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5298 adjust_zone_range_for_zone_movable(nid, zone_type,
5299 node_start_pfn, node_end_pfn,
5300 &zone_start_pfn, &zone_end_pfn);
5301 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5304 * ZONE_MOVABLE handling.
5305 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5308 if (zone_movable_pfn[nid]) {
5309 if (mirrored_kernelcore) {
5310 unsigned long start_pfn, end_pfn;
5311 struct memblock_region *r;
5313 for_each_memblock(memory, r) {
5314 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5315 zone_start_pfn, zone_end_pfn);
5316 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5317 zone_start_pfn, zone_end_pfn);
5319 if (zone_type == ZONE_MOVABLE &&
5320 memblock_is_mirror(r))
5321 nr_absent += end_pfn - start_pfn;
5323 if (zone_type == ZONE_NORMAL &&
5324 !memblock_is_mirror(r))
5325 nr_absent += end_pfn - start_pfn;
5328 if (zone_type == ZONE_NORMAL)
5329 nr_absent += node_end_pfn - zone_movable_pfn[nid];
5336 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5337 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5338 unsigned long zone_type,
5339 unsigned long node_start_pfn,
5340 unsigned long node_end_pfn,
5341 unsigned long *zone_start_pfn,
5342 unsigned long *zone_end_pfn,
5343 unsigned long *zones_size)
5347 *zone_start_pfn = node_start_pfn;
5348 for (zone = 0; zone < zone_type; zone++)
5349 *zone_start_pfn += zones_size[zone];
5351 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5353 return zones_size[zone_type];
5356 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5357 unsigned long zone_type,
5358 unsigned long node_start_pfn,
5359 unsigned long node_end_pfn,
5360 unsigned long *zholes_size)
5365 return zholes_size[zone_type];
5368 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5370 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5371 unsigned long node_start_pfn,
5372 unsigned long node_end_pfn,
5373 unsigned long *zones_size,
5374 unsigned long *zholes_size)
5376 unsigned long realtotalpages = 0, totalpages = 0;
5379 for (i = 0; i < MAX_NR_ZONES; i++) {
5380 struct zone *zone = pgdat->node_zones + i;
5381 unsigned long zone_start_pfn, zone_end_pfn;
5382 unsigned long size, real_size;
5384 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5390 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5391 node_start_pfn, node_end_pfn,
5394 zone->zone_start_pfn = zone_start_pfn;
5396 zone->zone_start_pfn = 0;
5397 zone->spanned_pages = size;
5398 zone->present_pages = real_size;
5401 realtotalpages += real_size;
5404 pgdat->node_spanned_pages = totalpages;
5405 pgdat->node_present_pages = realtotalpages;
5406 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5410 #ifndef CONFIG_SPARSEMEM
5412 * Calculate the size of the zone->blockflags rounded to an unsigned long
5413 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5414 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5415 * round what is now in bits to nearest long in bits, then return it in
5418 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5420 unsigned long usemapsize;
5422 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5423 usemapsize = roundup(zonesize, pageblock_nr_pages);
5424 usemapsize = usemapsize >> pageblock_order;
5425 usemapsize *= NR_PAGEBLOCK_BITS;
5426 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5428 return usemapsize / 8;
5431 static void __init setup_usemap(struct pglist_data *pgdat,
5433 unsigned long zone_start_pfn,
5434 unsigned long zonesize)
5436 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5437 zone->pageblock_flags = NULL;
5439 zone->pageblock_flags =
5440 memblock_virt_alloc_node_nopanic(usemapsize,
5444 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5445 unsigned long zone_start_pfn, unsigned long zonesize) {}
5446 #endif /* CONFIG_SPARSEMEM */
5448 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5450 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5451 void __paginginit set_pageblock_order(void)
5455 /* Check that pageblock_nr_pages has not already been setup */
5456 if (pageblock_order)
5459 if (HPAGE_SHIFT > PAGE_SHIFT)
5460 order = HUGETLB_PAGE_ORDER;
5462 order = MAX_ORDER - 1;
5465 * Assume the largest contiguous order of interest is a huge page.
5466 * This value may be variable depending on boot parameters on IA64 and
5469 pageblock_order = order;
5471 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5474 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5475 * is unused as pageblock_order is set at compile-time. See
5476 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5479 void __paginginit set_pageblock_order(void)
5483 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5485 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5486 unsigned long present_pages)
5488 unsigned long pages = spanned_pages;
5491 * Provide a more accurate estimation if there are holes within
5492 * the zone and SPARSEMEM is in use. If there are holes within the
5493 * zone, each populated memory region may cost us one or two extra
5494 * memmap pages due to alignment because memmap pages for each
5495 * populated regions may not naturally algined on page boundary.
5496 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5498 if (spanned_pages > present_pages + (present_pages >> 4) &&
5499 IS_ENABLED(CONFIG_SPARSEMEM))
5500 pages = present_pages;
5502 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5506 * Set up the zone data structures:
5507 * - mark all pages reserved
5508 * - mark all memory queues empty
5509 * - clear the memory bitmaps
5511 * NOTE: pgdat should get zeroed by caller.
5513 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5516 int nid = pgdat->node_id;
5519 pgdat_resize_init(pgdat);
5520 #ifdef CONFIG_NUMA_BALANCING
5521 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5522 pgdat->numabalancing_migrate_nr_pages = 0;
5523 pgdat->numabalancing_migrate_next_window = jiffies;
5525 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5526 spin_lock_init(&pgdat->split_queue_lock);
5527 INIT_LIST_HEAD(&pgdat->split_queue);
5528 pgdat->split_queue_len = 0;
5530 init_waitqueue_head(&pgdat->kswapd_wait);
5531 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5532 #ifdef CONFIG_COMPACTION
5533 init_waitqueue_head(&pgdat->kcompactd_wait);
5535 pgdat_page_ext_init(pgdat);
5537 for (j = 0; j < MAX_NR_ZONES; j++) {
5538 struct zone *zone = pgdat->node_zones + j;
5539 unsigned long size, realsize, freesize, memmap_pages;
5540 unsigned long zone_start_pfn = zone->zone_start_pfn;
5542 size = zone->spanned_pages;
5543 realsize = freesize = zone->present_pages;
5546 * Adjust freesize so that it accounts for how much memory
5547 * is used by this zone for memmap. This affects the watermark
5548 * and per-cpu initialisations
5550 memmap_pages = calc_memmap_size(size, realsize);
5551 if (!is_highmem_idx(j)) {
5552 if (freesize >= memmap_pages) {
5553 freesize -= memmap_pages;
5556 " %s zone: %lu pages used for memmap\n",
5557 zone_names[j], memmap_pages);
5559 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5560 zone_names[j], memmap_pages, freesize);
5563 /* Account for reserved pages */
5564 if (j == 0 && freesize > dma_reserve) {
5565 freesize -= dma_reserve;
5566 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5567 zone_names[0], dma_reserve);
5570 if (!is_highmem_idx(j))
5571 nr_kernel_pages += freesize;
5572 /* Charge for highmem memmap if there are enough kernel pages */
5573 else if (nr_kernel_pages > memmap_pages * 2)
5574 nr_kernel_pages -= memmap_pages;
5575 nr_all_pages += freesize;
5578 * Set an approximate value for lowmem here, it will be adjusted
5579 * when the bootmem allocator frees pages into the buddy system.
5580 * And all highmem pages will be managed by the buddy system.
5582 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5585 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5587 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5589 zone->name = zone_names[j];
5590 spin_lock_init(&zone->lock);
5591 spin_lock_init(&zone->lru_lock);
5592 zone_seqlock_init(zone);
5593 zone->zone_pgdat = pgdat;
5594 zone_pcp_init(zone);
5596 /* For bootup, initialized properly in watermark setup */
5597 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5599 lruvec_init(&zone->lruvec);
5603 set_pageblock_order();
5604 setup_usemap(pgdat, zone, zone_start_pfn, size);
5605 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5607 memmap_init(size, nid, j, zone_start_pfn);
5611 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5613 unsigned long __maybe_unused start = 0;
5614 unsigned long __maybe_unused offset = 0;
5616 /* Skip empty nodes */
5617 if (!pgdat->node_spanned_pages)
5620 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5621 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5622 offset = pgdat->node_start_pfn - start;
5623 /* ia64 gets its own node_mem_map, before this, without bootmem */
5624 if (!pgdat->node_mem_map) {
5625 unsigned long size, end;
5629 * The zone's endpoints aren't required to be MAX_ORDER
5630 * aligned but the node_mem_map endpoints must be in order
5631 * for the buddy allocator to function correctly.
5633 end = pgdat_end_pfn(pgdat);
5634 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5635 size = (end - start) * sizeof(struct page);
5636 map = alloc_remap(pgdat->node_id, size);
5638 map = memblock_virt_alloc_node_nopanic(size,
5640 pgdat->node_mem_map = map + offset;
5642 #ifndef CONFIG_NEED_MULTIPLE_NODES
5644 * With no DISCONTIG, the global mem_map is just set as node 0's
5646 if (pgdat == NODE_DATA(0)) {
5647 mem_map = NODE_DATA(0)->node_mem_map;
5648 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5649 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5651 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5654 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5657 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5658 unsigned long node_start_pfn, unsigned long *zholes_size)
5660 pg_data_t *pgdat = NODE_DATA(nid);
5661 unsigned long start_pfn = 0;
5662 unsigned long end_pfn = 0;
5664 /* pg_data_t should be reset to zero when it's allocated */
5665 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5667 reset_deferred_meminit(pgdat);
5668 pgdat->node_id = nid;
5669 pgdat->node_start_pfn = node_start_pfn;
5670 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5671 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5672 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5673 (u64)start_pfn << PAGE_SHIFT,
5674 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5676 start_pfn = node_start_pfn;
5678 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5679 zones_size, zholes_size);
5681 alloc_node_mem_map(pgdat);
5682 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5683 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5684 nid, (unsigned long)pgdat,
5685 (unsigned long)pgdat->node_mem_map);
5688 free_area_init_core(pgdat);
5691 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5693 #if MAX_NUMNODES > 1
5695 * Figure out the number of possible node ids.
5697 void __init setup_nr_node_ids(void)
5699 unsigned int highest;
5701 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5702 nr_node_ids = highest + 1;
5707 * node_map_pfn_alignment - determine the maximum internode alignment
5709 * This function should be called after node map is populated and sorted.
5710 * It calculates the maximum power of two alignment which can distinguish
5713 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5714 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5715 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5716 * shifted, 1GiB is enough and this function will indicate so.
5718 * This is used to test whether pfn -> nid mapping of the chosen memory
5719 * model has fine enough granularity to avoid incorrect mapping for the
5720 * populated node map.
5722 * Returns the determined alignment in pfn's. 0 if there is no alignment
5723 * requirement (single node).
5725 unsigned long __init node_map_pfn_alignment(void)
5727 unsigned long accl_mask = 0, last_end = 0;
5728 unsigned long start, end, mask;
5732 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5733 if (!start || last_nid < 0 || last_nid == nid) {
5740 * Start with a mask granular enough to pin-point to the
5741 * start pfn and tick off bits one-by-one until it becomes
5742 * too coarse to separate the current node from the last.
5744 mask = ~((1 << __ffs(start)) - 1);
5745 while (mask && last_end <= (start & (mask << 1)))
5748 /* accumulate all internode masks */
5752 /* convert mask to number of pages */
5753 return ~accl_mask + 1;
5756 /* Find the lowest pfn for a node */
5757 static unsigned long __init find_min_pfn_for_node(int nid)
5759 unsigned long min_pfn = ULONG_MAX;
5760 unsigned long start_pfn;
5763 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5764 min_pfn = min(min_pfn, start_pfn);
5766 if (min_pfn == ULONG_MAX) {
5767 pr_warn("Could not find start_pfn for node %d\n", nid);
5775 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5777 * It returns the minimum PFN based on information provided via
5778 * memblock_set_node().
5780 unsigned long __init find_min_pfn_with_active_regions(void)
5782 return find_min_pfn_for_node(MAX_NUMNODES);
5786 * early_calculate_totalpages()
5787 * Sum pages in active regions for movable zone.
5788 * Populate N_MEMORY for calculating usable_nodes.
5790 static unsigned long __init early_calculate_totalpages(void)
5792 unsigned long totalpages = 0;
5793 unsigned long start_pfn, end_pfn;
5796 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5797 unsigned long pages = end_pfn - start_pfn;
5799 totalpages += pages;
5801 node_set_state(nid, N_MEMORY);
5807 * Find the PFN the Movable zone begins in each node. Kernel memory
5808 * is spread evenly between nodes as long as the nodes have enough
5809 * memory. When they don't, some nodes will have more kernelcore than
5812 static void __init find_zone_movable_pfns_for_nodes(void)
5815 unsigned long usable_startpfn;
5816 unsigned long kernelcore_node, kernelcore_remaining;
5817 /* save the state before borrow the nodemask */
5818 nodemask_t saved_node_state = node_states[N_MEMORY];
5819 unsigned long totalpages = early_calculate_totalpages();
5820 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5821 struct memblock_region *r;
5823 /* Need to find movable_zone earlier when movable_node is specified. */
5824 find_usable_zone_for_movable();
5827 * If movable_node is specified, ignore kernelcore and movablecore
5830 if (movable_node_is_enabled()) {
5831 for_each_memblock(memory, r) {
5832 if (!memblock_is_hotpluggable(r))
5837 usable_startpfn = PFN_DOWN(r->base);
5838 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5839 min(usable_startpfn, zone_movable_pfn[nid]) :
5847 * If kernelcore=mirror is specified, ignore movablecore option
5849 if (mirrored_kernelcore) {
5850 bool mem_below_4gb_not_mirrored = false;
5852 for_each_memblock(memory, r) {
5853 if (memblock_is_mirror(r))
5858 usable_startpfn = memblock_region_memory_base_pfn(r);
5860 if (usable_startpfn < 0x100000) {
5861 mem_below_4gb_not_mirrored = true;
5865 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5866 min(usable_startpfn, zone_movable_pfn[nid]) :
5870 if (mem_below_4gb_not_mirrored)
5871 pr_warn("This configuration results in unmirrored kernel memory.");
5877 * If movablecore=nn[KMG] was specified, calculate what size of
5878 * kernelcore that corresponds so that memory usable for
5879 * any allocation type is evenly spread. If both kernelcore
5880 * and movablecore are specified, then the value of kernelcore
5881 * will be used for required_kernelcore if it's greater than
5882 * what movablecore would have allowed.
5884 if (required_movablecore) {
5885 unsigned long corepages;
5888 * Round-up so that ZONE_MOVABLE is at least as large as what
5889 * was requested by the user
5891 required_movablecore =
5892 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5893 required_movablecore = min(totalpages, required_movablecore);
5894 corepages = totalpages - required_movablecore;
5896 required_kernelcore = max(required_kernelcore, corepages);
5900 * If kernelcore was not specified or kernelcore size is larger
5901 * than totalpages, there is no ZONE_MOVABLE.
5903 if (!required_kernelcore || required_kernelcore >= totalpages)
5906 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5907 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5910 /* Spread kernelcore memory as evenly as possible throughout nodes */
5911 kernelcore_node = required_kernelcore / usable_nodes;
5912 for_each_node_state(nid, N_MEMORY) {
5913 unsigned long start_pfn, end_pfn;
5916 * Recalculate kernelcore_node if the division per node
5917 * now exceeds what is necessary to satisfy the requested
5918 * amount of memory for the kernel
5920 if (required_kernelcore < kernelcore_node)
5921 kernelcore_node = required_kernelcore / usable_nodes;
5924 * As the map is walked, we track how much memory is usable
5925 * by the kernel using kernelcore_remaining. When it is
5926 * 0, the rest of the node is usable by ZONE_MOVABLE
5928 kernelcore_remaining = kernelcore_node;
5930 /* Go through each range of PFNs within this node */
5931 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5932 unsigned long size_pages;
5934 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5935 if (start_pfn >= end_pfn)
5938 /* Account for what is only usable for kernelcore */
5939 if (start_pfn < usable_startpfn) {
5940 unsigned long kernel_pages;
5941 kernel_pages = min(end_pfn, usable_startpfn)
5944 kernelcore_remaining -= min(kernel_pages,
5945 kernelcore_remaining);
5946 required_kernelcore -= min(kernel_pages,
5947 required_kernelcore);
5949 /* Continue if range is now fully accounted */
5950 if (end_pfn <= usable_startpfn) {
5953 * Push zone_movable_pfn to the end so
5954 * that if we have to rebalance
5955 * kernelcore across nodes, we will
5956 * not double account here
5958 zone_movable_pfn[nid] = end_pfn;
5961 start_pfn = usable_startpfn;
5965 * The usable PFN range for ZONE_MOVABLE is from
5966 * start_pfn->end_pfn. Calculate size_pages as the
5967 * number of pages used as kernelcore
5969 size_pages = end_pfn - start_pfn;
5970 if (size_pages > kernelcore_remaining)
5971 size_pages = kernelcore_remaining;
5972 zone_movable_pfn[nid] = start_pfn + size_pages;
5975 * Some kernelcore has been met, update counts and
5976 * break if the kernelcore for this node has been
5979 required_kernelcore -= min(required_kernelcore,
5981 kernelcore_remaining -= size_pages;
5982 if (!kernelcore_remaining)
5988 * If there is still required_kernelcore, we do another pass with one
5989 * less node in the count. This will push zone_movable_pfn[nid] further
5990 * along on the nodes that still have memory until kernelcore is
5994 if (usable_nodes && required_kernelcore > usable_nodes)
5998 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5999 for (nid = 0; nid < MAX_NUMNODES; nid++)
6000 zone_movable_pfn[nid] =
6001 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6004 /* restore the node_state */
6005 node_states[N_MEMORY] = saved_node_state;
6008 /* Any regular or high memory on that node ? */
6009 static void check_for_memory(pg_data_t *pgdat, int nid)
6011 enum zone_type zone_type;
6013 if (N_MEMORY == N_NORMAL_MEMORY)
6016 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6017 struct zone *zone = &pgdat->node_zones[zone_type];
6018 if (populated_zone(zone)) {
6019 node_set_state(nid, N_HIGH_MEMORY);
6020 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6021 zone_type <= ZONE_NORMAL)
6022 node_set_state(nid, N_NORMAL_MEMORY);
6029 * free_area_init_nodes - Initialise all pg_data_t and zone data
6030 * @max_zone_pfn: an array of max PFNs for each zone
6032 * This will call free_area_init_node() for each active node in the system.
6033 * Using the page ranges provided by memblock_set_node(), the size of each
6034 * zone in each node and their holes is calculated. If the maximum PFN
6035 * between two adjacent zones match, it is assumed that the zone is empty.
6036 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6037 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6038 * starts where the previous one ended. For example, ZONE_DMA32 starts
6039 * at arch_max_dma_pfn.
6041 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6043 unsigned long start_pfn, end_pfn;
6046 /* Record where the zone boundaries are */
6047 memset(arch_zone_lowest_possible_pfn, 0,
6048 sizeof(arch_zone_lowest_possible_pfn));
6049 memset(arch_zone_highest_possible_pfn, 0,
6050 sizeof(arch_zone_highest_possible_pfn));
6051 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
6052 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
6053 for (i = 1; i < MAX_NR_ZONES; i++) {
6054 if (i == ZONE_MOVABLE)
6056 arch_zone_lowest_possible_pfn[i] =
6057 arch_zone_highest_possible_pfn[i-1];
6058 arch_zone_highest_possible_pfn[i] =
6059 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
6061 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
6062 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
6064 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6065 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6066 find_zone_movable_pfns_for_nodes();
6068 /* Print out the zone ranges */
6069 pr_info("Zone ranges:\n");
6070 for (i = 0; i < MAX_NR_ZONES; i++) {
6071 if (i == ZONE_MOVABLE)
6073 pr_info(" %-8s ", zone_names[i]);
6074 if (arch_zone_lowest_possible_pfn[i] ==
6075 arch_zone_highest_possible_pfn[i])
6078 pr_cont("[mem %#018Lx-%#018Lx]\n",
6079 (u64)arch_zone_lowest_possible_pfn[i]
6081 ((u64)arch_zone_highest_possible_pfn[i]
6082 << PAGE_SHIFT) - 1);
6085 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6086 pr_info("Movable zone start for each node\n");
6087 for (i = 0; i < MAX_NUMNODES; i++) {
6088 if (zone_movable_pfn[i])
6089 pr_info(" Node %d: %#018Lx\n", i,
6090 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6093 /* Print out the early node map */
6094 pr_info("Early memory node ranges\n");
6095 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6096 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6097 (u64)start_pfn << PAGE_SHIFT,
6098 ((u64)end_pfn << PAGE_SHIFT) - 1);
6100 /* Initialise every node */
6101 mminit_verify_pageflags_layout();
6102 setup_nr_node_ids();
6103 for_each_online_node(nid) {
6104 pg_data_t *pgdat = NODE_DATA(nid);
6105 free_area_init_node(nid, NULL,
6106 find_min_pfn_for_node(nid), NULL);
6108 /* Any memory on that node */
6109 if (pgdat->node_present_pages)
6110 node_set_state(nid, N_MEMORY);
6111 check_for_memory(pgdat, nid);
6115 static int __init cmdline_parse_core(char *p, unsigned long *core)
6117 unsigned long long coremem;
6121 coremem = memparse(p, &p);
6122 *core = coremem >> PAGE_SHIFT;
6124 /* Paranoid check that UL is enough for the coremem value */
6125 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6131 * kernelcore=size sets the amount of memory for use for allocations that
6132 * cannot be reclaimed or migrated.
6134 static int __init cmdline_parse_kernelcore(char *p)
6136 /* parse kernelcore=mirror */
6137 if (parse_option_str(p, "mirror")) {
6138 mirrored_kernelcore = true;
6142 return cmdline_parse_core(p, &required_kernelcore);
6146 * movablecore=size sets the amount of memory for use for allocations that
6147 * can be reclaimed or migrated.
6149 static int __init cmdline_parse_movablecore(char *p)
6151 return cmdline_parse_core(p, &required_movablecore);
6154 early_param("kernelcore", cmdline_parse_kernelcore);
6155 early_param("movablecore", cmdline_parse_movablecore);
6157 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6159 void adjust_managed_page_count(struct page *page, long count)
6161 spin_lock(&managed_page_count_lock);
6162 page_zone(page)->managed_pages += count;
6163 totalram_pages += count;
6164 #ifdef CONFIG_HIGHMEM
6165 if (PageHighMem(page))
6166 totalhigh_pages += count;
6168 spin_unlock(&managed_page_count_lock);
6170 EXPORT_SYMBOL(adjust_managed_page_count);
6172 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6175 unsigned long pages = 0;
6177 start = (void *)PAGE_ALIGN((unsigned long)start);
6178 end = (void *)((unsigned long)end & PAGE_MASK);
6179 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6180 if ((unsigned int)poison <= 0xFF)
6181 memset(pos, poison, PAGE_SIZE);
6182 free_reserved_page(virt_to_page(pos));
6186 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6187 s, pages << (PAGE_SHIFT - 10), start, end);
6191 EXPORT_SYMBOL(free_reserved_area);
6193 #ifdef CONFIG_HIGHMEM
6194 void free_highmem_page(struct page *page)
6196 __free_reserved_page(page);
6198 page_zone(page)->managed_pages++;
6204 void __init mem_init_print_info(const char *str)
6206 unsigned long physpages, codesize, datasize, rosize, bss_size;
6207 unsigned long init_code_size, init_data_size;
6209 physpages = get_num_physpages();
6210 codesize = _etext - _stext;
6211 datasize = _edata - _sdata;
6212 rosize = __end_rodata - __start_rodata;
6213 bss_size = __bss_stop - __bss_start;
6214 init_data_size = __init_end - __init_begin;
6215 init_code_size = _einittext - _sinittext;
6218 * Detect special cases and adjust section sizes accordingly:
6219 * 1) .init.* may be embedded into .data sections
6220 * 2) .init.text.* may be out of [__init_begin, __init_end],
6221 * please refer to arch/tile/kernel/vmlinux.lds.S.
6222 * 3) .rodata.* may be embedded into .text or .data sections.
6224 #define adj_init_size(start, end, size, pos, adj) \
6226 if (start <= pos && pos < end && size > adj) \
6230 adj_init_size(__init_begin, __init_end, init_data_size,
6231 _sinittext, init_code_size);
6232 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6233 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6234 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6235 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6237 #undef adj_init_size
6239 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6240 #ifdef CONFIG_HIGHMEM
6244 nr_free_pages() << (PAGE_SHIFT - 10),
6245 physpages << (PAGE_SHIFT - 10),
6246 codesize >> 10, datasize >> 10, rosize >> 10,
6247 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6248 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6249 totalcma_pages << (PAGE_SHIFT - 10),
6250 #ifdef CONFIG_HIGHMEM
6251 totalhigh_pages << (PAGE_SHIFT - 10),
6253 str ? ", " : "", str ? str : "");
6257 * set_dma_reserve - set the specified number of pages reserved in the first zone
6258 * @new_dma_reserve: The number of pages to mark reserved
6260 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6261 * In the DMA zone, a significant percentage may be consumed by kernel image
6262 * and other unfreeable allocations which can skew the watermarks badly. This
6263 * function may optionally be used to account for unfreeable pages in the
6264 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6265 * smaller per-cpu batchsize.
6267 void __init set_dma_reserve(unsigned long new_dma_reserve)
6269 dma_reserve = new_dma_reserve;
6272 void __init free_area_init(unsigned long *zones_size)
6274 free_area_init_node(0, zones_size,
6275 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6278 static int page_alloc_cpu_notify(struct notifier_block *self,
6279 unsigned long action, void *hcpu)
6281 int cpu = (unsigned long)hcpu;
6283 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6284 lru_add_drain_cpu(cpu);
6288 * Spill the event counters of the dead processor
6289 * into the current processors event counters.
6290 * This artificially elevates the count of the current
6293 vm_events_fold_cpu(cpu);
6296 * Zero the differential counters of the dead processor
6297 * so that the vm statistics are consistent.
6299 * This is only okay since the processor is dead and cannot
6300 * race with what we are doing.
6302 cpu_vm_stats_fold(cpu);
6307 void __init page_alloc_init(void)
6309 hotcpu_notifier(page_alloc_cpu_notify, 0);
6313 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6314 * or min_free_kbytes changes.
6316 static void calculate_totalreserve_pages(void)
6318 struct pglist_data *pgdat;
6319 unsigned long reserve_pages = 0;
6320 enum zone_type i, j;
6322 for_each_online_pgdat(pgdat) {
6323 for (i = 0; i < MAX_NR_ZONES; i++) {
6324 struct zone *zone = pgdat->node_zones + i;
6327 /* Find valid and maximum lowmem_reserve in the zone */
6328 for (j = i; j < MAX_NR_ZONES; j++) {
6329 if (zone->lowmem_reserve[j] > max)
6330 max = zone->lowmem_reserve[j];
6333 /* we treat the high watermark as reserved pages. */
6334 max += high_wmark_pages(zone);
6336 if (max > zone->managed_pages)
6337 max = zone->managed_pages;
6339 zone->totalreserve_pages = max;
6341 reserve_pages += max;
6344 totalreserve_pages = reserve_pages;
6348 * setup_per_zone_lowmem_reserve - called whenever
6349 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6350 * has a correct pages reserved value, so an adequate number of
6351 * pages are left in the zone after a successful __alloc_pages().
6353 static void setup_per_zone_lowmem_reserve(void)
6355 struct pglist_data *pgdat;
6356 enum zone_type j, idx;
6358 for_each_online_pgdat(pgdat) {
6359 for (j = 0; j < MAX_NR_ZONES; j++) {
6360 struct zone *zone = pgdat->node_zones + j;
6361 unsigned long managed_pages = zone->managed_pages;
6363 zone->lowmem_reserve[j] = 0;
6367 struct zone *lower_zone;
6371 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6372 sysctl_lowmem_reserve_ratio[idx] = 1;
6374 lower_zone = pgdat->node_zones + idx;
6375 lower_zone->lowmem_reserve[j] = managed_pages /
6376 sysctl_lowmem_reserve_ratio[idx];
6377 managed_pages += lower_zone->managed_pages;
6382 /* update totalreserve_pages */
6383 calculate_totalreserve_pages();
6386 static void __setup_per_zone_wmarks(void)
6388 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6389 unsigned long lowmem_pages = 0;
6391 unsigned long flags;
6393 /* Calculate total number of !ZONE_HIGHMEM pages */
6394 for_each_zone(zone) {
6395 if (!is_highmem(zone))
6396 lowmem_pages += zone->managed_pages;
6399 for_each_zone(zone) {
6402 spin_lock_irqsave(&zone->lock, flags);
6403 tmp = (u64)pages_min * zone->managed_pages;
6404 do_div(tmp, lowmem_pages);
6405 if (is_highmem(zone)) {
6407 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6408 * need highmem pages, so cap pages_min to a small
6411 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6412 * deltas control asynch page reclaim, and so should
6413 * not be capped for highmem.
6415 unsigned long min_pages;
6417 min_pages = zone->managed_pages / 1024;
6418 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6419 zone->watermark[WMARK_MIN] = min_pages;
6422 * If it's a lowmem zone, reserve a number of pages
6423 * proportionate to the zone's size.
6425 zone->watermark[WMARK_MIN] = tmp;
6429 * Set the kswapd watermarks distance according to the
6430 * scale factor in proportion to available memory, but
6431 * ensure a minimum size on small systems.
6433 tmp = max_t(u64, tmp >> 2,
6434 mult_frac(zone->managed_pages,
6435 watermark_scale_factor, 10000));
6437 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6438 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6440 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6441 high_wmark_pages(zone) - low_wmark_pages(zone) -
6442 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6444 spin_unlock_irqrestore(&zone->lock, flags);
6447 /* update totalreserve_pages */
6448 calculate_totalreserve_pages();
6452 * setup_per_zone_wmarks - called when min_free_kbytes changes
6453 * or when memory is hot-{added|removed}
6455 * Ensures that the watermark[min,low,high] values for each zone are set
6456 * correctly with respect to min_free_kbytes.
6458 void setup_per_zone_wmarks(void)
6460 mutex_lock(&zonelists_mutex);
6461 __setup_per_zone_wmarks();
6462 mutex_unlock(&zonelists_mutex);
6466 * The inactive anon list should be small enough that the VM never has to
6467 * do too much work, but large enough that each inactive page has a chance
6468 * to be referenced again before it is swapped out.
6470 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6471 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6472 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6473 * the anonymous pages are kept on the inactive list.
6476 * memory ratio inactive anon
6477 * -------------------------------------
6486 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6488 unsigned int gb, ratio;
6490 /* Zone size in gigabytes */
6491 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6493 ratio = int_sqrt(10 * gb);
6497 zone->inactive_ratio = ratio;
6500 static void __meminit setup_per_zone_inactive_ratio(void)
6505 calculate_zone_inactive_ratio(zone);
6509 * Initialise min_free_kbytes.
6511 * For small machines we want it small (128k min). For large machines
6512 * we want it large (64MB max). But it is not linear, because network
6513 * bandwidth does not increase linearly with machine size. We use
6515 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6516 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6532 int __meminit init_per_zone_wmark_min(void)
6534 unsigned long lowmem_kbytes;
6535 int new_min_free_kbytes;
6537 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6538 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6540 if (new_min_free_kbytes > user_min_free_kbytes) {
6541 min_free_kbytes = new_min_free_kbytes;
6542 if (min_free_kbytes < 128)
6543 min_free_kbytes = 128;
6544 if (min_free_kbytes > 65536)
6545 min_free_kbytes = 65536;
6547 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6548 new_min_free_kbytes, user_min_free_kbytes);
6550 setup_per_zone_wmarks();
6551 refresh_zone_stat_thresholds();
6552 setup_per_zone_lowmem_reserve();
6553 setup_per_zone_inactive_ratio();
6556 core_initcall(init_per_zone_wmark_min)
6559 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6560 * that we can call two helper functions whenever min_free_kbytes
6563 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6564 void __user *buffer, size_t *length, loff_t *ppos)
6568 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6573 user_min_free_kbytes = min_free_kbytes;
6574 setup_per_zone_wmarks();
6579 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6580 void __user *buffer, size_t *length, loff_t *ppos)
6584 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6589 setup_per_zone_wmarks();
6595 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6596 void __user *buffer, size_t *length, loff_t *ppos)
6601 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6606 zone->min_unmapped_pages = (zone->managed_pages *
6607 sysctl_min_unmapped_ratio) / 100;
6611 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6612 void __user *buffer, size_t *length, loff_t *ppos)
6617 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6622 zone->min_slab_pages = (zone->managed_pages *
6623 sysctl_min_slab_ratio) / 100;
6629 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6630 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6631 * whenever sysctl_lowmem_reserve_ratio changes.
6633 * The reserve ratio obviously has absolutely no relation with the
6634 * minimum watermarks. The lowmem reserve ratio can only make sense
6635 * if in function of the boot time zone sizes.
6637 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6638 void __user *buffer, size_t *length, loff_t *ppos)
6640 proc_dointvec_minmax(table, write, buffer, length, ppos);
6641 setup_per_zone_lowmem_reserve();
6646 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6647 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6648 * pagelist can have before it gets flushed back to buddy allocator.
6650 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6651 void __user *buffer, size_t *length, loff_t *ppos)
6654 int old_percpu_pagelist_fraction;
6657 mutex_lock(&pcp_batch_high_lock);
6658 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6660 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6661 if (!write || ret < 0)
6664 /* Sanity checking to avoid pcp imbalance */
6665 if (percpu_pagelist_fraction &&
6666 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6667 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6673 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6676 for_each_populated_zone(zone) {
6679 for_each_possible_cpu(cpu)
6680 pageset_set_high_and_batch(zone,
6681 per_cpu_ptr(zone->pageset, cpu));
6684 mutex_unlock(&pcp_batch_high_lock);
6689 int hashdist = HASHDIST_DEFAULT;
6691 static int __init set_hashdist(char *str)
6695 hashdist = simple_strtoul(str, &str, 0);
6698 __setup("hashdist=", set_hashdist);
6702 * allocate a large system hash table from bootmem
6703 * - it is assumed that the hash table must contain an exact power-of-2
6704 * quantity of entries
6705 * - limit is the number of hash buckets, not the total allocation size
6707 void *__init alloc_large_system_hash(const char *tablename,
6708 unsigned long bucketsize,
6709 unsigned long numentries,
6712 unsigned int *_hash_shift,
6713 unsigned int *_hash_mask,
6714 unsigned long low_limit,
6715 unsigned long high_limit)
6717 unsigned long long max = high_limit;
6718 unsigned long log2qty, size;
6721 /* allow the kernel cmdline to have a say */
6723 /* round applicable memory size up to nearest megabyte */
6724 numentries = nr_kernel_pages;
6726 /* It isn't necessary when PAGE_SIZE >= 1MB */
6727 if (PAGE_SHIFT < 20)
6728 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6730 /* limit to 1 bucket per 2^scale bytes of low memory */
6731 if (scale > PAGE_SHIFT)
6732 numentries >>= (scale - PAGE_SHIFT);
6734 numentries <<= (PAGE_SHIFT - scale);
6736 /* Make sure we've got at least a 0-order allocation.. */
6737 if (unlikely(flags & HASH_SMALL)) {
6738 /* Makes no sense without HASH_EARLY */
6739 WARN_ON(!(flags & HASH_EARLY));
6740 if (!(numentries >> *_hash_shift)) {
6741 numentries = 1UL << *_hash_shift;
6742 BUG_ON(!numentries);
6744 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6745 numentries = PAGE_SIZE / bucketsize;
6747 numentries = roundup_pow_of_two(numentries);
6749 /* limit allocation size to 1/16 total memory by default */
6751 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6752 do_div(max, bucketsize);
6754 max = min(max, 0x80000000ULL);
6756 if (numentries < low_limit)
6757 numentries = low_limit;
6758 if (numentries > max)
6761 log2qty = ilog2(numentries);
6764 size = bucketsize << log2qty;
6765 if (flags & HASH_EARLY)
6766 table = memblock_virt_alloc_nopanic(size, 0);
6768 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6771 * If bucketsize is not a power-of-two, we may free
6772 * some pages at the end of hash table which
6773 * alloc_pages_exact() automatically does
6775 if (get_order(size) < MAX_ORDER) {
6776 table = alloc_pages_exact(size, GFP_ATOMIC);
6777 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6780 } while (!table && size > PAGE_SIZE && --log2qty);
6783 panic("Failed to allocate %s hash table\n", tablename);
6785 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
6786 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
6789 *_hash_shift = log2qty;
6791 *_hash_mask = (1 << log2qty) - 1;
6796 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6797 static inline unsigned long *get_pageblock_bitmap(struct page *page,
6800 #ifdef CONFIG_SPARSEMEM
6801 return __pfn_to_section(pfn)->pageblock_flags;
6803 return page_zone(page)->pageblock_flags;
6804 #endif /* CONFIG_SPARSEMEM */
6807 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
6809 #ifdef CONFIG_SPARSEMEM
6810 pfn &= (PAGES_PER_SECTION-1);
6811 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6813 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
6814 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6815 #endif /* CONFIG_SPARSEMEM */
6819 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6820 * @page: The page within the block of interest
6821 * @pfn: The target page frame number
6822 * @end_bitidx: The last bit of interest to retrieve
6823 * @mask: mask of bits that the caller is interested in
6825 * Return: pageblock_bits flags
6827 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6828 unsigned long end_bitidx,
6831 unsigned long *bitmap;
6832 unsigned long bitidx, word_bitidx;
6835 bitmap = get_pageblock_bitmap(page, pfn);
6836 bitidx = pfn_to_bitidx(page, pfn);
6837 word_bitidx = bitidx / BITS_PER_LONG;
6838 bitidx &= (BITS_PER_LONG-1);
6840 word = bitmap[word_bitidx];
6841 bitidx += end_bitidx;
6842 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6846 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6847 * @page: The page within the block of interest
6848 * @flags: The flags to set
6849 * @pfn: The target page frame number
6850 * @end_bitidx: The last bit of interest
6851 * @mask: mask of bits that the caller is interested in
6853 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6855 unsigned long end_bitidx,
6858 unsigned long *bitmap;
6859 unsigned long bitidx, word_bitidx;
6860 unsigned long old_word, word;
6862 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6864 bitmap = get_pageblock_bitmap(page, pfn);
6865 bitidx = pfn_to_bitidx(page, pfn);
6866 word_bitidx = bitidx / BITS_PER_LONG;
6867 bitidx &= (BITS_PER_LONG-1);
6869 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
6871 bitidx += end_bitidx;
6872 mask <<= (BITS_PER_LONG - bitidx - 1);
6873 flags <<= (BITS_PER_LONG - bitidx - 1);
6875 word = READ_ONCE(bitmap[word_bitidx]);
6877 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6878 if (word == old_word)
6885 * This function checks whether pageblock includes unmovable pages or not.
6886 * If @count is not zero, it is okay to include less @count unmovable pages
6888 * PageLRU check without isolation or lru_lock could race so that
6889 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6890 * expect this function should be exact.
6892 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6893 bool skip_hwpoisoned_pages)
6895 unsigned long pfn, iter, found;
6899 * For avoiding noise data, lru_add_drain_all() should be called
6900 * If ZONE_MOVABLE, the zone never contains unmovable pages
6902 if (zone_idx(zone) == ZONE_MOVABLE)
6904 mt = get_pageblock_migratetype(page);
6905 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6908 pfn = page_to_pfn(page);
6909 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6910 unsigned long check = pfn + iter;
6912 if (!pfn_valid_within(check))
6915 page = pfn_to_page(check);
6918 * Hugepages are not in LRU lists, but they're movable.
6919 * We need not scan over tail pages bacause we don't
6920 * handle each tail page individually in migration.
6922 if (PageHuge(page)) {
6923 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6928 * We can't use page_count without pin a page
6929 * because another CPU can free compound page.
6930 * This check already skips compound tails of THP
6931 * because their page->_refcount is zero at all time.
6933 if (!page_ref_count(page)) {
6934 if (PageBuddy(page))
6935 iter += (1 << page_order(page)) - 1;
6940 * The HWPoisoned page may be not in buddy system, and
6941 * page_count() is not 0.
6943 if (skip_hwpoisoned_pages && PageHWPoison(page))
6949 * If there are RECLAIMABLE pages, we need to check
6950 * it. But now, memory offline itself doesn't call
6951 * shrink_node_slabs() and it still to be fixed.
6954 * If the page is not RAM, page_count()should be 0.
6955 * we don't need more check. This is an _used_ not-movable page.
6957 * The problematic thing here is PG_reserved pages. PG_reserved
6958 * is set to both of a memory hole page and a _used_ kernel
6967 bool is_pageblock_removable_nolock(struct page *page)
6973 * We have to be careful here because we are iterating over memory
6974 * sections which are not zone aware so we might end up outside of
6975 * the zone but still within the section.
6976 * We have to take care about the node as well. If the node is offline
6977 * its NODE_DATA will be NULL - see page_zone.
6979 if (!node_online(page_to_nid(page)))
6982 zone = page_zone(page);
6983 pfn = page_to_pfn(page);
6984 if (!zone_spans_pfn(zone, pfn))
6987 return !has_unmovable_pages(zone, page, 0, true);
6990 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
6992 static unsigned long pfn_max_align_down(unsigned long pfn)
6994 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6995 pageblock_nr_pages) - 1);
6998 static unsigned long pfn_max_align_up(unsigned long pfn)
7000 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7001 pageblock_nr_pages));
7004 /* [start, end) must belong to a single zone. */
7005 static int __alloc_contig_migrate_range(struct compact_control *cc,
7006 unsigned long start, unsigned long end)
7008 /* This function is based on compact_zone() from compaction.c. */
7009 unsigned long nr_reclaimed;
7010 unsigned long pfn = start;
7011 unsigned int tries = 0;
7016 while (pfn < end || !list_empty(&cc->migratepages)) {
7017 if (fatal_signal_pending(current)) {
7022 if (list_empty(&cc->migratepages)) {
7023 cc->nr_migratepages = 0;
7024 pfn = isolate_migratepages_range(cc, pfn, end);
7030 } else if (++tries == 5) {
7031 ret = ret < 0 ? ret : -EBUSY;
7035 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7037 cc->nr_migratepages -= nr_reclaimed;
7039 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7040 NULL, 0, cc->mode, MR_CMA);
7043 putback_movable_pages(&cc->migratepages);
7050 * alloc_contig_range() -- tries to allocate given range of pages
7051 * @start: start PFN to allocate
7052 * @end: one-past-the-last PFN to allocate
7053 * @migratetype: migratetype of the underlaying pageblocks (either
7054 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7055 * in range must have the same migratetype and it must
7056 * be either of the two.
7058 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7059 * aligned, however it's the caller's responsibility to guarantee that
7060 * we are the only thread that changes migrate type of pageblocks the
7063 * The PFN range must belong to a single zone.
7065 * Returns zero on success or negative error code. On success all
7066 * pages which PFN is in [start, end) are allocated for the caller and
7067 * need to be freed with free_contig_range().
7069 int alloc_contig_range(unsigned long start, unsigned long end,
7070 unsigned migratetype)
7072 unsigned long outer_start, outer_end;
7076 struct compact_control cc = {
7077 .nr_migratepages = 0,
7079 .zone = page_zone(pfn_to_page(start)),
7080 .mode = MIGRATE_SYNC,
7081 .ignore_skip_hint = true,
7083 INIT_LIST_HEAD(&cc.migratepages);
7086 * What we do here is we mark all pageblocks in range as
7087 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7088 * have different sizes, and due to the way page allocator
7089 * work, we align the range to biggest of the two pages so
7090 * that page allocator won't try to merge buddies from
7091 * different pageblocks and change MIGRATE_ISOLATE to some
7092 * other migration type.
7094 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7095 * migrate the pages from an unaligned range (ie. pages that
7096 * we are interested in). This will put all the pages in
7097 * range back to page allocator as MIGRATE_ISOLATE.
7099 * When this is done, we take the pages in range from page
7100 * allocator removing them from the buddy system. This way
7101 * page allocator will never consider using them.
7103 * This lets us mark the pageblocks back as
7104 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7105 * aligned range but not in the unaligned, original range are
7106 * put back to page allocator so that buddy can use them.
7109 ret = start_isolate_page_range(pfn_max_align_down(start),
7110 pfn_max_align_up(end), migratetype,
7116 * In case of -EBUSY, we'd like to know which page causes problem.
7117 * So, just fall through. We will check it in test_pages_isolated().
7119 ret = __alloc_contig_migrate_range(&cc, start, end);
7120 if (ret && ret != -EBUSY)
7124 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7125 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7126 * more, all pages in [start, end) are free in page allocator.
7127 * What we are going to do is to allocate all pages from
7128 * [start, end) (that is remove them from page allocator).
7130 * The only problem is that pages at the beginning and at the
7131 * end of interesting range may be not aligned with pages that
7132 * page allocator holds, ie. they can be part of higher order
7133 * pages. Because of this, we reserve the bigger range and
7134 * once this is done free the pages we are not interested in.
7136 * We don't have to hold zone->lock here because the pages are
7137 * isolated thus they won't get removed from buddy.
7140 lru_add_drain_all();
7141 drain_all_pages(cc.zone);
7144 outer_start = start;
7145 while (!PageBuddy(pfn_to_page(outer_start))) {
7146 if (++order >= MAX_ORDER) {
7147 outer_start = start;
7150 outer_start &= ~0UL << order;
7153 if (outer_start != start) {
7154 order = page_order(pfn_to_page(outer_start));
7157 * outer_start page could be small order buddy page and
7158 * it doesn't include start page. Adjust outer_start
7159 * in this case to report failed page properly
7160 * on tracepoint in test_pages_isolated()
7162 if (outer_start + (1UL << order) <= start)
7163 outer_start = start;
7166 /* Make sure the range is really isolated. */
7167 if (test_pages_isolated(outer_start, end, false)) {
7168 pr_info("%s: [%lx, %lx) PFNs busy\n",
7169 __func__, outer_start, end);
7174 /* Grab isolated pages from freelists. */
7175 outer_end = isolate_freepages_range(&cc, outer_start, end);
7181 /* Free head and tail (if any) */
7182 if (start != outer_start)
7183 free_contig_range(outer_start, start - outer_start);
7184 if (end != outer_end)
7185 free_contig_range(end, outer_end - end);
7188 undo_isolate_page_range(pfn_max_align_down(start),
7189 pfn_max_align_up(end), migratetype);
7193 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7195 unsigned int count = 0;
7197 for (; nr_pages--; pfn++) {
7198 struct page *page = pfn_to_page(pfn);
7200 count += page_count(page) != 1;
7203 WARN(count != 0, "%d pages are still in use!\n", count);
7207 #ifdef CONFIG_MEMORY_HOTPLUG
7209 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7210 * page high values need to be recalulated.
7212 void __meminit zone_pcp_update(struct zone *zone)
7215 mutex_lock(&pcp_batch_high_lock);
7216 for_each_possible_cpu(cpu)
7217 pageset_set_high_and_batch(zone,
7218 per_cpu_ptr(zone->pageset, cpu));
7219 mutex_unlock(&pcp_batch_high_lock);
7223 void zone_pcp_reset(struct zone *zone)
7225 unsigned long flags;
7227 struct per_cpu_pageset *pset;
7229 /* avoid races with drain_pages() */
7230 local_irq_save(flags);
7231 if (zone->pageset != &boot_pageset) {
7232 for_each_online_cpu(cpu) {
7233 pset = per_cpu_ptr(zone->pageset, cpu);
7234 drain_zonestat(zone, pset);
7236 free_percpu(zone->pageset);
7237 zone->pageset = &boot_pageset;
7239 local_irq_restore(flags);
7242 #ifdef CONFIG_MEMORY_HOTREMOVE
7244 * All pages in the range must be in a single zone and isolated
7245 * before calling this.
7248 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7252 unsigned int order, i;
7254 unsigned long flags;
7255 /* find the first valid pfn */
7256 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7261 zone = page_zone(pfn_to_page(pfn));
7262 spin_lock_irqsave(&zone->lock, flags);
7264 while (pfn < end_pfn) {
7265 if (!pfn_valid(pfn)) {
7269 page = pfn_to_page(pfn);
7271 * The HWPoisoned page may be not in buddy system, and
7272 * page_count() is not 0.
7274 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7276 SetPageReserved(page);
7280 BUG_ON(page_count(page));
7281 BUG_ON(!PageBuddy(page));
7282 order = page_order(page);
7283 #ifdef CONFIG_DEBUG_VM
7284 pr_info("remove from free list %lx %d %lx\n",
7285 pfn, 1 << order, end_pfn);
7287 list_del(&page->lru);
7288 rmv_page_order(page);
7289 zone->free_area[order].nr_free--;
7290 for (i = 0; i < (1 << order); i++)
7291 SetPageReserved((page+i));
7292 pfn += (1 << order);
7294 spin_unlock_irqrestore(&zone->lock, flags);
7298 bool is_free_buddy_page(struct page *page)
7300 struct zone *zone = page_zone(page);
7301 unsigned long pfn = page_to_pfn(page);
7302 unsigned long flags;
7305 spin_lock_irqsave(&zone->lock, flags);
7306 for (order = 0; order < MAX_ORDER; order++) {
7307 struct page *page_head = page - (pfn & ((1 << order) - 1));
7309 if (PageBuddy(page_head) && page_order(page_head) >= order)
7312 spin_unlock_irqrestore(&zone->lock, flags);
7314 return order < MAX_ORDER;