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/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/page_ext.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/locallock.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 * When calculating the number of globally allowed dirty pages, there
120 * is a certain number of per-zone reserves that should not be
121 * considered dirtyable memory. This is the sum of those reserves
122 * over all existing zones that contribute dirtyable memory.
124 unsigned long dirty_balance_reserve __read_mostly;
126 int percpu_pagelist_fraction;
127 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
130 * A cached value of the page's pageblock's migratetype, used when the page is
131 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
132 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
133 * Also the migratetype set in the page does not necessarily match the pcplist
134 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
135 * other index - this ensures that it will be put on the correct CMA freelist.
137 static inline int get_pcppage_migratetype(struct page *page)
142 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
144 page->index = migratetype;
147 #ifdef CONFIG_PM_SLEEP
149 * The following functions are used by the suspend/hibernate code to temporarily
150 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
151 * while devices are suspended. To avoid races with the suspend/hibernate code,
152 * they should always be called with pm_mutex held (gfp_allowed_mask also should
153 * only be modified with pm_mutex held, unless the suspend/hibernate code is
154 * guaranteed not to run in parallel with that modification).
157 static gfp_t saved_gfp_mask;
159 void pm_restore_gfp_mask(void)
161 WARN_ON(!mutex_is_locked(&pm_mutex));
162 if (saved_gfp_mask) {
163 gfp_allowed_mask = saved_gfp_mask;
168 void pm_restrict_gfp_mask(void)
170 WARN_ON(!mutex_is_locked(&pm_mutex));
171 WARN_ON(saved_gfp_mask);
172 saved_gfp_mask = gfp_allowed_mask;
173 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
176 bool pm_suspended_storage(void)
178 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
182 #endif /* CONFIG_PM_SLEEP */
184 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
185 unsigned int pageblock_order __read_mostly;
188 static void __free_pages_ok(struct page *page, unsigned int order);
191 * results with 256, 32 in the lowmem_reserve sysctl:
192 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
193 * 1G machine -> (16M dma, 784M normal, 224M high)
194 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
195 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
196 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
198 * TBD: should special case ZONE_DMA32 machines here - in those we normally
199 * don't need any ZONE_NORMAL reservation
201 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
202 #ifdef CONFIG_ZONE_DMA
205 #ifdef CONFIG_ZONE_DMA32
208 #ifdef CONFIG_HIGHMEM
214 EXPORT_SYMBOL(totalram_pages);
216 static char * const zone_names[MAX_NR_ZONES] = {
217 #ifdef CONFIG_ZONE_DMA
220 #ifdef CONFIG_ZONE_DMA32
224 #ifdef CONFIG_HIGHMEM
228 #ifdef CONFIG_ZONE_DEVICE
233 static void free_compound_page(struct page *page);
234 compound_page_dtor * const compound_page_dtors[] = {
237 #ifdef CONFIG_HUGETLB_PAGE
242 int min_free_kbytes = 1024;
243 int user_min_free_kbytes = -1;
245 static unsigned long __meminitdata nr_kernel_pages;
246 static unsigned long __meminitdata nr_all_pages;
247 static unsigned long __meminitdata dma_reserve;
249 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
250 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
251 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
252 static unsigned long __initdata required_kernelcore;
253 static unsigned long __initdata required_movablecore;
254 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
256 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
258 EXPORT_SYMBOL(movable_zone);
259 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
262 int nr_node_ids __read_mostly = MAX_NUMNODES;
263 int nr_online_nodes __read_mostly = 1;
264 EXPORT_SYMBOL(nr_node_ids);
265 EXPORT_SYMBOL(nr_online_nodes);
268 static DEFINE_LOCAL_IRQ_LOCK(pa_lock);
270 #ifdef CONFIG_PREEMPT_RT_BASE
271 # define cpu_lock_irqsave(cpu, flags) \
272 local_lock_irqsave_on(pa_lock, flags, cpu)
273 # define cpu_unlock_irqrestore(cpu, flags) \
274 local_unlock_irqrestore_on(pa_lock, flags, cpu)
276 # define cpu_lock_irqsave(cpu, flags) local_irq_save(flags)
277 # define cpu_unlock_irqrestore(cpu, flags) local_irq_restore(flags)
280 int page_group_by_mobility_disabled __read_mostly;
282 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
283 static inline void reset_deferred_meminit(pg_data_t *pgdat)
285 pgdat->first_deferred_pfn = ULONG_MAX;
288 /* Returns true if the struct page for the pfn is uninitialised */
289 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
291 if (pfn >= NODE_DATA(early_pfn_to_nid(pfn))->first_deferred_pfn)
297 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
299 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
306 * Returns false when the remaining initialisation should be deferred until
307 * later in the boot cycle when it can be parallelised.
309 static inline bool update_defer_init(pg_data_t *pgdat,
310 unsigned long pfn, unsigned long zone_end,
311 unsigned long *nr_initialised)
313 /* Always populate low zones for address-contrained allocations */
314 if (zone_end < pgdat_end_pfn(pgdat))
317 /* Initialise at least 2G of the highest zone */
319 if (*nr_initialised > (2UL << (30 - PAGE_SHIFT)) &&
320 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
321 pgdat->first_deferred_pfn = pfn;
328 static inline void reset_deferred_meminit(pg_data_t *pgdat)
332 static inline bool early_page_uninitialised(unsigned long pfn)
337 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
342 static inline bool update_defer_init(pg_data_t *pgdat,
343 unsigned long pfn, unsigned long zone_end,
344 unsigned long *nr_initialised)
351 void set_pageblock_migratetype(struct page *page, int migratetype)
353 if (unlikely(page_group_by_mobility_disabled &&
354 migratetype < MIGRATE_PCPTYPES))
355 migratetype = MIGRATE_UNMOVABLE;
357 set_pageblock_flags_group(page, (unsigned long)migratetype,
358 PB_migrate, PB_migrate_end);
361 #ifdef CONFIG_DEBUG_VM
362 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
366 unsigned long pfn = page_to_pfn(page);
367 unsigned long sp, start_pfn;
370 seq = zone_span_seqbegin(zone);
371 start_pfn = zone->zone_start_pfn;
372 sp = zone->spanned_pages;
373 if (!zone_spans_pfn(zone, pfn))
375 } while (zone_span_seqretry(zone, seq));
378 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
379 pfn, zone_to_nid(zone), zone->name,
380 start_pfn, start_pfn + sp);
385 static int page_is_consistent(struct zone *zone, struct page *page)
387 if (!pfn_valid_within(page_to_pfn(page)))
389 if (zone != page_zone(page))
395 * Temporary debugging check for pages not lying within a given zone.
397 static int bad_range(struct zone *zone, struct page *page)
399 if (page_outside_zone_boundaries(zone, page))
401 if (!page_is_consistent(zone, page))
407 static inline int bad_range(struct zone *zone, struct page *page)
413 static void bad_page(struct page *page, const char *reason,
414 unsigned long bad_flags)
416 static unsigned long resume;
417 static unsigned long nr_shown;
418 static unsigned long nr_unshown;
420 /* Don't complain about poisoned pages */
421 if (PageHWPoison(page)) {
422 page_mapcount_reset(page); /* remove PageBuddy */
427 * Allow a burst of 60 reports, then keep quiet for that minute;
428 * or allow a steady drip of one report per second.
430 if (nr_shown == 60) {
431 if (time_before(jiffies, resume)) {
437 "BUG: Bad page state: %lu messages suppressed\n",
444 resume = jiffies + 60 * HZ;
446 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
447 current->comm, page_to_pfn(page));
448 dump_page_badflags(page, reason, bad_flags);
453 /* Leave bad fields for debug, except PageBuddy could make trouble */
454 page_mapcount_reset(page); /* remove PageBuddy */
455 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
459 * Higher-order pages are called "compound pages". They are structured thusly:
461 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
463 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
464 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
466 * The first tail page's ->compound_dtor holds the offset in array of compound
467 * page destructors. See compound_page_dtors.
469 * The first tail page's ->compound_order holds the order of allocation.
470 * This usage means that zero-order pages may not be compound.
473 static void free_compound_page(struct page *page)
475 __free_pages_ok(page, compound_order(page));
478 void prep_compound_page(struct page *page, unsigned int order)
481 int nr_pages = 1 << order;
483 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
484 set_compound_order(page, order);
486 for (i = 1; i < nr_pages; i++) {
487 struct page *p = page + i;
488 set_page_count(p, 0);
489 set_compound_head(p, page);
493 #ifdef CONFIG_DEBUG_PAGEALLOC
494 unsigned int _debug_guardpage_minorder;
495 bool _debug_pagealloc_enabled __read_mostly;
496 bool _debug_guardpage_enabled __read_mostly;
498 static int __init early_debug_pagealloc(char *buf)
503 if (strcmp(buf, "on") == 0)
504 _debug_pagealloc_enabled = true;
508 early_param("debug_pagealloc", early_debug_pagealloc);
510 static bool need_debug_guardpage(void)
512 /* If we don't use debug_pagealloc, we don't need guard page */
513 if (!debug_pagealloc_enabled())
519 static void init_debug_guardpage(void)
521 if (!debug_pagealloc_enabled())
524 _debug_guardpage_enabled = true;
527 struct page_ext_operations debug_guardpage_ops = {
528 .need = need_debug_guardpage,
529 .init = init_debug_guardpage,
532 static int __init debug_guardpage_minorder_setup(char *buf)
536 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
537 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
540 _debug_guardpage_minorder = res;
541 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
544 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
546 static inline void set_page_guard(struct zone *zone, struct page *page,
547 unsigned int order, int migratetype)
549 struct page_ext *page_ext;
551 if (!debug_guardpage_enabled())
554 page_ext = lookup_page_ext(page);
555 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
557 INIT_LIST_HEAD(&page->lru);
558 set_page_private(page, order);
559 /* Guard pages are not available for any usage */
560 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
563 static inline void clear_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 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
574 set_page_private(page, 0);
575 if (!is_migrate_isolate(migratetype))
576 __mod_zone_freepage_state(zone, (1 << order), migratetype);
579 struct page_ext_operations debug_guardpage_ops = { NULL, };
580 static inline void set_page_guard(struct zone *zone, struct page *page,
581 unsigned int order, int migratetype) {}
582 static inline void clear_page_guard(struct zone *zone, struct page *page,
583 unsigned int order, int migratetype) {}
586 static inline void set_page_order(struct page *page, unsigned int order)
588 set_page_private(page, order);
589 __SetPageBuddy(page);
592 static inline void rmv_page_order(struct page *page)
594 __ClearPageBuddy(page);
595 set_page_private(page, 0);
599 * This function checks whether a page is free && is the buddy
600 * we can do coalesce a page and its buddy if
601 * (a) the buddy is not in a hole &&
602 * (b) the buddy is in the buddy system &&
603 * (c) a page and its buddy have the same order &&
604 * (d) a page and its buddy are in the same zone.
606 * For recording whether a page is in the buddy system, we set ->_mapcount
607 * PAGE_BUDDY_MAPCOUNT_VALUE.
608 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
609 * serialized by zone->lock.
611 * For recording page's order, we use page_private(page).
613 static inline int page_is_buddy(struct page *page, struct page *buddy,
616 if (!pfn_valid_within(page_to_pfn(buddy)))
619 if (page_is_guard(buddy) && page_order(buddy) == order) {
620 if (page_zone_id(page) != page_zone_id(buddy))
623 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
628 if (PageBuddy(buddy) && page_order(buddy) == order) {
630 * zone check is done late to avoid uselessly
631 * calculating zone/node ids for pages that could
634 if (page_zone_id(page) != page_zone_id(buddy))
637 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
645 * Freeing function for a buddy system allocator.
647 * The concept of a buddy system is to maintain direct-mapped table
648 * (containing bit values) for memory blocks of various "orders".
649 * The bottom level table contains the map for the smallest allocatable
650 * units of memory (here, pages), and each level above it describes
651 * pairs of units from the levels below, hence, "buddies".
652 * At a high level, all that happens here is marking the table entry
653 * at the bottom level available, and propagating the changes upward
654 * as necessary, plus some accounting needed to play nicely with other
655 * parts of the VM system.
656 * At each level, we keep a list of pages, which are heads of continuous
657 * free pages of length of (1 << order) and marked with _mapcount
658 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
660 * So when we are allocating or freeing one, we can derive the state of the
661 * other. That is, if we allocate a small block, and both were
662 * free, the remainder of the region must be split into blocks.
663 * If a block is freed, and its buddy is also free, then this
664 * triggers coalescing into a block of larger size.
669 static inline void __free_one_page(struct page *page,
671 struct zone *zone, unsigned int order,
674 unsigned long page_idx;
675 unsigned long combined_idx;
676 unsigned long uninitialized_var(buddy_idx);
678 unsigned int max_order = MAX_ORDER;
680 VM_BUG_ON(!zone_is_initialized(zone));
681 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
683 VM_BUG_ON(migratetype == -1);
684 if (is_migrate_isolate(migratetype)) {
686 * We restrict max order of merging to prevent merge
687 * between freepages on isolate pageblock and normal
688 * pageblock. Without this, pageblock isolation
689 * could cause incorrect freepage accounting.
691 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
693 __mod_zone_freepage_state(zone, 1 << order, migratetype);
696 page_idx = pfn & ((1 << max_order) - 1);
698 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
699 VM_BUG_ON_PAGE(bad_range(zone, page), page);
701 while (order < max_order - 1) {
702 buddy_idx = __find_buddy_index(page_idx, order);
703 buddy = page + (buddy_idx - page_idx);
704 if (!page_is_buddy(page, buddy, order))
707 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
708 * merge with it and move up one order.
710 if (page_is_guard(buddy)) {
711 clear_page_guard(zone, buddy, order, migratetype);
713 list_del(&buddy->lru);
714 zone->free_area[order].nr_free--;
715 rmv_page_order(buddy);
717 combined_idx = buddy_idx & page_idx;
718 page = page + (combined_idx - page_idx);
719 page_idx = combined_idx;
722 set_page_order(page, order);
725 * If this is not the largest possible page, check if the buddy
726 * of the next-highest order is free. If it is, it's possible
727 * that pages are being freed that will coalesce soon. In case,
728 * that is happening, add the free page to the tail of the list
729 * so it's less likely to be used soon and more likely to be merged
730 * as a higher order page
732 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
733 struct page *higher_page, *higher_buddy;
734 combined_idx = buddy_idx & page_idx;
735 higher_page = page + (combined_idx - page_idx);
736 buddy_idx = __find_buddy_index(combined_idx, order + 1);
737 higher_buddy = higher_page + (buddy_idx - combined_idx);
738 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
739 list_add_tail(&page->lru,
740 &zone->free_area[order].free_list[migratetype]);
745 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
747 zone->free_area[order].nr_free++;
750 static inline int free_pages_check(struct page *page)
752 const char *bad_reason = NULL;
753 unsigned long bad_flags = 0;
755 if (unlikely(page_mapcount(page)))
756 bad_reason = "nonzero mapcount";
757 if (unlikely(page->mapping != NULL))
758 bad_reason = "non-NULL mapping";
759 if (unlikely(atomic_read(&page->_count) != 0))
760 bad_reason = "nonzero _count";
761 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
762 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
763 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
766 if (unlikely(page->mem_cgroup))
767 bad_reason = "page still charged to cgroup";
769 if (unlikely(bad_reason)) {
770 bad_page(page, bad_reason, bad_flags);
773 page_cpupid_reset_last(page);
774 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
775 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
780 * Frees a number of pages which have been collected from the pcp lists.
781 * Assumes all pages on list are in same zone, and of same order.
782 * count is the number of pages to free.
784 * If the zone was previously in an "all pages pinned" state then look to
785 * see if this freeing clears that state.
787 * And clear the zone's pages_scanned counter, to hold off the "all pages are
788 * pinned" detection logic.
790 static void free_pcppages_bulk(struct zone *zone, int count,
791 struct list_head *list)
794 unsigned long nr_scanned;
797 spin_lock_irqsave(&zone->lock, flags);
799 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
801 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
803 while (!list_empty(list)) {
804 struct page *page = list_first_entry(list, struct page, lru);
805 int mt; /* migratetype of the to-be-freed page */
807 /* must delete as __free_one_page list manipulates */
808 list_del(&page->lru);
810 mt = get_pcppage_migratetype(page);
811 /* MIGRATE_ISOLATE page should not go to pcplists */
812 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
813 /* Pageblock could have been isolated meanwhile */
814 if (unlikely(has_isolate_pageblock(zone)))
815 mt = get_pageblock_migratetype(page);
817 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
818 trace_mm_page_pcpu_drain(page, 0, mt);
821 WARN_ON(to_free != 0);
822 spin_unlock_irqrestore(&zone->lock, flags);
826 * Moves a number of pages from the PCP lists to free list which
827 * is freed outside of the locked region.
829 * Assumes all pages on list are in same zone, and of same order.
830 * count is the number of pages to free.
832 static void isolate_pcp_pages(int to_free, struct per_cpu_pages *src,
833 struct list_head *dst)
840 struct list_head *list;
843 * Remove pages from lists in a round-robin fashion. A
844 * batch_free count is maintained that is incremented when an
845 * empty list is encountered. This is so more pages are freed
846 * off fuller lists instead of spinning excessively around empty
851 if (++migratetype == MIGRATE_PCPTYPES)
853 list = &src->lists[migratetype];
854 } while (list_empty(list));
856 /* This is the only non-empty list. Free them all. */
857 if (batch_free == MIGRATE_PCPTYPES)
858 batch_free = to_free;
861 page = list_last_entry(list, struct page, lru);
862 list_del(&page->lru);
864 list_add(&page->lru, dst);
865 } while (--to_free && --batch_free && !list_empty(list));
869 static void free_one_page(struct zone *zone,
870 struct page *page, unsigned long pfn,
874 unsigned long nr_scanned;
877 spin_lock_irqsave(&zone->lock, flags);
878 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
880 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
882 if (unlikely(has_isolate_pageblock(zone) ||
883 is_migrate_isolate(migratetype))) {
884 migratetype = get_pfnblock_migratetype(page, pfn);
886 __free_one_page(page, pfn, zone, order, migratetype);
887 spin_unlock_irqrestore(&zone->lock, flags);
890 static int free_tail_pages_check(struct page *head_page, struct page *page)
895 * We rely page->lru.next never has bit 0 set, unless the page
896 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
898 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
900 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
904 if (unlikely(!PageTail(page))) {
905 bad_page(page, "PageTail not set", 0);
908 if (unlikely(compound_head(page) != head_page)) {
909 bad_page(page, "compound_head not consistent", 0);
914 clear_compound_head(page);
918 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
919 unsigned long zone, int nid)
921 set_page_links(page, zone, nid, pfn);
922 init_page_count(page);
923 page_mapcount_reset(page);
924 page_cpupid_reset_last(page);
926 INIT_LIST_HEAD(&page->lru);
927 #ifdef WANT_PAGE_VIRTUAL
928 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
929 if (!is_highmem_idx(zone))
930 set_page_address(page, __va(pfn << PAGE_SHIFT));
934 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
937 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
940 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
941 static void init_reserved_page(unsigned long pfn)
946 if (!early_page_uninitialised(pfn))
949 nid = early_pfn_to_nid(pfn);
950 pgdat = NODE_DATA(nid);
952 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
953 struct zone *zone = &pgdat->node_zones[zid];
955 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
958 __init_single_pfn(pfn, zid, nid);
961 static inline void init_reserved_page(unsigned long pfn)
964 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
967 * Initialised pages do not have PageReserved set. This function is
968 * called for each range allocated by the bootmem allocator and
969 * marks the pages PageReserved. The remaining valid pages are later
970 * sent to the buddy page allocator.
972 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
974 unsigned long start_pfn = PFN_DOWN(start);
975 unsigned long end_pfn = PFN_UP(end);
977 for (; start_pfn < end_pfn; start_pfn++) {
978 if (pfn_valid(start_pfn)) {
979 struct page *page = pfn_to_page(start_pfn);
981 init_reserved_page(start_pfn);
983 /* Avoid false-positive PageTail() */
984 INIT_LIST_HEAD(&page->lru);
986 SetPageReserved(page);
991 static bool free_pages_prepare(struct page *page, unsigned int order)
993 bool compound = PageCompound(page);
996 VM_BUG_ON_PAGE(PageTail(page), page);
997 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
999 trace_mm_page_free(page, order);
1000 kmemcheck_free_shadow(page, order);
1001 kasan_free_pages(page, order);
1004 page->mapping = NULL;
1005 bad += free_pages_check(page);
1006 for (i = 1; i < (1 << order); i++) {
1008 bad += free_tail_pages_check(page, page + i);
1009 bad += free_pages_check(page + i);
1014 reset_page_owner(page, order);
1016 if (!PageHighMem(page)) {
1017 debug_check_no_locks_freed(page_address(page),
1018 PAGE_SIZE << order);
1019 debug_check_no_obj_freed(page_address(page),
1020 PAGE_SIZE << order);
1022 arch_free_page(page, order);
1023 kernel_map_pages(page, 1 << order, 0);
1028 static void __free_pages_ok(struct page *page, unsigned int order)
1030 unsigned long flags;
1032 unsigned long pfn = page_to_pfn(page);
1034 if (!free_pages_prepare(page, order))
1037 migratetype = get_pfnblock_migratetype(page, pfn);
1038 local_lock_irqsave(pa_lock, flags);
1039 __count_vm_events(PGFREE, 1 << order);
1040 free_one_page(page_zone(page), page, pfn, order, migratetype);
1041 local_unlock_irqrestore(pa_lock, flags);
1044 static void __init __free_pages_boot_core(struct page *page,
1045 unsigned long pfn, unsigned int order)
1047 unsigned int nr_pages = 1 << order;
1048 struct page *p = page;
1052 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1054 __ClearPageReserved(p);
1055 set_page_count(p, 0);
1057 __ClearPageReserved(p);
1058 set_page_count(p, 0);
1060 page_zone(page)->managed_pages += nr_pages;
1061 set_page_refcounted(page);
1062 __free_pages(page, order);
1065 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1066 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1068 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1070 int __meminit early_pfn_to_nid(unsigned long pfn)
1072 static DEFINE_SPINLOCK(early_pfn_lock);
1075 spin_lock(&early_pfn_lock);
1076 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1079 spin_unlock(&early_pfn_lock);
1085 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1086 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1087 struct mminit_pfnnid_cache *state)
1091 nid = __early_pfn_to_nid(pfn, state);
1092 if (nid >= 0 && nid != node)
1097 /* Only safe to use early in boot when initialisation is single-threaded */
1098 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1100 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1105 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1109 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1110 struct mminit_pfnnid_cache *state)
1117 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1120 if (early_page_uninitialised(pfn))
1122 return __free_pages_boot_core(page, pfn, order);
1125 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1126 static void __init deferred_free_range(struct page *page,
1127 unsigned long pfn, int nr_pages)
1134 /* Free a large naturally-aligned chunk if possible */
1135 if (nr_pages == MAX_ORDER_NR_PAGES &&
1136 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1137 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1138 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1142 for (i = 0; i < nr_pages; i++, page++, pfn++)
1143 __free_pages_boot_core(page, pfn, 0);
1146 /* Completion tracking for deferred_init_memmap() threads */
1147 static atomic_t pgdat_init_n_undone __initdata;
1148 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1150 static inline void __init pgdat_init_report_one_done(void)
1152 if (atomic_dec_and_test(&pgdat_init_n_undone))
1153 complete(&pgdat_init_all_done_comp);
1156 /* Initialise remaining memory on a node */
1157 static int __init deferred_init_memmap(void *data)
1159 pg_data_t *pgdat = data;
1160 int nid = pgdat->node_id;
1161 struct mminit_pfnnid_cache nid_init_state = { };
1162 unsigned long start = jiffies;
1163 unsigned long nr_pages = 0;
1164 unsigned long walk_start, walk_end;
1167 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1168 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1170 if (first_init_pfn == ULONG_MAX) {
1171 pgdat_init_report_one_done();
1175 /* Bind memory initialisation thread to a local node if possible */
1176 if (!cpumask_empty(cpumask))
1177 set_cpus_allowed_ptr(current, cpumask);
1179 /* Sanity check boundaries */
1180 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1181 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1182 pgdat->first_deferred_pfn = ULONG_MAX;
1184 /* Only the highest zone is deferred so find it */
1185 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1186 zone = pgdat->node_zones + zid;
1187 if (first_init_pfn < zone_end_pfn(zone))
1191 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1192 unsigned long pfn, end_pfn;
1193 struct page *page = NULL;
1194 struct page *free_base_page = NULL;
1195 unsigned long free_base_pfn = 0;
1198 end_pfn = min(walk_end, zone_end_pfn(zone));
1199 pfn = first_init_pfn;
1200 if (pfn < walk_start)
1202 if (pfn < zone->zone_start_pfn)
1203 pfn = zone->zone_start_pfn;
1205 for (; pfn < end_pfn; pfn++) {
1206 if (!pfn_valid_within(pfn))
1210 * Ensure pfn_valid is checked every
1211 * MAX_ORDER_NR_PAGES for memory holes
1213 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1214 if (!pfn_valid(pfn)) {
1220 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1225 /* Minimise pfn page lookups and scheduler checks */
1226 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1229 nr_pages += nr_to_free;
1230 deferred_free_range(free_base_page,
1231 free_base_pfn, nr_to_free);
1232 free_base_page = NULL;
1233 free_base_pfn = nr_to_free = 0;
1235 page = pfn_to_page(pfn);
1240 VM_BUG_ON(page_zone(page) != zone);
1244 __init_single_page(page, pfn, zid, nid);
1245 if (!free_base_page) {
1246 free_base_page = page;
1247 free_base_pfn = pfn;
1252 /* Where possible, batch up pages for a single free */
1255 /* Free the current block of pages to allocator */
1256 nr_pages += nr_to_free;
1257 deferred_free_range(free_base_page, free_base_pfn,
1259 free_base_page = NULL;
1260 free_base_pfn = nr_to_free = 0;
1263 first_init_pfn = max(end_pfn, first_init_pfn);
1266 /* Sanity check that the next zone really is unpopulated */
1267 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1269 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1270 jiffies_to_msecs(jiffies - start));
1272 pgdat_init_report_one_done();
1276 void __init page_alloc_init_late(void)
1280 /* There will be num_node_state(N_MEMORY) threads */
1281 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1282 for_each_node_state(nid, N_MEMORY) {
1283 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1286 /* Block until all are initialised */
1287 wait_for_completion(&pgdat_init_all_done_comp);
1289 /* Reinit limits that are based on free pages after the kernel is up */
1290 files_maxfiles_init();
1292 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1295 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1296 void __init init_cma_reserved_pageblock(struct page *page)
1298 unsigned i = pageblock_nr_pages;
1299 struct page *p = page;
1302 __ClearPageReserved(p);
1303 set_page_count(p, 0);
1306 set_pageblock_migratetype(page, MIGRATE_CMA);
1308 if (pageblock_order >= MAX_ORDER) {
1309 i = pageblock_nr_pages;
1312 set_page_refcounted(p);
1313 __free_pages(p, MAX_ORDER - 1);
1314 p += MAX_ORDER_NR_PAGES;
1315 } while (i -= MAX_ORDER_NR_PAGES);
1317 set_page_refcounted(page);
1318 __free_pages(page, pageblock_order);
1321 adjust_managed_page_count(page, pageblock_nr_pages);
1326 * The order of subdivision here is critical for the IO subsystem.
1327 * Please do not alter this order without good reasons and regression
1328 * testing. Specifically, as large blocks of memory are subdivided,
1329 * the order in which smaller blocks are delivered depends on the order
1330 * they're subdivided in this function. This is the primary factor
1331 * influencing the order in which pages are delivered to the IO
1332 * subsystem according to empirical testing, and this is also justified
1333 * by considering the behavior of a buddy system containing a single
1334 * large block of memory acted on by a series of small allocations.
1335 * This behavior is a critical factor in sglist merging's success.
1339 static inline void expand(struct zone *zone, struct page *page,
1340 int low, int high, struct free_area *area,
1343 unsigned long size = 1 << high;
1345 while (high > low) {
1349 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1351 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1352 debug_guardpage_enabled() &&
1353 high < debug_guardpage_minorder()) {
1355 * Mark as guard pages (or page), that will allow to
1356 * merge back to allocator when buddy will be freed.
1357 * Corresponding page table entries will not be touched,
1358 * pages will stay not present in virtual address space
1360 set_page_guard(zone, &page[size], high, migratetype);
1363 list_add(&page[size].lru, &area->free_list[migratetype]);
1365 set_page_order(&page[size], high);
1370 * This page is about to be returned from the page allocator
1372 static inline int check_new_page(struct page *page)
1374 const char *bad_reason = NULL;
1375 unsigned long bad_flags = 0;
1377 if (unlikely(page_mapcount(page)))
1378 bad_reason = "nonzero mapcount";
1379 if (unlikely(page->mapping != NULL))
1380 bad_reason = "non-NULL mapping";
1381 if (unlikely(atomic_read(&page->_count) != 0))
1382 bad_reason = "nonzero _count";
1383 if (unlikely(page->flags & __PG_HWPOISON)) {
1384 bad_reason = "HWPoisoned (hardware-corrupted)";
1385 bad_flags = __PG_HWPOISON;
1387 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1388 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1389 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1392 if (unlikely(page->mem_cgroup))
1393 bad_reason = "page still charged to cgroup";
1395 if (unlikely(bad_reason)) {
1396 bad_page(page, bad_reason, bad_flags);
1402 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1407 for (i = 0; i < (1 << order); i++) {
1408 struct page *p = page + i;
1409 if (unlikely(check_new_page(p)))
1413 set_page_private(page, 0);
1414 set_page_refcounted(page);
1416 arch_alloc_page(page, order);
1417 kernel_map_pages(page, 1 << order, 1);
1418 kasan_alloc_pages(page, order);
1420 if (gfp_flags & __GFP_ZERO)
1421 for (i = 0; i < (1 << order); i++)
1422 clear_highpage(page + i);
1424 if (order && (gfp_flags & __GFP_COMP))
1425 prep_compound_page(page, order);
1427 set_page_owner(page, order, gfp_flags);
1430 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1431 * allocate the page. The expectation is that the caller is taking
1432 * steps that will free more memory. The caller should avoid the page
1433 * being used for !PFMEMALLOC purposes.
1435 if (alloc_flags & ALLOC_NO_WATERMARKS)
1436 set_page_pfmemalloc(page);
1438 clear_page_pfmemalloc(page);
1444 * Go through the free lists for the given migratetype and remove
1445 * the smallest available page from the freelists
1448 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1451 unsigned int current_order;
1452 struct free_area *area;
1455 /* Find a page of the appropriate size in the preferred list */
1456 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1457 area = &(zone->free_area[current_order]);
1458 if (list_empty(&area->free_list[migratetype]))
1461 page = list_entry(area->free_list[migratetype].next,
1463 list_del(&page->lru);
1464 rmv_page_order(page);
1466 expand(zone, page, order, current_order, area, migratetype);
1467 set_pcppage_migratetype(page, migratetype);
1476 * This array describes the order lists are fallen back to when
1477 * the free lists for the desirable migrate type are depleted
1479 static int fallbacks[MIGRATE_TYPES][4] = {
1480 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1481 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1482 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1484 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1486 #ifdef CONFIG_MEMORY_ISOLATION
1487 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1492 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1495 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1498 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1499 unsigned int order) { return NULL; }
1503 * Move the free pages in a range to the free lists of the requested type.
1504 * Note that start_page and end_pages are not aligned on a pageblock
1505 * boundary. If alignment is required, use move_freepages_block()
1507 int move_freepages(struct zone *zone,
1508 struct page *start_page, struct page *end_page,
1513 int pages_moved = 0;
1515 #ifndef CONFIG_HOLES_IN_ZONE
1517 * page_zone is not safe to call in this context when
1518 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1519 * anyway as we check zone boundaries in move_freepages_block().
1520 * Remove at a later date when no bug reports exist related to
1521 * grouping pages by mobility
1523 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1526 for (page = start_page; page <= end_page;) {
1527 /* Make sure we are not inadvertently changing nodes */
1528 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1530 if (!pfn_valid_within(page_to_pfn(page))) {
1535 if (!PageBuddy(page)) {
1540 order = page_order(page);
1541 list_move(&page->lru,
1542 &zone->free_area[order].free_list[migratetype]);
1544 pages_moved += 1 << order;
1550 int move_freepages_block(struct zone *zone, struct page *page,
1553 unsigned long start_pfn, end_pfn;
1554 struct page *start_page, *end_page;
1556 start_pfn = page_to_pfn(page);
1557 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1558 start_page = pfn_to_page(start_pfn);
1559 end_page = start_page + pageblock_nr_pages - 1;
1560 end_pfn = start_pfn + pageblock_nr_pages - 1;
1562 /* Do not cross zone boundaries */
1563 if (!zone_spans_pfn(zone, start_pfn))
1565 if (!zone_spans_pfn(zone, end_pfn))
1568 return move_freepages(zone, start_page, end_page, migratetype);
1571 static void change_pageblock_range(struct page *pageblock_page,
1572 int start_order, int migratetype)
1574 int nr_pageblocks = 1 << (start_order - pageblock_order);
1576 while (nr_pageblocks--) {
1577 set_pageblock_migratetype(pageblock_page, migratetype);
1578 pageblock_page += pageblock_nr_pages;
1583 * When we are falling back to another migratetype during allocation, try to
1584 * steal extra free pages from the same pageblocks to satisfy further
1585 * allocations, instead of polluting multiple pageblocks.
1587 * If we are stealing a relatively large buddy page, it is likely there will
1588 * be more free pages in the pageblock, so try to steal them all. For
1589 * reclaimable and unmovable allocations, we steal regardless of page size,
1590 * as fragmentation caused by those allocations polluting movable pageblocks
1591 * is worse than movable allocations stealing from unmovable and reclaimable
1594 static bool can_steal_fallback(unsigned int order, int start_mt)
1597 * Leaving this order check is intended, although there is
1598 * relaxed order check in next check. The reason is that
1599 * we can actually steal whole pageblock if this condition met,
1600 * but, below check doesn't guarantee it and that is just heuristic
1601 * so could be changed anytime.
1603 if (order >= pageblock_order)
1606 if (order >= pageblock_order / 2 ||
1607 start_mt == MIGRATE_RECLAIMABLE ||
1608 start_mt == MIGRATE_UNMOVABLE ||
1609 page_group_by_mobility_disabled)
1616 * This function implements actual steal behaviour. If order is large enough,
1617 * we can steal whole pageblock. If not, we first move freepages in this
1618 * pageblock and check whether half of pages are moved or not. If half of
1619 * pages are moved, we can change migratetype of pageblock and permanently
1620 * use it's pages as requested migratetype in the future.
1622 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1625 unsigned int current_order = page_order(page);
1628 /* Take ownership for orders >= pageblock_order */
1629 if (current_order >= pageblock_order) {
1630 change_pageblock_range(page, current_order, start_type);
1634 pages = move_freepages_block(zone, page, start_type);
1636 /* Claim the whole block if over half of it is free */
1637 if (pages >= (1 << (pageblock_order-1)) ||
1638 page_group_by_mobility_disabled)
1639 set_pageblock_migratetype(page, start_type);
1643 * Check whether there is a suitable fallback freepage with requested order.
1644 * If only_stealable is true, this function returns fallback_mt only if
1645 * we can steal other freepages all together. This would help to reduce
1646 * fragmentation due to mixed migratetype pages in one pageblock.
1648 int find_suitable_fallback(struct free_area *area, unsigned int order,
1649 int migratetype, bool only_stealable, bool *can_steal)
1654 if (area->nr_free == 0)
1659 fallback_mt = fallbacks[migratetype][i];
1660 if (fallback_mt == MIGRATE_TYPES)
1663 if (list_empty(&area->free_list[fallback_mt]))
1666 if (can_steal_fallback(order, migratetype))
1669 if (!only_stealable)
1680 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1681 * there are no empty page blocks that contain a page with a suitable order
1683 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1684 unsigned int alloc_order)
1687 unsigned long max_managed, flags;
1690 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1691 * Check is race-prone but harmless.
1693 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1694 if (zone->nr_reserved_highatomic >= max_managed)
1697 spin_lock_irqsave(&zone->lock, flags);
1699 /* Recheck the nr_reserved_highatomic limit under the lock */
1700 if (zone->nr_reserved_highatomic >= max_managed)
1704 mt = get_pageblock_migratetype(page);
1705 if (mt != MIGRATE_HIGHATOMIC &&
1706 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1707 zone->nr_reserved_highatomic += pageblock_nr_pages;
1708 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1709 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1713 spin_unlock_irqrestore(&zone->lock, flags);
1717 * Used when an allocation is about to fail under memory pressure. This
1718 * potentially hurts the reliability of high-order allocations when under
1719 * intense memory pressure but failed atomic allocations should be easier
1720 * to recover from than an OOM.
1722 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1724 struct zonelist *zonelist = ac->zonelist;
1725 unsigned long flags;
1731 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1733 /* Preserve at least one pageblock */
1734 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1737 spin_lock_irqsave(&zone->lock, flags);
1738 for (order = 0; order < MAX_ORDER; order++) {
1739 struct free_area *area = &(zone->free_area[order]);
1741 if (list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
1744 page = list_entry(area->free_list[MIGRATE_HIGHATOMIC].next,
1748 * It should never happen but changes to locking could
1749 * inadvertently allow a per-cpu drain to add pages
1750 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1751 * and watch for underflows.
1753 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1754 zone->nr_reserved_highatomic);
1757 * Convert to ac->migratetype and avoid the normal
1758 * pageblock stealing heuristics. Minimally, the caller
1759 * is doing the work and needs the pages. More
1760 * importantly, if the block was always converted to
1761 * MIGRATE_UNMOVABLE or another type then the number
1762 * of pageblocks that cannot be completely freed
1765 set_pageblock_migratetype(page, ac->migratetype);
1766 move_freepages_block(zone, page, ac->migratetype);
1767 spin_unlock_irqrestore(&zone->lock, flags);
1770 spin_unlock_irqrestore(&zone->lock, flags);
1774 /* Remove an element from the buddy allocator from the fallback list */
1775 static inline struct page *
1776 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1778 struct free_area *area;
1779 unsigned int current_order;
1784 /* Find the largest possible block of pages in the other list */
1785 for (current_order = MAX_ORDER-1;
1786 current_order >= order && current_order <= MAX_ORDER-1;
1788 area = &(zone->free_area[current_order]);
1789 fallback_mt = find_suitable_fallback(area, current_order,
1790 start_migratetype, false, &can_steal);
1791 if (fallback_mt == -1)
1794 page = list_entry(area->free_list[fallback_mt].next,
1797 steal_suitable_fallback(zone, page, start_migratetype);
1799 /* Remove the page from the freelists */
1801 list_del(&page->lru);
1802 rmv_page_order(page);
1804 expand(zone, page, order, current_order, area,
1807 * The pcppage_migratetype may differ from pageblock's
1808 * migratetype depending on the decisions in
1809 * find_suitable_fallback(). This is OK as long as it does not
1810 * differ for MIGRATE_CMA pageblocks. Those can be used as
1811 * fallback only via special __rmqueue_cma_fallback() function
1813 set_pcppage_migratetype(page, start_migratetype);
1815 trace_mm_page_alloc_extfrag(page, order, current_order,
1816 start_migratetype, fallback_mt);
1825 * Do the hard work of removing an element from the buddy allocator.
1826 * Call me with the zone->lock already held.
1828 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1829 int migratetype, gfp_t gfp_flags)
1833 page = __rmqueue_smallest(zone, order, migratetype);
1834 if (unlikely(!page)) {
1835 if (migratetype == MIGRATE_MOVABLE)
1836 page = __rmqueue_cma_fallback(zone, order);
1839 page = __rmqueue_fallback(zone, order, migratetype);
1842 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1847 * Obtain a specified number of elements from the buddy allocator, all under
1848 * a single hold of the lock, for efficiency. Add them to the supplied list.
1849 * Returns the number of new pages which were placed at *list.
1851 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1852 unsigned long count, struct list_head *list,
1853 int migratetype, bool cold)
1857 spin_lock(&zone->lock);
1858 for (i = 0; i < count; ++i) {
1859 struct page *page = __rmqueue(zone, order, migratetype, 0);
1860 if (unlikely(page == NULL))
1864 * Split buddy pages returned by expand() are received here
1865 * in physical page order. The page is added to the callers and
1866 * list and the list head then moves forward. From the callers
1867 * perspective, the linked list is ordered by page number in
1868 * some conditions. This is useful for IO devices that can
1869 * merge IO requests if the physical pages are ordered
1873 list_add(&page->lru, list);
1875 list_add_tail(&page->lru, list);
1877 if (is_migrate_cma(get_pcppage_migratetype(page)))
1878 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1881 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1882 spin_unlock(&zone->lock);
1888 * Called from the vmstat counter updater to drain pagesets of this
1889 * currently executing processor on remote nodes after they have
1892 * Note that this function must be called with the thread pinned to
1893 * a single processor.
1895 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1897 unsigned long flags;
1899 int to_drain, batch;
1901 local_lock_irqsave(pa_lock, flags);
1902 batch = READ_ONCE(pcp->batch);
1903 to_drain = min(pcp->count, batch);
1905 isolate_pcp_pages(to_drain, pcp, &dst);
1906 pcp->count -= to_drain;
1908 local_unlock_irqrestore(pa_lock, flags);
1909 free_pcppages_bulk(zone, to_drain, &dst);
1914 * Drain pcplists of the indicated processor and zone.
1916 * The processor must either be the current processor and the
1917 * thread pinned to the current processor or a processor that
1920 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1922 unsigned long flags;
1923 struct per_cpu_pageset *pset;
1924 struct per_cpu_pages *pcp;
1928 cpu_lock_irqsave(cpu, flags);
1929 pset = per_cpu_ptr(zone->pageset, cpu);
1934 isolate_pcp_pages(count, pcp, &dst);
1937 cpu_unlock_irqrestore(cpu, flags);
1939 free_pcppages_bulk(zone, count, &dst);
1943 * Drain pcplists of all zones on the indicated processor.
1945 * The processor must either be the current processor and the
1946 * thread pinned to the current processor or a processor that
1949 static void drain_pages(unsigned int cpu)
1953 for_each_populated_zone(zone) {
1954 drain_pages_zone(cpu, zone);
1959 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1961 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1962 * the single zone's pages.
1964 void drain_local_pages(struct zone *zone)
1966 int cpu = smp_processor_id();
1969 drain_pages_zone(cpu, zone);
1975 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1977 * When zone parameter is non-NULL, spill just the single zone's pages.
1979 * Note that this code is protected against sending an IPI to an offline
1980 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1981 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1982 * nothing keeps CPUs from showing up after we populated the cpumask and
1983 * before the call to on_each_cpu_mask().
1985 void drain_all_pages(struct zone *zone)
1990 * Allocate in the BSS so we wont require allocation in
1991 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1993 static cpumask_t cpus_with_pcps;
1996 * We don't care about racing with CPU hotplug event
1997 * as offline notification will cause the notified
1998 * cpu to drain that CPU pcps and on_each_cpu_mask
1999 * disables preemption as part of its processing
2001 for_each_online_cpu(cpu) {
2002 struct per_cpu_pageset *pcp;
2004 bool has_pcps = false;
2007 pcp = per_cpu_ptr(zone->pageset, cpu);
2011 for_each_populated_zone(z) {
2012 pcp = per_cpu_ptr(z->pageset, cpu);
2013 if (pcp->pcp.count) {
2021 cpumask_set_cpu(cpu, &cpus_with_pcps);
2023 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2025 #ifndef CONFIG_PREEMPT_RT_BASE
2026 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2029 for_each_cpu(cpu, &cpus_with_pcps) {
2031 drain_pages_zone(cpu, zone);
2038 #ifdef CONFIG_HIBERNATION
2040 void mark_free_pages(struct zone *zone)
2042 unsigned long pfn, max_zone_pfn;
2043 unsigned long flags;
2044 unsigned int order, t;
2045 struct list_head *curr;
2047 if (zone_is_empty(zone))
2050 spin_lock_irqsave(&zone->lock, flags);
2052 max_zone_pfn = zone_end_pfn(zone);
2053 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2054 if (pfn_valid(pfn)) {
2055 struct page *page = pfn_to_page(pfn);
2057 if (!swsusp_page_is_forbidden(page))
2058 swsusp_unset_page_free(page);
2061 for_each_migratetype_order(order, t) {
2062 list_for_each(curr, &zone->free_area[order].free_list[t]) {
2065 pfn = page_to_pfn(list_entry(curr, struct page, lru));
2066 for (i = 0; i < (1UL << order); i++)
2067 swsusp_set_page_free(pfn_to_page(pfn + i));
2070 spin_unlock_irqrestore(&zone->lock, flags);
2072 #endif /* CONFIG_PM */
2075 * Free a 0-order page
2076 * cold == true ? free a cold page : free a hot page
2078 void free_hot_cold_page(struct page *page, bool cold)
2080 struct zone *zone = page_zone(page);
2081 struct per_cpu_pages *pcp;
2082 unsigned long flags;
2083 unsigned long pfn = page_to_pfn(page);
2086 if (!free_pages_prepare(page, 0))
2089 migratetype = get_pfnblock_migratetype(page, pfn);
2090 set_pcppage_migratetype(page, migratetype);
2091 local_lock_irqsave(pa_lock, flags);
2092 __count_vm_event(PGFREE);
2095 * We only track unmovable, reclaimable and movable on pcp lists.
2096 * Free ISOLATE pages back to the allocator because they are being
2097 * offlined but treat RESERVE as movable pages so we can get those
2098 * areas back if necessary. Otherwise, we may have to free
2099 * excessively into the page allocator
2101 if (migratetype >= MIGRATE_PCPTYPES) {
2102 if (unlikely(is_migrate_isolate(migratetype))) {
2103 free_one_page(zone, page, pfn, 0, migratetype);
2106 migratetype = MIGRATE_MOVABLE;
2109 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2111 list_add(&page->lru, &pcp->lists[migratetype]);
2113 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2115 if (pcp->count >= pcp->high) {
2116 unsigned long batch = READ_ONCE(pcp->batch);
2119 isolate_pcp_pages(batch, pcp, &dst);
2120 pcp->count -= batch;
2121 local_unlock_irqrestore(pa_lock, flags);
2122 free_pcppages_bulk(zone, batch, &dst);
2127 local_unlock_irqrestore(pa_lock, flags);
2131 * Free a list of 0-order pages
2133 void free_hot_cold_page_list(struct list_head *list, bool cold)
2135 struct page *page, *next;
2137 list_for_each_entry_safe(page, next, list, lru) {
2138 trace_mm_page_free_batched(page, cold);
2139 free_hot_cold_page(page, cold);
2144 * split_page takes a non-compound higher-order page, and splits it into
2145 * n (1<<order) sub-pages: page[0..n]
2146 * Each sub-page must be freed individually.
2148 * Note: this is probably too low level an operation for use in drivers.
2149 * Please consult with lkml before using this in your driver.
2151 void split_page(struct page *page, unsigned int order)
2156 VM_BUG_ON_PAGE(PageCompound(page), page);
2157 VM_BUG_ON_PAGE(!page_count(page), page);
2159 #ifdef CONFIG_KMEMCHECK
2161 * Split shadow pages too, because free(page[0]) would
2162 * otherwise free the whole shadow.
2164 if (kmemcheck_page_is_tracked(page))
2165 split_page(virt_to_page(page[0].shadow), order);
2168 gfp_mask = get_page_owner_gfp(page);
2169 set_page_owner(page, 0, gfp_mask);
2170 for (i = 1; i < (1 << order); i++) {
2171 set_page_refcounted(page + i);
2172 set_page_owner(page + i, 0, gfp_mask);
2175 EXPORT_SYMBOL_GPL(split_page);
2177 int __isolate_free_page(struct page *page, unsigned int order)
2179 unsigned long watermark;
2183 BUG_ON(!PageBuddy(page));
2185 zone = page_zone(page);
2186 mt = get_pageblock_migratetype(page);
2188 if (!is_migrate_isolate(mt)) {
2189 /* Obey watermarks as if the page was being allocated */
2190 watermark = low_wmark_pages(zone) + (1 << order);
2191 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2194 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2197 /* Remove page from free list */
2198 list_del(&page->lru);
2199 zone->free_area[order].nr_free--;
2200 rmv_page_order(page);
2202 set_page_owner(page, order, __GFP_MOVABLE);
2204 /* Set the pageblock if the isolated page is at least a pageblock */
2205 if (order >= pageblock_order - 1) {
2206 struct page *endpage = page + (1 << order) - 1;
2207 for (; page < endpage; page += pageblock_nr_pages) {
2208 int mt = get_pageblock_migratetype(page);
2209 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2210 set_pageblock_migratetype(page,
2216 return 1UL << order;
2220 * Similar to split_page except the page is already free. As this is only
2221 * being used for migration, the migratetype of the block also changes.
2222 * As this is called with interrupts disabled, the caller is responsible
2223 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2226 * Note: this is probably too low level an operation for use in drivers.
2227 * Please consult with lkml before using this in your driver.
2229 int split_free_page(struct page *page)
2234 order = page_order(page);
2236 nr_pages = __isolate_free_page(page, order);
2240 /* Split into individual pages */
2241 set_page_refcounted(page);
2242 split_page(page, order);
2247 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2250 struct page *buffered_rmqueue(struct zone *preferred_zone,
2251 struct zone *zone, unsigned int order,
2252 gfp_t gfp_flags, int alloc_flags, int migratetype)
2254 unsigned long flags;
2256 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2258 if (likely(order == 0)) {
2259 struct per_cpu_pages *pcp;
2260 struct list_head *list;
2262 local_lock_irqsave(pa_lock, flags);
2263 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2264 list = &pcp->lists[migratetype];
2265 if (list_empty(list)) {
2266 pcp->count += rmqueue_bulk(zone, 0,
2269 if (unlikely(list_empty(list)))
2274 page = list_entry(list->prev, struct page, lru);
2276 page = list_entry(list->next, struct page, lru);
2278 list_del(&page->lru);
2281 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2283 * __GFP_NOFAIL is not to be used in new code.
2285 * All __GFP_NOFAIL callers should be fixed so that they
2286 * properly detect and handle allocation failures.
2288 * We most definitely don't want callers attempting to
2289 * allocate greater than order-1 page units with
2292 WARN_ON_ONCE(order > 1);
2294 local_spin_lock_irqsave(pa_lock, &zone->lock, flags);
2297 if (alloc_flags & ALLOC_HARDER) {
2298 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2300 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2303 page = __rmqueue(zone, order, migratetype, gfp_flags);
2305 spin_unlock(&zone->lock);
2308 __mod_zone_freepage_state(zone, -(1 << order),
2309 get_pcppage_migratetype(page));
2310 spin_unlock(&zone->lock);
2313 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2314 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2315 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2316 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2318 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2319 zone_statistics(preferred_zone, zone, gfp_flags);
2320 local_unlock_irqrestore(pa_lock, flags);
2322 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2326 local_unlock_irqrestore(pa_lock, flags);
2330 #ifdef CONFIG_FAIL_PAGE_ALLOC
2333 struct fault_attr attr;
2335 bool ignore_gfp_highmem;
2336 bool ignore_gfp_reclaim;
2338 } fail_page_alloc = {
2339 .attr = FAULT_ATTR_INITIALIZER,
2340 .ignore_gfp_reclaim = true,
2341 .ignore_gfp_highmem = true,
2345 static int __init setup_fail_page_alloc(char *str)
2347 return setup_fault_attr(&fail_page_alloc.attr, str);
2349 __setup("fail_page_alloc=", setup_fail_page_alloc);
2351 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2353 if (order < fail_page_alloc.min_order)
2355 if (gfp_mask & __GFP_NOFAIL)
2357 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2359 if (fail_page_alloc.ignore_gfp_reclaim &&
2360 (gfp_mask & __GFP_DIRECT_RECLAIM))
2363 return should_fail(&fail_page_alloc.attr, 1 << order);
2366 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2368 static int __init fail_page_alloc_debugfs(void)
2370 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2373 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2374 &fail_page_alloc.attr);
2376 return PTR_ERR(dir);
2378 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2379 &fail_page_alloc.ignore_gfp_reclaim))
2381 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2382 &fail_page_alloc.ignore_gfp_highmem))
2384 if (!debugfs_create_u32("min-order", mode, dir,
2385 &fail_page_alloc.min_order))
2390 debugfs_remove_recursive(dir);
2395 late_initcall(fail_page_alloc_debugfs);
2397 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2399 #else /* CONFIG_FAIL_PAGE_ALLOC */
2401 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2406 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2409 * Return true if free base pages are above 'mark'. For high-order checks it
2410 * will return true of the order-0 watermark is reached and there is at least
2411 * one free page of a suitable size. Checking now avoids taking the zone lock
2412 * to check in the allocation paths if no pages are free.
2414 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2415 unsigned long mark, int classzone_idx, int alloc_flags,
2420 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2422 /* free_pages may go negative - that's OK */
2423 free_pages -= (1 << order) - 1;
2425 if (alloc_flags & ALLOC_HIGH)
2429 * If the caller does not have rights to ALLOC_HARDER then subtract
2430 * the high-atomic reserves. This will over-estimate the size of the
2431 * atomic reserve but it avoids a search.
2433 if (likely(!alloc_harder))
2434 free_pages -= z->nr_reserved_highatomic;
2439 /* If allocation can't use CMA areas don't use free CMA pages */
2440 if (!(alloc_flags & ALLOC_CMA))
2441 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2445 * Check watermarks for an order-0 allocation request. If these
2446 * are not met, then a high-order request also cannot go ahead
2447 * even if a suitable page happened to be free.
2449 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2452 /* If this is an order-0 request then the watermark is fine */
2456 /* For a high-order request, check at least one suitable page is free */
2457 for (o = order; o < MAX_ORDER; o++) {
2458 struct free_area *area = &z->free_area[o];
2467 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2468 if (!list_empty(&area->free_list[mt]))
2473 if ((alloc_flags & ALLOC_CMA) &&
2474 !list_empty(&area->free_list[MIGRATE_CMA])) {
2482 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2483 int classzone_idx, int alloc_flags)
2485 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2486 zone_page_state(z, NR_FREE_PAGES));
2489 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2490 unsigned long mark, int classzone_idx)
2492 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2494 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2495 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2497 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2502 static bool zone_local(struct zone *local_zone, struct zone *zone)
2504 return local_zone->node == zone->node;
2507 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2509 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2512 #else /* CONFIG_NUMA */
2513 static bool zone_local(struct zone *local_zone, struct zone *zone)
2518 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2522 #endif /* CONFIG_NUMA */
2524 static void reset_alloc_batches(struct zone *preferred_zone)
2526 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2529 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2530 high_wmark_pages(zone) - low_wmark_pages(zone) -
2531 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2532 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2533 } while (zone++ != preferred_zone);
2537 * get_page_from_freelist goes through the zonelist trying to allocate
2540 static struct page *
2541 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2542 const struct alloc_context *ac)
2544 struct zonelist *zonelist = ac->zonelist;
2546 struct page *page = NULL;
2548 int nr_fair_skipped = 0;
2549 bool zonelist_rescan;
2552 zonelist_rescan = false;
2555 * Scan zonelist, looking for a zone with enough free.
2556 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2558 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2562 if (cpusets_enabled() &&
2563 (alloc_flags & ALLOC_CPUSET) &&
2564 !cpuset_zone_allowed(zone, gfp_mask))
2567 * Distribute pages in proportion to the individual
2568 * zone size to ensure fair page aging. The zone a
2569 * page was allocated in should have no effect on the
2570 * time the page has in memory before being reclaimed.
2572 if (alloc_flags & ALLOC_FAIR) {
2573 if (!zone_local(ac->preferred_zone, zone))
2575 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2581 * When allocating a page cache page for writing, we
2582 * want to get it from a zone that is within its dirty
2583 * limit, such that no single zone holds more than its
2584 * proportional share of globally allowed dirty pages.
2585 * The dirty limits take into account the zone's
2586 * lowmem reserves and high watermark so that kswapd
2587 * should be able to balance it without having to
2588 * write pages from its LRU list.
2590 * This may look like it could increase pressure on
2591 * lower zones by failing allocations in higher zones
2592 * before they are full. But the pages that do spill
2593 * over are limited as the lower zones are protected
2594 * by this very same mechanism. It should not become
2595 * a practical burden to them.
2597 * XXX: For now, allow allocations to potentially
2598 * exceed the per-zone dirty limit in the slowpath
2599 * (spread_dirty_pages unset) before going into reclaim,
2600 * which is important when on a NUMA setup the allowed
2601 * zones are together not big enough to reach the
2602 * global limit. The proper fix for these situations
2603 * will require awareness of zones in the
2604 * dirty-throttling and the flusher threads.
2606 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2609 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2610 if (!zone_watermark_ok(zone, order, mark,
2611 ac->classzone_idx, alloc_flags)) {
2614 /* Checked here to keep the fast path fast */
2615 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2616 if (alloc_flags & ALLOC_NO_WATERMARKS)
2619 if (zone_reclaim_mode == 0 ||
2620 !zone_allows_reclaim(ac->preferred_zone, zone))
2623 ret = zone_reclaim(zone, gfp_mask, order);
2625 case ZONE_RECLAIM_NOSCAN:
2628 case ZONE_RECLAIM_FULL:
2629 /* scanned but unreclaimable */
2632 /* did we reclaim enough */
2633 if (zone_watermark_ok(zone, order, mark,
2634 ac->classzone_idx, alloc_flags))
2642 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2643 gfp_mask, alloc_flags, ac->migratetype);
2645 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2649 * If this is a high-order atomic allocation then check
2650 * if the pageblock should be reserved for the future
2652 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2653 reserve_highatomic_pageblock(page, zone, order);
2660 * The first pass makes sure allocations are spread fairly within the
2661 * local node. However, the local node might have free pages left
2662 * after the fairness batches are exhausted, and remote zones haven't
2663 * even been considered yet. Try once more without fairness, and
2664 * include remote zones now, before entering the slowpath and waking
2665 * kswapd: prefer spilling to a remote zone over swapping locally.
2667 if (alloc_flags & ALLOC_FAIR) {
2668 alloc_flags &= ~ALLOC_FAIR;
2669 if (nr_fair_skipped) {
2670 zonelist_rescan = true;
2671 reset_alloc_batches(ac->preferred_zone);
2673 if (nr_online_nodes > 1)
2674 zonelist_rescan = true;
2677 if (zonelist_rescan)
2684 * Large machines with many possible nodes should not always dump per-node
2685 * meminfo in irq context.
2687 static inline bool should_suppress_show_mem(void)
2692 ret = in_interrupt();
2697 static DEFINE_RATELIMIT_STATE(nopage_rs,
2698 DEFAULT_RATELIMIT_INTERVAL,
2699 DEFAULT_RATELIMIT_BURST);
2701 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2703 unsigned int filter = SHOW_MEM_FILTER_NODES;
2705 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2706 debug_guardpage_minorder() > 0)
2710 * This documents exceptions given to allocations in certain
2711 * contexts that are allowed to allocate outside current's set
2714 if (!(gfp_mask & __GFP_NOMEMALLOC))
2715 if (test_thread_flag(TIF_MEMDIE) ||
2716 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2717 filter &= ~SHOW_MEM_FILTER_NODES;
2718 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2719 filter &= ~SHOW_MEM_FILTER_NODES;
2722 struct va_format vaf;
2725 va_start(args, fmt);
2730 pr_warn("%pV", &vaf);
2735 pr_warn("%s: page allocation failure: order:%u, mode:0x%x\n",
2736 current->comm, order, gfp_mask);
2739 if (!should_suppress_show_mem())
2743 static inline struct page *
2744 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2745 const struct alloc_context *ac, unsigned long *did_some_progress)
2747 struct oom_control oc = {
2748 .zonelist = ac->zonelist,
2749 .nodemask = ac->nodemask,
2750 .gfp_mask = gfp_mask,
2755 *did_some_progress = 0;
2758 * Acquire the oom lock. If that fails, somebody else is
2759 * making progress for us.
2761 if (!mutex_trylock(&oom_lock)) {
2762 *did_some_progress = 1;
2763 schedule_timeout_uninterruptible(1);
2768 * Go through the zonelist yet one more time, keep very high watermark
2769 * here, this is only to catch a parallel oom killing, we must fail if
2770 * we're still under heavy pressure.
2772 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2773 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2777 if (!(gfp_mask & __GFP_NOFAIL)) {
2778 /* Coredumps can quickly deplete all memory reserves */
2779 if (current->flags & PF_DUMPCORE)
2781 /* The OOM killer will not help higher order allocs */
2782 if (order > PAGE_ALLOC_COSTLY_ORDER)
2784 /* The OOM killer does not needlessly kill tasks for lowmem */
2785 if (ac->high_zoneidx < ZONE_NORMAL)
2787 /* The OOM killer does not compensate for IO-less reclaim */
2788 if (!(gfp_mask & __GFP_FS)) {
2790 * XXX: Page reclaim didn't yield anything,
2791 * and the OOM killer can't be invoked, but
2792 * keep looping as per tradition.
2794 *did_some_progress = 1;
2797 if (pm_suspended_storage())
2799 /* The OOM killer may not free memory on a specific node */
2800 if (gfp_mask & __GFP_THISNODE)
2803 /* Exhausted what can be done so it's blamo time */
2804 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2805 *did_some_progress = 1;
2807 mutex_unlock(&oom_lock);
2811 #ifdef CONFIG_COMPACTION
2812 /* Try memory compaction for high-order allocations before reclaim */
2813 static struct page *
2814 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2815 int alloc_flags, const struct alloc_context *ac,
2816 enum migrate_mode mode, int *contended_compaction,
2817 bool *deferred_compaction)
2819 unsigned long compact_result;
2825 current->flags |= PF_MEMALLOC;
2826 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2827 mode, contended_compaction);
2828 current->flags &= ~PF_MEMALLOC;
2830 switch (compact_result) {
2831 case COMPACT_DEFERRED:
2832 *deferred_compaction = true;
2834 case COMPACT_SKIPPED:
2841 * At least in one zone compaction wasn't deferred or skipped, so let's
2842 * count a compaction stall
2844 count_vm_event(COMPACTSTALL);
2846 page = get_page_from_freelist(gfp_mask, order,
2847 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2850 struct zone *zone = page_zone(page);
2852 zone->compact_blockskip_flush = false;
2853 compaction_defer_reset(zone, order, true);
2854 count_vm_event(COMPACTSUCCESS);
2859 * It's bad if compaction run occurs and fails. The most likely reason
2860 * is that pages exist, but not enough to satisfy watermarks.
2862 count_vm_event(COMPACTFAIL);
2869 static inline struct page *
2870 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2871 int alloc_flags, const struct alloc_context *ac,
2872 enum migrate_mode mode, int *contended_compaction,
2873 bool *deferred_compaction)
2877 #endif /* CONFIG_COMPACTION */
2879 /* Perform direct synchronous page reclaim */
2881 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2882 const struct alloc_context *ac)
2884 struct reclaim_state reclaim_state;
2889 /* We now go into synchronous reclaim */
2890 cpuset_memory_pressure_bump();
2891 current->flags |= PF_MEMALLOC;
2892 lockdep_set_current_reclaim_state(gfp_mask);
2893 reclaim_state.reclaimed_slab = 0;
2894 current->reclaim_state = &reclaim_state;
2896 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2899 current->reclaim_state = NULL;
2900 lockdep_clear_current_reclaim_state();
2901 current->flags &= ~PF_MEMALLOC;
2908 /* The really slow allocator path where we enter direct reclaim */
2909 static inline struct page *
2910 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2911 int alloc_flags, const struct alloc_context *ac,
2912 unsigned long *did_some_progress)
2914 struct page *page = NULL;
2915 bool drained = false;
2917 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2918 if (unlikely(!(*did_some_progress)))
2922 page = get_page_from_freelist(gfp_mask, order,
2923 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2926 * If an allocation failed after direct reclaim, it could be because
2927 * pages are pinned on the per-cpu lists or in high alloc reserves.
2928 * Shrink them them and try again
2930 if (!page && !drained) {
2931 unreserve_highatomic_pageblock(ac);
2932 drain_all_pages(NULL);
2941 * This is called in the allocator slow-path if the allocation request is of
2942 * sufficient urgency to ignore watermarks and take other desperate measures
2944 static inline struct page *
2945 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2946 const struct alloc_context *ac)
2951 page = get_page_from_freelist(gfp_mask, order,
2952 ALLOC_NO_WATERMARKS, ac);
2954 if (!page && gfp_mask & __GFP_NOFAIL)
2955 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2957 } while (!page && (gfp_mask & __GFP_NOFAIL));
2962 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2967 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2968 ac->high_zoneidx, ac->nodemask)
2969 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2973 gfp_to_alloc_flags(gfp_t gfp_mask)
2975 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2977 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2978 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2981 * The caller may dip into page reserves a bit more if the caller
2982 * cannot run direct reclaim, or if the caller has realtime scheduling
2983 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2984 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
2986 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2988 if (gfp_mask & __GFP_ATOMIC) {
2990 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2991 * if it can't schedule.
2993 if (!(gfp_mask & __GFP_NOMEMALLOC))
2994 alloc_flags |= ALLOC_HARDER;
2996 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2997 * comment for __cpuset_node_allowed().
2999 alloc_flags &= ~ALLOC_CPUSET;
3000 } else if (unlikely(rt_task(current)) && !in_interrupt())
3001 alloc_flags |= ALLOC_HARDER;
3003 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
3004 if (gfp_mask & __GFP_MEMALLOC)
3005 alloc_flags |= ALLOC_NO_WATERMARKS;
3006 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3007 alloc_flags |= ALLOC_NO_WATERMARKS;
3008 else if (!in_interrupt() &&
3009 ((current->flags & PF_MEMALLOC) ||
3010 unlikely(test_thread_flag(TIF_MEMDIE))))
3011 alloc_flags |= ALLOC_NO_WATERMARKS;
3014 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3015 alloc_flags |= ALLOC_CMA;
3020 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3022 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
3025 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
3027 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
3030 static inline struct page *
3031 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3032 struct alloc_context *ac)
3034 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3035 struct page *page = NULL;
3037 unsigned long pages_reclaimed = 0;
3038 unsigned long did_some_progress;
3039 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3040 bool deferred_compaction = false;
3041 int contended_compaction = COMPACT_CONTENDED_NONE;
3044 * In the slowpath, we sanity check order to avoid ever trying to
3045 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3046 * be using allocators in order of preference for an area that is
3049 if (order >= MAX_ORDER) {
3050 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3055 * We also sanity check to catch abuse of atomic reserves being used by
3056 * callers that are not in atomic context.
3058 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3059 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3060 gfp_mask &= ~__GFP_ATOMIC;
3063 * If this allocation cannot block and it is for a specific node, then
3064 * fail early. There's no need to wakeup kswapd or retry for a
3065 * speculative node-specific allocation.
3067 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
3071 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3072 wake_all_kswapds(order, ac);
3075 * OK, we're below the kswapd watermark and have kicked background
3076 * reclaim. Now things get more complex, so set up alloc_flags according
3077 * to how we want to proceed.
3079 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3082 * Find the true preferred zone if the allocation is unconstrained by
3085 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3086 struct zoneref *preferred_zoneref;
3087 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3088 ac->high_zoneidx, NULL, &ac->preferred_zone);
3089 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3092 /* This is the last chance, in general, before the goto nopage. */
3093 page = get_page_from_freelist(gfp_mask, order,
3094 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3098 /* Allocate without watermarks if the context allows */
3099 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3101 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3102 * the allocation is high priority and these type of
3103 * allocations are system rather than user orientated
3105 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3107 page = __alloc_pages_high_priority(gfp_mask, order, ac);
3114 /* Caller is not willing to reclaim, we can't balance anything */
3115 if (!can_direct_reclaim) {
3117 * All existing users of the deprecated __GFP_NOFAIL are
3118 * blockable, so warn of any new users that actually allow this
3119 * type of allocation to fail.
3121 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3125 /* Avoid recursion of direct reclaim */
3126 if (current->flags & PF_MEMALLOC)
3129 /* Avoid allocations with no watermarks from looping endlessly */
3130 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3134 * Try direct compaction. The first pass is asynchronous. Subsequent
3135 * attempts after direct reclaim are synchronous
3137 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3139 &contended_compaction,
3140 &deferred_compaction);
3144 /* Checks for THP-specific high-order allocations */
3145 if (is_thp_gfp_mask(gfp_mask)) {
3147 * If compaction is deferred for high-order allocations, it is
3148 * because sync compaction recently failed. If this is the case
3149 * and the caller requested a THP allocation, we do not want
3150 * to heavily disrupt the system, so we fail the allocation
3151 * instead of entering direct reclaim.
3153 if (deferred_compaction)
3157 * In all zones where compaction was attempted (and not
3158 * deferred or skipped), lock contention has been detected.
3159 * For THP allocation we do not want to disrupt the others
3160 * so we fallback to base pages instead.
3162 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3166 * If compaction was aborted due to need_resched(), we do not
3167 * want to further increase allocation latency, unless it is
3168 * khugepaged trying to collapse.
3170 if (contended_compaction == COMPACT_CONTENDED_SCHED
3171 && !(current->flags & PF_KTHREAD))
3176 * It can become very expensive to allocate transparent hugepages at
3177 * fault, so use asynchronous memory compaction for THP unless it is
3178 * khugepaged trying to collapse.
3180 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3181 migration_mode = MIGRATE_SYNC_LIGHT;
3183 /* Try direct reclaim and then allocating */
3184 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3185 &did_some_progress);
3189 /* Do not loop if specifically requested */
3190 if (gfp_mask & __GFP_NORETRY)
3193 /* Keep reclaiming pages as long as there is reasonable progress */
3194 pages_reclaimed += did_some_progress;
3195 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3196 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3197 /* Wait for some write requests to complete then retry */
3198 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3202 /* Reclaim has failed us, start killing things */
3203 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3207 /* Retry as long as the OOM killer is making progress */
3208 if (did_some_progress)
3213 * High-order allocations do not necessarily loop after
3214 * direct reclaim and reclaim/compaction depends on compaction
3215 * being called after reclaim so call directly if necessary
3217 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3219 &contended_compaction,
3220 &deferred_compaction);
3224 warn_alloc_failed(gfp_mask, order, NULL);
3230 * This is the 'heart' of the zoned buddy allocator.
3233 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3234 struct zonelist *zonelist, nodemask_t *nodemask)
3236 struct zoneref *preferred_zoneref;
3237 struct page *page = NULL;
3238 unsigned int cpuset_mems_cookie;
3239 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3240 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3241 struct alloc_context ac = {
3242 .high_zoneidx = gfp_zone(gfp_mask),
3243 .nodemask = nodemask,
3244 .migratetype = gfpflags_to_migratetype(gfp_mask),
3247 gfp_mask &= gfp_allowed_mask;
3249 lockdep_trace_alloc(gfp_mask);
3251 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3253 if (should_fail_alloc_page(gfp_mask, order))
3257 * Check the zones suitable for the gfp_mask contain at least one
3258 * valid zone. It's possible to have an empty zonelist as a result
3259 * of __GFP_THISNODE and a memoryless node
3261 if (unlikely(!zonelist->_zonerefs->zone))
3264 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3265 alloc_flags |= ALLOC_CMA;
3268 cpuset_mems_cookie = read_mems_allowed_begin();
3270 /* We set it here, as __alloc_pages_slowpath might have changed it */
3271 ac.zonelist = zonelist;
3273 /* Dirty zone balancing only done in the fast path */
3274 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3276 /* The preferred zone is used for statistics later */
3277 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3278 ac.nodemask ? : &cpuset_current_mems_allowed,
3279 &ac.preferred_zone);
3280 if (!ac.preferred_zone)
3282 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3284 /* First allocation attempt */
3285 alloc_mask = gfp_mask|__GFP_HARDWALL;
3286 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3287 if (unlikely(!page)) {
3289 * Runtime PM, block IO and its error handling path
3290 * can deadlock because I/O on the device might not
3293 alloc_mask = memalloc_noio_flags(gfp_mask);
3294 ac.spread_dirty_pages = false;
3296 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3299 if (kmemcheck_enabled && page)
3300 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3302 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3306 * When updating a task's mems_allowed, it is possible to race with
3307 * parallel threads in such a way that an allocation can fail while
3308 * the mask is being updated. If a page allocation is about to fail,
3309 * check if the cpuset changed during allocation and if so, retry.
3311 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3316 EXPORT_SYMBOL(__alloc_pages_nodemask);
3319 * Common helper functions.
3321 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3326 * __get_free_pages() returns a 32-bit address, which cannot represent
3329 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3331 page = alloc_pages(gfp_mask, order);
3334 return (unsigned long) page_address(page);
3336 EXPORT_SYMBOL(__get_free_pages);
3338 unsigned long get_zeroed_page(gfp_t gfp_mask)
3340 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3342 EXPORT_SYMBOL(get_zeroed_page);
3344 void __free_pages(struct page *page, unsigned int order)
3346 if (put_page_testzero(page)) {
3348 free_hot_cold_page(page, false);
3350 __free_pages_ok(page, order);
3354 EXPORT_SYMBOL(__free_pages);
3356 void free_pages(unsigned long addr, unsigned int order)
3359 VM_BUG_ON(!virt_addr_valid((void *)addr));
3360 __free_pages(virt_to_page((void *)addr), order);
3364 EXPORT_SYMBOL(free_pages);
3368 * An arbitrary-length arbitrary-offset area of memory which resides
3369 * within a 0 or higher order page. Multiple fragments within that page
3370 * are individually refcounted, in the page's reference counter.
3372 * The page_frag functions below provide a simple allocation framework for
3373 * page fragments. This is used by the network stack and network device
3374 * drivers to provide a backing region of memory for use as either an
3375 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3377 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3380 struct page *page = NULL;
3381 gfp_t gfp = gfp_mask;
3383 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3384 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3386 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3387 PAGE_FRAG_CACHE_MAX_ORDER);
3388 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3390 if (unlikely(!page))
3391 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3393 nc->va = page ? page_address(page) : NULL;
3398 void *__alloc_page_frag(struct page_frag_cache *nc,
3399 unsigned int fragsz, gfp_t gfp_mask)
3401 unsigned int size = PAGE_SIZE;
3405 if (unlikely(!nc->va)) {
3407 page = __page_frag_refill(nc, gfp_mask);
3411 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3412 /* if size can vary use size else just use PAGE_SIZE */
3415 /* Even if we own the page, we do not use atomic_set().
3416 * This would break get_page_unless_zero() users.
3418 atomic_add(size - 1, &page->_count);
3420 /* reset page count bias and offset to start of new frag */
3421 nc->pfmemalloc = page_is_pfmemalloc(page);
3422 nc->pagecnt_bias = size;
3426 offset = nc->offset - fragsz;
3427 if (unlikely(offset < 0)) {
3428 page = virt_to_page(nc->va);
3430 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3433 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3434 /* if size can vary use size else just use PAGE_SIZE */
3437 /* OK, page count is 0, we can safely set it */
3438 atomic_set(&page->_count, size);
3440 /* reset page count bias and offset to start of new frag */
3441 nc->pagecnt_bias = size;
3442 offset = size - fragsz;
3446 nc->offset = offset;
3448 return nc->va + offset;
3450 EXPORT_SYMBOL(__alloc_page_frag);
3453 * Frees a page fragment allocated out of either a compound or order 0 page.
3455 void __free_page_frag(void *addr)
3457 struct page *page = virt_to_head_page(addr);
3459 if (unlikely(put_page_testzero(page)))
3460 __free_pages_ok(page, compound_order(page));
3462 EXPORT_SYMBOL(__free_page_frag);
3465 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3466 * of the current memory cgroup.
3468 * It should be used when the caller would like to use kmalloc, but since the
3469 * allocation is large, it has to fall back to the page allocator.
3471 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3475 page = alloc_pages(gfp_mask, order);
3476 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3477 __free_pages(page, order);
3483 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3487 page = alloc_pages_node(nid, gfp_mask, order);
3488 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3489 __free_pages(page, order);
3496 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3499 void __free_kmem_pages(struct page *page, unsigned int order)
3501 memcg_kmem_uncharge(page, order);
3502 __free_pages(page, order);
3505 void free_kmem_pages(unsigned long addr, unsigned int order)
3508 VM_BUG_ON(!virt_addr_valid((void *)addr));
3509 __free_kmem_pages(virt_to_page((void *)addr), order);
3513 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3517 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3518 unsigned long used = addr + PAGE_ALIGN(size);
3520 split_page(virt_to_page((void *)addr), order);
3521 while (used < alloc_end) {
3526 return (void *)addr;
3530 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3531 * @size: the number of bytes to allocate
3532 * @gfp_mask: GFP flags for the allocation
3534 * This function is similar to alloc_pages(), except that it allocates the
3535 * minimum number of pages to satisfy the request. alloc_pages() can only
3536 * allocate memory in power-of-two pages.
3538 * This function is also limited by MAX_ORDER.
3540 * Memory allocated by this function must be released by free_pages_exact().
3542 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3544 unsigned int order = get_order(size);
3547 addr = __get_free_pages(gfp_mask, order);
3548 return make_alloc_exact(addr, order, size);
3550 EXPORT_SYMBOL(alloc_pages_exact);
3553 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3555 * @nid: the preferred node ID where memory should be allocated
3556 * @size: the number of bytes to allocate
3557 * @gfp_mask: GFP flags for the allocation
3559 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3562 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3564 unsigned int order = get_order(size);
3565 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3568 return make_alloc_exact((unsigned long)page_address(p), order, size);
3572 * free_pages_exact - release memory allocated via alloc_pages_exact()
3573 * @virt: the value returned by alloc_pages_exact.
3574 * @size: size of allocation, same value as passed to alloc_pages_exact().
3576 * Release the memory allocated by a previous call to alloc_pages_exact.
3578 void free_pages_exact(void *virt, size_t size)
3580 unsigned long addr = (unsigned long)virt;
3581 unsigned long end = addr + PAGE_ALIGN(size);
3583 while (addr < end) {
3588 EXPORT_SYMBOL(free_pages_exact);
3591 * nr_free_zone_pages - count number of pages beyond high watermark
3592 * @offset: The zone index of the highest zone
3594 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3595 * high watermark within all zones at or below a given zone index. For each
3596 * zone, the number of pages is calculated as:
3597 * managed_pages - high_pages
3599 static unsigned long nr_free_zone_pages(int offset)
3604 /* Just pick one node, since fallback list is circular */
3605 unsigned long sum = 0;
3607 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3609 for_each_zone_zonelist(zone, z, zonelist, offset) {
3610 unsigned long size = zone->managed_pages;
3611 unsigned long high = high_wmark_pages(zone);
3620 * nr_free_buffer_pages - count number of pages beyond high watermark
3622 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3623 * watermark within ZONE_DMA and ZONE_NORMAL.
3625 unsigned long nr_free_buffer_pages(void)
3627 return nr_free_zone_pages(gfp_zone(GFP_USER));
3629 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3632 * nr_free_pagecache_pages - count number of pages beyond high watermark
3634 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3635 * high watermark within all zones.
3637 unsigned long nr_free_pagecache_pages(void)
3639 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3642 static inline void show_node(struct zone *zone)
3644 if (IS_ENABLED(CONFIG_NUMA))
3645 printk("Node %d ", zone_to_nid(zone));
3648 void si_meminfo(struct sysinfo *val)
3650 val->totalram = totalram_pages;
3651 val->sharedram = global_page_state(NR_SHMEM);
3652 val->freeram = global_page_state(NR_FREE_PAGES);
3653 val->bufferram = nr_blockdev_pages();
3654 val->totalhigh = totalhigh_pages;
3655 val->freehigh = nr_free_highpages();
3656 val->mem_unit = PAGE_SIZE;
3659 EXPORT_SYMBOL(si_meminfo);
3662 void si_meminfo_node(struct sysinfo *val, int nid)
3664 int zone_type; /* needs to be signed */
3665 unsigned long managed_pages = 0;
3666 pg_data_t *pgdat = NODE_DATA(nid);
3668 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3669 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3670 val->totalram = managed_pages;
3671 val->sharedram = node_page_state(nid, NR_SHMEM);
3672 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3673 #ifdef CONFIG_HIGHMEM
3674 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3675 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3681 val->mem_unit = PAGE_SIZE;
3686 * Determine whether the node should be displayed or not, depending on whether
3687 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3689 bool skip_free_areas_node(unsigned int flags, int nid)
3692 unsigned int cpuset_mems_cookie;
3694 if (!(flags & SHOW_MEM_FILTER_NODES))
3698 cpuset_mems_cookie = read_mems_allowed_begin();
3699 ret = !node_isset(nid, cpuset_current_mems_allowed);
3700 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3705 #define K(x) ((x) << (PAGE_SHIFT-10))
3707 static void show_migration_types(unsigned char type)
3709 static const char types[MIGRATE_TYPES] = {
3710 [MIGRATE_UNMOVABLE] = 'U',
3711 [MIGRATE_MOVABLE] = 'M',
3712 [MIGRATE_RECLAIMABLE] = 'E',
3713 [MIGRATE_HIGHATOMIC] = 'H',
3715 [MIGRATE_CMA] = 'C',
3717 #ifdef CONFIG_MEMORY_ISOLATION
3718 [MIGRATE_ISOLATE] = 'I',
3721 char tmp[MIGRATE_TYPES + 1];
3725 for (i = 0; i < MIGRATE_TYPES; i++) {
3726 if (type & (1 << i))
3731 printk("(%s) ", tmp);
3735 * Show free area list (used inside shift_scroll-lock stuff)
3736 * We also calculate the percentage fragmentation. We do this by counting the
3737 * memory on each free list with the exception of the first item on the list.
3740 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3743 void show_free_areas(unsigned int filter)
3745 unsigned long free_pcp = 0;
3749 for_each_populated_zone(zone) {
3750 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3753 for_each_online_cpu(cpu)
3754 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3757 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3758 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3759 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3760 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3761 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3762 " free:%lu free_pcp:%lu free_cma:%lu\n",
3763 global_page_state(NR_ACTIVE_ANON),
3764 global_page_state(NR_INACTIVE_ANON),
3765 global_page_state(NR_ISOLATED_ANON),
3766 global_page_state(NR_ACTIVE_FILE),
3767 global_page_state(NR_INACTIVE_FILE),
3768 global_page_state(NR_ISOLATED_FILE),
3769 global_page_state(NR_UNEVICTABLE),
3770 global_page_state(NR_FILE_DIRTY),
3771 global_page_state(NR_WRITEBACK),
3772 global_page_state(NR_UNSTABLE_NFS),
3773 global_page_state(NR_SLAB_RECLAIMABLE),
3774 global_page_state(NR_SLAB_UNRECLAIMABLE),
3775 global_page_state(NR_FILE_MAPPED),
3776 global_page_state(NR_SHMEM),
3777 global_page_state(NR_PAGETABLE),
3778 global_page_state(NR_BOUNCE),
3779 global_page_state(NR_FREE_PAGES),
3781 global_page_state(NR_FREE_CMA_PAGES));
3783 for_each_populated_zone(zone) {
3786 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3790 for_each_online_cpu(cpu)
3791 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3799 " active_anon:%lukB"
3800 " inactive_anon:%lukB"
3801 " active_file:%lukB"
3802 " inactive_file:%lukB"
3803 " unevictable:%lukB"
3804 " isolated(anon):%lukB"
3805 " isolated(file):%lukB"
3813 " slab_reclaimable:%lukB"
3814 " slab_unreclaimable:%lukB"
3815 " kernel_stack:%lukB"
3822 " writeback_tmp:%lukB"
3823 " pages_scanned:%lu"
3824 " all_unreclaimable? %s"
3827 K(zone_page_state(zone, NR_FREE_PAGES)),
3828 K(min_wmark_pages(zone)),
3829 K(low_wmark_pages(zone)),
3830 K(high_wmark_pages(zone)),
3831 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3832 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3833 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3834 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3835 K(zone_page_state(zone, NR_UNEVICTABLE)),
3836 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3837 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3838 K(zone->present_pages),
3839 K(zone->managed_pages),
3840 K(zone_page_state(zone, NR_MLOCK)),
3841 K(zone_page_state(zone, NR_FILE_DIRTY)),
3842 K(zone_page_state(zone, NR_WRITEBACK)),
3843 K(zone_page_state(zone, NR_FILE_MAPPED)),
3844 K(zone_page_state(zone, NR_SHMEM)),
3845 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3846 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3847 zone_page_state(zone, NR_KERNEL_STACK) *
3849 K(zone_page_state(zone, NR_PAGETABLE)),
3850 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3851 K(zone_page_state(zone, NR_BOUNCE)),
3853 K(this_cpu_read(zone->pageset->pcp.count)),
3854 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3855 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3856 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3857 (!zone_reclaimable(zone) ? "yes" : "no")
3859 printk("lowmem_reserve[]:");
3860 for (i = 0; i < MAX_NR_ZONES; i++)
3861 printk(" %ld", zone->lowmem_reserve[i]);
3865 for_each_populated_zone(zone) {
3867 unsigned long nr[MAX_ORDER], flags, total = 0;
3868 unsigned char types[MAX_ORDER];
3870 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3873 printk("%s: ", zone->name);
3875 spin_lock_irqsave(&zone->lock, flags);
3876 for (order = 0; order < MAX_ORDER; order++) {
3877 struct free_area *area = &zone->free_area[order];
3880 nr[order] = area->nr_free;
3881 total += nr[order] << order;
3884 for (type = 0; type < MIGRATE_TYPES; type++) {
3885 if (!list_empty(&area->free_list[type]))
3886 types[order] |= 1 << type;
3889 spin_unlock_irqrestore(&zone->lock, flags);
3890 for (order = 0; order < MAX_ORDER; order++) {
3891 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3893 show_migration_types(types[order]);
3895 printk("= %lukB\n", K(total));
3898 hugetlb_show_meminfo();
3900 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3902 show_swap_cache_info();
3905 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3907 zoneref->zone = zone;
3908 zoneref->zone_idx = zone_idx(zone);
3912 * Builds allocation fallback zone lists.
3914 * Add all populated zones of a node to the zonelist.
3916 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3920 enum zone_type zone_type = MAX_NR_ZONES;
3924 zone = pgdat->node_zones + zone_type;
3925 if (populated_zone(zone)) {
3926 zoneref_set_zone(zone,
3927 &zonelist->_zonerefs[nr_zones++]);
3928 check_highest_zone(zone_type);
3930 } while (zone_type);
3938 * 0 = automatic detection of better ordering.
3939 * 1 = order by ([node] distance, -zonetype)
3940 * 2 = order by (-zonetype, [node] distance)
3942 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3943 * the same zonelist. So only NUMA can configure this param.
3945 #define ZONELIST_ORDER_DEFAULT 0
3946 #define ZONELIST_ORDER_NODE 1
3947 #define ZONELIST_ORDER_ZONE 2
3949 /* zonelist order in the kernel.
3950 * set_zonelist_order() will set this to NODE or ZONE.
3952 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3953 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3957 /* The value user specified ....changed by config */
3958 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3959 /* string for sysctl */
3960 #define NUMA_ZONELIST_ORDER_LEN 16
3961 char numa_zonelist_order[16] = "default";
3964 * interface for configure zonelist ordering.
3965 * command line option "numa_zonelist_order"
3966 * = "[dD]efault - default, automatic configuration.
3967 * = "[nN]ode - order by node locality, then by zone within node
3968 * = "[zZ]one - order by zone, then by locality within zone
3971 static int __parse_numa_zonelist_order(char *s)
3973 if (*s == 'd' || *s == 'D') {
3974 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3975 } else if (*s == 'n' || *s == 'N') {
3976 user_zonelist_order = ZONELIST_ORDER_NODE;
3977 } else if (*s == 'z' || *s == 'Z') {
3978 user_zonelist_order = ZONELIST_ORDER_ZONE;
3981 "Ignoring invalid numa_zonelist_order value: "
3988 static __init int setup_numa_zonelist_order(char *s)
3995 ret = __parse_numa_zonelist_order(s);
3997 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4001 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4004 * sysctl handler for numa_zonelist_order
4006 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4007 void __user *buffer, size_t *length,
4010 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4012 static DEFINE_MUTEX(zl_order_mutex);
4014 mutex_lock(&zl_order_mutex);
4016 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4020 strcpy(saved_string, (char *)table->data);
4022 ret = proc_dostring(table, write, buffer, length, ppos);
4026 int oldval = user_zonelist_order;
4028 ret = __parse_numa_zonelist_order((char *)table->data);
4031 * bogus value. restore saved string
4033 strncpy((char *)table->data, saved_string,
4034 NUMA_ZONELIST_ORDER_LEN);
4035 user_zonelist_order = oldval;
4036 } else if (oldval != user_zonelist_order) {
4037 mutex_lock(&zonelists_mutex);
4038 build_all_zonelists(NULL, NULL);
4039 mutex_unlock(&zonelists_mutex);
4043 mutex_unlock(&zl_order_mutex);
4048 #define MAX_NODE_LOAD (nr_online_nodes)
4049 static int node_load[MAX_NUMNODES];
4052 * find_next_best_node - find the next node that should appear in a given node's fallback list
4053 * @node: node whose fallback list we're appending
4054 * @used_node_mask: nodemask_t of already used nodes
4056 * We use a number of factors to determine which is the next node that should
4057 * appear on a given node's fallback list. The node should not have appeared
4058 * already in @node's fallback list, and it should be the next closest node
4059 * according to the distance array (which contains arbitrary distance values
4060 * from each node to each node in the system), and should also prefer nodes
4061 * with no CPUs, since presumably they'll have very little allocation pressure
4062 * on them otherwise.
4063 * It returns -1 if no node is found.
4065 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4068 int min_val = INT_MAX;
4069 int best_node = NUMA_NO_NODE;
4070 const struct cpumask *tmp = cpumask_of_node(0);
4072 /* Use the local node if we haven't already */
4073 if (!node_isset(node, *used_node_mask)) {
4074 node_set(node, *used_node_mask);
4078 for_each_node_state(n, N_MEMORY) {
4080 /* Don't want a node to appear more than once */
4081 if (node_isset(n, *used_node_mask))
4084 /* Use the distance array to find the distance */
4085 val = node_distance(node, n);
4087 /* Penalize nodes under us ("prefer the next node") */
4090 /* Give preference to headless and unused nodes */
4091 tmp = cpumask_of_node(n);
4092 if (!cpumask_empty(tmp))
4093 val += PENALTY_FOR_NODE_WITH_CPUS;
4095 /* Slight preference for less loaded node */
4096 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4097 val += node_load[n];
4099 if (val < min_val) {
4106 node_set(best_node, *used_node_mask);
4113 * Build zonelists ordered by node and zones within node.
4114 * This results in maximum locality--normal zone overflows into local
4115 * DMA zone, if any--but risks exhausting DMA zone.
4117 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4120 struct zonelist *zonelist;
4122 zonelist = &pgdat->node_zonelists[0];
4123 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4125 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4126 zonelist->_zonerefs[j].zone = NULL;
4127 zonelist->_zonerefs[j].zone_idx = 0;
4131 * Build gfp_thisnode zonelists
4133 static void build_thisnode_zonelists(pg_data_t *pgdat)
4136 struct zonelist *zonelist;
4138 zonelist = &pgdat->node_zonelists[1];
4139 j = build_zonelists_node(pgdat, zonelist, 0);
4140 zonelist->_zonerefs[j].zone = NULL;
4141 zonelist->_zonerefs[j].zone_idx = 0;
4145 * Build zonelists ordered by zone and nodes within zones.
4146 * This results in conserving DMA zone[s] until all Normal memory is
4147 * exhausted, but results in overflowing to remote node while memory
4148 * may still exist in local DMA zone.
4150 static int node_order[MAX_NUMNODES];
4152 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4155 int zone_type; /* needs to be signed */
4157 struct zonelist *zonelist;
4159 zonelist = &pgdat->node_zonelists[0];
4161 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4162 for (j = 0; j < nr_nodes; j++) {
4163 node = node_order[j];
4164 z = &NODE_DATA(node)->node_zones[zone_type];
4165 if (populated_zone(z)) {
4167 &zonelist->_zonerefs[pos++]);
4168 check_highest_zone(zone_type);
4172 zonelist->_zonerefs[pos].zone = NULL;
4173 zonelist->_zonerefs[pos].zone_idx = 0;
4176 #if defined(CONFIG_64BIT)
4178 * Devices that require DMA32/DMA are relatively rare and do not justify a
4179 * penalty to every machine in case the specialised case applies. Default
4180 * to Node-ordering on 64-bit NUMA machines
4182 static int default_zonelist_order(void)
4184 return ZONELIST_ORDER_NODE;
4188 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4189 * by the kernel. If processes running on node 0 deplete the low memory zone
4190 * then reclaim will occur more frequency increasing stalls and potentially
4191 * be easier to OOM if a large percentage of the zone is under writeback or
4192 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4193 * Hence, default to zone ordering on 32-bit.
4195 static int default_zonelist_order(void)
4197 return ZONELIST_ORDER_ZONE;
4199 #endif /* CONFIG_64BIT */
4201 static void set_zonelist_order(void)
4203 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4204 current_zonelist_order = default_zonelist_order();
4206 current_zonelist_order = user_zonelist_order;
4209 static void build_zonelists(pg_data_t *pgdat)
4213 nodemask_t used_mask;
4214 int local_node, prev_node;
4215 struct zonelist *zonelist;
4216 unsigned int order = current_zonelist_order;
4218 /* initialize zonelists */
4219 for (i = 0; i < MAX_ZONELISTS; i++) {
4220 zonelist = pgdat->node_zonelists + i;
4221 zonelist->_zonerefs[0].zone = NULL;
4222 zonelist->_zonerefs[0].zone_idx = 0;
4225 /* NUMA-aware ordering of nodes */
4226 local_node = pgdat->node_id;
4227 load = nr_online_nodes;
4228 prev_node = local_node;
4229 nodes_clear(used_mask);
4231 memset(node_order, 0, sizeof(node_order));
4234 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4236 * We don't want to pressure a particular node.
4237 * So adding penalty to the first node in same
4238 * distance group to make it round-robin.
4240 if (node_distance(local_node, node) !=
4241 node_distance(local_node, prev_node))
4242 node_load[node] = load;
4246 if (order == ZONELIST_ORDER_NODE)
4247 build_zonelists_in_node_order(pgdat, node);
4249 node_order[j++] = node; /* remember order */
4252 if (order == ZONELIST_ORDER_ZONE) {
4253 /* calculate node order -- i.e., DMA last! */
4254 build_zonelists_in_zone_order(pgdat, j);
4257 build_thisnode_zonelists(pgdat);
4260 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4262 * Return node id of node used for "local" allocations.
4263 * I.e., first node id of first zone in arg node's generic zonelist.
4264 * Used for initializing percpu 'numa_mem', which is used primarily
4265 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4267 int local_memory_node(int node)
4271 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4272 gfp_zone(GFP_KERNEL),
4279 #else /* CONFIG_NUMA */
4281 static void set_zonelist_order(void)
4283 current_zonelist_order = ZONELIST_ORDER_ZONE;
4286 static void build_zonelists(pg_data_t *pgdat)
4288 int node, local_node;
4290 struct zonelist *zonelist;
4292 local_node = pgdat->node_id;
4294 zonelist = &pgdat->node_zonelists[0];
4295 j = build_zonelists_node(pgdat, zonelist, 0);
4298 * Now we build the zonelist so that it contains the zones
4299 * of all the other nodes.
4300 * We don't want to pressure a particular node, so when
4301 * building the zones for node N, we make sure that the
4302 * zones coming right after the local ones are those from
4303 * node N+1 (modulo N)
4305 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4306 if (!node_online(node))
4308 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4310 for (node = 0; node < local_node; node++) {
4311 if (!node_online(node))
4313 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4316 zonelist->_zonerefs[j].zone = NULL;
4317 zonelist->_zonerefs[j].zone_idx = 0;
4320 #endif /* CONFIG_NUMA */
4323 * Boot pageset table. One per cpu which is going to be used for all
4324 * zones and all nodes. The parameters will be set in such a way
4325 * that an item put on a list will immediately be handed over to
4326 * the buddy list. This is safe since pageset manipulation is done
4327 * with interrupts disabled.
4329 * The boot_pagesets must be kept even after bootup is complete for
4330 * unused processors and/or zones. They do play a role for bootstrapping
4331 * hotplugged processors.
4333 * zoneinfo_show() and maybe other functions do
4334 * not check if the processor is online before following the pageset pointer.
4335 * Other parts of the kernel may not check if the zone is available.
4337 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4338 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4339 static void setup_zone_pageset(struct zone *zone);
4342 * Global mutex to protect against size modification of zonelists
4343 * as well as to serialize pageset setup for the new populated zone.
4345 DEFINE_MUTEX(zonelists_mutex);
4347 /* return values int ....just for stop_machine() */
4348 static int __build_all_zonelists(void *data)
4352 pg_data_t *self = data;
4355 memset(node_load, 0, sizeof(node_load));
4358 if (self && !node_online(self->node_id)) {
4359 build_zonelists(self);
4362 for_each_online_node(nid) {
4363 pg_data_t *pgdat = NODE_DATA(nid);
4365 build_zonelists(pgdat);
4369 * Initialize the boot_pagesets that are going to be used
4370 * for bootstrapping processors. The real pagesets for
4371 * each zone will be allocated later when the per cpu
4372 * allocator is available.
4374 * boot_pagesets are used also for bootstrapping offline
4375 * cpus if the system is already booted because the pagesets
4376 * are needed to initialize allocators on a specific cpu too.
4377 * F.e. the percpu allocator needs the page allocator which
4378 * needs the percpu allocator in order to allocate its pagesets
4379 * (a chicken-egg dilemma).
4381 for_each_possible_cpu(cpu) {
4382 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4384 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4386 * We now know the "local memory node" for each node--
4387 * i.e., the node of the first zone in the generic zonelist.
4388 * Set up numa_mem percpu variable for on-line cpus. During
4389 * boot, only the boot cpu should be on-line; we'll init the
4390 * secondary cpus' numa_mem as they come on-line. During
4391 * node/memory hotplug, we'll fixup all on-line cpus.
4393 if (cpu_online(cpu))
4394 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4401 static noinline void __init
4402 build_all_zonelists_init(void)
4404 __build_all_zonelists(NULL);
4405 mminit_verify_zonelist();
4406 cpuset_init_current_mems_allowed();
4410 * Called with zonelists_mutex held always
4411 * unless system_state == SYSTEM_BOOTING.
4413 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4414 * [we're only called with non-NULL zone through __meminit paths] and
4415 * (2) call of __init annotated helper build_all_zonelists_init
4416 * [protected by SYSTEM_BOOTING].
4418 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4420 set_zonelist_order();
4422 if (system_state == SYSTEM_BOOTING) {
4423 build_all_zonelists_init();
4425 #ifdef CONFIG_MEMORY_HOTPLUG
4427 setup_zone_pageset(zone);
4429 /* we have to stop all cpus to guarantee there is no user
4431 stop_machine(__build_all_zonelists, pgdat, NULL);
4432 /* cpuset refresh routine should be here */
4434 vm_total_pages = nr_free_pagecache_pages();
4436 * Disable grouping by mobility if the number of pages in the
4437 * system is too low to allow the mechanism to work. It would be
4438 * more accurate, but expensive to check per-zone. This check is
4439 * made on memory-hotadd so a system can start with mobility
4440 * disabled and enable it later
4442 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4443 page_group_by_mobility_disabled = 1;
4445 page_group_by_mobility_disabled = 0;
4447 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4448 "Total pages: %ld\n",
4450 zonelist_order_name[current_zonelist_order],
4451 page_group_by_mobility_disabled ? "off" : "on",
4454 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4459 * Helper functions to size the waitqueue hash table.
4460 * Essentially these want to choose hash table sizes sufficiently
4461 * large so that collisions trying to wait on pages are rare.
4462 * But in fact, the number of active page waitqueues on typical
4463 * systems is ridiculously low, less than 200. So this is even
4464 * conservative, even though it seems large.
4466 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4467 * waitqueues, i.e. the size of the waitq table given the number of pages.
4469 #define PAGES_PER_WAITQUEUE 256
4471 #ifndef CONFIG_MEMORY_HOTPLUG
4472 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4474 unsigned long size = 1;
4476 pages /= PAGES_PER_WAITQUEUE;
4478 while (size < pages)
4482 * Once we have dozens or even hundreds of threads sleeping
4483 * on IO we've got bigger problems than wait queue collision.
4484 * Limit the size of the wait table to a reasonable size.
4486 size = min(size, 4096UL);
4488 return max(size, 4UL);
4492 * A zone's size might be changed by hot-add, so it is not possible to determine
4493 * a suitable size for its wait_table. So we use the maximum size now.
4495 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4497 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4498 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4499 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4501 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4502 * or more by the traditional way. (See above). It equals:
4504 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4505 * ia64(16K page size) : = ( 8G + 4M)byte.
4506 * powerpc (64K page size) : = (32G +16M)byte.
4508 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4515 * This is an integer logarithm so that shifts can be used later
4516 * to extract the more random high bits from the multiplicative
4517 * hash function before the remainder is taken.
4519 static inline unsigned long wait_table_bits(unsigned long size)
4525 * Initially all pages are reserved - free ones are freed
4526 * up by free_all_bootmem() once the early boot process is
4527 * done. Non-atomic initialization, single-pass.
4529 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4530 unsigned long start_pfn, enum memmap_context context)
4532 pg_data_t *pgdat = NODE_DATA(nid);
4533 unsigned long end_pfn = start_pfn + size;
4536 unsigned long nr_initialised = 0;
4538 if (highest_memmap_pfn < end_pfn - 1)
4539 highest_memmap_pfn = end_pfn - 1;
4541 z = &pgdat->node_zones[zone];
4542 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4544 * There can be holes in boot-time mem_map[]s
4545 * handed to this function. They do not
4546 * exist on hotplugged memory.
4548 if (context == MEMMAP_EARLY) {
4549 if (!early_pfn_valid(pfn))
4551 if (!early_pfn_in_nid(pfn, nid))
4553 if (!update_defer_init(pgdat, pfn, end_pfn,
4559 * Mark the block movable so that blocks are reserved for
4560 * movable at startup. This will force kernel allocations
4561 * to reserve their blocks rather than leaking throughout
4562 * the address space during boot when many long-lived
4563 * kernel allocations are made.
4565 * bitmap is created for zone's valid pfn range. but memmap
4566 * can be created for invalid pages (for alignment)
4567 * check here not to call set_pageblock_migratetype() against
4570 if (!(pfn & (pageblock_nr_pages - 1))) {
4571 struct page *page = pfn_to_page(pfn);
4573 __init_single_page(page, pfn, zone, nid);
4574 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4576 __init_single_pfn(pfn, zone, nid);
4581 static void __meminit zone_init_free_lists(struct zone *zone)
4583 unsigned int order, t;
4584 for_each_migratetype_order(order, t) {
4585 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4586 zone->free_area[order].nr_free = 0;
4590 #ifndef __HAVE_ARCH_MEMMAP_INIT
4591 #define memmap_init(size, nid, zone, start_pfn) \
4592 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4595 static int zone_batchsize(struct zone *zone)
4601 * The per-cpu-pages pools are set to around 1000th of the
4602 * size of the zone. But no more than 1/2 of a meg.
4604 * OK, so we don't know how big the cache is. So guess.
4606 batch = zone->managed_pages / 1024;
4607 if (batch * PAGE_SIZE > 512 * 1024)
4608 batch = (512 * 1024) / PAGE_SIZE;
4609 batch /= 4; /* We effectively *= 4 below */
4614 * Clamp the batch to a 2^n - 1 value. Having a power
4615 * of 2 value was found to be more likely to have
4616 * suboptimal cache aliasing properties in some cases.
4618 * For example if 2 tasks are alternately allocating
4619 * batches of pages, one task can end up with a lot
4620 * of pages of one half of the possible page colors
4621 * and the other with pages of the other colors.
4623 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4628 /* The deferral and batching of frees should be suppressed under NOMMU
4631 * The problem is that NOMMU needs to be able to allocate large chunks
4632 * of contiguous memory as there's no hardware page translation to
4633 * assemble apparent contiguous memory from discontiguous pages.
4635 * Queueing large contiguous runs of pages for batching, however,
4636 * causes the pages to actually be freed in smaller chunks. As there
4637 * can be a significant delay between the individual batches being
4638 * recycled, this leads to the once large chunks of space being
4639 * fragmented and becoming unavailable for high-order allocations.
4646 * pcp->high and pcp->batch values are related and dependent on one another:
4647 * ->batch must never be higher then ->high.
4648 * The following function updates them in a safe manner without read side
4651 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4652 * those fields changing asynchronously (acording the the above rule).
4654 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4655 * outside of boot time (or some other assurance that no concurrent updaters
4658 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4659 unsigned long batch)
4661 /* start with a fail safe value for batch */
4665 /* Update high, then batch, in order */
4672 /* a companion to pageset_set_high() */
4673 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4675 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4678 static void pageset_init(struct per_cpu_pageset *p)
4680 struct per_cpu_pages *pcp;
4683 memset(p, 0, sizeof(*p));
4687 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4688 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4691 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4694 pageset_set_batch(p, batch);
4698 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4699 * to the value high for the pageset p.
4701 static void pageset_set_high(struct per_cpu_pageset *p,
4704 unsigned long batch = max(1UL, high / 4);
4705 if ((high / 4) > (PAGE_SHIFT * 8))
4706 batch = PAGE_SHIFT * 8;
4708 pageset_update(&p->pcp, high, batch);
4711 static void pageset_set_high_and_batch(struct zone *zone,
4712 struct per_cpu_pageset *pcp)
4714 if (percpu_pagelist_fraction)
4715 pageset_set_high(pcp,
4716 (zone->managed_pages /
4717 percpu_pagelist_fraction));
4719 pageset_set_batch(pcp, zone_batchsize(zone));
4722 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4724 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4727 pageset_set_high_and_batch(zone, pcp);
4730 static void __meminit setup_zone_pageset(struct zone *zone)
4733 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4734 for_each_possible_cpu(cpu)
4735 zone_pageset_init(zone, cpu);
4739 * Allocate per cpu pagesets and initialize them.
4740 * Before this call only boot pagesets were available.
4742 void __init setup_per_cpu_pageset(void)
4746 for_each_populated_zone(zone)
4747 setup_zone_pageset(zone);
4750 static noinline __init_refok
4751 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4757 * The per-page waitqueue mechanism uses hashed waitqueues
4760 zone->wait_table_hash_nr_entries =
4761 wait_table_hash_nr_entries(zone_size_pages);
4762 zone->wait_table_bits =
4763 wait_table_bits(zone->wait_table_hash_nr_entries);
4764 alloc_size = zone->wait_table_hash_nr_entries
4765 * sizeof(wait_queue_head_t);
4767 if (!slab_is_available()) {
4768 zone->wait_table = (wait_queue_head_t *)
4769 memblock_virt_alloc_node_nopanic(
4770 alloc_size, zone->zone_pgdat->node_id);
4773 * This case means that a zone whose size was 0 gets new memory
4774 * via memory hot-add.
4775 * But it may be the case that a new node was hot-added. In
4776 * this case vmalloc() will not be able to use this new node's
4777 * memory - this wait_table must be initialized to use this new
4778 * node itself as well.
4779 * To use this new node's memory, further consideration will be
4782 zone->wait_table = vmalloc(alloc_size);
4784 if (!zone->wait_table)
4787 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4788 init_waitqueue_head(zone->wait_table + i);
4793 static __meminit void zone_pcp_init(struct zone *zone)
4796 * per cpu subsystem is not up at this point. The following code
4797 * relies on the ability of the linker to provide the
4798 * offset of a (static) per cpu variable into the per cpu area.
4800 zone->pageset = &boot_pageset;
4802 if (populated_zone(zone))
4803 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4804 zone->name, zone->present_pages,
4805 zone_batchsize(zone));
4808 int __meminit init_currently_empty_zone(struct zone *zone,
4809 unsigned long zone_start_pfn,
4812 struct pglist_data *pgdat = zone->zone_pgdat;
4814 ret = zone_wait_table_init(zone, size);
4817 pgdat->nr_zones = zone_idx(zone) + 1;
4819 zone->zone_start_pfn = zone_start_pfn;
4821 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4822 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4824 (unsigned long)zone_idx(zone),
4825 zone_start_pfn, (zone_start_pfn + size));
4827 zone_init_free_lists(zone);
4832 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4833 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4836 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4838 int __meminit __early_pfn_to_nid(unsigned long pfn,
4839 struct mminit_pfnnid_cache *state)
4841 unsigned long start_pfn, end_pfn;
4844 if (state->last_start <= pfn && pfn < state->last_end)
4845 return state->last_nid;
4847 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4849 state->last_start = start_pfn;
4850 state->last_end = end_pfn;
4851 state->last_nid = nid;
4856 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4859 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4860 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4861 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4863 * If an architecture guarantees that all ranges registered contain no holes
4864 * and may be freed, this this function may be used instead of calling
4865 * memblock_free_early_nid() manually.
4867 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4869 unsigned long start_pfn, end_pfn;
4872 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4873 start_pfn = min(start_pfn, max_low_pfn);
4874 end_pfn = min(end_pfn, max_low_pfn);
4876 if (start_pfn < end_pfn)
4877 memblock_free_early_nid(PFN_PHYS(start_pfn),
4878 (end_pfn - start_pfn) << PAGE_SHIFT,
4884 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4885 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4887 * If an architecture guarantees that all ranges registered contain no holes and may
4888 * be freed, this function may be used instead of calling memory_present() manually.
4890 void __init sparse_memory_present_with_active_regions(int nid)
4892 unsigned long start_pfn, end_pfn;
4895 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4896 memory_present(this_nid, start_pfn, end_pfn);
4900 * get_pfn_range_for_nid - Return the start and end page frames for a node
4901 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4902 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4903 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4905 * It returns the start and end page frame of a node based on information
4906 * provided by memblock_set_node(). If called for a node
4907 * with no available memory, a warning is printed and the start and end
4910 void __meminit get_pfn_range_for_nid(unsigned int nid,
4911 unsigned long *start_pfn, unsigned long *end_pfn)
4913 unsigned long this_start_pfn, this_end_pfn;
4919 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4920 *start_pfn = min(*start_pfn, this_start_pfn);
4921 *end_pfn = max(*end_pfn, this_end_pfn);
4924 if (*start_pfn == -1UL)
4929 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4930 * assumption is made that zones within a node are ordered in monotonic
4931 * increasing memory addresses so that the "highest" populated zone is used
4933 static void __init find_usable_zone_for_movable(void)
4936 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4937 if (zone_index == ZONE_MOVABLE)
4940 if (arch_zone_highest_possible_pfn[zone_index] >
4941 arch_zone_lowest_possible_pfn[zone_index])
4945 VM_BUG_ON(zone_index == -1);
4946 movable_zone = zone_index;
4950 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4951 * because it is sized independent of architecture. Unlike the other zones,
4952 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4953 * in each node depending on the size of each node and how evenly kernelcore
4954 * is distributed. This helper function adjusts the zone ranges
4955 * provided by the architecture for a given node by using the end of the
4956 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4957 * zones within a node are in order of monotonic increases memory addresses
4959 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4960 unsigned long zone_type,
4961 unsigned long node_start_pfn,
4962 unsigned long node_end_pfn,
4963 unsigned long *zone_start_pfn,
4964 unsigned long *zone_end_pfn)
4966 /* Only adjust if ZONE_MOVABLE is on this node */
4967 if (zone_movable_pfn[nid]) {
4968 /* Size ZONE_MOVABLE */
4969 if (zone_type == ZONE_MOVABLE) {
4970 *zone_start_pfn = zone_movable_pfn[nid];
4971 *zone_end_pfn = min(node_end_pfn,
4972 arch_zone_highest_possible_pfn[movable_zone]);
4974 /* Adjust for ZONE_MOVABLE starting within this range */
4975 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4976 *zone_end_pfn > zone_movable_pfn[nid]) {
4977 *zone_end_pfn = zone_movable_pfn[nid];
4979 /* Check if this whole range is within ZONE_MOVABLE */
4980 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4981 *zone_start_pfn = *zone_end_pfn;
4986 * Return the number of pages a zone spans in a node, including holes
4987 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4989 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4990 unsigned long zone_type,
4991 unsigned long node_start_pfn,
4992 unsigned long node_end_pfn,
4993 unsigned long *ignored)
4995 unsigned long zone_start_pfn, zone_end_pfn;
4997 /* When hotadd a new node from cpu_up(), the node should be empty */
4998 if (!node_start_pfn && !node_end_pfn)
5001 /* Get the start and end of the zone */
5002 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5003 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5004 adjust_zone_range_for_zone_movable(nid, zone_type,
5005 node_start_pfn, node_end_pfn,
5006 &zone_start_pfn, &zone_end_pfn);
5008 /* Check that this node has pages within the zone's required range */
5009 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
5012 /* Move the zone boundaries inside the node if necessary */
5013 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
5014 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
5016 /* Return the spanned pages */
5017 return zone_end_pfn - zone_start_pfn;
5021 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5022 * then all holes in the requested range will be accounted for.
5024 unsigned long __meminit __absent_pages_in_range(int nid,
5025 unsigned long range_start_pfn,
5026 unsigned long range_end_pfn)
5028 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5029 unsigned long start_pfn, end_pfn;
5032 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5033 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5034 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5035 nr_absent -= end_pfn - start_pfn;
5041 * absent_pages_in_range - Return number of page frames in holes within a range
5042 * @start_pfn: The start PFN to start searching for holes
5043 * @end_pfn: The end PFN to stop searching for holes
5045 * It returns the number of pages frames in memory holes within a range.
5047 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5048 unsigned long end_pfn)
5050 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5053 /* Return the number of page frames in holes in a zone on a node */
5054 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5055 unsigned long zone_type,
5056 unsigned long node_start_pfn,
5057 unsigned long node_end_pfn,
5058 unsigned long *ignored)
5060 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5061 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5062 unsigned long zone_start_pfn, zone_end_pfn;
5064 /* When hotadd a new node from cpu_up(), the node should be empty */
5065 if (!node_start_pfn && !node_end_pfn)
5068 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5069 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5071 adjust_zone_range_for_zone_movable(nid, zone_type,
5072 node_start_pfn, node_end_pfn,
5073 &zone_start_pfn, &zone_end_pfn);
5074 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5077 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5078 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5079 unsigned long zone_type,
5080 unsigned long node_start_pfn,
5081 unsigned long node_end_pfn,
5082 unsigned long *zones_size)
5084 return zones_size[zone_type];
5087 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5088 unsigned long zone_type,
5089 unsigned long node_start_pfn,
5090 unsigned long node_end_pfn,
5091 unsigned long *zholes_size)
5096 return zholes_size[zone_type];
5099 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5101 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5102 unsigned long node_start_pfn,
5103 unsigned long node_end_pfn,
5104 unsigned long *zones_size,
5105 unsigned long *zholes_size)
5107 unsigned long realtotalpages = 0, totalpages = 0;
5110 for (i = 0; i < MAX_NR_ZONES; i++) {
5111 struct zone *zone = pgdat->node_zones + i;
5112 unsigned long size, real_size;
5114 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5118 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5119 node_start_pfn, node_end_pfn,
5121 zone->spanned_pages = size;
5122 zone->present_pages = real_size;
5125 realtotalpages += real_size;
5128 pgdat->node_spanned_pages = totalpages;
5129 pgdat->node_present_pages = realtotalpages;
5130 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5134 #ifndef CONFIG_SPARSEMEM
5136 * Calculate the size of the zone->blockflags rounded to an unsigned long
5137 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5138 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5139 * round what is now in bits to nearest long in bits, then return it in
5142 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5144 unsigned long usemapsize;
5146 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5147 usemapsize = roundup(zonesize, pageblock_nr_pages);
5148 usemapsize = usemapsize >> pageblock_order;
5149 usemapsize *= NR_PAGEBLOCK_BITS;
5150 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5152 return usemapsize / 8;
5155 static void __init setup_usemap(struct pglist_data *pgdat,
5157 unsigned long zone_start_pfn,
5158 unsigned long zonesize)
5160 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5161 zone->pageblock_flags = NULL;
5163 zone->pageblock_flags =
5164 memblock_virt_alloc_node_nopanic(usemapsize,
5168 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5169 unsigned long zone_start_pfn, unsigned long zonesize) {}
5170 #endif /* CONFIG_SPARSEMEM */
5172 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5174 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5175 void __paginginit set_pageblock_order(void)
5179 /* Check that pageblock_nr_pages has not already been setup */
5180 if (pageblock_order)
5183 if (HPAGE_SHIFT > PAGE_SHIFT)
5184 order = HUGETLB_PAGE_ORDER;
5186 order = MAX_ORDER - 1;
5189 * Assume the largest contiguous order of interest is a huge page.
5190 * This value may be variable depending on boot parameters on IA64 and
5193 pageblock_order = order;
5195 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5198 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5199 * is unused as pageblock_order is set at compile-time. See
5200 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5203 void __paginginit set_pageblock_order(void)
5207 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5209 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5210 unsigned long present_pages)
5212 unsigned long pages = spanned_pages;
5215 * Provide a more accurate estimation if there are holes within
5216 * the zone and SPARSEMEM is in use. If there are holes within the
5217 * zone, each populated memory region may cost us one or two extra
5218 * memmap pages due to alignment because memmap pages for each
5219 * populated regions may not naturally algined on page boundary.
5220 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5222 if (spanned_pages > present_pages + (present_pages >> 4) &&
5223 IS_ENABLED(CONFIG_SPARSEMEM))
5224 pages = present_pages;
5226 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5230 * Set up the zone data structures:
5231 * - mark all pages reserved
5232 * - mark all memory queues empty
5233 * - clear the memory bitmaps
5235 * NOTE: pgdat should get zeroed by caller.
5237 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5240 int nid = pgdat->node_id;
5241 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5244 pgdat_resize_init(pgdat);
5245 #ifdef CONFIG_NUMA_BALANCING
5246 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5247 pgdat->numabalancing_migrate_nr_pages = 0;
5248 pgdat->numabalancing_migrate_next_window = jiffies;
5250 init_waitqueue_head(&pgdat->kswapd_wait);
5251 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5252 pgdat_page_ext_init(pgdat);
5254 for (j = 0; j < MAX_NR_ZONES; j++) {
5255 struct zone *zone = pgdat->node_zones + j;
5256 unsigned long size, realsize, freesize, memmap_pages;
5258 size = zone->spanned_pages;
5259 realsize = freesize = zone->present_pages;
5262 * Adjust freesize so that it accounts for how much memory
5263 * is used by this zone for memmap. This affects the watermark
5264 * and per-cpu initialisations
5266 memmap_pages = calc_memmap_size(size, realsize);
5267 if (!is_highmem_idx(j)) {
5268 if (freesize >= memmap_pages) {
5269 freesize -= memmap_pages;
5272 " %s zone: %lu pages used for memmap\n",
5273 zone_names[j], memmap_pages);
5276 " %s zone: %lu pages exceeds freesize %lu\n",
5277 zone_names[j], memmap_pages, freesize);
5280 /* Account for reserved pages */
5281 if (j == 0 && freesize > dma_reserve) {
5282 freesize -= dma_reserve;
5283 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5284 zone_names[0], dma_reserve);
5287 if (!is_highmem_idx(j))
5288 nr_kernel_pages += freesize;
5289 /* Charge for highmem memmap if there are enough kernel pages */
5290 else if (nr_kernel_pages > memmap_pages * 2)
5291 nr_kernel_pages -= memmap_pages;
5292 nr_all_pages += freesize;
5295 * Set an approximate value for lowmem here, it will be adjusted
5296 * when the bootmem allocator frees pages into the buddy system.
5297 * And all highmem pages will be managed by the buddy system.
5299 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5302 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5304 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5306 zone->name = zone_names[j];
5307 spin_lock_init(&zone->lock);
5308 spin_lock_init(&zone->lru_lock);
5309 zone_seqlock_init(zone);
5310 zone->zone_pgdat = pgdat;
5311 zone_pcp_init(zone);
5313 /* For bootup, initialized properly in watermark setup */
5314 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5316 lruvec_init(&zone->lruvec);
5320 set_pageblock_order();
5321 setup_usemap(pgdat, zone, zone_start_pfn, size);
5322 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5324 memmap_init(size, nid, j, zone_start_pfn);
5325 zone_start_pfn += size;
5329 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5331 unsigned long __maybe_unused start = 0;
5332 unsigned long __maybe_unused offset = 0;
5334 /* Skip empty nodes */
5335 if (!pgdat->node_spanned_pages)
5338 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5339 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5340 offset = pgdat->node_start_pfn - start;
5341 /* ia64 gets its own node_mem_map, before this, without bootmem */
5342 if (!pgdat->node_mem_map) {
5343 unsigned long size, end;
5347 * The zone's endpoints aren't required to be MAX_ORDER
5348 * aligned but the node_mem_map endpoints must be in order
5349 * for the buddy allocator to function correctly.
5351 end = pgdat_end_pfn(pgdat);
5352 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5353 size = (end - start) * sizeof(struct page);
5354 map = alloc_remap(pgdat->node_id, size);
5356 map = memblock_virt_alloc_node_nopanic(size,
5358 pgdat->node_mem_map = map + offset;
5360 #ifndef CONFIG_NEED_MULTIPLE_NODES
5362 * With no DISCONTIG, the global mem_map is just set as node 0's
5364 if (pgdat == NODE_DATA(0)) {
5365 mem_map = NODE_DATA(0)->node_mem_map;
5366 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5367 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5369 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5372 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5375 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5376 unsigned long node_start_pfn, unsigned long *zholes_size)
5378 pg_data_t *pgdat = NODE_DATA(nid);
5379 unsigned long start_pfn = 0;
5380 unsigned long end_pfn = 0;
5382 /* pg_data_t should be reset to zero when it's allocated */
5383 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5385 reset_deferred_meminit(pgdat);
5386 pgdat->node_id = nid;
5387 pgdat->node_start_pfn = node_start_pfn;
5388 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5389 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5390 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5391 (u64)start_pfn << PAGE_SHIFT,
5392 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5394 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5395 zones_size, zholes_size);
5397 alloc_node_mem_map(pgdat);
5398 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5399 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5400 nid, (unsigned long)pgdat,
5401 (unsigned long)pgdat->node_mem_map);
5404 free_area_init_core(pgdat);
5407 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5409 #if MAX_NUMNODES > 1
5411 * Figure out the number of possible node ids.
5413 void __init setup_nr_node_ids(void)
5415 unsigned int highest;
5417 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5418 nr_node_ids = highest + 1;
5423 * node_map_pfn_alignment - determine the maximum internode alignment
5425 * This function should be called after node map is populated and sorted.
5426 * It calculates the maximum power of two alignment which can distinguish
5429 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5430 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5431 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5432 * shifted, 1GiB is enough and this function will indicate so.
5434 * This is used to test whether pfn -> nid mapping of the chosen memory
5435 * model has fine enough granularity to avoid incorrect mapping for the
5436 * populated node map.
5438 * Returns the determined alignment in pfn's. 0 if there is no alignment
5439 * requirement (single node).
5441 unsigned long __init node_map_pfn_alignment(void)
5443 unsigned long accl_mask = 0, last_end = 0;
5444 unsigned long start, end, mask;
5448 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5449 if (!start || last_nid < 0 || last_nid == nid) {
5456 * Start with a mask granular enough to pin-point to the
5457 * start pfn and tick off bits one-by-one until it becomes
5458 * too coarse to separate the current node from the last.
5460 mask = ~((1 << __ffs(start)) - 1);
5461 while (mask && last_end <= (start & (mask << 1)))
5464 /* accumulate all internode masks */
5468 /* convert mask to number of pages */
5469 return ~accl_mask + 1;
5472 /* Find the lowest pfn for a node */
5473 static unsigned long __init find_min_pfn_for_node(int nid)
5475 unsigned long min_pfn = ULONG_MAX;
5476 unsigned long start_pfn;
5479 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5480 min_pfn = min(min_pfn, start_pfn);
5482 if (min_pfn == ULONG_MAX) {
5484 "Could not find start_pfn for node %d\n", nid);
5492 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5494 * It returns the minimum PFN based on information provided via
5495 * memblock_set_node().
5497 unsigned long __init find_min_pfn_with_active_regions(void)
5499 return find_min_pfn_for_node(MAX_NUMNODES);
5503 * early_calculate_totalpages()
5504 * Sum pages in active regions for movable zone.
5505 * Populate N_MEMORY for calculating usable_nodes.
5507 static unsigned long __init early_calculate_totalpages(void)
5509 unsigned long totalpages = 0;
5510 unsigned long start_pfn, end_pfn;
5513 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5514 unsigned long pages = end_pfn - start_pfn;
5516 totalpages += pages;
5518 node_set_state(nid, N_MEMORY);
5524 * Find the PFN the Movable zone begins in each node. Kernel memory
5525 * is spread evenly between nodes as long as the nodes have enough
5526 * memory. When they don't, some nodes will have more kernelcore than
5529 static void __init find_zone_movable_pfns_for_nodes(void)
5532 unsigned long usable_startpfn;
5533 unsigned long kernelcore_node, kernelcore_remaining;
5534 /* save the state before borrow the nodemask */
5535 nodemask_t saved_node_state = node_states[N_MEMORY];
5536 unsigned long totalpages = early_calculate_totalpages();
5537 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5538 struct memblock_region *r;
5540 /* Need to find movable_zone earlier when movable_node is specified. */
5541 find_usable_zone_for_movable();
5544 * If movable_node is specified, ignore kernelcore and movablecore
5547 if (movable_node_is_enabled()) {
5548 for_each_memblock(memory, r) {
5549 if (!memblock_is_hotpluggable(r))
5554 usable_startpfn = PFN_DOWN(r->base);
5555 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5556 min(usable_startpfn, zone_movable_pfn[nid]) :
5564 * If movablecore=nn[KMG] was specified, calculate what size of
5565 * kernelcore that corresponds so that memory usable for
5566 * any allocation type is evenly spread. If both kernelcore
5567 * and movablecore are specified, then the value of kernelcore
5568 * will be used for required_kernelcore if it's greater than
5569 * what movablecore would have allowed.
5571 if (required_movablecore) {
5572 unsigned long corepages;
5575 * Round-up so that ZONE_MOVABLE is at least as large as what
5576 * was requested by the user
5578 required_movablecore =
5579 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5580 required_movablecore = min(totalpages, required_movablecore);
5581 corepages = totalpages - required_movablecore;
5583 required_kernelcore = max(required_kernelcore, corepages);
5587 * If kernelcore was not specified or kernelcore size is larger
5588 * than totalpages, there is no ZONE_MOVABLE.
5590 if (!required_kernelcore || required_kernelcore >= totalpages)
5593 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5594 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5597 /* Spread kernelcore memory as evenly as possible throughout nodes */
5598 kernelcore_node = required_kernelcore / usable_nodes;
5599 for_each_node_state(nid, N_MEMORY) {
5600 unsigned long start_pfn, end_pfn;
5603 * Recalculate kernelcore_node if the division per node
5604 * now exceeds what is necessary to satisfy the requested
5605 * amount of memory for the kernel
5607 if (required_kernelcore < kernelcore_node)
5608 kernelcore_node = required_kernelcore / usable_nodes;
5611 * As the map is walked, we track how much memory is usable
5612 * by the kernel using kernelcore_remaining. When it is
5613 * 0, the rest of the node is usable by ZONE_MOVABLE
5615 kernelcore_remaining = kernelcore_node;
5617 /* Go through each range of PFNs within this node */
5618 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5619 unsigned long size_pages;
5621 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5622 if (start_pfn >= end_pfn)
5625 /* Account for what is only usable for kernelcore */
5626 if (start_pfn < usable_startpfn) {
5627 unsigned long kernel_pages;
5628 kernel_pages = min(end_pfn, usable_startpfn)
5631 kernelcore_remaining -= min(kernel_pages,
5632 kernelcore_remaining);
5633 required_kernelcore -= min(kernel_pages,
5634 required_kernelcore);
5636 /* Continue if range is now fully accounted */
5637 if (end_pfn <= usable_startpfn) {
5640 * Push zone_movable_pfn to the end so
5641 * that if we have to rebalance
5642 * kernelcore across nodes, we will
5643 * not double account here
5645 zone_movable_pfn[nid] = end_pfn;
5648 start_pfn = usable_startpfn;
5652 * The usable PFN range for ZONE_MOVABLE is from
5653 * start_pfn->end_pfn. Calculate size_pages as the
5654 * number of pages used as kernelcore
5656 size_pages = end_pfn - start_pfn;
5657 if (size_pages > kernelcore_remaining)
5658 size_pages = kernelcore_remaining;
5659 zone_movable_pfn[nid] = start_pfn + size_pages;
5662 * Some kernelcore has been met, update counts and
5663 * break if the kernelcore for this node has been
5666 required_kernelcore -= min(required_kernelcore,
5668 kernelcore_remaining -= size_pages;
5669 if (!kernelcore_remaining)
5675 * If there is still required_kernelcore, we do another pass with one
5676 * less node in the count. This will push zone_movable_pfn[nid] further
5677 * along on the nodes that still have memory until kernelcore is
5681 if (usable_nodes && required_kernelcore > usable_nodes)
5685 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5686 for (nid = 0; nid < MAX_NUMNODES; nid++)
5687 zone_movable_pfn[nid] =
5688 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5691 /* restore the node_state */
5692 node_states[N_MEMORY] = saved_node_state;
5695 /* Any regular or high memory on that node ? */
5696 static void check_for_memory(pg_data_t *pgdat, int nid)
5698 enum zone_type zone_type;
5700 if (N_MEMORY == N_NORMAL_MEMORY)
5703 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5704 struct zone *zone = &pgdat->node_zones[zone_type];
5705 if (populated_zone(zone)) {
5706 node_set_state(nid, N_HIGH_MEMORY);
5707 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5708 zone_type <= ZONE_NORMAL)
5709 node_set_state(nid, N_NORMAL_MEMORY);
5716 * free_area_init_nodes - Initialise all pg_data_t and zone data
5717 * @max_zone_pfn: an array of max PFNs for each zone
5719 * This will call free_area_init_node() for each active node in the system.
5720 * Using the page ranges provided by memblock_set_node(), the size of each
5721 * zone in each node and their holes is calculated. If the maximum PFN
5722 * between two adjacent zones match, it is assumed that the zone is empty.
5723 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5724 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5725 * starts where the previous one ended. For example, ZONE_DMA32 starts
5726 * at arch_max_dma_pfn.
5728 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5730 unsigned long start_pfn, end_pfn;
5733 /* Record where the zone boundaries are */
5734 memset(arch_zone_lowest_possible_pfn, 0,
5735 sizeof(arch_zone_lowest_possible_pfn));
5736 memset(arch_zone_highest_possible_pfn, 0,
5737 sizeof(arch_zone_highest_possible_pfn));
5738 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5739 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5740 for (i = 1; i < MAX_NR_ZONES; i++) {
5741 if (i == ZONE_MOVABLE)
5743 arch_zone_lowest_possible_pfn[i] =
5744 arch_zone_highest_possible_pfn[i-1];
5745 arch_zone_highest_possible_pfn[i] =
5746 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5748 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5749 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5751 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5752 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5753 find_zone_movable_pfns_for_nodes();
5755 /* Print out the zone ranges */
5756 pr_info("Zone ranges:\n");
5757 for (i = 0; i < MAX_NR_ZONES; i++) {
5758 if (i == ZONE_MOVABLE)
5760 pr_info(" %-8s ", zone_names[i]);
5761 if (arch_zone_lowest_possible_pfn[i] ==
5762 arch_zone_highest_possible_pfn[i])
5765 pr_cont("[mem %#018Lx-%#018Lx]\n",
5766 (u64)arch_zone_lowest_possible_pfn[i]
5768 ((u64)arch_zone_highest_possible_pfn[i]
5769 << PAGE_SHIFT) - 1);
5772 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5773 pr_info("Movable zone start for each node\n");
5774 for (i = 0; i < MAX_NUMNODES; i++) {
5775 if (zone_movable_pfn[i])
5776 pr_info(" Node %d: %#018Lx\n", i,
5777 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5780 /* Print out the early node map */
5781 pr_info("Early memory node ranges\n");
5782 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5783 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5784 (u64)start_pfn << PAGE_SHIFT,
5785 ((u64)end_pfn << PAGE_SHIFT) - 1);
5787 /* Initialise every node */
5788 mminit_verify_pageflags_layout();
5789 setup_nr_node_ids();
5790 for_each_online_node(nid) {
5791 pg_data_t *pgdat = NODE_DATA(nid);
5792 free_area_init_node(nid, NULL,
5793 find_min_pfn_for_node(nid), NULL);
5795 /* Any memory on that node */
5796 if (pgdat->node_present_pages)
5797 node_set_state(nid, N_MEMORY);
5798 check_for_memory(pgdat, nid);
5802 static int __init cmdline_parse_core(char *p, unsigned long *core)
5804 unsigned long long coremem;
5808 coremem = memparse(p, &p);
5809 *core = coremem >> PAGE_SHIFT;
5811 /* Paranoid check that UL is enough for the coremem value */
5812 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5818 * kernelcore=size sets the amount of memory for use for allocations that
5819 * cannot be reclaimed or migrated.
5821 static int __init cmdline_parse_kernelcore(char *p)
5823 return cmdline_parse_core(p, &required_kernelcore);
5827 * movablecore=size sets the amount of memory for use for allocations that
5828 * can be reclaimed or migrated.
5830 static int __init cmdline_parse_movablecore(char *p)
5832 return cmdline_parse_core(p, &required_movablecore);
5835 early_param("kernelcore", cmdline_parse_kernelcore);
5836 early_param("movablecore", cmdline_parse_movablecore);
5838 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5840 void adjust_managed_page_count(struct page *page, long count)
5842 spin_lock(&managed_page_count_lock);
5843 page_zone(page)->managed_pages += count;
5844 totalram_pages += count;
5845 #ifdef CONFIG_HIGHMEM
5846 if (PageHighMem(page))
5847 totalhigh_pages += count;
5849 spin_unlock(&managed_page_count_lock);
5851 EXPORT_SYMBOL(adjust_managed_page_count);
5853 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5856 unsigned long pages = 0;
5858 start = (void *)PAGE_ALIGN((unsigned long)start);
5859 end = (void *)((unsigned long)end & PAGE_MASK);
5860 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5861 if ((unsigned int)poison <= 0xFF)
5862 memset(pos, poison, PAGE_SIZE);
5863 free_reserved_page(virt_to_page(pos));
5867 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5868 s, pages << (PAGE_SHIFT - 10), start, end);
5872 EXPORT_SYMBOL(free_reserved_area);
5874 #ifdef CONFIG_HIGHMEM
5875 void free_highmem_page(struct page *page)
5877 __free_reserved_page(page);
5879 page_zone(page)->managed_pages++;
5885 void __init mem_init_print_info(const char *str)
5887 unsigned long physpages, codesize, datasize, rosize, bss_size;
5888 unsigned long init_code_size, init_data_size;
5890 physpages = get_num_physpages();
5891 codesize = _etext - _stext;
5892 datasize = _edata - _sdata;
5893 rosize = __end_rodata - __start_rodata;
5894 bss_size = __bss_stop - __bss_start;
5895 init_data_size = __init_end - __init_begin;
5896 init_code_size = _einittext - _sinittext;
5899 * Detect special cases and adjust section sizes accordingly:
5900 * 1) .init.* may be embedded into .data sections
5901 * 2) .init.text.* may be out of [__init_begin, __init_end],
5902 * please refer to arch/tile/kernel/vmlinux.lds.S.
5903 * 3) .rodata.* may be embedded into .text or .data sections.
5905 #define adj_init_size(start, end, size, pos, adj) \
5907 if (start <= pos && pos < end && size > adj) \
5911 adj_init_size(__init_begin, __init_end, init_data_size,
5912 _sinittext, init_code_size);
5913 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5914 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5915 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5916 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5918 #undef adj_init_size
5920 pr_info("Memory: %luK/%luK available "
5921 "(%luK kernel code, %luK rwdata, %luK rodata, "
5922 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5923 #ifdef CONFIG_HIGHMEM
5927 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5928 codesize >> 10, datasize >> 10, rosize >> 10,
5929 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5930 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5931 totalcma_pages << (PAGE_SHIFT-10),
5932 #ifdef CONFIG_HIGHMEM
5933 totalhigh_pages << (PAGE_SHIFT-10),
5935 str ? ", " : "", str ? str : "");
5939 * set_dma_reserve - set the specified number of pages reserved in the first zone
5940 * @new_dma_reserve: The number of pages to mark reserved
5942 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
5943 * In the DMA zone, a significant percentage may be consumed by kernel image
5944 * and other unfreeable allocations which can skew the watermarks badly. This
5945 * function may optionally be used to account for unfreeable pages in the
5946 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5947 * smaller per-cpu batchsize.
5949 void __init set_dma_reserve(unsigned long new_dma_reserve)
5951 dma_reserve = new_dma_reserve;
5954 void __init free_area_init(unsigned long *zones_size)
5956 free_area_init_node(0, zones_size,
5957 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5960 static int page_alloc_cpu_notify(struct notifier_block *self,
5961 unsigned long action, void *hcpu)
5963 int cpu = (unsigned long)hcpu;
5965 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5966 lru_add_drain_cpu(cpu);
5970 * Spill the event counters of the dead processor
5971 * into the current processors event counters.
5972 * This artificially elevates the count of the current
5975 vm_events_fold_cpu(cpu);
5978 * Zero the differential counters of the dead processor
5979 * so that the vm statistics are consistent.
5981 * This is only okay since the processor is dead and cannot
5982 * race with what we are doing.
5984 cpu_vm_stats_fold(cpu);
5989 void __init page_alloc_init(void)
5991 hotcpu_notifier(page_alloc_cpu_notify, 0);
5992 local_irq_lock_init(pa_lock);
5996 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5997 * or min_free_kbytes changes.
5999 static void calculate_totalreserve_pages(void)
6001 struct pglist_data *pgdat;
6002 unsigned long reserve_pages = 0;
6003 enum zone_type i, j;
6005 for_each_online_pgdat(pgdat) {
6006 for (i = 0; i < MAX_NR_ZONES; i++) {
6007 struct zone *zone = pgdat->node_zones + i;
6010 /* Find valid and maximum lowmem_reserve in the zone */
6011 for (j = i; j < MAX_NR_ZONES; j++) {
6012 if (zone->lowmem_reserve[j] > max)
6013 max = zone->lowmem_reserve[j];
6016 /* we treat the high watermark as reserved pages. */
6017 max += high_wmark_pages(zone);
6019 if (max > zone->managed_pages)
6020 max = zone->managed_pages;
6021 reserve_pages += max;
6023 * Lowmem reserves are not available to
6024 * GFP_HIGHUSER page cache allocations and
6025 * kswapd tries to balance zones to their high
6026 * watermark. As a result, neither should be
6027 * regarded as dirtyable memory, to prevent a
6028 * situation where reclaim has to clean pages
6029 * in order to balance the zones.
6031 zone->dirty_balance_reserve = max;
6034 dirty_balance_reserve = reserve_pages;
6035 totalreserve_pages = reserve_pages;
6039 * setup_per_zone_lowmem_reserve - called whenever
6040 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6041 * has a correct pages reserved value, so an adequate number of
6042 * pages are left in the zone after a successful __alloc_pages().
6044 static void setup_per_zone_lowmem_reserve(void)
6046 struct pglist_data *pgdat;
6047 enum zone_type j, idx;
6049 for_each_online_pgdat(pgdat) {
6050 for (j = 0; j < MAX_NR_ZONES; j++) {
6051 struct zone *zone = pgdat->node_zones + j;
6052 unsigned long managed_pages = zone->managed_pages;
6054 zone->lowmem_reserve[j] = 0;
6058 struct zone *lower_zone;
6062 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6063 sysctl_lowmem_reserve_ratio[idx] = 1;
6065 lower_zone = pgdat->node_zones + idx;
6066 lower_zone->lowmem_reserve[j] = managed_pages /
6067 sysctl_lowmem_reserve_ratio[idx];
6068 managed_pages += lower_zone->managed_pages;
6073 /* update totalreserve_pages */
6074 calculate_totalreserve_pages();
6077 static void __setup_per_zone_wmarks(void)
6079 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6080 unsigned long lowmem_pages = 0;
6082 unsigned long flags;
6084 /* Calculate total number of !ZONE_HIGHMEM pages */
6085 for_each_zone(zone) {
6086 if (!is_highmem(zone))
6087 lowmem_pages += zone->managed_pages;
6090 for_each_zone(zone) {
6093 spin_lock_irqsave(&zone->lock, flags);
6094 tmp = (u64)pages_min * zone->managed_pages;
6095 do_div(tmp, lowmem_pages);
6096 if (is_highmem(zone)) {
6098 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6099 * need highmem pages, so cap pages_min to a small
6102 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6103 * deltas control asynch page reclaim, and so should
6104 * not be capped for highmem.
6106 unsigned long min_pages;
6108 min_pages = zone->managed_pages / 1024;
6109 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6110 zone->watermark[WMARK_MIN] = min_pages;
6113 * If it's a lowmem zone, reserve a number of pages
6114 * proportionate to the zone's size.
6116 zone->watermark[WMARK_MIN] = tmp;
6119 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
6120 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6122 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6123 high_wmark_pages(zone) - low_wmark_pages(zone) -
6124 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6126 spin_unlock_irqrestore(&zone->lock, flags);
6129 /* update totalreserve_pages */
6130 calculate_totalreserve_pages();
6134 * setup_per_zone_wmarks - called when min_free_kbytes changes
6135 * or when memory is hot-{added|removed}
6137 * Ensures that the watermark[min,low,high] values for each zone are set
6138 * correctly with respect to min_free_kbytes.
6140 void setup_per_zone_wmarks(void)
6142 mutex_lock(&zonelists_mutex);
6143 __setup_per_zone_wmarks();
6144 mutex_unlock(&zonelists_mutex);
6148 * The inactive anon list should be small enough that the VM never has to
6149 * do too much work, but large enough that each inactive page has a chance
6150 * to be referenced again before it is swapped out.
6152 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6153 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6154 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6155 * the anonymous pages are kept on the inactive list.
6158 * memory ratio inactive anon
6159 * -------------------------------------
6168 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6170 unsigned int gb, ratio;
6172 /* Zone size in gigabytes */
6173 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6175 ratio = int_sqrt(10 * gb);
6179 zone->inactive_ratio = ratio;
6182 static void __meminit setup_per_zone_inactive_ratio(void)
6187 calculate_zone_inactive_ratio(zone);
6191 * Initialise min_free_kbytes.
6193 * For small machines we want it small (128k min). For large machines
6194 * we want it large (64MB max). But it is not linear, because network
6195 * bandwidth does not increase linearly with machine size. We use
6197 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6198 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6214 int __meminit init_per_zone_wmark_min(void)
6216 unsigned long lowmem_kbytes;
6217 int new_min_free_kbytes;
6219 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6220 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6222 if (new_min_free_kbytes > user_min_free_kbytes) {
6223 min_free_kbytes = new_min_free_kbytes;
6224 if (min_free_kbytes < 128)
6225 min_free_kbytes = 128;
6226 if (min_free_kbytes > 65536)
6227 min_free_kbytes = 65536;
6229 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6230 new_min_free_kbytes, user_min_free_kbytes);
6232 setup_per_zone_wmarks();
6233 refresh_zone_stat_thresholds();
6234 setup_per_zone_lowmem_reserve();
6235 setup_per_zone_inactive_ratio();
6238 module_init(init_per_zone_wmark_min)
6241 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6242 * that we can call two helper functions whenever min_free_kbytes
6245 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6246 void __user *buffer, size_t *length, loff_t *ppos)
6250 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6255 user_min_free_kbytes = min_free_kbytes;
6256 setup_per_zone_wmarks();
6262 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6263 void __user *buffer, size_t *length, loff_t *ppos)
6268 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6273 zone->min_unmapped_pages = (zone->managed_pages *
6274 sysctl_min_unmapped_ratio) / 100;
6278 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6279 void __user *buffer, size_t *length, loff_t *ppos)
6284 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6289 zone->min_slab_pages = (zone->managed_pages *
6290 sysctl_min_slab_ratio) / 100;
6296 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6297 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6298 * whenever sysctl_lowmem_reserve_ratio changes.
6300 * The reserve ratio obviously has absolutely no relation with the
6301 * minimum watermarks. The lowmem reserve ratio can only make sense
6302 * if in function of the boot time zone sizes.
6304 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6305 void __user *buffer, size_t *length, loff_t *ppos)
6307 proc_dointvec_minmax(table, write, buffer, length, ppos);
6308 setup_per_zone_lowmem_reserve();
6313 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6314 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6315 * pagelist can have before it gets flushed back to buddy allocator.
6317 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6318 void __user *buffer, size_t *length, loff_t *ppos)
6321 int old_percpu_pagelist_fraction;
6324 mutex_lock(&pcp_batch_high_lock);
6325 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6327 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6328 if (!write || ret < 0)
6331 /* Sanity checking to avoid pcp imbalance */
6332 if (percpu_pagelist_fraction &&
6333 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6334 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6340 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6343 for_each_populated_zone(zone) {
6346 for_each_possible_cpu(cpu)
6347 pageset_set_high_and_batch(zone,
6348 per_cpu_ptr(zone->pageset, cpu));
6351 mutex_unlock(&pcp_batch_high_lock);
6356 int hashdist = HASHDIST_DEFAULT;
6358 static int __init set_hashdist(char *str)
6362 hashdist = simple_strtoul(str, &str, 0);
6365 __setup("hashdist=", set_hashdist);
6369 * allocate a large system hash table from bootmem
6370 * - it is assumed that the hash table must contain an exact power-of-2
6371 * quantity of entries
6372 * - limit is the number of hash buckets, not the total allocation size
6374 void *__init alloc_large_system_hash(const char *tablename,
6375 unsigned long bucketsize,
6376 unsigned long numentries,
6379 unsigned int *_hash_shift,
6380 unsigned int *_hash_mask,
6381 unsigned long low_limit,
6382 unsigned long high_limit)
6384 unsigned long long max = high_limit;
6385 unsigned long log2qty, size;
6388 /* allow the kernel cmdline to have a say */
6390 /* round applicable memory size up to nearest megabyte */
6391 numentries = nr_kernel_pages;
6393 /* It isn't necessary when PAGE_SIZE >= 1MB */
6394 if (PAGE_SHIFT < 20)
6395 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6397 /* limit to 1 bucket per 2^scale bytes of low memory */
6398 if (scale > PAGE_SHIFT)
6399 numentries >>= (scale - PAGE_SHIFT);
6401 numentries <<= (PAGE_SHIFT - scale);
6403 /* Make sure we've got at least a 0-order allocation.. */
6404 if (unlikely(flags & HASH_SMALL)) {
6405 /* Makes no sense without HASH_EARLY */
6406 WARN_ON(!(flags & HASH_EARLY));
6407 if (!(numentries >> *_hash_shift)) {
6408 numentries = 1UL << *_hash_shift;
6409 BUG_ON(!numentries);
6411 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6412 numentries = PAGE_SIZE / bucketsize;
6414 numentries = roundup_pow_of_two(numentries);
6416 /* limit allocation size to 1/16 total memory by default */
6418 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6419 do_div(max, bucketsize);
6421 max = min(max, 0x80000000ULL);
6423 if (numentries < low_limit)
6424 numentries = low_limit;
6425 if (numentries > max)
6428 log2qty = ilog2(numentries);
6431 size = bucketsize << log2qty;
6432 if (flags & HASH_EARLY)
6433 table = memblock_virt_alloc_nopanic(size, 0);
6435 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6438 * If bucketsize is not a power-of-two, we may free
6439 * some pages at the end of hash table which
6440 * alloc_pages_exact() automatically does
6442 if (get_order(size) < MAX_ORDER) {
6443 table = alloc_pages_exact(size, GFP_ATOMIC);
6444 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6447 } while (!table && size > PAGE_SIZE && --log2qty);
6450 panic("Failed to allocate %s hash table\n", tablename);
6452 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6455 ilog2(size) - PAGE_SHIFT,
6459 *_hash_shift = log2qty;
6461 *_hash_mask = (1 << log2qty) - 1;
6466 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6467 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6470 #ifdef CONFIG_SPARSEMEM
6471 return __pfn_to_section(pfn)->pageblock_flags;
6473 return zone->pageblock_flags;
6474 #endif /* CONFIG_SPARSEMEM */
6477 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6479 #ifdef CONFIG_SPARSEMEM
6480 pfn &= (PAGES_PER_SECTION-1);
6481 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6483 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6484 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6485 #endif /* CONFIG_SPARSEMEM */
6489 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6490 * @page: The page within the block of interest
6491 * @pfn: The target page frame number
6492 * @end_bitidx: The last bit of interest to retrieve
6493 * @mask: mask of bits that the caller is interested in
6495 * Return: pageblock_bits flags
6497 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6498 unsigned long end_bitidx,
6502 unsigned long *bitmap;
6503 unsigned long bitidx, word_bitidx;
6506 zone = page_zone(page);
6507 bitmap = get_pageblock_bitmap(zone, pfn);
6508 bitidx = pfn_to_bitidx(zone, pfn);
6509 word_bitidx = bitidx / BITS_PER_LONG;
6510 bitidx &= (BITS_PER_LONG-1);
6512 word = bitmap[word_bitidx];
6513 bitidx += end_bitidx;
6514 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6518 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6519 * @page: The page within the block of interest
6520 * @flags: The flags to set
6521 * @pfn: The target page frame number
6522 * @end_bitidx: The last bit of interest
6523 * @mask: mask of bits that the caller is interested in
6525 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6527 unsigned long end_bitidx,
6531 unsigned long *bitmap;
6532 unsigned long bitidx, word_bitidx;
6533 unsigned long old_word, word;
6535 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6537 zone = page_zone(page);
6538 bitmap = get_pageblock_bitmap(zone, pfn);
6539 bitidx = pfn_to_bitidx(zone, pfn);
6540 word_bitidx = bitidx / BITS_PER_LONG;
6541 bitidx &= (BITS_PER_LONG-1);
6543 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6545 bitidx += end_bitidx;
6546 mask <<= (BITS_PER_LONG - bitidx - 1);
6547 flags <<= (BITS_PER_LONG - bitidx - 1);
6549 word = READ_ONCE(bitmap[word_bitidx]);
6551 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6552 if (word == old_word)
6559 * This function checks whether pageblock includes unmovable pages or not.
6560 * If @count is not zero, it is okay to include less @count unmovable pages
6562 * PageLRU check without isolation or lru_lock could race so that
6563 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6564 * expect this function should be exact.
6566 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6567 bool skip_hwpoisoned_pages)
6569 unsigned long pfn, iter, found;
6573 * For avoiding noise data, lru_add_drain_all() should be called
6574 * If ZONE_MOVABLE, the zone never contains unmovable pages
6576 if (zone_idx(zone) == ZONE_MOVABLE)
6578 mt = get_pageblock_migratetype(page);
6579 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6582 pfn = page_to_pfn(page);
6583 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6584 unsigned long check = pfn + iter;
6586 if (!pfn_valid_within(check))
6589 page = pfn_to_page(check);
6592 * Hugepages are not in LRU lists, but they're movable.
6593 * We need not scan over tail pages bacause we don't
6594 * handle each tail page individually in migration.
6596 if (PageHuge(page)) {
6597 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6602 * We can't use page_count without pin a page
6603 * because another CPU can free compound page.
6604 * This check already skips compound tails of THP
6605 * because their page->_count is zero at all time.
6607 if (!atomic_read(&page->_count)) {
6608 if (PageBuddy(page))
6609 iter += (1 << page_order(page)) - 1;
6614 * The HWPoisoned page may be not in buddy system, and
6615 * page_count() is not 0.
6617 if (skip_hwpoisoned_pages && PageHWPoison(page))
6623 * If there are RECLAIMABLE pages, we need to check
6624 * it. But now, memory offline itself doesn't call
6625 * shrink_node_slabs() and it still to be fixed.
6628 * If the page is not RAM, page_count()should be 0.
6629 * we don't need more check. This is an _used_ not-movable page.
6631 * The problematic thing here is PG_reserved pages. PG_reserved
6632 * is set to both of a memory hole page and a _used_ kernel
6641 bool is_pageblock_removable_nolock(struct page *page)
6647 * We have to be careful here because we are iterating over memory
6648 * sections which are not zone aware so we might end up outside of
6649 * the zone but still within the section.
6650 * We have to take care about the node as well. If the node is offline
6651 * its NODE_DATA will be NULL - see page_zone.
6653 if (!node_online(page_to_nid(page)))
6656 zone = page_zone(page);
6657 pfn = page_to_pfn(page);
6658 if (!zone_spans_pfn(zone, pfn))
6661 return !has_unmovable_pages(zone, page, 0, true);
6666 static unsigned long pfn_max_align_down(unsigned long pfn)
6668 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6669 pageblock_nr_pages) - 1);
6672 static unsigned long pfn_max_align_up(unsigned long pfn)
6674 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6675 pageblock_nr_pages));
6678 /* [start, end) must belong to a single zone. */
6679 static int __alloc_contig_migrate_range(struct compact_control *cc,
6680 unsigned long start, unsigned long end)
6682 /* This function is based on compact_zone() from compaction.c. */
6683 unsigned long nr_reclaimed;
6684 unsigned long pfn = start;
6685 unsigned int tries = 0;
6690 while (pfn < end || !list_empty(&cc->migratepages)) {
6691 if (fatal_signal_pending(current)) {
6696 if (list_empty(&cc->migratepages)) {
6697 cc->nr_migratepages = 0;
6698 pfn = isolate_migratepages_range(cc, pfn, end);
6704 } else if (++tries == 5) {
6705 ret = ret < 0 ? ret : -EBUSY;
6709 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6711 cc->nr_migratepages -= nr_reclaimed;
6713 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6714 NULL, 0, cc->mode, MR_CMA);
6717 putback_movable_pages(&cc->migratepages);
6724 * alloc_contig_range() -- tries to allocate given range of pages
6725 * @start: start PFN to allocate
6726 * @end: one-past-the-last PFN to allocate
6727 * @migratetype: migratetype of the underlaying pageblocks (either
6728 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6729 * in range must have the same migratetype and it must
6730 * be either of the two.
6732 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6733 * aligned, however it's the caller's responsibility to guarantee that
6734 * we are the only thread that changes migrate type of pageblocks the
6737 * The PFN range must belong to a single zone.
6739 * Returns zero on success or negative error code. On success all
6740 * pages which PFN is in [start, end) are allocated for the caller and
6741 * need to be freed with free_contig_range().
6743 int alloc_contig_range(unsigned long start, unsigned long end,
6744 unsigned migratetype)
6746 unsigned long outer_start, outer_end;
6750 struct compact_control cc = {
6751 .nr_migratepages = 0,
6753 .zone = page_zone(pfn_to_page(start)),
6754 .mode = MIGRATE_SYNC,
6755 .ignore_skip_hint = true,
6757 INIT_LIST_HEAD(&cc.migratepages);
6760 * What we do here is we mark all pageblocks in range as
6761 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6762 * have different sizes, and due to the way page allocator
6763 * work, we align the range to biggest of the two pages so
6764 * that page allocator won't try to merge buddies from
6765 * different pageblocks and change MIGRATE_ISOLATE to some
6766 * other migration type.
6768 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6769 * migrate the pages from an unaligned range (ie. pages that
6770 * we are interested in). This will put all the pages in
6771 * range back to page allocator as MIGRATE_ISOLATE.
6773 * When this is done, we take the pages in range from page
6774 * allocator removing them from the buddy system. This way
6775 * page allocator will never consider using them.
6777 * This lets us mark the pageblocks back as
6778 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6779 * aligned range but not in the unaligned, original range are
6780 * put back to page allocator so that buddy can use them.
6783 ret = start_isolate_page_range(pfn_max_align_down(start),
6784 pfn_max_align_up(end), migratetype,
6789 ret = __alloc_contig_migrate_range(&cc, start, end);
6794 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6795 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6796 * more, all pages in [start, end) are free in page allocator.
6797 * What we are going to do is to allocate all pages from
6798 * [start, end) (that is remove them from page allocator).
6800 * The only problem is that pages at the beginning and at the
6801 * end of interesting range may be not aligned with pages that
6802 * page allocator holds, ie. they can be part of higher order
6803 * pages. Because of this, we reserve the bigger range and
6804 * once this is done free the pages we are not interested in.
6806 * We don't have to hold zone->lock here because the pages are
6807 * isolated thus they won't get removed from buddy.
6810 lru_add_drain_all();
6811 drain_all_pages(cc.zone);
6814 outer_start = start;
6815 while (!PageBuddy(pfn_to_page(outer_start))) {
6816 if (++order >= MAX_ORDER) {
6820 outer_start &= ~0UL << order;
6823 /* Make sure the range is really isolated. */
6824 if (test_pages_isolated(outer_start, end, false)) {
6825 pr_info("%s: [%lx, %lx) PFNs busy\n",
6826 __func__, outer_start, end);
6831 /* Grab isolated pages from freelists. */
6832 outer_end = isolate_freepages_range(&cc, outer_start, end);
6838 /* Free head and tail (if any) */
6839 if (start != outer_start)
6840 free_contig_range(outer_start, start - outer_start);
6841 if (end != outer_end)
6842 free_contig_range(end, outer_end - end);
6845 undo_isolate_page_range(pfn_max_align_down(start),
6846 pfn_max_align_up(end), migratetype);
6850 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6852 unsigned int count = 0;
6854 for (; nr_pages--; pfn++) {
6855 struct page *page = pfn_to_page(pfn);
6857 count += page_count(page) != 1;
6860 WARN(count != 0, "%d pages are still in use!\n", count);
6864 #ifdef CONFIG_MEMORY_HOTPLUG
6866 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6867 * page high values need to be recalulated.
6869 void __meminit zone_pcp_update(struct zone *zone)
6872 mutex_lock(&pcp_batch_high_lock);
6873 for_each_possible_cpu(cpu)
6874 pageset_set_high_and_batch(zone,
6875 per_cpu_ptr(zone->pageset, cpu));
6876 mutex_unlock(&pcp_batch_high_lock);
6880 void zone_pcp_reset(struct zone *zone)
6882 unsigned long flags;
6884 struct per_cpu_pageset *pset;
6886 /* avoid races with drain_pages() */
6887 local_lock_irqsave(pa_lock, flags);
6888 if (zone->pageset != &boot_pageset) {
6889 for_each_online_cpu(cpu) {
6890 pset = per_cpu_ptr(zone->pageset, cpu);
6891 drain_zonestat(zone, pset);
6893 free_percpu(zone->pageset);
6894 zone->pageset = &boot_pageset;
6896 local_unlock_irqrestore(pa_lock, flags);
6899 #ifdef CONFIG_MEMORY_HOTREMOVE
6901 * All pages in the range must be isolated before calling this.
6904 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6908 unsigned int order, i;
6910 unsigned long flags;
6911 /* find the first valid pfn */
6912 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6917 zone = page_zone(pfn_to_page(pfn));
6918 spin_lock_irqsave(&zone->lock, flags);
6920 while (pfn < end_pfn) {
6921 if (!pfn_valid(pfn)) {
6925 page = pfn_to_page(pfn);
6927 * The HWPoisoned page may be not in buddy system, and
6928 * page_count() is not 0.
6930 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6932 SetPageReserved(page);
6936 BUG_ON(page_count(page));
6937 BUG_ON(!PageBuddy(page));
6938 order = page_order(page);
6939 #ifdef CONFIG_DEBUG_VM
6940 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6941 pfn, 1 << order, end_pfn);
6943 list_del(&page->lru);
6944 rmv_page_order(page);
6945 zone->free_area[order].nr_free--;
6946 for (i = 0; i < (1 << order); i++)
6947 SetPageReserved((page+i));
6948 pfn += (1 << order);
6950 spin_unlock_irqrestore(&zone->lock, flags);
6954 #ifdef CONFIG_MEMORY_FAILURE
6955 bool is_free_buddy_page(struct page *page)
6957 struct zone *zone = page_zone(page);
6958 unsigned long pfn = page_to_pfn(page);
6959 unsigned long flags;
6962 spin_lock_irqsave(&zone->lock, flags);
6963 for (order = 0; order < MAX_ORDER; order++) {
6964 struct page *page_head = page - (pfn & ((1 << order) - 1));
6966 if (PageBuddy(page_head) && page_order(page_head) >= order)
6969 spin_unlock_irqrestore(&zone->lock, flags);
6971 return order < MAX_ORDER;