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/compiler.h>
25 #include <linux/kernel.h>
26 #include <linux/module.h>
27 #include <linux/suspend.h>
28 #include <linux/pagevec.h>
29 #include <linux/blkdev.h>
30 #include <linux/slab.h>
31 #include <linux/oom.h>
32 #include <linux/notifier.h>
33 #include <linux/topology.h>
34 #include <linux/sysctl.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/memory_hotplug.h>
38 #include <linux/nodemask.h>
39 #include <linux/vmalloc.h>
40 #include <linux/mempolicy.h>
41 #include <linux/stop_machine.h>
42 #include <linux/sort.h>
43 #include <linux/pfn.h>
44 #include <linux/backing-dev.h>
45 #include <linux/fault-inject.h>
46 #include <linux/page-isolation.h>
47 #include <linux/page_cgroup.h>
48 #include <linux/debugobjects.h>
49 #include <linux/kmemleak.h>
51 #include <asm/tlbflush.h>
52 #include <asm/div64.h>
56 * Array of node states.
58 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
59 [N_POSSIBLE] = NODE_MASK_ALL,
60 [N_ONLINE] = { { [0] = 1UL } },
62 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
64 [N_HIGH_MEMORY] = { { [0] = 1UL } },
66 [N_CPU] = { { [0] = 1UL } },
69 EXPORT_SYMBOL(node_states);
71 unsigned long totalram_pages __read_mostly;
72 unsigned long totalreserve_pages __read_mostly;
73 unsigned long highest_memmap_pfn __read_mostly;
74 int percpu_pagelist_fraction;
76 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
77 int pageblock_order __read_mostly;
80 static void __free_pages_ok(struct page *page, unsigned int order);
83 * results with 256, 32 in the lowmem_reserve sysctl:
84 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
85 * 1G machine -> (16M dma, 784M normal, 224M high)
86 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
87 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
88 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
90 * TBD: should special case ZONE_DMA32 machines here - in those we normally
91 * don't need any ZONE_NORMAL reservation
93 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
94 #ifdef CONFIG_ZONE_DMA
97 #ifdef CONFIG_ZONE_DMA32
100 #ifdef CONFIG_HIGHMEM
106 EXPORT_SYMBOL(totalram_pages);
108 static char * const zone_names[MAX_NR_ZONES] = {
109 #ifdef CONFIG_ZONE_DMA
112 #ifdef CONFIG_ZONE_DMA32
116 #ifdef CONFIG_HIGHMEM
122 int min_free_kbytes = 1024;
124 unsigned long __meminitdata nr_kernel_pages;
125 unsigned long __meminitdata nr_all_pages;
126 static unsigned long __meminitdata dma_reserve;
128 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
130 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
131 * ranges of memory (RAM) that may be registered with add_active_range().
132 * Ranges passed to add_active_range() will be merged if possible
133 * so the number of times add_active_range() can be called is
134 * related to the number of nodes and the number of holes
136 #ifdef CONFIG_MAX_ACTIVE_REGIONS
137 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
138 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
140 #if MAX_NUMNODES >= 32
141 /* If there can be many nodes, allow up to 50 holes per node */
142 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
144 /* By default, allow up to 256 distinct regions */
145 #define MAX_ACTIVE_REGIONS 256
149 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
150 static int __meminitdata nr_nodemap_entries;
151 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
152 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
153 static unsigned long __initdata required_kernelcore;
154 static unsigned long __initdata required_movablecore;
155 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
157 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
159 EXPORT_SYMBOL(movable_zone);
160 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
163 int nr_node_ids __read_mostly = MAX_NUMNODES;
164 EXPORT_SYMBOL(nr_node_ids);
167 int page_group_by_mobility_disabled __read_mostly;
169 static void set_pageblock_migratetype(struct page *page, int migratetype)
172 if (unlikely(page_group_by_mobility_disabled))
173 migratetype = MIGRATE_UNMOVABLE;
175 set_pageblock_flags_group(page, (unsigned long)migratetype,
176 PB_migrate, PB_migrate_end);
179 #ifdef CONFIG_DEBUG_VM
180 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
184 unsigned long pfn = page_to_pfn(page);
187 seq = zone_span_seqbegin(zone);
188 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
190 else if (pfn < zone->zone_start_pfn)
192 } while (zone_span_seqretry(zone, seq));
197 static int page_is_consistent(struct zone *zone, struct page *page)
199 if (!pfn_valid_within(page_to_pfn(page)))
201 if (zone != page_zone(page))
207 * Temporary debugging check for pages not lying within a given zone.
209 static int bad_range(struct zone *zone, struct page *page)
211 if (page_outside_zone_boundaries(zone, page))
213 if (!page_is_consistent(zone, page))
219 static inline int bad_range(struct zone *zone, struct page *page)
225 static void bad_page(struct page *page)
227 static unsigned long resume;
228 static unsigned long nr_shown;
229 static unsigned long nr_unshown;
232 * Allow a burst of 60 reports, then keep quiet for that minute;
233 * or allow a steady drip of one report per second.
235 if (nr_shown == 60) {
236 if (time_before(jiffies, resume)) {
242 "BUG: Bad page state: %lu messages suppressed\n",
249 resume = jiffies + 60 * HZ;
251 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
252 current->comm, page_to_pfn(page));
254 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
255 page, (void *)page->flags, page_count(page),
256 page_mapcount(page), page->mapping, page->index);
260 /* Leave bad fields for debug, except PageBuddy could make trouble */
261 __ClearPageBuddy(page);
262 add_taint(TAINT_BAD_PAGE);
266 * Higher-order pages are called "compound pages". They are structured thusly:
268 * The first PAGE_SIZE page is called the "head page".
270 * The remaining PAGE_SIZE pages are called "tail pages".
272 * All pages have PG_compound set. All pages have their ->private pointing at
273 * the head page (even the head page has this).
275 * The first tail page's ->lru.next holds the address of the compound page's
276 * put_page() function. Its ->lru.prev holds the order of allocation.
277 * This usage means that zero-order pages may not be compound.
280 static void free_compound_page(struct page *page)
282 __free_pages_ok(page, compound_order(page));
285 void prep_compound_page(struct page *page, unsigned long order)
288 int nr_pages = 1 << order;
290 set_compound_page_dtor(page, free_compound_page);
291 set_compound_order(page, order);
293 for (i = 1; i < nr_pages; i++) {
294 struct page *p = page + i;
297 p->first_page = page;
301 #ifdef CONFIG_HUGETLBFS
302 void prep_compound_gigantic_page(struct page *page, unsigned long order)
305 int nr_pages = 1 << order;
306 struct page *p = page + 1;
308 set_compound_page_dtor(page, free_compound_page);
309 set_compound_order(page, order);
311 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
313 p->first_page = page;
318 static int destroy_compound_page(struct page *page, unsigned long order)
321 int nr_pages = 1 << order;
324 if (unlikely(compound_order(page) != order) ||
325 unlikely(!PageHead(page))) {
330 __ClearPageHead(page);
332 for (i = 1; i < nr_pages; i++) {
333 struct page *p = page + i;
335 if (unlikely(!PageTail(p) || (p->first_page != page))) {
345 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
350 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
351 * and __GFP_HIGHMEM from hard or soft interrupt context.
353 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
354 for (i = 0; i < (1 << order); i++)
355 clear_highpage(page + i);
358 static inline void set_page_order(struct page *page, int order)
360 set_page_private(page, order);
361 __SetPageBuddy(page);
364 static inline void rmv_page_order(struct page *page)
366 __ClearPageBuddy(page);
367 set_page_private(page, 0);
371 * Locate the struct page for both the matching buddy in our
372 * pair (buddy1) and the combined O(n+1) page they form (page).
374 * 1) Any buddy B1 will have an order O twin B2 which satisfies
375 * the following equation:
377 * For example, if the starting buddy (buddy2) is #8 its order
379 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
381 * 2) Any buddy B will have an order O+1 parent P which
382 * satisfies the following equation:
385 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
387 static inline struct page *
388 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
390 unsigned long buddy_idx = page_idx ^ (1 << order);
392 return page + (buddy_idx - page_idx);
395 static inline unsigned long
396 __find_combined_index(unsigned long page_idx, unsigned int order)
398 return (page_idx & ~(1 << order));
402 * This function checks whether a page is free && is the buddy
403 * we can do coalesce a page and its buddy if
404 * (a) the buddy is not in a hole &&
405 * (b) the buddy is in the buddy system &&
406 * (c) a page and its buddy have the same order &&
407 * (d) a page and its buddy are in the same zone.
409 * For recording whether a page is in the buddy system, we use PG_buddy.
410 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
412 * For recording page's order, we use page_private(page).
414 static inline int page_is_buddy(struct page *page, struct page *buddy,
417 if (!pfn_valid_within(page_to_pfn(buddy)))
420 if (page_zone_id(page) != page_zone_id(buddy))
423 if (PageBuddy(buddy) && page_order(buddy) == order) {
424 BUG_ON(page_count(buddy) != 0);
431 * Freeing function for a buddy system allocator.
433 * The concept of a buddy system is to maintain direct-mapped table
434 * (containing bit values) for memory blocks of various "orders".
435 * The bottom level table contains the map for the smallest allocatable
436 * units of memory (here, pages), and each level above it describes
437 * pairs of units from the levels below, hence, "buddies".
438 * At a high level, all that happens here is marking the table entry
439 * at the bottom level available, and propagating the changes upward
440 * as necessary, plus some accounting needed to play nicely with other
441 * parts of the VM system.
442 * At each level, we keep a list of pages, which are heads of continuous
443 * free pages of length of (1 << order) and marked with PG_buddy. Page's
444 * order is recorded in page_private(page) field.
445 * So when we are allocating or freeing one, we can derive the state of the
446 * other. That is, if we allocate a small block, and both were
447 * free, the remainder of the region must be split into blocks.
448 * If a block is freed, and its buddy is also free, then this
449 * triggers coalescing into a block of larger size.
454 static inline void __free_one_page(struct page *page,
455 struct zone *zone, unsigned int order)
457 unsigned long page_idx;
458 int order_size = 1 << order;
459 int migratetype = get_pageblock_migratetype(page);
461 if (unlikely(PageCompound(page)))
462 if (unlikely(destroy_compound_page(page, order)))
465 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
467 VM_BUG_ON(page_idx & (order_size - 1));
468 VM_BUG_ON(bad_range(zone, page));
470 __mod_zone_page_state(zone, NR_FREE_PAGES, order_size);
471 while (order < MAX_ORDER-1) {
472 unsigned long combined_idx;
475 buddy = __page_find_buddy(page, page_idx, order);
476 if (!page_is_buddy(page, buddy, order))
479 /* Our buddy is free, merge with it and move up one order. */
480 list_del(&buddy->lru);
481 zone->free_area[order].nr_free--;
482 rmv_page_order(buddy);
483 combined_idx = __find_combined_index(page_idx, order);
484 page = page + (combined_idx - page_idx);
485 page_idx = combined_idx;
488 set_page_order(page, order);
490 &zone->free_area[order].free_list[migratetype]);
491 zone->free_area[order].nr_free++;
494 static inline int free_pages_check(struct page *page)
496 free_page_mlock(page);
497 if (unlikely(page_mapcount(page) |
498 (page->mapping != NULL) |
499 (page_count(page) != 0) |
500 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) {
504 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
505 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
510 * Frees a list of pages.
511 * Assumes all pages on list are in same zone, and of same order.
512 * count is the number of pages to free.
514 * If the zone was previously in an "all pages pinned" state then look to
515 * see if this freeing clears that state.
517 * And clear the zone's pages_scanned counter, to hold off the "all pages are
518 * pinned" detection logic.
520 static void free_pages_bulk(struct zone *zone, int count,
521 struct list_head *list, int order)
523 spin_lock(&zone->lock);
524 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
525 zone->pages_scanned = 0;
529 VM_BUG_ON(list_empty(list));
530 page = list_entry(list->prev, struct page, lru);
531 /* have to delete it as __free_one_page list manipulates */
532 list_del(&page->lru);
533 __free_one_page(page, zone, order);
535 spin_unlock(&zone->lock);
538 static void free_one_page(struct zone *zone, struct page *page, int order)
540 spin_lock(&zone->lock);
541 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
542 zone->pages_scanned = 0;
543 __free_one_page(page, zone, order);
544 spin_unlock(&zone->lock);
547 static void __free_pages_ok(struct page *page, unsigned int order)
553 for (i = 0 ; i < (1 << order) ; ++i)
554 bad += free_pages_check(page + i);
558 if (!PageHighMem(page)) {
559 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
560 debug_check_no_obj_freed(page_address(page),
563 arch_free_page(page, order);
564 kernel_map_pages(page, 1 << order, 0);
566 local_irq_save(flags);
567 __count_vm_events(PGFREE, 1 << order);
568 free_one_page(page_zone(page), page, order);
569 local_irq_restore(flags);
573 * permit the bootmem allocator to evade page validation on high-order frees
575 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
578 __ClearPageReserved(page);
579 set_page_count(page, 0);
580 set_page_refcounted(page);
586 for (loop = 0; loop < BITS_PER_LONG; loop++) {
587 struct page *p = &page[loop];
589 if (loop + 1 < BITS_PER_LONG)
591 __ClearPageReserved(p);
592 set_page_count(p, 0);
595 set_page_refcounted(page);
596 __free_pages(page, order);
602 * The order of subdivision here is critical for the IO subsystem.
603 * Please do not alter this order without good reasons and regression
604 * testing. Specifically, as large blocks of memory are subdivided,
605 * the order in which smaller blocks are delivered depends on the order
606 * they're subdivided in this function. This is the primary factor
607 * influencing the order in which pages are delivered to the IO
608 * subsystem according to empirical testing, and this is also justified
609 * by considering the behavior of a buddy system containing a single
610 * large block of memory acted on by a series of small allocations.
611 * This behavior is a critical factor in sglist merging's success.
615 static inline void expand(struct zone *zone, struct page *page,
616 int low, int high, struct free_area *area,
619 unsigned long size = 1 << high;
625 VM_BUG_ON(bad_range(zone, &page[size]));
626 list_add(&page[size].lru, &area->free_list[migratetype]);
628 set_page_order(&page[size], high);
633 * This page is about to be returned from the page allocator
635 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
637 if (unlikely(page_mapcount(page) |
638 (page->mapping != NULL) |
639 (page_count(page) != 0) |
640 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) {
645 set_page_private(page, 0);
646 set_page_refcounted(page);
648 arch_alloc_page(page, order);
649 kernel_map_pages(page, 1 << order, 1);
651 if (gfp_flags & __GFP_ZERO)
652 prep_zero_page(page, order, gfp_flags);
654 if (order && (gfp_flags & __GFP_COMP))
655 prep_compound_page(page, order);
661 * Go through the free lists for the given migratetype and remove
662 * the smallest available page from the freelists
664 static struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
667 unsigned int current_order;
668 struct free_area * area;
671 /* Find a page of the appropriate size in the preferred list */
672 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
673 area = &(zone->free_area[current_order]);
674 if (list_empty(&area->free_list[migratetype]))
677 page = list_entry(area->free_list[migratetype].next,
679 list_del(&page->lru);
680 rmv_page_order(page);
682 __mod_zone_page_state(zone, NR_FREE_PAGES, - (1UL << order));
683 expand(zone, page, order, current_order, area, migratetype);
692 * This array describes the order lists are fallen back to when
693 * the free lists for the desirable migrate type are depleted
695 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
696 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
697 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
698 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
699 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
703 * Move the free pages in a range to the free lists of the requested type.
704 * Note that start_page and end_pages are not aligned on a pageblock
705 * boundary. If alignment is required, use move_freepages_block()
707 static int move_freepages(struct zone *zone,
708 struct page *start_page, struct page *end_page,
715 #ifndef CONFIG_HOLES_IN_ZONE
717 * page_zone is not safe to call in this context when
718 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
719 * anyway as we check zone boundaries in move_freepages_block().
720 * Remove at a later date when no bug reports exist related to
721 * grouping pages by mobility
723 BUG_ON(page_zone(start_page) != page_zone(end_page));
726 for (page = start_page; page <= end_page;) {
727 /* Make sure we are not inadvertently changing nodes */
728 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
730 if (!pfn_valid_within(page_to_pfn(page))) {
735 if (!PageBuddy(page)) {
740 order = page_order(page);
741 list_del(&page->lru);
743 &zone->free_area[order].free_list[migratetype]);
745 pages_moved += 1 << order;
751 static int move_freepages_block(struct zone *zone, struct page *page,
754 unsigned long start_pfn, end_pfn;
755 struct page *start_page, *end_page;
757 start_pfn = page_to_pfn(page);
758 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
759 start_page = pfn_to_page(start_pfn);
760 end_page = start_page + pageblock_nr_pages - 1;
761 end_pfn = start_pfn + pageblock_nr_pages - 1;
763 /* Do not cross zone boundaries */
764 if (start_pfn < zone->zone_start_pfn)
766 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
769 return move_freepages(zone, start_page, end_page, migratetype);
772 /* Remove an element from the buddy allocator from the fallback list */
773 static struct page *__rmqueue_fallback(struct zone *zone, int order,
774 int start_migratetype)
776 struct free_area * area;
781 /* Find the largest possible block of pages in the other list */
782 for (current_order = MAX_ORDER-1; current_order >= order;
784 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
785 migratetype = fallbacks[start_migratetype][i];
787 /* MIGRATE_RESERVE handled later if necessary */
788 if (migratetype == MIGRATE_RESERVE)
791 area = &(zone->free_area[current_order]);
792 if (list_empty(&area->free_list[migratetype]))
795 page = list_entry(area->free_list[migratetype].next,
800 * If breaking a large block of pages, move all free
801 * pages to the preferred allocation list. If falling
802 * back for a reclaimable kernel allocation, be more
803 * agressive about taking ownership of free pages
805 if (unlikely(current_order >= (pageblock_order >> 1)) ||
806 start_migratetype == MIGRATE_RECLAIMABLE) {
808 pages = move_freepages_block(zone, page,
811 /* Claim the whole block if over half of it is free */
812 if (pages >= (1 << (pageblock_order-1)))
813 set_pageblock_migratetype(page,
816 migratetype = start_migratetype;
819 /* Remove the page from the freelists */
820 list_del(&page->lru);
821 rmv_page_order(page);
822 __mod_zone_page_state(zone, NR_FREE_PAGES,
825 if (current_order == pageblock_order)
826 set_pageblock_migratetype(page,
829 expand(zone, page, order, current_order, area, migratetype);
834 /* Use MIGRATE_RESERVE rather than fail an allocation */
835 return __rmqueue_smallest(zone, order, MIGRATE_RESERVE);
839 * Do the hard work of removing an element from the buddy allocator.
840 * Call me with the zone->lock already held.
842 static struct page *__rmqueue(struct zone *zone, unsigned int order,
847 page = __rmqueue_smallest(zone, order, migratetype);
850 page = __rmqueue_fallback(zone, order, migratetype);
856 * Obtain a specified number of elements from the buddy allocator, all under
857 * a single hold of the lock, for efficiency. Add them to the supplied list.
858 * Returns the number of new pages which were placed at *list.
860 static int rmqueue_bulk(struct zone *zone, unsigned int order,
861 unsigned long count, struct list_head *list,
866 spin_lock(&zone->lock);
867 for (i = 0; i < count; ++i) {
868 struct page *page = __rmqueue(zone, order, migratetype);
869 if (unlikely(page == NULL))
873 * Split buddy pages returned by expand() are received here
874 * in physical page order. The page is added to the callers and
875 * list and the list head then moves forward. From the callers
876 * perspective, the linked list is ordered by page number in
877 * some conditions. This is useful for IO devices that can
878 * merge IO requests if the physical pages are ordered
881 list_add(&page->lru, list);
882 set_page_private(page, migratetype);
885 spin_unlock(&zone->lock);
891 * Called from the vmstat counter updater to drain pagesets of this
892 * currently executing processor on remote nodes after they have
895 * Note that this function must be called with the thread pinned to
896 * a single processor.
898 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
903 local_irq_save(flags);
904 if (pcp->count >= pcp->batch)
905 to_drain = pcp->batch;
907 to_drain = pcp->count;
908 free_pages_bulk(zone, to_drain, &pcp->list, 0);
909 pcp->count -= to_drain;
910 local_irq_restore(flags);
915 * Drain pages of the indicated processor.
917 * The processor must either be the current processor and the
918 * thread pinned to the current processor or a processor that
921 static void drain_pages(unsigned int cpu)
926 for_each_populated_zone(zone) {
927 struct per_cpu_pageset *pset;
928 struct per_cpu_pages *pcp;
930 pset = zone_pcp(zone, cpu);
933 local_irq_save(flags);
934 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
936 local_irq_restore(flags);
941 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
943 void drain_local_pages(void *arg)
945 drain_pages(smp_processor_id());
949 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
951 void drain_all_pages(void)
953 on_each_cpu(drain_local_pages, NULL, 1);
956 #ifdef CONFIG_HIBERNATION
958 void mark_free_pages(struct zone *zone)
960 unsigned long pfn, max_zone_pfn;
963 struct list_head *curr;
965 if (!zone->spanned_pages)
968 spin_lock_irqsave(&zone->lock, flags);
970 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
971 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
972 if (pfn_valid(pfn)) {
973 struct page *page = pfn_to_page(pfn);
975 if (!swsusp_page_is_forbidden(page))
976 swsusp_unset_page_free(page);
979 for_each_migratetype_order(order, t) {
980 list_for_each(curr, &zone->free_area[order].free_list[t]) {
983 pfn = page_to_pfn(list_entry(curr, struct page, lru));
984 for (i = 0; i < (1UL << order); i++)
985 swsusp_set_page_free(pfn_to_page(pfn + i));
988 spin_unlock_irqrestore(&zone->lock, flags);
990 #endif /* CONFIG_PM */
993 * Free a 0-order page
995 static void free_hot_cold_page(struct page *page, int cold)
997 struct zone *zone = page_zone(page);
998 struct per_cpu_pages *pcp;
1002 page->mapping = NULL;
1003 if (free_pages_check(page))
1006 if (!PageHighMem(page)) {
1007 debug_check_no_locks_freed(page_address(page), PAGE_SIZE);
1008 debug_check_no_obj_freed(page_address(page), PAGE_SIZE);
1010 arch_free_page(page, 0);
1011 kernel_map_pages(page, 1, 0);
1013 pcp = &zone_pcp(zone, get_cpu())->pcp;
1014 local_irq_save(flags);
1015 __count_vm_event(PGFREE);
1017 list_add_tail(&page->lru, &pcp->list);
1019 list_add(&page->lru, &pcp->list);
1020 set_page_private(page, get_pageblock_migratetype(page));
1022 if (pcp->count >= pcp->high) {
1023 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
1024 pcp->count -= pcp->batch;
1026 local_irq_restore(flags);
1030 void free_hot_page(struct page *page)
1032 free_hot_cold_page(page, 0);
1035 void free_cold_page(struct page *page)
1037 free_hot_cold_page(page, 1);
1041 * split_page takes a non-compound higher-order page, and splits it into
1042 * n (1<<order) sub-pages: page[0..n]
1043 * Each sub-page must be freed individually.
1045 * Note: this is probably too low level an operation for use in drivers.
1046 * Please consult with lkml before using this in your driver.
1048 void split_page(struct page *page, unsigned int order)
1052 VM_BUG_ON(PageCompound(page));
1053 VM_BUG_ON(!page_count(page));
1054 for (i = 1; i < (1 << order); i++)
1055 set_page_refcounted(page + i);
1059 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1060 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1063 static struct page *buffered_rmqueue(struct zone *preferred_zone,
1064 struct zone *zone, int order, gfp_t gfp_flags)
1066 unsigned long flags;
1068 int cold = !!(gfp_flags & __GFP_COLD);
1070 int migratetype = allocflags_to_migratetype(gfp_flags);
1074 if (likely(order == 0)) {
1075 struct per_cpu_pages *pcp;
1077 pcp = &zone_pcp(zone, cpu)->pcp;
1078 local_irq_save(flags);
1080 pcp->count = rmqueue_bulk(zone, 0,
1081 pcp->batch, &pcp->list, migratetype);
1082 if (unlikely(!pcp->count))
1086 /* Find a page of the appropriate migrate type */
1088 list_for_each_entry_reverse(page, &pcp->list, lru)
1089 if (page_private(page) == migratetype)
1092 list_for_each_entry(page, &pcp->list, lru)
1093 if (page_private(page) == migratetype)
1097 /* Allocate more to the pcp list if necessary */
1098 if (unlikely(&page->lru == &pcp->list)) {
1099 pcp->count += rmqueue_bulk(zone, 0,
1100 pcp->batch, &pcp->list, migratetype);
1101 page = list_entry(pcp->list.next, struct page, lru);
1104 list_del(&page->lru);
1107 spin_lock_irqsave(&zone->lock, flags);
1108 page = __rmqueue(zone, order, migratetype);
1109 spin_unlock(&zone->lock);
1114 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1115 zone_statistics(preferred_zone, zone);
1116 local_irq_restore(flags);
1119 VM_BUG_ON(bad_range(zone, page));
1120 if (prep_new_page(page, order, gfp_flags))
1125 local_irq_restore(flags);
1130 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
1131 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
1132 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
1133 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
1134 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1135 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1136 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1138 #ifdef CONFIG_FAIL_PAGE_ALLOC
1140 static struct fail_page_alloc_attr {
1141 struct fault_attr attr;
1143 u32 ignore_gfp_highmem;
1144 u32 ignore_gfp_wait;
1147 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1149 struct dentry *ignore_gfp_highmem_file;
1150 struct dentry *ignore_gfp_wait_file;
1151 struct dentry *min_order_file;
1153 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1155 } fail_page_alloc = {
1156 .attr = FAULT_ATTR_INITIALIZER,
1157 .ignore_gfp_wait = 1,
1158 .ignore_gfp_highmem = 1,
1162 static int __init setup_fail_page_alloc(char *str)
1164 return setup_fault_attr(&fail_page_alloc.attr, str);
1166 __setup("fail_page_alloc=", setup_fail_page_alloc);
1168 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1170 if (order < fail_page_alloc.min_order)
1172 if (gfp_mask & __GFP_NOFAIL)
1174 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1176 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1179 return should_fail(&fail_page_alloc.attr, 1 << order);
1182 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1184 static int __init fail_page_alloc_debugfs(void)
1186 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1190 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1194 dir = fail_page_alloc.attr.dentries.dir;
1196 fail_page_alloc.ignore_gfp_wait_file =
1197 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1198 &fail_page_alloc.ignore_gfp_wait);
1200 fail_page_alloc.ignore_gfp_highmem_file =
1201 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1202 &fail_page_alloc.ignore_gfp_highmem);
1203 fail_page_alloc.min_order_file =
1204 debugfs_create_u32("min-order", mode, dir,
1205 &fail_page_alloc.min_order);
1207 if (!fail_page_alloc.ignore_gfp_wait_file ||
1208 !fail_page_alloc.ignore_gfp_highmem_file ||
1209 !fail_page_alloc.min_order_file) {
1211 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1212 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1213 debugfs_remove(fail_page_alloc.min_order_file);
1214 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1220 late_initcall(fail_page_alloc_debugfs);
1222 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1224 #else /* CONFIG_FAIL_PAGE_ALLOC */
1226 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1231 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1234 * Return 1 if free pages are above 'mark'. This takes into account the order
1235 * of the allocation.
1237 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1238 int classzone_idx, int alloc_flags)
1240 /* free_pages my go negative - that's OK */
1242 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
1245 if (alloc_flags & ALLOC_HIGH)
1247 if (alloc_flags & ALLOC_HARDER)
1250 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1252 for (o = 0; o < order; o++) {
1253 /* At the next order, this order's pages become unavailable */
1254 free_pages -= z->free_area[o].nr_free << o;
1256 /* Require fewer higher order pages to be free */
1259 if (free_pages <= min)
1267 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1268 * skip over zones that are not allowed by the cpuset, or that have
1269 * been recently (in last second) found to be nearly full. See further
1270 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1271 * that have to skip over a lot of full or unallowed zones.
1273 * If the zonelist cache is present in the passed in zonelist, then
1274 * returns a pointer to the allowed node mask (either the current
1275 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1277 * If the zonelist cache is not available for this zonelist, does
1278 * nothing and returns NULL.
1280 * If the fullzones BITMAP in the zonelist cache is stale (more than
1281 * a second since last zap'd) then we zap it out (clear its bits.)
1283 * We hold off even calling zlc_setup, until after we've checked the
1284 * first zone in the zonelist, on the theory that most allocations will
1285 * be satisfied from that first zone, so best to examine that zone as
1286 * quickly as we can.
1288 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1290 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1291 nodemask_t *allowednodes; /* zonelist_cache approximation */
1293 zlc = zonelist->zlcache_ptr;
1297 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1298 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1299 zlc->last_full_zap = jiffies;
1302 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1303 &cpuset_current_mems_allowed :
1304 &node_states[N_HIGH_MEMORY];
1305 return allowednodes;
1309 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1310 * if it is worth looking at further for free memory:
1311 * 1) Check that the zone isn't thought to be full (doesn't have its
1312 * bit set in the zonelist_cache fullzones BITMAP).
1313 * 2) Check that the zones node (obtained from the zonelist_cache
1314 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1315 * Return true (non-zero) if zone is worth looking at further, or
1316 * else return false (zero) if it is not.
1318 * This check -ignores- the distinction between various watermarks,
1319 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1320 * found to be full for any variation of these watermarks, it will
1321 * be considered full for up to one second by all requests, unless
1322 * we are so low on memory on all allowed nodes that we are forced
1323 * into the second scan of the zonelist.
1325 * In the second scan we ignore this zonelist cache and exactly
1326 * apply the watermarks to all zones, even it is slower to do so.
1327 * We are low on memory in the second scan, and should leave no stone
1328 * unturned looking for a free page.
1330 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1331 nodemask_t *allowednodes)
1333 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1334 int i; /* index of *z in zonelist zones */
1335 int n; /* node that zone *z is on */
1337 zlc = zonelist->zlcache_ptr;
1341 i = z - zonelist->_zonerefs;
1344 /* This zone is worth trying if it is allowed but not full */
1345 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1349 * Given 'z' scanning a zonelist, set the corresponding bit in
1350 * zlc->fullzones, so that subsequent attempts to allocate a page
1351 * from that zone don't waste time re-examining it.
1353 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1355 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1356 int i; /* index of *z in zonelist zones */
1358 zlc = zonelist->zlcache_ptr;
1362 i = z - zonelist->_zonerefs;
1364 set_bit(i, zlc->fullzones);
1367 #else /* CONFIG_NUMA */
1369 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1374 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1375 nodemask_t *allowednodes)
1380 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1383 #endif /* CONFIG_NUMA */
1386 * get_page_from_freelist goes through the zonelist trying to allocate
1389 static struct page *
1390 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1391 struct zonelist *zonelist, int high_zoneidx, int alloc_flags)
1394 struct page *page = NULL;
1396 struct zone *zone, *preferred_zone;
1397 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1398 int zlc_active = 0; /* set if using zonelist_cache */
1399 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1401 (void)first_zones_zonelist(zonelist, high_zoneidx, nodemask,
1403 if (!preferred_zone)
1406 classzone_idx = zone_idx(preferred_zone);
1408 if (WARN_ON_ONCE(order >= MAX_ORDER))
1413 * Scan zonelist, looking for a zone with enough free.
1414 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1416 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1417 high_zoneidx, nodemask) {
1418 if (NUMA_BUILD && zlc_active &&
1419 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1421 if ((alloc_flags & ALLOC_CPUSET) &&
1422 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1425 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1427 if (alloc_flags & ALLOC_WMARK_MIN)
1428 mark = zone->pages_min;
1429 else if (alloc_flags & ALLOC_WMARK_LOW)
1430 mark = zone->pages_low;
1432 mark = zone->pages_high;
1433 if (!zone_watermark_ok(zone, order, mark,
1434 classzone_idx, alloc_flags)) {
1435 if (!zone_reclaim_mode ||
1436 !zone_reclaim(zone, gfp_mask, order))
1437 goto this_zone_full;
1441 page = buffered_rmqueue(preferred_zone, zone, order, gfp_mask);
1446 zlc_mark_zone_full(zonelist, z);
1448 if (NUMA_BUILD && !did_zlc_setup) {
1449 /* we do zlc_setup after the first zone is tried */
1450 allowednodes = zlc_setup(zonelist, alloc_flags);
1456 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1457 /* Disable zlc cache for second zonelist scan */
1465 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1466 unsigned long pages_reclaimed)
1468 /* Do not loop if specifically requested */
1469 if (gfp_mask & __GFP_NORETRY)
1473 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1474 * means __GFP_NOFAIL, but that may not be true in other
1477 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1481 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1482 * specified, then we retry until we no longer reclaim any pages
1483 * (above), or we've reclaimed an order of pages at least as
1484 * large as the allocation's order. In both cases, if the
1485 * allocation still fails, we stop retrying.
1487 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1491 * Don't let big-order allocations loop unless the caller
1492 * explicitly requests that.
1494 if (gfp_mask & __GFP_NOFAIL)
1500 static inline struct page *
1501 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1502 struct zonelist *zonelist, enum zone_type high_zoneidx,
1503 nodemask_t *nodemask)
1507 /* Acquire the OOM killer lock for the zones in zonelist */
1508 if (!try_set_zone_oom(zonelist, gfp_mask)) {
1509 schedule_timeout_uninterruptible(1);
1514 * Go through the zonelist yet one more time, keep very high watermark
1515 * here, this is only to catch a parallel oom killing, we must fail if
1516 * we're still under heavy pressure.
1518 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1519 order, zonelist, high_zoneidx,
1520 ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1524 /* The OOM killer will not help higher order allocs */
1525 if (order > PAGE_ALLOC_COSTLY_ORDER)
1528 /* Exhausted what can be done so it's blamo time */
1529 out_of_memory(zonelist, gfp_mask, order);
1532 clear_zonelist_oom(zonelist, gfp_mask);
1536 /* The really slow allocator path where we enter direct reclaim */
1537 static inline struct page *
1538 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1539 struct zonelist *zonelist, enum zone_type high_zoneidx,
1540 nodemask_t *nodemask, int alloc_flags, unsigned long *did_some_progress)
1542 struct page *page = NULL;
1543 struct reclaim_state reclaim_state;
1544 struct task_struct *p = current;
1548 /* We now go into synchronous reclaim */
1549 cpuset_memory_pressure_bump();
1552 * The task's cpuset might have expanded its set of allowable nodes
1554 p->flags |= PF_MEMALLOC;
1555 lockdep_set_current_reclaim_state(gfp_mask);
1556 reclaim_state.reclaimed_slab = 0;
1557 p->reclaim_state = &reclaim_state;
1559 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1561 p->reclaim_state = NULL;
1562 lockdep_clear_current_reclaim_state();
1563 p->flags &= ~PF_MEMALLOC;
1570 if (likely(*did_some_progress))
1571 page = get_page_from_freelist(gfp_mask, nodemask, order,
1572 zonelist, high_zoneidx, alloc_flags);
1577 is_allocation_high_priority(struct task_struct *p, gfp_t gfp_mask)
1579 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
1586 * This is called in the allocator slow-path if the allocation request is of
1587 * sufficient urgency to ignore watermarks and take other desperate measures
1589 static inline struct page *
1590 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1591 struct zonelist *zonelist, enum zone_type high_zoneidx,
1592 nodemask_t *nodemask)
1597 page = get_page_from_freelist(gfp_mask, nodemask, order,
1598 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS);
1600 if (!page && gfp_mask & __GFP_NOFAIL)
1601 congestion_wait(WRITE, HZ/50);
1602 } while (!page && (gfp_mask & __GFP_NOFAIL));
1608 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
1609 enum zone_type high_zoneidx)
1614 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1615 wakeup_kswapd(zone, order);
1618 static inline struct page *
1619 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
1620 struct zonelist *zonelist, enum zone_type high_zoneidx,
1621 nodemask_t *nodemask)
1623 const gfp_t wait = gfp_mask & __GFP_WAIT;
1624 struct page *page = NULL;
1626 unsigned long pages_reclaimed = 0;
1627 unsigned long did_some_progress;
1628 struct task_struct *p = current;
1631 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1632 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1633 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1634 * using a larger set of nodes after it has established that the
1635 * allowed per node queues are empty and that nodes are
1638 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1641 wake_all_kswapd(order, zonelist, high_zoneidx);
1644 * OK, we're below the kswapd watermark and have kicked background
1645 * reclaim. Now things get more complex, so set up alloc_flags according
1646 * to how we want to proceed.
1648 * The caller may dip into page reserves a bit more if the caller
1649 * cannot run direct reclaim, or if the caller has realtime scheduling
1650 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1651 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1653 alloc_flags = ALLOC_WMARK_MIN;
1654 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
1655 alloc_flags |= ALLOC_HARDER;
1656 if (gfp_mask & __GFP_HIGH)
1657 alloc_flags |= ALLOC_HIGH;
1659 alloc_flags |= ALLOC_CPUSET;
1663 * Go through the zonelist again. Let __GFP_HIGH and allocations
1664 * coming from realtime tasks go deeper into reserves.
1666 * This is the last chance, in general, before the goto nopage.
1667 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1668 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1670 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
1671 high_zoneidx, alloc_flags);
1676 /* Allocate without watermarks if the context allows */
1677 if (is_allocation_high_priority(p, gfp_mask)) {
1678 /* Do not dip into emergency reserves if specified */
1679 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
1680 page = __alloc_pages_high_priority(gfp_mask, order,
1681 zonelist, high_zoneidx, nodemask);
1686 /* Ensure no recursion into the allocator */
1690 /* Atomic allocations - we can't balance anything */
1694 /* Try direct reclaim and then allocating */
1695 page = __alloc_pages_direct_reclaim(gfp_mask, order,
1696 zonelist, high_zoneidx,
1698 alloc_flags, &did_some_progress);
1703 * If we failed to make any progress reclaiming, then we are
1704 * running out of options and have to consider going OOM
1706 if (!did_some_progress) {
1707 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1708 page = __alloc_pages_may_oom(gfp_mask, order,
1709 zonelist, high_zoneidx,
1715 * The OOM killer does not trigger for high-order allocations
1716 * but if no progress is being made, there are no other
1717 * options and retrying is unlikely to help
1719 if (order > PAGE_ALLOC_COSTLY_ORDER)
1726 /* Check if we should retry the allocation */
1727 pages_reclaimed += did_some_progress;
1728 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
1729 /* Wait for some write requests to complete then retry */
1730 congestion_wait(WRITE, HZ/50);
1735 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1736 printk(KERN_WARNING "%s: page allocation failure."
1737 " order:%d, mode:0x%x\n",
1738 p->comm, order, gfp_mask);
1748 * This is the 'heart' of the zoned buddy allocator.
1751 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
1752 struct zonelist *zonelist, nodemask_t *nodemask)
1754 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1757 lockdep_trace_alloc(gfp_mask);
1759 might_sleep_if(gfp_mask & __GFP_WAIT);
1761 if (should_fail_alloc_page(gfp_mask, order))
1765 * Check the zones suitable for the gfp_mask contain at least one
1766 * valid zone. It's possible to have an empty zonelist as a result
1767 * of GFP_THISNODE and a memoryless node
1769 if (unlikely(!zonelist->_zonerefs->zone))
1772 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
1773 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET);
1774 if (unlikely(!page))
1775 page = __alloc_pages_slowpath(gfp_mask, order,
1776 zonelist, high_zoneidx, nodemask);
1780 EXPORT_SYMBOL(__alloc_pages_nodemask);
1783 * Common helper functions.
1785 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1788 page = alloc_pages(gfp_mask, order);
1791 return (unsigned long) page_address(page);
1794 EXPORT_SYMBOL(__get_free_pages);
1796 unsigned long get_zeroed_page(gfp_t gfp_mask)
1801 * get_zeroed_page() returns a 32-bit address, which cannot represent
1804 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1806 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1808 return (unsigned long) page_address(page);
1812 EXPORT_SYMBOL(get_zeroed_page);
1814 void __pagevec_free(struct pagevec *pvec)
1816 int i = pagevec_count(pvec);
1819 free_hot_cold_page(pvec->pages[i], pvec->cold);
1822 void __free_pages(struct page *page, unsigned int order)
1824 if (put_page_testzero(page)) {
1826 free_hot_page(page);
1828 __free_pages_ok(page, order);
1832 EXPORT_SYMBOL(__free_pages);
1834 void free_pages(unsigned long addr, unsigned int order)
1837 VM_BUG_ON(!virt_addr_valid((void *)addr));
1838 __free_pages(virt_to_page((void *)addr), order);
1842 EXPORT_SYMBOL(free_pages);
1845 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
1846 * @size: the number of bytes to allocate
1847 * @gfp_mask: GFP flags for the allocation
1849 * This function is similar to alloc_pages(), except that it allocates the
1850 * minimum number of pages to satisfy the request. alloc_pages() can only
1851 * allocate memory in power-of-two pages.
1853 * This function is also limited by MAX_ORDER.
1855 * Memory allocated by this function must be released by free_pages_exact().
1857 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
1859 unsigned int order = get_order(size);
1862 addr = __get_free_pages(gfp_mask, order);
1864 unsigned long alloc_end = addr + (PAGE_SIZE << order);
1865 unsigned long used = addr + PAGE_ALIGN(size);
1867 split_page(virt_to_page(addr), order);
1868 while (used < alloc_end) {
1874 return (void *)addr;
1876 EXPORT_SYMBOL(alloc_pages_exact);
1879 * free_pages_exact - release memory allocated via alloc_pages_exact()
1880 * @virt: the value returned by alloc_pages_exact.
1881 * @size: size of allocation, same value as passed to alloc_pages_exact().
1883 * Release the memory allocated by a previous call to alloc_pages_exact.
1885 void free_pages_exact(void *virt, size_t size)
1887 unsigned long addr = (unsigned long)virt;
1888 unsigned long end = addr + PAGE_ALIGN(size);
1890 while (addr < end) {
1895 EXPORT_SYMBOL(free_pages_exact);
1897 static unsigned int nr_free_zone_pages(int offset)
1902 /* Just pick one node, since fallback list is circular */
1903 unsigned int sum = 0;
1905 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
1907 for_each_zone_zonelist(zone, z, zonelist, offset) {
1908 unsigned long size = zone->present_pages;
1909 unsigned long high = zone->pages_high;
1918 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1920 unsigned int nr_free_buffer_pages(void)
1922 return nr_free_zone_pages(gfp_zone(GFP_USER));
1924 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
1927 * Amount of free RAM allocatable within all zones
1929 unsigned int nr_free_pagecache_pages(void)
1931 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
1934 static inline void show_node(struct zone *zone)
1937 printk("Node %d ", zone_to_nid(zone));
1940 void si_meminfo(struct sysinfo *val)
1942 val->totalram = totalram_pages;
1944 val->freeram = global_page_state(NR_FREE_PAGES);
1945 val->bufferram = nr_blockdev_pages();
1946 val->totalhigh = totalhigh_pages;
1947 val->freehigh = nr_free_highpages();
1948 val->mem_unit = PAGE_SIZE;
1951 EXPORT_SYMBOL(si_meminfo);
1954 void si_meminfo_node(struct sysinfo *val, int nid)
1956 pg_data_t *pgdat = NODE_DATA(nid);
1958 val->totalram = pgdat->node_present_pages;
1959 val->freeram = node_page_state(nid, NR_FREE_PAGES);
1960 #ifdef CONFIG_HIGHMEM
1961 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1962 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
1968 val->mem_unit = PAGE_SIZE;
1972 #define K(x) ((x) << (PAGE_SHIFT-10))
1975 * Show free area list (used inside shift_scroll-lock stuff)
1976 * We also calculate the percentage fragmentation. We do this by counting the
1977 * memory on each free list with the exception of the first item on the list.
1979 void show_free_areas(void)
1984 for_each_populated_zone(zone) {
1986 printk("%s per-cpu:\n", zone->name);
1988 for_each_online_cpu(cpu) {
1989 struct per_cpu_pageset *pageset;
1991 pageset = zone_pcp(zone, cpu);
1993 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
1994 cpu, pageset->pcp.high,
1995 pageset->pcp.batch, pageset->pcp.count);
1999 printk("Active_anon:%lu active_file:%lu inactive_anon:%lu\n"
2000 " inactive_file:%lu"
2001 //TODO: check/adjust line lengths
2002 #ifdef CONFIG_UNEVICTABLE_LRU
2005 " dirty:%lu writeback:%lu unstable:%lu\n"
2006 " free:%lu slab:%lu mapped:%lu pagetables:%lu bounce:%lu\n",
2007 global_page_state(NR_ACTIVE_ANON),
2008 global_page_state(NR_ACTIVE_FILE),
2009 global_page_state(NR_INACTIVE_ANON),
2010 global_page_state(NR_INACTIVE_FILE),
2011 #ifdef CONFIG_UNEVICTABLE_LRU
2012 global_page_state(NR_UNEVICTABLE),
2014 global_page_state(NR_FILE_DIRTY),
2015 global_page_state(NR_WRITEBACK),
2016 global_page_state(NR_UNSTABLE_NFS),
2017 global_page_state(NR_FREE_PAGES),
2018 global_page_state(NR_SLAB_RECLAIMABLE) +
2019 global_page_state(NR_SLAB_UNRECLAIMABLE),
2020 global_page_state(NR_FILE_MAPPED),
2021 global_page_state(NR_PAGETABLE),
2022 global_page_state(NR_BOUNCE));
2024 for_each_populated_zone(zone) {
2033 " active_anon:%lukB"
2034 " inactive_anon:%lukB"
2035 " active_file:%lukB"
2036 " inactive_file:%lukB"
2037 #ifdef CONFIG_UNEVICTABLE_LRU
2038 " unevictable:%lukB"
2041 " pages_scanned:%lu"
2042 " all_unreclaimable? %s"
2045 K(zone_page_state(zone, NR_FREE_PAGES)),
2048 K(zone->pages_high),
2049 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2050 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2051 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2052 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2053 #ifdef CONFIG_UNEVICTABLE_LRU
2054 K(zone_page_state(zone, NR_UNEVICTABLE)),
2056 K(zone->present_pages),
2057 zone->pages_scanned,
2058 (zone_is_all_unreclaimable(zone) ? "yes" : "no")
2060 printk("lowmem_reserve[]:");
2061 for (i = 0; i < MAX_NR_ZONES; i++)
2062 printk(" %lu", zone->lowmem_reserve[i]);
2066 for_each_populated_zone(zone) {
2067 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2070 printk("%s: ", zone->name);
2072 spin_lock_irqsave(&zone->lock, flags);
2073 for (order = 0; order < MAX_ORDER; order++) {
2074 nr[order] = zone->free_area[order].nr_free;
2075 total += nr[order] << order;
2077 spin_unlock_irqrestore(&zone->lock, flags);
2078 for (order = 0; order < MAX_ORDER; order++)
2079 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2080 printk("= %lukB\n", K(total));
2083 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2085 show_swap_cache_info();
2088 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2090 zoneref->zone = zone;
2091 zoneref->zone_idx = zone_idx(zone);
2095 * Builds allocation fallback zone lists.
2097 * Add all populated zones of a node to the zonelist.
2099 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2100 int nr_zones, enum zone_type zone_type)
2104 BUG_ON(zone_type >= MAX_NR_ZONES);
2109 zone = pgdat->node_zones + zone_type;
2110 if (populated_zone(zone)) {
2111 zoneref_set_zone(zone,
2112 &zonelist->_zonerefs[nr_zones++]);
2113 check_highest_zone(zone_type);
2116 } while (zone_type);
2123 * 0 = automatic detection of better ordering.
2124 * 1 = order by ([node] distance, -zonetype)
2125 * 2 = order by (-zonetype, [node] distance)
2127 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2128 * the same zonelist. So only NUMA can configure this param.
2130 #define ZONELIST_ORDER_DEFAULT 0
2131 #define ZONELIST_ORDER_NODE 1
2132 #define ZONELIST_ORDER_ZONE 2
2134 /* zonelist order in the kernel.
2135 * set_zonelist_order() will set this to NODE or ZONE.
2137 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2138 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2142 /* The value user specified ....changed by config */
2143 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2144 /* string for sysctl */
2145 #define NUMA_ZONELIST_ORDER_LEN 16
2146 char numa_zonelist_order[16] = "default";
2149 * interface for configure zonelist ordering.
2150 * command line option "numa_zonelist_order"
2151 * = "[dD]efault - default, automatic configuration.
2152 * = "[nN]ode - order by node locality, then by zone within node
2153 * = "[zZ]one - order by zone, then by locality within zone
2156 static int __parse_numa_zonelist_order(char *s)
2158 if (*s == 'd' || *s == 'D') {
2159 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2160 } else if (*s == 'n' || *s == 'N') {
2161 user_zonelist_order = ZONELIST_ORDER_NODE;
2162 } else if (*s == 'z' || *s == 'Z') {
2163 user_zonelist_order = ZONELIST_ORDER_ZONE;
2166 "Ignoring invalid numa_zonelist_order value: "
2173 static __init int setup_numa_zonelist_order(char *s)
2176 return __parse_numa_zonelist_order(s);
2179 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2182 * sysctl handler for numa_zonelist_order
2184 int numa_zonelist_order_handler(ctl_table *table, int write,
2185 struct file *file, void __user *buffer, size_t *length,
2188 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2192 strncpy(saved_string, (char*)table->data,
2193 NUMA_ZONELIST_ORDER_LEN);
2194 ret = proc_dostring(table, write, file, buffer, length, ppos);
2198 int oldval = user_zonelist_order;
2199 if (__parse_numa_zonelist_order((char*)table->data)) {
2201 * bogus value. restore saved string
2203 strncpy((char*)table->data, saved_string,
2204 NUMA_ZONELIST_ORDER_LEN);
2205 user_zonelist_order = oldval;
2206 } else if (oldval != user_zonelist_order)
2207 build_all_zonelists();
2213 #define MAX_NODE_LOAD (num_online_nodes())
2214 static int node_load[MAX_NUMNODES];
2217 * find_next_best_node - find the next node that should appear in a given node's fallback list
2218 * @node: node whose fallback list we're appending
2219 * @used_node_mask: nodemask_t of already used nodes
2221 * We use a number of factors to determine which is the next node that should
2222 * appear on a given node's fallback list. The node should not have appeared
2223 * already in @node's fallback list, and it should be the next closest node
2224 * according to the distance array (which contains arbitrary distance values
2225 * from each node to each node in the system), and should also prefer nodes
2226 * with no CPUs, since presumably they'll have very little allocation pressure
2227 * on them otherwise.
2228 * It returns -1 if no node is found.
2230 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2233 int min_val = INT_MAX;
2235 const struct cpumask *tmp = cpumask_of_node(0);
2237 /* Use the local node if we haven't already */
2238 if (!node_isset(node, *used_node_mask)) {
2239 node_set(node, *used_node_mask);
2243 for_each_node_state(n, N_HIGH_MEMORY) {
2245 /* Don't want a node to appear more than once */
2246 if (node_isset(n, *used_node_mask))
2249 /* Use the distance array to find the distance */
2250 val = node_distance(node, n);
2252 /* Penalize nodes under us ("prefer the next node") */
2255 /* Give preference to headless and unused nodes */
2256 tmp = cpumask_of_node(n);
2257 if (!cpumask_empty(tmp))
2258 val += PENALTY_FOR_NODE_WITH_CPUS;
2260 /* Slight preference for less loaded node */
2261 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2262 val += node_load[n];
2264 if (val < min_val) {
2271 node_set(best_node, *used_node_mask);
2278 * Build zonelists ordered by node and zones within node.
2279 * This results in maximum locality--normal zone overflows into local
2280 * DMA zone, if any--but risks exhausting DMA zone.
2282 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2285 struct zonelist *zonelist;
2287 zonelist = &pgdat->node_zonelists[0];
2288 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2290 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2292 zonelist->_zonerefs[j].zone = NULL;
2293 zonelist->_zonerefs[j].zone_idx = 0;
2297 * Build gfp_thisnode zonelists
2299 static void build_thisnode_zonelists(pg_data_t *pgdat)
2302 struct zonelist *zonelist;
2304 zonelist = &pgdat->node_zonelists[1];
2305 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2306 zonelist->_zonerefs[j].zone = NULL;
2307 zonelist->_zonerefs[j].zone_idx = 0;
2311 * Build zonelists ordered by zone and nodes within zones.
2312 * This results in conserving DMA zone[s] until all Normal memory is
2313 * exhausted, but results in overflowing to remote node while memory
2314 * may still exist in local DMA zone.
2316 static int node_order[MAX_NUMNODES];
2318 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2321 int zone_type; /* needs to be signed */
2323 struct zonelist *zonelist;
2325 zonelist = &pgdat->node_zonelists[0];
2327 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2328 for (j = 0; j < nr_nodes; j++) {
2329 node = node_order[j];
2330 z = &NODE_DATA(node)->node_zones[zone_type];
2331 if (populated_zone(z)) {
2333 &zonelist->_zonerefs[pos++]);
2334 check_highest_zone(zone_type);
2338 zonelist->_zonerefs[pos].zone = NULL;
2339 zonelist->_zonerefs[pos].zone_idx = 0;
2342 static int default_zonelist_order(void)
2345 unsigned long low_kmem_size,total_size;
2349 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2350 * If they are really small and used heavily, the system can fall
2351 * into OOM very easily.
2352 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2354 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2357 for_each_online_node(nid) {
2358 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2359 z = &NODE_DATA(nid)->node_zones[zone_type];
2360 if (populated_zone(z)) {
2361 if (zone_type < ZONE_NORMAL)
2362 low_kmem_size += z->present_pages;
2363 total_size += z->present_pages;
2367 if (!low_kmem_size || /* there are no DMA area. */
2368 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2369 return ZONELIST_ORDER_NODE;
2371 * look into each node's config.
2372 * If there is a node whose DMA/DMA32 memory is very big area on
2373 * local memory, NODE_ORDER may be suitable.
2375 average_size = total_size /
2376 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2377 for_each_online_node(nid) {
2380 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2381 z = &NODE_DATA(nid)->node_zones[zone_type];
2382 if (populated_zone(z)) {
2383 if (zone_type < ZONE_NORMAL)
2384 low_kmem_size += z->present_pages;
2385 total_size += z->present_pages;
2388 if (low_kmem_size &&
2389 total_size > average_size && /* ignore small node */
2390 low_kmem_size > total_size * 70/100)
2391 return ZONELIST_ORDER_NODE;
2393 return ZONELIST_ORDER_ZONE;
2396 static void set_zonelist_order(void)
2398 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2399 current_zonelist_order = default_zonelist_order();
2401 current_zonelist_order = user_zonelist_order;
2404 static void build_zonelists(pg_data_t *pgdat)
2408 nodemask_t used_mask;
2409 int local_node, prev_node;
2410 struct zonelist *zonelist;
2411 int order = current_zonelist_order;
2413 /* initialize zonelists */
2414 for (i = 0; i < MAX_ZONELISTS; i++) {
2415 zonelist = pgdat->node_zonelists + i;
2416 zonelist->_zonerefs[0].zone = NULL;
2417 zonelist->_zonerefs[0].zone_idx = 0;
2420 /* NUMA-aware ordering of nodes */
2421 local_node = pgdat->node_id;
2422 load = num_online_nodes();
2423 prev_node = local_node;
2424 nodes_clear(used_mask);
2426 memset(node_load, 0, sizeof(node_load));
2427 memset(node_order, 0, sizeof(node_order));
2430 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2431 int distance = node_distance(local_node, node);
2434 * If another node is sufficiently far away then it is better
2435 * to reclaim pages in a zone before going off node.
2437 if (distance > RECLAIM_DISTANCE)
2438 zone_reclaim_mode = 1;
2441 * We don't want to pressure a particular node.
2442 * So adding penalty to the first node in same
2443 * distance group to make it round-robin.
2445 if (distance != node_distance(local_node, prev_node))
2446 node_load[node] = load;
2450 if (order == ZONELIST_ORDER_NODE)
2451 build_zonelists_in_node_order(pgdat, node);
2453 node_order[j++] = node; /* remember order */
2456 if (order == ZONELIST_ORDER_ZONE) {
2457 /* calculate node order -- i.e., DMA last! */
2458 build_zonelists_in_zone_order(pgdat, j);
2461 build_thisnode_zonelists(pgdat);
2464 /* Construct the zonelist performance cache - see further mmzone.h */
2465 static void build_zonelist_cache(pg_data_t *pgdat)
2467 struct zonelist *zonelist;
2468 struct zonelist_cache *zlc;
2471 zonelist = &pgdat->node_zonelists[0];
2472 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2473 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2474 for (z = zonelist->_zonerefs; z->zone; z++)
2475 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2479 #else /* CONFIG_NUMA */
2481 static void set_zonelist_order(void)
2483 current_zonelist_order = ZONELIST_ORDER_ZONE;
2486 static void build_zonelists(pg_data_t *pgdat)
2488 int node, local_node;
2490 struct zonelist *zonelist;
2492 local_node = pgdat->node_id;
2494 zonelist = &pgdat->node_zonelists[0];
2495 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2498 * Now we build the zonelist so that it contains the zones
2499 * of all the other nodes.
2500 * We don't want to pressure a particular node, so when
2501 * building the zones for node N, we make sure that the
2502 * zones coming right after the local ones are those from
2503 * node N+1 (modulo N)
2505 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2506 if (!node_online(node))
2508 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2511 for (node = 0; node < local_node; node++) {
2512 if (!node_online(node))
2514 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2518 zonelist->_zonerefs[j].zone = NULL;
2519 zonelist->_zonerefs[j].zone_idx = 0;
2522 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2523 static void build_zonelist_cache(pg_data_t *pgdat)
2525 pgdat->node_zonelists[0].zlcache_ptr = NULL;
2528 #endif /* CONFIG_NUMA */
2530 /* return values int ....just for stop_machine() */
2531 static int __build_all_zonelists(void *dummy)
2535 for_each_online_node(nid) {
2536 pg_data_t *pgdat = NODE_DATA(nid);
2538 build_zonelists(pgdat);
2539 build_zonelist_cache(pgdat);
2544 void build_all_zonelists(void)
2546 set_zonelist_order();
2548 if (system_state == SYSTEM_BOOTING) {
2549 __build_all_zonelists(NULL);
2550 mminit_verify_zonelist();
2551 cpuset_init_current_mems_allowed();
2553 /* we have to stop all cpus to guarantee there is no user
2555 stop_machine(__build_all_zonelists, NULL, NULL);
2556 /* cpuset refresh routine should be here */
2558 vm_total_pages = nr_free_pagecache_pages();
2560 * Disable grouping by mobility if the number of pages in the
2561 * system is too low to allow the mechanism to work. It would be
2562 * more accurate, but expensive to check per-zone. This check is
2563 * made on memory-hotadd so a system can start with mobility
2564 * disabled and enable it later
2566 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
2567 page_group_by_mobility_disabled = 1;
2569 page_group_by_mobility_disabled = 0;
2571 printk("Built %i zonelists in %s order, mobility grouping %s. "
2572 "Total pages: %ld\n",
2574 zonelist_order_name[current_zonelist_order],
2575 page_group_by_mobility_disabled ? "off" : "on",
2578 printk("Policy zone: %s\n", zone_names[policy_zone]);
2583 * Helper functions to size the waitqueue hash table.
2584 * Essentially these want to choose hash table sizes sufficiently
2585 * large so that collisions trying to wait on pages are rare.
2586 * But in fact, the number of active page waitqueues on typical
2587 * systems is ridiculously low, less than 200. So this is even
2588 * conservative, even though it seems large.
2590 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2591 * waitqueues, i.e. the size of the waitq table given the number of pages.
2593 #define PAGES_PER_WAITQUEUE 256
2595 #ifndef CONFIG_MEMORY_HOTPLUG
2596 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2598 unsigned long size = 1;
2600 pages /= PAGES_PER_WAITQUEUE;
2602 while (size < pages)
2606 * Once we have dozens or even hundreds of threads sleeping
2607 * on IO we've got bigger problems than wait queue collision.
2608 * Limit the size of the wait table to a reasonable size.
2610 size = min(size, 4096UL);
2612 return max(size, 4UL);
2616 * A zone's size might be changed by hot-add, so it is not possible to determine
2617 * a suitable size for its wait_table. So we use the maximum size now.
2619 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2621 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2622 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2623 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2625 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2626 * or more by the traditional way. (See above). It equals:
2628 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2629 * ia64(16K page size) : = ( 8G + 4M)byte.
2630 * powerpc (64K page size) : = (32G +16M)byte.
2632 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2639 * This is an integer logarithm so that shifts can be used later
2640 * to extract the more random high bits from the multiplicative
2641 * hash function before the remainder is taken.
2643 static inline unsigned long wait_table_bits(unsigned long size)
2648 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2651 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2652 * of blocks reserved is based on zone->pages_min. The memory within the
2653 * reserve will tend to store contiguous free pages. Setting min_free_kbytes
2654 * higher will lead to a bigger reserve which will get freed as contiguous
2655 * blocks as reclaim kicks in
2657 static void setup_zone_migrate_reserve(struct zone *zone)
2659 unsigned long start_pfn, pfn, end_pfn;
2661 unsigned long reserve, block_migratetype;
2663 /* Get the start pfn, end pfn and the number of blocks to reserve */
2664 start_pfn = zone->zone_start_pfn;
2665 end_pfn = start_pfn + zone->spanned_pages;
2666 reserve = roundup(zone->pages_min, pageblock_nr_pages) >>
2669 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
2670 if (!pfn_valid(pfn))
2672 page = pfn_to_page(pfn);
2674 /* Watch out for overlapping nodes */
2675 if (page_to_nid(page) != zone_to_nid(zone))
2678 /* Blocks with reserved pages will never free, skip them. */
2679 if (PageReserved(page))
2682 block_migratetype = get_pageblock_migratetype(page);
2684 /* If this block is reserved, account for it */
2685 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
2690 /* Suitable for reserving if this block is movable */
2691 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
2692 set_pageblock_migratetype(page, MIGRATE_RESERVE);
2693 move_freepages_block(zone, page, MIGRATE_RESERVE);
2699 * If the reserve is met and this is a previous reserved block,
2702 if (block_migratetype == MIGRATE_RESERVE) {
2703 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2704 move_freepages_block(zone, page, MIGRATE_MOVABLE);
2710 * Initially all pages are reserved - free ones are freed
2711 * up by free_all_bootmem() once the early boot process is
2712 * done. Non-atomic initialization, single-pass.
2714 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
2715 unsigned long start_pfn, enum memmap_context context)
2718 unsigned long end_pfn = start_pfn + size;
2722 if (highest_memmap_pfn < end_pfn - 1)
2723 highest_memmap_pfn = end_pfn - 1;
2725 z = &NODE_DATA(nid)->node_zones[zone];
2726 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
2728 * There can be holes in boot-time mem_map[]s
2729 * handed to this function. They do not
2730 * exist on hotplugged memory.
2732 if (context == MEMMAP_EARLY) {
2733 if (!early_pfn_valid(pfn))
2735 if (!early_pfn_in_nid(pfn, nid))
2738 page = pfn_to_page(pfn);
2739 set_page_links(page, zone, nid, pfn);
2740 mminit_verify_page_links(page, zone, nid, pfn);
2741 init_page_count(page);
2742 reset_page_mapcount(page);
2743 SetPageReserved(page);
2745 * Mark the block movable so that blocks are reserved for
2746 * movable at startup. This will force kernel allocations
2747 * to reserve their blocks rather than leaking throughout
2748 * the address space during boot when many long-lived
2749 * kernel allocations are made. Later some blocks near
2750 * the start are marked MIGRATE_RESERVE by
2751 * setup_zone_migrate_reserve()
2753 * bitmap is created for zone's valid pfn range. but memmap
2754 * can be created for invalid pages (for alignment)
2755 * check here not to call set_pageblock_migratetype() against
2758 if ((z->zone_start_pfn <= pfn)
2759 && (pfn < z->zone_start_pfn + z->spanned_pages)
2760 && !(pfn & (pageblock_nr_pages - 1)))
2761 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2763 INIT_LIST_HEAD(&page->lru);
2764 #ifdef WANT_PAGE_VIRTUAL
2765 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2766 if (!is_highmem_idx(zone))
2767 set_page_address(page, __va(pfn << PAGE_SHIFT));
2772 static void __meminit zone_init_free_lists(struct zone *zone)
2775 for_each_migratetype_order(order, t) {
2776 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
2777 zone->free_area[order].nr_free = 0;
2781 #ifndef __HAVE_ARCH_MEMMAP_INIT
2782 #define memmap_init(size, nid, zone, start_pfn) \
2783 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
2786 static int zone_batchsize(struct zone *zone)
2792 * The per-cpu-pages pools are set to around 1000th of the
2793 * size of the zone. But no more than 1/2 of a meg.
2795 * OK, so we don't know how big the cache is. So guess.
2797 batch = zone->present_pages / 1024;
2798 if (batch * PAGE_SIZE > 512 * 1024)
2799 batch = (512 * 1024) / PAGE_SIZE;
2800 batch /= 4; /* We effectively *= 4 below */
2805 * Clamp the batch to a 2^n - 1 value. Having a power
2806 * of 2 value was found to be more likely to have
2807 * suboptimal cache aliasing properties in some cases.
2809 * For example if 2 tasks are alternately allocating
2810 * batches of pages, one task can end up with a lot
2811 * of pages of one half of the possible page colors
2812 * and the other with pages of the other colors.
2814 batch = rounddown_pow_of_two(batch + batch/2) - 1;
2819 /* The deferral and batching of frees should be suppressed under NOMMU
2822 * The problem is that NOMMU needs to be able to allocate large chunks
2823 * of contiguous memory as there's no hardware page translation to
2824 * assemble apparent contiguous memory from discontiguous pages.
2826 * Queueing large contiguous runs of pages for batching, however,
2827 * causes the pages to actually be freed in smaller chunks. As there
2828 * can be a significant delay between the individual batches being
2829 * recycled, this leads to the once large chunks of space being
2830 * fragmented and becoming unavailable for high-order allocations.
2836 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
2838 struct per_cpu_pages *pcp;
2840 memset(p, 0, sizeof(*p));
2844 pcp->high = 6 * batch;
2845 pcp->batch = max(1UL, 1 * batch);
2846 INIT_LIST_HEAD(&pcp->list);
2850 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
2851 * to the value high for the pageset p.
2854 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
2857 struct per_cpu_pages *pcp;
2861 pcp->batch = max(1UL, high/4);
2862 if ((high/4) > (PAGE_SHIFT * 8))
2863 pcp->batch = PAGE_SHIFT * 8;
2869 * Boot pageset table. One per cpu which is going to be used for all
2870 * zones and all nodes. The parameters will be set in such a way
2871 * that an item put on a list will immediately be handed over to
2872 * the buddy list. This is safe since pageset manipulation is done
2873 * with interrupts disabled.
2875 * Some NUMA counter updates may also be caught by the boot pagesets.
2877 * The boot_pagesets must be kept even after bootup is complete for
2878 * unused processors and/or zones. They do play a role for bootstrapping
2879 * hotplugged processors.
2881 * zoneinfo_show() and maybe other functions do
2882 * not check if the processor is online before following the pageset pointer.
2883 * Other parts of the kernel may not check if the zone is available.
2885 static struct per_cpu_pageset boot_pageset[NR_CPUS];
2888 * Dynamically allocate memory for the
2889 * per cpu pageset array in struct zone.
2891 static int __cpuinit process_zones(int cpu)
2893 struct zone *zone, *dzone;
2894 int node = cpu_to_node(cpu);
2896 node_set_state(node, N_CPU); /* this node has a cpu */
2898 for_each_populated_zone(zone) {
2899 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
2901 if (!zone_pcp(zone, cpu))
2904 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
2906 if (percpu_pagelist_fraction)
2907 setup_pagelist_highmark(zone_pcp(zone, cpu),
2908 (zone->present_pages / percpu_pagelist_fraction));
2913 for_each_zone(dzone) {
2914 if (!populated_zone(dzone))
2918 kfree(zone_pcp(dzone, cpu));
2919 zone_pcp(dzone, cpu) = NULL;
2924 static inline void free_zone_pagesets(int cpu)
2928 for_each_zone(zone) {
2929 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
2931 /* Free per_cpu_pageset if it is slab allocated */
2932 if (pset != &boot_pageset[cpu])
2934 zone_pcp(zone, cpu) = NULL;
2938 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
2939 unsigned long action,
2942 int cpu = (long)hcpu;
2943 int ret = NOTIFY_OK;
2946 case CPU_UP_PREPARE:
2947 case CPU_UP_PREPARE_FROZEN:
2948 if (process_zones(cpu))
2951 case CPU_UP_CANCELED:
2952 case CPU_UP_CANCELED_FROZEN:
2954 case CPU_DEAD_FROZEN:
2955 free_zone_pagesets(cpu);
2963 static struct notifier_block __cpuinitdata pageset_notifier =
2964 { &pageset_cpuup_callback, NULL, 0 };
2966 void __init setup_per_cpu_pageset(void)
2970 /* Initialize per_cpu_pageset for cpu 0.
2971 * A cpuup callback will do this for every cpu
2972 * as it comes online
2974 err = process_zones(smp_processor_id());
2976 register_cpu_notifier(&pageset_notifier);
2981 static noinline __init_refok
2982 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
2985 struct pglist_data *pgdat = zone->zone_pgdat;
2989 * The per-page waitqueue mechanism uses hashed waitqueues
2992 zone->wait_table_hash_nr_entries =
2993 wait_table_hash_nr_entries(zone_size_pages);
2994 zone->wait_table_bits =
2995 wait_table_bits(zone->wait_table_hash_nr_entries);
2996 alloc_size = zone->wait_table_hash_nr_entries
2997 * sizeof(wait_queue_head_t);
2999 if (!slab_is_available()) {
3000 zone->wait_table = (wait_queue_head_t *)
3001 alloc_bootmem_node(pgdat, alloc_size);
3004 * This case means that a zone whose size was 0 gets new memory
3005 * via memory hot-add.
3006 * But it may be the case that a new node was hot-added. In
3007 * this case vmalloc() will not be able to use this new node's
3008 * memory - this wait_table must be initialized to use this new
3009 * node itself as well.
3010 * To use this new node's memory, further consideration will be
3013 zone->wait_table = vmalloc(alloc_size);
3015 if (!zone->wait_table)
3018 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3019 init_waitqueue_head(zone->wait_table + i);
3024 static __meminit void zone_pcp_init(struct zone *zone)
3027 unsigned long batch = zone_batchsize(zone);
3029 for (cpu = 0; cpu < NR_CPUS; cpu++) {
3031 /* Early boot. Slab allocator not functional yet */
3032 zone_pcp(zone, cpu) = &boot_pageset[cpu];
3033 setup_pageset(&boot_pageset[cpu],0);
3035 setup_pageset(zone_pcp(zone,cpu), batch);
3038 if (zone->present_pages)
3039 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
3040 zone->name, zone->present_pages, batch);
3043 __meminit int init_currently_empty_zone(struct zone *zone,
3044 unsigned long zone_start_pfn,
3046 enum memmap_context context)
3048 struct pglist_data *pgdat = zone->zone_pgdat;
3050 ret = zone_wait_table_init(zone, size);
3053 pgdat->nr_zones = zone_idx(zone) + 1;
3055 zone->zone_start_pfn = zone_start_pfn;
3057 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3058 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3060 (unsigned long)zone_idx(zone),
3061 zone_start_pfn, (zone_start_pfn + size));
3063 zone_init_free_lists(zone);
3068 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3070 * Basic iterator support. Return the first range of PFNs for a node
3071 * Note: nid == MAX_NUMNODES returns first region regardless of node
3073 static int __meminit first_active_region_index_in_nid(int nid)
3077 for (i = 0; i < nr_nodemap_entries; i++)
3078 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3085 * Basic iterator support. Return the next active range of PFNs for a node
3086 * Note: nid == MAX_NUMNODES returns next region regardless of node
3088 static int __meminit next_active_region_index_in_nid(int index, int nid)
3090 for (index = index + 1; index < nr_nodemap_entries; index++)
3091 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3097 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3099 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3100 * Architectures may implement their own version but if add_active_range()
3101 * was used and there are no special requirements, this is a convenient
3104 int __meminit __early_pfn_to_nid(unsigned long pfn)
3108 for (i = 0; i < nr_nodemap_entries; i++) {
3109 unsigned long start_pfn = early_node_map[i].start_pfn;
3110 unsigned long end_pfn = early_node_map[i].end_pfn;
3112 if (start_pfn <= pfn && pfn < end_pfn)
3113 return early_node_map[i].nid;
3115 /* This is a memory hole */
3118 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3120 int __meminit early_pfn_to_nid(unsigned long pfn)
3124 nid = __early_pfn_to_nid(pfn);
3127 /* just returns 0 */
3131 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3132 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3136 nid = __early_pfn_to_nid(pfn);
3137 if (nid >= 0 && nid != node)
3143 /* Basic iterator support to walk early_node_map[] */
3144 #define for_each_active_range_index_in_nid(i, nid) \
3145 for (i = first_active_region_index_in_nid(nid); i != -1; \
3146 i = next_active_region_index_in_nid(i, nid))
3149 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3150 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3151 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3153 * If an architecture guarantees that all ranges registered with
3154 * add_active_ranges() contain no holes and may be freed, this
3155 * this function may be used instead of calling free_bootmem() manually.
3157 void __init free_bootmem_with_active_regions(int nid,
3158 unsigned long max_low_pfn)
3162 for_each_active_range_index_in_nid(i, nid) {
3163 unsigned long size_pages = 0;
3164 unsigned long end_pfn = early_node_map[i].end_pfn;
3166 if (early_node_map[i].start_pfn >= max_low_pfn)
3169 if (end_pfn > max_low_pfn)
3170 end_pfn = max_low_pfn;
3172 size_pages = end_pfn - early_node_map[i].start_pfn;
3173 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3174 PFN_PHYS(early_node_map[i].start_pfn),
3175 size_pages << PAGE_SHIFT);
3179 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3184 for_each_active_range_index_in_nid(i, nid) {
3185 ret = work_fn(early_node_map[i].start_pfn,
3186 early_node_map[i].end_pfn, data);
3192 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3193 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3195 * If an architecture guarantees that all ranges registered with
3196 * add_active_ranges() contain no holes and may be freed, this
3197 * function may be used instead of calling memory_present() manually.
3199 void __init sparse_memory_present_with_active_regions(int nid)
3203 for_each_active_range_index_in_nid(i, nid)
3204 memory_present(early_node_map[i].nid,
3205 early_node_map[i].start_pfn,
3206 early_node_map[i].end_pfn);
3210 * get_pfn_range_for_nid - Return the start and end page frames for a node
3211 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3212 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3213 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3215 * It returns the start and end page frame of a node based on information
3216 * provided by an arch calling add_active_range(). If called for a node
3217 * with no available memory, a warning is printed and the start and end
3220 void __meminit get_pfn_range_for_nid(unsigned int nid,
3221 unsigned long *start_pfn, unsigned long *end_pfn)
3227 for_each_active_range_index_in_nid(i, nid) {
3228 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3229 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3232 if (*start_pfn == -1UL)
3237 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3238 * assumption is made that zones within a node are ordered in monotonic
3239 * increasing memory addresses so that the "highest" populated zone is used
3241 static void __init find_usable_zone_for_movable(void)
3244 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3245 if (zone_index == ZONE_MOVABLE)
3248 if (arch_zone_highest_possible_pfn[zone_index] >
3249 arch_zone_lowest_possible_pfn[zone_index])
3253 VM_BUG_ON(zone_index == -1);
3254 movable_zone = zone_index;
3258 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3259 * because it is sized independant of architecture. Unlike the other zones,
3260 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3261 * in each node depending on the size of each node and how evenly kernelcore
3262 * is distributed. This helper function adjusts the zone ranges
3263 * provided by the architecture for a given node by using the end of the
3264 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3265 * zones within a node are in order of monotonic increases memory addresses
3267 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3268 unsigned long zone_type,
3269 unsigned long node_start_pfn,
3270 unsigned long node_end_pfn,
3271 unsigned long *zone_start_pfn,
3272 unsigned long *zone_end_pfn)
3274 /* Only adjust if ZONE_MOVABLE is on this node */
3275 if (zone_movable_pfn[nid]) {
3276 /* Size ZONE_MOVABLE */
3277 if (zone_type == ZONE_MOVABLE) {
3278 *zone_start_pfn = zone_movable_pfn[nid];
3279 *zone_end_pfn = min(node_end_pfn,
3280 arch_zone_highest_possible_pfn[movable_zone]);
3282 /* Adjust for ZONE_MOVABLE starting within this range */
3283 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3284 *zone_end_pfn > zone_movable_pfn[nid]) {
3285 *zone_end_pfn = zone_movable_pfn[nid];
3287 /* Check if this whole range is within ZONE_MOVABLE */
3288 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3289 *zone_start_pfn = *zone_end_pfn;
3294 * Return the number of pages a zone spans in a node, including holes
3295 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3297 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3298 unsigned long zone_type,
3299 unsigned long *ignored)
3301 unsigned long node_start_pfn, node_end_pfn;
3302 unsigned long zone_start_pfn, zone_end_pfn;
3304 /* Get the start and end of the node and zone */
3305 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3306 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3307 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3308 adjust_zone_range_for_zone_movable(nid, zone_type,
3309 node_start_pfn, node_end_pfn,
3310 &zone_start_pfn, &zone_end_pfn);
3312 /* Check that this node has pages within the zone's required range */
3313 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3316 /* Move the zone boundaries inside the node if necessary */
3317 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3318 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3320 /* Return the spanned pages */
3321 return zone_end_pfn - zone_start_pfn;
3325 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3326 * then all holes in the requested range will be accounted for.
3328 static unsigned long __meminit __absent_pages_in_range(int nid,
3329 unsigned long range_start_pfn,
3330 unsigned long range_end_pfn)
3333 unsigned long prev_end_pfn = 0, hole_pages = 0;
3334 unsigned long start_pfn;
3336 /* Find the end_pfn of the first active range of pfns in the node */
3337 i = first_active_region_index_in_nid(nid);
3341 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3343 /* Account for ranges before physical memory on this node */
3344 if (early_node_map[i].start_pfn > range_start_pfn)
3345 hole_pages = prev_end_pfn - range_start_pfn;
3347 /* Find all holes for the zone within the node */
3348 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3350 /* No need to continue if prev_end_pfn is outside the zone */
3351 if (prev_end_pfn >= range_end_pfn)
3354 /* Make sure the end of the zone is not within the hole */
3355 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3356 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3358 /* Update the hole size cound and move on */
3359 if (start_pfn > range_start_pfn) {
3360 BUG_ON(prev_end_pfn > start_pfn);
3361 hole_pages += start_pfn - prev_end_pfn;
3363 prev_end_pfn = early_node_map[i].end_pfn;
3366 /* Account for ranges past physical memory on this node */
3367 if (range_end_pfn > prev_end_pfn)
3368 hole_pages += range_end_pfn -
3369 max(range_start_pfn, prev_end_pfn);
3375 * absent_pages_in_range - Return number of page frames in holes within a range
3376 * @start_pfn: The start PFN to start searching for holes
3377 * @end_pfn: The end PFN to stop searching for holes
3379 * It returns the number of pages frames in memory holes within a range.
3381 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3382 unsigned long end_pfn)
3384 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3387 /* Return the number of page frames in holes in a zone on a node */
3388 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3389 unsigned long zone_type,
3390 unsigned long *ignored)
3392 unsigned long node_start_pfn, node_end_pfn;
3393 unsigned long zone_start_pfn, zone_end_pfn;
3395 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3396 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3398 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3401 adjust_zone_range_for_zone_movable(nid, zone_type,
3402 node_start_pfn, node_end_pfn,
3403 &zone_start_pfn, &zone_end_pfn);
3404 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3408 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3409 unsigned long zone_type,
3410 unsigned long *zones_size)
3412 return zones_size[zone_type];
3415 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3416 unsigned long zone_type,
3417 unsigned long *zholes_size)
3422 return zholes_size[zone_type];
3427 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3428 unsigned long *zones_size, unsigned long *zholes_size)
3430 unsigned long realtotalpages, totalpages = 0;
3433 for (i = 0; i < MAX_NR_ZONES; i++)
3434 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3436 pgdat->node_spanned_pages = totalpages;
3438 realtotalpages = totalpages;
3439 for (i = 0; i < MAX_NR_ZONES; i++)
3441 zone_absent_pages_in_node(pgdat->node_id, i,
3443 pgdat->node_present_pages = realtotalpages;
3444 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3448 #ifndef CONFIG_SPARSEMEM
3450 * Calculate the size of the zone->blockflags rounded to an unsigned long
3451 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3452 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3453 * round what is now in bits to nearest long in bits, then return it in
3456 static unsigned long __init usemap_size(unsigned long zonesize)
3458 unsigned long usemapsize;
3460 usemapsize = roundup(zonesize, pageblock_nr_pages);
3461 usemapsize = usemapsize >> pageblock_order;
3462 usemapsize *= NR_PAGEBLOCK_BITS;
3463 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
3465 return usemapsize / 8;
3468 static void __init setup_usemap(struct pglist_data *pgdat,
3469 struct zone *zone, unsigned long zonesize)
3471 unsigned long usemapsize = usemap_size(zonesize);
3472 zone->pageblock_flags = NULL;
3474 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
3477 static void inline setup_usemap(struct pglist_data *pgdat,
3478 struct zone *zone, unsigned long zonesize) {}
3479 #endif /* CONFIG_SPARSEMEM */
3481 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3483 /* Return a sensible default order for the pageblock size. */
3484 static inline int pageblock_default_order(void)
3486 if (HPAGE_SHIFT > PAGE_SHIFT)
3487 return HUGETLB_PAGE_ORDER;
3492 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3493 static inline void __init set_pageblock_order(unsigned int order)
3495 /* Check that pageblock_nr_pages has not already been setup */
3496 if (pageblock_order)
3500 * Assume the largest contiguous order of interest is a huge page.
3501 * This value may be variable depending on boot parameters on IA64
3503 pageblock_order = order;
3505 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3508 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3509 * and pageblock_default_order() are unused as pageblock_order is set
3510 * at compile-time. See include/linux/pageblock-flags.h for the values of
3511 * pageblock_order based on the kernel config
3513 static inline int pageblock_default_order(unsigned int order)
3517 #define set_pageblock_order(x) do {} while (0)
3519 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3522 * Set up the zone data structures:
3523 * - mark all pages reserved
3524 * - mark all memory queues empty
3525 * - clear the memory bitmaps
3527 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
3528 unsigned long *zones_size, unsigned long *zholes_size)
3531 int nid = pgdat->node_id;
3532 unsigned long zone_start_pfn = pgdat->node_start_pfn;
3535 pgdat_resize_init(pgdat);
3536 pgdat->nr_zones = 0;
3537 init_waitqueue_head(&pgdat->kswapd_wait);
3538 pgdat->kswapd_max_order = 0;
3539 pgdat_page_cgroup_init(pgdat);
3541 for (j = 0; j < MAX_NR_ZONES; j++) {
3542 struct zone *zone = pgdat->node_zones + j;
3543 unsigned long size, realsize, memmap_pages;
3546 size = zone_spanned_pages_in_node(nid, j, zones_size);
3547 realsize = size - zone_absent_pages_in_node(nid, j,
3551 * Adjust realsize so that it accounts for how much memory
3552 * is used by this zone for memmap. This affects the watermark
3553 * and per-cpu initialisations
3556 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
3557 if (realsize >= memmap_pages) {
3558 realsize -= memmap_pages;
3561 " %s zone: %lu pages used for memmap\n",
3562 zone_names[j], memmap_pages);
3565 " %s zone: %lu pages exceeds realsize %lu\n",
3566 zone_names[j], memmap_pages, realsize);
3568 /* Account for reserved pages */
3569 if (j == 0 && realsize > dma_reserve) {
3570 realsize -= dma_reserve;
3571 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
3572 zone_names[0], dma_reserve);
3575 if (!is_highmem_idx(j))
3576 nr_kernel_pages += realsize;
3577 nr_all_pages += realsize;
3579 zone->spanned_pages = size;
3580 zone->present_pages = realsize;
3583 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
3585 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
3587 zone->name = zone_names[j];
3588 spin_lock_init(&zone->lock);
3589 spin_lock_init(&zone->lru_lock);
3590 zone_seqlock_init(zone);
3591 zone->zone_pgdat = pgdat;
3593 zone->prev_priority = DEF_PRIORITY;
3595 zone_pcp_init(zone);
3597 INIT_LIST_HEAD(&zone->lru[l].list);
3598 zone->lru[l].nr_scan = 0;
3600 zone->reclaim_stat.recent_rotated[0] = 0;
3601 zone->reclaim_stat.recent_rotated[1] = 0;
3602 zone->reclaim_stat.recent_scanned[0] = 0;
3603 zone->reclaim_stat.recent_scanned[1] = 0;
3604 zap_zone_vm_stats(zone);
3609 set_pageblock_order(pageblock_default_order());
3610 setup_usemap(pgdat, zone, size);
3611 ret = init_currently_empty_zone(zone, zone_start_pfn,
3612 size, MEMMAP_EARLY);
3614 memmap_init(size, nid, j, zone_start_pfn);
3615 zone_start_pfn += size;
3619 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
3621 /* Skip empty nodes */
3622 if (!pgdat->node_spanned_pages)
3625 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3626 /* ia64 gets its own node_mem_map, before this, without bootmem */
3627 if (!pgdat->node_mem_map) {
3628 unsigned long size, start, end;
3632 * The zone's endpoints aren't required to be MAX_ORDER
3633 * aligned but the node_mem_map endpoints must be in order
3634 * for the buddy allocator to function correctly.
3636 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
3637 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
3638 end = ALIGN(end, MAX_ORDER_NR_PAGES);
3639 size = (end - start) * sizeof(struct page);
3640 map = alloc_remap(pgdat->node_id, size);
3642 map = alloc_bootmem_node(pgdat, size);
3643 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
3645 #ifndef CONFIG_NEED_MULTIPLE_NODES
3647 * With no DISCONTIG, the global mem_map is just set as node 0's
3649 if (pgdat == NODE_DATA(0)) {
3650 mem_map = NODE_DATA(0)->node_mem_map;
3651 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3652 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
3653 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
3654 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3657 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3660 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
3661 unsigned long node_start_pfn, unsigned long *zholes_size)
3663 pg_data_t *pgdat = NODE_DATA(nid);
3665 pgdat->node_id = nid;
3666 pgdat->node_start_pfn = node_start_pfn;
3667 calculate_node_totalpages(pgdat, zones_size, zholes_size);
3669 alloc_node_mem_map(pgdat);
3670 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3671 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3672 nid, (unsigned long)pgdat,
3673 (unsigned long)pgdat->node_mem_map);
3676 free_area_init_core(pgdat, zones_size, zholes_size);
3679 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3681 #if MAX_NUMNODES > 1
3683 * Figure out the number of possible node ids.
3685 static void __init setup_nr_node_ids(void)
3688 unsigned int highest = 0;
3690 for_each_node_mask(node, node_possible_map)
3692 nr_node_ids = highest + 1;
3695 static inline void setup_nr_node_ids(void)
3701 * add_active_range - Register a range of PFNs backed by physical memory
3702 * @nid: The node ID the range resides on
3703 * @start_pfn: The start PFN of the available physical memory
3704 * @end_pfn: The end PFN of the available physical memory
3706 * These ranges are stored in an early_node_map[] and later used by
3707 * free_area_init_nodes() to calculate zone sizes and holes. If the
3708 * range spans a memory hole, it is up to the architecture to ensure
3709 * the memory is not freed by the bootmem allocator. If possible
3710 * the range being registered will be merged with existing ranges.
3712 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
3713 unsigned long end_pfn)
3717 mminit_dprintk(MMINIT_TRACE, "memory_register",
3718 "Entering add_active_range(%d, %#lx, %#lx) "
3719 "%d entries of %d used\n",
3720 nid, start_pfn, end_pfn,
3721 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
3723 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
3725 /* Merge with existing active regions if possible */
3726 for (i = 0; i < nr_nodemap_entries; i++) {
3727 if (early_node_map[i].nid != nid)
3730 /* Skip if an existing region covers this new one */
3731 if (start_pfn >= early_node_map[i].start_pfn &&
3732 end_pfn <= early_node_map[i].end_pfn)
3735 /* Merge forward if suitable */
3736 if (start_pfn <= early_node_map[i].end_pfn &&
3737 end_pfn > early_node_map[i].end_pfn) {
3738 early_node_map[i].end_pfn = end_pfn;
3742 /* Merge backward if suitable */
3743 if (start_pfn < early_node_map[i].end_pfn &&
3744 end_pfn >= early_node_map[i].start_pfn) {
3745 early_node_map[i].start_pfn = start_pfn;
3750 /* Check that early_node_map is large enough */
3751 if (i >= MAX_ACTIVE_REGIONS) {
3752 printk(KERN_CRIT "More than %d memory regions, truncating\n",
3753 MAX_ACTIVE_REGIONS);
3757 early_node_map[i].nid = nid;
3758 early_node_map[i].start_pfn = start_pfn;
3759 early_node_map[i].end_pfn = end_pfn;
3760 nr_nodemap_entries = i + 1;
3764 * remove_active_range - Shrink an existing registered range of PFNs
3765 * @nid: The node id the range is on that should be shrunk
3766 * @start_pfn: The new PFN of the range
3767 * @end_pfn: The new PFN of the range
3769 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
3770 * The map is kept near the end physical page range that has already been
3771 * registered. This function allows an arch to shrink an existing registered
3774 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
3775 unsigned long end_pfn)
3780 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
3781 nid, start_pfn, end_pfn);
3783 /* Find the old active region end and shrink */
3784 for_each_active_range_index_in_nid(i, nid) {
3785 if (early_node_map[i].start_pfn >= start_pfn &&
3786 early_node_map[i].end_pfn <= end_pfn) {
3788 early_node_map[i].start_pfn = 0;
3789 early_node_map[i].end_pfn = 0;
3793 if (early_node_map[i].start_pfn < start_pfn &&
3794 early_node_map[i].end_pfn > start_pfn) {
3795 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
3796 early_node_map[i].end_pfn = start_pfn;
3797 if (temp_end_pfn > end_pfn)
3798 add_active_range(nid, end_pfn, temp_end_pfn);
3801 if (early_node_map[i].start_pfn >= start_pfn &&
3802 early_node_map[i].end_pfn > end_pfn &&
3803 early_node_map[i].start_pfn < end_pfn) {
3804 early_node_map[i].start_pfn = end_pfn;
3812 /* remove the blank ones */
3813 for (i = nr_nodemap_entries - 1; i > 0; i--) {
3814 if (early_node_map[i].nid != nid)
3816 if (early_node_map[i].end_pfn)
3818 /* we found it, get rid of it */
3819 for (j = i; j < nr_nodemap_entries - 1; j++)
3820 memcpy(&early_node_map[j], &early_node_map[j+1],
3821 sizeof(early_node_map[j]));
3822 j = nr_nodemap_entries - 1;
3823 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
3824 nr_nodemap_entries--;
3829 * remove_all_active_ranges - Remove all currently registered regions
3831 * During discovery, it may be found that a table like SRAT is invalid
3832 * and an alternative discovery method must be used. This function removes
3833 * all currently registered regions.
3835 void __init remove_all_active_ranges(void)
3837 memset(early_node_map, 0, sizeof(early_node_map));
3838 nr_nodemap_entries = 0;
3841 /* Compare two active node_active_regions */
3842 static int __init cmp_node_active_region(const void *a, const void *b)
3844 struct node_active_region *arange = (struct node_active_region *)a;
3845 struct node_active_region *brange = (struct node_active_region *)b;
3847 /* Done this way to avoid overflows */
3848 if (arange->start_pfn > brange->start_pfn)
3850 if (arange->start_pfn < brange->start_pfn)
3856 /* sort the node_map by start_pfn */
3857 static void __init sort_node_map(void)
3859 sort(early_node_map, (size_t)nr_nodemap_entries,
3860 sizeof(struct node_active_region),
3861 cmp_node_active_region, NULL);
3864 /* Find the lowest pfn for a node */
3865 static unsigned long __init find_min_pfn_for_node(int nid)
3868 unsigned long min_pfn = ULONG_MAX;
3870 /* Assuming a sorted map, the first range found has the starting pfn */
3871 for_each_active_range_index_in_nid(i, nid)
3872 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
3874 if (min_pfn == ULONG_MAX) {
3876 "Could not find start_pfn for node %d\n", nid);
3884 * find_min_pfn_with_active_regions - Find the minimum PFN registered
3886 * It returns the minimum PFN based on information provided via
3887 * add_active_range().
3889 unsigned long __init find_min_pfn_with_active_regions(void)
3891 return find_min_pfn_for_node(MAX_NUMNODES);
3895 * early_calculate_totalpages()
3896 * Sum pages in active regions for movable zone.
3897 * Populate N_HIGH_MEMORY for calculating usable_nodes.
3899 static unsigned long __init early_calculate_totalpages(void)
3902 unsigned long totalpages = 0;
3904 for (i = 0; i < nr_nodemap_entries; i++) {
3905 unsigned long pages = early_node_map[i].end_pfn -
3906 early_node_map[i].start_pfn;
3907 totalpages += pages;
3909 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
3915 * Find the PFN the Movable zone begins in each node. Kernel memory
3916 * is spread evenly between nodes as long as the nodes have enough
3917 * memory. When they don't, some nodes will have more kernelcore than
3920 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
3923 unsigned long usable_startpfn;
3924 unsigned long kernelcore_node, kernelcore_remaining;
3925 unsigned long totalpages = early_calculate_totalpages();
3926 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
3929 * If movablecore was specified, calculate what size of
3930 * kernelcore that corresponds so that memory usable for
3931 * any allocation type is evenly spread. If both kernelcore
3932 * and movablecore are specified, then the value of kernelcore
3933 * will be used for required_kernelcore if it's greater than
3934 * what movablecore would have allowed.
3936 if (required_movablecore) {
3937 unsigned long corepages;
3940 * Round-up so that ZONE_MOVABLE is at least as large as what
3941 * was requested by the user
3943 required_movablecore =
3944 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
3945 corepages = totalpages - required_movablecore;
3947 required_kernelcore = max(required_kernelcore, corepages);
3950 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
3951 if (!required_kernelcore)
3954 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
3955 find_usable_zone_for_movable();
3956 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
3959 /* Spread kernelcore memory as evenly as possible throughout nodes */
3960 kernelcore_node = required_kernelcore / usable_nodes;
3961 for_each_node_state(nid, N_HIGH_MEMORY) {
3963 * Recalculate kernelcore_node if the division per node
3964 * now exceeds what is necessary to satisfy the requested
3965 * amount of memory for the kernel
3967 if (required_kernelcore < kernelcore_node)
3968 kernelcore_node = required_kernelcore / usable_nodes;
3971 * As the map is walked, we track how much memory is usable
3972 * by the kernel using kernelcore_remaining. When it is
3973 * 0, the rest of the node is usable by ZONE_MOVABLE
3975 kernelcore_remaining = kernelcore_node;
3977 /* Go through each range of PFNs within this node */
3978 for_each_active_range_index_in_nid(i, nid) {
3979 unsigned long start_pfn, end_pfn;
3980 unsigned long size_pages;
3982 start_pfn = max(early_node_map[i].start_pfn,
3983 zone_movable_pfn[nid]);
3984 end_pfn = early_node_map[i].end_pfn;
3985 if (start_pfn >= end_pfn)
3988 /* Account for what is only usable for kernelcore */
3989 if (start_pfn < usable_startpfn) {
3990 unsigned long kernel_pages;
3991 kernel_pages = min(end_pfn, usable_startpfn)
3994 kernelcore_remaining -= min(kernel_pages,
3995 kernelcore_remaining);
3996 required_kernelcore -= min(kernel_pages,
3997 required_kernelcore);
3999 /* Continue if range is now fully accounted */
4000 if (end_pfn <= usable_startpfn) {
4003 * Push zone_movable_pfn to the end so
4004 * that if we have to rebalance
4005 * kernelcore across nodes, we will
4006 * not double account here
4008 zone_movable_pfn[nid] = end_pfn;
4011 start_pfn = usable_startpfn;
4015 * The usable PFN range for ZONE_MOVABLE is from
4016 * start_pfn->end_pfn. Calculate size_pages as the
4017 * number of pages used as kernelcore
4019 size_pages = end_pfn - start_pfn;
4020 if (size_pages > kernelcore_remaining)
4021 size_pages = kernelcore_remaining;
4022 zone_movable_pfn[nid] = start_pfn + size_pages;
4025 * Some kernelcore has been met, update counts and
4026 * break if the kernelcore for this node has been
4029 required_kernelcore -= min(required_kernelcore,
4031 kernelcore_remaining -= size_pages;
4032 if (!kernelcore_remaining)
4038 * If there is still required_kernelcore, we do another pass with one
4039 * less node in the count. This will push zone_movable_pfn[nid] further
4040 * along on the nodes that still have memory until kernelcore is
4044 if (usable_nodes && required_kernelcore > usable_nodes)
4047 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4048 for (nid = 0; nid < MAX_NUMNODES; nid++)
4049 zone_movable_pfn[nid] =
4050 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4053 /* Any regular memory on that node ? */
4054 static void check_for_regular_memory(pg_data_t *pgdat)
4056 #ifdef CONFIG_HIGHMEM
4057 enum zone_type zone_type;
4059 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4060 struct zone *zone = &pgdat->node_zones[zone_type];
4061 if (zone->present_pages)
4062 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4068 * free_area_init_nodes - Initialise all pg_data_t and zone data
4069 * @max_zone_pfn: an array of max PFNs for each zone
4071 * This will call free_area_init_node() for each active node in the system.
4072 * Using the page ranges provided by add_active_range(), the size of each
4073 * zone in each node and their holes is calculated. If the maximum PFN
4074 * between two adjacent zones match, it is assumed that the zone is empty.
4075 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4076 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4077 * starts where the previous one ended. For example, ZONE_DMA32 starts
4078 * at arch_max_dma_pfn.
4080 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4085 /* Sort early_node_map as initialisation assumes it is sorted */
4088 /* Record where the zone boundaries are */
4089 memset(arch_zone_lowest_possible_pfn, 0,
4090 sizeof(arch_zone_lowest_possible_pfn));
4091 memset(arch_zone_highest_possible_pfn, 0,
4092 sizeof(arch_zone_highest_possible_pfn));
4093 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4094 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4095 for (i = 1; i < MAX_NR_ZONES; i++) {
4096 if (i == ZONE_MOVABLE)
4098 arch_zone_lowest_possible_pfn[i] =
4099 arch_zone_highest_possible_pfn[i-1];
4100 arch_zone_highest_possible_pfn[i] =
4101 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4103 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4104 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4106 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4107 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4108 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4110 /* Print out the zone ranges */
4111 printk("Zone PFN ranges:\n");
4112 for (i = 0; i < MAX_NR_ZONES; i++) {
4113 if (i == ZONE_MOVABLE)
4115 printk(" %-8s %0#10lx -> %0#10lx\n",
4117 arch_zone_lowest_possible_pfn[i],
4118 arch_zone_highest_possible_pfn[i]);
4121 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4122 printk("Movable zone start PFN for each node\n");
4123 for (i = 0; i < MAX_NUMNODES; i++) {
4124 if (zone_movable_pfn[i])
4125 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4128 /* Print out the early_node_map[] */
4129 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4130 for (i = 0; i < nr_nodemap_entries; i++)
4131 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4132 early_node_map[i].start_pfn,
4133 early_node_map[i].end_pfn);
4135 /* Initialise every node */
4136 mminit_verify_pageflags_layout();
4137 setup_nr_node_ids();
4138 for_each_online_node(nid) {
4139 pg_data_t *pgdat = NODE_DATA(nid);
4140 free_area_init_node(nid, NULL,
4141 find_min_pfn_for_node(nid), NULL);
4143 /* Any memory on that node */
4144 if (pgdat->node_present_pages)
4145 node_set_state(nid, N_HIGH_MEMORY);
4146 check_for_regular_memory(pgdat);
4150 static int __init cmdline_parse_core(char *p, unsigned long *core)
4152 unsigned long long coremem;
4156 coremem = memparse(p, &p);
4157 *core = coremem >> PAGE_SHIFT;
4159 /* Paranoid check that UL is enough for the coremem value */
4160 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4166 * kernelcore=size sets the amount of memory for use for allocations that
4167 * cannot be reclaimed or migrated.
4169 static int __init cmdline_parse_kernelcore(char *p)
4171 return cmdline_parse_core(p, &required_kernelcore);
4175 * movablecore=size sets the amount of memory for use for allocations that
4176 * can be reclaimed or migrated.
4178 static int __init cmdline_parse_movablecore(char *p)
4180 return cmdline_parse_core(p, &required_movablecore);
4183 early_param("kernelcore", cmdline_parse_kernelcore);
4184 early_param("movablecore", cmdline_parse_movablecore);
4186 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4189 * set_dma_reserve - set the specified number of pages reserved in the first zone
4190 * @new_dma_reserve: The number of pages to mark reserved
4192 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4193 * In the DMA zone, a significant percentage may be consumed by kernel image
4194 * and other unfreeable allocations which can skew the watermarks badly. This
4195 * function may optionally be used to account for unfreeable pages in the
4196 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4197 * smaller per-cpu batchsize.
4199 void __init set_dma_reserve(unsigned long new_dma_reserve)
4201 dma_reserve = new_dma_reserve;
4204 #ifndef CONFIG_NEED_MULTIPLE_NODES
4205 struct pglist_data __refdata contig_page_data = { .bdata = &bootmem_node_data[0] };
4206 EXPORT_SYMBOL(contig_page_data);
4209 void __init free_area_init(unsigned long *zones_size)
4211 free_area_init_node(0, zones_size,
4212 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4215 static int page_alloc_cpu_notify(struct notifier_block *self,
4216 unsigned long action, void *hcpu)
4218 int cpu = (unsigned long)hcpu;
4220 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4224 * Spill the event counters of the dead processor
4225 * into the current processors event counters.
4226 * This artificially elevates the count of the current
4229 vm_events_fold_cpu(cpu);
4232 * Zero the differential counters of the dead processor
4233 * so that the vm statistics are consistent.
4235 * This is only okay since the processor is dead and cannot
4236 * race with what we are doing.
4238 refresh_cpu_vm_stats(cpu);
4243 void __init page_alloc_init(void)
4245 hotcpu_notifier(page_alloc_cpu_notify, 0);
4249 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4250 * or min_free_kbytes changes.
4252 static void calculate_totalreserve_pages(void)
4254 struct pglist_data *pgdat;
4255 unsigned long reserve_pages = 0;
4256 enum zone_type i, j;
4258 for_each_online_pgdat(pgdat) {
4259 for (i = 0; i < MAX_NR_ZONES; i++) {
4260 struct zone *zone = pgdat->node_zones + i;
4261 unsigned long max = 0;
4263 /* Find valid and maximum lowmem_reserve in the zone */
4264 for (j = i; j < MAX_NR_ZONES; j++) {
4265 if (zone->lowmem_reserve[j] > max)
4266 max = zone->lowmem_reserve[j];
4269 /* we treat pages_high as reserved pages. */
4270 max += zone->pages_high;
4272 if (max > zone->present_pages)
4273 max = zone->present_pages;
4274 reserve_pages += max;
4277 totalreserve_pages = reserve_pages;
4281 * setup_per_zone_lowmem_reserve - called whenever
4282 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4283 * has a correct pages reserved value, so an adequate number of
4284 * pages are left in the zone after a successful __alloc_pages().
4286 static void setup_per_zone_lowmem_reserve(void)
4288 struct pglist_data *pgdat;
4289 enum zone_type j, idx;
4291 for_each_online_pgdat(pgdat) {
4292 for (j = 0; j < MAX_NR_ZONES; j++) {
4293 struct zone *zone = pgdat->node_zones + j;
4294 unsigned long present_pages = zone->present_pages;
4296 zone->lowmem_reserve[j] = 0;
4300 struct zone *lower_zone;
4304 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4305 sysctl_lowmem_reserve_ratio[idx] = 1;
4307 lower_zone = pgdat->node_zones + idx;
4308 lower_zone->lowmem_reserve[j] = present_pages /
4309 sysctl_lowmem_reserve_ratio[idx];
4310 present_pages += lower_zone->present_pages;
4315 /* update totalreserve_pages */
4316 calculate_totalreserve_pages();
4320 * setup_per_zone_pages_min - called when min_free_kbytes changes.
4322 * Ensures that the pages_{min,low,high} values for each zone are set correctly
4323 * with respect to min_free_kbytes.
4325 void setup_per_zone_pages_min(void)
4327 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4328 unsigned long lowmem_pages = 0;
4330 unsigned long flags;
4332 /* Calculate total number of !ZONE_HIGHMEM pages */
4333 for_each_zone(zone) {
4334 if (!is_highmem(zone))
4335 lowmem_pages += zone->present_pages;
4338 for_each_zone(zone) {
4341 spin_lock_irqsave(&zone->lock, flags);
4342 tmp = (u64)pages_min * zone->present_pages;
4343 do_div(tmp, lowmem_pages);
4344 if (is_highmem(zone)) {
4346 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4347 * need highmem pages, so cap pages_min to a small
4350 * The (pages_high-pages_low) and (pages_low-pages_min)
4351 * deltas controls asynch page reclaim, and so should
4352 * not be capped for highmem.
4356 min_pages = zone->present_pages / 1024;
4357 if (min_pages < SWAP_CLUSTER_MAX)
4358 min_pages = SWAP_CLUSTER_MAX;
4359 if (min_pages > 128)
4361 zone->pages_min = min_pages;
4364 * If it's a lowmem zone, reserve a number of pages
4365 * proportionate to the zone's size.
4367 zone->pages_min = tmp;
4370 zone->pages_low = zone->pages_min + (tmp >> 2);
4371 zone->pages_high = zone->pages_min + (tmp >> 1);
4372 setup_zone_migrate_reserve(zone);
4373 spin_unlock_irqrestore(&zone->lock, flags);
4376 /* update totalreserve_pages */
4377 calculate_totalreserve_pages();
4381 * setup_per_zone_inactive_ratio - called when min_free_kbytes changes.
4383 * The inactive anon list should be small enough that the VM never has to
4384 * do too much work, but large enough that each inactive page has a chance
4385 * to be referenced again before it is swapped out.
4387 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4388 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4389 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4390 * the anonymous pages are kept on the inactive list.
4393 * memory ratio inactive anon
4394 * -------------------------------------
4403 static void setup_per_zone_inactive_ratio(void)
4407 for_each_zone(zone) {
4408 unsigned int gb, ratio;
4410 /* Zone size in gigabytes */
4411 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4412 ratio = int_sqrt(10 * gb);
4416 zone->inactive_ratio = ratio;
4421 * Initialise min_free_kbytes.
4423 * For small machines we want it small (128k min). For large machines
4424 * we want it large (64MB max). But it is not linear, because network
4425 * bandwidth does not increase linearly with machine size. We use
4427 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4428 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4444 static int __init init_per_zone_pages_min(void)
4446 unsigned long lowmem_kbytes;
4448 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4450 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4451 if (min_free_kbytes < 128)
4452 min_free_kbytes = 128;
4453 if (min_free_kbytes > 65536)
4454 min_free_kbytes = 65536;
4455 setup_per_zone_pages_min();
4456 setup_per_zone_lowmem_reserve();
4457 setup_per_zone_inactive_ratio();
4460 module_init(init_per_zone_pages_min)
4463 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4464 * that we can call two helper functions whenever min_free_kbytes
4467 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
4468 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4470 proc_dointvec(table, write, file, buffer, length, ppos);
4472 setup_per_zone_pages_min();
4477 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
4478 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4483 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4488 zone->min_unmapped_pages = (zone->present_pages *
4489 sysctl_min_unmapped_ratio) / 100;
4493 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
4494 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4499 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4504 zone->min_slab_pages = (zone->present_pages *
4505 sysctl_min_slab_ratio) / 100;
4511 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4512 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4513 * whenever sysctl_lowmem_reserve_ratio changes.
4515 * The reserve ratio obviously has absolutely no relation with the
4516 * pages_min watermarks. The lowmem reserve ratio can only make sense
4517 * if in function of the boot time zone sizes.
4519 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
4520 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4522 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4523 setup_per_zone_lowmem_reserve();
4528 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4529 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4530 * can have before it gets flushed back to buddy allocator.
4533 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
4534 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4540 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4541 if (!write || (ret == -EINVAL))
4543 for_each_zone(zone) {
4544 for_each_online_cpu(cpu) {
4546 high = zone->present_pages / percpu_pagelist_fraction;
4547 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
4553 int hashdist = HASHDIST_DEFAULT;
4556 static int __init set_hashdist(char *str)
4560 hashdist = simple_strtoul(str, &str, 0);
4563 __setup("hashdist=", set_hashdist);
4567 * allocate a large system hash table from bootmem
4568 * - it is assumed that the hash table must contain an exact power-of-2
4569 * quantity of entries
4570 * - limit is the number of hash buckets, not the total allocation size
4572 void *__init alloc_large_system_hash(const char *tablename,
4573 unsigned long bucketsize,
4574 unsigned long numentries,
4577 unsigned int *_hash_shift,
4578 unsigned int *_hash_mask,
4579 unsigned long limit)
4581 unsigned long long max = limit;
4582 unsigned long log2qty, size;
4585 /* allow the kernel cmdline to have a say */
4587 /* round applicable memory size up to nearest megabyte */
4588 numentries = nr_kernel_pages;
4589 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
4590 numentries >>= 20 - PAGE_SHIFT;
4591 numentries <<= 20 - PAGE_SHIFT;
4593 /* limit to 1 bucket per 2^scale bytes of low memory */
4594 if (scale > PAGE_SHIFT)
4595 numentries >>= (scale - PAGE_SHIFT);
4597 numentries <<= (PAGE_SHIFT - scale);
4599 /* Make sure we've got at least a 0-order allocation.. */
4600 if (unlikely((numentries * bucketsize) < PAGE_SIZE))
4601 numentries = PAGE_SIZE / bucketsize;
4603 numentries = roundup_pow_of_two(numentries);
4605 /* limit allocation size to 1/16 total memory by default */
4607 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
4608 do_div(max, bucketsize);
4611 if (numentries > max)
4614 log2qty = ilog2(numentries);
4617 size = bucketsize << log2qty;
4618 if (flags & HASH_EARLY)
4619 table = alloc_bootmem_nopanic(size);
4621 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
4623 unsigned long order = get_order(size);
4625 if (order < MAX_ORDER)
4626 table = (void *)__get_free_pages(GFP_ATOMIC,
4629 * If bucketsize is not a power-of-two, we may free
4630 * some pages at the end of hash table.
4633 unsigned long alloc_end = (unsigned long)table +
4634 (PAGE_SIZE << order);
4635 unsigned long used = (unsigned long)table +
4637 split_page(virt_to_page(table), order);
4638 while (used < alloc_end) {
4644 } while (!table && size > PAGE_SIZE && --log2qty);
4647 panic("Failed to allocate %s hash table\n", tablename);
4649 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
4652 ilog2(size) - PAGE_SHIFT,
4656 *_hash_shift = log2qty;
4658 *_hash_mask = (1 << log2qty) - 1;
4661 * If hashdist is set, the table allocation is done with __vmalloc()
4662 * which invokes the kmemleak_alloc() callback. This function may also
4663 * be called before the slab and kmemleak are initialised when
4664 * kmemleak simply buffers the request to be executed later
4665 * (GFP_ATOMIC flag ignored in this case).
4668 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
4673 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4674 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
4677 #ifdef CONFIG_SPARSEMEM
4678 return __pfn_to_section(pfn)->pageblock_flags;
4680 return zone->pageblock_flags;
4681 #endif /* CONFIG_SPARSEMEM */
4684 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
4686 #ifdef CONFIG_SPARSEMEM
4687 pfn &= (PAGES_PER_SECTION-1);
4688 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4690 pfn = pfn - zone->zone_start_pfn;
4691 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4692 #endif /* CONFIG_SPARSEMEM */
4696 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4697 * @page: The page within the block of interest
4698 * @start_bitidx: The first bit of interest to retrieve
4699 * @end_bitidx: The last bit of interest
4700 * returns pageblock_bits flags
4702 unsigned long get_pageblock_flags_group(struct page *page,
4703 int start_bitidx, int end_bitidx)
4706 unsigned long *bitmap;
4707 unsigned long pfn, bitidx;
4708 unsigned long flags = 0;
4709 unsigned long value = 1;
4711 zone = page_zone(page);
4712 pfn = page_to_pfn(page);
4713 bitmap = get_pageblock_bitmap(zone, pfn);
4714 bitidx = pfn_to_bitidx(zone, pfn);
4716 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4717 if (test_bit(bitidx + start_bitidx, bitmap))
4724 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4725 * @page: The page within the block of interest
4726 * @start_bitidx: The first bit of interest
4727 * @end_bitidx: The last bit of interest
4728 * @flags: The flags to set
4730 void set_pageblock_flags_group(struct page *page, unsigned long flags,
4731 int start_bitidx, int end_bitidx)
4734 unsigned long *bitmap;
4735 unsigned long pfn, bitidx;
4736 unsigned long value = 1;
4738 zone = page_zone(page);
4739 pfn = page_to_pfn(page);
4740 bitmap = get_pageblock_bitmap(zone, pfn);
4741 bitidx = pfn_to_bitidx(zone, pfn);
4742 VM_BUG_ON(pfn < zone->zone_start_pfn);
4743 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
4745 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4747 __set_bit(bitidx + start_bitidx, bitmap);
4749 __clear_bit(bitidx + start_bitidx, bitmap);
4753 * This is designed as sub function...plz see page_isolation.c also.
4754 * set/clear page block's type to be ISOLATE.
4755 * page allocater never alloc memory from ISOLATE block.
4758 int set_migratetype_isolate(struct page *page)
4761 unsigned long flags;
4764 zone = page_zone(page);
4765 spin_lock_irqsave(&zone->lock, flags);
4767 * In future, more migrate types will be able to be isolation target.
4769 if (get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
4771 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
4772 move_freepages_block(zone, page, MIGRATE_ISOLATE);
4775 spin_unlock_irqrestore(&zone->lock, flags);
4781 void unset_migratetype_isolate(struct page *page)
4784 unsigned long flags;
4785 zone = page_zone(page);
4786 spin_lock_irqsave(&zone->lock, flags);
4787 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
4789 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4790 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4792 spin_unlock_irqrestore(&zone->lock, flags);
4795 #ifdef CONFIG_MEMORY_HOTREMOVE
4797 * All pages in the range must be isolated before calling this.
4800 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
4806 unsigned long flags;
4807 /* find the first valid pfn */
4808 for (pfn = start_pfn; pfn < end_pfn; pfn++)
4813 zone = page_zone(pfn_to_page(pfn));
4814 spin_lock_irqsave(&zone->lock, flags);
4816 while (pfn < end_pfn) {
4817 if (!pfn_valid(pfn)) {
4821 page = pfn_to_page(pfn);
4822 BUG_ON(page_count(page));
4823 BUG_ON(!PageBuddy(page));
4824 order = page_order(page);
4825 #ifdef CONFIG_DEBUG_VM
4826 printk(KERN_INFO "remove from free list %lx %d %lx\n",
4827 pfn, 1 << order, end_pfn);
4829 list_del(&page->lru);
4830 rmv_page_order(page);
4831 zone->free_area[order].nr_free--;
4832 __mod_zone_page_state(zone, NR_FREE_PAGES,
4834 for (i = 0; i < (1 << order); i++)
4835 SetPageReserved((page+i));
4836 pfn += (1 << order);
4838 spin_unlock_irqrestore(&zone->lock, flags);