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/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/ftrace_event.h>
57 #include <linux/memcontrol.h>
58 #include <linux/prefetch.h>
59 #include <linux/migrate.h>
60 #include <linux/page-debug-flags.h>
61 #include <linux/sched/rt.h>
63 #include <asm/tlbflush.h>
64 #include <asm/div64.h>
67 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
68 DEFINE_PER_CPU(int, numa_node);
69 EXPORT_PER_CPU_SYMBOL(numa_node);
72 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
74 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
75 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
76 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
77 * defined in <linux/topology.h>.
79 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
80 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
84 * Array of node states.
86 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
87 [N_POSSIBLE] = NODE_MASK_ALL,
88 [N_ONLINE] = { { [0] = 1UL } },
90 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
92 [N_HIGH_MEMORY] = { { [0] = 1UL } },
94 #ifdef CONFIG_MOVABLE_NODE
95 [N_MEMORY] = { { [0] = 1UL } },
97 [N_CPU] = { { [0] = 1UL } },
100 EXPORT_SYMBOL(node_states);
102 unsigned long totalram_pages __read_mostly;
103 unsigned long totalreserve_pages __read_mostly;
105 * When calculating the number of globally allowed dirty pages, there
106 * is a certain number of per-zone reserves that should not be
107 * considered dirtyable memory. This is the sum of those reserves
108 * over all existing zones that contribute dirtyable memory.
110 unsigned long dirty_balance_reserve __read_mostly;
112 int percpu_pagelist_fraction;
113 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
115 #ifdef CONFIG_PM_SLEEP
117 * The following functions are used by the suspend/hibernate code to temporarily
118 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
119 * while devices are suspended. To avoid races with the suspend/hibernate code,
120 * they should always be called with pm_mutex held (gfp_allowed_mask also should
121 * only be modified with pm_mutex held, unless the suspend/hibernate code is
122 * guaranteed not to run in parallel with that modification).
125 static gfp_t saved_gfp_mask;
127 void pm_restore_gfp_mask(void)
129 WARN_ON(!mutex_is_locked(&pm_mutex));
130 if (saved_gfp_mask) {
131 gfp_allowed_mask = saved_gfp_mask;
136 void pm_restrict_gfp_mask(void)
138 WARN_ON(!mutex_is_locked(&pm_mutex));
139 WARN_ON(saved_gfp_mask);
140 saved_gfp_mask = gfp_allowed_mask;
141 gfp_allowed_mask &= ~GFP_IOFS;
144 bool pm_suspended_storage(void)
146 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
150 #endif /* CONFIG_PM_SLEEP */
152 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
153 int pageblock_order __read_mostly;
156 static void __free_pages_ok(struct page *page, unsigned int order);
159 * results with 256, 32 in the lowmem_reserve sysctl:
160 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
161 * 1G machine -> (16M dma, 784M normal, 224M high)
162 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
163 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
164 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
166 * TBD: should special case ZONE_DMA32 machines here - in those we normally
167 * don't need any ZONE_NORMAL reservation
169 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
170 #ifdef CONFIG_ZONE_DMA
173 #ifdef CONFIG_ZONE_DMA32
176 #ifdef CONFIG_HIGHMEM
182 EXPORT_SYMBOL(totalram_pages);
184 static char * const zone_names[MAX_NR_ZONES] = {
185 #ifdef CONFIG_ZONE_DMA
188 #ifdef CONFIG_ZONE_DMA32
192 #ifdef CONFIG_HIGHMEM
198 int min_free_kbytes = 1024;
200 static unsigned long __meminitdata nr_kernel_pages;
201 static unsigned long __meminitdata nr_all_pages;
202 static unsigned long __meminitdata dma_reserve;
204 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
205 /* Movable memory ranges, will also be used by memblock subsystem. */
206 struct movablemem_map movablemem_map;
208 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
209 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
210 static unsigned long __initdata required_kernelcore;
211 static unsigned long __initdata required_movablecore;
212 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
213 static unsigned long __meminitdata zone_movable_limit[MAX_NUMNODES];
215 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
217 EXPORT_SYMBOL(movable_zone);
218 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
221 int nr_node_ids __read_mostly = MAX_NUMNODES;
222 int nr_online_nodes __read_mostly = 1;
223 EXPORT_SYMBOL(nr_node_ids);
224 EXPORT_SYMBOL(nr_online_nodes);
227 int page_group_by_mobility_disabled __read_mostly;
229 void set_pageblock_migratetype(struct page *page, int migratetype)
232 if (unlikely(page_group_by_mobility_disabled))
233 migratetype = MIGRATE_UNMOVABLE;
235 set_pageblock_flags_group(page, (unsigned long)migratetype,
236 PB_migrate, PB_migrate_end);
239 bool oom_killer_disabled __read_mostly;
241 #ifdef CONFIG_DEBUG_VM
242 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
246 unsigned long pfn = page_to_pfn(page);
249 seq = zone_span_seqbegin(zone);
250 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
252 else if (pfn < zone->zone_start_pfn)
254 } while (zone_span_seqretry(zone, seq));
259 static int page_is_consistent(struct zone *zone, struct page *page)
261 if (!pfn_valid_within(page_to_pfn(page)))
263 if (zone != page_zone(page))
269 * Temporary debugging check for pages not lying within a given zone.
271 static int bad_range(struct zone *zone, struct page *page)
273 if (page_outside_zone_boundaries(zone, page))
275 if (!page_is_consistent(zone, page))
281 static inline int bad_range(struct zone *zone, struct page *page)
287 static void bad_page(struct page *page)
289 static unsigned long resume;
290 static unsigned long nr_shown;
291 static unsigned long nr_unshown;
293 /* Don't complain about poisoned pages */
294 if (PageHWPoison(page)) {
295 reset_page_mapcount(page); /* remove PageBuddy */
300 * Allow a burst of 60 reports, then keep quiet for that minute;
301 * or allow a steady drip of one report per second.
303 if (nr_shown == 60) {
304 if (time_before(jiffies, resume)) {
310 "BUG: Bad page state: %lu messages suppressed\n",
317 resume = jiffies + 60 * HZ;
319 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
320 current->comm, page_to_pfn(page));
326 /* Leave bad fields for debug, except PageBuddy could make trouble */
327 reset_page_mapcount(page); /* remove PageBuddy */
328 add_taint(TAINT_BAD_PAGE);
332 * Higher-order pages are called "compound pages". They are structured thusly:
334 * The first PAGE_SIZE page is called the "head page".
336 * The remaining PAGE_SIZE pages are called "tail pages".
338 * All pages have PG_compound set. All tail pages have their ->first_page
339 * pointing at the head page.
341 * The first tail page's ->lru.next holds the address of the compound page's
342 * put_page() function. Its ->lru.prev holds the order of allocation.
343 * This usage means that zero-order pages may not be compound.
346 static void free_compound_page(struct page *page)
348 __free_pages_ok(page, compound_order(page));
351 void prep_compound_page(struct page *page, unsigned long order)
354 int nr_pages = 1 << order;
356 set_compound_page_dtor(page, free_compound_page);
357 set_compound_order(page, order);
359 for (i = 1; i < nr_pages; i++) {
360 struct page *p = page + i;
362 set_page_count(p, 0);
363 p->first_page = page;
367 /* update __split_huge_page_refcount if you change this function */
368 static int destroy_compound_page(struct page *page, unsigned long order)
371 int nr_pages = 1 << order;
374 if (unlikely(compound_order(page) != order)) {
379 __ClearPageHead(page);
381 for (i = 1; i < nr_pages; i++) {
382 struct page *p = page + i;
384 if (unlikely(!PageTail(p) || (p->first_page != page))) {
394 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
399 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
400 * and __GFP_HIGHMEM from hard or soft interrupt context.
402 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
403 for (i = 0; i < (1 << order); i++)
404 clear_highpage(page + i);
407 #ifdef CONFIG_DEBUG_PAGEALLOC
408 unsigned int _debug_guardpage_minorder;
410 static int __init debug_guardpage_minorder_setup(char *buf)
414 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
415 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
418 _debug_guardpage_minorder = res;
419 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
422 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
424 static inline void set_page_guard_flag(struct page *page)
426 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
429 static inline void clear_page_guard_flag(struct page *page)
431 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
434 static inline void set_page_guard_flag(struct page *page) { }
435 static inline void clear_page_guard_flag(struct page *page) { }
438 static inline void set_page_order(struct page *page, int order)
440 set_page_private(page, order);
441 __SetPageBuddy(page);
444 static inline void rmv_page_order(struct page *page)
446 __ClearPageBuddy(page);
447 set_page_private(page, 0);
451 * Locate the struct page for both the matching buddy in our
452 * pair (buddy1) and the combined O(n+1) page they form (page).
454 * 1) Any buddy B1 will have an order O twin B2 which satisfies
455 * the following equation:
457 * For example, if the starting buddy (buddy2) is #8 its order
459 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
461 * 2) Any buddy B will have an order O+1 parent P which
462 * satisfies the following equation:
465 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
467 static inline unsigned long
468 __find_buddy_index(unsigned long page_idx, unsigned int order)
470 return page_idx ^ (1 << order);
474 * This function checks whether a page is free && is the buddy
475 * we can do coalesce a page and its buddy if
476 * (a) the buddy is not in a hole &&
477 * (b) the buddy is in the buddy system &&
478 * (c) a page and its buddy have the same order &&
479 * (d) a page and its buddy are in the same zone.
481 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
482 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
484 * For recording page's order, we use page_private(page).
486 static inline int page_is_buddy(struct page *page, struct page *buddy,
489 if (!pfn_valid_within(page_to_pfn(buddy)))
492 if (page_zone_id(page) != page_zone_id(buddy))
495 if (page_is_guard(buddy) && page_order(buddy) == order) {
496 VM_BUG_ON(page_count(buddy) != 0);
500 if (PageBuddy(buddy) && page_order(buddy) == order) {
501 VM_BUG_ON(page_count(buddy) != 0);
508 * Freeing function for a buddy system allocator.
510 * The concept of a buddy system is to maintain direct-mapped table
511 * (containing bit values) for memory blocks of various "orders".
512 * The bottom level table contains the map for the smallest allocatable
513 * units of memory (here, pages), and each level above it describes
514 * pairs of units from the levels below, hence, "buddies".
515 * At a high level, all that happens here is marking the table entry
516 * at the bottom level available, and propagating the changes upward
517 * as necessary, plus some accounting needed to play nicely with other
518 * parts of the VM system.
519 * At each level, we keep a list of pages, which are heads of continuous
520 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
521 * order is recorded in page_private(page) field.
522 * So when we are allocating or freeing one, we can derive the state of the
523 * other. That is, if we allocate a small block, and both were
524 * free, the remainder of the region must be split into blocks.
525 * If a block is freed, and its buddy is also free, then this
526 * triggers coalescing into a block of larger size.
531 static inline void __free_one_page(struct page *page,
532 struct zone *zone, unsigned int order,
535 unsigned long page_idx;
536 unsigned long combined_idx;
537 unsigned long uninitialized_var(buddy_idx);
540 if (unlikely(PageCompound(page)))
541 if (unlikely(destroy_compound_page(page, order)))
544 VM_BUG_ON(migratetype == -1);
546 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
548 VM_BUG_ON(page_idx & ((1 << order) - 1));
549 VM_BUG_ON(bad_range(zone, page));
551 while (order < MAX_ORDER-1) {
552 buddy_idx = __find_buddy_index(page_idx, order);
553 buddy = page + (buddy_idx - page_idx);
554 if (!page_is_buddy(page, buddy, order))
557 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
558 * merge with it and move up one order.
560 if (page_is_guard(buddy)) {
561 clear_page_guard_flag(buddy);
562 set_page_private(page, 0);
563 __mod_zone_freepage_state(zone, 1 << order,
566 list_del(&buddy->lru);
567 zone->free_area[order].nr_free--;
568 rmv_page_order(buddy);
570 combined_idx = buddy_idx & page_idx;
571 page = page + (combined_idx - page_idx);
572 page_idx = combined_idx;
575 set_page_order(page, order);
578 * If this is not the largest possible page, check if the buddy
579 * of the next-highest order is free. If it is, it's possible
580 * that pages are being freed that will coalesce soon. In case,
581 * that is happening, add the free page to the tail of the list
582 * so it's less likely to be used soon and more likely to be merged
583 * as a higher order page
585 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
586 struct page *higher_page, *higher_buddy;
587 combined_idx = buddy_idx & page_idx;
588 higher_page = page + (combined_idx - page_idx);
589 buddy_idx = __find_buddy_index(combined_idx, order + 1);
590 higher_buddy = higher_page + (buddy_idx - combined_idx);
591 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
592 list_add_tail(&page->lru,
593 &zone->free_area[order].free_list[migratetype]);
598 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
600 zone->free_area[order].nr_free++;
603 static inline int free_pages_check(struct page *page)
605 if (unlikely(page_mapcount(page) |
606 (page->mapping != NULL) |
607 (atomic_read(&page->_count) != 0) |
608 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
609 (mem_cgroup_bad_page_check(page)))) {
613 reset_page_last_nid(page);
614 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
615 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
620 * Frees a number of pages from the PCP lists
621 * Assumes all pages on list are in same zone, and of same order.
622 * count is the number of pages to free.
624 * If the zone was previously in an "all pages pinned" state then look to
625 * see if this freeing clears that state.
627 * And clear the zone's pages_scanned counter, to hold off the "all pages are
628 * pinned" detection logic.
630 static void free_pcppages_bulk(struct zone *zone, int count,
631 struct per_cpu_pages *pcp)
637 spin_lock(&zone->lock);
638 zone->all_unreclaimable = 0;
639 zone->pages_scanned = 0;
643 struct list_head *list;
646 * Remove pages from lists in a round-robin fashion. A
647 * batch_free count is maintained that is incremented when an
648 * empty list is encountered. This is so more pages are freed
649 * off fuller lists instead of spinning excessively around empty
654 if (++migratetype == MIGRATE_PCPTYPES)
656 list = &pcp->lists[migratetype];
657 } while (list_empty(list));
659 /* This is the only non-empty list. Free them all. */
660 if (batch_free == MIGRATE_PCPTYPES)
661 batch_free = to_free;
664 int mt; /* migratetype of the to-be-freed page */
666 page = list_entry(list->prev, struct page, lru);
667 /* must delete as __free_one_page list manipulates */
668 list_del(&page->lru);
669 mt = get_freepage_migratetype(page);
670 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
671 __free_one_page(page, zone, 0, mt);
672 trace_mm_page_pcpu_drain(page, 0, mt);
673 if (likely(get_pageblock_migratetype(page) != MIGRATE_ISOLATE)) {
674 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
675 if (is_migrate_cma(mt))
676 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
678 } while (--to_free && --batch_free && !list_empty(list));
680 spin_unlock(&zone->lock);
683 static void free_one_page(struct zone *zone, struct page *page, int order,
686 spin_lock(&zone->lock);
687 zone->all_unreclaimable = 0;
688 zone->pages_scanned = 0;
690 __free_one_page(page, zone, order, migratetype);
691 if (unlikely(migratetype != MIGRATE_ISOLATE))
692 __mod_zone_freepage_state(zone, 1 << order, migratetype);
693 spin_unlock(&zone->lock);
696 static bool free_pages_prepare(struct page *page, unsigned int order)
701 trace_mm_page_free(page, order);
702 kmemcheck_free_shadow(page, order);
705 page->mapping = NULL;
706 for (i = 0; i < (1 << order); i++)
707 bad += free_pages_check(page + i);
711 if (!PageHighMem(page)) {
712 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
713 debug_check_no_obj_freed(page_address(page),
716 arch_free_page(page, order);
717 kernel_map_pages(page, 1 << order, 0);
722 static void __free_pages_ok(struct page *page, unsigned int order)
727 if (!free_pages_prepare(page, order))
730 local_irq_save(flags);
731 __count_vm_events(PGFREE, 1 << order);
732 migratetype = get_pageblock_migratetype(page);
733 set_freepage_migratetype(page, migratetype);
734 free_one_page(page_zone(page), page, order, migratetype);
735 local_irq_restore(flags);
739 * Read access to zone->managed_pages is safe because it's unsigned long,
740 * but we still need to serialize writers. Currently all callers of
741 * __free_pages_bootmem() except put_page_bootmem() should only be used
742 * at boot time. So for shorter boot time, we shift the burden to
743 * put_page_bootmem() to serialize writers.
745 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
747 unsigned int nr_pages = 1 << order;
751 for (loop = 0; loop < nr_pages; loop++) {
752 struct page *p = &page[loop];
754 if (loop + 1 < nr_pages)
756 __ClearPageReserved(p);
757 set_page_count(p, 0);
760 page_zone(page)->managed_pages += 1 << order;
761 set_page_refcounted(page);
762 __free_pages(page, order);
766 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
767 void __init init_cma_reserved_pageblock(struct page *page)
769 unsigned i = pageblock_nr_pages;
770 struct page *p = page;
773 __ClearPageReserved(p);
774 set_page_count(p, 0);
777 set_page_refcounted(page);
778 set_pageblock_migratetype(page, MIGRATE_CMA);
779 __free_pages(page, pageblock_order);
780 totalram_pages += pageblock_nr_pages;
781 #ifdef CONFIG_HIGHMEM
782 if (PageHighMem(page))
783 totalhigh_pages += pageblock_nr_pages;
789 * The order of subdivision here is critical for the IO subsystem.
790 * Please do not alter this order without good reasons and regression
791 * testing. Specifically, as large blocks of memory are subdivided,
792 * the order in which smaller blocks are delivered depends on the order
793 * they're subdivided in this function. This is the primary factor
794 * influencing the order in which pages are delivered to the IO
795 * subsystem according to empirical testing, and this is also justified
796 * by considering the behavior of a buddy system containing a single
797 * large block of memory acted on by a series of small allocations.
798 * This behavior is a critical factor in sglist merging's success.
802 static inline void expand(struct zone *zone, struct page *page,
803 int low, int high, struct free_area *area,
806 unsigned long size = 1 << high;
812 VM_BUG_ON(bad_range(zone, &page[size]));
814 #ifdef CONFIG_DEBUG_PAGEALLOC
815 if (high < debug_guardpage_minorder()) {
817 * Mark as guard pages (or page), that will allow to
818 * merge back to allocator when buddy will be freed.
819 * Corresponding page table entries will not be touched,
820 * pages will stay not present in virtual address space
822 INIT_LIST_HEAD(&page[size].lru);
823 set_page_guard_flag(&page[size]);
824 set_page_private(&page[size], high);
825 /* Guard pages are not available for any usage */
826 __mod_zone_freepage_state(zone, -(1 << high),
831 list_add(&page[size].lru, &area->free_list[migratetype]);
833 set_page_order(&page[size], high);
838 * This page is about to be returned from the page allocator
840 static inline int check_new_page(struct page *page)
842 if (unlikely(page_mapcount(page) |
843 (page->mapping != NULL) |
844 (atomic_read(&page->_count) != 0) |
845 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
846 (mem_cgroup_bad_page_check(page)))) {
853 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
857 for (i = 0; i < (1 << order); i++) {
858 struct page *p = page + i;
859 if (unlikely(check_new_page(p)))
863 set_page_private(page, 0);
864 set_page_refcounted(page);
866 arch_alloc_page(page, order);
867 kernel_map_pages(page, 1 << order, 1);
869 if (gfp_flags & __GFP_ZERO)
870 prep_zero_page(page, order, gfp_flags);
872 if (order && (gfp_flags & __GFP_COMP))
873 prep_compound_page(page, order);
879 * Go through the free lists for the given migratetype and remove
880 * the smallest available page from the freelists
883 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
886 unsigned int current_order;
887 struct free_area * area;
890 /* Find a page of the appropriate size in the preferred list */
891 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
892 area = &(zone->free_area[current_order]);
893 if (list_empty(&area->free_list[migratetype]))
896 page = list_entry(area->free_list[migratetype].next,
898 list_del(&page->lru);
899 rmv_page_order(page);
901 expand(zone, page, order, current_order, area, migratetype);
910 * This array describes the order lists are fallen back to when
911 * the free lists for the desirable migrate type are depleted
913 static int fallbacks[MIGRATE_TYPES][4] = {
914 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
915 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
917 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
918 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
920 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
922 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
923 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
927 * Move the free pages in a range to the free lists of the requested type.
928 * Note that start_page and end_pages are not aligned on a pageblock
929 * boundary. If alignment is required, use move_freepages_block()
931 int move_freepages(struct zone *zone,
932 struct page *start_page, struct page *end_page,
939 #ifndef CONFIG_HOLES_IN_ZONE
941 * page_zone is not safe to call in this context when
942 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
943 * anyway as we check zone boundaries in move_freepages_block().
944 * Remove at a later date when no bug reports exist related to
945 * grouping pages by mobility
947 BUG_ON(page_zone(start_page) != page_zone(end_page));
950 for (page = start_page; page <= end_page;) {
951 /* Make sure we are not inadvertently changing nodes */
952 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
954 if (!pfn_valid_within(page_to_pfn(page))) {
959 if (!PageBuddy(page)) {
964 order = page_order(page);
965 list_move(&page->lru,
966 &zone->free_area[order].free_list[migratetype]);
967 set_freepage_migratetype(page, migratetype);
969 pages_moved += 1 << order;
975 int move_freepages_block(struct zone *zone, struct page *page,
978 unsigned long start_pfn, end_pfn;
979 struct page *start_page, *end_page;
981 start_pfn = page_to_pfn(page);
982 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
983 start_page = pfn_to_page(start_pfn);
984 end_page = start_page + pageblock_nr_pages - 1;
985 end_pfn = start_pfn + pageblock_nr_pages - 1;
987 /* Do not cross zone boundaries */
988 if (start_pfn < zone->zone_start_pfn)
990 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
993 return move_freepages(zone, start_page, end_page, migratetype);
996 static void change_pageblock_range(struct page *pageblock_page,
997 int start_order, int migratetype)
999 int nr_pageblocks = 1 << (start_order - pageblock_order);
1001 while (nr_pageblocks--) {
1002 set_pageblock_migratetype(pageblock_page, migratetype);
1003 pageblock_page += pageblock_nr_pages;
1007 /* Remove an element from the buddy allocator from the fallback list */
1008 static inline struct page *
1009 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1011 struct free_area * area;
1016 /* Find the largest possible block of pages in the other list */
1017 for (current_order = MAX_ORDER-1; current_order >= order;
1020 migratetype = fallbacks[start_migratetype][i];
1022 /* MIGRATE_RESERVE handled later if necessary */
1023 if (migratetype == MIGRATE_RESERVE)
1026 area = &(zone->free_area[current_order]);
1027 if (list_empty(&area->free_list[migratetype]))
1030 page = list_entry(area->free_list[migratetype].next,
1035 * If breaking a large block of pages, move all free
1036 * pages to the preferred allocation list. If falling
1037 * back for a reclaimable kernel allocation, be more
1038 * aggressive about taking ownership of free pages
1040 * On the other hand, never change migration
1041 * type of MIGRATE_CMA pageblocks nor move CMA
1042 * pages on different free lists. We don't
1043 * want unmovable pages to be allocated from
1044 * MIGRATE_CMA areas.
1046 if (!is_migrate_cma(migratetype) &&
1047 (unlikely(current_order >= pageblock_order / 2) ||
1048 start_migratetype == MIGRATE_RECLAIMABLE ||
1049 page_group_by_mobility_disabled)) {
1051 pages = move_freepages_block(zone, page,
1054 /* Claim the whole block if over half of it is free */
1055 if (pages >= (1 << (pageblock_order-1)) ||
1056 page_group_by_mobility_disabled)
1057 set_pageblock_migratetype(page,
1060 migratetype = start_migratetype;
1063 /* Remove the page from the freelists */
1064 list_del(&page->lru);
1065 rmv_page_order(page);
1067 /* Take ownership for orders >= pageblock_order */
1068 if (current_order >= pageblock_order &&
1069 !is_migrate_cma(migratetype))
1070 change_pageblock_range(page, current_order,
1073 expand(zone, page, order, current_order, area,
1074 is_migrate_cma(migratetype)
1075 ? migratetype : start_migratetype);
1077 trace_mm_page_alloc_extfrag(page, order, current_order,
1078 start_migratetype, migratetype);
1088 * Do the hard work of removing an element from the buddy allocator.
1089 * Call me with the zone->lock already held.
1091 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1097 page = __rmqueue_smallest(zone, order, migratetype);
1099 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1100 page = __rmqueue_fallback(zone, order, migratetype);
1103 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1104 * is used because __rmqueue_smallest is an inline function
1105 * and we want just one call site
1108 migratetype = MIGRATE_RESERVE;
1113 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1118 * Obtain a specified number of elements from the buddy allocator, all under
1119 * a single hold of the lock, for efficiency. Add them to the supplied list.
1120 * Returns the number of new pages which were placed at *list.
1122 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1123 unsigned long count, struct list_head *list,
1124 int migratetype, int cold)
1126 int mt = migratetype, i;
1128 spin_lock(&zone->lock);
1129 for (i = 0; i < count; ++i) {
1130 struct page *page = __rmqueue(zone, order, migratetype);
1131 if (unlikely(page == NULL))
1135 * Split buddy pages returned by expand() are received here
1136 * in physical page order. The page is added to the callers and
1137 * list and the list head then moves forward. From the callers
1138 * perspective, the linked list is ordered by page number in
1139 * some conditions. This is useful for IO devices that can
1140 * merge IO requests if the physical pages are ordered
1143 if (likely(cold == 0))
1144 list_add(&page->lru, list);
1146 list_add_tail(&page->lru, list);
1147 if (IS_ENABLED(CONFIG_CMA)) {
1148 mt = get_pageblock_migratetype(page);
1149 if (!is_migrate_cma(mt) && mt != MIGRATE_ISOLATE)
1152 set_freepage_migratetype(page, mt);
1154 if (is_migrate_cma(mt))
1155 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1158 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1159 spin_unlock(&zone->lock);
1165 * Called from the vmstat counter updater to drain pagesets of this
1166 * currently executing processor on remote nodes after they have
1169 * Note that this function must be called with the thread pinned to
1170 * a single processor.
1172 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1174 unsigned long flags;
1177 local_irq_save(flags);
1178 if (pcp->count >= pcp->batch)
1179 to_drain = pcp->batch;
1181 to_drain = pcp->count;
1183 free_pcppages_bulk(zone, to_drain, pcp);
1184 pcp->count -= to_drain;
1186 local_irq_restore(flags);
1191 * Drain pages of the indicated processor.
1193 * The processor must either be the current processor and the
1194 * thread pinned to the current processor or a processor that
1197 static void drain_pages(unsigned int cpu)
1199 unsigned long flags;
1202 for_each_populated_zone(zone) {
1203 struct per_cpu_pageset *pset;
1204 struct per_cpu_pages *pcp;
1206 local_irq_save(flags);
1207 pset = per_cpu_ptr(zone->pageset, cpu);
1211 free_pcppages_bulk(zone, pcp->count, pcp);
1214 local_irq_restore(flags);
1219 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1221 void drain_local_pages(void *arg)
1223 drain_pages(smp_processor_id());
1227 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1229 * Note that this code is protected against sending an IPI to an offline
1230 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1231 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1232 * nothing keeps CPUs from showing up after we populated the cpumask and
1233 * before the call to on_each_cpu_mask().
1235 void drain_all_pages(void)
1238 struct per_cpu_pageset *pcp;
1242 * Allocate in the BSS so we wont require allocation in
1243 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1245 static cpumask_t cpus_with_pcps;
1248 * We don't care about racing with CPU hotplug event
1249 * as offline notification will cause the notified
1250 * cpu to drain that CPU pcps and on_each_cpu_mask
1251 * disables preemption as part of its processing
1253 for_each_online_cpu(cpu) {
1254 bool has_pcps = false;
1255 for_each_populated_zone(zone) {
1256 pcp = per_cpu_ptr(zone->pageset, cpu);
1257 if (pcp->pcp.count) {
1263 cpumask_set_cpu(cpu, &cpus_with_pcps);
1265 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1267 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1270 #ifdef CONFIG_HIBERNATION
1272 void mark_free_pages(struct zone *zone)
1274 unsigned long pfn, max_zone_pfn;
1275 unsigned long flags;
1277 struct list_head *curr;
1279 if (!zone->spanned_pages)
1282 spin_lock_irqsave(&zone->lock, flags);
1284 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1285 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1286 if (pfn_valid(pfn)) {
1287 struct page *page = pfn_to_page(pfn);
1289 if (!swsusp_page_is_forbidden(page))
1290 swsusp_unset_page_free(page);
1293 for_each_migratetype_order(order, t) {
1294 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1297 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1298 for (i = 0; i < (1UL << order); i++)
1299 swsusp_set_page_free(pfn_to_page(pfn + i));
1302 spin_unlock_irqrestore(&zone->lock, flags);
1304 #endif /* CONFIG_PM */
1307 * Free a 0-order page
1308 * cold == 1 ? free a cold page : free a hot page
1310 void free_hot_cold_page(struct page *page, int cold)
1312 struct zone *zone = page_zone(page);
1313 struct per_cpu_pages *pcp;
1314 unsigned long flags;
1317 if (!free_pages_prepare(page, 0))
1320 migratetype = get_pageblock_migratetype(page);
1321 set_freepage_migratetype(page, migratetype);
1322 local_irq_save(flags);
1323 __count_vm_event(PGFREE);
1326 * We only track unmovable, reclaimable and movable on pcp lists.
1327 * Free ISOLATE pages back to the allocator because they are being
1328 * offlined but treat RESERVE as movable pages so we can get those
1329 * areas back if necessary. Otherwise, we may have to free
1330 * excessively into the page allocator
1332 if (migratetype >= MIGRATE_PCPTYPES) {
1333 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1334 free_one_page(zone, page, 0, migratetype);
1337 migratetype = MIGRATE_MOVABLE;
1340 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1342 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1344 list_add(&page->lru, &pcp->lists[migratetype]);
1346 if (pcp->count >= pcp->high) {
1347 free_pcppages_bulk(zone, pcp->batch, pcp);
1348 pcp->count -= pcp->batch;
1352 local_irq_restore(flags);
1356 * Free a list of 0-order pages
1358 void free_hot_cold_page_list(struct list_head *list, int cold)
1360 struct page *page, *next;
1362 list_for_each_entry_safe(page, next, list, lru) {
1363 trace_mm_page_free_batched(page, cold);
1364 free_hot_cold_page(page, cold);
1369 * split_page takes a non-compound higher-order page, and splits it into
1370 * n (1<<order) sub-pages: page[0..n]
1371 * Each sub-page must be freed individually.
1373 * Note: this is probably too low level an operation for use in drivers.
1374 * Please consult with lkml before using this in your driver.
1376 void split_page(struct page *page, unsigned int order)
1380 VM_BUG_ON(PageCompound(page));
1381 VM_BUG_ON(!page_count(page));
1383 #ifdef CONFIG_KMEMCHECK
1385 * Split shadow pages too, because free(page[0]) would
1386 * otherwise free the whole shadow.
1388 if (kmemcheck_page_is_tracked(page))
1389 split_page(virt_to_page(page[0].shadow), order);
1392 for (i = 1; i < (1 << order); i++)
1393 set_page_refcounted(page + i);
1396 static int __isolate_free_page(struct page *page, unsigned int order)
1398 unsigned long watermark;
1402 BUG_ON(!PageBuddy(page));
1404 zone = page_zone(page);
1405 mt = get_pageblock_migratetype(page);
1407 if (mt != MIGRATE_ISOLATE) {
1408 /* Obey watermarks as if the page was being allocated */
1409 watermark = low_wmark_pages(zone) + (1 << order);
1410 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1413 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1416 /* Remove page from free list */
1417 list_del(&page->lru);
1418 zone->free_area[order].nr_free--;
1419 rmv_page_order(page);
1421 /* Set the pageblock if the isolated page is at least a pageblock */
1422 if (order >= pageblock_order - 1) {
1423 struct page *endpage = page + (1 << order) - 1;
1424 for (; page < endpage; page += pageblock_nr_pages) {
1425 int mt = get_pageblock_migratetype(page);
1426 if (mt != MIGRATE_ISOLATE && !is_migrate_cma(mt))
1427 set_pageblock_migratetype(page,
1432 return 1UL << order;
1436 * Similar to split_page except the page is already free. As this is only
1437 * being used for migration, the migratetype of the block also changes.
1438 * As this is called with interrupts disabled, the caller is responsible
1439 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1442 * Note: this is probably too low level an operation for use in drivers.
1443 * Please consult with lkml before using this in your driver.
1445 int split_free_page(struct page *page)
1450 order = page_order(page);
1452 nr_pages = __isolate_free_page(page, order);
1456 /* Split into individual pages */
1457 set_page_refcounted(page);
1458 split_page(page, order);
1463 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1464 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1468 struct page *buffered_rmqueue(struct zone *preferred_zone,
1469 struct zone *zone, int order, gfp_t gfp_flags,
1472 unsigned long flags;
1474 int cold = !!(gfp_flags & __GFP_COLD);
1477 if (likely(order == 0)) {
1478 struct per_cpu_pages *pcp;
1479 struct list_head *list;
1481 local_irq_save(flags);
1482 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1483 list = &pcp->lists[migratetype];
1484 if (list_empty(list)) {
1485 pcp->count += rmqueue_bulk(zone, 0,
1488 if (unlikely(list_empty(list)))
1493 page = list_entry(list->prev, struct page, lru);
1495 page = list_entry(list->next, struct page, lru);
1497 list_del(&page->lru);
1500 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1502 * __GFP_NOFAIL is not to be used in new code.
1504 * All __GFP_NOFAIL callers should be fixed so that they
1505 * properly detect and handle allocation failures.
1507 * We most definitely don't want callers attempting to
1508 * allocate greater than order-1 page units with
1511 WARN_ON_ONCE(order > 1);
1513 spin_lock_irqsave(&zone->lock, flags);
1514 page = __rmqueue(zone, order, migratetype);
1515 spin_unlock(&zone->lock);
1518 __mod_zone_freepage_state(zone, -(1 << order),
1519 get_pageblock_migratetype(page));
1522 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1523 zone_statistics(preferred_zone, zone, gfp_flags);
1524 local_irq_restore(flags);
1526 VM_BUG_ON(bad_range(zone, page));
1527 if (prep_new_page(page, order, gfp_flags))
1532 local_irq_restore(flags);
1536 #ifdef CONFIG_FAIL_PAGE_ALLOC
1539 struct fault_attr attr;
1541 u32 ignore_gfp_highmem;
1542 u32 ignore_gfp_wait;
1544 } fail_page_alloc = {
1545 .attr = FAULT_ATTR_INITIALIZER,
1546 .ignore_gfp_wait = 1,
1547 .ignore_gfp_highmem = 1,
1551 static int __init setup_fail_page_alloc(char *str)
1553 return setup_fault_attr(&fail_page_alloc.attr, str);
1555 __setup("fail_page_alloc=", setup_fail_page_alloc);
1557 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1559 if (order < fail_page_alloc.min_order)
1561 if (gfp_mask & __GFP_NOFAIL)
1563 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1565 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1568 return should_fail(&fail_page_alloc.attr, 1 << order);
1571 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1573 static int __init fail_page_alloc_debugfs(void)
1575 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1578 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1579 &fail_page_alloc.attr);
1581 return PTR_ERR(dir);
1583 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1584 &fail_page_alloc.ignore_gfp_wait))
1586 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1587 &fail_page_alloc.ignore_gfp_highmem))
1589 if (!debugfs_create_u32("min-order", mode, dir,
1590 &fail_page_alloc.min_order))
1595 debugfs_remove_recursive(dir);
1600 late_initcall(fail_page_alloc_debugfs);
1602 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1604 #else /* CONFIG_FAIL_PAGE_ALLOC */
1606 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1611 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1614 * Return true if free pages are above 'mark'. This takes into account the order
1615 * of the allocation.
1617 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1618 int classzone_idx, int alloc_flags, long free_pages)
1620 /* free_pages my go negative - that's OK */
1622 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1625 free_pages -= (1 << order) - 1;
1626 if (alloc_flags & ALLOC_HIGH)
1628 if (alloc_flags & ALLOC_HARDER)
1631 /* If allocation can't use CMA areas don't use free CMA pages */
1632 if (!(alloc_flags & ALLOC_CMA))
1633 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
1635 if (free_pages <= min + lowmem_reserve)
1637 for (o = 0; o < order; o++) {
1638 /* At the next order, this order's pages become unavailable */
1639 free_pages -= z->free_area[o].nr_free << o;
1641 /* Require fewer higher order pages to be free */
1644 if (free_pages <= min)
1650 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1651 int classzone_idx, int alloc_flags)
1653 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1654 zone_page_state(z, NR_FREE_PAGES));
1657 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1658 int classzone_idx, int alloc_flags)
1660 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1662 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1663 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1665 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1671 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1672 * skip over zones that are not allowed by the cpuset, or that have
1673 * been recently (in last second) found to be nearly full. See further
1674 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1675 * that have to skip over a lot of full or unallowed zones.
1677 * If the zonelist cache is present in the passed in zonelist, then
1678 * returns a pointer to the allowed node mask (either the current
1679 * tasks mems_allowed, or node_states[N_MEMORY].)
1681 * If the zonelist cache is not available for this zonelist, does
1682 * nothing and returns NULL.
1684 * If the fullzones BITMAP in the zonelist cache is stale (more than
1685 * a second since last zap'd) then we zap it out (clear its bits.)
1687 * We hold off even calling zlc_setup, until after we've checked the
1688 * first zone in the zonelist, on the theory that most allocations will
1689 * be satisfied from that first zone, so best to examine that zone as
1690 * quickly as we can.
1692 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1694 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1695 nodemask_t *allowednodes; /* zonelist_cache approximation */
1697 zlc = zonelist->zlcache_ptr;
1701 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1702 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1703 zlc->last_full_zap = jiffies;
1706 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1707 &cpuset_current_mems_allowed :
1708 &node_states[N_MEMORY];
1709 return allowednodes;
1713 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1714 * if it is worth looking at further for free memory:
1715 * 1) Check that the zone isn't thought to be full (doesn't have its
1716 * bit set in the zonelist_cache fullzones BITMAP).
1717 * 2) Check that the zones node (obtained from the zonelist_cache
1718 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1719 * Return true (non-zero) if zone is worth looking at further, or
1720 * else return false (zero) if it is not.
1722 * This check -ignores- the distinction between various watermarks,
1723 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1724 * found to be full for any variation of these watermarks, it will
1725 * be considered full for up to one second by all requests, unless
1726 * we are so low on memory on all allowed nodes that we are forced
1727 * into the second scan of the zonelist.
1729 * In the second scan we ignore this zonelist cache and exactly
1730 * apply the watermarks to all zones, even it is slower to do so.
1731 * We are low on memory in the second scan, and should leave no stone
1732 * unturned looking for a free page.
1734 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1735 nodemask_t *allowednodes)
1737 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1738 int i; /* index of *z in zonelist zones */
1739 int n; /* node that zone *z is on */
1741 zlc = zonelist->zlcache_ptr;
1745 i = z - zonelist->_zonerefs;
1748 /* This zone is worth trying if it is allowed but not full */
1749 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1753 * Given 'z' scanning a zonelist, set the corresponding bit in
1754 * zlc->fullzones, so that subsequent attempts to allocate a page
1755 * from that zone don't waste time re-examining it.
1757 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1759 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1760 int i; /* index of *z in zonelist zones */
1762 zlc = zonelist->zlcache_ptr;
1766 i = z - zonelist->_zonerefs;
1768 set_bit(i, zlc->fullzones);
1772 * clear all zones full, called after direct reclaim makes progress so that
1773 * a zone that was recently full is not skipped over for up to a second
1775 static void zlc_clear_zones_full(struct zonelist *zonelist)
1777 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1779 zlc = zonelist->zlcache_ptr;
1783 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1786 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1788 return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes);
1791 static void __paginginit init_zone_allows_reclaim(int nid)
1795 for_each_online_node(i)
1796 if (node_distance(nid, i) <= RECLAIM_DISTANCE)
1797 node_set(i, NODE_DATA(nid)->reclaim_nodes);
1799 zone_reclaim_mode = 1;
1802 #else /* CONFIG_NUMA */
1804 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1809 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1810 nodemask_t *allowednodes)
1815 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1819 static void zlc_clear_zones_full(struct zonelist *zonelist)
1823 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1828 static inline void init_zone_allows_reclaim(int nid)
1831 #endif /* CONFIG_NUMA */
1834 * get_page_from_freelist goes through the zonelist trying to allocate
1837 static struct page *
1838 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1839 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1840 struct zone *preferred_zone, int migratetype)
1843 struct page *page = NULL;
1846 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1847 int zlc_active = 0; /* set if using zonelist_cache */
1848 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1850 classzone_idx = zone_idx(preferred_zone);
1853 * Scan zonelist, looking for a zone with enough free.
1854 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1856 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1857 high_zoneidx, nodemask) {
1858 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1859 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1861 if ((alloc_flags & ALLOC_CPUSET) &&
1862 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1865 * When allocating a page cache page for writing, we
1866 * want to get it from a zone that is within its dirty
1867 * limit, such that no single zone holds more than its
1868 * proportional share of globally allowed dirty pages.
1869 * The dirty limits take into account the zone's
1870 * lowmem reserves and high watermark so that kswapd
1871 * should be able to balance it without having to
1872 * write pages from its LRU list.
1874 * This may look like it could increase pressure on
1875 * lower zones by failing allocations in higher zones
1876 * before they are full. But the pages that do spill
1877 * over are limited as the lower zones are protected
1878 * by this very same mechanism. It should not become
1879 * a practical burden to them.
1881 * XXX: For now, allow allocations to potentially
1882 * exceed the per-zone dirty limit in the slowpath
1883 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1884 * which is important when on a NUMA setup the allowed
1885 * zones are together not big enough to reach the
1886 * global limit. The proper fix for these situations
1887 * will require awareness of zones in the
1888 * dirty-throttling and the flusher threads.
1890 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1891 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1892 goto this_zone_full;
1894 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1895 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1899 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1900 if (zone_watermark_ok(zone, order, mark,
1901 classzone_idx, alloc_flags))
1904 if (IS_ENABLED(CONFIG_NUMA) &&
1905 !did_zlc_setup && nr_online_nodes > 1) {
1907 * we do zlc_setup if there are multiple nodes
1908 * and before considering the first zone allowed
1911 allowednodes = zlc_setup(zonelist, alloc_flags);
1916 if (zone_reclaim_mode == 0 ||
1917 !zone_allows_reclaim(preferred_zone, zone))
1918 goto this_zone_full;
1921 * As we may have just activated ZLC, check if the first
1922 * eligible zone has failed zone_reclaim recently.
1924 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1925 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1928 ret = zone_reclaim(zone, gfp_mask, order);
1930 case ZONE_RECLAIM_NOSCAN:
1933 case ZONE_RECLAIM_FULL:
1934 /* scanned but unreclaimable */
1937 /* did we reclaim enough */
1938 if (!zone_watermark_ok(zone, order, mark,
1939 classzone_idx, alloc_flags))
1940 goto this_zone_full;
1945 page = buffered_rmqueue(preferred_zone, zone, order,
1946 gfp_mask, migratetype);
1950 if (IS_ENABLED(CONFIG_NUMA))
1951 zlc_mark_zone_full(zonelist, z);
1954 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
1955 /* Disable zlc cache for second zonelist scan */
1962 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1963 * necessary to allocate the page. The expectation is
1964 * that the caller is taking steps that will free more
1965 * memory. The caller should avoid the page being used
1966 * for !PFMEMALLOC purposes.
1968 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
1974 * Large machines with many possible nodes should not always dump per-node
1975 * meminfo in irq context.
1977 static inline bool should_suppress_show_mem(void)
1982 ret = in_interrupt();
1987 static DEFINE_RATELIMIT_STATE(nopage_rs,
1988 DEFAULT_RATELIMIT_INTERVAL,
1989 DEFAULT_RATELIMIT_BURST);
1991 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1993 unsigned int filter = SHOW_MEM_FILTER_NODES;
1995 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
1996 debug_guardpage_minorder() > 0)
2000 * This documents exceptions given to allocations in certain
2001 * contexts that are allowed to allocate outside current's set
2004 if (!(gfp_mask & __GFP_NOMEMALLOC))
2005 if (test_thread_flag(TIF_MEMDIE) ||
2006 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2007 filter &= ~SHOW_MEM_FILTER_NODES;
2008 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2009 filter &= ~SHOW_MEM_FILTER_NODES;
2012 struct va_format vaf;
2015 va_start(args, fmt);
2020 pr_warn("%pV", &vaf);
2025 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2026 current->comm, order, gfp_mask);
2029 if (!should_suppress_show_mem())
2034 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2035 unsigned long did_some_progress,
2036 unsigned long pages_reclaimed)
2038 /* Do not loop if specifically requested */
2039 if (gfp_mask & __GFP_NORETRY)
2042 /* Always retry if specifically requested */
2043 if (gfp_mask & __GFP_NOFAIL)
2047 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2048 * making forward progress without invoking OOM. Suspend also disables
2049 * storage devices so kswapd will not help. Bail if we are suspending.
2051 if (!did_some_progress && pm_suspended_storage())
2055 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2056 * means __GFP_NOFAIL, but that may not be true in other
2059 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2063 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2064 * specified, then we retry until we no longer reclaim any pages
2065 * (above), or we've reclaimed an order of pages at least as
2066 * large as the allocation's order. In both cases, if the
2067 * allocation still fails, we stop retrying.
2069 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2075 static inline struct page *
2076 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2077 struct zonelist *zonelist, enum zone_type high_zoneidx,
2078 nodemask_t *nodemask, struct zone *preferred_zone,
2083 /* Acquire the OOM killer lock for the zones in zonelist */
2084 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2085 schedule_timeout_uninterruptible(1);
2090 * Go through the zonelist yet one more time, keep very high watermark
2091 * here, this is only to catch a parallel oom killing, we must fail if
2092 * we're still under heavy pressure.
2094 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2095 order, zonelist, high_zoneidx,
2096 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2097 preferred_zone, migratetype);
2101 if (!(gfp_mask & __GFP_NOFAIL)) {
2102 /* The OOM killer will not help higher order allocs */
2103 if (order > PAGE_ALLOC_COSTLY_ORDER)
2105 /* The OOM killer does not needlessly kill tasks for lowmem */
2106 if (high_zoneidx < ZONE_NORMAL)
2109 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2110 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2111 * The caller should handle page allocation failure by itself if
2112 * it specifies __GFP_THISNODE.
2113 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2115 if (gfp_mask & __GFP_THISNODE)
2118 /* Exhausted what can be done so it's blamo time */
2119 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2122 clear_zonelist_oom(zonelist, gfp_mask);
2126 #ifdef CONFIG_COMPACTION
2127 /* Try memory compaction for high-order allocations before reclaim */
2128 static struct page *
2129 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2130 struct zonelist *zonelist, enum zone_type high_zoneidx,
2131 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2132 int migratetype, bool sync_migration,
2133 bool *contended_compaction, bool *deferred_compaction,
2134 unsigned long *did_some_progress)
2139 if (compaction_deferred(preferred_zone, order)) {
2140 *deferred_compaction = true;
2144 current->flags |= PF_MEMALLOC;
2145 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2146 nodemask, sync_migration,
2147 contended_compaction);
2148 current->flags &= ~PF_MEMALLOC;
2150 if (*did_some_progress != COMPACT_SKIPPED) {
2153 /* Page migration frees to the PCP lists but we want merging */
2154 drain_pages(get_cpu());
2157 page = get_page_from_freelist(gfp_mask, nodemask,
2158 order, zonelist, high_zoneidx,
2159 alloc_flags & ~ALLOC_NO_WATERMARKS,
2160 preferred_zone, migratetype);
2162 preferred_zone->compact_blockskip_flush = false;
2163 preferred_zone->compact_considered = 0;
2164 preferred_zone->compact_defer_shift = 0;
2165 if (order >= preferred_zone->compact_order_failed)
2166 preferred_zone->compact_order_failed = order + 1;
2167 count_vm_event(COMPACTSUCCESS);
2172 * It's bad if compaction run occurs and fails.
2173 * The most likely reason is that pages exist,
2174 * but not enough to satisfy watermarks.
2176 count_vm_event(COMPACTFAIL);
2179 * As async compaction considers a subset of pageblocks, only
2180 * defer if the failure was a sync compaction failure.
2183 defer_compaction(preferred_zone, order);
2191 static inline struct page *
2192 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2193 struct zonelist *zonelist, enum zone_type high_zoneidx,
2194 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2195 int migratetype, bool sync_migration,
2196 bool *contended_compaction, bool *deferred_compaction,
2197 unsigned long *did_some_progress)
2201 #endif /* CONFIG_COMPACTION */
2203 /* Perform direct synchronous page reclaim */
2205 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2206 nodemask_t *nodemask)
2208 struct reclaim_state reclaim_state;
2213 /* We now go into synchronous reclaim */
2214 cpuset_memory_pressure_bump();
2215 current->flags |= PF_MEMALLOC;
2216 lockdep_set_current_reclaim_state(gfp_mask);
2217 reclaim_state.reclaimed_slab = 0;
2218 current->reclaim_state = &reclaim_state;
2220 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2222 current->reclaim_state = NULL;
2223 lockdep_clear_current_reclaim_state();
2224 current->flags &= ~PF_MEMALLOC;
2231 /* The really slow allocator path where we enter direct reclaim */
2232 static inline struct page *
2233 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2234 struct zonelist *zonelist, enum zone_type high_zoneidx,
2235 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2236 int migratetype, unsigned long *did_some_progress)
2238 struct page *page = NULL;
2239 bool drained = false;
2241 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2243 if (unlikely(!(*did_some_progress)))
2246 /* After successful reclaim, reconsider all zones for allocation */
2247 if (IS_ENABLED(CONFIG_NUMA))
2248 zlc_clear_zones_full(zonelist);
2251 page = get_page_from_freelist(gfp_mask, nodemask, order,
2252 zonelist, high_zoneidx,
2253 alloc_flags & ~ALLOC_NO_WATERMARKS,
2254 preferred_zone, migratetype);
2257 * If an allocation failed after direct reclaim, it could be because
2258 * pages are pinned on the per-cpu lists. Drain them and try again
2260 if (!page && !drained) {
2270 * This is called in the allocator slow-path if the allocation request is of
2271 * sufficient urgency to ignore watermarks and take other desperate measures
2273 static inline struct page *
2274 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2275 struct zonelist *zonelist, enum zone_type high_zoneidx,
2276 nodemask_t *nodemask, struct zone *preferred_zone,
2282 page = get_page_from_freelist(gfp_mask, nodemask, order,
2283 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2284 preferred_zone, migratetype);
2286 if (!page && gfp_mask & __GFP_NOFAIL)
2287 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2288 } while (!page && (gfp_mask & __GFP_NOFAIL));
2294 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2295 enum zone_type high_zoneidx,
2296 enum zone_type classzone_idx)
2301 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2302 wakeup_kswapd(zone, order, classzone_idx);
2306 gfp_to_alloc_flags(gfp_t gfp_mask)
2308 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2309 const gfp_t wait = gfp_mask & __GFP_WAIT;
2311 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2312 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2315 * The caller may dip into page reserves a bit more if the caller
2316 * cannot run direct reclaim, or if the caller has realtime scheduling
2317 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2318 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2320 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2324 * Not worth trying to allocate harder for
2325 * __GFP_NOMEMALLOC even if it can't schedule.
2327 if (!(gfp_mask & __GFP_NOMEMALLOC))
2328 alloc_flags |= ALLOC_HARDER;
2330 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2331 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2333 alloc_flags &= ~ALLOC_CPUSET;
2334 } else if (unlikely(rt_task(current)) && !in_interrupt())
2335 alloc_flags |= ALLOC_HARDER;
2337 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2338 if (gfp_mask & __GFP_MEMALLOC)
2339 alloc_flags |= ALLOC_NO_WATERMARKS;
2340 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2341 alloc_flags |= ALLOC_NO_WATERMARKS;
2342 else if (!in_interrupt() &&
2343 ((current->flags & PF_MEMALLOC) ||
2344 unlikely(test_thread_flag(TIF_MEMDIE))))
2345 alloc_flags |= ALLOC_NO_WATERMARKS;
2348 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2349 alloc_flags |= ALLOC_CMA;
2354 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2356 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2359 static inline struct page *
2360 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2361 struct zonelist *zonelist, enum zone_type high_zoneidx,
2362 nodemask_t *nodemask, struct zone *preferred_zone,
2365 const gfp_t wait = gfp_mask & __GFP_WAIT;
2366 struct page *page = NULL;
2368 unsigned long pages_reclaimed = 0;
2369 unsigned long did_some_progress;
2370 bool sync_migration = false;
2371 bool deferred_compaction = false;
2372 bool contended_compaction = false;
2375 * In the slowpath, we sanity check order to avoid ever trying to
2376 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2377 * be using allocators in order of preference for an area that is
2380 if (order >= MAX_ORDER) {
2381 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2386 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2387 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2388 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2389 * using a larger set of nodes after it has established that the
2390 * allowed per node queues are empty and that nodes are
2393 if (IS_ENABLED(CONFIG_NUMA) &&
2394 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2398 if (!(gfp_mask & __GFP_NO_KSWAPD))
2399 wake_all_kswapd(order, zonelist, high_zoneidx,
2400 zone_idx(preferred_zone));
2403 * OK, we're below the kswapd watermark and have kicked background
2404 * reclaim. Now things get more complex, so set up alloc_flags according
2405 * to how we want to proceed.
2407 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2410 * Find the true preferred zone if the allocation is unconstrained by
2413 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2414 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2418 /* This is the last chance, in general, before the goto nopage. */
2419 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2420 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2421 preferred_zone, migratetype);
2425 /* Allocate without watermarks if the context allows */
2426 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2428 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2429 * the allocation is high priority and these type of
2430 * allocations are system rather than user orientated
2432 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2434 page = __alloc_pages_high_priority(gfp_mask, order,
2435 zonelist, high_zoneidx, nodemask,
2436 preferred_zone, migratetype);
2442 /* Atomic allocations - we can't balance anything */
2446 /* Avoid recursion of direct reclaim */
2447 if (current->flags & PF_MEMALLOC)
2450 /* Avoid allocations with no watermarks from looping endlessly */
2451 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2455 * Try direct compaction. The first pass is asynchronous. Subsequent
2456 * attempts after direct reclaim are synchronous
2458 page = __alloc_pages_direct_compact(gfp_mask, order,
2459 zonelist, high_zoneidx,
2461 alloc_flags, preferred_zone,
2462 migratetype, sync_migration,
2463 &contended_compaction,
2464 &deferred_compaction,
2465 &did_some_progress);
2468 sync_migration = true;
2471 * If compaction is deferred for high-order allocations, it is because
2472 * sync compaction recently failed. In this is the case and the caller
2473 * requested a movable allocation that does not heavily disrupt the
2474 * system then fail the allocation instead of entering direct reclaim.
2476 if ((deferred_compaction || contended_compaction) &&
2477 (gfp_mask & __GFP_NO_KSWAPD))
2480 /* Try direct reclaim and then allocating */
2481 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2482 zonelist, high_zoneidx,
2484 alloc_flags, preferred_zone,
2485 migratetype, &did_some_progress);
2490 * If we failed to make any progress reclaiming, then we are
2491 * running out of options and have to consider going OOM
2493 if (!did_some_progress) {
2494 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2495 if (oom_killer_disabled)
2497 /* Coredumps can quickly deplete all memory reserves */
2498 if ((current->flags & PF_DUMPCORE) &&
2499 !(gfp_mask & __GFP_NOFAIL))
2501 page = __alloc_pages_may_oom(gfp_mask, order,
2502 zonelist, high_zoneidx,
2503 nodemask, preferred_zone,
2508 if (!(gfp_mask & __GFP_NOFAIL)) {
2510 * The oom killer is not called for high-order
2511 * allocations that may fail, so if no progress
2512 * is being made, there are no other options and
2513 * retrying is unlikely to help.
2515 if (order > PAGE_ALLOC_COSTLY_ORDER)
2518 * The oom killer is not called for lowmem
2519 * allocations to prevent needlessly killing
2522 if (high_zoneidx < ZONE_NORMAL)
2530 /* Check if we should retry the allocation */
2531 pages_reclaimed += did_some_progress;
2532 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2534 /* Wait for some write requests to complete then retry */
2535 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2539 * High-order allocations do not necessarily loop after
2540 * direct reclaim and reclaim/compaction depends on compaction
2541 * being called after reclaim so call directly if necessary
2543 page = __alloc_pages_direct_compact(gfp_mask, order,
2544 zonelist, high_zoneidx,
2546 alloc_flags, preferred_zone,
2547 migratetype, sync_migration,
2548 &contended_compaction,
2549 &deferred_compaction,
2550 &did_some_progress);
2556 warn_alloc_failed(gfp_mask, order, NULL);
2559 if (kmemcheck_enabled)
2560 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2566 * This is the 'heart' of the zoned buddy allocator.
2569 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2570 struct zonelist *zonelist, nodemask_t *nodemask)
2572 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2573 struct zone *preferred_zone;
2574 struct page *page = NULL;
2575 int migratetype = allocflags_to_migratetype(gfp_mask);
2576 unsigned int cpuset_mems_cookie;
2577 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
2578 struct mem_cgroup *memcg = NULL;
2580 gfp_mask &= gfp_allowed_mask;
2582 lockdep_trace_alloc(gfp_mask);
2584 might_sleep_if(gfp_mask & __GFP_WAIT);
2586 if (should_fail_alloc_page(gfp_mask, order))
2590 * Check the zones suitable for the gfp_mask contain at least one
2591 * valid zone. It's possible to have an empty zonelist as a result
2592 * of GFP_THISNODE and a memoryless node
2594 if (unlikely(!zonelist->_zonerefs->zone))
2598 * Will only have any effect when __GFP_KMEMCG is set. This is
2599 * verified in the (always inline) callee
2601 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2605 cpuset_mems_cookie = get_mems_allowed();
2607 /* The preferred zone is used for statistics later */
2608 first_zones_zonelist(zonelist, high_zoneidx,
2609 nodemask ? : &cpuset_current_mems_allowed,
2611 if (!preferred_zone)
2615 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2616 alloc_flags |= ALLOC_CMA;
2618 /* First allocation attempt */
2619 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2620 zonelist, high_zoneidx, alloc_flags,
2621 preferred_zone, migratetype);
2622 if (unlikely(!page))
2623 page = __alloc_pages_slowpath(gfp_mask, order,
2624 zonelist, high_zoneidx, nodemask,
2625 preferred_zone, migratetype);
2627 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2631 * When updating a task's mems_allowed, it is possible to race with
2632 * parallel threads in such a way that an allocation can fail while
2633 * the mask is being updated. If a page allocation is about to fail,
2634 * check if the cpuset changed during allocation and if so, retry.
2636 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2639 memcg_kmem_commit_charge(page, memcg, order);
2643 EXPORT_SYMBOL(__alloc_pages_nodemask);
2646 * Common helper functions.
2648 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2653 * __get_free_pages() returns a 32-bit address, which cannot represent
2656 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2658 page = alloc_pages(gfp_mask, order);
2661 return (unsigned long) page_address(page);
2663 EXPORT_SYMBOL(__get_free_pages);
2665 unsigned long get_zeroed_page(gfp_t gfp_mask)
2667 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2669 EXPORT_SYMBOL(get_zeroed_page);
2671 void __free_pages(struct page *page, unsigned int order)
2673 if (put_page_testzero(page)) {
2675 free_hot_cold_page(page, 0);
2677 __free_pages_ok(page, order);
2681 EXPORT_SYMBOL(__free_pages);
2683 void free_pages(unsigned long addr, unsigned int order)
2686 VM_BUG_ON(!virt_addr_valid((void *)addr));
2687 __free_pages(virt_to_page((void *)addr), order);
2691 EXPORT_SYMBOL(free_pages);
2694 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2695 * pages allocated with __GFP_KMEMCG.
2697 * Those pages are accounted to a particular memcg, embedded in the
2698 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2699 * for that information only to find out that it is NULL for users who have no
2700 * interest in that whatsoever, we provide these functions.
2702 * The caller knows better which flags it relies on.
2704 void __free_memcg_kmem_pages(struct page *page, unsigned int order)
2706 memcg_kmem_uncharge_pages(page, order);
2707 __free_pages(page, order);
2710 void free_memcg_kmem_pages(unsigned long addr, unsigned int order)
2713 VM_BUG_ON(!virt_addr_valid((void *)addr));
2714 __free_memcg_kmem_pages(virt_to_page((void *)addr), order);
2718 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2721 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2722 unsigned long used = addr + PAGE_ALIGN(size);
2724 split_page(virt_to_page((void *)addr), order);
2725 while (used < alloc_end) {
2730 return (void *)addr;
2734 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2735 * @size: the number of bytes to allocate
2736 * @gfp_mask: GFP flags for the allocation
2738 * This function is similar to alloc_pages(), except that it allocates the
2739 * minimum number of pages to satisfy the request. alloc_pages() can only
2740 * allocate memory in power-of-two pages.
2742 * This function is also limited by MAX_ORDER.
2744 * Memory allocated by this function must be released by free_pages_exact().
2746 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2748 unsigned int order = get_order(size);
2751 addr = __get_free_pages(gfp_mask, order);
2752 return make_alloc_exact(addr, order, size);
2754 EXPORT_SYMBOL(alloc_pages_exact);
2757 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2759 * @nid: the preferred node ID where memory should be allocated
2760 * @size: the number of bytes to allocate
2761 * @gfp_mask: GFP flags for the allocation
2763 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2765 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2768 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2770 unsigned order = get_order(size);
2771 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2774 return make_alloc_exact((unsigned long)page_address(p), order, size);
2776 EXPORT_SYMBOL(alloc_pages_exact_nid);
2779 * free_pages_exact - release memory allocated via alloc_pages_exact()
2780 * @virt: the value returned by alloc_pages_exact.
2781 * @size: size of allocation, same value as passed to alloc_pages_exact().
2783 * Release the memory allocated by a previous call to alloc_pages_exact.
2785 void free_pages_exact(void *virt, size_t size)
2787 unsigned long addr = (unsigned long)virt;
2788 unsigned long end = addr + PAGE_ALIGN(size);
2790 while (addr < end) {
2795 EXPORT_SYMBOL(free_pages_exact);
2797 static unsigned int nr_free_zone_pages(int offset)
2802 /* Just pick one node, since fallback list is circular */
2803 unsigned int sum = 0;
2805 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2807 for_each_zone_zonelist(zone, z, zonelist, offset) {
2808 unsigned long size = zone->present_pages;
2809 unsigned long high = high_wmark_pages(zone);
2818 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2820 unsigned int nr_free_buffer_pages(void)
2822 return nr_free_zone_pages(gfp_zone(GFP_USER));
2824 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2827 * Amount of free RAM allocatable within all zones
2829 unsigned int nr_free_pagecache_pages(void)
2831 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2834 static inline void show_node(struct zone *zone)
2836 if (IS_ENABLED(CONFIG_NUMA))
2837 printk("Node %d ", zone_to_nid(zone));
2840 void si_meminfo(struct sysinfo *val)
2842 val->totalram = totalram_pages;
2844 val->freeram = global_page_state(NR_FREE_PAGES);
2845 val->bufferram = nr_blockdev_pages();
2846 val->totalhigh = totalhigh_pages;
2847 val->freehigh = nr_free_highpages();
2848 val->mem_unit = PAGE_SIZE;
2851 EXPORT_SYMBOL(si_meminfo);
2854 void si_meminfo_node(struct sysinfo *val, int nid)
2856 pg_data_t *pgdat = NODE_DATA(nid);
2858 val->totalram = pgdat->node_present_pages;
2859 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2860 #ifdef CONFIG_HIGHMEM
2861 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2862 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2868 val->mem_unit = PAGE_SIZE;
2873 * Determine whether the node should be displayed or not, depending on whether
2874 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2876 bool skip_free_areas_node(unsigned int flags, int nid)
2879 unsigned int cpuset_mems_cookie;
2881 if (!(flags & SHOW_MEM_FILTER_NODES))
2885 cpuset_mems_cookie = get_mems_allowed();
2886 ret = !node_isset(nid, cpuset_current_mems_allowed);
2887 } while (!put_mems_allowed(cpuset_mems_cookie));
2892 #define K(x) ((x) << (PAGE_SHIFT-10))
2894 static void show_migration_types(unsigned char type)
2896 static const char types[MIGRATE_TYPES] = {
2897 [MIGRATE_UNMOVABLE] = 'U',
2898 [MIGRATE_RECLAIMABLE] = 'E',
2899 [MIGRATE_MOVABLE] = 'M',
2900 [MIGRATE_RESERVE] = 'R',
2902 [MIGRATE_CMA] = 'C',
2904 [MIGRATE_ISOLATE] = 'I',
2906 char tmp[MIGRATE_TYPES + 1];
2910 for (i = 0; i < MIGRATE_TYPES; i++) {
2911 if (type & (1 << i))
2916 printk("(%s) ", tmp);
2920 * Show free area list (used inside shift_scroll-lock stuff)
2921 * We also calculate the percentage fragmentation. We do this by counting the
2922 * memory on each free list with the exception of the first item on the list.
2923 * Suppresses nodes that are not allowed by current's cpuset if
2924 * SHOW_MEM_FILTER_NODES is passed.
2926 void show_free_areas(unsigned int filter)
2931 for_each_populated_zone(zone) {
2932 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2935 printk("%s per-cpu:\n", zone->name);
2937 for_each_online_cpu(cpu) {
2938 struct per_cpu_pageset *pageset;
2940 pageset = per_cpu_ptr(zone->pageset, cpu);
2942 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2943 cpu, pageset->pcp.high,
2944 pageset->pcp.batch, pageset->pcp.count);
2948 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2949 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2951 " dirty:%lu writeback:%lu unstable:%lu\n"
2952 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2953 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
2955 global_page_state(NR_ACTIVE_ANON),
2956 global_page_state(NR_INACTIVE_ANON),
2957 global_page_state(NR_ISOLATED_ANON),
2958 global_page_state(NR_ACTIVE_FILE),
2959 global_page_state(NR_INACTIVE_FILE),
2960 global_page_state(NR_ISOLATED_FILE),
2961 global_page_state(NR_UNEVICTABLE),
2962 global_page_state(NR_FILE_DIRTY),
2963 global_page_state(NR_WRITEBACK),
2964 global_page_state(NR_UNSTABLE_NFS),
2965 global_page_state(NR_FREE_PAGES),
2966 global_page_state(NR_SLAB_RECLAIMABLE),
2967 global_page_state(NR_SLAB_UNRECLAIMABLE),
2968 global_page_state(NR_FILE_MAPPED),
2969 global_page_state(NR_SHMEM),
2970 global_page_state(NR_PAGETABLE),
2971 global_page_state(NR_BOUNCE),
2972 global_page_state(NR_FREE_CMA_PAGES));
2974 for_each_populated_zone(zone) {
2977 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2985 " active_anon:%lukB"
2986 " inactive_anon:%lukB"
2987 " active_file:%lukB"
2988 " inactive_file:%lukB"
2989 " unevictable:%lukB"
2990 " isolated(anon):%lukB"
2991 " isolated(file):%lukB"
2999 " slab_reclaimable:%lukB"
3000 " slab_unreclaimable:%lukB"
3001 " kernel_stack:%lukB"
3006 " writeback_tmp:%lukB"
3007 " pages_scanned:%lu"
3008 " all_unreclaimable? %s"
3011 K(zone_page_state(zone, NR_FREE_PAGES)),
3012 K(min_wmark_pages(zone)),
3013 K(low_wmark_pages(zone)),
3014 K(high_wmark_pages(zone)),
3015 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3016 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3017 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3018 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3019 K(zone_page_state(zone, NR_UNEVICTABLE)),
3020 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3021 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3022 K(zone->present_pages),
3023 K(zone->managed_pages),
3024 K(zone_page_state(zone, NR_MLOCK)),
3025 K(zone_page_state(zone, NR_FILE_DIRTY)),
3026 K(zone_page_state(zone, NR_WRITEBACK)),
3027 K(zone_page_state(zone, NR_FILE_MAPPED)),
3028 K(zone_page_state(zone, NR_SHMEM)),
3029 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3030 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3031 zone_page_state(zone, NR_KERNEL_STACK) *
3033 K(zone_page_state(zone, NR_PAGETABLE)),
3034 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3035 K(zone_page_state(zone, NR_BOUNCE)),
3036 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3037 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3038 zone->pages_scanned,
3039 (zone->all_unreclaimable ? "yes" : "no")
3041 printk("lowmem_reserve[]:");
3042 for (i = 0; i < MAX_NR_ZONES; i++)
3043 printk(" %lu", zone->lowmem_reserve[i]);
3047 for_each_populated_zone(zone) {
3048 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3049 unsigned char types[MAX_ORDER];
3051 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3054 printk("%s: ", zone->name);
3056 spin_lock_irqsave(&zone->lock, flags);
3057 for (order = 0; order < MAX_ORDER; order++) {
3058 struct free_area *area = &zone->free_area[order];
3061 nr[order] = area->nr_free;
3062 total += nr[order] << order;
3065 for (type = 0; type < MIGRATE_TYPES; type++) {
3066 if (!list_empty(&area->free_list[type]))
3067 types[order] |= 1 << type;
3070 spin_unlock_irqrestore(&zone->lock, flags);
3071 for (order = 0; order < MAX_ORDER; order++) {
3072 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3074 show_migration_types(types[order]);
3076 printk("= %lukB\n", K(total));
3079 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3081 show_swap_cache_info();
3084 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3086 zoneref->zone = zone;
3087 zoneref->zone_idx = zone_idx(zone);
3091 * Builds allocation fallback zone lists.
3093 * Add all populated zones of a node to the zonelist.
3095 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3096 int nr_zones, enum zone_type zone_type)
3100 BUG_ON(zone_type >= MAX_NR_ZONES);
3105 zone = pgdat->node_zones + zone_type;
3106 if (populated_zone(zone)) {
3107 zoneref_set_zone(zone,
3108 &zonelist->_zonerefs[nr_zones++]);
3109 check_highest_zone(zone_type);
3112 } while (zone_type);
3119 * 0 = automatic detection of better ordering.
3120 * 1 = order by ([node] distance, -zonetype)
3121 * 2 = order by (-zonetype, [node] distance)
3123 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3124 * the same zonelist. So only NUMA can configure this param.
3126 #define ZONELIST_ORDER_DEFAULT 0
3127 #define ZONELIST_ORDER_NODE 1
3128 #define ZONELIST_ORDER_ZONE 2
3130 /* zonelist order in the kernel.
3131 * set_zonelist_order() will set this to NODE or ZONE.
3133 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3134 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3138 /* The value user specified ....changed by config */
3139 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3140 /* string for sysctl */
3141 #define NUMA_ZONELIST_ORDER_LEN 16
3142 char numa_zonelist_order[16] = "default";
3145 * interface for configure zonelist ordering.
3146 * command line option "numa_zonelist_order"
3147 * = "[dD]efault - default, automatic configuration.
3148 * = "[nN]ode - order by node locality, then by zone within node
3149 * = "[zZ]one - order by zone, then by locality within zone
3152 static int __parse_numa_zonelist_order(char *s)
3154 if (*s == 'd' || *s == 'D') {
3155 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3156 } else if (*s == 'n' || *s == 'N') {
3157 user_zonelist_order = ZONELIST_ORDER_NODE;
3158 } else if (*s == 'z' || *s == 'Z') {
3159 user_zonelist_order = ZONELIST_ORDER_ZONE;
3162 "Ignoring invalid numa_zonelist_order value: "
3169 static __init int setup_numa_zonelist_order(char *s)
3176 ret = __parse_numa_zonelist_order(s);
3178 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3182 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3185 * sysctl handler for numa_zonelist_order
3187 int numa_zonelist_order_handler(ctl_table *table, int write,
3188 void __user *buffer, size_t *length,
3191 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3193 static DEFINE_MUTEX(zl_order_mutex);
3195 mutex_lock(&zl_order_mutex);
3197 strcpy(saved_string, (char*)table->data);
3198 ret = proc_dostring(table, write, buffer, length, ppos);
3202 int oldval = user_zonelist_order;
3203 if (__parse_numa_zonelist_order((char*)table->data)) {
3205 * bogus value. restore saved string
3207 strncpy((char*)table->data, saved_string,
3208 NUMA_ZONELIST_ORDER_LEN);
3209 user_zonelist_order = oldval;
3210 } else if (oldval != user_zonelist_order) {
3211 mutex_lock(&zonelists_mutex);
3212 build_all_zonelists(NULL, NULL);
3213 mutex_unlock(&zonelists_mutex);
3217 mutex_unlock(&zl_order_mutex);
3222 #define MAX_NODE_LOAD (nr_online_nodes)
3223 static int node_load[MAX_NUMNODES];
3226 * find_next_best_node - find the next node that should appear in a given node's fallback list
3227 * @node: node whose fallback list we're appending
3228 * @used_node_mask: nodemask_t of already used nodes
3230 * We use a number of factors to determine which is the next node that should
3231 * appear on a given node's fallback list. The node should not have appeared
3232 * already in @node's fallback list, and it should be the next closest node
3233 * according to the distance array (which contains arbitrary distance values
3234 * from each node to each node in the system), and should also prefer nodes
3235 * with no CPUs, since presumably they'll have very little allocation pressure
3236 * on them otherwise.
3237 * It returns -1 if no node is found.
3239 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3242 int min_val = INT_MAX;
3244 const struct cpumask *tmp = cpumask_of_node(0);
3246 /* Use the local node if we haven't already */
3247 if (!node_isset(node, *used_node_mask)) {
3248 node_set(node, *used_node_mask);
3252 for_each_node_state(n, N_MEMORY) {
3254 /* Don't want a node to appear more than once */
3255 if (node_isset(n, *used_node_mask))
3258 /* Use the distance array to find the distance */
3259 val = node_distance(node, n);
3261 /* Penalize nodes under us ("prefer the next node") */
3264 /* Give preference to headless and unused nodes */
3265 tmp = cpumask_of_node(n);
3266 if (!cpumask_empty(tmp))
3267 val += PENALTY_FOR_NODE_WITH_CPUS;
3269 /* Slight preference for less loaded node */
3270 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3271 val += node_load[n];
3273 if (val < min_val) {
3280 node_set(best_node, *used_node_mask);
3287 * Build zonelists ordered by node and zones within node.
3288 * This results in maximum locality--normal zone overflows into local
3289 * DMA zone, if any--but risks exhausting DMA zone.
3291 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3294 struct zonelist *zonelist;
3296 zonelist = &pgdat->node_zonelists[0];
3297 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3299 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3301 zonelist->_zonerefs[j].zone = NULL;
3302 zonelist->_zonerefs[j].zone_idx = 0;
3306 * Build gfp_thisnode zonelists
3308 static void build_thisnode_zonelists(pg_data_t *pgdat)
3311 struct zonelist *zonelist;
3313 zonelist = &pgdat->node_zonelists[1];
3314 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3315 zonelist->_zonerefs[j].zone = NULL;
3316 zonelist->_zonerefs[j].zone_idx = 0;
3320 * Build zonelists ordered by zone and nodes within zones.
3321 * This results in conserving DMA zone[s] until all Normal memory is
3322 * exhausted, but results in overflowing to remote node while memory
3323 * may still exist in local DMA zone.
3325 static int node_order[MAX_NUMNODES];
3327 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3330 int zone_type; /* needs to be signed */
3332 struct zonelist *zonelist;
3334 zonelist = &pgdat->node_zonelists[0];
3336 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3337 for (j = 0; j < nr_nodes; j++) {
3338 node = node_order[j];
3339 z = &NODE_DATA(node)->node_zones[zone_type];
3340 if (populated_zone(z)) {
3342 &zonelist->_zonerefs[pos++]);
3343 check_highest_zone(zone_type);
3347 zonelist->_zonerefs[pos].zone = NULL;
3348 zonelist->_zonerefs[pos].zone_idx = 0;
3351 static int default_zonelist_order(void)
3354 unsigned long low_kmem_size,total_size;
3358 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3359 * If they are really small and used heavily, the system can fall
3360 * into OOM very easily.
3361 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3363 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3366 for_each_online_node(nid) {
3367 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3368 z = &NODE_DATA(nid)->node_zones[zone_type];
3369 if (populated_zone(z)) {
3370 if (zone_type < ZONE_NORMAL)
3371 low_kmem_size += z->present_pages;
3372 total_size += z->present_pages;
3373 } else if (zone_type == ZONE_NORMAL) {
3375 * If any node has only lowmem, then node order
3376 * is preferred to allow kernel allocations
3377 * locally; otherwise, they can easily infringe
3378 * on other nodes when there is an abundance of
3379 * lowmem available to allocate from.
3381 return ZONELIST_ORDER_NODE;
3385 if (!low_kmem_size || /* there are no DMA area. */
3386 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3387 return ZONELIST_ORDER_NODE;
3389 * look into each node's config.
3390 * If there is a node whose DMA/DMA32 memory is very big area on
3391 * local memory, NODE_ORDER may be suitable.
3393 average_size = total_size /
3394 (nodes_weight(node_states[N_MEMORY]) + 1);
3395 for_each_online_node(nid) {
3398 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3399 z = &NODE_DATA(nid)->node_zones[zone_type];
3400 if (populated_zone(z)) {
3401 if (zone_type < ZONE_NORMAL)
3402 low_kmem_size += z->present_pages;
3403 total_size += z->present_pages;
3406 if (low_kmem_size &&
3407 total_size > average_size && /* ignore small node */
3408 low_kmem_size > total_size * 70/100)
3409 return ZONELIST_ORDER_NODE;
3411 return ZONELIST_ORDER_ZONE;
3414 static void set_zonelist_order(void)
3416 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3417 current_zonelist_order = default_zonelist_order();
3419 current_zonelist_order = user_zonelist_order;
3422 static void build_zonelists(pg_data_t *pgdat)
3426 nodemask_t used_mask;
3427 int local_node, prev_node;
3428 struct zonelist *zonelist;
3429 int order = current_zonelist_order;
3431 /* initialize zonelists */
3432 for (i = 0; i < MAX_ZONELISTS; i++) {
3433 zonelist = pgdat->node_zonelists + i;
3434 zonelist->_zonerefs[0].zone = NULL;
3435 zonelist->_zonerefs[0].zone_idx = 0;
3438 /* NUMA-aware ordering of nodes */
3439 local_node = pgdat->node_id;
3440 load = nr_online_nodes;
3441 prev_node = local_node;
3442 nodes_clear(used_mask);
3444 memset(node_order, 0, sizeof(node_order));
3447 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3449 * We don't want to pressure a particular node.
3450 * So adding penalty to the first node in same
3451 * distance group to make it round-robin.
3453 if (node_distance(local_node, node) !=
3454 node_distance(local_node, prev_node))
3455 node_load[node] = load;
3459 if (order == ZONELIST_ORDER_NODE)
3460 build_zonelists_in_node_order(pgdat, node);
3462 node_order[j++] = node; /* remember order */
3465 if (order == ZONELIST_ORDER_ZONE) {
3466 /* calculate node order -- i.e., DMA last! */
3467 build_zonelists_in_zone_order(pgdat, j);
3470 build_thisnode_zonelists(pgdat);
3473 /* Construct the zonelist performance cache - see further mmzone.h */
3474 static void build_zonelist_cache(pg_data_t *pgdat)
3476 struct zonelist *zonelist;
3477 struct zonelist_cache *zlc;
3480 zonelist = &pgdat->node_zonelists[0];
3481 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3482 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3483 for (z = zonelist->_zonerefs; z->zone; z++)
3484 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3487 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3489 * Return node id of node used for "local" allocations.
3490 * I.e., first node id of first zone in arg node's generic zonelist.
3491 * Used for initializing percpu 'numa_mem', which is used primarily
3492 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3494 int local_memory_node(int node)
3498 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3499 gfp_zone(GFP_KERNEL),
3506 #else /* CONFIG_NUMA */
3508 static void set_zonelist_order(void)
3510 current_zonelist_order = ZONELIST_ORDER_ZONE;
3513 static void build_zonelists(pg_data_t *pgdat)
3515 int node, local_node;
3517 struct zonelist *zonelist;
3519 local_node = pgdat->node_id;
3521 zonelist = &pgdat->node_zonelists[0];
3522 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3525 * Now we build the zonelist so that it contains the zones
3526 * of all the other nodes.
3527 * We don't want to pressure a particular node, so when
3528 * building the zones for node N, we make sure that the
3529 * zones coming right after the local ones are those from
3530 * node N+1 (modulo N)
3532 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3533 if (!node_online(node))
3535 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3538 for (node = 0; node < local_node; node++) {
3539 if (!node_online(node))
3541 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3545 zonelist->_zonerefs[j].zone = NULL;
3546 zonelist->_zonerefs[j].zone_idx = 0;
3549 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3550 static void build_zonelist_cache(pg_data_t *pgdat)
3552 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3555 #endif /* CONFIG_NUMA */
3558 * Boot pageset table. One per cpu which is going to be used for all
3559 * zones and all nodes. The parameters will be set in such a way
3560 * that an item put on a list will immediately be handed over to
3561 * the buddy list. This is safe since pageset manipulation is done
3562 * with interrupts disabled.
3564 * The boot_pagesets must be kept even after bootup is complete for
3565 * unused processors and/or zones. They do play a role for bootstrapping
3566 * hotplugged processors.
3568 * zoneinfo_show() and maybe other functions do
3569 * not check if the processor is online before following the pageset pointer.
3570 * Other parts of the kernel may not check if the zone is available.
3572 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3573 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3574 static void setup_zone_pageset(struct zone *zone);
3577 * Global mutex to protect against size modification of zonelists
3578 * as well as to serialize pageset setup for the new populated zone.
3580 DEFINE_MUTEX(zonelists_mutex);
3582 /* return values int ....just for stop_machine() */
3583 static int __build_all_zonelists(void *data)
3587 pg_data_t *self = data;
3590 memset(node_load, 0, sizeof(node_load));
3593 if (self && !node_online(self->node_id)) {
3594 build_zonelists(self);
3595 build_zonelist_cache(self);
3598 for_each_online_node(nid) {
3599 pg_data_t *pgdat = NODE_DATA(nid);
3601 build_zonelists(pgdat);
3602 build_zonelist_cache(pgdat);
3606 * Initialize the boot_pagesets that are going to be used
3607 * for bootstrapping processors. The real pagesets for
3608 * each zone will be allocated later when the per cpu
3609 * allocator is available.
3611 * boot_pagesets are used also for bootstrapping offline
3612 * cpus if the system is already booted because the pagesets
3613 * are needed to initialize allocators on a specific cpu too.
3614 * F.e. the percpu allocator needs the page allocator which
3615 * needs the percpu allocator in order to allocate its pagesets
3616 * (a chicken-egg dilemma).
3618 for_each_possible_cpu(cpu) {
3619 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3621 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3623 * We now know the "local memory node" for each node--
3624 * i.e., the node of the first zone in the generic zonelist.
3625 * Set up numa_mem percpu variable for on-line cpus. During
3626 * boot, only the boot cpu should be on-line; we'll init the
3627 * secondary cpus' numa_mem as they come on-line. During
3628 * node/memory hotplug, we'll fixup all on-line cpus.
3630 if (cpu_online(cpu))
3631 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3639 * Called with zonelists_mutex held always
3640 * unless system_state == SYSTEM_BOOTING.
3642 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3644 set_zonelist_order();
3646 if (system_state == SYSTEM_BOOTING) {
3647 __build_all_zonelists(NULL);
3648 mminit_verify_zonelist();
3649 cpuset_init_current_mems_allowed();
3651 /* we have to stop all cpus to guarantee there is no user
3653 #ifdef CONFIG_MEMORY_HOTPLUG
3655 setup_zone_pageset(zone);
3657 stop_machine(__build_all_zonelists, pgdat, NULL);
3658 /* cpuset refresh routine should be here */
3660 vm_total_pages = nr_free_pagecache_pages();
3662 * Disable grouping by mobility if the number of pages in the
3663 * system is too low to allow the mechanism to work. It would be
3664 * more accurate, but expensive to check per-zone. This check is
3665 * made on memory-hotadd so a system can start with mobility
3666 * disabled and enable it later
3668 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3669 page_group_by_mobility_disabled = 1;
3671 page_group_by_mobility_disabled = 0;
3673 printk("Built %i zonelists in %s order, mobility grouping %s. "
3674 "Total pages: %ld\n",
3676 zonelist_order_name[current_zonelist_order],
3677 page_group_by_mobility_disabled ? "off" : "on",
3680 printk("Policy zone: %s\n", zone_names[policy_zone]);
3685 * Helper functions to size the waitqueue hash table.
3686 * Essentially these want to choose hash table sizes sufficiently
3687 * large so that collisions trying to wait on pages are rare.
3688 * But in fact, the number of active page waitqueues on typical
3689 * systems is ridiculously low, less than 200. So this is even
3690 * conservative, even though it seems large.
3692 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3693 * waitqueues, i.e. the size of the waitq table given the number of pages.
3695 #define PAGES_PER_WAITQUEUE 256
3697 #ifndef CONFIG_MEMORY_HOTPLUG
3698 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3700 unsigned long size = 1;
3702 pages /= PAGES_PER_WAITQUEUE;
3704 while (size < pages)
3708 * Once we have dozens or even hundreds of threads sleeping
3709 * on IO we've got bigger problems than wait queue collision.
3710 * Limit the size of the wait table to a reasonable size.
3712 size = min(size, 4096UL);
3714 return max(size, 4UL);
3718 * A zone's size might be changed by hot-add, so it is not possible to determine
3719 * a suitable size for its wait_table. So we use the maximum size now.
3721 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3723 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3724 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3725 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3727 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3728 * or more by the traditional way. (See above). It equals:
3730 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3731 * ia64(16K page size) : = ( 8G + 4M)byte.
3732 * powerpc (64K page size) : = (32G +16M)byte.
3734 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3741 * This is an integer logarithm so that shifts can be used later
3742 * to extract the more random high bits from the multiplicative
3743 * hash function before the remainder is taken.
3745 static inline unsigned long wait_table_bits(unsigned long size)
3750 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3753 * Check if a pageblock contains reserved pages
3755 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3759 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3760 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3767 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3768 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3769 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3770 * higher will lead to a bigger reserve which will get freed as contiguous
3771 * blocks as reclaim kicks in
3773 static void setup_zone_migrate_reserve(struct zone *zone)
3775 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3777 unsigned long block_migratetype;
3781 * Get the start pfn, end pfn and the number of blocks to reserve
3782 * We have to be careful to be aligned to pageblock_nr_pages to
3783 * make sure that we always check pfn_valid for the first page in
3786 start_pfn = zone->zone_start_pfn;
3787 end_pfn = start_pfn + zone->spanned_pages;
3788 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3789 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3793 * Reserve blocks are generally in place to help high-order atomic
3794 * allocations that are short-lived. A min_free_kbytes value that
3795 * would result in more than 2 reserve blocks for atomic allocations
3796 * is assumed to be in place to help anti-fragmentation for the
3797 * future allocation of hugepages at runtime.
3799 reserve = min(2, reserve);
3801 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3802 if (!pfn_valid(pfn))
3804 page = pfn_to_page(pfn);
3806 /* Watch out for overlapping nodes */
3807 if (page_to_nid(page) != zone_to_nid(zone))
3810 block_migratetype = get_pageblock_migratetype(page);
3812 /* Only test what is necessary when the reserves are not met */
3815 * Blocks with reserved pages will never free, skip
3818 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3819 if (pageblock_is_reserved(pfn, block_end_pfn))
3822 /* If this block is reserved, account for it */
3823 if (block_migratetype == MIGRATE_RESERVE) {
3828 /* Suitable for reserving if this block is movable */
3829 if (block_migratetype == MIGRATE_MOVABLE) {
3830 set_pageblock_migratetype(page,
3832 move_freepages_block(zone, page,
3840 * If the reserve is met and this is a previous reserved block,
3843 if (block_migratetype == MIGRATE_RESERVE) {
3844 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3845 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3851 * Initially all pages are reserved - free ones are freed
3852 * up by free_all_bootmem() once the early boot process is
3853 * done. Non-atomic initialization, single-pass.
3855 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3856 unsigned long start_pfn, enum memmap_context context)
3859 unsigned long end_pfn = start_pfn + size;
3863 if (highest_memmap_pfn < end_pfn - 1)
3864 highest_memmap_pfn = end_pfn - 1;
3866 z = &NODE_DATA(nid)->node_zones[zone];
3867 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3869 * There can be holes in boot-time mem_map[]s
3870 * handed to this function. They do not
3871 * exist on hotplugged memory.
3873 if (context == MEMMAP_EARLY) {
3874 if (!early_pfn_valid(pfn))
3876 if (!early_pfn_in_nid(pfn, nid))
3879 page = pfn_to_page(pfn);
3880 set_page_links(page, zone, nid, pfn);
3881 mminit_verify_page_links(page, zone, nid, pfn);
3882 init_page_count(page);
3883 reset_page_mapcount(page);
3884 reset_page_last_nid(page);
3885 SetPageReserved(page);
3887 * Mark the block movable so that blocks are reserved for
3888 * movable at startup. This will force kernel allocations
3889 * to reserve their blocks rather than leaking throughout
3890 * the address space during boot when many long-lived
3891 * kernel allocations are made. Later some blocks near
3892 * the start are marked MIGRATE_RESERVE by
3893 * setup_zone_migrate_reserve()
3895 * bitmap is created for zone's valid pfn range. but memmap
3896 * can be created for invalid pages (for alignment)
3897 * check here not to call set_pageblock_migratetype() against
3900 if ((z->zone_start_pfn <= pfn)
3901 && (pfn < z->zone_start_pfn + z->spanned_pages)
3902 && !(pfn & (pageblock_nr_pages - 1)))
3903 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3905 INIT_LIST_HEAD(&page->lru);
3906 #ifdef WANT_PAGE_VIRTUAL
3907 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3908 if (!is_highmem_idx(zone))
3909 set_page_address(page, __va(pfn << PAGE_SHIFT));
3914 static void __meminit zone_init_free_lists(struct zone *zone)
3917 for_each_migratetype_order(order, t) {
3918 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3919 zone->free_area[order].nr_free = 0;
3923 #ifndef __HAVE_ARCH_MEMMAP_INIT
3924 #define memmap_init(size, nid, zone, start_pfn) \
3925 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3928 static int __meminit zone_batchsize(struct zone *zone)
3934 * The per-cpu-pages pools are set to around 1000th of the
3935 * size of the zone. But no more than 1/2 of a meg.
3937 * OK, so we don't know how big the cache is. So guess.
3939 batch = zone->present_pages / 1024;
3940 if (batch * PAGE_SIZE > 512 * 1024)
3941 batch = (512 * 1024) / PAGE_SIZE;
3942 batch /= 4; /* We effectively *= 4 below */
3947 * Clamp the batch to a 2^n - 1 value. Having a power
3948 * of 2 value was found to be more likely to have
3949 * suboptimal cache aliasing properties in some cases.
3951 * For example if 2 tasks are alternately allocating
3952 * batches of pages, one task can end up with a lot
3953 * of pages of one half of the possible page colors
3954 * and the other with pages of the other colors.
3956 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3961 /* The deferral and batching of frees should be suppressed under NOMMU
3964 * The problem is that NOMMU needs to be able to allocate large chunks
3965 * of contiguous memory as there's no hardware page translation to
3966 * assemble apparent contiguous memory from discontiguous pages.
3968 * Queueing large contiguous runs of pages for batching, however,
3969 * causes the pages to actually be freed in smaller chunks. As there
3970 * can be a significant delay between the individual batches being
3971 * recycled, this leads to the once large chunks of space being
3972 * fragmented and becoming unavailable for high-order allocations.
3978 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3980 struct per_cpu_pages *pcp;
3983 memset(p, 0, sizeof(*p));
3987 pcp->high = 6 * batch;
3988 pcp->batch = max(1UL, 1 * batch);
3989 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3990 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3994 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3995 * to the value high for the pageset p.
3998 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
4001 struct per_cpu_pages *pcp;
4005 pcp->batch = max(1UL, high/4);
4006 if ((high/4) > (PAGE_SHIFT * 8))
4007 pcp->batch = PAGE_SHIFT * 8;
4010 static void __meminit setup_zone_pageset(struct zone *zone)
4014 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4016 for_each_possible_cpu(cpu) {
4017 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4019 setup_pageset(pcp, zone_batchsize(zone));
4021 if (percpu_pagelist_fraction)
4022 setup_pagelist_highmark(pcp,
4023 (zone->present_pages /
4024 percpu_pagelist_fraction));
4029 * Allocate per cpu pagesets and initialize them.
4030 * Before this call only boot pagesets were available.
4032 void __init setup_per_cpu_pageset(void)
4036 for_each_populated_zone(zone)
4037 setup_zone_pageset(zone);
4040 static noinline __init_refok
4041 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4044 struct pglist_data *pgdat = zone->zone_pgdat;
4048 * The per-page waitqueue mechanism uses hashed waitqueues
4051 zone->wait_table_hash_nr_entries =
4052 wait_table_hash_nr_entries(zone_size_pages);
4053 zone->wait_table_bits =
4054 wait_table_bits(zone->wait_table_hash_nr_entries);
4055 alloc_size = zone->wait_table_hash_nr_entries
4056 * sizeof(wait_queue_head_t);
4058 if (!slab_is_available()) {
4059 zone->wait_table = (wait_queue_head_t *)
4060 alloc_bootmem_node_nopanic(pgdat, alloc_size);
4063 * This case means that a zone whose size was 0 gets new memory
4064 * via memory hot-add.
4065 * But it may be the case that a new node was hot-added. In
4066 * this case vmalloc() will not be able to use this new node's
4067 * memory - this wait_table must be initialized to use this new
4068 * node itself as well.
4069 * To use this new node's memory, further consideration will be
4072 zone->wait_table = vmalloc(alloc_size);
4074 if (!zone->wait_table)
4077 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4078 init_waitqueue_head(zone->wait_table + i);
4083 static __meminit void zone_pcp_init(struct zone *zone)
4086 * per cpu subsystem is not up at this point. The following code
4087 * relies on the ability of the linker to provide the
4088 * offset of a (static) per cpu variable into the per cpu area.
4090 zone->pageset = &boot_pageset;
4092 if (zone->present_pages)
4093 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4094 zone->name, zone->present_pages,
4095 zone_batchsize(zone));
4098 int __meminit init_currently_empty_zone(struct zone *zone,
4099 unsigned long zone_start_pfn,
4101 enum memmap_context context)
4103 struct pglist_data *pgdat = zone->zone_pgdat;
4105 ret = zone_wait_table_init(zone, size);
4108 pgdat->nr_zones = zone_idx(zone) + 1;
4110 zone->zone_start_pfn = zone_start_pfn;
4112 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4113 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4115 (unsigned long)zone_idx(zone),
4116 zone_start_pfn, (zone_start_pfn + size));
4118 zone_init_free_lists(zone);
4123 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4124 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4126 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4127 * Architectures may implement their own version but if add_active_range()
4128 * was used and there are no special requirements, this is a convenient
4131 int __meminit __early_pfn_to_nid(unsigned long pfn)
4133 unsigned long start_pfn, end_pfn;
4136 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
4137 if (start_pfn <= pfn && pfn < end_pfn)
4139 /* This is a memory hole */
4142 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4144 int __meminit early_pfn_to_nid(unsigned long pfn)
4148 nid = __early_pfn_to_nid(pfn);
4151 /* just returns 0 */
4155 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4156 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4160 nid = __early_pfn_to_nid(pfn);
4161 if (nid >= 0 && nid != node)
4168 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4169 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4170 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4172 * If an architecture guarantees that all ranges registered with
4173 * add_active_ranges() contain no holes and may be freed, this
4174 * this function may be used instead of calling free_bootmem() manually.
4176 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4178 unsigned long start_pfn, end_pfn;
4181 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4182 start_pfn = min(start_pfn, max_low_pfn);
4183 end_pfn = min(end_pfn, max_low_pfn);
4185 if (start_pfn < end_pfn)
4186 free_bootmem_node(NODE_DATA(this_nid),
4187 PFN_PHYS(start_pfn),
4188 (end_pfn - start_pfn) << PAGE_SHIFT);
4193 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4194 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4196 * If an architecture guarantees that all ranges registered with
4197 * add_active_ranges() contain no holes and may be freed, this
4198 * function may be used instead of calling memory_present() manually.
4200 void __init sparse_memory_present_with_active_regions(int nid)
4202 unsigned long start_pfn, end_pfn;
4205 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4206 memory_present(this_nid, start_pfn, end_pfn);
4210 * get_pfn_range_for_nid - Return the start and end page frames for a node
4211 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4212 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4213 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4215 * It returns the start and end page frame of a node based on information
4216 * provided by an arch calling add_active_range(). If called for a node
4217 * with no available memory, a warning is printed and the start and end
4220 void __meminit get_pfn_range_for_nid(unsigned int nid,
4221 unsigned long *start_pfn, unsigned long *end_pfn)
4223 unsigned long this_start_pfn, this_end_pfn;
4229 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4230 *start_pfn = min(*start_pfn, this_start_pfn);
4231 *end_pfn = max(*end_pfn, this_end_pfn);
4234 if (*start_pfn == -1UL)
4239 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4240 * assumption is made that zones within a node are ordered in monotonic
4241 * increasing memory addresses so that the "highest" populated zone is used
4243 static void __init find_usable_zone_for_movable(void)
4246 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4247 if (zone_index == ZONE_MOVABLE)
4250 if (arch_zone_highest_possible_pfn[zone_index] >
4251 arch_zone_lowest_possible_pfn[zone_index])
4255 VM_BUG_ON(zone_index == -1);
4256 movable_zone = zone_index;
4260 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4261 * because it is sized independent of architecture. Unlike the other zones,
4262 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4263 * in each node depending on the size of each node and how evenly kernelcore
4264 * is distributed. This helper function adjusts the zone ranges
4265 * provided by the architecture for a given node by using the end of the
4266 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4267 * zones within a node are in order of monotonic increases memory addresses
4269 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4270 unsigned long zone_type,
4271 unsigned long node_start_pfn,
4272 unsigned long node_end_pfn,
4273 unsigned long *zone_start_pfn,
4274 unsigned long *zone_end_pfn)
4276 /* Only adjust if ZONE_MOVABLE is on this node */
4277 if (zone_movable_pfn[nid]) {
4278 /* Size ZONE_MOVABLE */
4279 if (zone_type == ZONE_MOVABLE) {
4280 *zone_start_pfn = zone_movable_pfn[nid];
4281 *zone_end_pfn = min(node_end_pfn,
4282 arch_zone_highest_possible_pfn[movable_zone]);
4284 /* Adjust for ZONE_MOVABLE starting within this range */
4285 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4286 *zone_end_pfn > zone_movable_pfn[nid]) {
4287 *zone_end_pfn = zone_movable_pfn[nid];
4289 /* Check if this whole range is within ZONE_MOVABLE */
4290 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4291 *zone_start_pfn = *zone_end_pfn;
4296 * Return the number of pages a zone spans in a node, including holes
4297 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4299 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4300 unsigned long zone_type,
4301 unsigned long *ignored)
4303 unsigned long node_start_pfn, node_end_pfn;
4304 unsigned long zone_start_pfn, zone_end_pfn;
4306 /* Get the start and end of the node and zone */
4307 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4308 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4309 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4310 adjust_zone_range_for_zone_movable(nid, zone_type,
4311 node_start_pfn, node_end_pfn,
4312 &zone_start_pfn, &zone_end_pfn);
4314 /* Check that this node has pages within the zone's required range */
4315 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4318 /* Move the zone boundaries inside the node if necessary */
4319 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4320 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4322 /* Return the spanned pages */
4323 return zone_end_pfn - zone_start_pfn;
4327 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4328 * then all holes in the requested range will be accounted for.
4330 unsigned long __meminit __absent_pages_in_range(int nid,
4331 unsigned long range_start_pfn,
4332 unsigned long range_end_pfn)
4334 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4335 unsigned long start_pfn, end_pfn;
4338 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4339 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4340 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4341 nr_absent -= end_pfn - start_pfn;
4347 * absent_pages_in_range - Return number of page frames in holes within a range
4348 * @start_pfn: The start PFN to start searching for holes
4349 * @end_pfn: The end PFN to stop searching for holes
4351 * It returns the number of pages frames in memory holes within a range.
4353 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4354 unsigned long end_pfn)
4356 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4359 /* Return the number of page frames in holes in a zone on a node */
4360 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4361 unsigned long zone_type,
4362 unsigned long *ignored)
4364 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4365 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4366 unsigned long node_start_pfn, node_end_pfn;
4367 unsigned long zone_start_pfn, zone_end_pfn;
4369 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4370 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4371 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4373 adjust_zone_range_for_zone_movable(nid, zone_type,
4374 node_start_pfn, node_end_pfn,
4375 &zone_start_pfn, &zone_end_pfn);
4376 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4380 * sanitize_zone_movable_limit - Sanitize the zone_movable_limit array.
4382 * zone_movable_limit is initialized as 0. This function will try to get
4383 * the first ZONE_MOVABLE pfn of each node from movablemem_map, and
4384 * assigne them to zone_movable_limit.
4385 * zone_movable_limit[nid] == 0 means no limit for the node.
4387 * Note: Each range is represented as [start_pfn, end_pfn)
4389 static void __meminit sanitize_zone_movable_limit(void)
4391 int map_pos = 0, i, nid;
4392 unsigned long start_pfn, end_pfn;
4394 if (!movablemem_map.nr_map)
4397 /* Iterate all ranges from minimum to maximum */
4398 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4400 * If we have found lowest pfn of ZONE_MOVABLE of the node
4401 * specified by user, just go on to check next range.
4403 if (zone_movable_limit[nid])
4406 #ifdef CONFIG_ZONE_DMA
4407 /* Skip DMA memory. */
4408 if (start_pfn < arch_zone_highest_possible_pfn[ZONE_DMA])
4409 start_pfn = arch_zone_highest_possible_pfn[ZONE_DMA];
4412 #ifdef CONFIG_ZONE_DMA32
4413 /* Skip DMA32 memory. */
4414 if (start_pfn < arch_zone_highest_possible_pfn[ZONE_DMA32])
4415 start_pfn = arch_zone_highest_possible_pfn[ZONE_DMA32];
4418 #ifdef CONFIG_HIGHMEM
4419 /* Skip lowmem if ZONE_MOVABLE is highmem. */
4420 if (zone_movable_is_highmem() &&
4421 start_pfn < arch_zone_lowest_possible_pfn[ZONE_HIGHMEM])
4422 start_pfn = arch_zone_lowest_possible_pfn[ZONE_HIGHMEM];
4425 if (start_pfn >= end_pfn)
4428 while (map_pos < movablemem_map.nr_map) {
4429 if (end_pfn <= movablemem_map.map[map_pos].start_pfn)
4432 if (start_pfn >= movablemem_map.map[map_pos].end_pfn) {
4438 * The start_pfn of ZONE_MOVABLE is either the minimum
4439 * pfn specified by movablemem_map, or 0, which means
4440 * the node has no ZONE_MOVABLE.
4442 zone_movable_limit[nid] = max(start_pfn,
4443 movablemem_map.map[map_pos].start_pfn);
4450 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4451 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4452 unsigned long zone_type,
4453 unsigned long *zones_size)
4455 return zones_size[zone_type];
4458 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4459 unsigned long zone_type,
4460 unsigned long *zholes_size)
4465 return zholes_size[zone_type];
4467 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4469 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4470 unsigned long *zones_size, unsigned long *zholes_size)
4472 unsigned long realtotalpages, totalpages = 0;
4475 for (i = 0; i < MAX_NR_ZONES; i++)
4476 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4478 pgdat->node_spanned_pages = totalpages;
4480 realtotalpages = totalpages;
4481 for (i = 0; i < MAX_NR_ZONES; i++)
4483 zone_absent_pages_in_node(pgdat->node_id, i,
4485 pgdat->node_present_pages = realtotalpages;
4486 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4490 #ifndef CONFIG_SPARSEMEM
4492 * Calculate the size of the zone->blockflags rounded to an unsigned long
4493 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4494 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4495 * round what is now in bits to nearest long in bits, then return it in
4498 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4500 unsigned long usemapsize;
4502 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4503 usemapsize = roundup(zonesize, pageblock_nr_pages);
4504 usemapsize = usemapsize >> pageblock_order;
4505 usemapsize *= NR_PAGEBLOCK_BITS;
4506 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4508 return usemapsize / 8;
4511 static void __init setup_usemap(struct pglist_data *pgdat,
4513 unsigned long zone_start_pfn,
4514 unsigned long zonesize)
4516 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4517 zone->pageblock_flags = NULL;
4519 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4523 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4524 unsigned long zone_start_pfn, unsigned long zonesize) {}
4525 #endif /* CONFIG_SPARSEMEM */
4527 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4529 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4530 void __init set_pageblock_order(void)
4534 /* Check that pageblock_nr_pages has not already been setup */
4535 if (pageblock_order)
4538 if (HPAGE_SHIFT > PAGE_SHIFT)
4539 order = HUGETLB_PAGE_ORDER;
4541 order = MAX_ORDER - 1;
4544 * Assume the largest contiguous order of interest is a huge page.
4545 * This value may be variable depending on boot parameters on IA64 and
4548 pageblock_order = order;
4550 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4553 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4554 * is unused as pageblock_order is set at compile-time. See
4555 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4558 void __init set_pageblock_order(void)
4562 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4564 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4565 unsigned long present_pages)
4567 unsigned long pages = spanned_pages;
4570 * Provide a more accurate estimation if there are holes within
4571 * the zone and SPARSEMEM is in use. If there are holes within the
4572 * zone, each populated memory region may cost us one or two extra
4573 * memmap pages due to alignment because memmap pages for each
4574 * populated regions may not naturally algined on page boundary.
4575 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4577 if (spanned_pages > present_pages + (present_pages >> 4) &&
4578 IS_ENABLED(CONFIG_SPARSEMEM))
4579 pages = present_pages;
4581 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4585 * Set up the zone data structures:
4586 * - mark all pages reserved
4587 * - mark all memory queues empty
4588 * - clear the memory bitmaps
4590 * NOTE: pgdat should get zeroed by caller.
4592 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4593 unsigned long *zones_size, unsigned long *zholes_size)
4596 int nid = pgdat->node_id;
4597 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4600 pgdat_resize_init(pgdat);
4601 #ifdef CONFIG_NUMA_BALANCING
4602 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4603 pgdat->numabalancing_migrate_nr_pages = 0;
4604 pgdat->numabalancing_migrate_next_window = jiffies;
4606 init_waitqueue_head(&pgdat->kswapd_wait);
4607 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4608 pgdat_page_cgroup_init(pgdat);
4610 for (j = 0; j < MAX_NR_ZONES; j++) {
4611 struct zone *zone = pgdat->node_zones + j;
4612 unsigned long size, realsize, freesize, memmap_pages;
4614 size = zone_spanned_pages_in_node(nid, j, zones_size);
4615 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4619 * Adjust freesize so that it accounts for how much memory
4620 * is used by this zone for memmap. This affects the watermark
4621 * and per-cpu initialisations
4623 memmap_pages = calc_memmap_size(size, realsize);
4624 if (freesize >= memmap_pages) {
4625 freesize -= memmap_pages;
4628 " %s zone: %lu pages used for memmap\n",
4629 zone_names[j], memmap_pages);
4632 " %s zone: %lu pages exceeds freesize %lu\n",
4633 zone_names[j], memmap_pages, freesize);
4635 /* Account for reserved pages */
4636 if (j == 0 && freesize > dma_reserve) {
4637 freesize -= dma_reserve;
4638 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4639 zone_names[0], dma_reserve);
4642 if (!is_highmem_idx(j))
4643 nr_kernel_pages += freesize;
4644 /* Charge for highmem memmap if there are enough kernel pages */
4645 else if (nr_kernel_pages > memmap_pages * 2)
4646 nr_kernel_pages -= memmap_pages;
4647 nr_all_pages += freesize;
4649 zone->spanned_pages = size;
4650 zone->present_pages = freesize;
4652 * Set an approximate value for lowmem here, it will be adjusted
4653 * when the bootmem allocator frees pages into the buddy system.
4654 * And all highmem pages will be managed by the buddy system.
4656 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4659 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4661 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4663 zone->name = zone_names[j];
4664 spin_lock_init(&zone->lock);
4665 spin_lock_init(&zone->lru_lock);
4666 zone_seqlock_init(zone);
4667 zone->zone_pgdat = pgdat;
4669 zone_pcp_init(zone);
4670 lruvec_init(&zone->lruvec);
4674 set_pageblock_order();
4675 setup_usemap(pgdat, zone, zone_start_pfn, size);
4676 ret = init_currently_empty_zone(zone, zone_start_pfn,
4677 size, MEMMAP_EARLY);
4679 memmap_init(size, nid, j, zone_start_pfn);
4680 zone_start_pfn += size;
4684 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4686 /* Skip empty nodes */
4687 if (!pgdat->node_spanned_pages)
4690 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4691 /* ia64 gets its own node_mem_map, before this, without bootmem */
4692 if (!pgdat->node_mem_map) {
4693 unsigned long size, start, end;
4697 * The zone's endpoints aren't required to be MAX_ORDER
4698 * aligned but the node_mem_map endpoints must be in order
4699 * for the buddy allocator to function correctly.
4701 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4702 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4703 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4704 size = (end - start) * sizeof(struct page);
4705 map = alloc_remap(pgdat->node_id, size);
4707 map = alloc_bootmem_node_nopanic(pgdat, size);
4708 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4710 #ifndef CONFIG_NEED_MULTIPLE_NODES
4712 * With no DISCONTIG, the global mem_map is just set as node 0's
4714 if (pgdat == NODE_DATA(0)) {
4715 mem_map = NODE_DATA(0)->node_mem_map;
4716 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4717 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4718 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4719 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4722 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4725 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4726 unsigned long node_start_pfn, unsigned long *zholes_size)
4728 pg_data_t *pgdat = NODE_DATA(nid);
4730 /* pg_data_t should be reset to zero when it's allocated */
4731 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4733 pgdat->node_id = nid;
4734 pgdat->node_start_pfn = node_start_pfn;
4735 init_zone_allows_reclaim(nid);
4736 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4738 alloc_node_mem_map(pgdat);
4739 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4740 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4741 nid, (unsigned long)pgdat,
4742 (unsigned long)pgdat->node_mem_map);
4745 free_area_init_core(pgdat, zones_size, zholes_size);
4748 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4750 #if MAX_NUMNODES > 1
4752 * Figure out the number of possible node ids.
4754 static void __init setup_nr_node_ids(void)
4757 unsigned int highest = 0;
4759 for_each_node_mask(node, node_possible_map)
4761 nr_node_ids = highest + 1;
4764 static inline void setup_nr_node_ids(void)
4770 * node_map_pfn_alignment - determine the maximum internode alignment
4772 * This function should be called after node map is populated and sorted.
4773 * It calculates the maximum power of two alignment which can distinguish
4776 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4777 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4778 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4779 * shifted, 1GiB is enough and this function will indicate so.
4781 * This is used to test whether pfn -> nid mapping of the chosen memory
4782 * model has fine enough granularity to avoid incorrect mapping for the
4783 * populated node map.
4785 * Returns the determined alignment in pfn's. 0 if there is no alignment
4786 * requirement (single node).
4788 unsigned long __init node_map_pfn_alignment(void)
4790 unsigned long accl_mask = 0, last_end = 0;
4791 unsigned long start, end, mask;
4795 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4796 if (!start || last_nid < 0 || last_nid == nid) {
4803 * Start with a mask granular enough to pin-point to the
4804 * start pfn and tick off bits one-by-one until it becomes
4805 * too coarse to separate the current node from the last.
4807 mask = ~((1 << __ffs(start)) - 1);
4808 while (mask && last_end <= (start & (mask << 1)))
4811 /* accumulate all internode masks */
4815 /* convert mask to number of pages */
4816 return ~accl_mask + 1;
4819 /* Find the lowest pfn for a node */
4820 static unsigned long __init find_min_pfn_for_node(int nid)
4822 unsigned long min_pfn = ULONG_MAX;
4823 unsigned long start_pfn;
4826 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4827 min_pfn = min(min_pfn, start_pfn);
4829 if (min_pfn == ULONG_MAX) {
4831 "Could not find start_pfn for node %d\n", nid);
4839 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4841 * It returns the minimum PFN based on information provided via
4842 * add_active_range().
4844 unsigned long __init find_min_pfn_with_active_regions(void)
4846 return find_min_pfn_for_node(MAX_NUMNODES);
4850 * early_calculate_totalpages()
4851 * Sum pages in active regions for movable zone.
4852 * Populate N_MEMORY for calculating usable_nodes.
4854 static unsigned long __init early_calculate_totalpages(void)
4856 unsigned long totalpages = 0;
4857 unsigned long start_pfn, end_pfn;
4860 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
4861 unsigned long pages = end_pfn - start_pfn;
4863 totalpages += pages;
4865 node_set_state(nid, N_MEMORY);
4871 * Find the PFN the Movable zone begins in each node. Kernel memory
4872 * is spread evenly between nodes as long as the nodes have enough
4873 * memory. When they don't, some nodes will have more kernelcore than
4876 static void __init find_zone_movable_pfns_for_nodes(void)
4879 unsigned long usable_startpfn;
4880 unsigned long kernelcore_node, kernelcore_remaining;
4881 /* save the state before borrow the nodemask */
4882 nodemask_t saved_node_state = node_states[N_MEMORY];
4883 unsigned long totalpages = early_calculate_totalpages();
4884 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
4887 * If movablecore was specified, calculate what size of
4888 * kernelcore that corresponds so that memory usable for
4889 * any allocation type is evenly spread. If both kernelcore
4890 * and movablecore are specified, then the value of kernelcore
4891 * will be used for required_kernelcore if it's greater than
4892 * what movablecore would have allowed.
4894 if (required_movablecore) {
4895 unsigned long corepages;
4898 * Round-up so that ZONE_MOVABLE is at least as large as what
4899 * was requested by the user
4901 required_movablecore =
4902 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4903 corepages = totalpages - required_movablecore;
4905 required_kernelcore = max(required_kernelcore, corepages);
4908 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4909 if (!required_kernelcore)
4912 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4913 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4916 /* Spread kernelcore memory as evenly as possible throughout nodes */
4917 kernelcore_node = required_kernelcore / usable_nodes;
4918 for_each_node_state(nid, N_MEMORY) {
4919 unsigned long start_pfn, end_pfn;
4922 * Recalculate kernelcore_node if the division per node
4923 * now exceeds what is necessary to satisfy the requested
4924 * amount of memory for the kernel
4926 if (required_kernelcore < kernelcore_node)
4927 kernelcore_node = required_kernelcore / usable_nodes;
4930 * As the map is walked, we track how much memory is usable
4931 * by the kernel using kernelcore_remaining. When it is
4932 * 0, the rest of the node is usable by ZONE_MOVABLE
4934 kernelcore_remaining = kernelcore_node;
4936 /* Go through each range of PFNs within this node */
4937 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4938 unsigned long size_pages;
4940 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
4941 if (start_pfn >= end_pfn)
4944 /* Account for what is only usable for kernelcore */
4945 if (start_pfn < usable_startpfn) {
4946 unsigned long kernel_pages;
4947 kernel_pages = min(end_pfn, usable_startpfn)
4950 kernelcore_remaining -= min(kernel_pages,
4951 kernelcore_remaining);
4952 required_kernelcore -= min(kernel_pages,
4953 required_kernelcore);
4955 /* Continue if range is now fully accounted */
4956 if (end_pfn <= usable_startpfn) {
4959 * Push zone_movable_pfn to the end so
4960 * that if we have to rebalance
4961 * kernelcore across nodes, we will
4962 * not double account here
4964 zone_movable_pfn[nid] = end_pfn;
4967 start_pfn = usable_startpfn;
4971 * The usable PFN range for ZONE_MOVABLE is from
4972 * start_pfn->end_pfn. Calculate size_pages as the
4973 * number of pages used as kernelcore
4975 size_pages = end_pfn - start_pfn;
4976 if (size_pages > kernelcore_remaining)
4977 size_pages = kernelcore_remaining;
4978 zone_movable_pfn[nid] = start_pfn + size_pages;
4981 * Some kernelcore has been met, update counts and
4982 * break if the kernelcore for this node has been
4985 required_kernelcore -= min(required_kernelcore,
4987 kernelcore_remaining -= size_pages;
4988 if (!kernelcore_remaining)
4994 * If there is still required_kernelcore, we do another pass with one
4995 * less node in the count. This will push zone_movable_pfn[nid] further
4996 * along on the nodes that still have memory until kernelcore is
5000 if (usable_nodes && required_kernelcore > usable_nodes)
5003 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5004 for (nid = 0; nid < MAX_NUMNODES; nid++)
5005 zone_movable_pfn[nid] =
5006 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5009 /* restore the node_state */
5010 node_states[N_MEMORY] = saved_node_state;
5013 /* Any regular or high memory on that node ? */
5014 static void check_for_memory(pg_data_t *pgdat, int nid)
5016 enum zone_type zone_type;
5018 if (N_MEMORY == N_NORMAL_MEMORY)
5021 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5022 struct zone *zone = &pgdat->node_zones[zone_type];
5023 if (zone->present_pages) {
5024 node_set_state(nid, N_HIGH_MEMORY);
5025 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5026 zone_type <= ZONE_NORMAL)
5027 node_set_state(nid, N_NORMAL_MEMORY);
5034 * free_area_init_nodes - Initialise all pg_data_t and zone data
5035 * @max_zone_pfn: an array of max PFNs for each zone
5037 * This will call free_area_init_node() for each active node in the system.
5038 * Using the page ranges provided by add_active_range(), the size of each
5039 * zone in each node and their holes is calculated. If the maximum PFN
5040 * between two adjacent zones match, it is assumed that the zone is empty.
5041 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5042 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5043 * starts where the previous one ended. For example, ZONE_DMA32 starts
5044 * at arch_max_dma_pfn.
5046 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5048 unsigned long start_pfn, end_pfn;
5051 /* Record where the zone boundaries are */
5052 memset(arch_zone_lowest_possible_pfn, 0,
5053 sizeof(arch_zone_lowest_possible_pfn));
5054 memset(arch_zone_highest_possible_pfn, 0,
5055 sizeof(arch_zone_highest_possible_pfn));
5056 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5057 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5058 for (i = 1; i < MAX_NR_ZONES; i++) {
5059 if (i == ZONE_MOVABLE)
5061 arch_zone_lowest_possible_pfn[i] =
5062 arch_zone_highest_possible_pfn[i-1];
5063 arch_zone_highest_possible_pfn[i] =
5064 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5066 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5067 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5069 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5070 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5071 find_usable_zone_for_movable();
5072 sanitize_zone_movable_limit();
5073 find_zone_movable_pfns_for_nodes();
5075 /* Print out the zone ranges */
5076 printk("Zone ranges:\n");
5077 for (i = 0; i < MAX_NR_ZONES; i++) {
5078 if (i == ZONE_MOVABLE)
5080 printk(KERN_CONT " %-8s ", zone_names[i]);
5081 if (arch_zone_lowest_possible_pfn[i] ==
5082 arch_zone_highest_possible_pfn[i])
5083 printk(KERN_CONT "empty\n");
5085 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5086 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5087 (arch_zone_highest_possible_pfn[i]
5088 << PAGE_SHIFT) - 1);
5091 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5092 printk("Movable zone start for each node\n");
5093 for (i = 0; i < MAX_NUMNODES; i++) {
5094 if (zone_movable_pfn[i])
5095 printk(" Node %d: %#010lx\n", i,
5096 zone_movable_pfn[i] << PAGE_SHIFT);
5099 /* Print out the early node map */
5100 printk("Early memory node ranges\n");
5101 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5102 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5103 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5105 /* Initialise every node */
5106 mminit_verify_pageflags_layout();
5107 setup_nr_node_ids();
5108 for_each_online_node(nid) {
5109 pg_data_t *pgdat = NODE_DATA(nid);
5110 free_area_init_node(nid, NULL,
5111 find_min_pfn_for_node(nid), NULL);
5113 /* Any memory on that node */
5114 if (pgdat->node_present_pages)
5115 node_set_state(nid, N_MEMORY);
5116 check_for_memory(pgdat, nid);
5120 static int __init cmdline_parse_core(char *p, unsigned long *core)
5122 unsigned long long coremem;
5126 coremem = memparse(p, &p);
5127 *core = coremem >> PAGE_SHIFT;
5129 /* Paranoid check that UL is enough for the coremem value */
5130 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5136 * kernelcore=size sets the amount of memory for use for allocations that
5137 * cannot be reclaimed or migrated.
5139 static int __init cmdline_parse_kernelcore(char *p)
5141 return cmdline_parse_core(p, &required_kernelcore);
5145 * movablecore=size sets the amount of memory for use for allocations that
5146 * can be reclaimed or migrated.
5148 static int __init cmdline_parse_movablecore(char *p)
5150 return cmdline_parse_core(p, &required_movablecore);
5153 early_param("kernelcore", cmdline_parse_kernelcore);
5154 early_param("movablecore", cmdline_parse_movablecore);
5157 * insert_movablemem_map - Insert a memory range in to movablemem_map.map.
5158 * @start_pfn: start pfn of the range
5159 * @end_pfn: end pfn of the range
5161 * This function will also merge the overlapped ranges, and sort the array
5162 * by start_pfn in monotonic increasing order.
5164 static void __init insert_movablemem_map(unsigned long start_pfn,
5165 unsigned long end_pfn)
5170 * pos will be at the 1st overlapped range, or the position
5171 * where the element should be inserted.
5173 for (pos = 0; pos < movablemem_map.nr_map; pos++)
5174 if (start_pfn <= movablemem_map.map[pos].end_pfn)
5177 /* If there is no overlapped range, just insert the element. */
5178 if (pos == movablemem_map.nr_map ||
5179 end_pfn < movablemem_map.map[pos].start_pfn) {
5181 * If pos is not the end of array, we need to move all
5182 * the rest elements backward.
5184 if (pos < movablemem_map.nr_map)
5185 memmove(&movablemem_map.map[pos+1],
5186 &movablemem_map.map[pos],
5187 sizeof(struct movablemem_entry) *
5188 (movablemem_map.nr_map - pos));
5189 movablemem_map.map[pos].start_pfn = start_pfn;
5190 movablemem_map.map[pos].end_pfn = end_pfn;
5191 movablemem_map.nr_map++;
5195 /* overlap will be at the last overlapped range */
5196 for (overlap = pos + 1; overlap < movablemem_map.nr_map; overlap++)
5197 if (end_pfn < movablemem_map.map[overlap].start_pfn)
5201 * If there are more ranges overlapped, we need to merge them,
5202 * and move the rest elements forward.
5205 movablemem_map.map[pos].start_pfn = min(start_pfn,
5206 movablemem_map.map[pos].start_pfn);
5207 movablemem_map.map[pos].end_pfn = max(end_pfn,
5208 movablemem_map.map[overlap].end_pfn);
5210 if (pos != overlap && overlap + 1 != movablemem_map.nr_map)
5211 memmove(&movablemem_map.map[pos+1],
5212 &movablemem_map.map[overlap+1],
5213 sizeof(struct movablemem_entry) *
5214 (movablemem_map.nr_map - overlap - 1));
5216 movablemem_map.nr_map -= overlap - pos;
5220 * movablemem_map_add_region - Add a memory range into movablemem_map.
5221 * @start: physical start address of range
5222 * @end: physical end address of range
5224 * This function transform the physical address into pfn, and then add the
5225 * range into movablemem_map by calling insert_movablemem_map().
5227 static void __init movablemem_map_add_region(u64 start, u64 size)
5229 unsigned long start_pfn, end_pfn;
5231 /* In case size == 0 or start + size overflows */
5232 if (start + size <= start)
5235 if (movablemem_map.nr_map >= ARRAY_SIZE(movablemem_map.map)) {
5236 pr_err("movablemem_map: too many entries;"
5237 " ignoring [mem %#010llx-%#010llx]\n",
5238 (unsigned long long) start,
5239 (unsigned long long) (start + size - 1));
5243 start_pfn = PFN_DOWN(start);
5244 end_pfn = PFN_UP(start + size);
5245 insert_movablemem_map(start_pfn, end_pfn);
5249 * cmdline_parse_movablemem_map - Parse boot option movablemem_map.
5250 * @p: The boot option of the following format:
5251 * movablemem_map=nn[KMG]@ss[KMG]
5253 * This option sets the memory range [ss, ss+nn) to be used as movable memory.
5255 * Return: 0 on success or -EINVAL on failure.
5257 static int __init cmdline_parse_movablemem_map(char *p)
5260 u64 start_at, mem_size;
5266 mem_size = memparse(p, &p);
5272 start_at = memparse(p, &p);
5273 if (p == oldp || *p != '\0')
5276 movablemem_map_add_region(start_at, mem_size);
5282 early_param("movablemem_map", cmdline_parse_movablemem_map);
5284 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5287 * set_dma_reserve - set the specified number of pages reserved in the first zone
5288 * @new_dma_reserve: The number of pages to mark reserved
5290 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5291 * In the DMA zone, a significant percentage may be consumed by kernel image
5292 * and other unfreeable allocations which can skew the watermarks badly. This
5293 * function may optionally be used to account for unfreeable pages in the
5294 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5295 * smaller per-cpu batchsize.
5297 void __init set_dma_reserve(unsigned long new_dma_reserve)
5299 dma_reserve = new_dma_reserve;
5302 void __init free_area_init(unsigned long *zones_size)
5304 free_area_init_node(0, zones_size,
5305 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5308 static int page_alloc_cpu_notify(struct notifier_block *self,
5309 unsigned long action, void *hcpu)
5311 int cpu = (unsigned long)hcpu;
5313 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5314 lru_add_drain_cpu(cpu);
5318 * Spill the event counters of the dead processor
5319 * into the current processors event counters.
5320 * This artificially elevates the count of the current
5323 vm_events_fold_cpu(cpu);
5326 * Zero the differential counters of the dead processor
5327 * so that the vm statistics are consistent.
5329 * This is only okay since the processor is dead and cannot
5330 * race with what we are doing.
5332 refresh_cpu_vm_stats(cpu);
5337 void __init page_alloc_init(void)
5339 hotcpu_notifier(page_alloc_cpu_notify, 0);
5343 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5344 * or min_free_kbytes changes.
5346 static void calculate_totalreserve_pages(void)
5348 struct pglist_data *pgdat;
5349 unsigned long reserve_pages = 0;
5350 enum zone_type i, j;
5352 for_each_online_pgdat(pgdat) {
5353 for (i = 0; i < MAX_NR_ZONES; i++) {
5354 struct zone *zone = pgdat->node_zones + i;
5355 unsigned long max = 0;
5357 /* Find valid and maximum lowmem_reserve in the zone */
5358 for (j = i; j < MAX_NR_ZONES; j++) {
5359 if (zone->lowmem_reserve[j] > max)
5360 max = zone->lowmem_reserve[j];
5363 /* we treat the high watermark as reserved pages. */
5364 max += high_wmark_pages(zone);
5366 if (max > zone->present_pages)
5367 max = zone->present_pages;
5368 reserve_pages += max;
5370 * Lowmem reserves are not available to
5371 * GFP_HIGHUSER page cache allocations and
5372 * kswapd tries to balance zones to their high
5373 * watermark. As a result, neither should be
5374 * regarded as dirtyable memory, to prevent a
5375 * situation where reclaim has to clean pages
5376 * in order to balance the zones.
5378 zone->dirty_balance_reserve = max;
5381 dirty_balance_reserve = reserve_pages;
5382 totalreserve_pages = reserve_pages;
5386 * setup_per_zone_lowmem_reserve - called whenever
5387 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5388 * has a correct pages reserved value, so an adequate number of
5389 * pages are left in the zone after a successful __alloc_pages().
5391 static void setup_per_zone_lowmem_reserve(void)
5393 struct pglist_data *pgdat;
5394 enum zone_type j, idx;
5396 for_each_online_pgdat(pgdat) {
5397 for (j = 0; j < MAX_NR_ZONES; j++) {
5398 struct zone *zone = pgdat->node_zones + j;
5399 unsigned long present_pages = zone->present_pages;
5401 zone->lowmem_reserve[j] = 0;
5405 struct zone *lower_zone;
5409 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5410 sysctl_lowmem_reserve_ratio[idx] = 1;
5412 lower_zone = pgdat->node_zones + idx;
5413 lower_zone->lowmem_reserve[j] = present_pages /
5414 sysctl_lowmem_reserve_ratio[idx];
5415 present_pages += lower_zone->present_pages;
5420 /* update totalreserve_pages */
5421 calculate_totalreserve_pages();
5424 static void __setup_per_zone_wmarks(void)
5426 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5427 unsigned long lowmem_pages = 0;
5429 unsigned long flags;
5431 /* Calculate total number of !ZONE_HIGHMEM pages */
5432 for_each_zone(zone) {
5433 if (!is_highmem(zone))
5434 lowmem_pages += zone->present_pages;
5437 for_each_zone(zone) {
5440 spin_lock_irqsave(&zone->lock, flags);
5441 tmp = (u64)pages_min * zone->present_pages;
5442 do_div(tmp, lowmem_pages);
5443 if (is_highmem(zone)) {
5445 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5446 * need highmem pages, so cap pages_min to a small
5449 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5450 * deltas controls asynch page reclaim, and so should
5451 * not be capped for highmem.
5453 unsigned long min_pages;
5455 min_pages = zone->present_pages / 1024;
5456 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5457 zone->watermark[WMARK_MIN] = min_pages;
5460 * If it's a lowmem zone, reserve a number of pages
5461 * proportionate to the zone's size.
5463 zone->watermark[WMARK_MIN] = tmp;
5466 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5467 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5469 setup_zone_migrate_reserve(zone);
5470 spin_unlock_irqrestore(&zone->lock, flags);
5473 /* update totalreserve_pages */
5474 calculate_totalreserve_pages();
5478 * setup_per_zone_wmarks - called when min_free_kbytes changes
5479 * or when memory is hot-{added|removed}
5481 * Ensures that the watermark[min,low,high] values for each zone are set
5482 * correctly with respect to min_free_kbytes.
5484 void setup_per_zone_wmarks(void)
5486 mutex_lock(&zonelists_mutex);
5487 __setup_per_zone_wmarks();
5488 mutex_unlock(&zonelists_mutex);
5492 * The inactive anon list should be small enough that the VM never has to
5493 * do too much work, but large enough that each inactive page has a chance
5494 * to be referenced again before it is swapped out.
5496 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5497 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5498 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5499 * the anonymous pages are kept on the inactive list.
5502 * memory ratio inactive anon
5503 * -------------------------------------
5512 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5514 unsigned int gb, ratio;
5516 /* Zone size in gigabytes */
5517 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5519 ratio = int_sqrt(10 * gb);
5523 zone->inactive_ratio = ratio;
5526 static void __meminit setup_per_zone_inactive_ratio(void)
5531 calculate_zone_inactive_ratio(zone);
5535 * Initialise min_free_kbytes.
5537 * For small machines we want it small (128k min). For large machines
5538 * we want it large (64MB max). But it is not linear, because network
5539 * bandwidth does not increase linearly with machine size. We use
5541 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5542 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5558 int __meminit init_per_zone_wmark_min(void)
5560 unsigned long lowmem_kbytes;
5562 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5564 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5565 if (min_free_kbytes < 128)
5566 min_free_kbytes = 128;
5567 if (min_free_kbytes > 65536)
5568 min_free_kbytes = 65536;
5569 setup_per_zone_wmarks();
5570 refresh_zone_stat_thresholds();
5571 setup_per_zone_lowmem_reserve();
5572 setup_per_zone_inactive_ratio();
5575 module_init(init_per_zone_wmark_min)
5578 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5579 * that we can call two helper functions whenever min_free_kbytes
5582 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5583 void __user *buffer, size_t *length, loff_t *ppos)
5585 proc_dointvec(table, write, buffer, length, ppos);
5587 setup_per_zone_wmarks();
5592 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5593 void __user *buffer, size_t *length, loff_t *ppos)
5598 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5603 zone->min_unmapped_pages = (zone->present_pages *
5604 sysctl_min_unmapped_ratio) / 100;
5608 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5609 void __user *buffer, size_t *length, loff_t *ppos)
5614 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5619 zone->min_slab_pages = (zone->present_pages *
5620 sysctl_min_slab_ratio) / 100;
5626 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5627 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5628 * whenever sysctl_lowmem_reserve_ratio changes.
5630 * The reserve ratio obviously has absolutely no relation with the
5631 * minimum watermarks. The lowmem reserve ratio can only make sense
5632 * if in function of the boot time zone sizes.
5634 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5635 void __user *buffer, size_t *length, loff_t *ppos)
5637 proc_dointvec_minmax(table, write, buffer, length, ppos);
5638 setup_per_zone_lowmem_reserve();
5643 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5644 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5645 * can have before it gets flushed back to buddy allocator.
5648 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5649 void __user *buffer, size_t *length, loff_t *ppos)
5655 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5656 if (!write || (ret < 0))
5658 for_each_populated_zone(zone) {
5659 for_each_possible_cpu(cpu) {
5661 high = zone->present_pages / percpu_pagelist_fraction;
5662 setup_pagelist_highmark(
5663 per_cpu_ptr(zone->pageset, cpu), high);
5669 int hashdist = HASHDIST_DEFAULT;
5672 static int __init set_hashdist(char *str)
5676 hashdist = simple_strtoul(str, &str, 0);
5679 __setup("hashdist=", set_hashdist);
5683 * allocate a large system hash table from bootmem
5684 * - it is assumed that the hash table must contain an exact power-of-2
5685 * quantity of entries
5686 * - limit is the number of hash buckets, not the total allocation size
5688 void *__init alloc_large_system_hash(const char *tablename,
5689 unsigned long bucketsize,
5690 unsigned long numentries,
5693 unsigned int *_hash_shift,
5694 unsigned int *_hash_mask,
5695 unsigned long low_limit,
5696 unsigned long high_limit)
5698 unsigned long long max = high_limit;
5699 unsigned long log2qty, size;
5702 /* allow the kernel cmdline to have a say */
5704 /* round applicable memory size up to nearest megabyte */
5705 numentries = nr_kernel_pages;
5706 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5707 numentries >>= 20 - PAGE_SHIFT;
5708 numentries <<= 20 - PAGE_SHIFT;
5710 /* limit to 1 bucket per 2^scale bytes of low memory */
5711 if (scale > PAGE_SHIFT)
5712 numentries >>= (scale - PAGE_SHIFT);
5714 numentries <<= (PAGE_SHIFT - scale);
5716 /* Make sure we've got at least a 0-order allocation.. */
5717 if (unlikely(flags & HASH_SMALL)) {
5718 /* Makes no sense without HASH_EARLY */
5719 WARN_ON(!(flags & HASH_EARLY));
5720 if (!(numentries >> *_hash_shift)) {
5721 numentries = 1UL << *_hash_shift;
5722 BUG_ON(!numentries);
5724 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5725 numentries = PAGE_SIZE / bucketsize;
5727 numentries = roundup_pow_of_two(numentries);
5729 /* limit allocation size to 1/16 total memory by default */
5731 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5732 do_div(max, bucketsize);
5734 max = min(max, 0x80000000ULL);
5736 if (numentries < low_limit)
5737 numentries = low_limit;
5738 if (numentries > max)
5741 log2qty = ilog2(numentries);
5744 size = bucketsize << log2qty;
5745 if (flags & HASH_EARLY)
5746 table = alloc_bootmem_nopanic(size);
5748 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5751 * If bucketsize is not a power-of-two, we may free
5752 * some pages at the end of hash table which
5753 * alloc_pages_exact() automatically does
5755 if (get_order(size) < MAX_ORDER) {
5756 table = alloc_pages_exact(size, GFP_ATOMIC);
5757 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5760 } while (!table && size > PAGE_SIZE && --log2qty);
5763 panic("Failed to allocate %s hash table\n", tablename);
5765 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5768 ilog2(size) - PAGE_SHIFT,
5772 *_hash_shift = log2qty;
5774 *_hash_mask = (1 << log2qty) - 1;
5779 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5780 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5783 #ifdef CONFIG_SPARSEMEM
5784 return __pfn_to_section(pfn)->pageblock_flags;
5786 return zone->pageblock_flags;
5787 #endif /* CONFIG_SPARSEMEM */
5790 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5792 #ifdef CONFIG_SPARSEMEM
5793 pfn &= (PAGES_PER_SECTION-1);
5794 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5796 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
5797 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5798 #endif /* CONFIG_SPARSEMEM */
5802 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5803 * @page: The page within the block of interest
5804 * @start_bitidx: The first bit of interest to retrieve
5805 * @end_bitidx: The last bit of interest
5806 * returns pageblock_bits flags
5808 unsigned long get_pageblock_flags_group(struct page *page,
5809 int start_bitidx, int end_bitidx)
5812 unsigned long *bitmap;
5813 unsigned long pfn, bitidx;
5814 unsigned long flags = 0;
5815 unsigned long value = 1;
5817 zone = page_zone(page);
5818 pfn = page_to_pfn(page);
5819 bitmap = get_pageblock_bitmap(zone, pfn);
5820 bitidx = pfn_to_bitidx(zone, pfn);
5822 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5823 if (test_bit(bitidx + start_bitidx, bitmap))
5830 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5831 * @page: The page within the block of interest
5832 * @start_bitidx: The first bit of interest
5833 * @end_bitidx: The last bit of interest
5834 * @flags: The flags to set
5836 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5837 int start_bitidx, int end_bitidx)
5840 unsigned long *bitmap;
5841 unsigned long pfn, bitidx;
5842 unsigned long value = 1;
5844 zone = page_zone(page);
5845 pfn = page_to_pfn(page);
5846 bitmap = get_pageblock_bitmap(zone, pfn);
5847 bitidx = pfn_to_bitidx(zone, pfn);
5848 VM_BUG_ON(pfn < zone->zone_start_pfn);
5849 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5851 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5853 __set_bit(bitidx + start_bitidx, bitmap);
5855 __clear_bit(bitidx + start_bitidx, bitmap);
5859 * This function checks whether pageblock includes unmovable pages or not.
5860 * If @count is not zero, it is okay to include less @count unmovable pages
5862 * PageLRU check wihtout isolation or lru_lock could race so that
5863 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5864 * expect this function should be exact.
5866 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
5867 bool skip_hwpoisoned_pages)
5869 unsigned long pfn, iter, found;
5873 * For avoiding noise data, lru_add_drain_all() should be called
5874 * If ZONE_MOVABLE, the zone never contains unmovable pages
5876 if (zone_idx(zone) == ZONE_MOVABLE)
5878 mt = get_pageblock_migratetype(page);
5879 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
5882 pfn = page_to_pfn(page);
5883 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5884 unsigned long check = pfn + iter;
5886 if (!pfn_valid_within(check))
5889 page = pfn_to_page(check);
5891 * We can't use page_count without pin a page
5892 * because another CPU can free compound page.
5893 * This check already skips compound tails of THP
5894 * because their page->_count is zero at all time.
5896 if (!atomic_read(&page->_count)) {
5897 if (PageBuddy(page))
5898 iter += (1 << page_order(page)) - 1;
5903 * The HWPoisoned page may be not in buddy system, and
5904 * page_count() is not 0.
5906 if (skip_hwpoisoned_pages && PageHWPoison(page))
5912 * If there are RECLAIMABLE pages, we need to check it.
5913 * But now, memory offline itself doesn't call shrink_slab()
5914 * and it still to be fixed.
5917 * If the page is not RAM, page_count()should be 0.
5918 * we don't need more check. This is an _used_ not-movable page.
5920 * The problematic thing here is PG_reserved pages. PG_reserved
5921 * is set to both of a memory hole page and a _used_ kernel
5930 bool is_pageblock_removable_nolock(struct page *page)
5936 * We have to be careful here because we are iterating over memory
5937 * sections which are not zone aware so we might end up outside of
5938 * the zone but still within the section.
5939 * We have to take care about the node as well. If the node is offline
5940 * its NODE_DATA will be NULL - see page_zone.
5942 if (!node_online(page_to_nid(page)))
5945 zone = page_zone(page);
5946 pfn = page_to_pfn(page);
5947 if (zone->zone_start_pfn > pfn ||
5948 zone->zone_start_pfn + zone->spanned_pages <= pfn)
5951 return !has_unmovable_pages(zone, page, 0, true);
5956 static unsigned long pfn_max_align_down(unsigned long pfn)
5958 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
5959 pageblock_nr_pages) - 1);
5962 static unsigned long pfn_max_align_up(unsigned long pfn)
5964 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
5965 pageblock_nr_pages));
5968 /* [start, end) must belong to a single zone. */
5969 static int __alloc_contig_migrate_range(struct compact_control *cc,
5970 unsigned long start, unsigned long end)
5972 /* This function is based on compact_zone() from compaction.c. */
5973 unsigned long nr_reclaimed;
5974 unsigned long pfn = start;
5975 unsigned int tries = 0;
5980 while (pfn < end || !list_empty(&cc->migratepages)) {
5981 if (fatal_signal_pending(current)) {
5986 if (list_empty(&cc->migratepages)) {
5987 cc->nr_migratepages = 0;
5988 pfn = isolate_migratepages_range(cc->zone, cc,
5995 } else if (++tries == 5) {
5996 ret = ret < 0 ? ret : -EBUSY;
6000 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6002 cc->nr_migratepages -= nr_reclaimed;
6004 ret = migrate_pages(&cc->migratepages,
6005 alloc_migrate_target,
6006 0, false, MIGRATE_SYNC,
6010 putback_movable_pages(&cc->migratepages);
6017 * alloc_contig_range() -- tries to allocate given range of pages
6018 * @start: start PFN to allocate
6019 * @end: one-past-the-last PFN to allocate
6020 * @migratetype: migratetype of the underlaying pageblocks (either
6021 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6022 * in range must have the same migratetype and it must
6023 * be either of the two.
6025 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6026 * aligned, however it's the caller's responsibility to guarantee that
6027 * we are the only thread that changes migrate type of pageblocks the
6030 * The PFN range must belong to a single zone.
6032 * Returns zero on success or negative error code. On success all
6033 * pages which PFN is in [start, end) are allocated for the caller and
6034 * need to be freed with free_contig_range().
6036 int alloc_contig_range(unsigned long start, unsigned long end,
6037 unsigned migratetype)
6039 unsigned long outer_start, outer_end;
6042 struct compact_control cc = {
6043 .nr_migratepages = 0,
6045 .zone = page_zone(pfn_to_page(start)),
6047 .ignore_skip_hint = true,
6049 INIT_LIST_HEAD(&cc.migratepages);
6052 * What we do here is we mark all pageblocks in range as
6053 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6054 * have different sizes, and due to the way page allocator
6055 * work, we align the range to biggest of the two pages so
6056 * that page allocator won't try to merge buddies from
6057 * different pageblocks and change MIGRATE_ISOLATE to some
6058 * other migration type.
6060 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6061 * migrate the pages from an unaligned range (ie. pages that
6062 * we are interested in). This will put all the pages in
6063 * range back to page allocator as MIGRATE_ISOLATE.
6065 * When this is done, we take the pages in range from page
6066 * allocator removing them from the buddy system. This way
6067 * page allocator will never consider using them.
6069 * This lets us mark the pageblocks back as
6070 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6071 * aligned range but not in the unaligned, original range are
6072 * put back to page allocator so that buddy can use them.
6075 ret = start_isolate_page_range(pfn_max_align_down(start),
6076 pfn_max_align_up(end), migratetype,
6081 ret = __alloc_contig_migrate_range(&cc, start, end);
6086 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6087 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6088 * more, all pages in [start, end) are free in page allocator.
6089 * What we are going to do is to allocate all pages from
6090 * [start, end) (that is remove them from page allocator).
6092 * The only problem is that pages at the beginning and at the
6093 * end of interesting range may be not aligned with pages that
6094 * page allocator holds, ie. they can be part of higher order
6095 * pages. Because of this, we reserve the bigger range and
6096 * once this is done free the pages we are not interested in.
6098 * We don't have to hold zone->lock here because the pages are
6099 * isolated thus they won't get removed from buddy.
6102 lru_add_drain_all();
6106 outer_start = start;
6107 while (!PageBuddy(pfn_to_page(outer_start))) {
6108 if (++order >= MAX_ORDER) {
6112 outer_start &= ~0UL << order;
6115 /* Make sure the range is really isolated. */
6116 if (test_pages_isolated(outer_start, end, false)) {
6117 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6124 /* Grab isolated pages from freelists. */
6125 outer_end = isolate_freepages_range(&cc, outer_start, end);
6131 /* Free head and tail (if any) */
6132 if (start != outer_start)
6133 free_contig_range(outer_start, start - outer_start);
6134 if (end != outer_end)
6135 free_contig_range(end, outer_end - end);
6138 undo_isolate_page_range(pfn_max_align_down(start),
6139 pfn_max_align_up(end), migratetype);
6143 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6145 unsigned int count = 0;
6147 for (; nr_pages--; pfn++) {
6148 struct page *page = pfn_to_page(pfn);
6150 count += page_count(page) != 1;
6153 WARN(count != 0, "%d pages are still in use!\n", count);
6157 #ifdef CONFIG_MEMORY_HOTPLUG
6158 static int __meminit __zone_pcp_update(void *data)
6160 struct zone *zone = data;
6162 unsigned long batch = zone_batchsize(zone), flags;
6164 for_each_possible_cpu(cpu) {
6165 struct per_cpu_pageset *pset;
6166 struct per_cpu_pages *pcp;
6168 pset = per_cpu_ptr(zone->pageset, cpu);
6171 local_irq_save(flags);
6173 free_pcppages_bulk(zone, pcp->count, pcp);
6174 drain_zonestat(zone, pset);
6175 setup_pageset(pset, batch);
6176 local_irq_restore(flags);
6181 void __meminit zone_pcp_update(struct zone *zone)
6183 stop_machine(__zone_pcp_update, zone, NULL);
6187 void zone_pcp_reset(struct zone *zone)
6189 unsigned long flags;
6191 struct per_cpu_pageset *pset;
6193 /* avoid races with drain_pages() */
6194 local_irq_save(flags);
6195 if (zone->pageset != &boot_pageset) {
6196 for_each_online_cpu(cpu) {
6197 pset = per_cpu_ptr(zone->pageset, cpu);
6198 drain_zonestat(zone, pset);
6200 free_percpu(zone->pageset);
6201 zone->pageset = &boot_pageset;
6203 local_irq_restore(flags);
6206 #ifdef CONFIG_MEMORY_HOTREMOVE
6208 * All pages in the range must be isolated before calling this.
6211 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6217 unsigned long flags;
6218 /* find the first valid pfn */
6219 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6224 zone = page_zone(pfn_to_page(pfn));
6225 spin_lock_irqsave(&zone->lock, flags);
6227 while (pfn < end_pfn) {
6228 if (!pfn_valid(pfn)) {
6232 page = pfn_to_page(pfn);
6234 * The HWPoisoned page may be not in buddy system, and
6235 * page_count() is not 0.
6237 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6239 SetPageReserved(page);
6243 BUG_ON(page_count(page));
6244 BUG_ON(!PageBuddy(page));
6245 order = page_order(page);
6246 #ifdef CONFIG_DEBUG_VM
6247 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6248 pfn, 1 << order, end_pfn);
6250 list_del(&page->lru);
6251 rmv_page_order(page);
6252 zone->free_area[order].nr_free--;
6253 for (i = 0; i < (1 << order); i++)
6254 SetPageReserved((page+i));
6255 pfn += (1 << order);
6257 spin_unlock_irqrestore(&zone->lock, flags);
6261 #ifdef CONFIG_MEMORY_FAILURE
6262 bool is_free_buddy_page(struct page *page)
6264 struct zone *zone = page_zone(page);
6265 unsigned long pfn = page_to_pfn(page);
6266 unsigned long flags;
6269 spin_lock_irqsave(&zone->lock, flags);
6270 for (order = 0; order < MAX_ORDER; order++) {
6271 struct page *page_head = page - (pfn & ((1 << order) - 1));
6273 if (PageBuddy(page_head) && page_order(page_head) >= order)
6276 spin_unlock_irqrestore(&zone->lock, flags);
6278 return order < MAX_ORDER;
6282 static const struct trace_print_flags pageflag_names[] = {
6283 {1UL << PG_locked, "locked" },
6284 {1UL << PG_error, "error" },
6285 {1UL << PG_referenced, "referenced" },
6286 {1UL << PG_uptodate, "uptodate" },
6287 {1UL << PG_dirty, "dirty" },
6288 {1UL << PG_lru, "lru" },
6289 {1UL << PG_active, "active" },
6290 {1UL << PG_slab, "slab" },
6291 {1UL << PG_owner_priv_1, "owner_priv_1" },
6292 {1UL << PG_arch_1, "arch_1" },
6293 {1UL << PG_reserved, "reserved" },
6294 {1UL << PG_private, "private" },
6295 {1UL << PG_private_2, "private_2" },
6296 {1UL << PG_writeback, "writeback" },
6297 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6298 {1UL << PG_head, "head" },
6299 {1UL << PG_tail, "tail" },
6301 {1UL << PG_compound, "compound" },
6303 {1UL << PG_swapcache, "swapcache" },
6304 {1UL << PG_mappedtodisk, "mappedtodisk" },
6305 {1UL << PG_reclaim, "reclaim" },
6306 {1UL << PG_swapbacked, "swapbacked" },
6307 {1UL << PG_unevictable, "unevictable" },
6309 {1UL << PG_mlocked, "mlocked" },
6311 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6312 {1UL << PG_uncached, "uncached" },
6314 #ifdef CONFIG_MEMORY_FAILURE
6315 {1UL << PG_hwpoison, "hwpoison" },
6317 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6318 {1UL << PG_compound_lock, "compound_lock" },
6322 static void dump_page_flags(unsigned long flags)
6324 const char *delim = "";
6328 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6330 printk(KERN_ALERT "page flags: %#lx(", flags);
6332 /* remove zone id */
6333 flags &= (1UL << NR_PAGEFLAGS) - 1;
6335 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6337 mask = pageflag_names[i].mask;
6338 if ((flags & mask) != mask)
6342 printk("%s%s", delim, pageflag_names[i].name);
6346 /* check for left over flags */
6348 printk("%s%#lx", delim, flags);
6353 void dump_page(struct page *page)
6356 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6357 page, atomic_read(&page->_count), page_mapcount(page),
6358 page->mapping, page->index);
6359 dump_page_flags(page->flags);
6360 mem_cgroup_print_bad_page(page);