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)
1027 bool compound = PageCompound(page);
1030 VM_BUG_ON_PAGE(PageTail(page), page);
1031 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1033 trace_mm_page_free(page, order);
1034 kmemcheck_free_shadow(page, order);
1035 kasan_free_pages(page, order);
1038 page->mapping = NULL;
1039 bad += free_pages_check(page);
1040 for (i = 1; i < (1 << order); i++) {
1042 bad += free_tail_pages_check(page, page + i);
1043 bad += free_pages_check(page + i);
1048 reset_page_owner(page, order);
1050 if (!PageHighMem(page)) {
1051 debug_check_no_locks_freed(page_address(page),
1052 PAGE_SIZE << order);
1053 debug_check_no_obj_freed(page_address(page),
1054 PAGE_SIZE << order);
1056 arch_free_page(page, order);
1057 kernel_poison_pages(page, 1 << order, 0);
1058 kernel_map_pages(page, 1 << order, 0);
1063 static void __free_pages_ok(struct page *page, unsigned int order)
1065 unsigned long flags;
1067 unsigned long pfn = page_to_pfn(page);
1069 if (!free_pages_prepare(page, order))
1072 migratetype = get_pfnblock_migratetype(page, pfn);
1073 local_irq_save(flags);
1074 __count_vm_events(PGFREE, 1 << order);
1075 free_one_page(page_zone(page), page, pfn, order, migratetype);
1076 local_irq_restore(flags);
1079 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1081 unsigned int nr_pages = 1 << order;
1082 struct page *p = page;
1086 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1088 __ClearPageReserved(p);
1089 set_page_count(p, 0);
1091 __ClearPageReserved(p);
1092 set_page_count(p, 0);
1094 page_zone(page)->managed_pages += nr_pages;
1095 set_page_refcounted(page);
1096 __free_pages(page, order);
1099 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1100 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1102 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1104 int __meminit early_pfn_to_nid(unsigned long pfn)
1106 static DEFINE_SPINLOCK(early_pfn_lock);
1109 spin_lock(&early_pfn_lock);
1110 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1113 spin_unlock(&early_pfn_lock);
1119 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1120 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1121 struct mminit_pfnnid_cache *state)
1125 nid = __early_pfn_to_nid(pfn, state);
1126 if (nid >= 0 && nid != node)
1131 /* Only safe to use early in boot when initialisation is single-threaded */
1132 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1134 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1139 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1143 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1144 struct mminit_pfnnid_cache *state)
1151 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1154 if (early_page_uninitialised(pfn))
1156 return __free_pages_boot_core(page, order);
1160 * Check that the whole (or subset of) a pageblock given by the interval of
1161 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1162 * with the migration of free compaction scanner. The scanners then need to
1163 * use only pfn_valid_within() check for arches that allow holes within
1166 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1168 * It's possible on some configurations to have a setup like node0 node1 node0
1169 * i.e. it's possible that all pages within a zones range of pages do not
1170 * belong to a single zone. We assume that a border between node0 and node1
1171 * can occur within a single pageblock, but not a node0 node1 node0
1172 * interleaving within a single pageblock. It is therefore sufficient to check
1173 * the first and last page of a pageblock and avoid checking each individual
1174 * page in a pageblock.
1176 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1177 unsigned long end_pfn, struct zone *zone)
1179 struct page *start_page;
1180 struct page *end_page;
1182 /* end_pfn is one past the range we are checking */
1185 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1188 start_page = pfn_to_page(start_pfn);
1190 if (page_zone(start_page) != zone)
1193 end_page = pfn_to_page(end_pfn);
1195 /* This gives a shorter code than deriving page_zone(end_page) */
1196 if (page_zone_id(start_page) != page_zone_id(end_page))
1202 void set_zone_contiguous(struct zone *zone)
1204 unsigned long block_start_pfn = zone->zone_start_pfn;
1205 unsigned long block_end_pfn;
1207 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1208 for (; block_start_pfn < zone_end_pfn(zone);
1209 block_start_pfn = block_end_pfn,
1210 block_end_pfn += pageblock_nr_pages) {
1212 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1214 if (!__pageblock_pfn_to_page(block_start_pfn,
1215 block_end_pfn, zone))
1219 /* We confirm that there is no hole */
1220 zone->contiguous = true;
1223 void clear_zone_contiguous(struct zone *zone)
1225 zone->contiguous = false;
1228 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1229 static void __init deferred_free_range(struct page *page,
1230 unsigned long pfn, int nr_pages)
1237 /* Free a large naturally-aligned chunk if possible */
1238 if (nr_pages == MAX_ORDER_NR_PAGES &&
1239 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1240 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1241 __free_pages_boot_core(page, MAX_ORDER-1);
1245 for (i = 0; i < nr_pages; i++, page++)
1246 __free_pages_boot_core(page, 0);
1249 /* Completion tracking for deferred_init_memmap() threads */
1250 static atomic_t pgdat_init_n_undone __initdata;
1251 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1253 static inline void __init pgdat_init_report_one_done(void)
1255 if (atomic_dec_and_test(&pgdat_init_n_undone))
1256 complete(&pgdat_init_all_done_comp);
1259 /* Initialise remaining memory on a node */
1260 static int __init deferred_init_memmap(void *data)
1262 pg_data_t *pgdat = data;
1263 int nid = pgdat->node_id;
1264 struct mminit_pfnnid_cache nid_init_state = { };
1265 unsigned long start = jiffies;
1266 unsigned long nr_pages = 0;
1267 unsigned long walk_start, walk_end;
1270 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1271 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1273 if (first_init_pfn == ULONG_MAX) {
1274 pgdat_init_report_one_done();
1278 /* Bind memory initialisation thread to a local node if possible */
1279 if (!cpumask_empty(cpumask))
1280 set_cpus_allowed_ptr(current, cpumask);
1282 /* Sanity check boundaries */
1283 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1284 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1285 pgdat->first_deferred_pfn = ULONG_MAX;
1287 /* Only the highest zone is deferred so find it */
1288 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1289 zone = pgdat->node_zones + zid;
1290 if (first_init_pfn < zone_end_pfn(zone))
1294 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1295 unsigned long pfn, end_pfn;
1296 struct page *page = NULL;
1297 struct page *free_base_page = NULL;
1298 unsigned long free_base_pfn = 0;
1301 end_pfn = min(walk_end, zone_end_pfn(zone));
1302 pfn = first_init_pfn;
1303 if (pfn < walk_start)
1305 if (pfn < zone->zone_start_pfn)
1306 pfn = zone->zone_start_pfn;
1308 for (; pfn < end_pfn; pfn++) {
1309 if (!pfn_valid_within(pfn))
1313 * Ensure pfn_valid is checked every
1314 * MAX_ORDER_NR_PAGES for memory holes
1316 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1317 if (!pfn_valid(pfn)) {
1323 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1328 /* Minimise pfn page lookups and scheduler checks */
1329 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1332 nr_pages += nr_to_free;
1333 deferred_free_range(free_base_page,
1334 free_base_pfn, nr_to_free);
1335 free_base_page = NULL;
1336 free_base_pfn = nr_to_free = 0;
1338 page = pfn_to_page(pfn);
1343 VM_BUG_ON(page_zone(page) != zone);
1347 __init_single_page(page, pfn, zid, nid);
1348 if (!free_base_page) {
1349 free_base_page = page;
1350 free_base_pfn = pfn;
1355 /* Where possible, batch up pages for a single free */
1358 /* Free the current block of pages to allocator */
1359 nr_pages += nr_to_free;
1360 deferred_free_range(free_base_page, free_base_pfn,
1362 free_base_page = NULL;
1363 free_base_pfn = nr_to_free = 0;
1366 first_init_pfn = max(end_pfn, first_init_pfn);
1369 /* Sanity check that the next zone really is unpopulated */
1370 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1372 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1373 jiffies_to_msecs(jiffies - start));
1375 pgdat_init_report_one_done();
1378 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1380 void __init page_alloc_init_late(void)
1384 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1387 /* There will be num_node_state(N_MEMORY) threads */
1388 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1389 for_each_node_state(nid, N_MEMORY) {
1390 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1393 /* Block until all are initialised */
1394 wait_for_completion(&pgdat_init_all_done_comp);
1396 /* Reinit limits that are based on free pages after the kernel is up */
1397 files_maxfiles_init();
1400 for_each_populated_zone(zone)
1401 set_zone_contiguous(zone);
1405 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1406 void __init init_cma_reserved_pageblock(struct page *page)
1408 unsigned i = pageblock_nr_pages;
1409 struct page *p = page;
1412 __ClearPageReserved(p);
1413 set_page_count(p, 0);
1416 set_pageblock_migratetype(page, MIGRATE_CMA);
1418 if (pageblock_order >= MAX_ORDER) {
1419 i = pageblock_nr_pages;
1422 set_page_refcounted(p);
1423 __free_pages(p, MAX_ORDER - 1);
1424 p += MAX_ORDER_NR_PAGES;
1425 } while (i -= MAX_ORDER_NR_PAGES);
1427 set_page_refcounted(page);
1428 __free_pages(page, pageblock_order);
1431 adjust_managed_page_count(page, pageblock_nr_pages);
1436 * The order of subdivision here is critical for the IO subsystem.
1437 * Please do not alter this order without good reasons and regression
1438 * testing. Specifically, as large blocks of memory are subdivided,
1439 * the order in which smaller blocks are delivered depends on the order
1440 * they're subdivided in this function. This is the primary factor
1441 * influencing the order in which pages are delivered to the IO
1442 * subsystem according to empirical testing, and this is also justified
1443 * by considering the behavior of a buddy system containing a single
1444 * large block of memory acted on by a series of small allocations.
1445 * This behavior is a critical factor in sglist merging's success.
1449 static inline void expand(struct zone *zone, struct page *page,
1450 int low, int high, struct free_area *area,
1453 unsigned long size = 1 << high;
1455 while (high > low) {
1459 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1461 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1462 debug_guardpage_enabled() &&
1463 high < debug_guardpage_minorder()) {
1465 * Mark as guard pages (or page), that will allow to
1466 * merge back to allocator when buddy will be freed.
1467 * Corresponding page table entries will not be touched,
1468 * pages will stay not present in virtual address space
1470 set_page_guard(zone, &page[size], high, migratetype);
1473 list_add(&page[size].lru, &area->free_list[migratetype]);
1475 set_page_order(&page[size], high);
1480 * This page is about to be returned from the page allocator
1482 static inline int check_new_page(struct page *page)
1484 const char *bad_reason = NULL;
1485 unsigned long bad_flags = 0;
1487 if (unlikely(atomic_read(&page->_mapcount) != -1))
1488 bad_reason = "nonzero mapcount";
1489 if (unlikely(page->mapping != NULL))
1490 bad_reason = "non-NULL mapping";
1491 if (unlikely(page_ref_count(page) != 0))
1492 bad_reason = "nonzero _count";
1493 if (unlikely(page->flags & __PG_HWPOISON)) {
1494 bad_reason = "HWPoisoned (hardware-corrupted)";
1495 bad_flags = __PG_HWPOISON;
1497 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1498 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1499 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1502 if (unlikely(page->mem_cgroup))
1503 bad_reason = "page still charged to cgroup";
1505 if (unlikely(bad_reason)) {
1506 bad_page(page, bad_reason, bad_flags);
1512 static inline bool free_pages_prezeroed(bool poisoned)
1514 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1515 page_poisoning_enabled() && poisoned;
1518 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1522 bool poisoned = true;
1524 for (i = 0; i < (1 << order); i++) {
1525 struct page *p = page + i;
1526 if (unlikely(check_new_page(p)))
1529 poisoned &= page_is_poisoned(p);
1532 set_page_private(page, 0);
1533 set_page_refcounted(page);
1535 arch_alloc_page(page, order);
1536 kernel_map_pages(page, 1 << order, 1);
1537 kernel_poison_pages(page, 1 << order, 1);
1538 kasan_alloc_pages(page, order);
1540 if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO))
1541 for (i = 0; i < (1 << order); i++)
1542 clear_highpage(page + i);
1544 if (order && (gfp_flags & __GFP_COMP))
1545 prep_compound_page(page, order);
1547 set_page_owner(page, order, gfp_flags);
1550 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1551 * allocate the page. The expectation is that the caller is taking
1552 * steps that will free more memory. The caller should avoid the page
1553 * being used for !PFMEMALLOC purposes.
1555 if (alloc_flags & ALLOC_NO_WATERMARKS)
1556 set_page_pfmemalloc(page);
1558 clear_page_pfmemalloc(page);
1564 * Go through the free lists for the given migratetype and remove
1565 * the smallest available page from the freelists
1568 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1571 unsigned int current_order;
1572 struct free_area *area;
1575 /* Find a page of the appropriate size in the preferred list */
1576 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1577 area = &(zone->free_area[current_order]);
1578 page = list_first_entry_or_null(&area->free_list[migratetype],
1582 list_del(&page->lru);
1583 rmv_page_order(page);
1585 expand(zone, page, order, current_order, area, migratetype);
1586 set_pcppage_migratetype(page, migratetype);
1595 * This array describes the order lists are fallen back to when
1596 * the free lists for the desirable migrate type are depleted
1598 static int fallbacks[MIGRATE_TYPES][4] = {
1599 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1600 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1601 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1603 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1605 #ifdef CONFIG_MEMORY_ISOLATION
1606 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1611 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1614 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1617 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1618 unsigned int order) { return NULL; }
1622 * Move the free pages in a range to the free lists of the requested type.
1623 * Note that start_page and end_pages are not aligned on a pageblock
1624 * boundary. If alignment is required, use move_freepages_block()
1626 int move_freepages(struct zone *zone,
1627 struct page *start_page, struct page *end_page,
1632 int pages_moved = 0;
1634 #ifndef CONFIG_HOLES_IN_ZONE
1636 * page_zone is not safe to call in this context when
1637 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1638 * anyway as we check zone boundaries in move_freepages_block().
1639 * Remove at a later date when no bug reports exist related to
1640 * grouping pages by mobility
1642 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1645 for (page = start_page; page <= end_page;) {
1646 /* Make sure we are not inadvertently changing nodes */
1647 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1649 if (!pfn_valid_within(page_to_pfn(page))) {
1654 if (!PageBuddy(page)) {
1659 order = page_order(page);
1660 list_move(&page->lru,
1661 &zone->free_area[order].free_list[migratetype]);
1663 pages_moved += 1 << order;
1669 int move_freepages_block(struct zone *zone, struct page *page,
1672 unsigned long start_pfn, end_pfn;
1673 struct page *start_page, *end_page;
1675 start_pfn = page_to_pfn(page);
1676 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1677 start_page = pfn_to_page(start_pfn);
1678 end_page = start_page + pageblock_nr_pages - 1;
1679 end_pfn = start_pfn + pageblock_nr_pages - 1;
1681 /* Do not cross zone boundaries */
1682 if (!zone_spans_pfn(zone, start_pfn))
1684 if (!zone_spans_pfn(zone, end_pfn))
1687 return move_freepages(zone, start_page, end_page, migratetype);
1690 static void change_pageblock_range(struct page *pageblock_page,
1691 int start_order, int migratetype)
1693 int nr_pageblocks = 1 << (start_order - pageblock_order);
1695 while (nr_pageblocks--) {
1696 set_pageblock_migratetype(pageblock_page, migratetype);
1697 pageblock_page += pageblock_nr_pages;
1702 * When we are falling back to another migratetype during allocation, try to
1703 * steal extra free pages from the same pageblocks to satisfy further
1704 * allocations, instead of polluting multiple pageblocks.
1706 * If we are stealing a relatively large buddy page, it is likely there will
1707 * be more free pages in the pageblock, so try to steal them all. For
1708 * reclaimable and unmovable allocations, we steal regardless of page size,
1709 * as fragmentation caused by those allocations polluting movable pageblocks
1710 * is worse than movable allocations stealing from unmovable and reclaimable
1713 static bool can_steal_fallback(unsigned int order, int start_mt)
1716 * Leaving this order check is intended, although there is
1717 * relaxed order check in next check. The reason is that
1718 * we can actually steal whole pageblock if this condition met,
1719 * but, below check doesn't guarantee it and that is just heuristic
1720 * so could be changed anytime.
1722 if (order >= pageblock_order)
1725 if (order >= pageblock_order / 2 ||
1726 start_mt == MIGRATE_RECLAIMABLE ||
1727 start_mt == MIGRATE_UNMOVABLE ||
1728 page_group_by_mobility_disabled)
1735 * This function implements actual steal behaviour. If order is large enough,
1736 * we can steal whole pageblock. If not, we first move freepages in this
1737 * pageblock and check whether half of pages are moved or not. If half of
1738 * pages are moved, we can change migratetype of pageblock and permanently
1739 * use it's pages as requested migratetype in the future.
1741 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1744 unsigned int current_order = page_order(page);
1747 /* Take ownership for orders >= pageblock_order */
1748 if (current_order >= pageblock_order) {
1749 change_pageblock_range(page, current_order, start_type);
1753 pages = move_freepages_block(zone, page, start_type);
1755 /* Claim the whole block if over half of it is free */
1756 if (pages >= (1 << (pageblock_order-1)) ||
1757 page_group_by_mobility_disabled)
1758 set_pageblock_migratetype(page, start_type);
1762 * Check whether there is a suitable fallback freepage with requested order.
1763 * If only_stealable is true, this function returns fallback_mt only if
1764 * we can steal other freepages all together. This would help to reduce
1765 * fragmentation due to mixed migratetype pages in one pageblock.
1767 int find_suitable_fallback(struct free_area *area, unsigned int order,
1768 int migratetype, bool only_stealable, bool *can_steal)
1773 if (area->nr_free == 0)
1778 fallback_mt = fallbacks[migratetype][i];
1779 if (fallback_mt == MIGRATE_TYPES)
1782 if (list_empty(&area->free_list[fallback_mt]))
1785 if (can_steal_fallback(order, migratetype))
1788 if (!only_stealable)
1799 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1800 * there are no empty page blocks that contain a page with a suitable order
1802 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1803 unsigned int alloc_order)
1806 unsigned long max_managed, flags;
1809 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1810 * Check is race-prone but harmless.
1812 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1813 if (zone->nr_reserved_highatomic >= max_managed)
1816 spin_lock_irqsave(&zone->lock, flags);
1818 /* Recheck the nr_reserved_highatomic limit under the lock */
1819 if (zone->nr_reserved_highatomic >= max_managed)
1823 mt = get_pageblock_migratetype(page);
1824 if (mt != MIGRATE_HIGHATOMIC &&
1825 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1826 zone->nr_reserved_highatomic += pageblock_nr_pages;
1827 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1828 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1832 spin_unlock_irqrestore(&zone->lock, flags);
1836 * Used when an allocation is about to fail under memory pressure. This
1837 * potentially hurts the reliability of high-order allocations when under
1838 * intense memory pressure but failed atomic allocations should be easier
1839 * to recover from than an OOM.
1841 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1843 struct zonelist *zonelist = ac->zonelist;
1844 unsigned long flags;
1850 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1852 /* Preserve at least one pageblock */
1853 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1856 spin_lock_irqsave(&zone->lock, flags);
1857 for (order = 0; order < MAX_ORDER; order++) {
1858 struct free_area *area = &(zone->free_area[order]);
1860 page = list_first_entry_or_null(
1861 &area->free_list[MIGRATE_HIGHATOMIC],
1867 * It should never happen but changes to locking could
1868 * inadvertently allow a per-cpu drain to add pages
1869 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1870 * and watch for underflows.
1872 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1873 zone->nr_reserved_highatomic);
1876 * Convert to ac->migratetype and avoid the normal
1877 * pageblock stealing heuristics. Minimally, the caller
1878 * is doing the work and needs the pages. More
1879 * importantly, if the block was always converted to
1880 * MIGRATE_UNMOVABLE or another type then the number
1881 * of pageblocks that cannot be completely freed
1884 set_pageblock_migratetype(page, ac->migratetype);
1885 move_freepages_block(zone, page, ac->migratetype);
1886 spin_unlock_irqrestore(&zone->lock, flags);
1889 spin_unlock_irqrestore(&zone->lock, flags);
1893 /* Remove an element from the buddy allocator from the fallback list */
1894 static inline struct page *
1895 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1897 struct free_area *area;
1898 unsigned int current_order;
1903 /* Find the largest possible block of pages in the other list */
1904 for (current_order = MAX_ORDER-1;
1905 current_order >= order && current_order <= MAX_ORDER-1;
1907 area = &(zone->free_area[current_order]);
1908 fallback_mt = find_suitable_fallback(area, current_order,
1909 start_migratetype, false, &can_steal);
1910 if (fallback_mt == -1)
1913 page = list_first_entry(&area->free_list[fallback_mt],
1916 steal_suitable_fallback(zone, page, start_migratetype);
1918 /* Remove the page from the freelists */
1920 list_del(&page->lru);
1921 rmv_page_order(page);
1923 expand(zone, page, order, current_order, area,
1926 * The pcppage_migratetype may differ from pageblock's
1927 * migratetype depending on the decisions in
1928 * find_suitable_fallback(). This is OK as long as it does not
1929 * differ for MIGRATE_CMA pageblocks. Those can be used as
1930 * fallback only via special __rmqueue_cma_fallback() function
1932 set_pcppage_migratetype(page, start_migratetype);
1934 trace_mm_page_alloc_extfrag(page, order, current_order,
1935 start_migratetype, fallback_mt);
1944 * Do the hard work of removing an element from the buddy allocator.
1945 * Call me with the zone->lock already held.
1947 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1952 page = __rmqueue_smallest(zone, order, migratetype);
1953 if (unlikely(!page)) {
1954 if (migratetype == MIGRATE_MOVABLE)
1955 page = __rmqueue_cma_fallback(zone, order);
1958 page = __rmqueue_fallback(zone, order, migratetype);
1961 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1966 * Obtain a specified number of elements from the buddy allocator, all under
1967 * a single hold of the lock, for efficiency. Add them to the supplied list.
1968 * Returns the number of new pages which were placed at *list.
1970 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1971 unsigned long count, struct list_head *list,
1972 int migratetype, bool cold)
1976 spin_lock(&zone->lock);
1977 for (i = 0; i < count; ++i) {
1978 struct page *page = __rmqueue(zone, order, migratetype);
1979 if (unlikely(page == NULL))
1983 * Split buddy pages returned by expand() are received here
1984 * in physical page order. The page is added to the callers and
1985 * list and the list head then moves forward. From the callers
1986 * perspective, the linked list is ordered by page number in
1987 * some conditions. This is useful for IO devices that can
1988 * merge IO requests if the physical pages are ordered
1992 list_add(&page->lru, list);
1994 list_add_tail(&page->lru, list);
1996 if (is_migrate_cma(get_pcppage_migratetype(page)))
1997 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2000 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2001 spin_unlock(&zone->lock);
2007 * Called from the vmstat counter updater to drain pagesets of this
2008 * currently executing processor on remote nodes after they have
2011 * Note that this function must be called with the thread pinned to
2012 * a single processor.
2014 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2016 unsigned long flags;
2017 int to_drain, batch;
2019 local_irq_save(flags);
2020 batch = READ_ONCE(pcp->batch);
2021 to_drain = min(pcp->count, batch);
2023 free_pcppages_bulk(zone, to_drain, pcp);
2024 pcp->count -= to_drain;
2026 local_irq_restore(flags);
2031 * Drain pcplists of the indicated processor and zone.
2033 * The processor must either be the current processor and the
2034 * thread pinned to the current processor or a processor that
2037 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2039 unsigned long flags;
2040 struct per_cpu_pageset *pset;
2041 struct per_cpu_pages *pcp;
2043 local_irq_save(flags);
2044 pset = per_cpu_ptr(zone->pageset, cpu);
2048 free_pcppages_bulk(zone, pcp->count, pcp);
2051 local_irq_restore(flags);
2055 * Drain pcplists of all zones on the indicated processor.
2057 * The processor must either be the current processor and the
2058 * thread pinned to the current processor or a processor that
2061 static void drain_pages(unsigned int cpu)
2065 for_each_populated_zone(zone) {
2066 drain_pages_zone(cpu, zone);
2071 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2073 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2074 * the single zone's pages.
2076 void drain_local_pages(struct zone *zone)
2078 int cpu = smp_processor_id();
2081 drain_pages_zone(cpu, zone);
2087 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2089 * When zone parameter is non-NULL, spill just the single zone's pages.
2091 * Note that this code is protected against sending an IPI to an offline
2092 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2093 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2094 * nothing keeps CPUs from showing up after we populated the cpumask and
2095 * before the call to on_each_cpu_mask().
2097 void drain_all_pages(struct zone *zone)
2102 * Allocate in the BSS so we wont require allocation in
2103 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2105 static cpumask_t cpus_with_pcps;
2108 * We don't care about racing with CPU hotplug event
2109 * as offline notification will cause the notified
2110 * cpu to drain that CPU pcps and on_each_cpu_mask
2111 * disables preemption as part of its processing
2113 for_each_online_cpu(cpu) {
2114 struct per_cpu_pageset *pcp;
2116 bool has_pcps = false;
2119 pcp = per_cpu_ptr(zone->pageset, cpu);
2123 for_each_populated_zone(z) {
2124 pcp = per_cpu_ptr(z->pageset, cpu);
2125 if (pcp->pcp.count) {
2133 cpumask_set_cpu(cpu, &cpus_with_pcps);
2135 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2137 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2141 #ifdef CONFIG_HIBERNATION
2143 void mark_free_pages(struct zone *zone)
2145 unsigned long pfn, max_zone_pfn;
2146 unsigned long flags;
2147 unsigned int order, t;
2150 if (zone_is_empty(zone))
2153 spin_lock_irqsave(&zone->lock, flags);
2155 max_zone_pfn = zone_end_pfn(zone);
2156 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2157 if (pfn_valid(pfn)) {
2158 page = pfn_to_page(pfn);
2160 if (page_zone(page) != zone)
2163 if (!swsusp_page_is_forbidden(page))
2164 swsusp_unset_page_free(page);
2167 for_each_migratetype_order(order, t) {
2168 list_for_each_entry(page,
2169 &zone->free_area[order].free_list[t], lru) {
2172 pfn = page_to_pfn(page);
2173 for (i = 0; i < (1UL << order); i++)
2174 swsusp_set_page_free(pfn_to_page(pfn + i));
2177 spin_unlock_irqrestore(&zone->lock, flags);
2179 #endif /* CONFIG_PM */
2182 * Free a 0-order page
2183 * cold == true ? free a cold page : free a hot page
2185 void free_hot_cold_page(struct page *page, bool cold)
2187 struct zone *zone = page_zone(page);
2188 struct per_cpu_pages *pcp;
2189 unsigned long flags;
2190 unsigned long pfn = page_to_pfn(page);
2193 if (!free_pages_prepare(page, 0))
2196 migratetype = get_pfnblock_migratetype(page, pfn);
2197 set_pcppage_migratetype(page, migratetype);
2198 local_irq_save(flags);
2199 __count_vm_event(PGFREE);
2202 * We only track unmovable, reclaimable and movable on pcp lists.
2203 * Free ISOLATE pages back to the allocator because they are being
2204 * offlined but treat RESERVE as movable pages so we can get those
2205 * areas back if necessary. Otherwise, we may have to free
2206 * excessively into the page allocator
2208 if (migratetype >= MIGRATE_PCPTYPES) {
2209 if (unlikely(is_migrate_isolate(migratetype))) {
2210 free_one_page(zone, page, pfn, 0, migratetype);
2213 migratetype = MIGRATE_MOVABLE;
2216 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2218 list_add(&page->lru, &pcp->lists[migratetype]);
2220 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2222 if (pcp->count >= pcp->high) {
2223 unsigned long batch = READ_ONCE(pcp->batch);
2224 free_pcppages_bulk(zone, batch, pcp);
2225 pcp->count -= batch;
2229 local_irq_restore(flags);
2233 * Free a list of 0-order pages
2235 void free_hot_cold_page_list(struct list_head *list, bool cold)
2237 struct page *page, *next;
2239 list_for_each_entry_safe(page, next, list, lru) {
2240 trace_mm_page_free_batched(page, cold);
2241 free_hot_cold_page(page, cold);
2246 * split_page takes a non-compound higher-order page, and splits it into
2247 * n (1<<order) sub-pages: page[0..n]
2248 * Each sub-page must be freed individually.
2250 * Note: this is probably too low level an operation for use in drivers.
2251 * Please consult with lkml before using this in your driver.
2253 void split_page(struct page *page, unsigned int order)
2258 VM_BUG_ON_PAGE(PageCompound(page), page);
2259 VM_BUG_ON_PAGE(!page_count(page), page);
2261 #ifdef CONFIG_KMEMCHECK
2263 * Split shadow pages too, because free(page[0]) would
2264 * otherwise free the whole shadow.
2266 if (kmemcheck_page_is_tracked(page))
2267 split_page(virt_to_page(page[0].shadow), order);
2270 gfp_mask = get_page_owner_gfp(page);
2271 set_page_owner(page, 0, gfp_mask);
2272 for (i = 1; i < (1 << order); i++) {
2273 set_page_refcounted(page + i);
2274 set_page_owner(page + i, 0, gfp_mask);
2277 EXPORT_SYMBOL_GPL(split_page);
2279 int __isolate_free_page(struct page *page, unsigned int order)
2281 unsigned long watermark;
2285 BUG_ON(!PageBuddy(page));
2287 zone = page_zone(page);
2288 mt = get_pageblock_migratetype(page);
2290 if (!is_migrate_isolate(mt)) {
2291 /* Obey watermarks as if the page was being allocated */
2292 watermark = low_wmark_pages(zone) + (1 << order);
2293 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2296 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2299 /* Remove page from free list */
2300 list_del(&page->lru);
2301 zone->free_area[order].nr_free--;
2302 rmv_page_order(page);
2304 set_page_owner(page, order, __GFP_MOVABLE);
2306 /* Set the pageblock if the isolated page is at least a pageblock */
2307 if (order >= pageblock_order - 1) {
2308 struct page *endpage = page + (1 << order) - 1;
2309 for (; page < endpage; page += pageblock_nr_pages) {
2310 int mt = get_pageblock_migratetype(page);
2311 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2312 set_pageblock_migratetype(page,
2318 return 1UL << order;
2322 * Similar to split_page except the page is already free. As this is only
2323 * being used for migration, the migratetype of the block also changes.
2324 * As this is called with interrupts disabled, the caller is responsible
2325 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2328 * Note: this is probably too low level an operation for use in drivers.
2329 * Please consult with lkml before using this in your driver.
2331 int split_free_page(struct page *page)
2336 order = page_order(page);
2338 nr_pages = __isolate_free_page(page, order);
2342 /* Split into individual pages */
2343 set_page_refcounted(page);
2344 split_page(page, order);
2349 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2352 struct page *buffered_rmqueue(struct zone *preferred_zone,
2353 struct zone *zone, unsigned int order,
2354 gfp_t gfp_flags, int alloc_flags, int migratetype)
2356 unsigned long flags;
2358 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2360 if (likely(order == 0)) {
2361 struct per_cpu_pages *pcp;
2362 struct list_head *list;
2364 local_irq_save(flags);
2365 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2366 list = &pcp->lists[migratetype];
2367 if (list_empty(list)) {
2368 pcp->count += rmqueue_bulk(zone, 0,
2371 if (unlikely(list_empty(list)))
2376 page = list_last_entry(list, struct page, lru);
2378 page = list_first_entry(list, struct page, lru);
2380 list_del(&page->lru);
2384 * We most definitely don't want callers attempting to
2385 * allocate greater than order-1 page units with __GFP_NOFAIL.
2387 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2388 spin_lock_irqsave(&zone->lock, flags);
2391 if (alloc_flags & ALLOC_HARDER) {
2392 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2394 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2397 page = __rmqueue(zone, order, migratetype);
2398 spin_unlock(&zone->lock);
2401 __mod_zone_freepage_state(zone, -(1 << order),
2402 get_pcppage_migratetype(page));
2405 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2406 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2407 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2408 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2410 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2411 zone_statistics(preferred_zone, zone, gfp_flags);
2412 local_irq_restore(flags);
2414 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2418 local_irq_restore(flags);
2422 #ifdef CONFIG_FAIL_PAGE_ALLOC
2425 struct fault_attr attr;
2427 bool ignore_gfp_highmem;
2428 bool ignore_gfp_reclaim;
2430 } fail_page_alloc = {
2431 .attr = FAULT_ATTR_INITIALIZER,
2432 .ignore_gfp_reclaim = true,
2433 .ignore_gfp_highmem = true,
2437 static int __init setup_fail_page_alloc(char *str)
2439 return setup_fault_attr(&fail_page_alloc.attr, str);
2441 __setup("fail_page_alloc=", setup_fail_page_alloc);
2443 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2445 if (order < fail_page_alloc.min_order)
2447 if (gfp_mask & __GFP_NOFAIL)
2449 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2451 if (fail_page_alloc.ignore_gfp_reclaim &&
2452 (gfp_mask & __GFP_DIRECT_RECLAIM))
2455 return should_fail(&fail_page_alloc.attr, 1 << order);
2458 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2460 static int __init fail_page_alloc_debugfs(void)
2462 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2465 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2466 &fail_page_alloc.attr);
2468 return PTR_ERR(dir);
2470 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2471 &fail_page_alloc.ignore_gfp_reclaim))
2473 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2474 &fail_page_alloc.ignore_gfp_highmem))
2476 if (!debugfs_create_u32("min-order", mode, dir,
2477 &fail_page_alloc.min_order))
2482 debugfs_remove_recursive(dir);
2487 late_initcall(fail_page_alloc_debugfs);
2489 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2491 #else /* CONFIG_FAIL_PAGE_ALLOC */
2493 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2498 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2501 * Return true if free base pages are above 'mark'. For high-order checks it
2502 * will return true of the order-0 watermark is reached and there is at least
2503 * one free page of a suitable size. Checking now avoids taking the zone lock
2504 * to check in the allocation paths if no pages are free.
2506 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2507 unsigned long mark, int classzone_idx, int alloc_flags,
2512 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2514 /* free_pages may go negative - that's OK */
2515 free_pages -= (1 << order) - 1;
2517 if (alloc_flags & ALLOC_HIGH)
2521 * If the caller does not have rights to ALLOC_HARDER then subtract
2522 * the high-atomic reserves. This will over-estimate the size of the
2523 * atomic reserve but it avoids a search.
2525 if (likely(!alloc_harder))
2526 free_pages -= z->nr_reserved_highatomic;
2531 /* If allocation can't use CMA areas don't use free CMA pages */
2532 if (!(alloc_flags & ALLOC_CMA))
2533 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2537 * Check watermarks for an order-0 allocation request. If these
2538 * are not met, then a high-order request also cannot go ahead
2539 * even if a suitable page happened to be free.
2541 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2544 /* If this is an order-0 request then the watermark is fine */
2548 /* For a high-order request, check at least one suitable page is free */
2549 for (o = order; o < MAX_ORDER; o++) {
2550 struct free_area *area = &z->free_area[o];
2559 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2560 if (!list_empty(&area->free_list[mt]))
2565 if ((alloc_flags & ALLOC_CMA) &&
2566 !list_empty(&area->free_list[MIGRATE_CMA])) {
2574 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2575 int classzone_idx, int alloc_flags)
2577 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2578 zone_page_state(z, NR_FREE_PAGES));
2581 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2582 unsigned long mark, int classzone_idx)
2584 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2586 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2587 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2589 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2594 static bool zone_local(struct zone *local_zone, struct zone *zone)
2596 return local_zone->node == zone->node;
2599 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2601 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2604 #else /* CONFIG_NUMA */
2605 static bool zone_local(struct zone *local_zone, struct zone *zone)
2610 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2614 #endif /* CONFIG_NUMA */
2616 static void reset_alloc_batches(struct zone *preferred_zone)
2618 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2621 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2622 high_wmark_pages(zone) - low_wmark_pages(zone) -
2623 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2624 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2625 } while (zone++ != preferred_zone);
2629 * get_page_from_freelist goes through the zonelist trying to allocate
2632 static struct page *
2633 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2634 const struct alloc_context *ac)
2636 struct zonelist *zonelist = ac->zonelist;
2638 struct page *page = NULL;
2640 int nr_fair_skipped = 0;
2641 bool zonelist_rescan;
2644 zonelist_rescan = false;
2647 * Scan zonelist, looking for a zone with enough free.
2648 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2650 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2654 if (cpusets_enabled() &&
2655 (alloc_flags & ALLOC_CPUSET) &&
2656 !cpuset_zone_allowed(zone, gfp_mask))
2659 * Distribute pages in proportion to the individual
2660 * zone size to ensure fair page aging. The zone a
2661 * page was allocated in should have no effect on the
2662 * time the page has in memory before being reclaimed.
2664 if (alloc_flags & ALLOC_FAIR) {
2665 if (!zone_local(ac->preferred_zone, zone))
2667 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2673 * When allocating a page cache page for writing, we
2674 * want to get it from a zone that is within its dirty
2675 * limit, such that no single zone holds more than its
2676 * proportional share of globally allowed dirty pages.
2677 * The dirty limits take into account the zone's
2678 * lowmem reserves and high watermark so that kswapd
2679 * should be able to balance it without having to
2680 * write pages from its LRU list.
2682 * This may look like it could increase pressure on
2683 * lower zones by failing allocations in higher zones
2684 * before they are full. But the pages that do spill
2685 * over are limited as the lower zones are protected
2686 * by this very same mechanism. It should not become
2687 * a practical burden to them.
2689 * XXX: For now, allow allocations to potentially
2690 * exceed the per-zone dirty limit in the slowpath
2691 * (spread_dirty_pages unset) before going into reclaim,
2692 * which is important when on a NUMA setup the allowed
2693 * zones are together not big enough to reach the
2694 * global limit. The proper fix for these situations
2695 * will require awareness of zones in the
2696 * dirty-throttling and the flusher threads.
2698 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2701 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2702 if (!zone_watermark_ok(zone, order, mark,
2703 ac->classzone_idx, alloc_flags)) {
2706 /* Checked here to keep the fast path fast */
2707 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2708 if (alloc_flags & ALLOC_NO_WATERMARKS)
2711 if (zone_reclaim_mode == 0 ||
2712 !zone_allows_reclaim(ac->preferred_zone, zone))
2715 ret = zone_reclaim(zone, gfp_mask, order);
2717 case ZONE_RECLAIM_NOSCAN:
2720 case ZONE_RECLAIM_FULL:
2721 /* scanned but unreclaimable */
2724 /* did we reclaim enough */
2725 if (zone_watermark_ok(zone, order, mark,
2726 ac->classzone_idx, alloc_flags))
2734 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2735 gfp_mask, alloc_flags, ac->migratetype);
2737 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2741 * If this is a high-order atomic allocation then check
2742 * if the pageblock should be reserved for the future
2744 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2745 reserve_highatomic_pageblock(page, zone, order);
2752 * The first pass makes sure allocations are spread fairly within the
2753 * local node. However, the local node might have free pages left
2754 * after the fairness batches are exhausted, and remote zones haven't
2755 * even been considered yet. Try once more without fairness, and
2756 * include remote zones now, before entering the slowpath and waking
2757 * kswapd: prefer spilling to a remote zone over swapping locally.
2759 if (alloc_flags & ALLOC_FAIR) {
2760 alloc_flags &= ~ALLOC_FAIR;
2761 if (nr_fair_skipped) {
2762 zonelist_rescan = true;
2763 reset_alloc_batches(ac->preferred_zone);
2765 if (nr_online_nodes > 1)
2766 zonelist_rescan = true;
2769 if (zonelist_rescan)
2776 * Large machines with many possible nodes should not always dump per-node
2777 * meminfo in irq context.
2779 static inline bool should_suppress_show_mem(void)
2784 ret = in_interrupt();
2789 static DEFINE_RATELIMIT_STATE(nopage_rs,
2790 DEFAULT_RATELIMIT_INTERVAL,
2791 DEFAULT_RATELIMIT_BURST);
2793 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2795 unsigned int filter = SHOW_MEM_FILTER_NODES;
2797 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2798 debug_guardpage_minorder() > 0)
2802 * This documents exceptions given to allocations in certain
2803 * contexts that are allowed to allocate outside current's set
2806 if (!(gfp_mask & __GFP_NOMEMALLOC))
2807 if (test_thread_flag(TIF_MEMDIE) ||
2808 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2809 filter &= ~SHOW_MEM_FILTER_NODES;
2810 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2811 filter &= ~SHOW_MEM_FILTER_NODES;
2814 struct va_format vaf;
2817 va_start(args, fmt);
2822 pr_warn("%pV", &vaf);
2827 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
2828 current->comm, order, gfp_mask, &gfp_mask);
2830 if (!should_suppress_show_mem())
2834 static inline struct page *
2835 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2836 const struct alloc_context *ac, unsigned long *did_some_progress)
2838 struct oom_control oc = {
2839 .zonelist = ac->zonelist,
2840 .nodemask = ac->nodemask,
2841 .gfp_mask = gfp_mask,
2846 *did_some_progress = 0;
2849 * Acquire the oom lock. If that fails, somebody else is
2850 * making progress for us.
2852 if (!mutex_trylock(&oom_lock)) {
2853 *did_some_progress = 1;
2854 schedule_timeout_uninterruptible(1);
2859 * Go through the zonelist yet one more time, keep very high watermark
2860 * here, this is only to catch a parallel oom killing, we must fail if
2861 * we're still under heavy pressure.
2863 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2864 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2868 if (!(gfp_mask & __GFP_NOFAIL)) {
2869 /* Coredumps can quickly deplete all memory reserves */
2870 if (current->flags & PF_DUMPCORE)
2872 /* The OOM killer will not help higher order allocs */
2873 if (order > PAGE_ALLOC_COSTLY_ORDER)
2875 /* The OOM killer does not needlessly kill tasks for lowmem */
2876 if (ac->high_zoneidx < ZONE_NORMAL)
2878 if (pm_suspended_storage())
2881 * XXX: GFP_NOFS allocations should rather fail than rely on
2882 * other request to make a forward progress.
2883 * We are in an unfortunate situation where out_of_memory cannot
2884 * do much for this context but let's try it to at least get
2885 * access to memory reserved if the current task is killed (see
2886 * out_of_memory). Once filesystems are ready to handle allocation
2887 * failures more gracefully we should just bail out here.
2890 /* The OOM killer may not free memory on a specific node */
2891 if (gfp_mask & __GFP_THISNODE)
2894 /* Exhausted what can be done so it's blamo time */
2895 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
2896 *did_some_progress = 1;
2898 if (gfp_mask & __GFP_NOFAIL) {
2899 page = get_page_from_freelist(gfp_mask, order,
2900 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
2902 * fallback to ignore cpuset restriction if our nodes
2906 page = get_page_from_freelist(gfp_mask, order,
2907 ALLOC_NO_WATERMARKS, ac);
2911 mutex_unlock(&oom_lock);
2915 #ifdef CONFIG_COMPACTION
2916 /* Try memory compaction for high-order allocations before reclaim */
2917 static struct page *
2918 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2919 int alloc_flags, const struct alloc_context *ac,
2920 enum migrate_mode mode, int *contended_compaction,
2921 bool *deferred_compaction)
2923 unsigned long compact_result;
2929 current->flags |= PF_MEMALLOC;
2930 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2931 mode, contended_compaction);
2932 current->flags &= ~PF_MEMALLOC;
2934 switch (compact_result) {
2935 case COMPACT_DEFERRED:
2936 *deferred_compaction = true;
2938 case COMPACT_SKIPPED:
2945 * At least in one zone compaction wasn't deferred or skipped, so let's
2946 * count a compaction stall
2948 count_vm_event(COMPACTSTALL);
2950 page = get_page_from_freelist(gfp_mask, order,
2951 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2954 struct zone *zone = page_zone(page);
2956 zone->compact_blockskip_flush = false;
2957 compaction_defer_reset(zone, order, true);
2958 count_vm_event(COMPACTSUCCESS);
2963 * It's bad if compaction run occurs and fails. The most likely reason
2964 * is that pages exist, but not enough to satisfy watermarks.
2966 count_vm_event(COMPACTFAIL);
2973 static inline struct page *
2974 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2975 int alloc_flags, const struct alloc_context *ac,
2976 enum migrate_mode mode, int *contended_compaction,
2977 bool *deferred_compaction)
2981 #endif /* CONFIG_COMPACTION */
2983 /* Perform direct synchronous page reclaim */
2985 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2986 const struct alloc_context *ac)
2988 struct reclaim_state reclaim_state;
2993 /* We now go into synchronous reclaim */
2994 cpuset_memory_pressure_bump();
2995 current->flags |= PF_MEMALLOC;
2996 lockdep_set_current_reclaim_state(gfp_mask);
2997 reclaim_state.reclaimed_slab = 0;
2998 current->reclaim_state = &reclaim_state;
3000 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3003 current->reclaim_state = NULL;
3004 lockdep_clear_current_reclaim_state();
3005 current->flags &= ~PF_MEMALLOC;
3012 /* The really slow allocator path where we enter direct reclaim */
3013 static inline struct page *
3014 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3015 int alloc_flags, const struct alloc_context *ac,
3016 unsigned long *did_some_progress)
3018 struct page *page = NULL;
3019 bool drained = false;
3021 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3022 if (unlikely(!(*did_some_progress)))
3026 page = get_page_from_freelist(gfp_mask, order,
3027 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3030 * If an allocation failed after direct reclaim, it could be because
3031 * pages are pinned on the per-cpu lists or in high alloc reserves.
3032 * Shrink them them and try again
3034 if (!page && !drained) {
3035 unreserve_highatomic_pageblock(ac);
3036 drain_all_pages(NULL);
3044 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3049 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3050 ac->high_zoneidx, ac->nodemask)
3051 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
3055 gfp_to_alloc_flags(gfp_t gfp_mask)
3057 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3059 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3060 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3063 * The caller may dip into page reserves a bit more if the caller
3064 * cannot run direct reclaim, or if the caller has realtime scheduling
3065 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3066 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3068 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3070 if (gfp_mask & __GFP_ATOMIC) {
3072 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3073 * if it can't schedule.
3075 if (!(gfp_mask & __GFP_NOMEMALLOC))
3076 alloc_flags |= ALLOC_HARDER;
3078 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3079 * comment for __cpuset_node_allowed().
3081 alloc_flags &= ~ALLOC_CPUSET;
3082 } else if (unlikely(rt_task(current)) && !in_interrupt())
3083 alloc_flags |= ALLOC_HARDER;
3085 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
3086 if (gfp_mask & __GFP_MEMALLOC)
3087 alloc_flags |= ALLOC_NO_WATERMARKS;
3088 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3089 alloc_flags |= ALLOC_NO_WATERMARKS;
3090 else if (!in_interrupt() &&
3091 ((current->flags & PF_MEMALLOC) ||
3092 unlikely(test_thread_flag(TIF_MEMDIE))))
3093 alloc_flags |= ALLOC_NO_WATERMARKS;
3096 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3097 alloc_flags |= ALLOC_CMA;
3102 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3104 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
3107 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
3109 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
3112 static inline struct page *
3113 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3114 struct alloc_context *ac)
3116 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3117 struct page *page = NULL;
3119 unsigned long pages_reclaimed = 0;
3120 unsigned long did_some_progress;
3121 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3122 bool deferred_compaction = false;
3123 int contended_compaction = COMPACT_CONTENDED_NONE;
3126 * In the slowpath, we sanity check order to avoid ever trying to
3127 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3128 * be using allocators in order of preference for an area that is
3131 if (order >= MAX_ORDER) {
3132 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3137 * We also sanity check to catch abuse of atomic reserves being used by
3138 * callers that are not in atomic context.
3140 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3141 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3142 gfp_mask &= ~__GFP_ATOMIC;
3145 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3146 wake_all_kswapds(order, ac);
3149 * OK, we're below the kswapd watermark and have kicked background
3150 * reclaim. Now things get more complex, so set up alloc_flags according
3151 * to how we want to proceed.
3153 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3156 * Find the true preferred zone if the allocation is unconstrained by
3159 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3160 struct zoneref *preferred_zoneref;
3161 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3162 ac->high_zoneidx, NULL, &ac->preferred_zone);
3163 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3166 /* This is the last chance, in general, before the goto nopage. */
3167 page = get_page_from_freelist(gfp_mask, order,
3168 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3172 /* Allocate without watermarks if the context allows */
3173 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3175 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3176 * the allocation is high priority and these type of
3177 * allocations are system rather than user orientated
3179 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3180 page = get_page_from_freelist(gfp_mask, order,
3181 ALLOC_NO_WATERMARKS, ac);
3186 /* Caller is not willing to reclaim, we can't balance anything */
3187 if (!can_direct_reclaim) {
3189 * All existing users of the __GFP_NOFAIL are blockable, so warn
3190 * of any new users that actually allow this type of allocation
3193 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3197 /* Avoid recursion of direct reclaim */
3198 if (current->flags & PF_MEMALLOC) {
3200 * __GFP_NOFAIL request from this context is rather bizarre
3201 * because we cannot reclaim anything and only can loop waiting
3202 * for somebody to do a work for us.
3204 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3211 /* Avoid allocations with no watermarks from looping endlessly */
3212 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3216 * Try direct compaction. The first pass is asynchronous. Subsequent
3217 * attempts after direct reclaim are synchronous
3219 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3221 &contended_compaction,
3222 &deferred_compaction);
3226 /* Checks for THP-specific high-order allocations */
3227 if (is_thp_gfp_mask(gfp_mask)) {
3229 * If compaction is deferred for high-order allocations, it is
3230 * because sync compaction recently failed. If this is the case
3231 * and the caller requested a THP allocation, we do not want
3232 * to heavily disrupt the system, so we fail the allocation
3233 * instead of entering direct reclaim.
3235 if (deferred_compaction)
3239 * In all zones where compaction was attempted (and not
3240 * deferred or skipped), lock contention has been detected.
3241 * For THP allocation we do not want to disrupt the others
3242 * so we fallback to base pages instead.
3244 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3248 * If compaction was aborted due to need_resched(), we do not
3249 * want to further increase allocation latency, unless it is
3250 * khugepaged trying to collapse.
3252 if (contended_compaction == COMPACT_CONTENDED_SCHED
3253 && !(current->flags & PF_KTHREAD))
3258 * It can become very expensive to allocate transparent hugepages at
3259 * fault, so use asynchronous memory compaction for THP unless it is
3260 * khugepaged trying to collapse.
3262 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3263 migration_mode = MIGRATE_SYNC_LIGHT;
3265 /* Try direct reclaim and then allocating */
3266 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3267 &did_some_progress);
3271 /* Do not loop if specifically requested */
3272 if (gfp_mask & __GFP_NORETRY)
3275 /* Keep reclaiming pages as long as there is reasonable progress */
3276 pages_reclaimed += did_some_progress;
3277 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3278 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3279 /* Wait for some write requests to complete then retry */
3280 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3284 /* Reclaim has failed us, start killing things */
3285 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3289 /* Retry as long as the OOM killer is making progress */
3290 if (did_some_progress)
3295 * High-order allocations do not necessarily loop after
3296 * direct reclaim and reclaim/compaction depends on compaction
3297 * being called after reclaim so call directly if necessary
3299 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3301 &contended_compaction,
3302 &deferred_compaction);
3306 warn_alloc_failed(gfp_mask, order, NULL);
3312 * This is the 'heart' of the zoned buddy allocator.
3315 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3316 struct zonelist *zonelist, nodemask_t *nodemask)
3318 struct zoneref *preferred_zoneref;
3319 struct page *page = NULL;
3320 unsigned int cpuset_mems_cookie;
3321 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3322 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3323 struct alloc_context ac = {
3324 .high_zoneidx = gfp_zone(gfp_mask),
3325 .nodemask = nodemask,
3326 .migratetype = gfpflags_to_migratetype(gfp_mask),
3329 gfp_mask &= gfp_allowed_mask;
3331 lockdep_trace_alloc(gfp_mask);
3333 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3335 if (should_fail_alloc_page(gfp_mask, order))
3339 * Check the zones suitable for the gfp_mask contain at least one
3340 * valid zone. It's possible to have an empty zonelist as a result
3341 * of __GFP_THISNODE and a memoryless node
3343 if (unlikely(!zonelist->_zonerefs->zone))
3346 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3347 alloc_flags |= ALLOC_CMA;
3350 cpuset_mems_cookie = read_mems_allowed_begin();
3352 /* We set it here, as __alloc_pages_slowpath might have changed it */
3353 ac.zonelist = zonelist;
3355 /* Dirty zone balancing only done in the fast path */
3356 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3358 /* The preferred zone is used for statistics later */
3359 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3360 ac.nodemask ? : &cpuset_current_mems_allowed,
3361 &ac.preferred_zone);
3362 if (!ac.preferred_zone)
3364 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3366 /* First allocation attempt */
3367 alloc_mask = gfp_mask|__GFP_HARDWALL;
3368 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3369 if (unlikely(!page)) {
3371 * Runtime PM, block IO and its error handling path
3372 * can deadlock because I/O on the device might not
3375 alloc_mask = memalloc_noio_flags(gfp_mask);
3376 ac.spread_dirty_pages = false;
3378 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3381 if (kmemcheck_enabled && page)
3382 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3384 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3388 * When updating a task's mems_allowed, it is possible to race with
3389 * parallel threads in such a way that an allocation can fail while
3390 * the mask is being updated. If a page allocation is about to fail,
3391 * check if the cpuset changed during allocation and if so, retry.
3393 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3398 EXPORT_SYMBOL(__alloc_pages_nodemask);
3401 * Common helper functions.
3403 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3408 * __get_free_pages() returns a 32-bit address, which cannot represent
3411 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3413 page = alloc_pages(gfp_mask, order);
3416 return (unsigned long) page_address(page);
3418 EXPORT_SYMBOL(__get_free_pages);
3420 unsigned long get_zeroed_page(gfp_t gfp_mask)
3422 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3424 EXPORT_SYMBOL(get_zeroed_page);
3426 void __free_pages(struct page *page, unsigned int order)
3428 if (put_page_testzero(page)) {
3430 free_hot_cold_page(page, false);
3432 __free_pages_ok(page, order);
3436 EXPORT_SYMBOL(__free_pages);
3438 void free_pages(unsigned long addr, unsigned int order)
3441 VM_BUG_ON(!virt_addr_valid((void *)addr));
3442 __free_pages(virt_to_page((void *)addr), order);
3446 EXPORT_SYMBOL(free_pages);
3450 * An arbitrary-length arbitrary-offset area of memory which resides
3451 * within a 0 or higher order page. Multiple fragments within that page
3452 * are individually refcounted, in the page's reference counter.
3454 * The page_frag functions below provide a simple allocation framework for
3455 * page fragments. This is used by the network stack and network device
3456 * drivers to provide a backing region of memory for use as either an
3457 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3459 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3462 struct page *page = NULL;
3463 gfp_t gfp = gfp_mask;
3465 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3466 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3468 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3469 PAGE_FRAG_CACHE_MAX_ORDER);
3470 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3472 if (unlikely(!page))
3473 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3475 nc->va = page ? page_address(page) : NULL;
3480 void *__alloc_page_frag(struct page_frag_cache *nc,
3481 unsigned int fragsz, gfp_t gfp_mask)
3483 unsigned int size = PAGE_SIZE;
3487 if (unlikely(!nc->va)) {
3489 page = __page_frag_refill(nc, gfp_mask);
3493 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3494 /* if size can vary use size else just use PAGE_SIZE */
3497 /* Even if we own the page, we do not use atomic_set().
3498 * This would break get_page_unless_zero() users.
3500 page_ref_add(page, size - 1);
3502 /* reset page count bias and offset to start of new frag */
3503 nc->pfmemalloc = page_is_pfmemalloc(page);
3504 nc->pagecnt_bias = size;
3508 offset = nc->offset - fragsz;
3509 if (unlikely(offset < 0)) {
3510 page = virt_to_page(nc->va);
3512 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
3515 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3516 /* if size can vary use size else just use PAGE_SIZE */
3519 /* OK, page count is 0, we can safely set it */
3520 set_page_count(page, size);
3522 /* reset page count bias and offset to start of new frag */
3523 nc->pagecnt_bias = size;
3524 offset = size - fragsz;
3528 nc->offset = offset;
3530 return nc->va + offset;
3532 EXPORT_SYMBOL(__alloc_page_frag);
3535 * Frees a page fragment allocated out of either a compound or order 0 page.
3537 void __free_page_frag(void *addr)
3539 struct page *page = virt_to_head_page(addr);
3541 if (unlikely(put_page_testzero(page)))
3542 __free_pages_ok(page, compound_order(page));
3544 EXPORT_SYMBOL(__free_page_frag);
3547 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3548 * of the current memory cgroup if __GFP_ACCOUNT is set, other than that it is
3549 * equivalent to alloc_pages.
3551 * It should be used when the caller would like to use kmalloc, but since the
3552 * allocation is large, it has to fall back to the page allocator.
3554 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3558 page = alloc_pages(gfp_mask, order);
3559 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3560 __free_pages(page, order);
3566 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3570 page = alloc_pages_node(nid, gfp_mask, order);
3571 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3572 __free_pages(page, order);
3579 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3582 void __free_kmem_pages(struct page *page, unsigned int order)
3584 memcg_kmem_uncharge(page, order);
3585 __free_pages(page, order);
3588 void free_kmem_pages(unsigned long addr, unsigned int order)
3591 VM_BUG_ON(!virt_addr_valid((void *)addr));
3592 __free_kmem_pages(virt_to_page((void *)addr), order);
3596 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3600 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3601 unsigned long used = addr + PAGE_ALIGN(size);
3603 split_page(virt_to_page((void *)addr), order);
3604 while (used < alloc_end) {
3609 return (void *)addr;
3613 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3614 * @size: the number of bytes to allocate
3615 * @gfp_mask: GFP flags for the allocation
3617 * This function is similar to alloc_pages(), except that it allocates the
3618 * minimum number of pages to satisfy the request. alloc_pages() can only
3619 * allocate memory in power-of-two pages.
3621 * This function is also limited by MAX_ORDER.
3623 * Memory allocated by this function must be released by free_pages_exact().
3625 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3627 unsigned int order = get_order(size);
3630 addr = __get_free_pages(gfp_mask, order);
3631 return make_alloc_exact(addr, order, size);
3633 EXPORT_SYMBOL(alloc_pages_exact);
3636 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3638 * @nid: the preferred node ID where memory should be allocated
3639 * @size: the number of bytes to allocate
3640 * @gfp_mask: GFP flags for the allocation
3642 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3645 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3647 unsigned int order = get_order(size);
3648 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3651 return make_alloc_exact((unsigned long)page_address(p), order, size);
3655 * free_pages_exact - release memory allocated via alloc_pages_exact()
3656 * @virt: the value returned by alloc_pages_exact.
3657 * @size: size of allocation, same value as passed to alloc_pages_exact().
3659 * Release the memory allocated by a previous call to alloc_pages_exact.
3661 void free_pages_exact(void *virt, size_t size)
3663 unsigned long addr = (unsigned long)virt;
3664 unsigned long end = addr + PAGE_ALIGN(size);
3666 while (addr < end) {
3671 EXPORT_SYMBOL(free_pages_exact);
3674 * nr_free_zone_pages - count number of pages beyond high watermark
3675 * @offset: The zone index of the highest zone
3677 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3678 * high watermark within all zones at or below a given zone index. For each
3679 * zone, the number of pages is calculated as:
3680 * managed_pages - high_pages
3682 static unsigned long nr_free_zone_pages(int offset)
3687 /* Just pick one node, since fallback list is circular */
3688 unsigned long sum = 0;
3690 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3692 for_each_zone_zonelist(zone, z, zonelist, offset) {
3693 unsigned long size = zone->managed_pages;
3694 unsigned long high = high_wmark_pages(zone);
3703 * nr_free_buffer_pages - count number of pages beyond high watermark
3705 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3706 * watermark within ZONE_DMA and ZONE_NORMAL.
3708 unsigned long nr_free_buffer_pages(void)
3710 return nr_free_zone_pages(gfp_zone(GFP_USER));
3712 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3715 * nr_free_pagecache_pages - count number of pages beyond high watermark
3717 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3718 * high watermark within all zones.
3720 unsigned long nr_free_pagecache_pages(void)
3722 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3725 static inline void show_node(struct zone *zone)
3727 if (IS_ENABLED(CONFIG_NUMA))
3728 printk("Node %d ", zone_to_nid(zone));
3731 long si_mem_available(void)
3734 unsigned long pagecache;
3735 unsigned long wmark_low = 0;
3736 unsigned long pages[NR_LRU_LISTS];
3740 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
3741 pages[lru] = global_page_state(NR_LRU_BASE + lru);
3744 wmark_low += zone->watermark[WMARK_LOW];
3747 * Estimate the amount of memory available for userspace allocations,
3748 * without causing swapping.
3750 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
3753 * Not all the page cache can be freed, otherwise the system will
3754 * start swapping. Assume at least half of the page cache, or the
3755 * low watermark worth of cache, needs to stay.
3757 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
3758 pagecache -= min(pagecache / 2, wmark_low);
3759 available += pagecache;
3762 * Part of the reclaimable slab consists of items that are in use,
3763 * and cannot be freed. Cap this estimate at the low watermark.
3765 available += global_page_state(NR_SLAB_RECLAIMABLE) -
3766 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
3772 EXPORT_SYMBOL_GPL(si_mem_available);
3774 void si_meminfo(struct sysinfo *val)
3776 val->totalram = totalram_pages;
3777 val->sharedram = global_page_state(NR_SHMEM);
3778 val->freeram = global_page_state(NR_FREE_PAGES);
3779 val->bufferram = nr_blockdev_pages();
3780 val->totalhigh = totalhigh_pages;
3781 val->freehigh = nr_free_highpages();
3782 val->mem_unit = PAGE_SIZE;
3785 EXPORT_SYMBOL(si_meminfo);
3788 void si_meminfo_node(struct sysinfo *val, int nid)
3790 int zone_type; /* needs to be signed */
3791 unsigned long managed_pages = 0;
3792 unsigned long managed_highpages = 0;
3793 unsigned long free_highpages = 0;
3794 pg_data_t *pgdat = NODE_DATA(nid);
3796 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3797 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3798 val->totalram = managed_pages;
3799 val->sharedram = node_page_state(nid, NR_SHMEM);
3800 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3801 #ifdef CONFIG_HIGHMEM
3802 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3803 struct zone *zone = &pgdat->node_zones[zone_type];
3805 if (is_highmem(zone)) {
3806 managed_highpages += zone->managed_pages;
3807 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
3810 val->totalhigh = managed_highpages;
3811 val->freehigh = free_highpages;
3813 val->totalhigh = managed_highpages;
3814 val->freehigh = free_highpages;
3816 val->mem_unit = PAGE_SIZE;
3821 * Determine whether the node should be displayed or not, depending on whether
3822 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3824 bool skip_free_areas_node(unsigned int flags, int nid)
3827 unsigned int cpuset_mems_cookie;
3829 if (!(flags & SHOW_MEM_FILTER_NODES))
3833 cpuset_mems_cookie = read_mems_allowed_begin();
3834 ret = !node_isset(nid, cpuset_current_mems_allowed);
3835 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3840 #define K(x) ((x) << (PAGE_SHIFT-10))
3842 static void show_migration_types(unsigned char type)
3844 static const char types[MIGRATE_TYPES] = {
3845 [MIGRATE_UNMOVABLE] = 'U',
3846 [MIGRATE_MOVABLE] = 'M',
3847 [MIGRATE_RECLAIMABLE] = 'E',
3848 [MIGRATE_HIGHATOMIC] = 'H',
3850 [MIGRATE_CMA] = 'C',
3852 #ifdef CONFIG_MEMORY_ISOLATION
3853 [MIGRATE_ISOLATE] = 'I',
3856 char tmp[MIGRATE_TYPES + 1];
3860 for (i = 0; i < MIGRATE_TYPES; i++) {
3861 if (type & (1 << i))
3866 printk("(%s) ", tmp);
3870 * Show free area list (used inside shift_scroll-lock stuff)
3871 * We also calculate the percentage fragmentation. We do this by counting the
3872 * memory on each free list with the exception of the first item on the list.
3875 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3878 void show_free_areas(unsigned int filter)
3880 unsigned long free_pcp = 0;
3884 for_each_populated_zone(zone) {
3885 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3888 for_each_online_cpu(cpu)
3889 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3892 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3893 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3894 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3895 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3896 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3897 " free:%lu free_pcp:%lu free_cma:%lu\n",
3898 global_page_state(NR_ACTIVE_ANON),
3899 global_page_state(NR_INACTIVE_ANON),
3900 global_page_state(NR_ISOLATED_ANON),
3901 global_page_state(NR_ACTIVE_FILE),
3902 global_page_state(NR_INACTIVE_FILE),
3903 global_page_state(NR_ISOLATED_FILE),
3904 global_page_state(NR_UNEVICTABLE),
3905 global_page_state(NR_FILE_DIRTY),
3906 global_page_state(NR_WRITEBACK),
3907 global_page_state(NR_UNSTABLE_NFS),
3908 global_page_state(NR_SLAB_RECLAIMABLE),
3909 global_page_state(NR_SLAB_UNRECLAIMABLE),
3910 global_page_state(NR_FILE_MAPPED),
3911 global_page_state(NR_SHMEM),
3912 global_page_state(NR_PAGETABLE),
3913 global_page_state(NR_BOUNCE),
3914 global_page_state(NR_FREE_PAGES),
3916 global_page_state(NR_FREE_CMA_PAGES));
3918 for_each_populated_zone(zone) {
3921 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3925 for_each_online_cpu(cpu)
3926 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3934 " active_anon:%lukB"
3935 " inactive_anon:%lukB"
3936 " active_file:%lukB"
3937 " inactive_file:%lukB"
3938 " unevictable:%lukB"
3939 " isolated(anon):%lukB"
3940 " isolated(file):%lukB"
3948 " slab_reclaimable:%lukB"
3949 " slab_unreclaimable:%lukB"
3950 " kernel_stack:%lukB"
3957 " writeback_tmp:%lukB"
3958 " pages_scanned:%lu"
3959 " all_unreclaimable? %s"
3962 K(zone_page_state(zone, NR_FREE_PAGES)),
3963 K(min_wmark_pages(zone)),
3964 K(low_wmark_pages(zone)),
3965 K(high_wmark_pages(zone)),
3966 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3967 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3968 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3969 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3970 K(zone_page_state(zone, NR_UNEVICTABLE)),
3971 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3972 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3973 K(zone->present_pages),
3974 K(zone->managed_pages),
3975 K(zone_page_state(zone, NR_MLOCK)),
3976 K(zone_page_state(zone, NR_FILE_DIRTY)),
3977 K(zone_page_state(zone, NR_WRITEBACK)),
3978 K(zone_page_state(zone, NR_FILE_MAPPED)),
3979 K(zone_page_state(zone, NR_SHMEM)),
3980 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3981 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3982 zone_page_state(zone, NR_KERNEL_STACK) *
3984 K(zone_page_state(zone, NR_PAGETABLE)),
3985 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3986 K(zone_page_state(zone, NR_BOUNCE)),
3988 K(this_cpu_read(zone->pageset->pcp.count)),
3989 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3990 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3991 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3992 (!zone_reclaimable(zone) ? "yes" : "no")
3994 printk("lowmem_reserve[]:");
3995 for (i = 0; i < MAX_NR_ZONES; i++)
3996 printk(" %ld", zone->lowmem_reserve[i]);
4000 for_each_populated_zone(zone) {
4002 unsigned long nr[MAX_ORDER], flags, total = 0;
4003 unsigned char types[MAX_ORDER];
4005 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4008 printk("%s: ", zone->name);
4010 spin_lock_irqsave(&zone->lock, flags);
4011 for (order = 0; order < MAX_ORDER; order++) {
4012 struct free_area *area = &zone->free_area[order];
4015 nr[order] = area->nr_free;
4016 total += nr[order] << order;
4019 for (type = 0; type < MIGRATE_TYPES; type++) {
4020 if (!list_empty(&area->free_list[type]))
4021 types[order] |= 1 << type;
4024 spin_unlock_irqrestore(&zone->lock, flags);
4025 for (order = 0; order < MAX_ORDER; order++) {
4026 printk("%lu*%lukB ", nr[order], K(1UL) << order);
4028 show_migration_types(types[order]);
4030 printk("= %lukB\n", K(total));
4033 hugetlb_show_meminfo();
4035 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
4037 show_swap_cache_info();
4040 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4042 zoneref->zone = zone;
4043 zoneref->zone_idx = zone_idx(zone);
4047 * Builds allocation fallback zone lists.
4049 * Add all populated zones of a node to the zonelist.
4051 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4055 enum zone_type zone_type = MAX_NR_ZONES;
4059 zone = pgdat->node_zones + zone_type;
4060 if (populated_zone(zone)) {
4061 zoneref_set_zone(zone,
4062 &zonelist->_zonerefs[nr_zones++]);
4063 check_highest_zone(zone_type);
4065 } while (zone_type);
4073 * 0 = automatic detection of better ordering.
4074 * 1 = order by ([node] distance, -zonetype)
4075 * 2 = order by (-zonetype, [node] distance)
4077 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4078 * the same zonelist. So only NUMA can configure this param.
4080 #define ZONELIST_ORDER_DEFAULT 0
4081 #define ZONELIST_ORDER_NODE 1
4082 #define ZONELIST_ORDER_ZONE 2
4084 /* zonelist order in the kernel.
4085 * set_zonelist_order() will set this to NODE or ZONE.
4087 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4088 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4092 /* The value user specified ....changed by config */
4093 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4094 /* string for sysctl */
4095 #define NUMA_ZONELIST_ORDER_LEN 16
4096 char numa_zonelist_order[16] = "default";
4099 * interface for configure zonelist ordering.
4100 * command line option "numa_zonelist_order"
4101 * = "[dD]efault - default, automatic configuration.
4102 * = "[nN]ode - order by node locality, then by zone within node
4103 * = "[zZ]one - order by zone, then by locality within zone
4106 static int __parse_numa_zonelist_order(char *s)
4108 if (*s == 'd' || *s == 'D') {
4109 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4110 } else if (*s == 'n' || *s == 'N') {
4111 user_zonelist_order = ZONELIST_ORDER_NODE;
4112 } else if (*s == 'z' || *s == 'Z') {
4113 user_zonelist_order = ZONELIST_ORDER_ZONE;
4115 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4121 static __init int setup_numa_zonelist_order(char *s)
4128 ret = __parse_numa_zonelist_order(s);
4130 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4134 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4137 * sysctl handler for numa_zonelist_order
4139 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4140 void __user *buffer, size_t *length,
4143 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4145 static DEFINE_MUTEX(zl_order_mutex);
4147 mutex_lock(&zl_order_mutex);
4149 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4153 strcpy(saved_string, (char *)table->data);
4155 ret = proc_dostring(table, write, buffer, length, ppos);
4159 int oldval = user_zonelist_order;
4161 ret = __parse_numa_zonelist_order((char *)table->data);
4164 * bogus value. restore saved string
4166 strncpy((char *)table->data, saved_string,
4167 NUMA_ZONELIST_ORDER_LEN);
4168 user_zonelist_order = oldval;
4169 } else if (oldval != user_zonelist_order) {
4170 mutex_lock(&zonelists_mutex);
4171 build_all_zonelists(NULL, NULL);
4172 mutex_unlock(&zonelists_mutex);
4176 mutex_unlock(&zl_order_mutex);
4181 #define MAX_NODE_LOAD (nr_online_nodes)
4182 static int node_load[MAX_NUMNODES];
4185 * find_next_best_node - find the next node that should appear in a given node's fallback list
4186 * @node: node whose fallback list we're appending
4187 * @used_node_mask: nodemask_t of already used nodes
4189 * We use a number of factors to determine which is the next node that should
4190 * appear on a given node's fallback list. The node should not have appeared
4191 * already in @node's fallback list, and it should be the next closest node
4192 * according to the distance array (which contains arbitrary distance values
4193 * from each node to each node in the system), and should also prefer nodes
4194 * with no CPUs, since presumably they'll have very little allocation pressure
4195 * on them otherwise.
4196 * It returns -1 if no node is found.
4198 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4201 int min_val = INT_MAX;
4202 int best_node = NUMA_NO_NODE;
4203 const struct cpumask *tmp = cpumask_of_node(0);
4205 /* Use the local node if we haven't already */
4206 if (!node_isset(node, *used_node_mask)) {
4207 node_set(node, *used_node_mask);
4211 for_each_node_state(n, N_MEMORY) {
4213 /* Don't want a node to appear more than once */
4214 if (node_isset(n, *used_node_mask))
4217 /* Use the distance array to find the distance */
4218 val = node_distance(node, n);
4220 /* Penalize nodes under us ("prefer the next node") */
4223 /* Give preference to headless and unused nodes */
4224 tmp = cpumask_of_node(n);
4225 if (!cpumask_empty(tmp))
4226 val += PENALTY_FOR_NODE_WITH_CPUS;
4228 /* Slight preference for less loaded node */
4229 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4230 val += node_load[n];
4232 if (val < min_val) {
4239 node_set(best_node, *used_node_mask);
4246 * Build zonelists ordered by node and zones within node.
4247 * This results in maximum locality--normal zone overflows into local
4248 * DMA zone, if any--but risks exhausting DMA zone.
4250 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4253 struct zonelist *zonelist;
4255 zonelist = &pgdat->node_zonelists[0];
4256 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4258 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4259 zonelist->_zonerefs[j].zone = NULL;
4260 zonelist->_zonerefs[j].zone_idx = 0;
4264 * Build gfp_thisnode zonelists
4266 static void build_thisnode_zonelists(pg_data_t *pgdat)
4269 struct zonelist *zonelist;
4271 zonelist = &pgdat->node_zonelists[1];
4272 j = build_zonelists_node(pgdat, zonelist, 0);
4273 zonelist->_zonerefs[j].zone = NULL;
4274 zonelist->_zonerefs[j].zone_idx = 0;
4278 * Build zonelists ordered by zone and nodes within zones.
4279 * This results in conserving DMA zone[s] until all Normal memory is
4280 * exhausted, but results in overflowing to remote node while memory
4281 * may still exist in local DMA zone.
4283 static int node_order[MAX_NUMNODES];
4285 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4288 int zone_type; /* needs to be signed */
4290 struct zonelist *zonelist;
4292 zonelist = &pgdat->node_zonelists[0];
4294 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4295 for (j = 0; j < nr_nodes; j++) {
4296 node = node_order[j];
4297 z = &NODE_DATA(node)->node_zones[zone_type];
4298 if (populated_zone(z)) {
4300 &zonelist->_zonerefs[pos++]);
4301 check_highest_zone(zone_type);
4305 zonelist->_zonerefs[pos].zone = NULL;
4306 zonelist->_zonerefs[pos].zone_idx = 0;
4309 #if defined(CONFIG_64BIT)
4311 * Devices that require DMA32/DMA are relatively rare and do not justify a
4312 * penalty to every machine in case the specialised case applies. Default
4313 * to Node-ordering on 64-bit NUMA machines
4315 static int default_zonelist_order(void)
4317 return ZONELIST_ORDER_NODE;
4321 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4322 * by the kernel. If processes running on node 0 deplete the low memory zone
4323 * then reclaim will occur more frequency increasing stalls and potentially
4324 * be easier to OOM if a large percentage of the zone is under writeback or
4325 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4326 * Hence, default to zone ordering on 32-bit.
4328 static int default_zonelist_order(void)
4330 return ZONELIST_ORDER_ZONE;
4332 #endif /* CONFIG_64BIT */
4334 static void set_zonelist_order(void)
4336 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4337 current_zonelist_order = default_zonelist_order();
4339 current_zonelist_order = user_zonelist_order;
4342 static void build_zonelists(pg_data_t *pgdat)
4345 nodemask_t used_mask;
4346 int local_node, prev_node;
4347 struct zonelist *zonelist;
4348 unsigned int order = current_zonelist_order;
4350 /* initialize zonelists */
4351 for (i = 0; i < MAX_ZONELISTS; i++) {
4352 zonelist = pgdat->node_zonelists + i;
4353 zonelist->_zonerefs[0].zone = NULL;
4354 zonelist->_zonerefs[0].zone_idx = 0;
4357 /* NUMA-aware ordering of nodes */
4358 local_node = pgdat->node_id;
4359 load = nr_online_nodes;
4360 prev_node = local_node;
4361 nodes_clear(used_mask);
4363 memset(node_order, 0, sizeof(node_order));
4366 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4368 * We don't want to pressure a particular node.
4369 * So adding penalty to the first node in same
4370 * distance group to make it round-robin.
4372 if (node_distance(local_node, node) !=
4373 node_distance(local_node, prev_node))
4374 node_load[node] = load;
4378 if (order == ZONELIST_ORDER_NODE)
4379 build_zonelists_in_node_order(pgdat, node);
4381 node_order[i++] = node; /* remember order */
4384 if (order == ZONELIST_ORDER_ZONE) {
4385 /* calculate node order -- i.e., DMA last! */
4386 build_zonelists_in_zone_order(pgdat, i);
4389 build_thisnode_zonelists(pgdat);
4392 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4394 * Return node id of node used for "local" allocations.
4395 * I.e., first node id of first zone in arg node's generic zonelist.
4396 * Used for initializing percpu 'numa_mem', which is used primarily
4397 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4399 int local_memory_node(int node)
4403 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4404 gfp_zone(GFP_KERNEL),
4411 #else /* CONFIG_NUMA */
4413 static void set_zonelist_order(void)
4415 current_zonelist_order = ZONELIST_ORDER_ZONE;
4418 static void build_zonelists(pg_data_t *pgdat)
4420 int node, local_node;
4422 struct zonelist *zonelist;
4424 local_node = pgdat->node_id;
4426 zonelist = &pgdat->node_zonelists[0];
4427 j = build_zonelists_node(pgdat, zonelist, 0);
4430 * Now we build the zonelist so that it contains the zones
4431 * of all the other nodes.
4432 * We don't want to pressure a particular node, so when
4433 * building the zones for node N, we make sure that the
4434 * zones coming right after the local ones are those from
4435 * node N+1 (modulo N)
4437 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4438 if (!node_online(node))
4440 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4442 for (node = 0; node < local_node; node++) {
4443 if (!node_online(node))
4445 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4448 zonelist->_zonerefs[j].zone = NULL;
4449 zonelist->_zonerefs[j].zone_idx = 0;
4452 #endif /* CONFIG_NUMA */
4455 * Boot pageset table. One per cpu which is going to be used for all
4456 * zones and all nodes. The parameters will be set in such a way
4457 * that an item put on a list will immediately be handed over to
4458 * the buddy list. This is safe since pageset manipulation is done
4459 * with interrupts disabled.
4461 * The boot_pagesets must be kept even after bootup is complete for
4462 * unused processors and/or zones. They do play a role for bootstrapping
4463 * hotplugged processors.
4465 * zoneinfo_show() and maybe other functions do
4466 * not check if the processor is online before following the pageset pointer.
4467 * Other parts of the kernel may not check if the zone is available.
4469 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4470 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4471 static void setup_zone_pageset(struct zone *zone);
4474 * Global mutex to protect against size modification of zonelists
4475 * as well as to serialize pageset setup for the new populated zone.
4477 DEFINE_MUTEX(zonelists_mutex);
4479 /* return values int ....just for stop_machine() */
4480 static int __build_all_zonelists(void *data)
4484 pg_data_t *self = data;
4487 memset(node_load, 0, sizeof(node_load));
4490 if (self && !node_online(self->node_id)) {
4491 build_zonelists(self);
4494 for_each_online_node(nid) {
4495 pg_data_t *pgdat = NODE_DATA(nid);
4497 build_zonelists(pgdat);
4501 * Initialize the boot_pagesets that are going to be used
4502 * for bootstrapping processors. The real pagesets for
4503 * each zone will be allocated later when the per cpu
4504 * allocator is available.
4506 * boot_pagesets are used also for bootstrapping offline
4507 * cpus if the system is already booted because the pagesets
4508 * are needed to initialize allocators on a specific cpu too.
4509 * F.e. the percpu allocator needs the page allocator which
4510 * needs the percpu allocator in order to allocate its pagesets
4511 * (a chicken-egg dilemma).
4513 for_each_possible_cpu(cpu) {
4514 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4516 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4518 * We now know the "local memory node" for each node--
4519 * i.e., the node of the first zone in the generic zonelist.
4520 * Set up numa_mem percpu variable for on-line cpus. During
4521 * boot, only the boot cpu should be on-line; we'll init the
4522 * secondary cpus' numa_mem as they come on-line. During
4523 * node/memory hotplug, we'll fixup all on-line cpus.
4525 if (cpu_online(cpu))
4526 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4533 static noinline void __init
4534 build_all_zonelists_init(void)
4536 __build_all_zonelists(NULL);
4537 mminit_verify_zonelist();
4538 cpuset_init_current_mems_allowed();
4542 * Called with zonelists_mutex held always
4543 * unless system_state == SYSTEM_BOOTING.
4545 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4546 * [we're only called with non-NULL zone through __meminit paths] and
4547 * (2) call of __init annotated helper build_all_zonelists_init
4548 * [protected by SYSTEM_BOOTING].
4550 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4552 set_zonelist_order();
4554 if (system_state == SYSTEM_BOOTING) {
4555 build_all_zonelists_init();
4557 #ifdef CONFIG_MEMORY_HOTPLUG
4559 setup_zone_pageset(zone);
4561 /* we have to stop all cpus to guarantee there is no user
4563 stop_machine(__build_all_zonelists, pgdat, NULL);
4564 /* cpuset refresh routine should be here */
4566 vm_total_pages = nr_free_pagecache_pages();
4568 * Disable grouping by mobility if the number of pages in the
4569 * system is too low to allow the mechanism to work. It would be
4570 * more accurate, but expensive to check per-zone. This check is
4571 * made on memory-hotadd so a system can start with mobility
4572 * disabled and enable it later
4574 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4575 page_group_by_mobility_disabled = 1;
4577 page_group_by_mobility_disabled = 0;
4579 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
4581 zonelist_order_name[current_zonelist_order],
4582 page_group_by_mobility_disabled ? "off" : "on",
4585 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4590 * Helper functions to size the waitqueue hash table.
4591 * Essentially these want to choose hash table sizes sufficiently
4592 * large so that collisions trying to wait on pages are rare.
4593 * But in fact, the number of active page waitqueues on typical
4594 * systems is ridiculously low, less than 200. So this is even
4595 * conservative, even though it seems large.
4597 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4598 * waitqueues, i.e. the size of the waitq table given the number of pages.
4600 #define PAGES_PER_WAITQUEUE 256
4602 #ifndef CONFIG_MEMORY_HOTPLUG
4603 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4605 unsigned long size = 1;
4607 pages /= PAGES_PER_WAITQUEUE;
4609 while (size < pages)
4613 * Once we have dozens or even hundreds of threads sleeping
4614 * on IO we've got bigger problems than wait queue collision.
4615 * Limit the size of the wait table to a reasonable size.
4617 size = min(size, 4096UL);
4619 return max(size, 4UL);
4623 * A zone's size might be changed by hot-add, so it is not possible to determine
4624 * a suitable size for its wait_table. So we use the maximum size now.
4626 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4628 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4629 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4630 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4632 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4633 * or more by the traditional way. (See above). It equals:
4635 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4636 * ia64(16K page size) : = ( 8G + 4M)byte.
4637 * powerpc (64K page size) : = (32G +16M)byte.
4639 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4646 * This is an integer logarithm so that shifts can be used later
4647 * to extract the more random high bits from the multiplicative
4648 * hash function before the remainder is taken.
4650 static inline unsigned long wait_table_bits(unsigned long size)
4656 * Initially all pages are reserved - free ones are freed
4657 * up by free_all_bootmem() once the early boot process is
4658 * done. Non-atomic initialization, single-pass.
4660 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4661 unsigned long start_pfn, enum memmap_context context)
4663 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
4664 unsigned long end_pfn = start_pfn + size;
4665 pg_data_t *pgdat = NODE_DATA(nid);
4667 unsigned long nr_initialised = 0;
4668 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4669 struct memblock_region *r = NULL, *tmp;
4672 if (highest_memmap_pfn < end_pfn - 1)
4673 highest_memmap_pfn = end_pfn - 1;
4676 * Honor reservation requested by the driver for this ZONE_DEVICE
4679 if (altmap && start_pfn == altmap->base_pfn)
4680 start_pfn += altmap->reserve;
4682 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4684 * There can be holes in boot-time mem_map[]s handed to this
4685 * function. They do not exist on hotplugged memory.
4687 if (context != MEMMAP_EARLY)
4690 if (!early_pfn_valid(pfn))
4692 if (!early_pfn_in_nid(pfn, nid))
4694 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
4697 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4699 * If not mirrored_kernelcore and ZONE_MOVABLE exists, range
4700 * from zone_movable_pfn[nid] to end of each node should be
4701 * ZONE_MOVABLE not ZONE_NORMAL. skip it.
4703 if (!mirrored_kernelcore && zone_movable_pfn[nid])
4704 if (zone == ZONE_NORMAL && pfn >= zone_movable_pfn[nid])
4708 * Check given memblock attribute by firmware which can affect
4709 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
4710 * mirrored, it's an overlapped memmap init. skip it.
4712 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
4713 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
4714 for_each_memblock(memory, tmp)
4715 if (pfn < memblock_region_memory_end_pfn(tmp))
4719 if (pfn >= memblock_region_memory_base_pfn(r) &&
4720 memblock_is_mirror(r)) {
4721 /* already initialized as NORMAL */
4722 pfn = memblock_region_memory_end_pfn(r);
4730 * Mark the block movable so that blocks are reserved for
4731 * movable at startup. This will force kernel allocations
4732 * to reserve their blocks rather than leaking throughout
4733 * the address space during boot when many long-lived
4734 * kernel allocations are made.
4736 * bitmap is created for zone's valid pfn range. but memmap
4737 * can be created for invalid pages (for alignment)
4738 * check here not to call set_pageblock_migratetype() against
4741 if (!(pfn & (pageblock_nr_pages - 1))) {
4742 struct page *page = pfn_to_page(pfn);
4744 __init_single_page(page, pfn, zone, nid);
4745 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4747 __init_single_pfn(pfn, zone, nid);
4752 static void __meminit zone_init_free_lists(struct zone *zone)
4754 unsigned int order, t;
4755 for_each_migratetype_order(order, t) {
4756 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4757 zone->free_area[order].nr_free = 0;
4761 #ifndef __HAVE_ARCH_MEMMAP_INIT
4762 #define memmap_init(size, nid, zone, start_pfn) \
4763 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4766 static int zone_batchsize(struct zone *zone)
4772 * The per-cpu-pages pools are set to around 1000th of the
4773 * size of the zone. But no more than 1/2 of a meg.
4775 * OK, so we don't know how big the cache is. So guess.
4777 batch = zone->managed_pages / 1024;
4778 if (batch * PAGE_SIZE > 512 * 1024)
4779 batch = (512 * 1024) / PAGE_SIZE;
4780 batch /= 4; /* We effectively *= 4 below */
4785 * Clamp the batch to a 2^n - 1 value. Having a power
4786 * of 2 value was found to be more likely to have
4787 * suboptimal cache aliasing properties in some cases.
4789 * For example if 2 tasks are alternately allocating
4790 * batches of pages, one task can end up with a lot
4791 * of pages of one half of the possible page colors
4792 * and the other with pages of the other colors.
4794 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4799 /* The deferral and batching of frees should be suppressed under NOMMU
4802 * The problem is that NOMMU needs to be able to allocate large chunks
4803 * of contiguous memory as there's no hardware page translation to
4804 * assemble apparent contiguous memory from discontiguous pages.
4806 * Queueing large contiguous runs of pages for batching, however,
4807 * causes the pages to actually be freed in smaller chunks. As there
4808 * can be a significant delay between the individual batches being
4809 * recycled, this leads to the once large chunks of space being
4810 * fragmented and becoming unavailable for high-order allocations.
4817 * pcp->high and pcp->batch values are related and dependent on one another:
4818 * ->batch must never be higher then ->high.
4819 * The following function updates them in a safe manner without read side
4822 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4823 * those fields changing asynchronously (acording the the above rule).
4825 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4826 * outside of boot time (or some other assurance that no concurrent updaters
4829 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4830 unsigned long batch)
4832 /* start with a fail safe value for batch */
4836 /* Update high, then batch, in order */
4843 /* a companion to pageset_set_high() */
4844 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4846 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4849 static void pageset_init(struct per_cpu_pageset *p)
4851 struct per_cpu_pages *pcp;
4854 memset(p, 0, sizeof(*p));
4858 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4859 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4862 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4865 pageset_set_batch(p, batch);
4869 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4870 * to the value high for the pageset p.
4872 static void pageset_set_high(struct per_cpu_pageset *p,
4875 unsigned long batch = max(1UL, high / 4);
4876 if ((high / 4) > (PAGE_SHIFT * 8))
4877 batch = PAGE_SHIFT * 8;
4879 pageset_update(&p->pcp, high, batch);
4882 static void pageset_set_high_and_batch(struct zone *zone,
4883 struct per_cpu_pageset *pcp)
4885 if (percpu_pagelist_fraction)
4886 pageset_set_high(pcp,
4887 (zone->managed_pages /
4888 percpu_pagelist_fraction));
4890 pageset_set_batch(pcp, zone_batchsize(zone));
4893 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4895 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4898 pageset_set_high_and_batch(zone, pcp);
4901 static void __meminit setup_zone_pageset(struct zone *zone)
4904 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4905 for_each_possible_cpu(cpu)
4906 zone_pageset_init(zone, cpu);
4910 * Allocate per cpu pagesets and initialize them.
4911 * Before this call only boot pagesets were available.
4913 void __init setup_per_cpu_pageset(void)
4917 for_each_populated_zone(zone)
4918 setup_zone_pageset(zone);
4921 static noinline __init_refok
4922 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4928 * The per-page waitqueue mechanism uses hashed waitqueues
4931 zone->wait_table_hash_nr_entries =
4932 wait_table_hash_nr_entries(zone_size_pages);
4933 zone->wait_table_bits =
4934 wait_table_bits(zone->wait_table_hash_nr_entries);
4935 alloc_size = zone->wait_table_hash_nr_entries
4936 * sizeof(wait_queue_head_t);
4938 if (!slab_is_available()) {
4939 zone->wait_table = (wait_queue_head_t *)
4940 memblock_virt_alloc_node_nopanic(
4941 alloc_size, zone->zone_pgdat->node_id);
4944 * This case means that a zone whose size was 0 gets new memory
4945 * via memory hot-add.
4946 * But it may be the case that a new node was hot-added. In
4947 * this case vmalloc() will not be able to use this new node's
4948 * memory - this wait_table must be initialized to use this new
4949 * node itself as well.
4950 * To use this new node's memory, further consideration will be
4953 zone->wait_table = vmalloc(alloc_size);
4955 if (!zone->wait_table)
4958 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4959 init_waitqueue_head(zone->wait_table + i);
4964 static __meminit void zone_pcp_init(struct zone *zone)
4967 * per cpu subsystem is not up at this point. The following code
4968 * relies on the ability of the linker to provide the
4969 * offset of a (static) per cpu variable into the per cpu area.
4971 zone->pageset = &boot_pageset;
4973 if (populated_zone(zone))
4974 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4975 zone->name, zone->present_pages,
4976 zone_batchsize(zone));
4979 int __meminit init_currently_empty_zone(struct zone *zone,
4980 unsigned long zone_start_pfn,
4983 struct pglist_data *pgdat = zone->zone_pgdat;
4985 ret = zone_wait_table_init(zone, size);
4988 pgdat->nr_zones = zone_idx(zone) + 1;
4990 zone->zone_start_pfn = zone_start_pfn;
4992 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4993 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4995 (unsigned long)zone_idx(zone),
4996 zone_start_pfn, (zone_start_pfn + size));
4998 zone_init_free_lists(zone);
5003 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5004 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5007 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5009 int __meminit __early_pfn_to_nid(unsigned long pfn,
5010 struct mminit_pfnnid_cache *state)
5012 unsigned long start_pfn, end_pfn;
5015 if (state->last_start <= pfn && pfn < state->last_end)
5016 return state->last_nid;
5018 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5020 state->last_start = start_pfn;
5021 state->last_end = end_pfn;
5022 state->last_nid = nid;
5027 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5030 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5031 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5032 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5034 * If an architecture guarantees that all ranges registered contain no holes
5035 * and may be freed, this this function may be used instead of calling
5036 * memblock_free_early_nid() manually.
5038 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5040 unsigned long start_pfn, end_pfn;
5043 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5044 start_pfn = min(start_pfn, max_low_pfn);
5045 end_pfn = min(end_pfn, max_low_pfn);
5047 if (start_pfn < end_pfn)
5048 memblock_free_early_nid(PFN_PHYS(start_pfn),
5049 (end_pfn - start_pfn) << PAGE_SHIFT,
5055 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5056 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5058 * If an architecture guarantees that all ranges registered contain no holes and may
5059 * be freed, this function may be used instead of calling memory_present() manually.
5061 void __init sparse_memory_present_with_active_regions(int nid)
5063 unsigned long start_pfn, end_pfn;
5066 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5067 memory_present(this_nid, start_pfn, end_pfn);
5071 * get_pfn_range_for_nid - Return the start and end page frames for a node
5072 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5073 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5074 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5076 * It returns the start and end page frame of a node based on information
5077 * provided by memblock_set_node(). If called for a node
5078 * with no available memory, a warning is printed and the start and end
5081 void __meminit get_pfn_range_for_nid(unsigned int nid,
5082 unsigned long *start_pfn, unsigned long *end_pfn)
5084 unsigned long this_start_pfn, this_end_pfn;
5090 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5091 *start_pfn = min(*start_pfn, this_start_pfn);
5092 *end_pfn = max(*end_pfn, this_end_pfn);
5095 if (*start_pfn == -1UL)
5100 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5101 * assumption is made that zones within a node are ordered in monotonic
5102 * increasing memory addresses so that the "highest" populated zone is used
5104 static void __init find_usable_zone_for_movable(void)
5107 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5108 if (zone_index == ZONE_MOVABLE)
5111 if (arch_zone_highest_possible_pfn[zone_index] >
5112 arch_zone_lowest_possible_pfn[zone_index])
5116 VM_BUG_ON(zone_index == -1);
5117 movable_zone = zone_index;
5121 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5122 * because it is sized independent of architecture. Unlike the other zones,
5123 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5124 * in each node depending on the size of each node and how evenly kernelcore
5125 * is distributed. This helper function adjusts the zone ranges
5126 * provided by the architecture for a given node by using the end of the
5127 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5128 * zones within a node are in order of monotonic increases memory addresses
5130 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5131 unsigned long zone_type,
5132 unsigned long node_start_pfn,
5133 unsigned long node_end_pfn,
5134 unsigned long *zone_start_pfn,
5135 unsigned long *zone_end_pfn)
5137 /* Only adjust if ZONE_MOVABLE is on this node */
5138 if (zone_movable_pfn[nid]) {
5139 /* Size ZONE_MOVABLE */
5140 if (zone_type == ZONE_MOVABLE) {
5141 *zone_start_pfn = zone_movable_pfn[nid];
5142 *zone_end_pfn = min(node_end_pfn,
5143 arch_zone_highest_possible_pfn[movable_zone]);
5145 /* Check if this whole range is within ZONE_MOVABLE */
5146 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5147 *zone_start_pfn = *zone_end_pfn;
5152 * Return the number of pages a zone spans in a node, including holes
5153 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5155 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5156 unsigned long zone_type,
5157 unsigned long node_start_pfn,
5158 unsigned long node_end_pfn,
5159 unsigned long *zone_start_pfn,
5160 unsigned long *zone_end_pfn,
5161 unsigned long *ignored)
5163 /* When hotadd a new node from cpu_up(), the node should be empty */
5164 if (!node_start_pfn && !node_end_pfn)
5167 /* Get the start and end of the zone */
5168 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5169 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5170 adjust_zone_range_for_zone_movable(nid, zone_type,
5171 node_start_pfn, node_end_pfn,
5172 zone_start_pfn, zone_end_pfn);
5174 /* Check that this node has pages within the zone's required range */
5175 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5178 /* Move the zone boundaries inside the node if necessary */
5179 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5180 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5182 /* Return the spanned pages */
5183 return *zone_end_pfn - *zone_start_pfn;
5187 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5188 * then all holes in the requested range will be accounted for.
5190 unsigned long __meminit __absent_pages_in_range(int nid,
5191 unsigned long range_start_pfn,
5192 unsigned long range_end_pfn)
5194 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5195 unsigned long start_pfn, end_pfn;
5198 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5199 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5200 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5201 nr_absent -= end_pfn - start_pfn;
5207 * absent_pages_in_range - Return number of page frames in holes within a range
5208 * @start_pfn: The start PFN to start searching for holes
5209 * @end_pfn: The end PFN to stop searching for holes
5211 * It returns the number of pages frames in memory holes within a range.
5213 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5214 unsigned long end_pfn)
5216 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5219 /* Return the number of page frames in holes in a zone on a node */
5220 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5221 unsigned long zone_type,
5222 unsigned long node_start_pfn,
5223 unsigned long node_end_pfn,
5224 unsigned long *ignored)
5226 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5227 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5228 unsigned long zone_start_pfn, zone_end_pfn;
5229 unsigned long nr_absent;
5231 /* When hotadd a new node from cpu_up(), the node should be empty */
5232 if (!node_start_pfn && !node_end_pfn)
5235 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5236 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5238 adjust_zone_range_for_zone_movable(nid, zone_type,
5239 node_start_pfn, node_end_pfn,
5240 &zone_start_pfn, &zone_end_pfn);
5241 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5244 * ZONE_MOVABLE handling.
5245 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5248 if (zone_movable_pfn[nid]) {
5249 if (mirrored_kernelcore) {
5250 unsigned long start_pfn, end_pfn;
5251 struct memblock_region *r;
5253 for_each_memblock(memory, r) {
5254 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5255 zone_start_pfn, zone_end_pfn);
5256 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5257 zone_start_pfn, zone_end_pfn);
5259 if (zone_type == ZONE_MOVABLE &&
5260 memblock_is_mirror(r))
5261 nr_absent += end_pfn - start_pfn;
5263 if (zone_type == ZONE_NORMAL &&
5264 !memblock_is_mirror(r))
5265 nr_absent += end_pfn - start_pfn;
5268 if (zone_type == ZONE_NORMAL)
5269 nr_absent += node_end_pfn - zone_movable_pfn[nid];
5276 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5277 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5278 unsigned long zone_type,
5279 unsigned long node_start_pfn,
5280 unsigned long node_end_pfn,
5281 unsigned long *zone_start_pfn,
5282 unsigned long *zone_end_pfn,
5283 unsigned long *zones_size)
5287 *zone_start_pfn = node_start_pfn;
5288 for (zone = 0; zone < zone_type; zone++)
5289 *zone_start_pfn += zones_size[zone];
5291 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5293 return zones_size[zone_type];
5296 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5297 unsigned long zone_type,
5298 unsigned long node_start_pfn,
5299 unsigned long node_end_pfn,
5300 unsigned long *zholes_size)
5305 return zholes_size[zone_type];
5308 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5310 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5311 unsigned long node_start_pfn,
5312 unsigned long node_end_pfn,
5313 unsigned long *zones_size,
5314 unsigned long *zholes_size)
5316 unsigned long realtotalpages = 0, totalpages = 0;
5319 for (i = 0; i < MAX_NR_ZONES; i++) {
5320 struct zone *zone = pgdat->node_zones + i;
5321 unsigned long zone_start_pfn, zone_end_pfn;
5322 unsigned long size, real_size;
5324 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5330 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5331 node_start_pfn, node_end_pfn,
5334 zone->zone_start_pfn = zone_start_pfn;
5336 zone->zone_start_pfn = 0;
5337 zone->spanned_pages = size;
5338 zone->present_pages = real_size;
5341 realtotalpages += real_size;
5344 pgdat->node_spanned_pages = totalpages;
5345 pgdat->node_present_pages = realtotalpages;
5346 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5350 #ifndef CONFIG_SPARSEMEM
5352 * Calculate the size of the zone->blockflags rounded to an unsigned long
5353 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5354 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5355 * round what is now in bits to nearest long in bits, then return it in
5358 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5360 unsigned long usemapsize;
5362 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5363 usemapsize = roundup(zonesize, pageblock_nr_pages);
5364 usemapsize = usemapsize >> pageblock_order;
5365 usemapsize *= NR_PAGEBLOCK_BITS;
5366 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5368 return usemapsize / 8;
5371 static void __init setup_usemap(struct pglist_data *pgdat,
5373 unsigned long zone_start_pfn,
5374 unsigned long zonesize)
5376 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5377 zone->pageblock_flags = NULL;
5379 zone->pageblock_flags =
5380 memblock_virt_alloc_node_nopanic(usemapsize,
5384 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5385 unsigned long zone_start_pfn, unsigned long zonesize) {}
5386 #endif /* CONFIG_SPARSEMEM */
5388 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5390 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5391 void __paginginit set_pageblock_order(void)
5395 /* Check that pageblock_nr_pages has not already been setup */
5396 if (pageblock_order)
5399 if (HPAGE_SHIFT > PAGE_SHIFT)
5400 order = HUGETLB_PAGE_ORDER;
5402 order = MAX_ORDER - 1;
5405 * Assume the largest contiguous order of interest is a huge page.
5406 * This value may be variable depending on boot parameters on IA64 and
5409 pageblock_order = order;
5411 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5414 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5415 * is unused as pageblock_order is set at compile-time. See
5416 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5419 void __paginginit set_pageblock_order(void)
5423 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5425 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5426 unsigned long present_pages)
5428 unsigned long pages = spanned_pages;
5431 * Provide a more accurate estimation if there are holes within
5432 * the zone and SPARSEMEM is in use. If there are holes within the
5433 * zone, each populated memory region may cost us one or two extra
5434 * memmap pages due to alignment because memmap pages for each
5435 * populated regions may not naturally algined on page boundary.
5436 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5438 if (spanned_pages > present_pages + (present_pages >> 4) &&
5439 IS_ENABLED(CONFIG_SPARSEMEM))
5440 pages = present_pages;
5442 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5446 * Set up the zone data structures:
5447 * - mark all pages reserved
5448 * - mark all memory queues empty
5449 * - clear the memory bitmaps
5451 * NOTE: pgdat should get zeroed by caller.
5453 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5456 int nid = pgdat->node_id;
5459 pgdat_resize_init(pgdat);
5460 #ifdef CONFIG_NUMA_BALANCING
5461 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5462 pgdat->numabalancing_migrate_nr_pages = 0;
5463 pgdat->numabalancing_migrate_next_window = jiffies;
5465 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5466 spin_lock_init(&pgdat->split_queue_lock);
5467 INIT_LIST_HEAD(&pgdat->split_queue);
5468 pgdat->split_queue_len = 0;
5470 init_waitqueue_head(&pgdat->kswapd_wait);
5471 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5472 #ifdef CONFIG_COMPACTION
5473 init_waitqueue_head(&pgdat->kcompactd_wait);
5475 pgdat_page_ext_init(pgdat);
5477 for (j = 0; j < MAX_NR_ZONES; j++) {
5478 struct zone *zone = pgdat->node_zones + j;
5479 unsigned long size, realsize, freesize, memmap_pages;
5480 unsigned long zone_start_pfn = zone->zone_start_pfn;
5482 size = zone->spanned_pages;
5483 realsize = freesize = zone->present_pages;
5486 * Adjust freesize so that it accounts for how much memory
5487 * is used by this zone for memmap. This affects the watermark
5488 * and per-cpu initialisations
5490 memmap_pages = calc_memmap_size(size, realsize);
5491 if (!is_highmem_idx(j)) {
5492 if (freesize >= memmap_pages) {
5493 freesize -= memmap_pages;
5496 " %s zone: %lu pages used for memmap\n",
5497 zone_names[j], memmap_pages);
5499 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5500 zone_names[j], memmap_pages, freesize);
5503 /* Account for reserved pages */
5504 if (j == 0 && freesize > dma_reserve) {
5505 freesize -= dma_reserve;
5506 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5507 zone_names[0], dma_reserve);
5510 if (!is_highmem_idx(j))
5511 nr_kernel_pages += freesize;
5512 /* Charge for highmem memmap if there are enough kernel pages */
5513 else if (nr_kernel_pages > memmap_pages * 2)
5514 nr_kernel_pages -= memmap_pages;
5515 nr_all_pages += freesize;
5518 * Set an approximate value for lowmem here, it will be adjusted
5519 * when the bootmem allocator frees pages into the buddy system.
5520 * And all highmem pages will be managed by the buddy system.
5522 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5525 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5527 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5529 zone->name = zone_names[j];
5530 spin_lock_init(&zone->lock);
5531 spin_lock_init(&zone->lru_lock);
5532 zone_seqlock_init(zone);
5533 zone->zone_pgdat = pgdat;
5534 zone_pcp_init(zone);
5536 /* For bootup, initialized properly in watermark setup */
5537 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5539 lruvec_init(&zone->lruvec);
5543 set_pageblock_order();
5544 setup_usemap(pgdat, zone, zone_start_pfn, size);
5545 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5547 memmap_init(size, nid, j, zone_start_pfn);
5551 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5553 unsigned long __maybe_unused start = 0;
5554 unsigned long __maybe_unused offset = 0;
5556 /* Skip empty nodes */
5557 if (!pgdat->node_spanned_pages)
5560 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5561 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5562 offset = pgdat->node_start_pfn - start;
5563 /* ia64 gets its own node_mem_map, before this, without bootmem */
5564 if (!pgdat->node_mem_map) {
5565 unsigned long size, end;
5569 * The zone's endpoints aren't required to be MAX_ORDER
5570 * aligned but the node_mem_map endpoints must be in order
5571 * for the buddy allocator to function correctly.
5573 end = pgdat_end_pfn(pgdat);
5574 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5575 size = (end - start) * sizeof(struct page);
5576 map = alloc_remap(pgdat->node_id, size);
5578 map = memblock_virt_alloc_node_nopanic(size,
5580 pgdat->node_mem_map = map + offset;
5582 #ifndef CONFIG_NEED_MULTIPLE_NODES
5584 * With no DISCONTIG, the global mem_map is just set as node 0's
5586 if (pgdat == NODE_DATA(0)) {
5587 mem_map = NODE_DATA(0)->node_mem_map;
5588 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5589 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5591 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5594 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5597 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5598 unsigned long node_start_pfn, unsigned long *zholes_size)
5600 pg_data_t *pgdat = NODE_DATA(nid);
5601 unsigned long start_pfn = 0;
5602 unsigned long end_pfn = 0;
5604 /* pg_data_t should be reset to zero when it's allocated */
5605 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5607 reset_deferred_meminit(pgdat);
5608 pgdat->node_id = nid;
5609 pgdat->node_start_pfn = node_start_pfn;
5610 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5611 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5612 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5613 (u64)start_pfn << PAGE_SHIFT,
5614 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5616 start_pfn = node_start_pfn;
5618 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5619 zones_size, zholes_size);
5621 alloc_node_mem_map(pgdat);
5622 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5623 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5624 nid, (unsigned long)pgdat,
5625 (unsigned long)pgdat->node_mem_map);
5628 free_area_init_core(pgdat);
5631 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5633 #if MAX_NUMNODES > 1
5635 * Figure out the number of possible node ids.
5637 void __init setup_nr_node_ids(void)
5639 unsigned int highest;
5641 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5642 nr_node_ids = highest + 1;
5647 * node_map_pfn_alignment - determine the maximum internode alignment
5649 * This function should be called after node map is populated and sorted.
5650 * It calculates the maximum power of two alignment which can distinguish
5653 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5654 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5655 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5656 * shifted, 1GiB is enough and this function will indicate so.
5658 * This is used to test whether pfn -> nid mapping of the chosen memory
5659 * model has fine enough granularity to avoid incorrect mapping for the
5660 * populated node map.
5662 * Returns the determined alignment in pfn's. 0 if there is no alignment
5663 * requirement (single node).
5665 unsigned long __init node_map_pfn_alignment(void)
5667 unsigned long accl_mask = 0, last_end = 0;
5668 unsigned long start, end, mask;
5672 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5673 if (!start || last_nid < 0 || last_nid == nid) {
5680 * Start with a mask granular enough to pin-point to the
5681 * start pfn and tick off bits one-by-one until it becomes
5682 * too coarse to separate the current node from the last.
5684 mask = ~((1 << __ffs(start)) - 1);
5685 while (mask && last_end <= (start & (mask << 1)))
5688 /* accumulate all internode masks */
5692 /* convert mask to number of pages */
5693 return ~accl_mask + 1;
5696 /* Find the lowest pfn for a node */
5697 static unsigned long __init find_min_pfn_for_node(int nid)
5699 unsigned long min_pfn = ULONG_MAX;
5700 unsigned long start_pfn;
5703 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5704 min_pfn = min(min_pfn, start_pfn);
5706 if (min_pfn == ULONG_MAX) {
5707 pr_warn("Could not find start_pfn for node %d\n", nid);
5715 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5717 * It returns the minimum PFN based on information provided via
5718 * memblock_set_node().
5720 unsigned long __init find_min_pfn_with_active_regions(void)
5722 return find_min_pfn_for_node(MAX_NUMNODES);
5726 * early_calculate_totalpages()
5727 * Sum pages in active regions for movable zone.
5728 * Populate N_MEMORY for calculating usable_nodes.
5730 static unsigned long __init early_calculate_totalpages(void)
5732 unsigned long totalpages = 0;
5733 unsigned long start_pfn, end_pfn;
5736 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5737 unsigned long pages = end_pfn - start_pfn;
5739 totalpages += pages;
5741 node_set_state(nid, N_MEMORY);
5747 * Find the PFN the Movable zone begins in each node. Kernel memory
5748 * is spread evenly between nodes as long as the nodes have enough
5749 * memory. When they don't, some nodes will have more kernelcore than
5752 static void __init find_zone_movable_pfns_for_nodes(void)
5755 unsigned long usable_startpfn;
5756 unsigned long kernelcore_node, kernelcore_remaining;
5757 /* save the state before borrow the nodemask */
5758 nodemask_t saved_node_state = node_states[N_MEMORY];
5759 unsigned long totalpages = early_calculate_totalpages();
5760 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5761 struct memblock_region *r;
5763 /* Need to find movable_zone earlier when movable_node is specified. */
5764 find_usable_zone_for_movable();
5767 * If movable_node is specified, ignore kernelcore and movablecore
5770 if (movable_node_is_enabled()) {
5771 for_each_memblock(memory, r) {
5772 if (!memblock_is_hotpluggable(r))
5777 usable_startpfn = PFN_DOWN(r->base);
5778 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5779 min(usable_startpfn, zone_movable_pfn[nid]) :
5787 * If kernelcore=mirror is specified, ignore movablecore option
5789 if (mirrored_kernelcore) {
5790 bool mem_below_4gb_not_mirrored = false;
5792 for_each_memblock(memory, r) {
5793 if (memblock_is_mirror(r))
5798 usable_startpfn = memblock_region_memory_base_pfn(r);
5800 if (usable_startpfn < 0x100000) {
5801 mem_below_4gb_not_mirrored = true;
5805 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5806 min(usable_startpfn, zone_movable_pfn[nid]) :
5810 if (mem_below_4gb_not_mirrored)
5811 pr_warn("This configuration results in unmirrored kernel memory.");
5817 * If movablecore=nn[KMG] was specified, calculate what size of
5818 * kernelcore that corresponds so that memory usable for
5819 * any allocation type is evenly spread. If both kernelcore
5820 * and movablecore are specified, then the value of kernelcore
5821 * will be used for required_kernelcore if it's greater than
5822 * what movablecore would have allowed.
5824 if (required_movablecore) {
5825 unsigned long corepages;
5828 * Round-up so that ZONE_MOVABLE is at least as large as what
5829 * was requested by the user
5831 required_movablecore =
5832 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5833 required_movablecore = min(totalpages, required_movablecore);
5834 corepages = totalpages - required_movablecore;
5836 required_kernelcore = max(required_kernelcore, corepages);
5840 * If kernelcore was not specified or kernelcore size is larger
5841 * than totalpages, there is no ZONE_MOVABLE.
5843 if (!required_kernelcore || required_kernelcore >= totalpages)
5846 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5847 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5850 /* Spread kernelcore memory as evenly as possible throughout nodes */
5851 kernelcore_node = required_kernelcore / usable_nodes;
5852 for_each_node_state(nid, N_MEMORY) {
5853 unsigned long start_pfn, end_pfn;
5856 * Recalculate kernelcore_node if the division per node
5857 * now exceeds what is necessary to satisfy the requested
5858 * amount of memory for the kernel
5860 if (required_kernelcore < kernelcore_node)
5861 kernelcore_node = required_kernelcore / usable_nodes;
5864 * As the map is walked, we track how much memory is usable
5865 * by the kernel using kernelcore_remaining. When it is
5866 * 0, the rest of the node is usable by ZONE_MOVABLE
5868 kernelcore_remaining = kernelcore_node;
5870 /* Go through each range of PFNs within this node */
5871 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5872 unsigned long size_pages;
5874 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5875 if (start_pfn >= end_pfn)
5878 /* Account for what is only usable for kernelcore */
5879 if (start_pfn < usable_startpfn) {
5880 unsigned long kernel_pages;
5881 kernel_pages = min(end_pfn, usable_startpfn)
5884 kernelcore_remaining -= min(kernel_pages,
5885 kernelcore_remaining);
5886 required_kernelcore -= min(kernel_pages,
5887 required_kernelcore);
5889 /* Continue if range is now fully accounted */
5890 if (end_pfn <= usable_startpfn) {
5893 * Push zone_movable_pfn to the end so
5894 * that if we have to rebalance
5895 * kernelcore across nodes, we will
5896 * not double account here
5898 zone_movable_pfn[nid] = end_pfn;
5901 start_pfn = usable_startpfn;
5905 * The usable PFN range for ZONE_MOVABLE is from
5906 * start_pfn->end_pfn. Calculate size_pages as the
5907 * number of pages used as kernelcore
5909 size_pages = end_pfn - start_pfn;
5910 if (size_pages > kernelcore_remaining)
5911 size_pages = kernelcore_remaining;
5912 zone_movable_pfn[nid] = start_pfn + size_pages;
5915 * Some kernelcore has been met, update counts and
5916 * break if the kernelcore for this node has been
5919 required_kernelcore -= min(required_kernelcore,
5921 kernelcore_remaining -= size_pages;
5922 if (!kernelcore_remaining)
5928 * If there is still required_kernelcore, we do another pass with one
5929 * less node in the count. This will push zone_movable_pfn[nid] further
5930 * along on the nodes that still have memory until kernelcore is
5934 if (usable_nodes && required_kernelcore > usable_nodes)
5938 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5939 for (nid = 0; nid < MAX_NUMNODES; nid++)
5940 zone_movable_pfn[nid] =
5941 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5944 /* restore the node_state */
5945 node_states[N_MEMORY] = saved_node_state;
5948 /* Any regular or high memory on that node ? */
5949 static void check_for_memory(pg_data_t *pgdat, int nid)
5951 enum zone_type zone_type;
5953 if (N_MEMORY == N_NORMAL_MEMORY)
5956 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5957 struct zone *zone = &pgdat->node_zones[zone_type];
5958 if (populated_zone(zone)) {
5959 node_set_state(nid, N_HIGH_MEMORY);
5960 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5961 zone_type <= ZONE_NORMAL)
5962 node_set_state(nid, N_NORMAL_MEMORY);
5969 * free_area_init_nodes - Initialise all pg_data_t and zone data
5970 * @max_zone_pfn: an array of max PFNs for each zone
5972 * This will call free_area_init_node() for each active node in the system.
5973 * Using the page ranges provided by memblock_set_node(), the size of each
5974 * zone in each node and their holes is calculated. If the maximum PFN
5975 * between two adjacent zones match, it is assumed that the zone is empty.
5976 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5977 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5978 * starts where the previous one ended. For example, ZONE_DMA32 starts
5979 * at arch_max_dma_pfn.
5981 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5983 unsigned long start_pfn, end_pfn;
5986 /* Record where the zone boundaries are */
5987 memset(arch_zone_lowest_possible_pfn, 0,
5988 sizeof(arch_zone_lowest_possible_pfn));
5989 memset(arch_zone_highest_possible_pfn, 0,
5990 sizeof(arch_zone_highest_possible_pfn));
5991 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5992 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5993 for (i = 1; i < MAX_NR_ZONES; i++) {
5994 if (i == ZONE_MOVABLE)
5996 arch_zone_lowest_possible_pfn[i] =
5997 arch_zone_highest_possible_pfn[i-1];
5998 arch_zone_highest_possible_pfn[i] =
5999 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
6001 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
6002 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
6004 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6005 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6006 find_zone_movable_pfns_for_nodes();
6008 /* Print out the zone ranges */
6009 pr_info("Zone ranges:\n");
6010 for (i = 0; i < MAX_NR_ZONES; i++) {
6011 if (i == ZONE_MOVABLE)
6013 pr_info(" %-8s ", zone_names[i]);
6014 if (arch_zone_lowest_possible_pfn[i] ==
6015 arch_zone_highest_possible_pfn[i])
6018 pr_cont("[mem %#018Lx-%#018Lx]\n",
6019 (u64)arch_zone_lowest_possible_pfn[i]
6021 ((u64)arch_zone_highest_possible_pfn[i]
6022 << PAGE_SHIFT) - 1);
6025 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6026 pr_info("Movable zone start for each node\n");
6027 for (i = 0; i < MAX_NUMNODES; i++) {
6028 if (zone_movable_pfn[i])
6029 pr_info(" Node %d: %#018Lx\n", i,
6030 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6033 /* Print out the early node map */
6034 pr_info("Early memory node ranges\n");
6035 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6036 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6037 (u64)start_pfn << PAGE_SHIFT,
6038 ((u64)end_pfn << PAGE_SHIFT) - 1);
6040 /* Initialise every node */
6041 mminit_verify_pageflags_layout();
6042 setup_nr_node_ids();
6043 for_each_online_node(nid) {
6044 pg_data_t *pgdat = NODE_DATA(nid);
6045 free_area_init_node(nid, NULL,
6046 find_min_pfn_for_node(nid), NULL);
6048 /* Any memory on that node */
6049 if (pgdat->node_present_pages)
6050 node_set_state(nid, N_MEMORY);
6051 check_for_memory(pgdat, nid);
6055 static int __init cmdline_parse_core(char *p, unsigned long *core)
6057 unsigned long long coremem;
6061 coremem = memparse(p, &p);
6062 *core = coremem >> PAGE_SHIFT;
6064 /* Paranoid check that UL is enough for the coremem value */
6065 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6071 * kernelcore=size sets the amount of memory for use for allocations that
6072 * cannot be reclaimed or migrated.
6074 static int __init cmdline_parse_kernelcore(char *p)
6076 /* parse kernelcore=mirror */
6077 if (parse_option_str(p, "mirror")) {
6078 mirrored_kernelcore = true;
6082 return cmdline_parse_core(p, &required_kernelcore);
6086 * movablecore=size sets the amount of memory for use for allocations that
6087 * can be reclaimed or migrated.
6089 static int __init cmdline_parse_movablecore(char *p)
6091 return cmdline_parse_core(p, &required_movablecore);
6094 early_param("kernelcore", cmdline_parse_kernelcore);
6095 early_param("movablecore", cmdline_parse_movablecore);
6097 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6099 void adjust_managed_page_count(struct page *page, long count)
6101 spin_lock(&managed_page_count_lock);
6102 page_zone(page)->managed_pages += count;
6103 totalram_pages += count;
6104 #ifdef CONFIG_HIGHMEM
6105 if (PageHighMem(page))
6106 totalhigh_pages += count;
6108 spin_unlock(&managed_page_count_lock);
6110 EXPORT_SYMBOL(adjust_managed_page_count);
6112 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6115 unsigned long pages = 0;
6117 start = (void *)PAGE_ALIGN((unsigned long)start);
6118 end = (void *)((unsigned long)end & PAGE_MASK);
6119 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6120 if ((unsigned int)poison <= 0xFF)
6121 memset(pos, poison, PAGE_SIZE);
6122 free_reserved_page(virt_to_page(pos));
6126 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6127 s, pages << (PAGE_SHIFT - 10), start, end);
6131 EXPORT_SYMBOL(free_reserved_area);
6133 #ifdef CONFIG_HIGHMEM
6134 void free_highmem_page(struct page *page)
6136 __free_reserved_page(page);
6138 page_zone(page)->managed_pages++;
6144 void __init mem_init_print_info(const char *str)
6146 unsigned long physpages, codesize, datasize, rosize, bss_size;
6147 unsigned long init_code_size, init_data_size;
6149 physpages = get_num_physpages();
6150 codesize = _etext - _stext;
6151 datasize = _edata - _sdata;
6152 rosize = __end_rodata - __start_rodata;
6153 bss_size = __bss_stop - __bss_start;
6154 init_data_size = __init_end - __init_begin;
6155 init_code_size = _einittext - _sinittext;
6158 * Detect special cases and adjust section sizes accordingly:
6159 * 1) .init.* may be embedded into .data sections
6160 * 2) .init.text.* may be out of [__init_begin, __init_end],
6161 * please refer to arch/tile/kernel/vmlinux.lds.S.
6162 * 3) .rodata.* may be embedded into .text or .data sections.
6164 #define adj_init_size(start, end, size, pos, adj) \
6166 if (start <= pos && pos < end && size > adj) \
6170 adj_init_size(__init_begin, __init_end, init_data_size,
6171 _sinittext, init_code_size);
6172 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6173 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6174 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6175 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6177 #undef adj_init_size
6179 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6180 #ifdef CONFIG_HIGHMEM
6184 nr_free_pages() << (PAGE_SHIFT - 10),
6185 physpages << (PAGE_SHIFT - 10),
6186 codesize >> 10, datasize >> 10, rosize >> 10,
6187 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6188 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6189 totalcma_pages << (PAGE_SHIFT - 10),
6190 #ifdef CONFIG_HIGHMEM
6191 totalhigh_pages << (PAGE_SHIFT - 10),
6193 str ? ", " : "", str ? str : "");
6197 * set_dma_reserve - set the specified number of pages reserved in the first zone
6198 * @new_dma_reserve: The number of pages to mark reserved
6200 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6201 * In the DMA zone, a significant percentage may be consumed by kernel image
6202 * and other unfreeable allocations which can skew the watermarks badly. This
6203 * function may optionally be used to account for unfreeable pages in the
6204 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6205 * smaller per-cpu batchsize.
6207 void __init set_dma_reserve(unsigned long new_dma_reserve)
6209 dma_reserve = new_dma_reserve;
6212 void __init free_area_init(unsigned long *zones_size)
6214 free_area_init_node(0, zones_size,
6215 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6218 static int page_alloc_cpu_notify(struct notifier_block *self,
6219 unsigned long action, void *hcpu)
6221 int cpu = (unsigned long)hcpu;
6223 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6224 lru_add_drain_cpu(cpu);
6228 * Spill the event counters of the dead processor
6229 * into the current processors event counters.
6230 * This artificially elevates the count of the current
6233 vm_events_fold_cpu(cpu);
6236 * Zero the differential counters of the dead processor
6237 * so that the vm statistics are consistent.
6239 * This is only okay since the processor is dead and cannot
6240 * race with what we are doing.
6242 cpu_vm_stats_fold(cpu);
6247 void __init page_alloc_init(void)
6249 hotcpu_notifier(page_alloc_cpu_notify, 0);
6253 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6254 * or min_free_kbytes changes.
6256 static void calculate_totalreserve_pages(void)
6258 struct pglist_data *pgdat;
6259 unsigned long reserve_pages = 0;
6260 enum zone_type i, j;
6262 for_each_online_pgdat(pgdat) {
6263 for (i = 0; i < MAX_NR_ZONES; i++) {
6264 struct zone *zone = pgdat->node_zones + i;
6267 /* Find valid and maximum lowmem_reserve in the zone */
6268 for (j = i; j < MAX_NR_ZONES; j++) {
6269 if (zone->lowmem_reserve[j] > max)
6270 max = zone->lowmem_reserve[j];
6273 /* we treat the high watermark as reserved pages. */
6274 max += high_wmark_pages(zone);
6276 if (max > zone->managed_pages)
6277 max = zone->managed_pages;
6279 zone->totalreserve_pages = max;
6281 reserve_pages += max;
6284 totalreserve_pages = reserve_pages;
6288 * setup_per_zone_lowmem_reserve - called whenever
6289 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6290 * has a correct pages reserved value, so an adequate number of
6291 * pages are left in the zone after a successful __alloc_pages().
6293 static void setup_per_zone_lowmem_reserve(void)
6295 struct pglist_data *pgdat;
6296 enum zone_type j, idx;
6298 for_each_online_pgdat(pgdat) {
6299 for (j = 0; j < MAX_NR_ZONES; j++) {
6300 struct zone *zone = pgdat->node_zones + j;
6301 unsigned long managed_pages = zone->managed_pages;
6303 zone->lowmem_reserve[j] = 0;
6307 struct zone *lower_zone;
6311 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6312 sysctl_lowmem_reserve_ratio[idx] = 1;
6314 lower_zone = pgdat->node_zones + idx;
6315 lower_zone->lowmem_reserve[j] = managed_pages /
6316 sysctl_lowmem_reserve_ratio[idx];
6317 managed_pages += lower_zone->managed_pages;
6322 /* update totalreserve_pages */
6323 calculate_totalreserve_pages();
6326 static void __setup_per_zone_wmarks(void)
6328 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6329 unsigned long lowmem_pages = 0;
6331 unsigned long flags;
6333 /* Calculate total number of !ZONE_HIGHMEM pages */
6334 for_each_zone(zone) {
6335 if (!is_highmem(zone))
6336 lowmem_pages += zone->managed_pages;
6339 for_each_zone(zone) {
6342 spin_lock_irqsave(&zone->lock, flags);
6343 tmp = (u64)pages_min * zone->managed_pages;
6344 do_div(tmp, lowmem_pages);
6345 if (is_highmem(zone)) {
6347 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6348 * need highmem pages, so cap pages_min to a small
6351 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6352 * deltas control asynch page reclaim, and so should
6353 * not be capped for highmem.
6355 unsigned long min_pages;
6357 min_pages = zone->managed_pages / 1024;
6358 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6359 zone->watermark[WMARK_MIN] = min_pages;
6362 * If it's a lowmem zone, reserve a number of pages
6363 * proportionate to the zone's size.
6365 zone->watermark[WMARK_MIN] = tmp;
6369 * Set the kswapd watermarks distance according to the
6370 * scale factor in proportion to available memory, but
6371 * ensure a minimum size on small systems.
6373 tmp = max_t(u64, tmp >> 2,
6374 mult_frac(zone->managed_pages,
6375 watermark_scale_factor, 10000));
6377 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6378 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6380 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6381 high_wmark_pages(zone) - low_wmark_pages(zone) -
6382 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6384 spin_unlock_irqrestore(&zone->lock, flags);
6387 /* update totalreserve_pages */
6388 calculate_totalreserve_pages();
6392 * setup_per_zone_wmarks - called when min_free_kbytes changes
6393 * or when memory is hot-{added|removed}
6395 * Ensures that the watermark[min,low,high] values for each zone are set
6396 * correctly with respect to min_free_kbytes.
6398 void setup_per_zone_wmarks(void)
6400 mutex_lock(&zonelists_mutex);
6401 __setup_per_zone_wmarks();
6402 mutex_unlock(&zonelists_mutex);
6406 * The inactive anon list should be small enough that the VM never has to
6407 * do too much work, but large enough that each inactive page has a chance
6408 * to be referenced again before it is swapped out.
6410 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6411 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6412 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6413 * the anonymous pages are kept on the inactive list.
6416 * memory ratio inactive anon
6417 * -------------------------------------
6426 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6428 unsigned int gb, ratio;
6430 /* Zone size in gigabytes */
6431 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6433 ratio = int_sqrt(10 * gb);
6437 zone->inactive_ratio = ratio;
6440 static void __meminit setup_per_zone_inactive_ratio(void)
6445 calculate_zone_inactive_ratio(zone);
6449 * Initialise min_free_kbytes.
6451 * For small machines we want it small (128k min). For large machines
6452 * we want it large (64MB max). But it is not linear, because network
6453 * bandwidth does not increase linearly with machine size. We use
6455 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6456 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6472 int __meminit init_per_zone_wmark_min(void)
6474 unsigned long lowmem_kbytes;
6475 int new_min_free_kbytes;
6477 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6478 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6480 if (new_min_free_kbytes > user_min_free_kbytes) {
6481 min_free_kbytes = new_min_free_kbytes;
6482 if (min_free_kbytes < 128)
6483 min_free_kbytes = 128;
6484 if (min_free_kbytes > 65536)
6485 min_free_kbytes = 65536;
6487 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6488 new_min_free_kbytes, user_min_free_kbytes);
6490 setup_per_zone_wmarks();
6491 refresh_zone_stat_thresholds();
6492 setup_per_zone_lowmem_reserve();
6493 setup_per_zone_inactive_ratio();
6496 core_initcall(init_per_zone_wmark_min)
6499 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6500 * that we can call two helper functions whenever min_free_kbytes
6503 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6504 void __user *buffer, size_t *length, loff_t *ppos)
6508 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6513 user_min_free_kbytes = min_free_kbytes;
6514 setup_per_zone_wmarks();
6519 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6520 void __user *buffer, size_t *length, loff_t *ppos)
6524 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6529 setup_per_zone_wmarks();
6535 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6536 void __user *buffer, size_t *length, loff_t *ppos)
6541 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6546 zone->min_unmapped_pages = (zone->managed_pages *
6547 sysctl_min_unmapped_ratio) / 100;
6551 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6552 void __user *buffer, size_t *length, loff_t *ppos)
6557 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6562 zone->min_slab_pages = (zone->managed_pages *
6563 sysctl_min_slab_ratio) / 100;
6569 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6570 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6571 * whenever sysctl_lowmem_reserve_ratio changes.
6573 * The reserve ratio obviously has absolutely no relation with the
6574 * minimum watermarks. The lowmem reserve ratio can only make sense
6575 * if in function of the boot time zone sizes.
6577 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6578 void __user *buffer, size_t *length, loff_t *ppos)
6580 proc_dointvec_minmax(table, write, buffer, length, ppos);
6581 setup_per_zone_lowmem_reserve();
6586 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6587 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6588 * pagelist can have before it gets flushed back to buddy allocator.
6590 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6591 void __user *buffer, size_t *length, loff_t *ppos)
6594 int old_percpu_pagelist_fraction;
6597 mutex_lock(&pcp_batch_high_lock);
6598 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6600 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6601 if (!write || ret < 0)
6604 /* Sanity checking to avoid pcp imbalance */
6605 if (percpu_pagelist_fraction &&
6606 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6607 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6613 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6616 for_each_populated_zone(zone) {
6619 for_each_possible_cpu(cpu)
6620 pageset_set_high_and_batch(zone,
6621 per_cpu_ptr(zone->pageset, cpu));
6624 mutex_unlock(&pcp_batch_high_lock);
6629 int hashdist = HASHDIST_DEFAULT;
6631 static int __init set_hashdist(char *str)
6635 hashdist = simple_strtoul(str, &str, 0);
6638 __setup("hashdist=", set_hashdist);
6642 * allocate a large system hash table from bootmem
6643 * - it is assumed that the hash table must contain an exact power-of-2
6644 * quantity of entries
6645 * - limit is the number of hash buckets, not the total allocation size
6647 void *__init alloc_large_system_hash(const char *tablename,
6648 unsigned long bucketsize,
6649 unsigned long numentries,
6652 unsigned int *_hash_shift,
6653 unsigned int *_hash_mask,
6654 unsigned long low_limit,
6655 unsigned long high_limit)
6657 unsigned long long max = high_limit;
6658 unsigned long log2qty, size;
6661 /* allow the kernel cmdline to have a say */
6663 /* round applicable memory size up to nearest megabyte */
6664 numentries = nr_kernel_pages;
6666 /* It isn't necessary when PAGE_SIZE >= 1MB */
6667 if (PAGE_SHIFT < 20)
6668 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6670 /* limit to 1 bucket per 2^scale bytes of low memory */
6671 if (scale > PAGE_SHIFT)
6672 numentries >>= (scale - PAGE_SHIFT);
6674 numentries <<= (PAGE_SHIFT - scale);
6676 /* Make sure we've got at least a 0-order allocation.. */
6677 if (unlikely(flags & HASH_SMALL)) {
6678 /* Makes no sense without HASH_EARLY */
6679 WARN_ON(!(flags & HASH_EARLY));
6680 if (!(numentries >> *_hash_shift)) {
6681 numentries = 1UL << *_hash_shift;
6682 BUG_ON(!numentries);
6684 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6685 numentries = PAGE_SIZE / bucketsize;
6687 numentries = roundup_pow_of_two(numentries);
6689 /* limit allocation size to 1/16 total memory by default */
6691 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6692 do_div(max, bucketsize);
6694 max = min(max, 0x80000000ULL);
6696 if (numentries < low_limit)
6697 numentries = low_limit;
6698 if (numentries > max)
6701 log2qty = ilog2(numentries);
6704 size = bucketsize << log2qty;
6705 if (flags & HASH_EARLY)
6706 table = memblock_virt_alloc_nopanic(size, 0);
6708 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6711 * If bucketsize is not a power-of-two, we may free
6712 * some pages at the end of hash table which
6713 * alloc_pages_exact() automatically does
6715 if (get_order(size) < MAX_ORDER) {
6716 table = alloc_pages_exact(size, GFP_ATOMIC);
6717 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6720 } while (!table && size > PAGE_SIZE && --log2qty);
6723 panic("Failed to allocate %s hash table\n", tablename);
6725 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
6726 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
6729 *_hash_shift = log2qty;
6731 *_hash_mask = (1 << log2qty) - 1;
6736 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6737 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6740 #ifdef CONFIG_SPARSEMEM
6741 return __pfn_to_section(pfn)->pageblock_flags;
6743 return zone->pageblock_flags;
6744 #endif /* CONFIG_SPARSEMEM */
6747 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6749 #ifdef CONFIG_SPARSEMEM
6750 pfn &= (PAGES_PER_SECTION-1);
6751 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6753 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6754 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6755 #endif /* CONFIG_SPARSEMEM */
6759 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6760 * @page: The page within the block of interest
6761 * @pfn: The target page frame number
6762 * @end_bitidx: The last bit of interest to retrieve
6763 * @mask: mask of bits that the caller is interested in
6765 * Return: pageblock_bits flags
6767 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6768 unsigned long end_bitidx,
6772 unsigned long *bitmap;
6773 unsigned long bitidx, word_bitidx;
6776 zone = page_zone(page);
6777 bitmap = get_pageblock_bitmap(zone, pfn);
6778 bitidx = pfn_to_bitidx(zone, pfn);
6779 word_bitidx = bitidx / BITS_PER_LONG;
6780 bitidx &= (BITS_PER_LONG-1);
6782 word = bitmap[word_bitidx];
6783 bitidx += end_bitidx;
6784 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6788 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6789 * @page: The page within the block of interest
6790 * @flags: The flags to set
6791 * @pfn: The target page frame number
6792 * @end_bitidx: The last bit of interest
6793 * @mask: mask of bits that the caller is interested in
6795 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6797 unsigned long end_bitidx,
6801 unsigned long *bitmap;
6802 unsigned long bitidx, word_bitidx;
6803 unsigned long old_word, word;
6805 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6807 zone = page_zone(page);
6808 bitmap = get_pageblock_bitmap(zone, pfn);
6809 bitidx = pfn_to_bitidx(zone, pfn);
6810 word_bitidx = bitidx / BITS_PER_LONG;
6811 bitidx &= (BITS_PER_LONG-1);
6813 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6815 bitidx += end_bitidx;
6816 mask <<= (BITS_PER_LONG - bitidx - 1);
6817 flags <<= (BITS_PER_LONG - bitidx - 1);
6819 word = READ_ONCE(bitmap[word_bitidx]);
6821 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6822 if (word == old_word)
6829 * This function checks whether pageblock includes unmovable pages or not.
6830 * If @count is not zero, it is okay to include less @count unmovable pages
6832 * PageLRU check without isolation or lru_lock could race so that
6833 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6834 * expect this function should be exact.
6836 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6837 bool skip_hwpoisoned_pages)
6839 unsigned long pfn, iter, found;
6843 * For avoiding noise data, lru_add_drain_all() should be called
6844 * If ZONE_MOVABLE, the zone never contains unmovable pages
6846 if (zone_idx(zone) == ZONE_MOVABLE)
6848 mt = get_pageblock_migratetype(page);
6849 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6852 pfn = page_to_pfn(page);
6853 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6854 unsigned long check = pfn + iter;
6856 if (!pfn_valid_within(check))
6859 page = pfn_to_page(check);
6862 * Hugepages are not in LRU lists, but they're movable.
6863 * We need not scan over tail pages bacause we don't
6864 * handle each tail page individually in migration.
6866 if (PageHuge(page)) {
6867 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6872 * We can't use page_count without pin a page
6873 * because another CPU can free compound page.
6874 * This check already skips compound tails of THP
6875 * because their page->_refcount is zero at all time.
6877 if (!page_ref_count(page)) {
6878 if (PageBuddy(page))
6879 iter += (1 << page_order(page)) - 1;
6884 * The HWPoisoned page may be not in buddy system, and
6885 * page_count() is not 0.
6887 if (skip_hwpoisoned_pages && PageHWPoison(page))
6893 * If there are RECLAIMABLE pages, we need to check
6894 * it. But now, memory offline itself doesn't call
6895 * shrink_node_slabs() and it still to be fixed.
6898 * If the page is not RAM, page_count()should be 0.
6899 * we don't need more check. This is an _used_ not-movable page.
6901 * The problematic thing here is PG_reserved pages. PG_reserved
6902 * is set to both of a memory hole page and a _used_ kernel
6911 bool is_pageblock_removable_nolock(struct page *page)
6917 * We have to be careful here because we are iterating over memory
6918 * sections which are not zone aware so we might end up outside of
6919 * the zone but still within the section.
6920 * We have to take care about the node as well. If the node is offline
6921 * its NODE_DATA will be NULL - see page_zone.
6923 if (!node_online(page_to_nid(page)))
6926 zone = page_zone(page);
6927 pfn = page_to_pfn(page);
6928 if (!zone_spans_pfn(zone, pfn))
6931 return !has_unmovable_pages(zone, page, 0, true);
6934 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
6936 static unsigned long pfn_max_align_down(unsigned long pfn)
6938 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6939 pageblock_nr_pages) - 1);
6942 static unsigned long pfn_max_align_up(unsigned long pfn)
6944 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6945 pageblock_nr_pages));
6948 /* [start, end) must belong to a single zone. */
6949 static int __alloc_contig_migrate_range(struct compact_control *cc,
6950 unsigned long start, unsigned long end)
6952 /* This function is based on compact_zone() from compaction.c. */
6953 unsigned long nr_reclaimed;
6954 unsigned long pfn = start;
6955 unsigned int tries = 0;
6960 while (pfn < end || !list_empty(&cc->migratepages)) {
6961 if (fatal_signal_pending(current)) {
6966 if (list_empty(&cc->migratepages)) {
6967 cc->nr_migratepages = 0;
6968 pfn = isolate_migratepages_range(cc, pfn, end);
6974 } else if (++tries == 5) {
6975 ret = ret < 0 ? ret : -EBUSY;
6979 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6981 cc->nr_migratepages -= nr_reclaimed;
6983 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6984 NULL, 0, cc->mode, MR_CMA);
6987 putback_movable_pages(&cc->migratepages);
6994 * alloc_contig_range() -- tries to allocate given range of pages
6995 * @start: start PFN to allocate
6996 * @end: one-past-the-last PFN to allocate
6997 * @migratetype: migratetype of the underlaying pageblocks (either
6998 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6999 * in range must have the same migratetype and it must
7000 * be either of the two.
7002 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7003 * aligned, however it's the caller's responsibility to guarantee that
7004 * we are the only thread that changes migrate type of pageblocks the
7007 * The PFN range must belong to a single zone.
7009 * Returns zero on success or negative error code. On success all
7010 * pages which PFN is in [start, end) are allocated for the caller and
7011 * need to be freed with free_contig_range().
7013 int alloc_contig_range(unsigned long start, unsigned long end,
7014 unsigned migratetype)
7016 unsigned long outer_start, outer_end;
7020 struct compact_control cc = {
7021 .nr_migratepages = 0,
7023 .zone = page_zone(pfn_to_page(start)),
7024 .mode = MIGRATE_SYNC,
7025 .ignore_skip_hint = true,
7027 INIT_LIST_HEAD(&cc.migratepages);
7030 * What we do here is we mark all pageblocks in range as
7031 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7032 * have different sizes, and due to the way page allocator
7033 * work, we align the range to biggest of the two pages so
7034 * that page allocator won't try to merge buddies from
7035 * different pageblocks and change MIGRATE_ISOLATE to some
7036 * other migration type.
7038 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7039 * migrate the pages from an unaligned range (ie. pages that
7040 * we are interested in). This will put all the pages in
7041 * range back to page allocator as MIGRATE_ISOLATE.
7043 * When this is done, we take the pages in range from page
7044 * allocator removing them from the buddy system. This way
7045 * page allocator will never consider using them.
7047 * This lets us mark the pageblocks back as
7048 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7049 * aligned range but not in the unaligned, original range are
7050 * put back to page allocator so that buddy can use them.
7053 ret = start_isolate_page_range(pfn_max_align_down(start),
7054 pfn_max_align_up(end), migratetype,
7060 * In case of -EBUSY, we'd like to know which page causes problem.
7061 * So, just fall through. We will check it in test_pages_isolated().
7063 ret = __alloc_contig_migrate_range(&cc, start, end);
7064 if (ret && ret != -EBUSY)
7068 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7069 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7070 * more, all pages in [start, end) are free in page allocator.
7071 * What we are going to do is to allocate all pages from
7072 * [start, end) (that is remove them from page allocator).
7074 * The only problem is that pages at the beginning and at the
7075 * end of interesting range may be not aligned with pages that
7076 * page allocator holds, ie. they can be part of higher order
7077 * pages. Because of this, we reserve the bigger range and
7078 * once this is done free the pages we are not interested in.
7080 * We don't have to hold zone->lock here because the pages are
7081 * isolated thus they won't get removed from buddy.
7084 lru_add_drain_all();
7085 drain_all_pages(cc.zone);
7088 outer_start = start;
7089 while (!PageBuddy(pfn_to_page(outer_start))) {
7090 if (++order >= MAX_ORDER) {
7091 outer_start = start;
7094 outer_start &= ~0UL << order;
7097 if (outer_start != start) {
7098 order = page_order(pfn_to_page(outer_start));
7101 * outer_start page could be small order buddy page and
7102 * it doesn't include start page. Adjust outer_start
7103 * in this case to report failed page properly
7104 * on tracepoint in test_pages_isolated()
7106 if (outer_start + (1UL << order) <= start)
7107 outer_start = start;
7110 /* Make sure the range is really isolated. */
7111 if (test_pages_isolated(outer_start, end, false)) {
7112 pr_info("%s: [%lx, %lx) PFNs busy\n",
7113 __func__, outer_start, end);
7118 /* Grab isolated pages from freelists. */
7119 outer_end = isolate_freepages_range(&cc, outer_start, end);
7125 /* Free head and tail (if any) */
7126 if (start != outer_start)
7127 free_contig_range(outer_start, start - outer_start);
7128 if (end != outer_end)
7129 free_contig_range(end, outer_end - end);
7132 undo_isolate_page_range(pfn_max_align_down(start),
7133 pfn_max_align_up(end), migratetype);
7137 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7139 unsigned int count = 0;
7141 for (; nr_pages--; pfn++) {
7142 struct page *page = pfn_to_page(pfn);
7144 count += page_count(page) != 1;
7147 WARN(count != 0, "%d pages are still in use!\n", count);
7151 #ifdef CONFIG_MEMORY_HOTPLUG
7153 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7154 * page high values need to be recalulated.
7156 void __meminit zone_pcp_update(struct zone *zone)
7159 mutex_lock(&pcp_batch_high_lock);
7160 for_each_possible_cpu(cpu)
7161 pageset_set_high_and_batch(zone,
7162 per_cpu_ptr(zone->pageset, cpu));
7163 mutex_unlock(&pcp_batch_high_lock);
7167 void zone_pcp_reset(struct zone *zone)
7169 unsigned long flags;
7171 struct per_cpu_pageset *pset;
7173 /* avoid races with drain_pages() */
7174 local_irq_save(flags);
7175 if (zone->pageset != &boot_pageset) {
7176 for_each_online_cpu(cpu) {
7177 pset = per_cpu_ptr(zone->pageset, cpu);
7178 drain_zonestat(zone, pset);
7180 free_percpu(zone->pageset);
7181 zone->pageset = &boot_pageset;
7183 local_irq_restore(flags);
7186 #ifdef CONFIG_MEMORY_HOTREMOVE
7188 * All pages in the range must be in a single zone and isolated
7189 * before calling this.
7192 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7196 unsigned int order, i;
7198 unsigned long flags;
7199 /* find the first valid pfn */
7200 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7205 zone = page_zone(pfn_to_page(pfn));
7206 spin_lock_irqsave(&zone->lock, flags);
7208 while (pfn < end_pfn) {
7209 if (!pfn_valid(pfn)) {
7213 page = pfn_to_page(pfn);
7215 * The HWPoisoned page may be not in buddy system, and
7216 * page_count() is not 0.
7218 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7220 SetPageReserved(page);
7224 BUG_ON(page_count(page));
7225 BUG_ON(!PageBuddy(page));
7226 order = page_order(page);
7227 #ifdef CONFIG_DEBUG_VM
7228 pr_info("remove from free list %lx %d %lx\n",
7229 pfn, 1 << order, end_pfn);
7231 list_del(&page->lru);
7232 rmv_page_order(page);
7233 zone->free_area[order].nr_free--;
7234 for (i = 0; i < (1 << order); i++)
7235 SetPageReserved((page+i));
7236 pfn += (1 << order);
7238 spin_unlock_irqrestore(&zone->lock, flags);
7242 bool is_free_buddy_page(struct page *page)
7244 struct zone *zone = page_zone(page);
7245 unsigned long pfn = page_to_pfn(page);
7246 unsigned long flags;
7249 spin_lock_irqsave(&zone->lock, flags);
7250 for (order = 0; order < MAX_ORDER; order++) {
7251 struct page *page_head = page - (pfn & ((1 << order) - 1));
7253 if (PageBuddy(page_head) && page_order(page_head) >= order)
7256 spin_unlock_irqrestore(&zone->lock, flags);
7258 return order < MAX_ORDER;