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;
835 spin_lock(&zone->lock);
836 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
838 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
842 struct list_head *list;
845 * Remove pages from lists in a round-robin fashion. A
846 * batch_free count is maintained that is incremented when an
847 * empty list is encountered. This is so more pages are freed
848 * off fuller lists instead of spinning excessively around empty
853 if (++migratetype == MIGRATE_PCPTYPES)
855 list = &pcp->lists[migratetype];
856 } while (list_empty(list));
858 /* This is the only non-empty list. Free them all. */
859 if (batch_free == MIGRATE_PCPTYPES)
860 batch_free = to_free;
863 int mt; /* migratetype of the to-be-freed page */
865 page = list_last_entry(list, struct page, lru);
866 /* must delete as __free_one_page list manipulates */
867 list_del(&page->lru);
869 mt = get_pcppage_migratetype(page);
870 /* MIGRATE_ISOLATE page should not go to pcplists */
871 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
872 /* Pageblock could have been isolated meanwhile */
873 if (unlikely(has_isolate_pageblock(zone)))
874 mt = get_pageblock_migratetype(page);
876 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
877 trace_mm_page_pcpu_drain(page, 0, mt);
878 } while (--to_free && --batch_free && !list_empty(list));
880 spin_unlock(&zone->lock);
883 static void free_one_page(struct zone *zone,
884 struct page *page, unsigned long pfn,
888 unsigned long nr_scanned;
889 spin_lock(&zone->lock);
890 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
892 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
894 if (unlikely(has_isolate_pageblock(zone) ||
895 is_migrate_isolate(migratetype))) {
896 migratetype = get_pfnblock_migratetype(page, pfn);
898 __free_one_page(page, pfn, zone, order, migratetype);
899 spin_unlock(&zone->lock);
902 static int free_tail_pages_check(struct page *head_page, struct page *page)
907 * We rely page->lru.next never has bit 0 set, unless the page
908 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
910 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
912 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
916 switch (page - head_page) {
918 /* the first tail page: ->mapping is compound_mapcount() */
919 if (unlikely(compound_mapcount(page))) {
920 bad_page(page, "nonzero compound_mapcount", 0);
926 * the second tail page: ->mapping is
927 * page_deferred_list().next -- ignore value.
931 if (page->mapping != TAIL_MAPPING) {
932 bad_page(page, "corrupted mapping in tail page", 0);
937 if (unlikely(!PageTail(page))) {
938 bad_page(page, "PageTail not set", 0);
941 if (unlikely(compound_head(page) != head_page)) {
942 bad_page(page, "compound_head not consistent", 0);
947 page->mapping = NULL;
948 clear_compound_head(page);
952 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
953 unsigned long zone, int nid)
955 set_page_links(page, zone, nid, pfn);
956 init_page_count(page);
957 page_mapcount_reset(page);
958 page_cpupid_reset_last(page);
960 INIT_LIST_HEAD(&page->lru);
961 #ifdef WANT_PAGE_VIRTUAL
962 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
963 if (!is_highmem_idx(zone))
964 set_page_address(page, __va(pfn << PAGE_SHIFT));
968 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
971 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
974 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
975 static void init_reserved_page(unsigned long pfn)
980 if (!early_page_uninitialised(pfn))
983 nid = early_pfn_to_nid(pfn);
984 pgdat = NODE_DATA(nid);
986 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
987 struct zone *zone = &pgdat->node_zones[zid];
989 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
992 __init_single_pfn(pfn, zid, nid);
995 static inline void init_reserved_page(unsigned long pfn)
998 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1001 * Initialised pages do not have PageReserved set. This function is
1002 * called for each range allocated by the bootmem allocator and
1003 * marks the pages PageReserved. The remaining valid pages are later
1004 * sent to the buddy page allocator.
1006 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
1008 unsigned long start_pfn = PFN_DOWN(start);
1009 unsigned long end_pfn = PFN_UP(end);
1011 for (; start_pfn < end_pfn; start_pfn++) {
1012 if (pfn_valid(start_pfn)) {
1013 struct page *page = pfn_to_page(start_pfn);
1015 init_reserved_page(start_pfn);
1017 /* Avoid false-positive PageTail() */
1018 INIT_LIST_HEAD(&page->lru);
1020 SetPageReserved(page);
1025 static bool free_pages_prepare(struct page *page, unsigned int order)
1029 VM_BUG_ON_PAGE(PageTail(page), page);
1031 trace_mm_page_free(page, order);
1032 kmemcheck_free_shadow(page, order);
1033 kasan_free_pages(page, order);
1036 * Check tail pages before head page information is cleared to
1037 * avoid checking PageCompound for order-0 pages.
1039 if (unlikely(order)) {
1040 bool compound = PageCompound(page);
1043 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1045 for (i = 1; i < (1 << order); i++) {
1047 bad += free_tail_pages_check(page, page + i);
1048 bad += free_pages_check(page + i);
1051 if (PageAnonHead(page))
1052 page->mapping = NULL;
1053 bad += free_pages_check(page);
1057 reset_page_owner(page, order);
1059 if (!PageHighMem(page)) {
1060 debug_check_no_locks_freed(page_address(page),
1061 PAGE_SIZE << order);
1062 debug_check_no_obj_freed(page_address(page),
1063 PAGE_SIZE << order);
1065 arch_free_page(page, order);
1066 kernel_poison_pages(page, 1 << order, 0);
1067 kernel_map_pages(page, 1 << order, 0);
1072 static void __free_pages_ok(struct page *page, unsigned int order)
1074 unsigned long flags;
1076 unsigned long pfn = page_to_pfn(page);
1078 if (!free_pages_prepare(page, order))
1081 migratetype = get_pfnblock_migratetype(page, pfn);
1082 local_irq_save(flags);
1083 __count_vm_events(PGFREE, 1 << order);
1084 free_one_page(page_zone(page), page, pfn, order, migratetype);
1085 local_irq_restore(flags);
1088 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1090 unsigned int nr_pages = 1 << order;
1091 struct page *p = page;
1095 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1097 __ClearPageReserved(p);
1098 set_page_count(p, 0);
1100 __ClearPageReserved(p);
1101 set_page_count(p, 0);
1103 page_zone(page)->managed_pages += nr_pages;
1104 set_page_refcounted(page);
1105 __free_pages(page, order);
1108 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1109 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1111 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1113 int __meminit early_pfn_to_nid(unsigned long pfn)
1115 static DEFINE_SPINLOCK(early_pfn_lock);
1118 spin_lock(&early_pfn_lock);
1119 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1122 spin_unlock(&early_pfn_lock);
1128 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1129 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1130 struct mminit_pfnnid_cache *state)
1134 nid = __early_pfn_to_nid(pfn, state);
1135 if (nid >= 0 && nid != node)
1140 /* Only safe to use early in boot when initialisation is single-threaded */
1141 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1143 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1148 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1152 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1153 struct mminit_pfnnid_cache *state)
1160 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1163 if (early_page_uninitialised(pfn))
1165 return __free_pages_boot_core(page, order);
1169 * Check that the whole (or subset of) a pageblock given by the interval of
1170 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1171 * with the migration of free compaction scanner. The scanners then need to
1172 * use only pfn_valid_within() check for arches that allow holes within
1175 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1177 * It's possible on some configurations to have a setup like node0 node1 node0
1178 * i.e. it's possible that all pages within a zones range of pages do not
1179 * belong to a single zone. We assume that a border between node0 and node1
1180 * can occur within a single pageblock, but not a node0 node1 node0
1181 * interleaving within a single pageblock. It is therefore sufficient to check
1182 * the first and last page of a pageblock and avoid checking each individual
1183 * page in a pageblock.
1185 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1186 unsigned long end_pfn, struct zone *zone)
1188 struct page *start_page;
1189 struct page *end_page;
1191 /* end_pfn is one past the range we are checking */
1194 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1197 start_page = pfn_to_page(start_pfn);
1199 if (page_zone(start_page) != zone)
1202 end_page = pfn_to_page(end_pfn);
1204 /* This gives a shorter code than deriving page_zone(end_page) */
1205 if (page_zone_id(start_page) != page_zone_id(end_page))
1211 void set_zone_contiguous(struct zone *zone)
1213 unsigned long block_start_pfn = zone->zone_start_pfn;
1214 unsigned long block_end_pfn;
1216 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1217 for (; block_start_pfn < zone_end_pfn(zone);
1218 block_start_pfn = block_end_pfn,
1219 block_end_pfn += pageblock_nr_pages) {
1221 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1223 if (!__pageblock_pfn_to_page(block_start_pfn,
1224 block_end_pfn, zone))
1228 /* We confirm that there is no hole */
1229 zone->contiguous = true;
1232 void clear_zone_contiguous(struct zone *zone)
1234 zone->contiguous = false;
1237 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1238 static void __init deferred_free_range(struct page *page,
1239 unsigned long pfn, int nr_pages)
1246 /* Free a large naturally-aligned chunk if possible */
1247 if (nr_pages == MAX_ORDER_NR_PAGES &&
1248 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1249 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1250 __free_pages_boot_core(page, MAX_ORDER-1);
1254 for (i = 0; i < nr_pages; i++, page++)
1255 __free_pages_boot_core(page, 0);
1258 /* Completion tracking for deferred_init_memmap() threads */
1259 static atomic_t pgdat_init_n_undone __initdata;
1260 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1262 static inline void __init pgdat_init_report_one_done(void)
1264 if (atomic_dec_and_test(&pgdat_init_n_undone))
1265 complete(&pgdat_init_all_done_comp);
1268 /* Initialise remaining memory on a node */
1269 static int __init deferred_init_memmap(void *data)
1271 pg_data_t *pgdat = data;
1272 int nid = pgdat->node_id;
1273 struct mminit_pfnnid_cache nid_init_state = { };
1274 unsigned long start = jiffies;
1275 unsigned long nr_pages = 0;
1276 unsigned long walk_start, walk_end;
1279 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1280 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1282 if (first_init_pfn == ULONG_MAX) {
1283 pgdat_init_report_one_done();
1287 /* Bind memory initialisation thread to a local node if possible */
1288 if (!cpumask_empty(cpumask))
1289 set_cpus_allowed_ptr(current, cpumask);
1291 /* Sanity check boundaries */
1292 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1293 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1294 pgdat->first_deferred_pfn = ULONG_MAX;
1296 /* Only the highest zone is deferred so find it */
1297 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1298 zone = pgdat->node_zones + zid;
1299 if (first_init_pfn < zone_end_pfn(zone))
1303 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1304 unsigned long pfn, end_pfn;
1305 struct page *page = NULL;
1306 struct page *free_base_page = NULL;
1307 unsigned long free_base_pfn = 0;
1310 end_pfn = min(walk_end, zone_end_pfn(zone));
1311 pfn = first_init_pfn;
1312 if (pfn < walk_start)
1314 if (pfn < zone->zone_start_pfn)
1315 pfn = zone->zone_start_pfn;
1317 for (; pfn < end_pfn; pfn++) {
1318 if (!pfn_valid_within(pfn))
1322 * Ensure pfn_valid is checked every
1323 * MAX_ORDER_NR_PAGES for memory holes
1325 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1326 if (!pfn_valid(pfn)) {
1332 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1337 /* Minimise pfn page lookups and scheduler checks */
1338 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1341 nr_pages += nr_to_free;
1342 deferred_free_range(free_base_page,
1343 free_base_pfn, nr_to_free);
1344 free_base_page = NULL;
1345 free_base_pfn = nr_to_free = 0;
1347 page = pfn_to_page(pfn);
1352 VM_BUG_ON(page_zone(page) != zone);
1356 __init_single_page(page, pfn, zid, nid);
1357 if (!free_base_page) {
1358 free_base_page = page;
1359 free_base_pfn = pfn;
1364 /* Where possible, batch up pages for a single free */
1367 /* Free the current block of pages to allocator */
1368 nr_pages += nr_to_free;
1369 deferred_free_range(free_base_page, free_base_pfn,
1371 free_base_page = NULL;
1372 free_base_pfn = nr_to_free = 0;
1375 first_init_pfn = max(end_pfn, first_init_pfn);
1378 /* Sanity check that the next zone really is unpopulated */
1379 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1381 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1382 jiffies_to_msecs(jiffies - start));
1384 pgdat_init_report_one_done();
1387 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1389 void __init page_alloc_init_late(void)
1393 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1396 /* There will be num_node_state(N_MEMORY) threads */
1397 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1398 for_each_node_state(nid, N_MEMORY) {
1399 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1402 /* Block until all are initialised */
1403 wait_for_completion(&pgdat_init_all_done_comp);
1405 /* Reinit limits that are based on free pages after the kernel is up */
1406 files_maxfiles_init();
1409 for_each_populated_zone(zone)
1410 set_zone_contiguous(zone);
1414 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1415 void __init init_cma_reserved_pageblock(struct page *page)
1417 unsigned i = pageblock_nr_pages;
1418 struct page *p = page;
1421 __ClearPageReserved(p);
1422 set_page_count(p, 0);
1425 set_pageblock_migratetype(page, MIGRATE_CMA);
1427 if (pageblock_order >= MAX_ORDER) {
1428 i = pageblock_nr_pages;
1431 set_page_refcounted(p);
1432 __free_pages(p, MAX_ORDER - 1);
1433 p += MAX_ORDER_NR_PAGES;
1434 } while (i -= MAX_ORDER_NR_PAGES);
1436 set_page_refcounted(page);
1437 __free_pages(page, pageblock_order);
1440 adjust_managed_page_count(page, pageblock_nr_pages);
1445 * The order of subdivision here is critical for the IO subsystem.
1446 * Please do not alter this order without good reasons and regression
1447 * testing. Specifically, as large blocks of memory are subdivided,
1448 * the order in which smaller blocks are delivered depends on the order
1449 * they're subdivided in this function. This is the primary factor
1450 * influencing the order in which pages are delivered to the IO
1451 * subsystem according to empirical testing, and this is also justified
1452 * by considering the behavior of a buddy system containing a single
1453 * large block of memory acted on by a series of small allocations.
1454 * This behavior is a critical factor in sglist merging's success.
1458 static inline void expand(struct zone *zone, struct page *page,
1459 int low, int high, struct free_area *area,
1462 unsigned long size = 1 << high;
1464 while (high > low) {
1468 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1470 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1471 debug_guardpage_enabled() &&
1472 high < debug_guardpage_minorder()) {
1474 * Mark as guard pages (or page), that will allow to
1475 * merge back to allocator when buddy will be freed.
1476 * Corresponding page table entries will not be touched,
1477 * pages will stay not present in virtual address space
1479 set_page_guard(zone, &page[size], high, migratetype);
1482 list_add(&page[size].lru, &area->free_list[migratetype]);
1484 set_page_order(&page[size], high);
1489 * This page is about to be returned from the page allocator
1491 static inline int check_new_page(struct page *page)
1493 const char *bad_reason = NULL;
1494 unsigned long bad_flags = 0;
1496 if (unlikely(atomic_read(&page->_mapcount) != -1))
1497 bad_reason = "nonzero mapcount";
1498 if (unlikely(page->mapping != NULL))
1499 bad_reason = "non-NULL mapping";
1500 if (unlikely(page_ref_count(page) != 0))
1501 bad_reason = "nonzero _count";
1502 if (unlikely(page->flags & __PG_HWPOISON)) {
1503 bad_reason = "HWPoisoned (hardware-corrupted)";
1504 bad_flags = __PG_HWPOISON;
1506 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1507 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1508 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1511 if (unlikely(page->mem_cgroup))
1512 bad_reason = "page still charged to cgroup";
1514 if (unlikely(bad_reason)) {
1515 bad_page(page, bad_reason, bad_flags);
1521 static inline bool free_pages_prezeroed(bool poisoned)
1523 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1524 page_poisoning_enabled() && poisoned;
1527 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1531 bool poisoned = true;
1533 for (i = 0; i < (1 << order); i++) {
1534 struct page *p = page + i;
1535 if (unlikely(check_new_page(p)))
1538 poisoned &= page_is_poisoned(p);
1541 set_page_private(page, 0);
1542 set_page_refcounted(page);
1544 arch_alloc_page(page, order);
1545 kernel_map_pages(page, 1 << order, 1);
1546 kernel_poison_pages(page, 1 << order, 1);
1547 kasan_alloc_pages(page, order);
1549 if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO))
1550 for (i = 0; i < (1 << order); i++)
1551 clear_highpage(page + i);
1553 if (order && (gfp_flags & __GFP_COMP))
1554 prep_compound_page(page, order);
1556 set_page_owner(page, order, gfp_flags);
1559 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1560 * allocate the page. The expectation is that the caller is taking
1561 * steps that will free more memory. The caller should avoid the page
1562 * being used for !PFMEMALLOC purposes.
1564 if (alloc_flags & ALLOC_NO_WATERMARKS)
1565 set_page_pfmemalloc(page);
1567 clear_page_pfmemalloc(page);
1573 * Go through the free lists for the given migratetype and remove
1574 * the smallest available page from the freelists
1577 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1580 unsigned int current_order;
1581 struct free_area *area;
1584 /* Find a page of the appropriate size in the preferred list */
1585 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1586 area = &(zone->free_area[current_order]);
1587 page = list_first_entry_or_null(&area->free_list[migratetype],
1591 list_del(&page->lru);
1592 rmv_page_order(page);
1594 expand(zone, page, order, current_order, area, migratetype);
1595 set_pcppage_migratetype(page, migratetype);
1604 * This array describes the order lists are fallen back to when
1605 * the free lists for the desirable migrate type are depleted
1607 static int fallbacks[MIGRATE_TYPES][4] = {
1608 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1609 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1610 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1612 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1614 #ifdef CONFIG_MEMORY_ISOLATION
1615 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1620 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1623 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1626 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1627 unsigned int order) { return NULL; }
1631 * Move the free pages in a range to the free lists of the requested type.
1632 * Note that start_page and end_pages are not aligned on a pageblock
1633 * boundary. If alignment is required, use move_freepages_block()
1635 int move_freepages(struct zone *zone,
1636 struct page *start_page, struct page *end_page,
1641 int pages_moved = 0;
1643 #ifndef CONFIG_HOLES_IN_ZONE
1645 * page_zone is not safe to call in this context when
1646 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1647 * anyway as we check zone boundaries in move_freepages_block().
1648 * Remove at a later date when no bug reports exist related to
1649 * grouping pages by mobility
1651 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1654 for (page = start_page; page <= end_page;) {
1655 /* Make sure we are not inadvertently changing nodes */
1656 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1658 if (!pfn_valid_within(page_to_pfn(page))) {
1663 if (!PageBuddy(page)) {
1668 order = page_order(page);
1669 list_move(&page->lru,
1670 &zone->free_area[order].free_list[migratetype]);
1672 pages_moved += 1 << order;
1678 int move_freepages_block(struct zone *zone, struct page *page,
1681 unsigned long start_pfn, end_pfn;
1682 struct page *start_page, *end_page;
1684 start_pfn = page_to_pfn(page);
1685 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1686 start_page = pfn_to_page(start_pfn);
1687 end_page = start_page + pageblock_nr_pages - 1;
1688 end_pfn = start_pfn + pageblock_nr_pages - 1;
1690 /* Do not cross zone boundaries */
1691 if (!zone_spans_pfn(zone, start_pfn))
1693 if (!zone_spans_pfn(zone, end_pfn))
1696 return move_freepages(zone, start_page, end_page, migratetype);
1699 static void change_pageblock_range(struct page *pageblock_page,
1700 int start_order, int migratetype)
1702 int nr_pageblocks = 1 << (start_order - pageblock_order);
1704 while (nr_pageblocks--) {
1705 set_pageblock_migratetype(pageblock_page, migratetype);
1706 pageblock_page += pageblock_nr_pages;
1711 * When we are falling back to another migratetype during allocation, try to
1712 * steal extra free pages from the same pageblocks to satisfy further
1713 * allocations, instead of polluting multiple pageblocks.
1715 * If we are stealing a relatively large buddy page, it is likely there will
1716 * be more free pages in the pageblock, so try to steal them all. For
1717 * reclaimable and unmovable allocations, we steal regardless of page size,
1718 * as fragmentation caused by those allocations polluting movable pageblocks
1719 * is worse than movable allocations stealing from unmovable and reclaimable
1722 static bool can_steal_fallback(unsigned int order, int start_mt)
1725 * Leaving this order check is intended, although there is
1726 * relaxed order check in next check. The reason is that
1727 * we can actually steal whole pageblock if this condition met,
1728 * but, below check doesn't guarantee it and that is just heuristic
1729 * so could be changed anytime.
1731 if (order >= pageblock_order)
1734 if (order >= pageblock_order / 2 ||
1735 start_mt == MIGRATE_RECLAIMABLE ||
1736 start_mt == MIGRATE_UNMOVABLE ||
1737 page_group_by_mobility_disabled)
1744 * This function implements actual steal behaviour. If order is large enough,
1745 * we can steal whole pageblock. If not, we first move freepages in this
1746 * pageblock and check whether half of pages are moved or not. If half of
1747 * pages are moved, we can change migratetype of pageblock and permanently
1748 * use it's pages as requested migratetype in the future.
1750 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1753 unsigned int current_order = page_order(page);
1756 /* Take ownership for orders >= pageblock_order */
1757 if (current_order >= pageblock_order) {
1758 change_pageblock_range(page, current_order, start_type);
1762 pages = move_freepages_block(zone, page, start_type);
1764 /* Claim the whole block if over half of it is free */
1765 if (pages >= (1 << (pageblock_order-1)) ||
1766 page_group_by_mobility_disabled)
1767 set_pageblock_migratetype(page, start_type);
1771 * Check whether there is a suitable fallback freepage with requested order.
1772 * If only_stealable is true, this function returns fallback_mt only if
1773 * we can steal other freepages all together. This would help to reduce
1774 * fragmentation due to mixed migratetype pages in one pageblock.
1776 int find_suitable_fallback(struct free_area *area, unsigned int order,
1777 int migratetype, bool only_stealable, bool *can_steal)
1782 if (area->nr_free == 0)
1787 fallback_mt = fallbacks[migratetype][i];
1788 if (fallback_mt == MIGRATE_TYPES)
1791 if (list_empty(&area->free_list[fallback_mt]))
1794 if (can_steal_fallback(order, migratetype))
1797 if (!only_stealable)
1808 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1809 * there are no empty page blocks that contain a page with a suitable order
1811 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1812 unsigned int alloc_order)
1815 unsigned long max_managed, flags;
1818 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1819 * Check is race-prone but harmless.
1821 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1822 if (zone->nr_reserved_highatomic >= max_managed)
1825 spin_lock_irqsave(&zone->lock, flags);
1827 /* Recheck the nr_reserved_highatomic limit under the lock */
1828 if (zone->nr_reserved_highatomic >= max_managed)
1832 mt = get_pageblock_migratetype(page);
1833 if (mt != MIGRATE_HIGHATOMIC &&
1834 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1835 zone->nr_reserved_highatomic += pageblock_nr_pages;
1836 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1837 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1841 spin_unlock_irqrestore(&zone->lock, flags);
1845 * Used when an allocation is about to fail under memory pressure. This
1846 * potentially hurts the reliability of high-order allocations when under
1847 * intense memory pressure but failed atomic allocations should be easier
1848 * to recover from than an OOM.
1850 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1852 struct zonelist *zonelist = ac->zonelist;
1853 unsigned long flags;
1859 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1861 /* Preserve at least one pageblock */
1862 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1865 spin_lock_irqsave(&zone->lock, flags);
1866 for (order = 0; order < MAX_ORDER; order++) {
1867 struct free_area *area = &(zone->free_area[order]);
1869 page = list_first_entry_or_null(
1870 &area->free_list[MIGRATE_HIGHATOMIC],
1876 * It should never happen but changes to locking could
1877 * inadvertently allow a per-cpu drain to add pages
1878 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1879 * and watch for underflows.
1881 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1882 zone->nr_reserved_highatomic);
1885 * Convert to ac->migratetype and avoid the normal
1886 * pageblock stealing heuristics. Minimally, the caller
1887 * is doing the work and needs the pages. More
1888 * importantly, if the block was always converted to
1889 * MIGRATE_UNMOVABLE or another type then the number
1890 * of pageblocks that cannot be completely freed
1893 set_pageblock_migratetype(page, ac->migratetype);
1894 move_freepages_block(zone, page, ac->migratetype);
1895 spin_unlock_irqrestore(&zone->lock, flags);
1898 spin_unlock_irqrestore(&zone->lock, flags);
1902 /* Remove an element from the buddy allocator from the fallback list */
1903 static inline struct page *
1904 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1906 struct free_area *area;
1907 unsigned int current_order;
1912 /* Find the largest possible block of pages in the other list */
1913 for (current_order = MAX_ORDER-1;
1914 current_order >= order && current_order <= MAX_ORDER-1;
1916 area = &(zone->free_area[current_order]);
1917 fallback_mt = find_suitable_fallback(area, current_order,
1918 start_migratetype, false, &can_steal);
1919 if (fallback_mt == -1)
1922 page = list_first_entry(&area->free_list[fallback_mt],
1925 steal_suitable_fallback(zone, page, start_migratetype);
1927 /* Remove the page from the freelists */
1929 list_del(&page->lru);
1930 rmv_page_order(page);
1932 expand(zone, page, order, current_order, area,
1935 * The pcppage_migratetype may differ from pageblock's
1936 * migratetype depending on the decisions in
1937 * find_suitable_fallback(). This is OK as long as it does not
1938 * differ for MIGRATE_CMA pageblocks. Those can be used as
1939 * fallback only via special __rmqueue_cma_fallback() function
1941 set_pcppage_migratetype(page, start_migratetype);
1943 trace_mm_page_alloc_extfrag(page, order, current_order,
1944 start_migratetype, fallback_mt);
1953 * Do the hard work of removing an element from the buddy allocator.
1954 * Call me with the zone->lock already held.
1956 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1961 page = __rmqueue_smallest(zone, order, migratetype);
1962 if (unlikely(!page)) {
1963 if (migratetype == MIGRATE_MOVABLE)
1964 page = __rmqueue_cma_fallback(zone, order);
1967 page = __rmqueue_fallback(zone, order, migratetype);
1970 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1975 * Obtain a specified number of elements from the buddy allocator, all under
1976 * a single hold of the lock, for efficiency. Add them to the supplied list.
1977 * Returns the number of new pages which were placed at *list.
1979 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1980 unsigned long count, struct list_head *list,
1981 int migratetype, bool cold)
1985 spin_lock(&zone->lock);
1986 for (i = 0; i < count; ++i) {
1987 struct page *page = __rmqueue(zone, order, migratetype);
1988 if (unlikely(page == NULL))
1992 * Split buddy pages returned by expand() are received here
1993 * in physical page order. The page is added to the callers and
1994 * list and the list head then moves forward. From the callers
1995 * perspective, the linked list is ordered by page number in
1996 * some conditions. This is useful for IO devices that can
1997 * merge IO requests if the physical pages are ordered
2001 list_add(&page->lru, list);
2003 list_add_tail(&page->lru, list);
2005 if (is_migrate_cma(get_pcppage_migratetype(page)))
2006 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2009 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2010 spin_unlock(&zone->lock);
2016 * Called from the vmstat counter updater to drain pagesets of this
2017 * currently executing processor on remote nodes after they have
2020 * Note that this function must be called with the thread pinned to
2021 * a single processor.
2023 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2025 unsigned long flags;
2026 int to_drain, batch;
2028 local_irq_save(flags);
2029 batch = READ_ONCE(pcp->batch);
2030 to_drain = min(pcp->count, batch);
2032 free_pcppages_bulk(zone, to_drain, pcp);
2033 pcp->count -= to_drain;
2035 local_irq_restore(flags);
2040 * Drain pcplists of the indicated processor and zone.
2042 * The processor must either be the current processor and the
2043 * thread pinned to the current processor or a processor that
2046 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2048 unsigned long flags;
2049 struct per_cpu_pageset *pset;
2050 struct per_cpu_pages *pcp;
2052 local_irq_save(flags);
2053 pset = per_cpu_ptr(zone->pageset, cpu);
2057 free_pcppages_bulk(zone, pcp->count, pcp);
2060 local_irq_restore(flags);
2064 * Drain pcplists of all zones on the indicated processor.
2066 * The processor must either be the current processor and the
2067 * thread pinned to the current processor or a processor that
2070 static void drain_pages(unsigned int cpu)
2074 for_each_populated_zone(zone) {
2075 drain_pages_zone(cpu, zone);
2080 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2082 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2083 * the single zone's pages.
2085 void drain_local_pages(struct zone *zone)
2087 int cpu = smp_processor_id();
2090 drain_pages_zone(cpu, zone);
2096 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2098 * When zone parameter is non-NULL, spill just the single zone's pages.
2100 * Note that this code is protected against sending an IPI to an offline
2101 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2102 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2103 * nothing keeps CPUs from showing up after we populated the cpumask and
2104 * before the call to on_each_cpu_mask().
2106 void drain_all_pages(struct zone *zone)
2111 * Allocate in the BSS so we wont require allocation in
2112 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2114 static cpumask_t cpus_with_pcps;
2117 * We don't care about racing with CPU hotplug event
2118 * as offline notification will cause the notified
2119 * cpu to drain that CPU pcps and on_each_cpu_mask
2120 * disables preemption as part of its processing
2122 for_each_online_cpu(cpu) {
2123 struct per_cpu_pageset *pcp;
2125 bool has_pcps = false;
2128 pcp = per_cpu_ptr(zone->pageset, cpu);
2132 for_each_populated_zone(z) {
2133 pcp = per_cpu_ptr(z->pageset, cpu);
2134 if (pcp->pcp.count) {
2142 cpumask_set_cpu(cpu, &cpus_with_pcps);
2144 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2146 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2150 #ifdef CONFIG_HIBERNATION
2152 void mark_free_pages(struct zone *zone)
2154 unsigned long pfn, max_zone_pfn;
2155 unsigned long flags;
2156 unsigned int order, t;
2159 if (zone_is_empty(zone))
2162 spin_lock_irqsave(&zone->lock, flags);
2164 max_zone_pfn = zone_end_pfn(zone);
2165 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2166 if (pfn_valid(pfn)) {
2167 page = pfn_to_page(pfn);
2169 if (page_zone(page) != zone)
2172 if (!swsusp_page_is_forbidden(page))
2173 swsusp_unset_page_free(page);
2176 for_each_migratetype_order(order, t) {
2177 list_for_each_entry(page,
2178 &zone->free_area[order].free_list[t], lru) {
2181 pfn = page_to_pfn(page);
2182 for (i = 0; i < (1UL << order); i++)
2183 swsusp_set_page_free(pfn_to_page(pfn + i));
2186 spin_unlock_irqrestore(&zone->lock, flags);
2188 #endif /* CONFIG_PM */
2191 * Free a 0-order page
2192 * cold == true ? free a cold page : free a hot page
2194 void free_hot_cold_page(struct page *page, bool cold)
2196 struct zone *zone = page_zone(page);
2197 struct per_cpu_pages *pcp;
2198 unsigned long flags;
2199 unsigned long pfn = page_to_pfn(page);
2202 if (!free_pages_prepare(page, 0))
2205 migratetype = get_pfnblock_migratetype(page, pfn);
2206 set_pcppage_migratetype(page, migratetype);
2207 local_irq_save(flags);
2208 __count_vm_event(PGFREE);
2211 * We only track unmovable, reclaimable and movable on pcp lists.
2212 * Free ISOLATE pages back to the allocator because they are being
2213 * offlined but treat RESERVE as movable pages so we can get those
2214 * areas back if necessary. Otherwise, we may have to free
2215 * excessively into the page allocator
2217 if (migratetype >= MIGRATE_PCPTYPES) {
2218 if (unlikely(is_migrate_isolate(migratetype))) {
2219 free_one_page(zone, page, pfn, 0, migratetype);
2222 migratetype = MIGRATE_MOVABLE;
2225 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2227 list_add(&page->lru, &pcp->lists[migratetype]);
2229 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2231 if (pcp->count >= pcp->high) {
2232 unsigned long batch = READ_ONCE(pcp->batch);
2233 free_pcppages_bulk(zone, batch, pcp);
2234 pcp->count -= batch;
2238 local_irq_restore(flags);
2242 * Free a list of 0-order pages
2244 void free_hot_cold_page_list(struct list_head *list, bool cold)
2246 struct page *page, *next;
2248 list_for_each_entry_safe(page, next, list, lru) {
2249 trace_mm_page_free_batched(page, cold);
2250 free_hot_cold_page(page, cold);
2255 * split_page takes a non-compound higher-order page, and splits it into
2256 * n (1<<order) sub-pages: page[0..n]
2257 * Each sub-page must be freed individually.
2259 * Note: this is probably too low level an operation for use in drivers.
2260 * Please consult with lkml before using this in your driver.
2262 void split_page(struct page *page, unsigned int order)
2267 VM_BUG_ON_PAGE(PageCompound(page), page);
2268 VM_BUG_ON_PAGE(!page_count(page), page);
2270 #ifdef CONFIG_KMEMCHECK
2272 * Split shadow pages too, because free(page[0]) would
2273 * otherwise free the whole shadow.
2275 if (kmemcheck_page_is_tracked(page))
2276 split_page(virt_to_page(page[0].shadow), order);
2279 gfp_mask = get_page_owner_gfp(page);
2280 set_page_owner(page, 0, gfp_mask);
2281 for (i = 1; i < (1 << order); i++) {
2282 set_page_refcounted(page + i);
2283 set_page_owner(page + i, 0, gfp_mask);
2286 EXPORT_SYMBOL_GPL(split_page);
2288 int __isolate_free_page(struct page *page, unsigned int order)
2290 unsigned long watermark;
2294 BUG_ON(!PageBuddy(page));
2296 zone = page_zone(page);
2297 mt = get_pageblock_migratetype(page);
2299 if (!is_migrate_isolate(mt)) {
2300 /* Obey watermarks as if the page was being allocated */
2301 watermark = low_wmark_pages(zone) + (1 << order);
2302 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2305 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2308 /* Remove page from free list */
2309 list_del(&page->lru);
2310 zone->free_area[order].nr_free--;
2311 rmv_page_order(page);
2313 set_page_owner(page, order, __GFP_MOVABLE);
2315 /* Set the pageblock if the isolated page is at least a pageblock */
2316 if (order >= pageblock_order - 1) {
2317 struct page *endpage = page + (1 << order) - 1;
2318 for (; page < endpage; page += pageblock_nr_pages) {
2319 int mt = get_pageblock_migratetype(page);
2320 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2321 set_pageblock_migratetype(page,
2327 return 1UL << order;
2331 * Similar to split_page except the page is already free. As this is only
2332 * being used for migration, the migratetype of the block also changes.
2333 * As this is called with interrupts disabled, the caller is responsible
2334 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2337 * Note: this is probably too low level an operation for use in drivers.
2338 * Please consult with lkml before using this in your driver.
2340 int split_free_page(struct page *page)
2345 order = page_order(page);
2347 nr_pages = __isolate_free_page(page, order);
2351 /* Split into individual pages */
2352 set_page_refcounted(page);
2353 split_page(page, order);
2358 * Update NUMA hit/miss statistics
2360 * Must be called with interrupts disabled.
2362 * When __GFP_OTHER_NODE is set assume the node of the preferred
2363 * zone is the local node. This is useful for daemons who allocate
2364 * memory on behalf of other processes.
2366 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2370 int local_nid = numa_node_id();
2371 enum zone_stat_item local_stat = NUMA_LOCAL;
2373 if (unlikely(flags & __GFP_OTHER_NODE)) {
2374 local_stat = NUMA_OTHER;
2375 local_nid = preferred_zone->node;
2378 if (z->node == local_nid) {
2379 __inc_zone_state(z, NUMA_HIT);
2380 __inc_zone_state(z, local_stat);
2382 __inc_zone_state(z, NUMA_MISS);
2383 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
2389 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2392 struct page *buffered_rmqueue(struct zone *preferred_zone,
2393 struct zone *zone, unsigned int order,
2394 gfp_t gfp_flags, int alloc_flags, int migratetype)
2396 unsigned long flags;
2398 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2400 if (likely(order == 0)) {
2401 struct per_cpu_pages *pcp;
2402 struct list_head *list;
2404 local_irq_save(flags);
2405 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2406 list = &pcp->lists[migratetype];
2407 if (list_empty(list)) {
2408 pcp->count += rmqueue_bulk(zone, 0,
2411 if (unlikely(list_empty(list)))
2416 page = list_last_entry(list, struct page, lru);
2418 page = list_first_entry(list, struct page, lru);
2420 list_del(&page->lru);
2424 * We most definitely don't want callers attempting to
2425 * allocate greater than order-1 page units with __GFP_NOFAIL.
2427 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2428 spin_lock_irqsave(&zone->lock, flags);
2431 if (alloc_flags & ALLOC_HARDER) {
2432 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2434 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2437 page = __rmqueue(zone, order, migratetype);
2438 spin_unlock(&zone->lock);
2441 __mod_zone_freepage_state(zone, -(1 << order),
2442 get_pcppage_migratetype(page));
2445 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2446 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2447 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2448 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2450 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2451 zone_statistics(preferred_zone, zone, gfp_flags);
2452 local_irq_restore(flags);
2454 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2458 local_irq_restore(flags);
2462 #ifdef CONFIG_FAIL_PAGE_ALLOC
2465 struct fault_attr attr;
2467 bool ignore_gfp_highmem;
2468 bool ignore_gfp_reclaim;
2470 } fail_page_alloc = {
2471 .attr = FAULT_ATTR_INITIALIZER,
2472 .ignore_gfp_reclaim = true,
2473 .ignore_gfp_highmem = true,
2477 static int __init setup_fail_page_alloc(char *str)
2479 return setup_fault_attr(&fail_page_alloc.attr, str);
2481 __setup("fail_page_alloc=", setup_fail_page_alloc);
2483 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2485 if (order < fail_page_alloc.min_order)
2487 if (gfp_mask & __GFP_NOFAIL)
2489 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2491 if (fail_page_alloc.ignore_gfp_reclaim &&
2492 (gfp_mask & __GFP_DIRECT_RECLAIM))
2495 return should_fail(&fail_page_alloc.attr, 1 << order);
2498 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2500 static int __init fail_page_alloc_debugfs(void)
2502 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2505 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2506 &fail_page_alloc.attr);
2508 return PTR_ERR(dir);
2510 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2511 &fail_page_alloc.ignore_gfp_reclaim))
2513 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2514 &fail_page_alloc.ignore_gfp_highmem))
2516 if (!debugfs_create_u32("min-order", mode, dir,
2517 &fail_page_alloc.min_order))
2522 debugfs_remove_recursive(dir);
2527 late_initcall(fail_page_alloc_debugfs);
2529 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2531 #else /* CONFIG_FAIL_PAGE_ALLOC */
2533 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2538 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2541 * Return true if free base pages are above 'mark'. For high-order checks it
2542 * will return true of the order-0 watermark is reached and there is at least
2543 * one free page of a suitable size. Checking now avoids taking the zone lock
2544 * to check in the allocation paths if no pages are free.
2546 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2547 unsigned long mark, int classzone_idx, int alloc_flags,
2552 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2554 /* free_pages may go negative - that's OK */
2555 free_pages -= (1 << order) - 1;
2557 if (alloc_flags & ALLOC_HIGH)
2561 * If the caller does not have rights to ALLOC_HARDER then subtract
2562 * the high-atomic reserves. This will over-estimate the size of the
2563 * atomic reserve but it avoids a search.
2565 if (likely(!alloc_harder))
2566 free_pages -= z->nr_reserved_highatomic;
2571 /* If allocation can't use CMA areas don't use free CMA pages */
2572 if (!(alloc_flags & ALLOC_CMA))
2573 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2577 * Check watermarks for an order-0 allocation request. If these
2578 * are not met, then a high-order request also cannot go ahead
2579 * even if a suitable page happened to be free.
2581 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2584 /* If this is an order-0 request then the watermark is fine */
2588 /* For a high-order request, check at least one suitable page is free */
2589 for (o = order; o < MAX_ORDER; o++) {
2590 struct free_area *area = &z->free_area[o];
2599 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2600 if (!list_empty(&area->free_list[mt]))
2605 if ((alloc_flags & ALLOC_CMA) &&
2606 !list_empty(&area->free_list[MIGRATE_CMA])) {
2614 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2615 int classzone_idx, int alloc_flags)
2617 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2618 zone_page_state(z, NR_FREE_PAGES));
2621 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2622 unsigned long mark, int classzone_idx)
2624 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2626 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2627 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2629 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2634 static bool zone_local(struct zone *local_zone, struct zone *zone)
2636 return local_zone->node == zone->node;
2639 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2641 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2644 #else /* CONFIG_NUMA */
2645 static bool zone_local(struct zone *local_zone, struct zone *zone)
2650 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2654 #endif /* CONFIG_NUMA */
2656 static void reset_alloc_batches(struct zone *preferred_zone)
2658 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2661 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2662 high_wmark_pages(zone) - low_wmark_pages(zone) -
2663 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2664 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2665 } while (zone++ != preferred_zone);
2669 * get_page_from_freelist goes through the zonelist trying to allocate
2672 static struct page *
2673 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2674 const struct alloc_context *ac)
2676 struct zonelist *zonelist = ac->zonelist;
2678 struct page *page = NULL;
2680 int nr_fair_skipped = 0;
2681 bool zonelist_rescan;
2684 zonelist_rescan = false;
2687 * Scan zonelist, looking for a zone with enough free.
2688 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2690 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2694 if (cpusets_enabled() &&
2695 (alloc_flags & ALLOC_CPUSET) &&
2696 !cpuset_zone_allowed(zone, gfp_mask))
2699 * Distribute pages in proportion to the individual
2700 * zone size to ensure fair page aging. The zone a
2701 * page was allocated in should have no effect on the
2702 * time the page has in memory before being reclaimed.
2704 if (alloc_flags & ALLOC_FAIR) {
2705 if (!zone_local(ac->preferred_zone, zone))
2707 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2713 * When allocating a page cache page for writing, we
2714 * want to get it from a zone that is within its dirty
2715 * limit, such that no single zone holds more than its
2716 * proportional share of globally allowed dirty pages.
2717 * The dirty limits take into account the zone's
2718 * lowmem reserves and high watermark so that kswapd
2719 * should be able to balance it without having to
2720 * write pages from its LRU list.
2722 * This may look like it could increase pressure on
2723 * lower zones by failing allocations in higher zones
2724 * before they are full. But the pages that do spill
2725 * over are limited as the lower zones are protected
2726 * by this very same mechanism. It should not become
2727 * a practical burden to them.
2729 * XXX: For now, allow allocations to potentially
2730 * exceed the per-zone dirty limit in the slowpath
2731 * (spread_dirty_pages unset) before going into reclaim,
2732 * which is important when on a NUMA setup the allowed
2733 * zones are together not big enough to reach the
2734 * global limit. The proper fix for these situations
2735 * will require awareness of zones in the
2736 * dirty-throttling and the flusher threads.
2738 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2741 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2742 if (!zone_watermark_ok(zone, order, mark,
2743 ac->classzone_idx, alloc_flags)) {
2746 /* Checked here to keep the fast path fast */
2747 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2748 if (alloc_flags & ALLOC_NO_WATERMARKS)
2751 if (zone_reclaim_mode == 0 ||
2752 !zone_allows_reclaim(ac->preferred_zone, zone))
2755 ret = zone_reclaim(zone, gfp_mask, order);
2757 case ZONE_RECLAIM_NOSCAN:
2760 case ZONE_RECLAIM_FULL:
2761 /* scanned but unreclaimable */
2764 /* did we reclaim enough */
2765 if (zone_watermark_ok(zone, order, mark,
2766 ac->classzone_idx, alloc_flags))
2774 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2775 gfp_mask, alloc_flags, ac->migratetype);
2777 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2781 * If this is a high-order atomic allocation then check
2782 * if the pageblock should be reserved for the future
2784 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2785 reserve_highatomic_pageblock(page, zone, order);
2792 * The first pass makes sure allocations are spread fairly within the
2793 * local node. However, the local node might have free pages left
2794 * after the fairness batches are exhausted, and remote zones haven't
2795 * even been considered yet. Try once more without fairness, and
2796 * include remote zones now, before entering the slowpath and waking
2797 * kswapd: prefer spilling to a remote zone over swapping locally.
2799 if (alloc_flags & ALLOC_FAIR) {
2800 alloc_flags &= ~ALLOC_FAIR;
2801 if (nr_fair_skipped) {
2802 zonelist_rescan = true;
2803 reset_alloc_batches(ac->preferred_zone);
2805 if (nr_online_nodes > 1)
2806 zonelist_rescan = true;
2809 if (zonelist_rescan)
2816 * Large machines with many possible nodes should not always dump per-node
2817 * meminfo in irq context.
2819 static inline bool should_suppress_show_mem(void)
2824 ret = in_interrupt();
2829 static DEFINE_RATELIMIT_STATE(nopage_rs,
2830 DEFAULT_RATELIMIT_INTERVAL,
2831 DEFAULT_RATELIMIT_BURST);
2833 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2835 unsigned int filter = SHOW_MEM_FILTER_NODES;
2837 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2838 debug_guardpage_minorder() > 0)
2842 * This documents exceptions given to allocations in certain
2843 * contexts that are allowed to allocate outside current's set
2846 if (!(gfp_mask & __GFP_NOMEMALLOC))
2847 if (test_thread_flag(TIF_MEMDIE) ||
2848 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2849 filter &= ~SHOW_MEM_FILTER_NODES;
2850 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2851 filter &= ~SHOW_MEM_FILTER_NODES;
2854 struct va_format vaf;
2857 va_start(args, fmt);
2862 pr_warn("%pV", &vaf);
2867 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
2868 current->comm, order, gfp_mask, &gfp_mask);
2870 if (!should_suppress_show_mem())
2874 static inline struct page *
2875 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2876 const struct alloc_context *ac, unsigned long *did_some_progress)
2878 struct oom_control oc = {
2879 .zonelist = ac->zonelist,
2880 .nodemask = ac->nodemask,
2881 .gfp_mask = gfp_mask,
2886 *did_some_progress = 0;
2889 * Acquire the oom lock. If that fails, somebody else is
2890 * making progress for us.
2892 if (!mutex_trylock(&oom_lock)) {
2893 *did_some_progress = 1;
2894 schedule_timeout_uninterruptible(1);
2899 * Go through the zonelist yet one more time, keep very high watermark
2900 * here, this is only to catch a parallel oom killing, we must fail if
2901 * we're still under heavy pressure.
2903 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2904 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2908 if (!(gfp_mask & __GFP_NOFAIL)) {
2909 /* Coredumps can quickly deplete all memory reserves */
2910 if (current->flags & PF_DUMPCORE)
2912 /* The OOM killer will not help higher order allocs */
2913 if (order > PAGE_ALLOC_COSTLY_ORDER)
2915 /* The OOM killer does not needlessly kill tasks for lowmem */
2916 if (ac->high_zoneidx < ZONE_NORMAL)
2918 if (pm_suspended_storage())
2921 * XXX: GFP_NOFS allocations should rather fail than rely on
2922 * other request to make a forward progress.
2923 * We are in an unfortunate situation where out_of_memory cannot
2924 * do much for this context but let's try it to at least get
2925 * access to memory reserved if the current task is killed (see
2926 * out_of_memory). Once filesystems are ready to handle allocation
2927 * failures more gracefully we should just bail out here.
2930 /* The OOM killer may not free memory on a specific node */
2931 if (gfp_mask & __GFP_THISNODE)
2934 /* Exhausted what can be done so it's blamo time */
2935 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
2936 *did_some_progress = 1;
2938 if (gfp_mask & __GFP_NOFAIL) {
2939 page = get_page_from_freelist(gfp_mask, order,
2940 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
2942 * fallback to ignore cpuset restriction if our nodes
2946 page = get_page_from_freelist(gfp_mask, order,
2947 ALLOC_NO_WATERMARKS, ac);
2951 mutex_unlock(&oom_lock);
2955 #ifdef CONFIG_COMPACTION
2956 /* Try memory compaction for high-order allocations before reclaim */
2957 static struct page *
2958 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2959 int alloc_flags, const struct alloc_context *ac,
2960 enum migrate_mode mode, int *contended_compaction,
2961 bool *deferred_compaction)
2963 unsigned long compact_result;
2969 current->flags |= PF_MEMALLOC;
2970 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2971 mode, contended_compaction);
2972 current->flags &= ~PF_MEMALLOC;
2974 switch (compact_result) {
2975 case COMPACT_DEFERRED:
2976 *deferred_compaction = true;
2978 case COMPACT_SKIPPED:
2985 * At least in one zone compaction wasn't deferred or skipped, so let's
2986 * count a compaction stall
2988 count_vm_event(COMPACTSTALL);
2990 page = get_page_from_freelist(gfp_mask, order,
2991 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2994 struct zone *zone = page_zone(page);
2996 zone->compact_blockskip_flush = false;
2997 compaction_defer_reset(zone, order, true);
2998 count_vm_event(COMPACTSUCCESS);
3003 * It's bad if compaction run occurs and fails. The most likely reason
3004 * is that pages exist, but not enough to satisfy watermarks.
3006 count_vm_event(COMPACTFAIL);
3013 static inline struct page *
3014 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3015 int alloc_flags, const struct alloc_context *ac,
3016 enum migrate_mode mode, int *contended_compaction,
3017 bool *deferred_compaction)
3021 #endif /* CONFIG_COMPACTION */
3023 /* Perform direct synchronous page reclaim */
3025 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3026 const struct alloc_context *ac)
3028 struct reclaim_state reclaim_state;
3033 /* We now go into synchronous reclaim */
3034 cpuset_memory_pressure_bump();
3035 current->flags |= PF_MEMALLOC;
3036 lockdep_set_current_reclaim_state(gfp_mask);
3037 reclaim_state.reclaimed_slab = 0;
3038 current->reclaim_state = &reclaim_state;
3040 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3043 current->reclaim_state = NULL;
3044 lockdep_clear_current_reclaim_state();
3045 current->flags &= ~PF_MEMALLOC;
3052 /* The really slow allocator path where we enter direct reclaim */
3053 static inline struct page *
3054 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3055 int alloc_flags, const struct alloc_context *ac,
3056 unsigned long *did_some_progress)
3058 struct page *page = NULL;
3059 bool drained = false;
3061 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3062 if (unlikely(!(*did_some_progress)))
3066 page = get_page_from_freelist(gfp_mask, order,
3067 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3070 * If an allocation failed after direct reclaim, it could be because
3071 * pages are pinned on the per-cpu lists or in high alloc reserves.
3072 * Shrink them them and try again
3074 if (!page && !drained) {
3075 unreserve_highatomic_pageblock(ac);
3076 drain_all_pages(NULL);
3084 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3089 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3090 ac->high_zoneidx, ac->nodemask)
3091 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
3095 gfp_to_alloc_flags(gfp_t gfp_mask)
3097 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3099 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3100 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3103 * The caller may dip into page reserves a bit more if the caller
3104 * cannot run direct reclaim, or if the caller has realtime scheduling
3105 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3106 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3108 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3110 if (gfp_mask & __GFP_ATOMIC) {
3112 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3113 * if it can't schedule.
3115 if (!(gfp_mask & __GFP_NOMEMALLOC))
3116 alloc_flags |= ALLOC_HARDER;
3118 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3119 * comment for __cpuset_node_allowed().
3121 alloc_flags &= ~ALLOC_CPUSET;
3122 } else if (unlikely(rt_task(current)) && !in_interrupt())
3123 alloc_flags |= ALLOC_HARDER;
3125 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
3126 if (gfp_mask & __GFP_MEMALLOC)
3127 alloc_flags |= ALLOC_NO_WATERMARKS;
3128 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3129 alloc_flags |= ALLOC_NO_WATERMARKS;
3130 else if (!in_interrupt() &&
3131 ((current->flags & PF_MEMALLOC) ||
3132 unlikely(test_thread_flag(TIF_MEMDIE))))
3133 alloc_flags |= ALLOC_NO_WATERMARKS;
3136 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3137 alloc_flags |= ALLOC_CMA;
3142 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3144 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
3147 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
3149 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
3152 static inline struct page *
3153 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3154 struct alloc_context *ac)
3156 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3157 struct page *page = NULL;
3159 unsigned long pages_reclaimed = 0;
3160 unsigned long did_some_progress;
3161 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3162 bool deferred_compaction = false;
3163 int contended_compaction = COMPACT_CONTENDED_NONE;
3166 * In the slowpath, we sanity check order to avoid ever trying to
3167 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3168 * be using allocators in order of preference for an area that is
3171 if (order >= MAX_ORDER) {
3172 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3177 * We also sanity check to catch abuse of atomic reserves being used by
3178 * callers that are not in atomic context.
3180 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3181 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3182 gfp_mask &= ~__GFP_ATOMIC;
3185 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3186 wake_all_kswapds(order, ac);
3189 * OK, we're below the kswapd watermark and have kicked background
3190 * reclaim. Now things get more complex, so set up alloc_flags according
3191 * to how we want to proceed.
3193 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3196 * Find the true preferred zone if the allocation is unconstrained by
3199 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3200 struct zoneref *preferred_zoneref;
3201 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3202 ac->high_zoneidx, NULL, &ac->preferred_zone);
3203 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3206 /* This is the last chance, in general, before the goto nopage. */
3207 page = get_page_from_freelist(gfp_mask, order,
3208 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3212 /* Allocate without watermarks if the context allows */
3213 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3215 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3216 * the allocation is high priority and these type of
3217 * allocations are system rather than user orientated
3219 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3220 page = get_page_from_freelist(gfp_mask, order,
3221 ALLOC_NO_WATERMARKS, ac);
3226 /* Caller is not willing to reclaim, we can't balance anything */
3227 if (!can_direct_reclaim) {
3229 * All existing users of the __GFP_NOFAIL are blockable, so warn
3230 * of any new users that actually allow this type of allocation
3233 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3237 /* Avoid recursion of direct reclaim */
3238 if (current->flags & PF_MEMALLOC) {
3240 * __GFP_NOFAIL request from this context is rather bizarre
3241 * because we cannot reclaim anything and only can loop waiting
3242 * for somebody to do a work for us.
3244 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3251 /* Avoid allocations with no watermarks from looping endlessly */
3252 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3256 * Try direct compaction. The first pass is asynchronous. Subsequent
3257 * attempts after direct reclaim are synchronous
3259 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3261 &contended_compaction,
3262 &deferred_compaction);
3266 /* Checks for THP-specific high-order allocations */
3267 if (is_thp_gfp_mask(gfp_mask)) {
3269 * If compaction is deferred for high-order allocations, it is
3270 * because sync compaction recently failed. If this is the case
3271 * and the caller requested a THP allocation, we do not want
3272 * to heavily disrupt the system, so we fail the allocation
3273 * instead of entering direct reclaim.
3275 if (deferred_compaction)
3279 * In all zones where compaction was attempted (and not
3280 * deferred or skipped), lock contention has been detected.
3281 * For THP allocation we do not want to disrupt the others
3282 * so we fallback to base pages instead.
3284 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3288 * If compaction was aborted due to need_resched(), we do not
3289 * want to further increase allocation latency, unless it is
3290 * khugepaged trying to collapse.
3292 if (contended_compaction == COMPACT_CONTENDED_SCHED
3293 && !(current->flags & PF_KTHREAD))
3298 * It can become very expensive to allocate transparent hugepages at
3299 * fault, so use asynchronous memory compaction for THP unless it is
3300 * khugepaged trying to collapse.
3302 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3303 migration_mode = MIGRATE_SYNC_LIGHT;
3305 /* Try direct reclaim and then allocating */
3306 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3307 &did_some_progress);
3311 /* Do not loop if specifically requested */
3312 if (gfp_mask & __GFP_NORETRY)
3315 /* Keep reclaiming pages as long as there is reasonable progress */
3316 pages_reclaimed += did_some_progress;
3317 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3318 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3319 /* Wait for some write requests to complete then retry */
3320 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3324 /* Reclaim has failed us, start killing things */
3325 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3329 /* Retry as long as the OOM killer is making progress */
3330 if (did_some_progress)
3335 * High-order allocations do not necessarily loop after
3336 * direct reclaim and reclaim/compaction depends on compaction
3337 * being called after reclaim so call directly if necessary
3339 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3341 &contended_compaction,
3342 &deferred_compaction);
3346 warn_alloc_failed(gfp_mask, order, NULL);
3352 * This is the 'heart' of the zoned buddy allocator.
3355 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3356 struct zonelist *zonelist, nodemask_t *nodemask)
3358 struct zoneref *preferred_zoneref;
3359 struct page *page = NULL;
3360 unsigned int cpuset_mems_cookie;
3361 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3362 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3363 struct alloc_context ac = {
3364 .high_zoneidx = gfp_zone(gfp_mask),
3365 .nodemask = nodemask,
3366 .migratetype = gfpflags_to_migratetype(gfp_mask),
3369 gfp_mask &= gfp_allowed_mask;
3371 lockdep_trace_alloc(gfp_mask);
3373 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3375 if (should_fail_alloc_page(gfp_mask, order))
3379 * Check the zones suitable for the gfp_mask contain at least one
3380 * valid zone. It's possible to have an empty zonelist as a result
3381 * of __GFP_THISNODE and a memoryless node
3383 if (unlikely(!zonelist->_zonerefs->zone))
3386 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3387 alloc_flags |= ALLOC_CMA;
3390 cpuset_mems_cookie = read_mems_allowed_begin();
3392 /* We set it here, as __alloc_pages_slowpath might have changed it */
3393 ac.zonelist = zonelist;
3395 /* Dirty zone balancing only done in the fast path */
3396 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3398 /* The preferred zone is used for statistics later */
3399 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3400 ac.nodemask ? : &cpuset_current_mems_allowed,
3401 &ac.preferred_zone);
3402 if (!ac.preferred_zone)
3404 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3406 /* First allocation attempt */
3407 alloc_mask = gfp_mask|__GFP_HARDWALL;
3408 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3409 if (unlikely(!page)) {
3411 * Runtime PM, block IO and its error handling path
3412 * can deadlock because I/O on the device might not
3415 alloc_mask = memalloc_noio_flags(gfp_mask);
3416 ac.spread_dirty_pages = false;
3418 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3421 if (kmemcheck_enabled && page)
3422 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3424 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3428 * When updating a task's mems_allowed, it is possible to race with
3429 * parallel threads in such a way that an allocation can fail while
3430 * the mask is being updated. If a page allocation is about to fail,
3431 * check if the cpuset changed during allocation and if so, retry.
3433 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3438 EXPORT_SYMBOL(__alloc_pages_nodemask);
3441 * Common helper functions.
3443 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3448 * __get_free_pages() returns a 32-bit address, which cannot represent
3451 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3453 page = alloc_pages(gfp_mask, order);
3456 return (unsigned long) page_address(page);
3458 EXPORT_SYMBOL(__get_free_pages);
3460 unsigned long get_zeroed_page(gfp_t gfp_mask)
3462 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3464 EXPORT_SYMBOL(get_zeroed_page);
3466 void __free_pages(struct page *page, unsigned int order)
3468 if (put_page_testzero(page)) {
3470 free_hot_cold_page(page, false);
3472 __free_pages_ok(page, order);
3476 EXPORT_SYMBOL(__free_pages);
3478 void free_pages(unsigned long addr, unsigned int order)
3481 VM_BUG_ON(!virt_addr_valid((void *)addr));
3482 __free_pages(virt_to_page((void *)addr), order);
3486 EXPORT_SYMBOL(free_pages);
3490 * An arbitrary-length arbitrary-offset area of memory which resides
3491 * within a 0 or higher order page. Multiple fragments within that page
3492 * are individually refcounted, in the page's reference counter.
3494 * The page_frag functions below provide a simple allocation framework for
3495 * page fragments. This is used by the network stack and network device
3496 * drivers to provide a backing region of memory for use as either an
3497 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3499 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3502 struct page *page = NULL;
3503 gfp_t gfp = gfp_mask;
3505 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3506 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3508 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3509 PAGE_FRAG_CACHE_MAX_ORDER);
3510 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3512 if (unlikely(!page))
3513 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3515 nc->va = page ? page_address(page) : NULL;
3520 void *__alloc_page_frag(struct page_frag_cache *nc,
3521 unsigned int fragsz, gfp_t gfp_mask)
3523 unsigned int size = PAGE_SIZE;
3527 if (unlikely(!nc->va)) {
3529 page = __page_frag_refill(nc, gfp_mask);
3533 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3534 /* if size can vary use size else just use PAGE_SIZE */
3537 /* Even if we own the page, we do not use atomic_set().
3538 * This would break get_page_unless_zero() users.
3540 page_ref_add(page, size - 1);
3542 /* reset page count bias and offset to start of new frag */
3543 nc->pfmemalloc = page_is_pfmemalloc(page);
3544 nc->pagecnt_bias = size;
3548 offset = nc->offset - fragsz;
3549 if (unlikely(offset < 0)) {
3550 page = virt_to_page(nc->va);
3552 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
3555 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3556 /* if size can vary use size else just use PAGE_SIZE */
3559 /* OK, page count is 0, we can safely set it */
3560 set_page_count(page, size);
3562 /* reset page count bias and offset to start of new frag */
3563 nc->pagecnt_bias = size;
3564 offset = size - fragsz;
3568 nc->offset = offset;
3570 return nc->va + offset;
3572 EXPORT_SYMBOL(__alloc_page_frag);
3575 * Frees a page fragment allocated out of either a compound or order 0 page.
3577 void __free_page_frag(void *addr)
3579 struct page *page = virt_to_head_page(addr);
3581 if (unlikely(put_page_testzero(page)))
3582 __free_pages_ok(page, compound_order(page));
3584 EXPORT_SYMBOL(__free_page_frag);
3587 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3588 * of the current memory cgroup if __GFP_ACCOUNT is set, other than that it is
3589 * equivalent to alloc_pages.
3591 * It should be used when the caller would like to use kmalloc, but since the
3592 * allocation is large, it has to fall back to the page allocator.
3594 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3598 page = alloc_pages(gfp_mask, order);
3599 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3600 __free_pages(page, order);
3606 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3610 page = alloc_pages_node(nid, gfp_mask, order);
3611 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3612 __free_pages(page, order);
3619 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3622 void __free_kmem_pages(struct page *page, unsigned int order)
3624 memcg_kmem_uncharge(page, order);
3625 __free_pages(page, order);
3628 void free_kmem_pages(unsigned long addr, unsigned int order)
3631 VM_BUG_ON(!virt_addr_valid((void *)addr));
3632 __free_kmem_pages(virt_to_page((void *)addr), order);
3636 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3640 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3641 unsigned long used = addr + PAGE_ALIGN(size);
3643 split_page(virt_to_page((void *)addr), order);
3644 while (used < alloc_end) {
3649 return (void *)addr;
3653 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3654 * @size: the number of bytes to allocate
3655 * @gfp_mask: GFP flags for the allocation
3657 * This function is similar to alloc_pages(), except that it allocates the
3658 * minimum number of pages to satisfy the request. alloc_pages() can only
3659 * allocate memory in power-of-two pages.
3661 * This function is also limited by MAX_ORDER.
3663 * Memory allocated by this function must be released by free_pages_exact().
3665 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3667 unsigned int order = get_order(size);
3670 addr = __get_free_pages(gfp_mask, order);
3671 return make_alloc_exact(addr, order, size);
3673 EXPORT_SYMBOL(alloc_pages_exact);
3676 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3678 * @nid: the preferred node ID where memory should be allocated
3679 * @size: the number of bytes to allocate
3680 * @gfp_mask: GFP flags for the allocation
3682 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3685 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3687 unsigned int order = get_order(size);
3688 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3691 return make_alloc_exact((unsigned long)page_address(p), order, size);
3695 * free_pages_exact - release memory allocated via alloc_pages_exact()
3696 * @virt: the value returned by alloc_pages_exact.
3697 * @size: size of allocation, same value as passed to alloc_pages_exact().
3699 * Release the memory allocated by a previous call to alloc_pages_exact.
3701 void free_pages_exact(void *virt, size_t size)
3703 unsigned long addr = (unsigned long)virt;
3704 unsigned long end = addr + PAGE_ALIGN(size);
3706 while (addr < end) {
3711 EXPORT_SYMBOL(free_pages_exact);
3714 * nr_free_zone_pages - count number of pages beyond high watermark
3715 * @offset: The zone index of the highest zone
3717 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3718 * high watermark within all zones at or below a given zone index. For each
3719 * zone, the number of pages is calculated as:
3720 * managed_pages - high_pages
3722 static unsigned long nr_free_zone_pages(int offset)
3727 /* Just pick one node, since fallback list is circular */
3728 unsigned long sum = 0;
3730 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3732 for_each_zone_zonelist(zone, z, zonelist, offset) {
3733 unsigned long size = zone->managed_pages;
3734 unsigned long high = high_wmark_pages(zone);
3743 * nr_free_buffer_pages - count number of pages beyond high watermark
3745 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3746 * watermark within ZONE_DMA and ZONE_NORMAL.
3748 unsigned long nr_free_buffer_pages(void)
3750 return nr_free_zone_pages(gfp_zone(GFP_USER));
3752 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3755 * nr_free_pagecache_pages - count number of pages beyond high watermark
3757 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3758 * high watermark within all zones.
3760 unsigned long nr_free_pagecache_pages(void)
3762 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3765 static inline void show_node(struct zone *zone)
3767 if (IS_ENABLED(CONFIG_NUMA))
3768 printk("Node %d ", zone_to_nid(zone));
3771 long si_mem_available(void)
3774 unsigned long pagecache;
3775 unsigned long wmark_low = 0;
3776 unsigned long pages[NR_LRU_LISTS];
3780 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
3781 pages[lru] = global_page_state(NR_LRU_BASE + lru);
3784 wmark_low += zone->watermark[WMARK_LOW];
3787 * Estimate the amount of memory available for userspace allocations,
3788 * without causing swapping.
3790 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
3793 * Not all the page cache can be freed, otherwise the system will
3794 * start swapping. Assume at least half of the page cache, or the
3795 * low watermark worth of cache, needs to stay.
3797 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
3798 pagecache -= min(pagecache / 2, wmark_low);
3799 available += pagecache;
3802 * Part of the reclaimable slab consists of items that are in use,
3803 * and cannot be freed. Cap this estimate at the low watermark.
3805 available += global_page_state(NR_SLAB_RECLAIMABLE) -
3806 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
3812 EXPORT_SYMBOL_GPL(si_mem_available);
3814 void si_meminfo(struct sysinfo *val)
3816 val->totalram = totalram_pages;
3817 val->sharedram = global_page_state(NR_SHMEM);
3818 val->freeram = global_page_state(NR_FREE_PAGES);
3819 val->bufferram = nr_blockdev_pages();
3820 val->totalhigh = totalhigh_pages;
3821 val->freehigh = nr_free_highpages();
3822 val->mem_unit = PAGE_SIZE;
3825 EXPORT_SYMBOL(si_meminfo);
3828 void si_meminfo_node(struct sysinfo *val, int nid)
3830 int zone_type; /* needs to be signed */
3831 unsigned long managed_pages = 0;
3832 unsigned long managed_highpages = 0;
3833 unsigned long free_highpages = 0;
3834 pg_data_t *pgdat = NODE_DATA(nid);
3836 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3837 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3838 val->totalram = managed_pages;
3839 val->sharedram = node_page_state(nid, NR_SHMEM);
3840 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3841 #ifdef CONFIG_HIGHMEM
3842 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3843 struct zone *zone = &pgdat->node_zones[zone_type];
3845 if (is_highmem(zone)) {
3846 managed_highpages += zone->managed_pages;
3847 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
3850 val->totalhigh = managed_highpages;
3851 val->freehigh = free_highpages;
3853 val->totalhigh = managed_highpages;
3854 val->freehigh = free_highpages;
3856 val->mem_unit = PAGE_SIZE;
3861 * Determine whether the node should be displayed or not, depending on whether
3862 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3864 bool skip_free_areas_node(unsigned int flags, int nid)
3867 unsigned int cpuset_mems_cookie;
3869 if (!(flags & SHOW_MEM_FILTER_NODES))
3873 cpuset_mems_cookie = read_mems_allowed_begin();
3874 ret = !node_isset(nid, cpuset_current_mems_allowed);
3875 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3880 #define K(x) ((x) << (PAGE_SHIFT-10))
3882 static void show_migration_types(unsigned char type)
3884 static const char types[MIGRATE_TYPES] = {
3885 [MIGRATE_UNMOVABLE] = 'U',
3886 [MIGRATE_MOVABLE] = 'M',
3887 [MIGRATE_RECLAIMABLE] = 'E',
3888 [MIGRATE_HIGHATOMIC] = 'H',
3890 [MIGRATE_CMA] = 'C',
3892 #ifdef CONFIG_MEMORY_ISOLATION
3893 [MIGRATE_ISOLATE] = 'I',
3896 char tmp[MIGRATE_TYPES + 1];
3900 for (i = 0; i < MIGRATE_TYPES; i++) {
3901 if (type & (1 << i))
3906 printk("(%s) ", tmp);
3910 * Show free area list (used inside shift_scroll-lock stuff)
3911 * We also calculate the percentage fragmentation. We do this by counting the
3912 * memory on each free list with the exception of the first item on the list.
3915 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3918 void show_free_areas(unsigned int filter)
3920 unsigned long free_pcp = 0;
3924 for_each_populated_zone(zone) {
3925 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3928 for_each_online_cpu(cpu)
3929 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3932 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3933 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3934 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3935 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3936 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3937 " free:%lu free_pcp:%lu free_cma:%lu\n",
3938 global_page_state(NR_ACTIVE_ANON),
3939 global_page_state(NR_INACTIVE_ANON),
3940 global_page_state(NR_ISOLATED_ANON),
3941 global_page_state(NR_ACTIVE_FILE),
3942 global_page_state(NR_INACTIVE_FILE),
3943 global_page_state(NR_ISOLATED_FILE),
3944 global_page_state(NR_UNEVICTABLE),
3945 global_page_state(NR_FILE_DIRTY),
3946 global_page_state(NR_WRITEBACK),
3947 global_page_state(NR_UNSTABLE_NFS),
3948 global_page_state(NR_SLAB_RECLAIMABLE),
3949 global_page_state(NR_SLAB_UNRECLAIMABLE),
3950 global_page_state(NR_FILE_MAPPED),
3951 global_page_state(NR_SHMEM),
3952 global_page_state(NR_PAGETABLE),
3953 global_page_state(NR_BOUNCE),
3954 global_page_state(NR_FREE_PAGES),
3956 global_page_state(NR_FREE_CMA_PAGES));
3958 for_each_populated_zone(zone) {
3961 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3965 for_each_online_cpu(cpu)
3966 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3974 " active_anon:%lukB"
3975 " inactive_anon:%lukB"
3976 " active_file:%lukB"
3977 " inactive_file:%lukB"
3978 " unevictable:%lukB"
3979 " isolated(anon):%lukB"
3980 " isolated(file):%lukB"
3988 " slab_reclaimable:%lukB"
3989 " slab_unreclaimable:%lukB"
3990 " kernel_stack:%lukB"
3997 " writeback_tmp:%lukB"
3998 " pages_scanned:%lu"
3999 " all_unreclaimable? %s"
4002 K(zone_page_state(zone, NR_FREE_PAGES)),
4003 K(min_wmark_pages(zone)),
4004 K(low_wmark_pages(zone)),
4005 K(high_wmark_pages(zone)),
4006 K(zone_page_state(zone, NR_ACTIVE_ANON)),
4007 K(zone_page_state(zone, NR_INACTIVE_ANON)),
4008 K(zone_page_state(zone, NR_ACTIVE_FILE)),
4009 K(zone_page_state(zone, NR_INACTIVE_FILE)),
4010 K(zone_page_state(zone, NR_UNEVICTABLE)),
4011 K(zone_page_state(zone, NR_ISOLATED_ANON)),
4012 K(zone_page_state(zone, NR_ISOLATED_FILE)),
4013 K(zone->present_pages),
4014 K(zone->managed_pages),
4015 K(zone_page_state(zone, NR_MLOCK)),
4016 K(zone_page_state(zone, NR_FILE_DIRTY)),
4017 K(zone_page_state(zone, NR_WRITEBACK)),
4018 K(zone_page_state(zone, NR_FILE_MAPPED)),
4019 K(zone_page_state(zone, NR_SHMEM)),
4020 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
4021 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
4022 zone_page_state(zone, NR_KERNEL_STACK) *
4024 K(zone_page_state(zone, NR_PAGETABLE)),
4025 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
4026 K(zone_page_state(zone, NR_BOUNCE)),
4028 K(this_cpu_read(zone->pageset->pcp.count)),
4029 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
4030 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
4031 K(zone_page_state(zone, NR_PAGES_SCANNED)),
4032 (!zone_reclaimable(zone) ? "yes" : "no")
4034 printk("lowmem_reserve[]:");
4035 for (i = 0; i < MAX_NR_ZONES; i++)
4036 printk(" %ld", zone->lowmem_reserve[i]);
4040 for_each_populated_zone(zone) {
4042 unsigned long nr[MAX_ORDER], flags, total = 0;
4043 unsigned char types[MAX_ORDER];
4045 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4048 printk("%s: ", zone->name);
4050 spin_lock_irqsave(&zone->lock, flags);
4051 for (order = 0; order < MAX_ORDER; order++) {
4052 struct free_area *area = &zone->free_area[order];
4055 nr[order] = area->nr_free;
4056 total += nr[order] << order;
4059 for (type = 0; type < MIGRATE_TYPES; type++) {
4060 if (!list_empty(&area->free_list[type]))
4061 types[order] |= 1 << type;
4064 spin_unlock_irqrestore(&zone->lock, flags);
4065 for (order = 0; order < MAX_ORDER; order++) {
4066 printk("%lu*%lukB ", nr[order], K(1UL) << order);
4068 show_migration_types(types[order]);
4070 printk("= %lukB\n", K(total));
4073 hugetlb_show_meminfo();
4075 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
4077 show_swap_cache_info();
4080 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4082 zoneref->zone = zone;
4083 zoneref->zone_idx = zone_idx(zone);
4087 * Builds allocation fallback zone lists.
4089 * Add all populated zones of a node to the zonelist.
4091 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4095 enum zone_type zone_type = MAX_NR_ZONES;
4099 zone = pgdat->node_zones + zone_type;
4100 if (populated_zone(zone)) {
4101 zoneref_set_zone(zone,
4102 &zonelist->_zonerefs[nr_zones++]);
4103 check_highest_zone(zone_type);
4105 } while (zone_type);
4113 * 0 = automatic detection of better ordering.
4114 * 1 = order by ([node] distance, -zonetype)
4115 * 2 = order by (-zonetype, [node] distance)
4117 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4118 * the same zonelist. So only NUMA can configure this param.
4120 #define ZONELIST_ORDER_DEFAULT 0
4121 #define ZONELIST_ORDER_NODE 1
4122 #define ZONELIST_ORDER_ZONE 2
4124 /* zonelist order in the kernel.
4125 * set_zonelist_order() will set this to NODE or ZONE.
4127 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4128 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4132 /* The value user specified ....changed by config */
4133 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4134 /* string for sysctl */
4135 #define NUMA_ZONELIST_ORDER_LEN 16
4136 char numa_zonelist_order[16] = "default";
4139 * interface for configure zonelist ordering.
4140 * command line option "numa_zonelist_order"
4141 * = "[dD]efault - default, automatic configuration.
4142 * = "[nN]ode - order by node locality, then by zone within node
4143 * = "[zZ]one - order by zone, then by locality within zone
4146 static int __parse_numa_zonelist_order(char *s)
4148 if (*s == 'd' || *s == 'D') {
4149 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4150 } else if (*s == 'n' || *s == 'N') {
4151 user_zonelist_order = ZONELIST_ORDER_NODE;
4152 } else if (*s == 'z' || *s == 'Z') {
4153 user_zonelist_order = ZONELIST_ORDER_ZONE;
4155 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4161 static __init int setup_numa_zonelist_order(char *s)
4168 ret = __parse_numa_zonelist_order(s);
4170 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4174 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4177 * sysctl handler for numa_zonelist_order
4179 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4180 void __user *buffer, size_t *length,
4183 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4185 static DEFINE_MUTEX(zl_order_mutex);
4187 mutex_lock(&zl_order_mutex);
4189 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4193 strcpy(saved_string, (char *)table->data);
4195 ret = proc_dostring(table, write, buffer, length, ppos);
4199 int oldval = user_zonelist_order;
4201 ret = __parse_numa_zonelist_order((char *)table->data);
4204 * bogus value. restore saved string
4206 strncpy((char *)table->data, saved_string,
4207 NUMA_ZONELIST_ORDER_LEN);
4208 user_zonelist_order = oldval;
4209 } else if (oldval != user_zonelist_order) {
4210 mutex_lock(&zonelists_mutex);
4211 build_all_zonelists(NULL, NULL);
4212 mutex_unlock(&zonelists_mutex);
4216 mutex_unlock(&zl_order_mutex);
4221 #define MAX_NODE_LOAD (nr_online_nodes)
4222 static int node_load[MAX_NUMNODES];
4225 * find_next_best_node - find the next node that should appear in a given node's fallback list
4226 * @node: node whose fallback list we're appending
4227 * @used_node_mask: nodemask_t of already used nodes
4229 * We use a number of factors to determine which is the next node that should
4230 * appear on a given node's fallback list. The node should not have appeared
4231 * already in @node's fallback list, and it should be the next closest node
4232 * according to the distance array (which contains arbitrary distance values
4233 * from each node to each node in the system), and should also prefer nodes
4234 * with no CPUs, since presumably they'll have very little allocation pressure
4235 * on them otherwise.
4236 * It returns -1 if no node is found.
4238 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4241 int min_val = INT_MAX;
4242 int best_node = NUMA_NO_NODE;
4243 const struct cpumask *tmp = cpumask_of_node(0);
4245 /* Use the local node if we haven't already */
4246 if (!node_isset(node, *used_node_mask)) {
4247 node_set(node, *used_node_mask);
4251 for_each_node_state(n, N_MEMORY) {
4253 /* Don't want a node to appear more than once */
4254 if (node_isset(n, *used_node_mask))
4257 /* Use the distance array to find the distance */
4258 val = node_distance(node, n);
4260 /* Penalize nodes under us ("prefer the next node") */
4263 /* Give preference to headless and unused nodes */
4264 tmp = cpumask_of_node(n);
4265 if (!cpumask_empty(tmp))
4266 val += PENALTY_FOR_NODE_WITH_CPUS;
4268 /* Slight preference for less loaded node */
4269 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4270 val += node_load[n];
4272 if (val < min_val) {
4279 node_set(best_node, *used_node_mask);
4286 * Build zonelists ordered by node and zones within node.
4287 * This results in maximum locality--normal zone overflows into local
4288 * DMA zone, if any--but risks exhausting DMA zone.
4290 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4293 struct zonelist *zonelist;
4295 zonelist = &pgdat->node_zonelists[0];
4296 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4298 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4299 zonelist->_zonerefs[j].zone = NULL;
4300 zonelist->_zonerefs[j].zone_idx = 0;
4304 * Build gfp_thisnode zonelists
4306 static void build_thisnode_zonelists(pg_data_t *pgdat)
4309 struct zonelist *zonelist;
4311 zonelist = &pgdat->node_zonelists[1];
4312 j = build_zonelists_node(pgdat, zonelist, 0);
4313 zonelist->_zonerefs[j].zone = NULL;
4314 zonelist->_zonerefs[j].zone_idx = 0;
4318 * Build zonelists ordered by zone and nodes within zones.
4319 * This results in conserving DMA zone[s] until all Normal memory is
4320 * exhausted, but results in overflowing to remote node while memory
4321 * may still exist in local DMA zone.
4323 static int node_order[MAX_NUMNODES];
4325 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4328 int zone_type; /* needs to be signed */
4330 struct zonelist *zonelist;
4332 zonelist = &pgdat->node_zonelists[0];
4334 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4335 for (j = 0; j < nr_nodes; j++) {
4336 node = node_order[j];
4337 z = &NODE_DATA(node)->node_zones[zone_type];
4338 if (populated_zone(z)) {
4340 &zonelist->_zonerefs[pos++]);
4341 check_highest_zone(zone_type);
4345 zonelist->_zonerefs[pos].zone = NULL;
4346 zonelist->_zonerefs[pos].zone_idx = 0;
4349 #if defined(CONFIG_64BIT)
4351 * Devices that require DMA32/DMA are relatively rare and do not justify a
4352 * penalty to every machine in case the specialised case applies. Default
4353 * to Node-ordering on 64-bit NUMA machines
4355 static int default_zonelist_order(void)
4357 return ZONELIST_ORDER_NODE;
4361 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4362 * by the kernel. If processes running on node 0 deplete the low memory zone
4363 * then reclaim will occur more frequency increasing stalls and potentially
4364 * be easier to OOM if a large percentage of the zone is under writeback or
4365 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4366 * Hence, default to zone ordering on 32-bit.
4368 static int default_zonelist_order(void)
4370 return ZONELIST_ORDER_ZONE;
4372 #endif /* CONFIG_64BIT */
4374 static void set_zonelist_order(void)
4376 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4377 current_zonelist_order = default_zonelist_order();
4379 current_zonelist_order = user_zonelist_order;
4382 static void build_zonelists(pg_data_t *pgdat)
4385 nodemask_t used_mask;
4386 int local_node, prev_node;
4387 struct zonelist *zonelist;
4388 unsigned int order = current_zonelist_order;
4390 /* initialize zonelists */
4391 for (i = 0; i < MAX_ZONELISTS; i++) {
4392 zonelist = pgdat->node_zonelists + i;
4393 zonelist->_zonerefs[0].zone = NULL;
4394 zonelist->_zonerefs[0].zone_idx = 0;
4397 /* NUMA-aware ordering of nodes */
4398 local_node = pgdat->node_id;
4399 load = nr_online_nodes;
4400 prev_node = local_node;
4401 nodes_clear(used_mask);
4403 memset(node_order, 0, sizeof(node_order));
4406 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4408 * We don't want to pressure a particular node.
4409 * So adding penalty to the first node in same
4410 * distance group to make it round-robin.
4412 if (node_distance(local_node, node) !=
4413 node_distance(local_node, prev_node))
4414 node_load[node] = load;
4418 if (order == ZONELIST_ORDER_NODE)
4419 build_zonelists_in_node_order(pgdat, node);
4421 node_order[i++] = node; /* remember order */
4424 if (order == ZONELIST_ORDER_ZONE) {
4425 /* calculate node order -- i.e., DMA last! */
4426 build_zonelists_in_zone_order(pgdat, i);
4429 build_thisnode_zonelists(pgdat);
4432 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4434 * Return node id of node used for "local" allocations.
4435 * I.e., first node id of first zone in arg node's generic zonelist.
4436 * Used for initializing percpu 'numa_mem', which is used primarily
4437 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4439 int local_memory_node(int node)
4443 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4444 gfp_zone(GFP_KERNEL),
4451 #else /* CONFIG_NUMA */
4453 static void set_zonelist_order(void)
4455 current_zonelist_order = ZONELIST_ORDER_ZONE;
4458 static void build_zonelists(pg_data_t *pgdat)
4460 int node, local_node;
4462 struct zonelist *zonelist;
4464 local_node = pgdat->node_id;
4466 zonelist = &pgdat->node_zonelists[0];
4467 j = build_zonelists_node(pgdat, zonelist, 0);
4470 * Now we build the zonelist so that it contains the zones
4471 * of all the other nodes.
4472 * We don't want to pressure a particular node, so when
4473 * building the zones for node N, we make sure that the
4474 * zones coming right after the local ones are those from
4475 * node N+1 (modulo N)
4477 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4478 if (!node_online(node))
4480 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4482 for (node = 0; node < local_node; node++) {
4483 if (!node_online(node))
4485 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4488 zonelist->_zonerefs[j].zone = NULL;
4489 zonelist->_zonerefs[j].zone_idx = 0;
4492 #endif /* CONFIG_NUMA */
4495 * Boot pageset table. One per cpu which is going to be used for all
4496 * zones and all nodes. The parameters will be set in such a way
4497 * that an item put on a list will immediately be handed over to
4498 * the buddy list. This is safe since pageset manipulation is done
4499 * with interrupts disabled.
4501 * The boot_pagesets must be kept even after bootup is complete for
4502 * unused processors and/or zones. They do play a role for bootstrapping
4503 * hotplugged processors.
4505 * zoneinfo_show() and maybe other functions do
4506 * not check if the processor is online before following the pageset pointer.
4507 * Other parts of the kernel may not check if the zone is available.
4509 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4510 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4511 static void setup_zone_pageset(struct zone *zone);
4514 * Global mutex to protect against size modification of zonelists
4515 * as well as to serialize pageset setup for the new populated zone.
4517 DEFINE_MUTEX(zonelists_mutex);
4519 /* return values int ....just for stop_machine() */
4520 static int __build_all_zonelists(void *data)
4524 pg_data_t *self = data;
4527 memset(node_load, 0, sizeof(node_load));
4530 if (self && !node_online(self->node_id)) {
4531 build_zonelists(self);
4534 for_each_online_node(nid) {
4535 pg_data_t *pgdat = NODE_DATA(nid);
4537 build_zonelists(pgdat);
4541 * Initialize the boot_pagesets that are going to be used
4542 * for bootstrapping processors. The real pagesets for
4543 * each zone will be allocated later when the per cpu
4544 * allocator is available.
4546 * boot_pagesets are used also for bootstrapping offline
4547 * cpus if the system is already booted because the pagesets
4548 * are needed to initialize allocators on a specific cpu too.
4549 * F.e. the percpu allocator needs the page allocator which
4550 * needs the percpu allocator in order to allocate its pagesets
4551 * (a chicken-egg dilemma).
4553 for_each_possible_cpu(cpu) {
4554 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4556 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4558 * We now know the "local memory node" for each node--
4559 * i.e., the node of the first zone in the generic zonelist.
4560 * Set up numa_mem percpu variable for on-line cpus. During
4561 * boot, only the boot cpu should be on-line; we'll init the
4562 * secondary cpus' numa_mem as they come on-line. During
4563 * node/memory hotplug, we'll fixup all on-line cpus.
4565 if (cpu_online(cpu))
4566 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4573 static noinline void __init
4574 build_all_zonelists_init(void)
4576 __build_all_zonelists(NULL);
4577 mminit_verify_zonelist();
4578 cpuset_init_current_mems_allowed();
4582 * Called with zonelists_mutex held always
4583 * unless system_state == SYSTEM_BOOTING.
4585 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4586 * [we're only called with non-NULL zone through __meminit paths] and
4587 * (2) call of __init annotated helper build_all_zonelists_init
4588 * [protected by SYSTEM_BOOTING].
4590 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4592 set_zonelist_order();
4594 if (system_state == SYSTEM_BOOTING) {
4595 build_all_zonelists_init();
4597 #ifdef CONFIG_MEMORY_HOTPLUG
4599 setup_zone_pageset(zone);
4601 /* we have to stop all cpus to guarantee there is no user
4603 stop_machine(__build_all_zonelists, pgdat, NULL);
4604 /* cpuset refresh routine should be here */
4606 vm_total_pages = nr_free_pagecache_pages();
4608 * Disable grouping by mobility if the number of pages in the
4609 * system is too low to allow the mechanism to work. It would be
4610 * more accurate, but expensive to check per-zone. This check is
4611 * made on memory-hotadd so a system can start with mobility
4612 * disabled and enable it later
4614 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4615 page_group_by_mobility_disabled = 1;
4617 page_group_by_mobility_disabled = 0;
4619 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
4621 zonelist_order_name[current_zonelist_order],
4622 page_group_by_mobility_disabled ? "off" : "on",
4625 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4630 * Helper functions to size the waitqueue hash table.
4631 * Essentially these want to choose hash table sizes sufficiently
4632 * large so that collisions trying to wait on pages are rare.
4633 * But in fact, the number of active page waitqueues on typical
4634 * systems is ridiculously low, less than 200. So this is even
4635 * conservative, even though it seems large.
4637 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4638 * waitqueues, i.e. the size of the waitq table given the number of pages.
4640 #define PAGES_PER_WAITQUEUE 256
4642 #ifndef CONFIG_MEMORY_HOTPLUG
4643 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4645 unsigned long size = 1;
4647 pages /= PAGES_PER_WAITQUEUE;
4649 while (size < pages)
4653 * Once we have dozens or even hundreds of threads sleeping
4654 * on IO we've got bigger problems than wait queue collision.
4655 * Limit the size of the wait table to a reasonable size.
4657 size = min(size, 4096UL);
4659 return max(size, 4UL);
4663 * A zone's size might be changed by hot-add, so it is not possible to determine
4664 * a suitable size for its wait_table. So we use the maximum size now.
4666 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4668 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4669 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4670 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4672 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4673 * or more by the traditional way. (See above). It equals:
4675 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4676 * ia64(16K page size) : = ( 8G + 4M)byte.
4677 * powerpc (64K page size) : = (32G +16M)byte.
4679 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4686 * This is an integer logarithm so that shifts can be used later
4687 * to extract the more random high bits from the multiplicative
4688 * hash function before the remainder is taken.
4690 static inline unsigned long wait_table_bits(unsigned long size)
4696 * Initially all pages are reserved - free ones are freed
4697 * up by free_all_bootmem() once the early boot process is
4698 * done. Non-atomic initialization, single-pass.
4700 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4701 unsigned long start_pfn, enum memmap_context context)
4703 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
4704 unsigned long end_pfn = start_pfn + size;
4705 pg_data_t *pgdat = NODE_DATA(nid);
4707 unsigned long nr_initialised = 0;
4708 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4709 struct memblock_region *r = NULL, *tmp;
4712 if (highest_memmap_pfn < end_pfn - 1)
4713 highest_memmap_pfn = end_pfn - 1;
4716 * Honor reservation requested by the driver for this ZONE_DEVICE
4719 if (altmap && start_pfn == altmap->base_pfn)
4720 start_pfn += altmap->reserve;
4722 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4724 * There can be holes in boot-time mem_map[]s handed to this
4725 * function. They do not exist on hotplugged memory.
4727 if (context != MEMMAP_EARLY)
4730 if (!early_pfn_valid(pfn))
4732 if (!early_pfn_in_nid(pfn, nid))
4734 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
4737 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4739 * If not mirrored_kernelcore and ZONE_MOVABLE exists, range
4740 * from zone_movable_pfn[nid] to end of each node should be
4741 * ZONE_MOVABLE not ZONE_NORMAL. skip it.
4743 if (!mirrored_kernelcore && zone_movable_pfn[nid])
4744 if (zone == ZONE_NORMAL && pfn >= zone_movable_pfn[nid])
4748 * Check given memblock attribute by firmware which can affect
4749 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
4750 * mirrored, it's an overlapped memmap init. skip it.
4752 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
4753 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
4754 for_each_memblock(memory, tmp)
4755 if (pfn < memblock_region_memory_end_pfn(tmp))
4759 if (pfn >= memblock_region_memory_base_pfn(r) &&
4760 memblock_is_mirror(r)) {
4761 /* already initialized as NORMAL */
4762 pfn = memblock_region_memory_end_pfn(r);
4770 * Mark the block movable so that blocks are reserved for
4771 * movable at startup. This will force kernel allocations
4772 * to reserve their blocks rather than leaking throughout
4773 * the address space during boot when many long-lived
4774 * kernel allocations are made.
4776 * bitmap is created for zone's valid pfn range. but memmap
4777 * can be created for invalid pages (for alignment)
4778 * check here not to call set_pageblock_migratetype() against
4781 if (!(pfn & (pageblock_nr_pages - 1))) {
4782 struct page *page = pfn_to_page(pfn);
4784 __init_single_page(page, pfn, zone, nid);
4785 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4787 __init_single_pfn(pfn, zone, nid);
4792 static void __meminit zone_init_free_lists(struct zone *zone)
4794 unsigned int order, t;
4795 for_each_migratetype_order(order, t) {
4796 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4797 zone->free_area[order].nr_free = 0;
4801 #ifndef __HAVE_ARCH_MEMMAP_INIT
4802 #define memmap_init(size, nid, zone, start_pfn) \
4803 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4806 static int zone_batchsize(struct zone *zone)
4812 * The per-cpu-pages pools are set to around 1000th of the
4813 * size of the zone. But no more than 1/2 of a meg.
4815 * OK, so we don't know how big the cache is. So guess.
4817 batch = zone->managed_pages / 1024;
4818 if (batch * PAGE_SIZE > 512 * 1024)
4819 batch = (512 * 1024) / PAGE_SIZE;
4820 batch /= 4; /* We effectively *= 4 below */
4825 * Clamp the batch to a 2^n - 1 value. Having a power
4826 * of 2 value was found to be more likely to have
4827 * suboptimal cache aliasing properties in some cases.
4829 * For example if 2 tasks are alternately allocating
4830 * batches of pages, one task can end up with a lot
4831 * of pages of one half of the possible page colors
4832 * and the other with pages of the other colors.
4834 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4839 /* The deferral and batching of frees should be suppressed under NOMMU
4842 * The problem is that NOMMU needs to be able to allocate large chunks
4843 * of contiguous memory as there's no hardware page translation to
4844 * assemble apparent contiguous memory from discontiguous pages.
4846 * Queueing large contiguous runs of pages for batching, however,
4847 * causes the pages to actually be freed in smaller chunks. As there
4848 * can be a significant delay between the individual batches being
4849 * recycled, this leads to the once large chunks of space being
4850 * fragmented and becoming unavailable for high-order allocations.
4857 * pcp->high and pcp->batch values are related and dependent on one another:
4858 * ->batch must never be higher then ->high.
4859 * The following function updates them in a safe manner without read side
4862 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4863 * those fields changing asynchronously (acording the the above rule).
4865 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4866 * outside of boot time (or some other assurance that no concurrent updaters
4869 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4870 unsigned long batch)
4872 /* start with a fail safe value for batch */
4876 /* Update high, then batch, in order */
4883 /* a companion to pageset_set_high() */
4884 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4886 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4889 static void pageset_init(struct per_cpu_pageset *p)
4891 struct per_cpu_pages *pcp;
4894 memset(p, 0, sizeof(*p));
4898 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4899 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4902 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4905 pageset_set_batch(p, batch);
4909 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4910 * to the value high for the pageset p.
4912 static void pageset_set_high(struct per_cpu_pageset *p,
4915 unsigned long batch = max(1UL, high / 4);
4916 if ((high / 4) > (PAGE_SHIFT * 8))
4917 batch = PAGE_SHIFT * 8;
4919 pageset_update(&p->pcp, high, batch);
4922 static void pageset_set_high_and_batch(struct zone *zone,
4923 struct per_cpu_pageset *pcp)
4925 if (percpu_pagelist_fraction)
4926 pageset_set_high(pcp,
4927 (zone->managed_pages /
4928 percpu_pagelist_fraction));
4930 pageset_set_batch(pcp, zone_batchsize(zone));
4933 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4935 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4938 pageset_set_high_and_batch(zone, pcp);
4941 static void __meminit setup_zone_pageset(struct zone *zone)
4944 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4945 for_each_possible_cpu(cpu)
4946 zone_pageset_init(zone, cpu);
4950 * Allocate per cpu pagesets and initialize them.
4951 * Before this call only boot pagesets were available.
4953 void __init setup_per_cpu_pageset(void)
4957 for_each_populated_zone(zone)
4958 setup_zone_pageset(zone);
4961 static noinline __init_refok
4962 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4968 * The per-page waitqueue mechanism uses hashed waitqueues
4971 zone->wait_table_hash_nr_entries =
4972 wait_table_hash_nr_entries(zone_size_pages);
4973 zone->wait_table_bits =
4974 wait_table_bits(zone->wait_table_hash_nr_entries);
4975 alloc_size = zone->wait_table_hash_nr_entries
4976 * sizeof(wait_queue_head_t);
4978 if (!slab_is_available()) {
4979 zone->wait_table = (wait_queue_head_t *)
4980 memblock_virt_alloc_node_nopanic(
4981 alloc_size, zone->zone_pgdat->node_id);
4984 * This case means that a zone whose size was 0 gets new memory
4985 * via memory hot-add.
4986 * But it may be the case that a new node was hot-added. In
4987 * this case vmalloc() will not be able to use this new node's
4988 * memory - this wait_table must be initialized to use this new
4989 * node itself as well.
4990 * To use this new node's memory, further consideration will be
4993 zone->wait_table = vmalloc(alloc_size);
4995 if (!zone->wait_table)
4998 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4999 init_waitqueue_head(zone->wait_table + i);
5004 static __meminit void zone_pcp_init(struct zone *zone)
5007 * per cpu subsystem is not up at this point. The following code
5008 * relies on the ability of the linker to provide the
5009 * offset of a (static) per cpu variable into the per cpu area.
5011 zone->pageset = &boot_pageset;
5013 if (populated_zone(zone))
5014 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5015 zone->name, zone->present_pages,
5016 zone_batchsize(zone));
5019 int __meminit init_currently_empty_zone(struct zone *zone,
5020 unsigned long zone_start_pfn,
5023 struct pglist_data *pgdat = zone->zone_pgdat;
5025 ret = zone_wait_table_init(zone, size);
5028 pgdat->nr_zones = zone_idx(zone) + 1;
5030 zone->zone_start_pfn = zone_start_pfn;
5032 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5033 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5035 (unsigned long)zone_idx(zone),
5036 zone_start_pfn, (zone_start_pfn + size));
5038 zone_init_free_lists(zone);
5043 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5044 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5047 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5049 int __meminit __early_pfn_to_nid(unsigned long pfn,
5050 struct mminit_pfnnid_cache *state)
5052 unsigned long start_pfn, end_pfn;
5055 if (state->last_start <= pfn && pfn < state->last_end)
5056 return state->last_nid;
5058 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5060 state->last_start = start_pfn;
5061 state->last_end = end_pfn;
5062 state->last_nid = nid;
5067 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5070 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5071 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5072 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5074 * If an architecture guarantees that all ranges registered contain no holes
5075 * and may be freed, this this function may be used instead of calling
5076 * memblock_free_early_nid() manually.
5078 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5080 unsigned long start_pfn, end_pfn;
5083 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5084 start_pfn = min(start_pfn, max_low_pfn);
5085 end_pfn = min(end_pfn, max_low_pfn);
5087 if (start_pfn < end_pfn)
5088 memblock_free_early_nid(PFN_PHYS(start_pfn),
5089 (end_pfn - start_pfn) << PAGE_SHIFT,
5095 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5096 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5098 * If an architecture guarantees that all ranges registered contain no holes and may
5099 * be freed, this function may be used instead of calling memory_present() manually.
5101 void __init sparse_memory_present_with_active_regions(int nid)
5103 unsigned long start_pfn, end_pfn;
5106 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5107 memory_present(this_nid, start_pfn, end_pfn);
5111 * get_pfn_range_for_nid - Return the start and end page frames for a node
5112 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5113 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5114 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5116 * It returns the start and end page frame of a node based on information
5117 * provided by memblock_set_node(). If called for a node
5118 * with no available memory, a warning is printed and the start and end
5121 void __meminit get_pfn_range_for_nid(unsigned int nid,
5122 unsigned long *start_pfn, unsigned long *end_pfn)
5124 unsigned long this_start_pfn, this_end_pfn;
5130 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5131 *start_pfn = min(*start_pfn, this_start_pfn);
5132 *end_pfn = max(*end_pfn, this_end_pfn);
5135 if (*start_pfn == -1UL)
5140 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5141 * assumption is made that zones within a node are ordered in monotonic
5142 * increasing memory addresses so that the "highest" populated zone is used
5144 static void __init find_usable_zone_for_movable(void)
5147 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5148 if (zone_index == ZONE_MOVABLE)
5151 if (arch_zone_highest_possible_pfn[zone_index] >
5152 arch_zone_lowest_possible_pfn[zone_index])
5156 VM_BUG_ON(zone_index == -1);
5157 movable_zone = zone_index;
5161 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5162 * because it is sized independent of architecture. Unlike the other zones,
5163 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5164 * in each node depending on the size of each node and how evenly kernelcore
5165 * is distributed. This helper function adjusts the zone ranges
5166 * provided by the architecture for a given node by using the end of the
5167 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5168 * zones within a node are in order of monotonic increases memory addresses
5170 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5171 unsigned long zone_type,
5172 unsigned long node_start_pfn,
5173 unsigned long node_end_pfn,
5174 unsigned long *zone_start_pfn,
5175 unsigned long *zone_end_pfn)
5177 /* Only adjust if ZONE_MOVABLE is on this node */
5178 if (zone_movable_pfn[nid]) {
5179 /* Size ZONE_MOVABLE */
5180 if (zone_type == ZONE_MOVABLE) {
5181 *zone_start_pfn = zone_movable_pfn[nid];
5182 *zone_end_pfn = min(node_end_pfn,
5183 arch_zone_highest_possible_pfn[movable_zone]);
5185 /* Check if this whole range is within ZONE_MOVABLE */
5186 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5187 *zone_start_pfn = *zone_end_pfn;
5192 * Return the number of pages a zone spans in a node, including holes
5193 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5195 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5196 unsigned long zone_type,
5197 unsigned long node_start_pfn,
5198 unsigned long node_end_pfn,
5199 unsigned long *zone_start_pfn,
5200 unsigned long *zone_end_pfn,
5201 unsigned long *ignored)
5203 /* When hotadd a new node from cpu_up(), the node should be empty */
5204 if (!node_start_pfn && !node_end_pfn)
5207 /* Get the start and end of the zone */
5208 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5209 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5210 adjust_zone_range_for_zone_movable(nid, zone_type,
5211 node_start_pfn, node_end_pfn,
5212 zone_start_pfn, zone_end_pfn);
5214 /* Check that this node has pages within the zone's required range */
5215 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5218 /* Move the zone boundaries inside the node if necessary */
5219 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5220 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5222 /* Return the spanned pages */
5223 return *zone_end_pfn - *zone_start_pfn;
5227 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5228 * then all holes in the requested range will be accounted for.
5230 unsigned long __meminit __absent_pages_in_range(int nid,
5231 unsigned long range_start_pfn,
5232 unsigned long range_end_pfn)
5234 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5235 unsigned long start_pfn, end_pfn;
5238 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5239 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5240 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5241 nr_absent -= end_pfn - start_pfn;
5247 * absent_pages_in_range - Return number of page frames in holes within a range
5248 * @start_pfn: The start PFN to start searching for holes
5249 * @end_pfn: The end PFN to stop searching for holes
5251 * It returns the number of pages frames in memory holes within a range.
5253 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5254 unsigned long end_pfn)
5256 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5259 /* Return the number of page frames in holes in a zone on a node */
5260 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5261 unsigned long zone_type,
5262 unsigned long node_start_pfn,
5263 unsigned long node_end_pfn,
5264 unsigned long *ignored)
5266 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5267 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5268 unsigned long zone_start_pfn, zone_end_pfn;
5269 unsigned long nr_absent;
5271 /* When hotadd a new node from cpu_up(), the node should be empty */
5272 if (!node_start_pfn && !node_end_pfn)
5275 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5276 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5278 adjust_zone_range_for_zone_movable(nid, zone_type,
5279 node_start_pfn, node_end_pfn,
5280 &zone_start_pfn, &zone_end_pfn);
5281 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5284 * ZONE_MOVABLE handling.
5285 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5288 if (zone_movable_pfn[nid]) {
5289 if (mirrored_kernelcore) {
5290 unsigned long start_pfn, end_pfn;
5291 struct memblock_region *r;
5293 for_each_memblock(memory, r) {
5294 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5295 zone_start_pfn, zone_end_pfn);
5296 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5297 zone_start_pfn, zone_end_pfn);
5299 if (zone_type == ZONE_MOVABLE &&
5300 memblock_is_mirror(r))
5301 nr_absent += end_pfn - start_pfn;
5303 if (zone_type == ZONE_NORMAL &&
5304 !memblock_is_mirror(r))
5305 nr_absent += end_pfn - start_pfn;
5308 if (zone_type == ZONE_NORMAL)
5309 nr_absent += node_end_pfn - zone_movable_pfn[nid];
5316 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5317 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5318 unsigned long zone_type,
5319 unsigned long node_start_pfn,
5320 unsigned long node_end_pfn,
5321 unsigned long *zone_start_pfn,
5322 unsigned long *zone_end_pfn,
5323 unsigned long *zones_size)
5327 *zone_start_pfn = node_start_pfn;
5328 for (zone = 0; zone < zone_type; zone++)
5329 *zone_start_pfn += zones_size[zone];
5331 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5333 return zones_size[zone_type];
5336 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5337 unsigned long zone_type,
5338 unsigned long node_start_pfn,
5339 unsigned long node_end_pfn,
5340 unsigned long *zholes_size)
5345 return zholes_size[zone_type];
5348 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5350 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5351 unsigned long node_start_pfn,
5352 unsigned long node_end_pfn,
5353 unsigned long *zones_size,
5354 unsigned long *zholes_size)
5356 unsigned long realtotalpages = 0, totalpages = 0;
5359 for (i = 0; i < MAX_NR_ZONES; i++) {
5360 struct zone *zone = pgdat->node_zones + i;
5361 unsigned long zone_start_pfn, zone_end_pfn;
5362 unsigned long size, real_size;
5364 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5370 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5371 node_start_pfn, node_end_pfn,
5374 zone->zone_start_pfn = zone_start_pfn;
5376 zone->zone_start_pfn = 0;
5377 zone->spanned_pages = size;
5378 zone->present_pages = real_size;
5381 realtotalpages += real_size;
5384 pgdat->node_spanned_pages = totalpages;
5385 pgdat->node_present_pages = realtotalpages;
5386 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5390 #ifndef CONFIG_SPARSEMEM
5392 * Calculate the size of the zone->blockflags rounded to an unsigned long
5393 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5394 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5395 * round what is now in bits to nearest long in bits, then return it in
5398 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5400 unsigned long usemapsize;
5402 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5403 usemapsize = roundup(zonesize, pageblock_nr_pages);
5404 usemapsize = usemapsize >> pageblock_order;
5405 usemapsize *= NR_PAGEBLOCK_BITS;
5406 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5408 return usemapsize / 8;
5411 static void __init setup_usemap(struct pglist_data *pgdat,
5413 unsigned long zone_start_pfn,
5414 unsigned long zonesize)
5416 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5417 zone->pageblock_flags = NULL;
5419 zone->pageblock_flags =
5420 memblock_virt_alloc_node_nopanic(usemapsize,
5424 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5425 unsigned long zone_start_pfn, unsigned long zonesize) {}
5426 #endif /* CONFIG_SPARSEMEM */
5428 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5430 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5431 void __paginginit set_pageblock_order(void)
5435 /* Check that pageblock_nr_pages has not already been setup */
5436 if (pageblock_order)
5439 if (HPAGE_SHIFT > PAGE_SHIFT)
5440 order = HUGETLB_PAGE_ORDER;
5442 order = MAX_ORDER - 1;
5445 * Assume the largest contiguous order of interest is a huge page.
5446 * This value may be variable depending on boot parameters on IA64 and
5449 pageblock_order = order;
5451 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5454 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5455 * is unused as pageblock_order is set at compile-time. See
5456 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5459 void __paginginit set_pageblock_order(void)
5463 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5465 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5466 unsigned long present_pages)
5468 unsigned long pages = spanned_pages;
5471 * Provide a more accurate estimation if there are holes within
5472 * the zone and SPARSEMEM is in use. If there are holes within the
5473 * zone, each populated memory region may cost us one or two extra
5474 * memmap pages due to alignment because memmap pages for each
5475 * populated regions may not naturally algined on page boundary.
5476 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5478 if (spanned_pages > present_pages + (present_pages >> 4) &&
5479 IS_ENABLED(CONFIG_SPARSEMEM))
5480 pages = present_pages;
5482 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5486 * Set up the zone data structures:
5487 * - mark all pages reserved
5488 * - mark all memory queues empty
5489 * - clear the memory bitmaps
5491 * NOTE: pgdat should get zeroed by caller.
5493 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5496 int nid = pgdat->node_id;
5499 pgdat_resize_init(pgdat);
5500 #ifdef CONFIG_NUMA_BALANCING
5501 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5502 pgdat->numabalancing_migrate_nr_pages = 0;
5503 pgdat->numabalancing_migrate_next_window = jiffies;
5505 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5506 spin_lock_init(&pgdat->split_queue_lock);
5507 INIT_LIST_HEAD(&pgdat->split_queue);
5508 pgdat->split_queue_len = 0;
5510 init_waitqueue_head(&pgdat->kswapd_wait);
5511 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5512 #ifdef CONFIG_COMPACTION
5513 init_waitqueue_head(&pgdat->kcompactd_wait);
5515 pgdat_page_ext_init(pgdat);
5517 for (j = 0; j < MAX_NR_ZONES; j++) {
5518 struct zone *zone = pgdat->node_zones + j;
5519 unsigned long size, realsize, freesize, memmap_pages;
5520 unsigned long zone_start_pfn = zone->zone_start_pfn;
5522 size = zone->spanned_pages;
5523 realsize = freesize = zone->present_pages;
5526 * Adjust freesize so that it accounts for how much memory
5527 * is used by this zone for memmap. This affects the watermark
5528 * and per-cpu initialisations
5530 memmap_pages = calc_memmap_size(size, realsize);
5531 if (!is_highmem_idx(j)) {
5532 if (freesize >= memmap_pages) {
5533 freesize -= memmap_pages;
5536 " %s zone: %lu pages used for memmap\n",
5537 zone_names[j], memmap_pages);
5539 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5540 zone_names[j], memmap_pages, freesize);
5543 /* Account for reserved pages */
5544 if (j == 0 && freesize > dma_reserve) {
5545 freesize -= dma_reserve;
5546 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5547 zone_names[0], dma_reserve);
5550 if (!is_highmem_idx(j))
5551 nr_kernel_pages += freesize;
5552 /* Charge for highmem memmap if there are enough kernel pages */
5553 else if (nr_kernel_pages > memmap_pages * 2)
5554 nr_kernel_pages -= memmap_pages;
5555 nr_all_pages += freesize;
5558 * Set an approximate value for lowmem here, it will be adjusted
5559 * when the bootmem allocator frees pages into the buddy system.
5560 * And all highmem pages will be managed by the buddy system.
5562 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5565 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5567 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5569 zone->name = zone_names[j];
5570 spin_lock_init(&zone->lock);
5571 spin_lock_init(&zone->lru_lock);
5572 zone_seqlock_init(zone);
5573 zone->zone_pgdat = pgdat;
5574 zone_pcp_init(zone);
5576 /* For bootup, initialized properly in watermark setup */
5577 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5579 lruvec_init(&zone->lruvec);
5583 set_pageblock_order();
5584 setup_usemap(pgdat, zone, zone_start_pfn, size);
5585 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5587 memmap_init(size, nid, j, zone_start_pfn);
5591 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5593 unsigned long __maybe_unused start = 0;
5594 unsigned long __maybe_unused offset = 0;
5596 /* Skip empty nodes */
5597 if (!pgdat->node_spanned_pages)
5600 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5601 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5602 offset = pgdat->node_start_pfn - start;
5603 /* ia64 gets its own node_mem_map, before this, without bootmem */
5604 if (!pgdat->node_mem_map) {
5605 unsigned long size, end;
5609 * The zone's endpoints aren't required to be MAX_ORDER
5610 * aligned but the node_mem_map endpoints must be in order
5611 * for the buddy allocator to function correctly.
5613 end = pgdat_end_pfn(pgdat);
5614 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5615 size = (end - start) * sizeof(struct page);
5616 map = alloc_remap(pgdat->node_id, size);
5618 map = memblock_virt_alloc_node_nopanic(size,
5620 pgdat->node_mem_map = map + offset;
5622 #ifndef CONFIG_NEED_MULTIPLE_NODES
5624 * With no DISCONTIG, the global mem_map is just set as node 0's
5626 if (pgdat == NODE_DATA(0)) {
5627 mem_map = NODE_DATA(0)->node_mem_map;
5628 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5629 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5631 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5634 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5637 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5638 unsigned long node_start_pfn, unsigned long *zholes_size)
5640 pg_data_t *pgdat = NODE_DATA(nid);
5641 unsigned long start_pfn = 0;
5642 unsigned long end_pfn = 0;
5644 /* pg_data_t should be reset to zero when it's allocated */
5645 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5647 reset_deferred_meminit(pgdat);
5648 pgdat->node_id = nid;
5649 pgdat->node_start_pfn = node_start_pfn;
5650 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5651 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5652 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5653 (u64)start_pfn << PAGE_SHIFT,
5654 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5656 start_pfn = node_start_pfn;
5658 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5659 zones_size, zholes_size);
5661 alloc_node_mem_map(pgdat);
5662 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5663 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5664 nid, (unsigned long)pgdat,
5665 (unsigned long)pgdat->node_mem_map);
5668 free_area_init_core(pgdat);
5671 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5673 #if MAX_NUMNODES > 1
5675 * Figure out the number of possible node ids.
5677 void __init setup_nr_node_ids(void)
5679 unsigned int highest;
5681 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5682 nr_node_ids = highest + 1;
5687 * node_map_pfn_alignment - determine the maximum internode alignment
5689 * This function should be called after node map is populated and sorted.
5690 * It calculates the maximum power of two alignment which can distinguish
5693 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5694 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5695 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5696 * shifted, 1GiB is enough and this function will indicate so.
5698 * This is used to test whether pfn -> nid mapping of the chosen memory
5699 * model has fine enough granularity to avoid incorrect mapping for the
5700 * populated node map.
5702 * Returns the determined alignment in pfn's. 0 if there is no alignment
5703 * requirement (single node).
5705 unsigned long __init node_map_pfn_alignment(void)
5707 unsigned long accl_mask = 0, last_end = 0;
5708 unsigned long start, end, mask;
5712 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5713 if (!start || last_nid < 0 || last_nid == nid) {
5720 * Start with a mask granular enough to pin-point to the
5721 * start pfn and tick off bits one-by-one until it becomes
5722 * too coarse to separate the current node from the last.
5724 mask = ~((1 << __ffs(start)) - 1);
5725 while (mask && last_end <= (start & (mask << 1)))
5728 /* accumulate all internode masks */
5732 /* convert mask to number of pages */
5733 return ~accl_mask + 1;
5736 /* Find the lowest pfn for a node */
5737 static unsigned long __init find_min_pfn_for_node(int nid)
5739 unsigned long min_pfn = ULONG_MAX;
5740 unsigned long start_pfn;
5743 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5744 min_pfn = min(min_pfn, start_pfn);
5746 if (min_pfn == ULONG_MAX) {
5747 pr_warn("Could not find start_pfn for node %d\n", nid);
5755 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5757 * It returns the minimum PFN based on information provided via
5758 * memblock_set_node().
5760 unsigned long __init find_min_pfn_with_active_regions(void)
5762 return find_min_pfn_for_node(MAX_NUMNODES);
5766 * early_calculate_totalpages()
5767 * Sum pages in active regions for movable zone.
5768 * Populate N_MEMORY for calculating usable_nodes.
5770 static unsigned long __init early_calculate_totalpages(void)
5772 unsigned long totalpages = 0;
5773 unsigned long start_pfn, end_pfn;
5776 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5777 unsigned long pages = end_pfn - start_pfn;
5779 totalpages += pages;
5781 node_set_state(nid, N_MEMORY);
5787 * Find the PFN the Movable zone begins in each node. Kernel memory
5788 * is spread evenly between nodes as long as the nodes have enough
5789 * memory. When they don't, some nodes will have more kernelcore than
5792 static void __init find_zone_movable_pfns_for_nodes(void)
5795 unsigned long usable_startpfn;
5796 unsigned long kernelcore_node, kernelcore_remaining;
5797 /* save the state before borrow the nodemask */
5798 nodemask_t saved_node_state = node_states[N_MEMORY];
5799 unsigned long totalpages = early_calculate_totalpages();
5800 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5801 struct memblock_region *r;
5803 /* Need to find movable_zone earlier when movable_node is specified. */
5804 find_usable_zone_for_movable();
5807 * If movable_node is specified, ignore kernelcore and movablecore
5810 if (movable_node_is_enabled()) {
5811 for_each_memblock(memory, r) {
5812 if (!memblock_is_hotpluggable(r))
5817 usable_startpfn = PFN_DOWN(r->base);
5818 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5819 min(usable_startpfn, zone_movable_pfn[nid]) :
5827 * If kernelcore=mirror is specified, ignore movablecore option
5829 if (mirrored_kernelcore) {
5830 bool mem_below_4gb_not_mirrored = false;
5832 for_each_memblock(memory, r) {
5833 if (memblock_is_mirror(r))
5838 usable_startpfn = memblock_region_memory_base_pfn(r);
5840 if (usable_startpfn < 0x100000) {
5841 mem_below_4gb_not_mirrored = true;
5845 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5846 min(usable_startpfn, zone_movable_pfn[nid]) :
5850 if (mem_below_4gb_not_mirrored)
5851 pr_warn("This configuration results in unmirrored kernel memory.");
5857 * If movablecore=nn[KMG] was specified, calculate what size of
5858 * kernelcore that corresponds so that memory usable for
5859 * any allocation type is evenly spread. If both kernelcore
5860 * and movablecore are specified, then the value of kernelcore
5861 * will be used for required_kernelcore if it's greater than
5862 * what movablecore would have allowed.
5864 if (required_movablecore) {
5865 unsigned long corepages;
5868 * Round-up so that ZONE_MOVABLE is at least as large as what
5869 * was requested by the user
5871 required_movablecore =
5872 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5873 required_movablecore = min(totalpages, required_movablecore);
5874 corepages = totalpages - required_movablecore;
5876 required_kernelcore = max(required_kernelcore, corepages);
5880 * If kernelcore was not specified or kernelcore size is larger
5881 * than totalpages, there is no ZONE_MOVABLE.
5883 if (!required_kernelcore || required_kernelcore >= totalpages)
5886 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5887 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5890 /* Spread kernelcore memory as evenly as possible throughout nodes */
5891 kernelcore_node = required_kernelcore / usable_nodes;
5892 for_each_node_state(nid, N_MEMORY) {
5893 unsigned long start_pfn, end_pfn;
5896 * Recalculate kernelcore_node if the division per node
5897 * now exceeds what is necessary to satisfy the requested
5898 * amount of memory for the kernel
5900 if (required_kernelcore < kernelcore_node)
5901 kernelcore_node = required_kernelcore / usable_nodes;
5904 * As the map is walked, we track how much memory is usable
5905 * by the kernel using kernelcore_remaining. When it is
5906 * 0, the rest of the node is usable by ZONE_MOVABLE
5908 kernelcore_remaining = kernelcore_node;
5910 /* Go through each range of PFNs within this node */
5911 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5912 unsigned long size_pages;
5914 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5915 if (start_pfn >= end_pfn)
5918 /* Account for what is only usable for kernelcore */
5919 if (start_pfn < usable_startpfn) {
5920 unsigned long kernel_pages;
5921 kernel_pages = min(end_pfn, usable_startpfn)
5924 kernelcore_remaining -= min(kernel_pages,
5925 kernelcore_remaining);
5926 required_kernelcore -= min(kernel_pages,
5927 required_kernelcore);
5929 /* Continue if range is now fully accounted */
5930 if (end_pfn <= usable_startpfn) {
5933 * Push zone_movable_pfn to the end so
5934 * that if we have to rebalance
5935 * kernelcore across nodes, we will
5936 * not double account here
5938 zone_movable_pfn[nid] = end_pfn;
5941 start_pfn = usable_startpfn;
5945 * The usable PFN range for ZONE_MOVABLE is from
5946 * start_pfn->end_pfn. Calculate size_pages as the
5947 * number of pages used as kernelcore
5949 size_pages = end_pfn - start_pfn;
5950 if (size_pages > kernelcore_remaining)
5951 size_pages = kernelcore_remaining;
5952 zone_movable_pfn[nid] = start_pfn + size_pages;
5955 * Some kernelcore has been met, update counts and
5956 * break if the kernelcore for this node has been
5959 required_kernelcore -= min(required_kernelcore,
5961 kernelcore_remaining -= size_pages;
5962 if (!kernelcore_remaining)
5968 * If there is still required_kernelcore, we do another pass with one
5969 * less node in the count. This will push zone_movable_pfn[nid] further
5970 * along on the nodes that still have memory until kernelcore is
5974 if (usable_nodes && required_kernelcore > usable_nodes)
5978 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5979 for (nid = 0; nid < MAX_NUMNODES; nid++)
5980 zone_movable_pfn[nid] =
5981 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5984 /* restore the node_state */
5985 node_states[N_MEMORY] = saved_node_state;
5988 /* Any regular or high memory on that node ? */
5989 static void check_for_memory(pg_data_t *pgdat, int nid)
5991 enum zone_type zone_type;
5993 if (N_MEMORY == N_NORMAL_MEMORY)
5996 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5997 struct zone *zone = &pgdat->node_zones[zone_type];
5998 if (populated_zone(zone)) {
5999 node_set_state(nid, N_HIGH_MEMORY);
6000 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6001 zone_type <= ZONE_NORMAL)
6002 node_set_state(nid, N_NORMAL_MEMORY);
6009 * free_area_init_nodes - Initialise all pg_data_t and zone data
6010 * @max_zone_pfn: an array of max PFNs for each zone
6012 * This will call free_area_init_node() for each active node in the system.
6013 * Using the page ranges provided by memblock_set_node(), the size of each
6014 * zone in each node and their holes is calculated. If the maximum PFN
6015 * between two adjacent zones match, it is assumed that the zone is empty.
6016 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6017 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6018 * starts where the previous one ended. For example, ZONE_DMA32 starts
6019 * at arch_max_dma_pfn.
6021 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6023 unsigned long start_pfn, end_pfn;
6026 /* Record where the zone boundaries are */
6027 memset(arch_zone_lowest_possible_pfn, 0,
6028 sizeof(arch_zone_lowest_possible_pfn));
6029 memset(arch_zone_highest_possible_pfn, 0,
6030 sizeof(arch_zone_highest_possible_pfn));
6031 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
6032 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
6033 for (i = 1; i < MAX_NR_ZONES; i++) {
6034 if (i == ZONE_MOVABLE)
6036 arch_zone_lowest_possible_pfn[i] =
6037 arch_zone_highest_possible_pfn[i-1];
6038 arch_zone_highest_possible_pfn[i] =
6039 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
6041 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
6042 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
6044 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6045 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6046 find_zone_movable_pfns_for_nodes();
6048 /* Print out the zone ranges */
6049 pr_info("Zone ranges:\n");
6050 for (i = 0; i < MAX_NR_ZONES; i++) {
6051 if (i == ZONE_MOVABLE)
6053 pr_info(" %-8s ", zone_names[i]);
6054 if (arch_zone_lowest_possible_pfn[i] ==
6055 arch_zone_highest_possible_pfn[i])
6058 pr_cont("[mem %#018Lx-%#018Lx]\n",
6059 (u64)arch_zone_lowest_possible_pfn[i]
6061 ((u64)arch_zone_highest_possible_pfn[i]
6062 << PAGE_SHIFT) - 1);
6065 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6066 pr_info("Movable zone start for each node\n");
6067 for (i = 0; i < MAX_NUMNODES; i++) {
6068 if (zone_movable_pfn[i])
6069 pr_info(" Node %d: %#018Lx\n", i,
6070 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6073 /* Print out the early node map */
6074 pr_info("Early memory node ranges\n");
6075 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6076 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6077 (u64)start_pfn << PAGE_SHIFT,
6078 ((u64)end_pfn << PAGE_SHIFT) - 1);
6080 /* Initialise every node */
6081 mminit_verify_pageflags_layout();
6082 setup_nr_node_ids();
6083 for_each_online_node(nid) {
6084 pg_data_t *pgdat = NODE_DATA(nid);
6085 free_area_init_node(nid, NULL,
6086 find_min_pfn_for_node(nid), NULL);
6088 /* Any memory on that node */
6089 if (pgdat->node_present_pages)
6090 node_set_state(nid, N_MEMORY);
6091 check_for_memory(pgdat, nid);
6095 static int __init cmdline_parse_core(char *p, unsigned long *core)
6097 unsigned long long coremem;
6101 coremem = memparse(p, &p);
6102 *core = coremem >> PAGE_SHIFT;
6104 /* Paranoid check that UL is enough for the coremem value */
6105 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6111 * kernelcore=size sets the amount of memory for use for allocations that
6112 * cannot be reclaimed or migrated.
6114 static int __init cmdline_parse_kernelcore(char *p)
6116 /* parse kernelcore=mirror */
6117 if (parse_option_str(p, "mirror")) {
6118 mirrored_kernelcore = true;
6122 return cmdline_parse_core(p, &required_kernelcore);
6126 * movablecore=size sets the amount of memory for use for allocations that
6127 * can be reclaimed or migrated.
6129 static int __init cmdline_parse_movablecore(char *p)
6131 return cmdline_parse_core(p, &required_movablecore);
6134 early_param("kernelcore", cmdline_parse_kernelcore);
6135 early_param("movablecore", cmdline_parse_movablecore);
6137 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6139 void adjust_managed_page_count(struct page *page, long count)
6141 spin_lock(&managed_page_count_lock);
6142 page_zone(page)->managed_pages += count;
6143 totalram_pages += count;
6144 #ifdef CONFIG_HIGHMEM
6145 if (PageHighMem(page))
6146 totalhigh_pages += count;
6148 spin_unlock(&managed_page_count_lock);
6150 EXPORT_SYMBOL(adjust_managed_page_count);
6152 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6155 unsigned long pages = 0;
6157 start = (void *)PAGE_ALIGN((unsigned long)start);
6158 end = (void *)((unsigned long)end & PAGE_MASK);
6159 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6160 if ((unsigned int)poison <= 0xFF)
6161 memset(pos, poison, PAGE_SIZE);
6162 free_reserved_page(virt_to_page(pos));
6166 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6167 s, pages << (PAGE_SHIFT - 10), start, end);
6171 EXPORT_SYMBOL(free_reserved_area);
6173 #ifdef CONFIG_HIGHMEM
6174 void free_highmem_page(struct page *page)
6176 __free_reserved_page(page);
6178 page_zone(page)->managed_pages++;
6184 void __init mem_init_print_info(const char *str)
6186 unsigned long physpages, codesize, datasize, rosize, bss_size;
6187 unsigned long init_code_size, init_data_size;
6189 physpages = get_num_physpages();
6190 codesize = _etext - _stext;
6191 datasize = _edata - _sdata;
6192 rosize = __end_rodata - __start_rodata;
6193 bss_size = __bss_stop - __bss_start;
6194 init_data_size = __init_end - __init_begin;
6195 init_code_size = _einittext - _sinittext;
6198 * Detect special cases and adjust section sizes accordingly:
6199 * 1) .init.* may be embedded into .data sections
6200 * 2) .init.text.* may be out of [__init_begin, __init_end],
6201 * please refer to arch/tile/kernel/vmlinux.lds.S.
6202 * 3) .rodata.* may be embedded into .text or .data sections.
6204 #define adj_init_size(start, end, size, pos, adj) \
6206 if (start <= pos && pos < end && size > adj) \
6210 adj_init_size(__init_begin, __init_end, init_data_size,
6211 _sinittext, init_code_size);
6212 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6213 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6214 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6215 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6217 #undef adj_init_size
6219 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6220 #ifdef CONFIG_HIGHMEM
6224 nr_free_pages() << (PAGE_SHIFT - 10),
6225 physpages << (PAGE_SHIFT - 10),
6226 codesize >> 10, datasize >> 10, rosize >> 10,
6227 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6228 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6229 totalcma_pages << (PAGE_SHIFT - 10),
6230 #ifdef CONFIG_HIGHMEM
6231 totalhigh_pages << (PAGE_SHIFT - 10),
6233 str ? ", " : "", str ? str : "");
6237 * set_dma_reserve - set the specified number of pages reserved in the first zone
6238 * @new_dma_reserve: The number of pages to mark reserved
6240 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6241 * In the DMA zone, a significant percentage may be consumed by kernel image
6242 * and other unfreeable allocations which can skew the watermarks badly. This
6243 * function may optionally be used to account for unfreeable pages in the
6244 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6245 * smaller per-cpu batchsize.
6247 void __init set_dma_reserve(unsigned long new_dma_reserve)
6249 dma_reserve = new_dma_reserve;
6252 void __init free_area_init(unsigned long *zones_size)
6254 free_area_init_node(0, zones_size,
6255 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6258 static int page_alloc_cpu_notify(struct notifier_block *self,
6259 unsigned long action, void *hcpu)
6261 int cpu = (unsigned long)hcpu;
6263 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6264 lru_add_drain_cpu(cpu);
6268 * Spill the event counters of the dead processor
6269 * into the current processors event counters.
6270 * This artificially elevates the count of the current
6273 vm_events_fold_cpu(cpu);
6276 * Zero the differential counters of the dead processor
6277 * so that the vm statistics are consistent.
6279 * This is only okay since the processor is dead and cannot
6280 * race with what we are doing.
6282 cpu_vm_stats_fold(cpu);
6287 void __init page_alloc_init(void)
6289 hotcpu_notifier(page_alloc_cpu_notify, 0);
6293 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6294 * or min_free_kbytes changes.
6296 static void calculate_totalreserve_pages(void)
6298 struct pglist_data *pgdat;
6299 unsigned long reserve_pages = 0;
6300 enum zone_type i, j;
6302 for_each_online_pgdat(pgdat) {
6303 for (i = 0; i < MAX_NR_ZONES; i++) {
6304 struct zone *zone = pgdat->node_zones + i;
6307 /* Find valid and maximum lowmem_reserve in the zone */
6308 for (j = i; j < MAX_NR_ZONES; j++) {
6309 if (zone->lowmem_reserve[j] > max)
6310 max = zone->lowmem_reserve[j];
6313 /* we treat the high watermark as reserved pages. */
6314 max += high_wmark_pages(zone);
6316 if (max > zone->managed_pages)
6317 max = zone->managed_pages;
6319 zone->totalreserve_pages = max;
6321 reserve_pages += max;
6324 totalreserve_pages = reserve_pages;
6328 * setup_per_zone_lowmem_reserve - called whenever
6329 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6330 * has a correct pages reserved value, so an adequate number of
6331 * pages are left in the zone after a successful __alloc_pages().
6333 static void setup_per_zone_lowmem_reserve(void)
6335 struct pglist_data *pgdat;
6336 enum zone_type j, idx;
6338 for_each_online_pgdat(pgdat) {
6339 for (j = 0; j < MAX_NR_ZONES; j++) {
6340 struct zone *zone = pgdat->node_zones + j;
6341 unsigned long managed_pages = zone->managed_pages;
6343 zone->lowmem_reserve[j] = 0;
6347 struct zone *lower_zone;
6351 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6352 sysctl_lowmem_reserve_ratio[idx] = 1;
6354 lower_zone = pgdat->node_zones + idx;
6355 lower_zone->lowmem_reserve[j] = managed_pages /
6356 sysctl_lowmem_reserve_ratio[idx];
6357 managed_pages += lower_zone->managed_pages;
6362 /* update totalreserve_pages */
6363 calculate_totalreserve_pages();
6366 static void __setup_per_zone_wmarks(void)
6368 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6369 unsigned long lowmem_pages = 0;
6371 unsigned long flags;
6373 /* Calculate total number of !ZONE_HIGHMEM pages */
6374 for_each_zone(zone) {
6375 if (!is_highmem(zone))
6376 lowmem_pages += zone->managed_pages;
6379 for_each_zone(zone) {
6382 spin_lock_irqsave(&zone->lock, flags);
6383 tmp = (u64)pages_min * zone->managed_pages;
6384 do_div(tmp, lowmem_pages);
6385 if (is_highmem(zone)) {
6387 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6388 * need highmem pages, so cap pages_min to a small
6391 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6392 * deltas control asynch page reclaim, and so should
6393 * not be capped for highmem.
6395 unsigned long min_pages;
6397 min_pages = zone->managed_pages / 1024;
6398 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6399 zone->watermark[WMARK_MIN] = min_pages;
6402 * If it's a lowmem zone, reserve a number of pages
6403 * proportionate to the zone's size.
6405 zone->watermark[WMARK_MIN] = tmp;
6409 * Set the kswapd watermarks distance according to the
6410 * scale factor in proportion to available memory, but
6411 * ensure a minimum size on small systems.
6413 tmp = max_t(u64, tmp >> 2,
6414 mult_frac(zone->managed_pages,
6415 watermark_scale_factor, 10000));
6417 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6418 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6420 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6421 high_wmark_pages(zone) - low_wmark_pages(zone) -
6422 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6424 spin_unlock_irqrestore(&zone->lock, flags);
6427 /* update totalreserve_pages */
6428 calculate_totalreserve_pages();
6432 * setup_per_zone_wmarks - called when min_free_kbytes changes
6433 * or when memory is hot-{added|removed}
6435 * Ensures that the watermark[min,low,high] values for each zone are set
6436 * correctly with respect to min_free_kbytes.
6438 void setup_per_zone_wmarks(void)
6440 mutex_lock(&zonelists_mutex);
6441 __setup_per_zone_wmarks();
6442 mutex_unlock(&zonelists_mutex);
6446 * The inactive anon list should be small enough that the VM never has to
6447 * do too much work, but large enough that each inactive page has a chance
6448 * to be referenced again before it is swapped out.
6450 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6451 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6452 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6453 * the anonymous pages are kept on the inactive list.
6456 * memory ratio inactive anon
6457 * -------------------------------------
6466 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6468 unsigned int gb, ratio;
6470 /* Zone size in gigabytes */
6471 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6473 ratio = int_sqrt(10 * gb);
6477 zone->inactive_ratio = ratio;
6480 static void __meminit setup_per_zone_inactive_ratio(void)
6485 calculate_zone_inactive_ratio(zone);
6489 * Initialise min_free_kbytes.
6491 * For small machines we want it small (128k min). For large machines
6492 * we want it large (64MB max). But it is not linear, because network
6493 * bandwidth does not increase linearly with machine size. We use
6495 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6496 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6512 int __meminit init_per_zone_wmark_min(void)
6514 unsigned long lowmem_kbytes;
6515 int new_min_free_kbytes;
6517 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6518 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6520 if (new_min_free_kbytes > user_min_free_kbytes) {
6521 min_free_kbytes = new_min_free_kbytes;
6522 if (min_free_kbytes < 128)
6523 min_free_kbytes = 128;
6524 if (min_free_kbytes > 65536)
6525 min_free_kbytes = 65536;
6527 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6528 new_min_free_kbytes, user_min_free_kbytes);
6530 setup_per_zone_wmarks();
6531 refresh_zone_stat_thresholds();
6532 setup_per_zone_lowmem_reserve();
6533 setup_per_zone_inactive_ratio();
6536 core_initcall(init_per_zone_wmark_min)
6539 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6540 * that we can call two helper functions whenever min_free_kbytes
6543 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6544 void __user *buffer, size_t *length, loff_t *ppos)
6548 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6553 user_min_free_kbytes = min_free_kbytes;
6554 setup_per_zone_wmarks();
6559 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6560 void __user *buffer, size_t *length, loff_t *ppos)
6564 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6569 setup_per_zone_wmarks();
6575 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6576 void __user *buffer, size_t *length, loff_t *ppos)
6581 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6586 zone->min_unmapped_pages = (zone->managed_pages *
6587 sysctl_min_unmapped_ratio) / 100;
6591 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6592 void __user *buffer, size_t *length, loff_t *ppos)
6597 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6602 zone->min_slab_pages = (zone->managed_pages *
6603 sysctl_min_slab_ratio) / 100;
6609 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6610 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6611 * whenever sysctl_lowmem_reserve_ratio changes.
6613 * The reserve ratio obviously has absolutely no relation with the
6614 * minimum watermarks. The lowmem reserve ratio can only make sense
6615 * if in function of the boot time zone sizes.
6617 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6618 void __user *buffer, size_t *length, loff_t *ppos)
6620 proc_dointvec_minmax(table, write, buffer, length, ppos);
6621 setup_per_zone_lowmem_reserve();
6626 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6627 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6628 * pagelist can have before it gets flushed back to buddy allocator.
6630 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6631 void __user *buffer, size_t *length, loff_t *ppos)
6634 int old_percpu_pagelist_fraction;
6637 mutex_lock(&pcp_batch_high_lock);
6638 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6640 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6641 if (!write || ret < 0)
6644 /* Sanity checking to avoid pcp imbalance */
6645 if (percpu_pagelist_fraction &&
6646 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6647 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6653 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6656 for_each_populated_zone(zone) {
6659 for_each_possible_cpu(cpu)
6660 pageset_set_high_and_batch(zone,
6661 per_cpu_ptr(zone->pageset, cpu));
6664 mutex_unlock(&pcp_batch_high_lock);
6669 int hashdist = HASHDIST_DEFAULT;
6671 static int __init set_hashdist(char *str)
6675 hashdist = simple_strtoul(str, &str, 0);
6678 __setup("hashdist=", set_hashdist);
6682 * allocate a large system hash table from bootmem
6683 * - it is assumed that the hash table must contain an exact power-of-2
6684 * quantity of entries
6685 * - limit is the number of hash buckets, not the total allocation size
6687 void *__init alloc_large_system_hash(const char *tablename,
6688 unsigned long bucketsize,
6689 unsigned long numentries,
6692 unsigned int *_hash_shift,
6693 unsigned int *_hash_mask,
6694 unsigned long low_limit,
6695 unsigned long high_limit)
6697 unsigned long long max = high_limit;
6698 unsigned long log2qty, size;
6701 /* allow the kernel cmdline to have a say */
6703 /* round applicable memory size up to nearest megabyte */
6704 numentries = nr_kernel_pages;
6706 /* It isn't necessary when PAGE_SIZE >= 1MB */
6707 if (PAGE_SHIFT < 20)
6708 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6710 /* limit to 1 bucket per 2^scale bytes of low memory */
6711 if (scale > PAGE_SHIFT)
6712 numentries >>= (scale - PAGE_SHIFT);
6714 numentries <<= (PAGE_SHIFT - scale);
6716 /* Make sure we've got at least a 0-order allocation.. */
6717 if (unlikely(flags & HASH_SMALL)) {
6718 /* Makes no sense without HASH_EARLY */
6719 WARN_ON(!(flags & HASH_EARLY));
6720 if (!(numentries >> *_hash_shift)) {
6721 numentries = 1UL << *_hash_shift;
6722 BUG_ON(!numentries);
6724 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6725 numentries = PAGE_SIZE / bucketsize;
6727 numentries = roundup_pow_of_two(numentries);
6729 /* limit allocation size to 1/16 total memory by default */
6731 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6732 do_div(max, bucketsize);
6734 max = min(max, 0x80000000ULL);
6736 if (numentries < low_limit)
6737 numentries = low_limit;
6738 if (numentries > max)
6741 log2qty = ilog2(numentries);
6744 size = bucketsize << log2qty;
6745 if (flags & HASH_EARLY)
6746 table = memblock_virt_alloc_nopanic(size, 0);
6748 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6751 * If bucketsize is not a power-of-two, we may free
6752 * some pages at the end of hash table which
6753 * alloc_pages_exact() automatically does
6755 if (get_order(size) < MAX_ORDER) {
6756 table = alloc_pages_exact(size, GFP_ATOMIC);
6757 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6760 } while (!table && size > PAGE_SIZE && --log2qty);
6763 panic("Failed to allocate %s hash table\n", tablename);
6765 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
6766 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
6769 *_hash_shift = log2qty;
6771 *_hash_mask = (1 << log2qty) - 1;
6776 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6777 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6780 #ifdef CONFIG_SPARSEMEM
6781 return __pfn_to_section(pfn)->pageblock_flags;
6783 return zone->pageblock_flags;
6784 #endif /* CONFIG_SPARSEMEM */
6787 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6789 #ifdef CONFIG_SPARSEMEM
6790 pfn &= (PAGES_PER_SECTION-1);
6791 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6793 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6794 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6795 #endif /* CONFIG_SPARSEMEM */
6799 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6800 * @page: The page within the block of interest
6801 * @pfn: The target page frame number
6802 * @end_bitidx: The last bit of interest to retrieve
6803 * @mask: mask of bits that the caller is interested in
6805 * Return: pageblock_bits flags
6807 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6808 unsigned long end_bitidx,
6812 unsigned long *bitmap;
6813 unsigned long bitidx, word_bitidx;
6816 zone = page_zone(page);
6817 bitmap = get_pageblock_bitmap(zone, pfn);
6818 bitidx = pfn_to_bitidx(zone, pfn);
6819 word_bitidx = bitidx / BITS_PER_LONG;
6820 bitidx &= (BITS_PER_LONG-1);
6822 word = bitmap[word_bitidx];
6823 bitidx += end_bitidx;
6824 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6828 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6829 * @page: The page within the block of interest
6830 * @flags: The flags to set
6831 * @pfn: The target page frame number
6832 * @end_bitidx: The last bit of interest
6833 * @mask: mask of bits that the caller is interested in
6835 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6837 unsigned long end_bitidx,
6841 unsigned long *bitmap;
6842 unsigned long bitidx, word_bitidx;
6843 unsigned long old_word, word;
6845 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6847 zone = page_zone(page);
6848 bitmap = get_pageblock_bitmap(zone, pfn);
6849 bitidx = pfn_to_bitidx(zone, pfn);
6850 word_bitidx = bitidx / BITS_PER_LONG;
6851 bitidx &= (BITS_PER_LONG-1);
6853 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6855 bitidx += end_bitidx;
6856 mask <<= (BITS_PER_LONG - bitidx - 1);
6857 flags <<= (BITS_PER_LONG - bitidx - 1);
6859 word = READ_ONCE(bitmap[word_bitidx]);
6861 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6862 if (word == old_word)
6869 * This function checks whether pageblock includes unmovable pages or not.
6870 * If @count is not zero, it is okay to include less @count unmovable pages
6872 * PageLRU check without isolation or lru_lock could race so that
6873 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6874 * expect this function should be exact.
6876 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6877 bool skip_hwpoisoned_pages)
6879 unsigned long pfn, iter, found;
6883 * For avoiding noise data, lru_add_drain_all() should be called
6884 * If ZONE_MOVABLE, the zone never contains unmovable pages
6886 if (zone_idx(zone) == ZONE_MOVABLE)
6888 mt = get_pageblock_migratetype(page);
6889 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6892 pfn = page_to_pfn(page);
6893 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6894 unsigned long check = pfn + iter;
6896 if (!pfn_valid_within(check))
6899 page = pfn_to_page(check);
6902 * Hugepages are not in LRU lists, but they're movable.
6903 * We need not scan over tail pages bacause we don't
6904 * handle each tail page individually in migration.
6906 if (PageHuge(page)) {
6907 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6912 * We can't use page_count without pin a page
6913 * because another CPU can free compound page.
6914 * This check already skips compound tails of THP
6915 * because their page->_refcount is zero at all time.
6917 if (!page_ref_count(page)) {
6918 if (PageBuddy(page))
6919 iter += (1 << page_order(page)) - 1;
6924 * The HWPoisoned page may be not in buddy system, and
6925 * page_count() is not 0.
6927 if (skip_hwpoisoned_pages && PageHWPoison(page))
6933 * If there are RECLAIMABLE pages, we need to check
6934 * it. But now, memory offline itself doesn't call
6935 * shrink_node_slabs() and it still to be fixed.
6938 * If the page is not RAM, page_count()should be 0.
6939 * we don't need more check. This is an _used_ not-movable page.
6941 * The problematic thing here is PG_reserved pages. PG_reserved
6942 * is set to both of a memory hole page and a _used_ kernel
6951 bool is_pageblock_removable_nolock(struct page *page)
6957 * We have to be careful here because we are iterating over memory
6958 * sections which are not zone aware so we might end up outside of
6959 * the zone but still within the section.
6960 * We have to take care about the node as well. If the node is offline
6961 * its NODE_DATA will be NULL - see page_zone.
6963 if (!node_online(page_to_nid(page)))
6966 zone = page_zone(page);
6967 pfn = page_to_pfn(page);
6968 if (!zone_spans_pfn(zone, pfn))
6971 return !has_unmovable_pages(zone, page, 0, true);
6974 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
6976 static unsigned long pfn_max_align_down(unsigned long pfn)
6978 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6979 pageblock_nr_pages) - 1);
6982 static unsigned long pfn_max_align_up(unsigned long pfn)
6984 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6985 pageblock_nr_pages));
6988 /* [start, end) must belong to a single zone. */
6989 static int __alloc_contig_migrate_range(struct compact_control *cc,
6990 unsigned long start, unsigned long end)
6992 /* This function is based on compact_zone() from compaction.c. */
6993 unsigned long nr_reclaimed;
6994 unsigned long pfn = start;
6995 unsigned int tries = 0;
7000 while (pfn < end || !list_empty(&cc->migratepages)) {
7001 if (fatal_signal_pending(current)) {
7006 if (list_empty(&cc->migratepages)) {
7007 cc->nr_migratepages = 0;
7008 pfn = isolate_migratepages_range(cc, pfn, end);
7014 } else if (++tries == 5) {
7015 ret = ret < 0 ? ret : -EBUSY;
7019 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7021 cc->nr_migratepages -= nr_reclaimed;
7023 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7024 NULL, 0, cc->mode, MR_CMA);
7027 putback_movable_pages(&cc->migratepages);
7034 * alloc_contig_range() -- tries to allocate given range of pages
7035 * @start: start PFN to allocate
7036 * @end: one-past-the-last PFN to allocate
7037 * @migratetype: migratetype of the underlaying pageblocks (either
7038 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7039 * in range must have the same migratetype and it must
7040 * be either of the two.
7042 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7043 * aligned, however it's the caller's responsibility to guarantee that
7044 * we are the only thread that changes migrate type of pageblocks the
7047 * The PFN range must belong to a single zone.
7049 * Returns zero on success or negative error code. On success all
7050 * pages which PFN is in [start, end) are allocated for the caller and
7051 * need to be freed with free_contig_range().
7053 int alloc_contig_range(unsigned long start, unsigned long end,
7054 unsigned migratetype)
7056 unsigned long outer_start, outer_end;
7060 struct compact_control cc = {
7061 .nr_migratepages = 0,
7063 .zone = page_zone(pfn_to_page(start)),
7064 .mode = MIGRATE_SYNC,
7065 .ignore_skip_hint = true,
7067 INIT_LIST_HEAD(&cc.migratepages);
7070 * What we do here is we mark all pageblocks in range as
7071 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7072 * have different sizes, and due to the way page allocator
7073 * work, we align the range to biggest of the two pages so
7074 * that page allocator won't try to merge buddies from
7075 * different pageblocks and change MIGRATE_ISOLATE to some
7076 * other migration type.
7078 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7079 * migrate the pages from an unaligned range (ie. pages that
7080 * we are interested in). This will put all the pages in
7081 * range back to page allocator as MIGRATE_ISOLATE.
7083 * When this is done, we take the pages in range from page
7084 * allocator removing them from the buddy system. This way
7085 * page allocator will never consider using them.
7087 * This lets us mark the pageblocks back as
7088 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7089 * aligned range but not in the unaligned, original range are
7090 * put back to page allocator so that buddy can use them.
7093 ret = start_isolate_page_range(pfn_max_align_down(start),
7094 pfn_max_align_up(end), migratetype,
7100 * In case of -EBUSY, we'd like to know which page causes problem.
7101 * So, just fall through. We will check it in test_pages_isolated().
7103 ret = __alloc_contig_migrate_range(&cc, start, end);
7104 if (ret && ret != -EBUSY)
7108 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7109 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7110 * more, all pages in [start, end) are free in page allocator.
7111 * What we are going to do is to allocate all pages from
7112 * [start, end) (that is remove them from page allocator).
7114 * The only problem is that pages at the beginning and at the
7115 * end of interesting range may be not aligned with pages that
7116 * page allocator holds, ie. they can be part of higher order
7117 * pages. Because of this, we reserve the bigger range and
7118 * once this is done free the pages we are not interested in.
7120 * We don't have to hold zone->lock here because the pages are
7121 * isolated thus they won't get removed from buddy.
7124 lru_add_drain_all();
7125 drain_all_pages(cc.zone);
7128 outer_start = start;
7129 while (!PageBuddy(pfn_to_page(outer_start))) {
7130 if (++order >= MAX_ORDER) {
7131 outer_start = start;
7134 outer_start &= ~0UL << order;
7137 if (outer_start != start) {
7138 order = page_order(pfn_to_page(outer_start));
7141 * outer_start page could be small order buddy page and
7142 * it doesn't include start page. Adjust outer_start
7143 * in this case to report failed page properly
7144 * on tracepoint in test_pages_isolated()
7146 if (outer_start + (1UL << order) <= start)
7147 outer_start = start;
7150 /* Make sure the range is really isolated. */
7151 if (test_pages_isolated(outer_start, end, false)) {
7152 pr_info("%s: [%lx, %lx) PFNs busy\n",
7153 __func__, outer_start, end);
7158 /* Grab isolated pages from freelists. */
7159 outer_end = isolate_freepages_range(&cc, outer_start, end);
7165 /* Free head and tail (if any) */
7166 if (start != outer_start)
7167 free_contig_range(outer_start, start - outer_start);
7168 if (end != outer_end)
7169 free_contig_range(end, outer_end - end);
7172 undo_isolate_page_range(pfn_max_align_down(start),
7173 pfn_max_align_up(end), migratetype);
7177 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7179 unsigned int count = 0;
7181 for (; nr_pages--; pfn++) {
7182 struct page *page = pfn_to_page(pfn);
7184 count += page_count(page) != 1;
7187 WARN(count != 0, "%d pages are still in use!\n", count);
7191 #ifdef CONFIG_MEMORY_HOTPLUG
7193 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7194 * page high values need to be recalulated.
7196 void __meminit zone_pcp_update(struct zone *zone)
7199 mutex_lock(&pcp_batch_high_lock);
7200 for_each_possible_cpu(cpu)
7201 pageset_set_high_and_batch(zone,
7202 per_cpu_ptr(zone->pageset, cpu));
7203 mutex_unlock(&pcp_batch_high_lock);
7207 void zone_pcp_reset(struct zone *zone)
7209 unsigned long flags;
7211 struct per_cpu_pageset *pset;
7213 /* avoid races with drain_pages() */
7214 local_irq_save(flags);
7215 if (zone->pageset != &boot_pageset) {
7216 for_each_online_cpu(cpu) {
7217 pset = per_cpu_ptr(zone->pageset, cpu);
7218 drain_zonestat(zone, pset);
7220 free_percpu(zone->pageset);
7221 zone->pageset = &boot_pageset;
7223 local_irq_restore(flags);
7226 #ifdef CONFIG_MEMORY_HOTREMOVE
7228 * All pages in the range must be in a single zone and isolated
7229 * before calling this.
7232 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7236 unsigned int order, i;
7238 unsigned long flags;
7239 /* find the first valid pfn */
7240 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7245 zone = page_zone(pfn_to_page(pfn));
7246 spin_lock_irqsave(&zone->lock, flags);
7248 while (pfn < end_pfn) {
7249 if (!pfn_valid(pfn)) {
7253 page = pfn_to_page(pfn);
7255 * The HWPoisoned page may be not in buddy system, and
7256 * page_count() is not 0.
7258 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7260 SetPageReserved(page);
7264 BUG_ON(page_count(page));
7265 BUG_ON(!PageBuddy(page));
7266 order = page_order(page);
7267 #ifdef CONFIG_DEBUG_VM
7268 pr_info("remove from free list %lx %d %lx\n",
7269 pfn, 1 << order, end_pfn);
7271 list_del(&page->lru);
7272 rmv_page_order(page);
7273 zone->free_area[order].nr_free--;
7274 for (i = 0; i < (1 << order); i++)
7275 SetPageReserved((page+i));
7276 pfn += (1 << order);
7278 spin_unlock_irqrestore(&zone->lock, flags);
7282 bool is_free_buddy_page(struct page *page)
7284 struct zone *zone = page_zone(page);
7285 unsigned long pfn = page_to_pfn(page);
7286 unsigned long flags;
7289 spin_lock_irqsave(&zone->lock, flags);
7290 for (order = 0; order < MAX_ORDER; order++) {
7291 struct page *page_head = page - (pfn & ((1 << order) - 1));
7293 if (PageBuddy(page_head) && page_order(page_head) >= order)
7296 spin_unlock_irqrestore(&zone->lock, flags);
7298 return order < MAX_ORDER;