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>
50 #include <asm/tlbflush.h>
51 #include <asm/div64.h>
55 * Array of node states.
57 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
58 [N_POSSIBLE] = NODE_MASK_ALL,
59 [N_ONLINE] = { { [0] = 1UL } },
61 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
63 [N_HIGH_MEMORY] = { { [0] = 1UL } },
65 [N_CPU] = { { [0] = 1UL } },
68 EXPORT_SYMBOL(node_states);
70 unsigned long totalram_pages __read_mostly;
71 unsigned long totalreserve_pages __read_mostly;
72 unsigned long highest_memmap_pfn __read_mostly;
73 int percpu_pagelist_fraction;
75 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
76 int pageblock_order __read_mostly;
79 static void __free_pages_ok(struct page *page, unsigned int order);
82 * results with 256, 32 in the lowmem_reserve sysctl:
83 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
84 * 1G machine -> (16M dma, 784M normal, 224M high)
85 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
86 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
87 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
89 * TBD: should special case ZONE_DMA32 machines here - in those we normally
90 * don't need any ZONE_NORMAL reservation
92 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
93 #ifdef CONFIG_ZONE_DMA
96 #ifdef CONFIG_ZONE_DMA32
105 EXPORT_SYMBOL(totalram_pages);
107 static char * const zone_names[MAX_NR_ZONES] = {
108 #ifdef CONFIG_ZONE_DMA
111 #ifdef CONFIG_ZONE_DMA32
115 #ifdef CONFIG_HIGHMEM
121 int min_free_kbytes = 1024;
123 unsigned long __meminitdata nr_kernel_pages;
124 unsigned long __meminitdata nr_all_pages;
125 static unsigned long __meminitdata dma_reserve;
127 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
129 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
130 * ranges of memory (RAM) that may be registered with add_active_range().
131 * Ranges passed to add_active_range() will be merged if possible
132 * so the number of times add_active_range() can be called is
133 * related to the number of nodes and the number of holes
135 #ifdef CONFIG_MAX_ACTIVE_REGIONS
136 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
137 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
139 #if MAX_NUMNODES >= 32
140 /* If there can be many nodes, allow up to 50 holes per node */
141 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
143 /* By default, allow up to 256 distinct regions */
144 #define MAX_ACTIVE_REGIONS 256
148 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
149 static int __meminitdata nr_nodemap_entries;
150 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
151 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
152 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
153 static unsigned long __meminitdata node_boundary_start_pfn[MAX_NUMNODES];
154 static unsigned long __meminitdata node_boundary_end_pfn[MAX_NUMNODES];
155 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
156 static unsigned long __initdata required_kernelcore;
157 static unsigned long __initdata required_movablecore;
158 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
160 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
162 EXPORT_SYMBOL(movable_zone);
163 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
166 int nr_node_ids __read_mostly = MAX_NUMNODES;
167 EXPORT_SYMBOL(nr_node_ids);
170 int page_group_by_mobility_disabled __read_mostly;
172 static void set_pageblock_migratetype(struct page *page, int migratetype)
174 set_pageblock_flags_group(page, (unsigned long)migratetype,
175 PB_migrate, PB_migrate_end);
178 #ifdef CONFIG_DEBUG_VM
179 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
183 unsigned long pfn = page_to_pfn(page);
186 seq = zone_span_seqbegin(zone);
187 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
189 else if (pfn < zone->zone_start_pfn)
191 } while (zone_span_seqretry(zone, seq));
196 static int page_is_consistent(struct zone *zone, struct page *page)
198 if (!pfn_valid_within(page_to_pfn(page)))
200 if (zone != page_zone(page))
206 * Temporary debugging check for pages not lying within a given zone.
208 static int bad_range(struct zone *zone, struct page *page)
210 if (page_outside_zone_boundaries(zone, page))
212 if (!page_is_consistent(zone, page))
218 static inline int bad_range(struct zone *zone, struct page *page)
224 static void bad_page(struct page *page)
226 static unsigned long resume;
227 static unsigned long nr_shown;
228 static unsigned long nr_unshown;
231 * Allow a burst of 60 reports, then keep quiet for that minute;
232 * or allow a steady drip of one report per second.
234 if (nr_shown == 60) {
235 if (time_before(jiffies, resume)) {
241 "BUG: Bad page state: %lu messages suppressed\n",
248 resume = jiffies + 60 * HZ;
250 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
251 current->comm, page_to_pfn(page));
253 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
254 page, (void *)page->flags, page_count(page),
255 page_mapcount(page), page->mapping, page->index);
259 /* Leave bad fields for debug, except PageBuddy could make trouble */
260 __ClearPageBuddy(page);
261 add_taint(TAINT_BAD_PAGE);
265 * Higher-order pages are called "compound pages". They are structured thusly:
267 * The first PAGE_SIZE page is called the "head page".
269 * The remaining PAGE_SIZE pages are called "tail pages".
271 * All pages have PG_compound set. All pages have their ->private pointing at
272 * the head page (even the head page has this).
274 * The first tail page's ->lru.next holds the address of the compound page's
275 * put_page() function. Its ->lru.prev holds the order of allocation.
276 * This usage means that zero-order pages may not be compound.
279 static void free_compound_page(struct page *page)
281 __free_pages_ok(page, compound_order(page));
284 void prep_compound_page(struct page *page, unsigned long order)
287 int nr_pages = 1 << order;
289 set_compound_page_dtor(page, free_compound_page);
290 set_compound_order(page, order);
292 for (i = 1; i < nr_pages; i++) {
293 struct page *p = page + i;
296 p->first_page = page;
300 #ifdef CONFIG_HUGETLBFS
301 void prep_compound_gigantic_page(struct page *page, unsigned long order)
304 int nr_pages = 1 << order;
305 struct page *p = page + 1;
307 set_compound_page_dtor(page, free_compound_page);
308 set_compound_order(page, order);
310 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
312 p->first_page = page;
317 static int destroy_compound_page(struct page *page, unsigned long order)
320 int nr_pages = 1 << order;
323 if (unlikely(compound_order(page) != order) ||
324 unlikely(!PageHead(page))) {
329 __ClearPageHead(page);
331 for (i = 1; i < nr_pages; i++) {
332 struct page *p = page + i;
334 if (unlikely(!PageTail(p) | (p->first_page != page))) {
344 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
349 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
350 * and __GFP_HIGHMEM from hard or soft interrupt context.
352 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
353 for (i = 0; i < (1 << order); i++)
354 clear_highpage(page + i);
357 static inline void set_page_order(struct page *page, int order)
359 set_page_private(page, order);
360 __SetPageBuddy(page);
363 static inline void rmv_page_order(struct page *page)
365 __ClearPageBuddy(page);
366 set_page_private(page, 0);
370 * Locate the struct page for both the matching buddy in our
371 * pair (buddy1) and the combined O(n+1) page they form (page).
373 * 1) Any buddy B1 will have an order O twin B2 which satisfies
374 * the following equation:
376 * For example, if the starting buddy (buddy2) is #8 its order
378 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
380 * 2) Any buddy B will have an order O+1 parent P which
381 * satisfies the following equation:
384 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
386 static inline struct page *
387 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
389 unsigned long buddy_idx = page_idx ^ (1 << order);
391 return page + (buddy_idx - page_idx);
394 static inline unsigned long
395 __find_combined_index(unsigned long page_idx, unsigned int order)
397 return (page_idx & ~(1 << order));
401 * This function checks whether a page is free && is the buddy
402 * we can do coalesce a page and its buddy if
403 * (a) the buddy is not in a hole &&
404 * (b) the buddy is in the buddy system &&
405 * (c) a page and its buddy have the same order &&
406 * (d) a page and its buddy are in the same zone.
408 * For recording whether a page is in the buddy system, we use PG_buddy.
409 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
411 * For recording page's order, we use page_private(page).
413 static inline int page_is_buddy(struct page *page, struct page *buddy,
416 if (!pfn_valid_within(page_to_pfn(buddy)))
419 if (page_zone_id(page) != page_zone_id(buddy))
422 if (PageBuddy(buddy) && page_order(buddy) == order) {
423 BUG_ON(page_count(buddy) != 0);
430 * Freeing function for a buddy system allocator.
432 * The concept of a buddy system is to maintain direct-mapped table
433 * (containing bit values) for memory blocks of various "orders".
434 * The bottom level table contains the map for the smallest allocatable
435 * units of memory (here, pages), and each level above it describes
436 * pairs of units from the levels below, hence, "buddies".
437 * At a high level, all that happens here is marking the table entry
438 * at the bottom level available, and propagating the changes upward
439 * as necessary, plus some accounting needed to play nicely with other
440 * parts of the VM system.
441 * At each level, we keep a list of pages, which are heads of continuous
442 * free pages of length of (1 << order) and marked with PG_buddy. Page's
443 * order is recorded in page_private(page) field.
444 * So when we are allocating or freeing one, we can derive the state of the
445 * other. That is, if we allocate a small block, and both were
446 * free, the remainder of the region must be split into blocks.
447 * If a block is freed, and its buddy is also free, then this
448 * triggers coalescing into a block of larger size.
453 static inline void __free_one_page(struct page *page,
454 struct zone *zone, unsigned int order)
456 unsigned long page_idx;
457 int order_size = 1 << order;
458 int migratetype = get_pageblock_migratetype(page);
460 if (unlikely(PageCompound(page)))
461 if (unlikely(destroy_compound_page(page, order)))
464 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
466 VM_BUG_ON(page_idx & (order_size - 1));
467 VM_BUG_ON(bad_range(zone, page));
469 __mod_zone_page_state(zone, NR_FREE_PAGES, order_size);
470 while (order < MAX_ORDER-1) {
471 unsigned long combined_idx;
474 buddy = __page_find_buddy(page, page_idx, order);
475 if (!page_is_buddy(page, buddy, order))
478 /* Our buddy is free, merge with it and move up one order. */
479 list_del(&buddy->lru);
480 zone->free_area[order].nr_free--;
481 rmv_page_order(buddy);
482 combined_idx = __find_combined_index(page_idx, order);
483 page = page + (combined_idx - page_idx);
484 page_idx = combined_idx;
487 set_page_order(page, order);
489 &zone->free_area[order].free_list[migratetype]);
490 zone->free_area[order].nr_free++;
493 static inline int free_pages_check(struct page *page)
495 free_page_mlock(page);
496 if (unlikely(page_mapcount(page) |
497 (page->mapping != NULL) |
498 (page_count(page) != 0) |
499 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) {
503 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
504 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
509 * Frees a list of pages.
510 * Assumes all pages on list are in same zone, and of same order.
511 * count is the number of pages to free.
513 * If the zone was previously in an "all pages pinned" state then look to
514 * see if this freeing clears that state.
516 * And clear the zone's pages_scanned counter, to hold off the "all pages are
517 * pinned" detection logic.
519 static void free_pages_bulk(struct zone *zone, int count,
520 struct list_head *list, int order)
522 spin_lock(&zone->lock);
523 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
524 zone->pages_scanned = 0;
528 VM_BUG_ON(list_empty(list));
529 page = list_entry(list->prev, struct page, lru);
530 /* have to delete it as __free_one_page list manipulates */
531 list_del(&page->lru);
532 __free_one_page(page, zone, order);
534 spin_unlock(&zone->lock);
537 static void free_one_page(struct zone *zone, struct page *page, int order)
539 spin_lock(&zone->lock);
540 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
541 zone->pages_scanned = 0;
542 __free_one_page(page, zone, order);
543 spin_unlock(&zone->lock);
546 static void __free_pages_ok(struct page *page, unsigned int order)
552 for (i = 0 ; i < (1 << order) ; ++i)
553 bad += free_pages_check(page + i);
557 if (!PageHighMem(page)) {
558 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
559 debug_check_no_obj_freed(page_address(page),
562 arch_free_page(page, order);
563 kernel_map_pages(page, 1 << order, 0);
565 local_irq_save(flags);
566 __count_vm_events(PGFREE, 1 << order);
567 free_one_page(page_zone(page), page, order);
568 local_irq_restore(flags);
572 * permit the bootmem allocator to evade page validation on high-order frees
574 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
577 __ClearPageReserved(page);
578 set_page_count(page, 0);
579 set_page_refcounted(page);
585 for (loop = 0; loop < BITS_PER_LONG; loop++) {
586 struct page *p = &page[loop];
588 if (loop + 1 < BITS_PER_LONG)
590 __ClearPageReserved(p);
591 set_page_count(p, 0);
594 set_page_refcounted(page);
595 __free_pages(page, order);
601 * The order of subdivision here is critical for the IO subsystem.
602 * Please do not alter this order without good reasons and regression
603 * testing. Specifically, as large blocks of memory are subdivided,
604 * the order in which smaller blocks are delivered depends on the order
605 * they're subdivided in this function. This is the primary factor
606 * influencing the order in which pages are delivered to the IO
607 * subsystem according to empirical testing, and this is also justified
608 * by considering the behavior of a buddy system containing a single
609 * large block of memory acted on by a series of small allocations.
610 * This behavior is a critical factor in sglist merging's success.
614 static inline void expand(struct zone *zone, struct page *page,
615 int low, int high, struct free_area *area,
618 unsigned long size = 1 << high;
624 VM_BUG_ON(bad_range(zone, &page[size]));
625 list_add(&page[size].lru, &area->free_list[migratetype]);
627 set_page_order(&page[size], high);
632 * This page is about to be returned from the page allocator
634 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
636 if (unlikely(page_mapcount(page) |
637 (page->mapping != NULL) |
638 (page_count(page) != 0) |
639 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) {
644 set_page_private(page, 0);
645 set_page_refcounted(page);
647 arch_alloc_page(page, order);
648 kernel_map_pages(page, 1 << order, 1);
650 if (gfp_flags & __GFP_ZERO)
651 prep_zero_page(page, order, gfp_flags);
653 if (order && (gfp_flags & __GFP_COMP))
654 prep_compound_page(page, order);
660 * Go through the free lists for the given migratetype and remove
661 * the smallest available page from the freelists
663 static struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
666 unsigned int current_order;
667 struct free_area * area;
670 /* Find a page of the appropriate size in the preferred list */
671 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
672 area = &(zone->free_area[current_order]);
673 if (list_empty(&area->free_list[migratetype]))
676 page = list_entry(area->free_list[migratetype].next,
678 list_del(&page->lru);
679 rmv_page_order(page);
681 __mod_zone_page_state(zone, NR_FREE_PAGES, - (1UL << order));
682 expand(zone, page, order, current_order, area, migratetype);
691 * This array describes the order lists are fallen back to when
692 * the free lists for the desirable migrate type are depleted
694 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
695 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
696 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
697 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
698 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
702 * Move the free pages in a range to the free lists of the requested type.
703 * Note that start_page and end_pages are not aligned on a pageblock
704 * boundary. If alignment is required, use move_freepages_block()
706 static int move_freepages(struct zone *zone,
707 struct page *start_page, struct page *end_page,
714 #ifndef CONFIG_HOLES_IN_ZONE
716 * page_zone is not safe to call in this context when
717 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
718 * anyway as we check zone boundaries in move_freepages_block().
719 * Remove at a later date when no bug reports exist related to
720 * grouping pages by mobility
722 BUG_ON(page_zone(start_page) != page_zone(end_page));
725 for (page = start_page; page <= end_page;) {
726 /* Make sure we are not inadvertently changing nodes */
727 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
729 if (!pfn_valid_within(page_to_pfn(page))) {
734 if (!PageBuddy(page)) {
739 order = page_order(page);
740 list_del(&page->lru);
742 &zone->free_area[order].free_list[migratetype]);
744 pages_moved += 1 << order;
750 static int move_freepages_block(struct zone *zone, struct page *page,
753 unsigned long start_pfn, end_pfn;
754 struct page *start_page, *end_page;
756 start_pfn = page_to_pfn(page);
757 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
758 start_page = pfn_to_page(start_pfn);
759 end_page = start_page + pageblock_nr_pages - 1;
760 end_pfn = start_pfn + pageblock_nr_pages - 1;
762 /* Do not cross zone boundaries */
763 if (start_pfn < zone->zone_start_pfn)
765 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
768 return move_freepages(zone, start_page, end_page, migratetype);
771 /* Remove an element from the buddy allocator from the fallback list */
772 static struct page *__rmqueue_fallback(struct zone *zone, int order,
773 int start_migratetype)
775 struct free_area * area;
780 /* Find the largest possible block of pages in the other list */
781 for (current_order = MAX_ORDER-1; current_order >= order;
783 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
784 migratetype = fallbacks[start_migratetype][i];
786 /* MIGRATE_RESERVE handled later if necessary */
787 if (migratetype == MIGRATE_RESERVE)
790 area = &(zone->free_area[current_order]);
791 if (list_empty(&area->free_list[migratetype]))
794 page = list_entry(area->free_list[migratetype].next,
799 * If breaking a large block of pages, move all free
800 * pages to the preferred allocation list. If falling
801 * back for a reclaimable kernel allocation, be more
802 * agressive about taking ownership of free pages
804 if (unlikely(current_order >= (pageblock_order >> 1)) ||
805 start_migratetype == MIGRATE_RECLAIMABLE) {
807 pages = move_freepages_block(zone, page,
810 /* Claim the whole block if over half of it is free */
811 if (pages >= (1 << (pageblock_order-1)))
812 set_pageblock_migratetype(page,
815 migratetype = start_migratetype;
818 /* Remove the page from the freelists */
819 list_del(&page->lru);
820 rmv_page_order(page);
821 __mod_zone_page_state(zone, NR_FREE_PAGES,
824 if (current_order == pageblock_order)
825 set_pageblock_migratetype(page,
828 expand(zone, page, order, current_order, area, migratetype);
833 /* Use MIGRATE_RESERVE rather than fail an allocation */
834 return __rmqueue_smallest(zone, order, MIGRATE_RESERVE);
838 * Do the hard work of removing an element from the buddy allocator.
839 * Call me with the zone->lock already held.
841 static struct page *__rmqueue(struct zone *zone, unsigned int order,
846 page = __rmqueue_smallest(zone, order, migratetype);
849 page = __rmqueue_fallback(zone, order, migratetype);
855 * Obtain a specified number of elements from the buddy allocator, all under
856 * a single hold of the lock, for efficiency. Add them to the supplied list.
857 * Returns the number of new pages which were placed at *list.
859 static int rmqueue_bulk(struct zone *zone, unsigned int order,
860 unsigned long count, struct list_head *list,
865 spin_lock(&zone->lock);
866 for (i = 0; i < count; ++i) {
867 struct page *page = __rmqueue(zone, order, migratetype);
868 if (unlikely(page == NULL))
872 * Split buddy pages returned by expand() are received here
873 * in physical page order. The page is added to the callers and
874 * list and the list head then moves forward. From the callers
875 * perspective, the linked list is ordered by page number in
876 * some conditions. This is useful for IO devices that can
877 * merge IO requests if the physical pages are ordered
880 list_add(&page->lru, list);
881 set_page_private(page, migratetype);
884 spin_unlock(&zone->lock);
890 * Called from the vmstat counter updater to drain pagesets of this
891 * currently executing processor on remote nodes after they have
894 * Note that this function must be called with the thread pinned to
895 * a single processor.
897 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
902 local_irq_save(flags);
903 if (pcp->count >= pcp->batch)
904 to_drain = pcp->batch;
906 to_drain = pcp->count;
907 free_pages_bulk(zone, to_drain, &pcp->list, 0);
908 pcp->count -= to_drain;
909 local_irq_restore(flags);
914 * Drain pages of the indicated processor.
916 * The processor must either be the current processor and the
917 * thread pinned to the current processor or a processor that
920 static void drain_pages(unsigned int cpu)
925 for_each_zone(zone) {
926 struct per_cpu_pageset *pset;
927 struct per_cpu_pages *pcp;
929 if (!populated_zone(zone))
932 pset = zone_pcp(zone, cpu);
935 local_irq_save(flags);
936 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
938 local_irq_restore(flags);
943 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
945 void drain_local_pages(void *arg)
947 drain_pages(smp_processor_id());
951 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
953 void drain_all_pages(void)
955 on_each_cpu(drain_local_pages, NULL, 1);
958 #ifdef CONFIG_HIBERNATION
960 void mark_free_pages(struct zone *zone)
962 unsigned long pfn, max_zone_pfn;
965 struct list_head *curr;
967 if (!zone->spanned_pages)
970 spin_lock_irqsave(&zone->lock, flags);
972 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
973 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
974 if (pfn_valid(pfn)) {
975 struct page *page = pfn_to_page(pfn);
977 if (!swsusp_page_is_forbidden(page))
978 swsusp_unset_page_free(page);
981 for_each_migratetype_order(order, t) {
982 list_for_each(curr, &zone->free_area[order].free_list[t]) {
985 pfn = page_to_pfn(list_entry(curr, struct page, lru));
986 for (i = 0; i < (1UL << order); i++)
987 swsusp_set_page_free(pfn_to_page(pfn + i));
990 spin_unlock_irqrestore(&zone->lock, flags);
992 #endif /* CONFIG_PM */
995 * Free a 0-order page
997 static void free_hot_cold_page(struct page *page, int cold)
999 struct zone *zone = page_zone(page);
1000 struct per_cpu_pages *pcp;
1001 unsigned long flags;
1004 page->mapping = NULL;
1005 if (free_pages_check(page))
1008 if (!PageHighMem(page)) {
1009 debug_check_no_locks_freed(page_address(page), PAGE_SIZE);
1010 debug_check_no_obj_freed(page_address(page), PAGE_SIZE);
1012 arch_free_page(page, 0);
1013 kernel_map_pages(page, 1, 0);
1015 pcp = &zone_pcp(zone, get_cpu())->pcp;
1016 local_irq_save(flags);
1017 __count_vm_event(PGFREE);
1019 list_add_tail(&page->lru, &pcp->list);
1021 list_add(&page->lru, &pcp->list);
1022 set_page_private(page, get_pageblock_migratetype(page));
1024 if (pcp->count >= pcp->high) {
1025 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
1026 pcp->count -= pcp->batch;
1028 local_irq_restore(flags);
1032 void free_hot_page(struct page *page)
1034 free_hot_cold_page(page, 0);
1037 void free_cold_page(struct page *page)
1039 free_hot_cold_page(page, 1);
1043 * split_page takes a non-compound higher-order page, and splits it into
1044 * n (1<<order) sub-pages: page[0..n]
1045 * Each sub-page must be freed individually.
1047 * Note: this is probably too low level an operation for use in drivers.
1048 * Please consult with lkml before using this in your driver.
1050 void split_page(struct page *page, unsigned int order)
1054 VM_BUG_ON(PageCompound(page));
1055 VM_BUG_ON(!page_count(page));
1056 for (i = 1; i < (1 << order); i++)
1057 set_page_refcounted(page + i);
1061 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1062 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1065 static struct page *buffered_rmqueue(struct zone *preferred_zone,
1066 struct zone *zone, int order, gfp_t gfp_flags)
1068 unsigned long flags;
1070 int cold = !!(gfp_flags & __GFP_COLD);
1072 int migratetype = allocflags_to_migratetype(gfp_flags);
1076 if (likely(order == 0)) {
1077 struct per_cpu_pages *pcp;
1079 pcp = &zone_pcp(zone, cpu)->pcp;
1080 local_irq_save(flags);
1082 pcp->count = rmqueue_bulk(zone, 0,
1083 pcp->batch, &pcp->list, migratetype);
1084 if (unlikely(!pcp->count))
1088 /* Find a page of the appropriate migrate type */
1090 list_for_each_entry_reverse(page, &pcp->list, lru)
1091 if (page_private(page) == migratetype)
1094 list_for_each_entry(page, &pcp->list, lru)
1095 if (page_private(page) == migratetype)
1099 /* Allocate more to the pcp list if necessary */
1100 if (unlikely(&page->lru == &pcp->list)) {
1101 pcp->count += rmqueue_bulk(zone, 0,
1102 pcp->batch, &pcp->list, migratetype);
1103 page = list_entry(pcp->list.next, struct page, lru);
1106 list_del(&page->lru);
1109 spin_lock_irqsave(&zone->lock, flags);
1110 page = __rmqueue(zone, order, migratetype);
1111 spin_unlock(&zone->lock);
1116 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1117 zone_statistics(preferred_zone, zone);
1118 local_irq_restore(flags);
1121 VM_BUG_ON(bad_range(zone, page));
1122 if (prep_new_page(page, order, gfp_flags))
1127 local_irq_restore(flags);
1132 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
1133 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
1134 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
1135 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
1136 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1137 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1138 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1140 #ifdef CONFIG_FAIL_PAGE_ALLOC
1142 static struct fail_page_alloc_attr {
1143 struct fault_attr attr;
1145 u32 ignore_gfp_highmem;
1146 u32 ignore_gfp_wait;
1149 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1151 struct dentry *ignore_gfp_highmem_file;
1152 struct dentry *ignore_gfp_wait_file;
1153 struct dentry *min_order_file;
1155 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1157 } fail_page_alloc = {
1158 .attr = FAULT_ATTR_INITIALIZER,
1159 .ignore_gfp_wait = 1,
1160 .ignore_gfp_highmem = 1,
1164 static int __init setup_fail_page_alloc(char *str)
1166 return setup_fault_attr(&fail_page_alloc.attr, str);
1168 __setup("fail_page_alloc=", setup_fail_page_alloc);
1170 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1172 if (order < fail_page_alloc.min_order)
1174 if (gfp_mask & __GFP_NOFAIL)
1176 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1178 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1181 return should_fail(&fail_page_alloc.attr, 1 << order);
1184 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1186 static int __init fail_page_alloc_debugfs(void)
1188 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1192 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1196 dir = fail_page_alloc.attr.dentries.dir;
1198 fail_page_alloc.ignore_gfp_wait_file =
1199 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1200 &fail_page_alloc.ignore_gfp_wait);
1202 fail_page_alloc.ignore_gfp_highmem_file =
1203 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1204 &fail_page_alloc.ignore_gfp_highmem);
1205 fail_page_alloc.min_order_file =
1206 debugfs_create_u32("min-order", mode, dir,
1207 &fail_page_alloc.min_order);
1209 if (!fail_page_alloc.ignore_gfp_wait_file ||
1210 !fail_page_alloc.ignore_gfp_highmem_file ||
1211 !fail_page_alloc.min_order_file) {
1213 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1214 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1215 debugfs_remove(fail_page_alloc.min_order_file);
1216 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1222 late_initcall(fail_page_alloc_debugfs);
1224 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1226 #else /* CONFIG_FAIL_PAGE_ALLOC */
1228 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1233 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1236 * Return 1 if free pages are above 'mark'. This takes into account the order
1237 * of the allocation.
1239 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1240 int classzone_idx, int alloc_flags)
1242 /* free_pages my go negative - that's OK */
1244 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
1247 if (alloc_flags & ALLOC_HIGH)
1249 if (alloc_flags & ALLOC_HARDER)
1252 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1254 for (o = 0; o < order; o++) {
1255 /* At the next order, this order's pages become unavailable */
1256 free_pages -= z->free_area[o].nr_free << o;
1258 /* Require fewer higher order pages to be free */
1261 if (free_pages <= min)
1269 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1270 * skip over zones that are not allowed by the cpuset, or that have
1271 * been recently (in last second) found to be nearly full. See further
1272 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1273 * that have to skip over a lot of full or unallowed zones.
1275 * If the zonelist cache is present in the passed in zonelist, then
1276 * returns a pointer to the allowed node mask (either the current
1277 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1279 * If the zonelist cache is not available for this zonelist, does
1280 * nothing and returns NULL.
1282 * If the fullzones BITMAP in the zonelist cache is stale (more than
1283 * a second since last zap'd) then we zap it out (clear its bits.)
1285 * We hold off even calling zlc_setup, until after we've checked the
1286 * first zone in the zonelist, on the theory that most allocations will
1287 * be satisfied from that first zone, so best to examine that zone as
1288 * quickly as we can.
1290 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1292 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1293 nodemask_t *allowednodes; /* zonelist_cache approximation */
1295 zlc = zonelist->zlcache_ptr;
1299 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1300 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1301 zlc->last_full_zap = jiffies;
1304 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1305 &cpuset_current_mems_allowed :
1306 &node_states[N_HIGH_MEMORY];
1307 return allowednodes;
1311 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1312 * if it is worth looking at further for free memory:
1313 * 1) Check that the zone isn't thought to be full (doesn't have its
1314 * bit set in the zonelist_cache fullzones BITMAP).
1315 * 2) Check that the zones node (obtained from the zonelist_cache
1316 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1317 * Return true (non-zero) if zone is worth looking at further, or
1318 * else return false (zero) if it is not.
1320 * This check -ignores- the distinction between various watermarks,
1321 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1322 * found to be full for any variation of these watermarks, it will
1323 * be considered full for up to one second by all requests, unless
1324 * we are so low on memory on all allowed nodes that we are forced
1325 * into the second scan of the zonelist.
1327 * In the second scan we ignore this zonelist cache and exactly
1328 * apply the watermarks to all zones, even it is slower to do so.
1329 * We are low on memory in the second scan, and should leave no stone
1330 * unturned looking for a free page.
1332 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1333 nodemask_t *allowednodes)
1335 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1336 int i; /* index of *z in zonelist zones */
1337 int n; /* node that zone *z is on */
1339 zlc = zonelist->zlcache_ptr;
1343 i = z - zonelist->_zonerefs;
1346 /* This zone is worth trying if it is allowed but not full */
1347 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1351 * Given 'z' scanning a zonelist, set the corresponding bit in
1352 * zlc->fullzones, so that subsequent attempts to allocate a page
1353 * from that zone don't waste time re-examining it.
1355 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1357 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1358 int i; /* index of *z in zonelist zones */
1360 zlc = zonelist->zlcache_ptr;
1364 i = z - zonelist->_zonerefs;
1366 set_bit(i, zlc->fullzones);
1369 #else /* CONFIG_NUMA */
1371 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1376 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1377 nodemask_t *allowednodes)
1382 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1385 #endif /* CONFIG_NUMA */
1388 * get_page_from_freelist goes through the zonelist trying to allocate
1391 static struct page *
1392 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1393 struct zonelist *zonelist, int high_zoneidx, int alloc_flags)
1396 struct page *page = NULL;
1398 struct zone *zone, *preferred_zone;
1399 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1400 int zlc_active = 0; /* set if using zonelist_cache */
1401 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1403 (void)first_zones_zonelist(zonelist, high_zoneidx, nodemask,
1405 if (!preferred_zone)
1408 classzone_idx = zone_idx(preferred_zone);
1412 * Scan zonelist, looking for a zone with enough free.
1413 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1415 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1416 high_zoneidx, nodemask) {
1417 if (NUMA_BUILD && zlc_active &&
1418 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1420 if ((alloc_flags & ALLOC_CPUSET) &&
1421 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1424 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1426 if (alloc_flags & ALLOC_WMARK_MIN)
1427 mark = zone->pages_min;
1428 else if (alloc_flags & ALLOC_WMARK_LOW)
1429 mark = zone->pages_low;
1431 mark = zone->pages_high;
1432 if (!zone_watermark_ok(zone, order, mark,
1433 classzone_idx, alloc_flags)) {
1434 if (!zone_reclaim_mode ||
1435 !zone_reclaim(zone, gfp_mask, order))
1436 goto this_zone_full;
1440 page = buffered_rmqueue(preferred_zone, zone, order, gfp_mask);
1445 zlc_mark_zone_full(zonelist, z);
1447 if (NUMA_BUILD && !did_zlc_setup) {
1448 /* we do zlc_setup after the first zone is tried */
1449 allowednodes = zlc_setup(zonelist, alloc_flags);
1455 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1456 /* Disable zlc cache for second zonelist scan */
1464 * This is the 'heart' of the zoned buddy allocator.
1467 __alloc_pages_internal(gfp_t gfp_mask, unsigned int order,
1468 struct zonelist *zonelist, nodemask_t *nodemask)
1470 const gfp_t wait = gfp_mask & __GFP_WAIT;
1471 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1475 struct reclaim_state reclaim_state;
1476 struct task_struct *p = current;
1479 unsigned long did_some_progress;
1480 unsigned long pages_reclaimed = 0;
1482 might_sleep_if(wait);
1484 if (should_fail_alloc_page(gfp_mask, order))
1488 z = zonelist->_zonerefs; /* the list of zones suitable for gfp_mask */
1490 if (unlikely(!z->zone)) {
1492 * Happens if we have an empty zonelist as a result of
1493 * GFP_THISNODE being used on a memoryless node
1498 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
1499 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET);
1504 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1505 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1506 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1507 * using a larger set of nodes after it has established that the
1508 * allowed per node queues are empty and that nodes are
1511 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1514 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1515 wakeup_kswapd(zone, order);
1518 * OK, we're below the kswapd watermark and have kicked background
1519 * reclaim. Now things get more complex, so set up alloc_flags according
1520 * to how we want to proceed.
1522 * The caller may dip into page reserves a bit more if the caller
1523 * cannot run direct reclaim, or if the caller has realtime scheduling
1524 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1525 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1527 alloc_flags = ALLOC_WMARK_MIN;
1528 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
1529 alloc_flags |= ALLOC_HARDER;
1530 if (gfp_mask & __GFP_HIGH)
1531 alloc_flags |= ALLOC_HIGH;
1533 alloc_flags |= ALLOC_CPUSET;
1536 * Go through the zonelist again. Let __GFP_HIGH and allocations
1537 * coming from realtime tasks go deeper into reserves.
1539 * This is the last chance, in general, before the goto nopage.
1540 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1541 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1543 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
1544 high_zoneidx, alloc_flags);
1548 /* This allocation should allow future memory freeing. */
1551 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
1552 && !in_interrupt()) {
1553 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
1555 /* go through the zonelist yet again, ignoring mins */
1556 page = get_page_from_freelist(gfp_mask, nodemask, order,
1557 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS);
1560 if (gfp_mask & __GFP_NOFAIL) {
1561 congestion_wait(WRITE, HZ/50);
1568 /* Atomic allocations - we can't balance anything */
1574 /* We now go into synchronous reclaim */
1575 cpuset_memory_pressure_bump();
1577 * The task's cpuset might have expanded its set of allowable nodes
1579 cpuset_update_task_memory_state();
1580 p->flags |= PF_MEMALLOC;
1581 reclaim_state.reclaimed_slab = 0;
1582 p->reclaim_state = &reclaim_state;
1584 did_some_progress = try_to_free_pages(zonelist, order, gfp_mask);
1586 p->reclaim_state = NULL;
1587 p->flags &= ~PF_MEMALLOC;
1594 if (likely(did_some_progress)) {
1595 page = get_page_from_freelist(gfp_mask, nodemask, order,
1596 zonelist, high_zoneidx, alloc_flags);
1599 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1600 if (!try_set_zone_oom(zonelist, gfp_mask)) {
1601 schedule_timeout_uninterruptible(1);
1606 * Go through the zonelist yet one more time, keep
1607 * very high watermark here, this is only to catch
1608 * a parallel oom killing, we must fail if we're still
1609 * under heavy pressure.
1611 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1612 order, zonelist, high_zoneidx,
1613 ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1615 clear_zonelist_oom(zonelist, gfp_mask);
1619 /* The OOM killer will not help higher order allocs so fail */
1620 if (order > PAGE_ALLOC_COSTLY_ORDER) {
1621 clear_zonelist_oom(zonelist, gfp_mask);
1625 out_of_memory(zonelist, gfp_mask, order);
1626 clear_zonelist_oom(zonelist, gfp_mask);
1631 * Don't let big-order allocations loop unless the caller explicitly
1632 * requests that. Wait for some write requests to complete then retry.
1634 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1635 * means __GFP_NOFAIL, but that may not be true in other
1638 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1639 * specified, then we retry until we no longer reclaim any pages
1640 * (above), or we've reclaimed an order of pages at least as
1641 * large as the allocation's order. In both cases, if the
1642 * allocation still fails, we stop retrying.
1644 pages_reclaimed += did_some_progress;
1646 if (!(gfp_mask & __GFP_NORETRY)) {
1647 if (order <= PAGE_ALLOC_COSTLY_ORDER) {
1650 if (gfp_mask & __GFP_REPEAT &&
1651 pages_reclaimed < (1 << order))
1654 if (gfp_mask & __GFP_NOFAIL)
1658 congestion_wait(WRITE, HZ/50);
1663 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1664 printk(KERN_WARNING "%s: page allocation failure."
1665 " order:%d, mode:0x%x\n",
1666 p->comm, order, gfp_mask);
1673 EXPORT_SYMBOL(__alloc_pages_internal);
1676 * Common helper functions.
1678 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1681 page = alloc_pages(gfp_mask, order);
1684 return (unsigned long) page_address(page);
1687 EXPORT_SYMBOL(__get_free_pages);
1689 unsigned long get_zeroed_page(gfp_t gfp_mask)
1694 * get_zeroed_page() returns a 32-bit address, which cannot represent
1697 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1699 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1701 return (unsigned long) page_address(page);
1705 EXPORT_SYMBOL(get_zeroed_page);
1707 void __pagevec_free(struct pagevec *pvec)
1709 int i = pagevec_count(pvec);
1712 free_hot_cold_page(pvec->pages[i], pvec->cold);
1715 void __free_pages(struct page *page, unsigned int order)
1717 if (put_page_testzero(page)) {
1719 free_hot_page(page);
1721 __free_pages_ok(page, order);
1725 EXPORT_SYMBOL(__free_pages);
1727 void free_pages(unsigned long addr, unsigned int order)
1730 VM_BUG_ON(!virt_addr_valid((void *)addr));
1731 __free_pages(virt_to_page((void *)addr), order);
1735 EXPORT_SYMBOL(free_pages);
1738 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
1739 * @size: the number of bytes to allocate
1740 * @gfp_mask: GFP flags for the allocation
1742 * This function is similar to alloc_pages(), except that it allocates the
1743 * minimum number of pages to satisfy the request. alloc_pages() can only
1744 * allocate memory in power-of-two pages.
1746 * This function is also limited by MAX_ORDER.
1748 * Memory allocated by this function must be released by free_pages_exact().
1750 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
1752 unsigned int order = get_order(size);
1755 addr = __get_free_pages(gfp_mask, order);
1757 unsigned long alloc_end = addr + (PAGE_SIZE << order);
1758 unsigned long used = addr + PAGE_ALIGN(size);
1760 split_page(virt_to_page(addr), order);
1761 while (used < alloc_end) {
1767 return (void *)addr;
1769 EXPORT_SYMBOL(alloc_pages_exact);
1772 * free_pages_exact - release memory allocated via alloc_pages_exact()
1773 * @virt: the value returned by alloc_pages_exact.
1774 * @size: size of allocation, same value as passed to alloc_pages_exact().
1776 * Release the memory allocated by a previous call to alloc_pages_exact.
1778 void free_pages_exact(void *virt, size_t size)
1780 unsigned long addr = (unsigned long)virt;
1781 unsigned long end = addr + PAGE_ALIGN(size);
1783 while (addr < end) {
1788 EXPORT_SYMBOL(free_pages_exact);
1790 static unsigned int nr_free_zone_pages(int offset)
1795 /* Just pick one node, since fallback list is circular */
1796 unsigned int sum = 0;
1798 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
1800 for_each_zone_zonelist(zone, z, zonelist, offset) {
1801 unsigned long size = zone->present_pages;
1802 unsigned long high = zone->pages_high;
1811 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1813 unsigned int nr_free_buffer_pages(void)
1815 return nr_free_zone_pages(gfp_zone(GFP_USER));
1817 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
1820 * Amount of free RAM allocatable within all zones
1822 unsigned int nr_free_pagecache_pages(void)
1824 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
1827 static inline void show_node(struct zone *zone)
1830 printk("Node %d ", zone_to_nid(zone));
1833 void si_meminfo(struct sysinfo *val)
1835 val->totalram = totalram_pages;
1837 val->freeram = global_page_state(NR_FREE_PAGES);
1838 val->bufferram = nr_blockdev_pages();
1839 val->totalhigh = totalhigh_pages;
1840 val->freehigh = nr_free_highpages();
1841 val->mem_unit = PAGE_SIZE;
1844 EXPORT_SYMBOL(si_meminfo);
1847 void si_meminfo_node(struct sysinfo *val, int nid)
1849 pg_data_t *pgdat = NODE_DATA(nid);
1851 val->totalram = pgdat->node_present_pages;
1852 val->freeram = node_page_state(nid, NR_FREE_PAGES);
1853 #ifdef CONFIG_HIGHMEM
1854 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1855 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
1861 val->mem_unit = PAGE_SIZE;
1865 #define K(x) ((x) << (PAGE_SHIFT-10))
1868 * Show free area list (used inside shift_scroll-lock stuff)
1869 * We also calculate the percentage fragmentation. We do this by counting the
1870 * memory on each free list with the exception of the first item on the list.
1872 void show_free_areas(void)
1877 for_each_zone(zone) {
1878 if (!populated_zone(zone))
1882 printk("%s per-cpu:\n", zone->name);
1884 for_each_online_cpu(cpu) {
1885 struct per_cpu_pageset *pageset;
1887 pageset = zone_pcp(zone, cpu);
1889 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
1890 cpu, pageset->pcp.high,
1891 pageset->pcp.batch, pageset->pcp.count);
1895 printk("Active_anon:%lu active_file:%lu inactive_anon:%lu\n"
1896 " inactive_file:%lu"
1897 //TODO: check/adjust line lengths
1898 #ifdef CONFIG_UNEVICTABLE_LRU
1901 " dirty:%lu writeback:%lu unstable:%lu\n"
1902 " free:%lu slab:%lu mapped:%lu pagetables:%lu bounce:%lu\n",
1903 global_page_state(NR_ACTIVE_ANON),
1904 global_page_state(NR_ACTIVE_FILE),
1905 global_page_state(NR_INACTIVE_ANON),
1906 global_page_state(NR_INACTIVE_FILE),
1907 #ifdef CONFIG_UNEVICTABLE_LRU
1908 global_page_state(NR_UNEVICTABLE),
1910 global_page_state(NR_FILE_DIRTY),
1911 global_page_state(NR_WRITEBACK),
1912 global_page_state(NR_UNSTABLE_NFS),
1913 global_page_state(NR_FREE_PAGES),
1914 global_page_state(NR_SLAB_RECLAIMABLE) +
1915 global_page_state(NR_SLAB_UNRECLAIMABLE),
1916 global_page_state(NR_FILE_MAPPED),
1917 global_page_state(NR_PAGETABLE),
1918 global_page_state(NR_BOUNCE));
1920 for_each_zone(zone) {
1923 if (!populated_zone(zone))
1932 " active_anon:%lukB"
1933 " inactive_anon:%lukB"
1934 " active_file:%lukB"
1935 " inactive_file:%lukB"
1936 #ifdef CONFIG_UNEVICTABLE_LRU
1937 " unevictable:%lukB"
1940 " pages_scanned:%lu"
1941 " all_unreclaimable? %s"
1944 K(zone_page_state(zone, NR_FREE_PAGES)),
1947 K(zone->pages_high),
1948 K(zone_page_state(zone, NR_ACTIVE_ANON)),
1949 K(zone_page_state(zone, NR_INACTIVE_ANON)),
1950 K(zone_page_state(zone, NR_ACTIVE_FILE)),
1951 K(zone_page_state(zone, NR_INACTIVE_FILE)),
1952 #ifdef CONFIG_UNEVICTABLE_LRU
1953 K(zone_page_state(zone, NR_UNEVICTABLE)),
1955 K(zone->present_pages),
1956 zone->pages_scanned,
1957 (zone_is_all_unreclaimable(zone) ? "yes" : "no")
1959 printk("lowmem_reserve[]:");
1960 for (i = 0; i < MAX_NR_ZONES; i++)
1961 printk(" %lu", zone->lowmem_reserve[i]);
1965 for_each_zone(zone) {
1966 unsigned long nr[MAX_ORDER], flags, order, total = 0;
1968 if (!populated_zone(zone))
1972 printk("%s: ", zone->name);
1974 spin_lock_irqsave(&zone->lock, flags);
1975 for (order = 0; order < MAX_ORDER; order++) {
1976 nr[order] = zone->free_area[order].nr_free;
1977 total += nr[order] << order;
1979 spin_unlock_irqrestore(&zone->lock, flags);
1980 for (order = 0; order < MAX_ORDER; order++)
1981 printk("%lu*%lukB ", nr[order], K(1UL) << order);
1982 printk("= %lukB\n", K(total));
1985 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
1987 show_swap_cache_info();
1990 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
1992 zoneref->zone = zone;
1993 zoneref->zone_idx = zone_idx(zone);
1997 * Builds allocation fallback zone lists.
1999 * Add all populated zones of a node to the zonelist.
2001 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2002 int nr_zones, enum zone_type zone_type)
2006 BUG_ON(zone_type >= MAX_NR_ZONES);
2011 zone = pgdat->node_zones + zone_type;
2012 if (populated_zone(zone)) {
2013 zoneref_set_zone(zone,
2014 &zonelist->_zonerefs[nr_zones++]);
2015 check_highest_zone(zone_type);
2018 } while (zone_type);
2025 * 0 = automatic detection of better ordering.
2026 * 1 = order by ([node] distance, -zonetype)
2027 * 2 = order by (-zonetype, [node] distance)
2029 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2030 * the same zonelist. So only NUMA can configure this param.
2032 #define ZONELIST_ORDER_DEFAULT 0
2033 #define ZONELIST_ORDER_NODE 1
2034 #define ZONELIST_ORDER_ZONE 2
2036 /* zonelist order in the kernel.
2037 * set_zonelist_order() will set this to NODE or ZONE.
2039 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2040 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2044 /* The value user specified ....changed by config */
2045 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2046 /* string for sysctl */
2047 #define NUMA_ZONELIST_ORDER_LEN 16
2048 char numa_zonelist_order[16] = "default";
2051 * interface for configure zonelist ordering.
2052 * command line option "numa_zonelist_order"
2053 * = "[dD]efault - default, automatic configuration.
2054 * = "[nN]ode - order by node locality, then by zone within node
2055 * = "[zZ]one - order by zone, then by locality within zone
2058 static int __parse_numa_zonelist_order(char *s)
2060 if (*s == 'd' || *s == 'D') {
2061 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2062 } else if (*s == 'n' || *s == 'N') {
2063 user_zonelist_order = ZONELIST_ORDER_NODE;
2064 } else if (*s == 'z' || *s == 'Z') {
2065 user_zonelist_order = ZONELIST_ORDER_ZONE;
2068 "Ignoring invalid numa_zonelist_order value: "
2075 static __init int setup_numa_zonelist_order(char *s)
2078 return __parse_numa_zonelist_order(s);
2081 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2084 * sysctl handler for numa_zonelist_order
2086 int numa_zonelist_order_handler(ctl_table *table, int write,
2087 struct file *file, void __user *buffer, size_t *length,
2090 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2094 strncpy(saved_string, (char*)table->data,
2095 NUMA_ZONELIST_ORDER_LEN);
2096 ret = proc_dostring(table, write, file, buffer, length, ppos);
2100 int oldval = user_zonelist_order;
2101 if (__parse_numa_zonelist_order((char*)table->data)) {
2103 * bogus value. restore saved string
2105 strncpy((char*)table->data, saved_string,
2106 NUMA_ZONELIST_ORDER_LEN);
2107 user_zonelist_order = oldval;
2108 } else if (oldval != user_zonelist_order)
2109 build_all_zonelists();
2115 #define MAX_NODE_LOAD (num_online_nodes())
2116 static int node_load[MAX_NUMNODES];
2119 * find_next_best_node - find the next node that should appear in a given node's fallback list
2120 * @node: node whose fallback list we're appending
2121 * @used_node_mask: nodemask_t of already used nodes
2123 * We use a number of factors to determine which is the next node that should
2124 * appear on a given node's fallback list. The node should not have appeared
2125 * already in @node's fallback list, and it should be the next closest node
2126 * according to the distance array (which contains arbitrary distance values
2127 * from each node to each node in the system), and should also prefer nodes
2128 * with no CPUs, since presumably they'll have very little allocation pressure
2129 * on them otherwise.
2130 * It returns -1 if no node is found.
2132 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2135 int min_val = INT_MAX;
2137 node_to_cpumask_ptr(tmp, 0);
2139 /* Use the local node if we haven't already */
2140 if (!node_isset(node, *used_node_mask)) {
2141 node_set(node, *used_node_mask);
2145 for_each_node_state(n, N_HIGH_MEMORY) {
2147 /* Don't want a node to appear more than once */
2148 if (node_isset(n, *used_node_mask))
2151 /* Use the distance array to find the distance */
2152 val = node_distance(node, n);
2154 /* Penalize nodes under us ("prefer the next node") */
2157 /* Give preference to headless and unused nodes */
2158 node_to_cpumask_ptr_next(tmp, n);
2159 if (!cpus_empty(*tmp))
2160 val += PENALTY_FOR_NODE_WITH_CPUS;
2162 /* Slight preference for less loaded node */
2163 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2164 val += node_load[n];
2166 if (val < min_val) {
2173 node_set(best_node, *used_node_mask);
2180 * Build zonelists ordered by node and zones within node.
2181 * This results in maximum locality--normal zone overflows into local
2182 * DMA zone, if any--but risks exhausting DMA zone.
2184 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2187 struct zonelist *zonelist;
2189 zonelist = &pgdat->node_zonelists[0];
2190 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2192 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2194 zonelist->_zonerefs[j].zone = NULL;
2195 zonelist->_zonerefs[j].zone_idx = 0;
2199 * Build gfp_thisnode zonelists
2201 static void build_thisnode_zonelists(pg_data_t *pgdat)
2204 struct zonelist *zonelist;
2206 zonelist = &pgdat->node_zonelists[1];
2207 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2208 zonelist->_zonerefs[j].zone = NULL;
2209 zonelist->_zonerefs[j].zone_idx = 0;
2213 * Build zonelists ordered by zone and nodes within zones.
2214 * This results in conserving DMA zone[s] until all Normal memory is
2215 * exhausted, but results in overflowing to remote node while memory
2216 * may still exist in local DMA zone.
2218 static int node_order[MAX_NUMNODES];
2220 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2223 int zone_type; /* needs to be signed */
2225 struct zonelist *zonelist;
2227 zonelist = &pgdat->node_zonelists[0];
2229 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2230 for (j = 0; j < nr_nodes; j++) {
2231 node = node_order[j];
2232 z = &NODE_DATA(node)->node_zones[zone_type];
2233 if (populated_zone(z)) {
2235 &zonelist->_zonerefs[pos++]);
2236 check_highest_zone(zone_type);
2240 zonelist->_zonerefs[pos].zone = NULL;
2241 zonelist->_zonerefs[pos].zone_idx = 0;
2244 static int default_zonelist_order(void)
2247 unsigned long low_kmem_size,total_size;
2251 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2252 * If they are really small and used heavily, the system can fall
2253 * into OOM very easily.
2254 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2256 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2259 for_each_online_node(nid) {
2260 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2261 z = &NODE_DATA(nid)->node_zones[zone_type];
2262 if (populated_zone(z)) {
2263 if (zone_type < ZONE_NORMAL)
2264 low_kmem_size += z->present_pages;
2265 total_size += z->present_pages;
2269 if (!low_kmem_size || /* there are no DMA area. */
2270 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2271 return ZONELIST_ORDER_NODE;
2273 * look into each node's config.
2274 * If there is a node whose DMA/DMA32 memory is very big area on
2275 * local memory, NODE_ORDER may be suitable.
2277 average_size = total_size /
2278 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2279 for_each_online_node(nid) {
2282 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2283 z = &NODE_DATA(nid)->node_zones[zone_type];
2284 if (populated_zone(z)) {
2285 if (zone_type < ZONE_NORMAL)
2286 low_kmem_size += z->present_pages;
2287 total_size += z->present_pages;
2290 if (low_kmem_size &&
2291 total_size > average_size && /* ignore small node */
2292 low_kmem_size > total_size * 70/100)
2293 return ZONELIST_ORDER_NODE;
2295 return ZONELIST_ORDER_ZONE;
2298 static void set_zonelist_order(void)
2300 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2301 current_zonelist_order = default_zonelist_order();
2303 current_zonelist_order = user_zonelist_order;
2306 static void build_zonelists(pg_data_t *pgdat)
2310 nodemask_t used_mask;
2311 int local_node, prev_node;
2312 struct zonelist *zonelist;
2313 int order = current_zonelist_order;
2315 /* initialize zonelists */
2316 for (i = 0; i < MAX_ZONELISTS; i++) {
2317 zonelist = pgdat->node_zonelists + i;
2318 zonelist->_zonerefs[0].zone = NULL;
2319 zonelist->_zonerefs[0].zone_idx = 0;
2322 /* NUMA-aware ordering of nodes */
2323 local_node = pgdat->node_id;
2324 load = num_online_nodes();
2325 prev_node = local_node;
2326 nodes_clear(used_mask);
2328 memset(node_load, 0, sizeof(node_load));
2329 memset(node_order, 0, sizeof(node_order));
2332 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2333 int distance = node_distance(local_node, node);
2336 * If another node is sufficiently far away then it is better
2337 * to reclaim pages in a zone before going off node.
2339 if (distance > RECLAIM_DISTANCE)
2340 zone_reclaim_mode = 1;
2343 * We don't want to pressure a particular node.
2344 * So adding penalty to the first node in same
2345 * distance group to make it round-robin.
2347 if (distance != node_distance(local_node, prev_node))
2348 node_load[node] = load;
2352 if (order == ZONELIST_ORDER_NODE)
2353 build_zonelists_in_node_order(pgdat, node);
2355 node_order[j++] = node; /* remember order */
2358 if (order == ZONELIST_ORDER_ZONE) {
2359 /* calculate node order -- i.e., DMA last! */
2360 build_zonelists_in_zone_order(pgdat, j);
2363 build_thisnode_zonelists(pgdat);
2366 /* Construct the zonelist performance cache - see further mmzone.h */
2367 static void build_zonelist_cache(pg_data_t *pgdat)
2369 struct zonelist *zonelist;
2370 struct zonelist_cache *zlc;
2373 zonelist = &pgdat->node_zonelists[0];
2374 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2375 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2376 for (z = zonelist->_zonerefs; z->zone; z++)
2377 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2381 #else /* CONFIG_NUMA */
2383 static void set_zonelist_order(void)
2385 current_zonelist_order = ZONELIST_ORDER_ZONE;
2388 static void build_zonelists(pg_data_t *pgdat)
2390 int node, local_node;
2392 struct zonelist *zonelist;
2394 local_node = pgdat->node_id;
2396 zonelist = &pgdat->node_zonelists[0];
2397 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2400 * Now we build the zonelist so that it contains the zones
2401 * of all the other nodes.
2402 * We don't want to pressure a particular node, so when
2403 * building the zones for node N, we make sure that the
2404 * zones coming right after the local ones are those from
2405 * node N+1 (modulo N)
2407 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2408 if (!node_online(node))
2410 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2413 for (node = 0; node < local_node; node++) {
2414 if (!node_online(node))
2416 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2420 zonelist->_zonerefs[j].zone = NULL;
2421 zonelist->_zonerefs[j].zone_idx = 0;
2424 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2425 static void build_zonelist_cache(pg_data_t *pgdat)
2427 pgdat->node_zonelists[0].zlcache_ptr = NULL;
2430 #endif /* CONFIG_NUMA */
2432 /* return values int ....just for stop_machine() */
2433 static int __build_all_zonelists(void *dummy)
2437 for_each_online_node(nid) {
2438 pg_data_t *pgdat = NODE_DATA(nid);
2440 build_zonelists(pgdat);
2441 build_zonelist_cache(pgdat);
2446 void build_all_zonelists(void)
2448 set_zonelist_order();
2450 if (system_state == SYSTEM_BOOTING) {
2451 __build_all_zonelists(NULL);
2452 mminit_verify_zonelist();
2453 cpuset_init_current_mems_allowed();
2455 /* we have to stop all cpus to guarantee there is no user
2457 stop_machine(__build_all_zonelists, NULL, NULL);
2458 /* cpuset refresh routine should be here */
2460 vm_total_pages = nr_free_pagecache_pages();
2462 * Disable grouping by mobility if the number of pages in the
2463 * system is too low to allow the mechanism to work. It would be
2464 * more accurate, but expensive to check per-zone. This check is
2465 * made on memory-hotadd so a system can start with mobility
2466 * disabled and enable it later
2468 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
2469 page_group_by_mobility_disabled = 1;
2471 page_group_by_mobility_disabled = 0;
2473 printk("Built %i zonelists in %s order, mobility grouping %s. "
2474 "Total pages: %ld\n",
2476 zonelist_order_name[current_zonelist_order],
2477 page_group_by_mobility_disabled ? "off" : "on",
2480 printk("Policy zone: %s\n", zone_names[policy_zone]);
2485 * Helper functions to size the waitqueue hash table.
2486 * Essentially these want to choose hash table sizes sufficiently
2487 * large so that collisions trying to wait on pages are rare.
2488 * But in fact, the number of active page waitqueues on typical
2489 * systems is ridiculously low, less than 200. So this is even
2490 * conservative, even though it seems large.
2492 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2493 * waitqueues, i.e. the size of the waitq table given the number of pages.
2495 #define PAGES_PER_WAITQUEUE 256
2497 #ifndef CONFIG_MEMORY_HOTPLUG
2498 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2500 unsigned long size = 1;
2502 pages /= PAGES_PER_WAITQUEUE;
2504 while (size < pages)
2508 * Once we have dozens or even hundreds of threads sleeping
2509 * on IO we've got bigger problems than wait queue collision.
2510 * Limit the size of the wait table to a reasonable size.
2512 size = min(size, 4096UL);
2514 return max(size, 4UL);
2518 * A zone's size might be changed by hot-add, so it is not possible to determine
2519 * a suitable size for its wait_table. So we use the maximum size now.
2521 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2523 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2524 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2525 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2527 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2528 * or more by the traditional way. (See above). It equals:
2530 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2531 * ia64(16K page size) : = ( 8G + 4M)byte.
2532 * powerpc (64K page size) : = (32G +16M)byte.
2534 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2541 * This is an integer logarithm so that shifts can be used later
2542 * to extract the more random high bits from the multiplicative
2543 * hash function before the remainder is taken.
2545 static inline unsigned long wait_table_bits(unsigned long size)
2550 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2553 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2554 * of blocks reserved is based on zone->pages_min. The memory within the
2555 * reserve will tend to store contiguous free pages. Setting min_free_kbytes
2556 * higher will lead to a bigger reserve which will get freed as contiguous
2557 * blocks as reclaim kicks in
2559 static void setup_zone_migrate_reserve(struct zone *zone)
2561 unsigned long start_pfn, pfn, end_pfn;
2563 unsigned long reserve, block_migratetype;
2565 /* Get the start pfn, end pfn and the number of blocks to reserve */
2566 start_pfn = zone->zone_start_pfn;
2567 end_pfn = start_pfn + zone->spanned_pages;
2568 reserve = roundup(zone->pages_min, pageblock_nr_pages) >>
2571 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
2572 if (!pfn_valid(pfn))
2574 page = pfn_to_page(pfn);
2576 /* Watch out for overlapping nodes */
2577 if (page_to_nid(page) != zone_to_nid(zone))
2580 /* Blocks with reserved pages will never free, skip them. */
2581 if (PageReserved(page))
2584 block_migratetype = get_pageblock_migratetype(page);
2586 /* If this block is reserved, account for it */
2587 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
2592 /* Suitable for reserving if this block is movable */
2593 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
2594 set_pageblock_migratetype(page, MIGRATE_RESERVE);
2595 move_freepages_block(zone, page, MIGRATE_RESERVE);
2601 * If the reserve is met and this is a previous reserved block,
2604 if (block_migratetype == MIGRATE_RESERVE) {
2605 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2606 move_freepages_block(zone, page, MIGRATE_MOVABLE);
2612 * Initially all pages are reserved - free ones are freed
2613 * up by free_all_bootmem() once the early boot process is
2614 * done. Non-atomic initialization, single-pass.
2616 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
2617 unsigned long start_pfn, enum memmap_context context)
2620 unsigned long end_pfn = start_pfn + size;
2624 if (highest_memmap_pfn < end_pfn - 1)
2625 highest_memmap_pfn = end_pfn - 1;
2627 z = &NODE_DATA(nid)->node_zones[zone];
2628 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
2630 * There can be holes in boot-time mem_map[]s
2631 * handed to this function. They do not
2632 * exist on hotplugged memory.
2634 if (context == MEMMAP_EARLY) {
2635 if (!early_pfn_valid(pfn))
2637 if (!early_pfn_in_nid(pfn, nid))
2640 page = pfn_to_page(pfn);
2641 set_page_links(page, zone, nid, pfn);
2642 mminit_verify_page_links(page, zone, nid, pfn);
2643 init_page_count(page);
2644 reset_page_mapcount(page);
2645 SetPageReserved(page);
2647 * Mark the block movable so that blocks are reserved for
2648 * movable at startup. This will force kernel allocations
2649 * to reserve their blocks rather than leaking throughout
2650 * the address space during boot when many long-lived
2651 * kernel allocations are made. Later some blocks near
2652 * the start are marked MIGRATE_RESERVE by
2653 * setup_zone_migrate_reserve()
2655 * bitmap is created for zone's valid pfn range. but memmap
2656 * can be created for invalid pages (for alignment)
2657 * check here not to call set_pageblock_migratetype() against
2660 if ((z->zone_start_pfn <= pfn)
2661 && (pfn < z->zone_start_pfn + z->spanned_pages)
2662 && !(pfn & (pageblock_nr_pages - 1)))
2663 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2665 INIT_LIST_HEAD(&page->lru);
2666 #ifdef WANT_PAGE_VIRTUAL
2667 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2668 if (!is_highmem_idx(zone))
2669 set_page_address(page, __va(pfn << PAGE_SHIFT));
2674 static void __meminit zone_init_free_lists(struct zone *zone)
2677 for_each_migratetype_order(order, t) {
2678 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
2679 zone->free_area[order].nr_free = 0;
2683 #ifndef __HAVE_ARCH_MEMMAP_INIT
2684 #define memmap_init(size, nid, zone, start_pfn) \
2685 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
2688 static int zone_batchsize(struct zone *zone)
2693 * The per-cpu-pages pools are set to around 1000th of the
2694 * size of the zone. But no more than 1/2 of a meg.
2696 * OK, so we don't know how big the cache is. So guess.
2698 batch = zone->present_pages / 1024;
2699 if (batch * PAGE_SIZE > 512 * 1024)
2700 batch = (512 * 1024) / PAGE_SIZE;
2701 batch /= 4; /* We effectively *= 4 below */
2706 * Clamp the batch to a 2^n - 1 value. Having a power
2707 * of 2 value was found to be more likely to have
2708 * suboptimal cache aliasing properties in some cases.
2710 * For example if 2 tasks are alternately allocating
2711 * batches of pages, one task can end up with a lot
2712 * of pages of one half of the possible page colors
2713 * and the other with pages of the other colors.
2715 batch = (1 << (fls(batch + batch/2)-1)) - 1;
2720 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
2722 struct per_cpu_pages *pcp;
2724 memset(p, 0, sizeof(*p));
2728 pcp->high = 6 * batch;
2729 pcp->batch = max(1UL, 1 * batch);
2730 INIT_LIST_HEAD(&pcp->list);
2734 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
2735 * to the value high for the pageset p.
2738 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
2741 struct per_cpu_pages *pcp;
2745 pcp->batch = max(1UL, high/4);
2746 if ((high/4) > (PAGE_SHIFT * 8))
2747 pcp->batch = PAGE_SHIFT * 8;
2753 * Boot pageset table. One per cpu which is going to be used for all
2754 * zones and all nodes. The parameters will be set in such a way
2755 * that an item put on a list will immediately be handed over to
2756 * the buddy list. This is safe since pageset manipulation is done
2757 * with interrupts disabled.
2759 * Some NUMA counter updates may also be caught by the boot pagesets.
2761 * The boot_pagesets must be kept even after bootup is complete for
2762 * unused processors and/or zones. They do play a role for bootstrapping
2763 * hotplugged processors.
2765 * zoneinfo_show() and maybe other functions do
2766 * not check if the processor is online before following the pageset pointer.
2767 * Other parts of the kernel may not check if the zone is available.
2769 static struct per_cpu_pageset boot_pageset[NR_CPUS];
2772 * Dynamically allocate memory for the
2773 * per cpu pageset array in struct zone.
2775 static int __cpuinit process_zones(int cpu)
2777 struct zone *zone, *dzone;
2778 int node = cpu_to_node(cpu);
2780 node_set_state(node, N_CPU); /* this node has a cpu */
2782 for_each_zone(zone) {
2784 if (!populated_zone(zone))
2787 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
2789 if (!zone_pcp(zone, cpu))
2792 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
2794 if (percpu_pagelist_fraction)
2795 setup_pagelist_highmark(zone_pcp(zone, cpu),
2796 (zone->present_pages / percpu_pagelist_fraction));
2801 for_each_zone(dzone) {
2802 if (!populated_zone(dzone))
2806 kfree(zone_pcp(dzone, cpu));
2807 zone_pcp(dzone, cpu) = NULL;
2812 static inline void free_zone_pagesets(int cpu)
2816 for_each_zone(zone) {
2817 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
2819 /* Free per_cpu_pageset if it is slab allocated */
2820 if (pset != &boot_pageset[cpu])
2822 zone_pcp(zone, cpu) = NULL;
2826 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
2827 unsigned long action,
2830 int cpu = (long)hcpu;
2831 int ret = NOTIFY_OK;
2834 case CPU_UP_PREPARE:
2835 case CPU_UP_PREPARE_FROZEN:
2836 if (process_zones(cpu))
2839 case CPU_UP_CANCELED:
2840 case CPU_UP_CANCELED_FROZEN:
2842 case CPU_DEAD_FROZEN:
2843 free_zone_pagesets(cpu);
2851 static struct notifier_block __cpuinitdata pageset_notifier =
2852 { &pageset_cpuup_callback, NULL, 0 };
2854 void __init setup_per_cpu_pageset(void)
2858 /* Initialize per_cpu_pageset for cpu 0.
2859 * A cpuup callback will do this for every cpu
2860 * as it comes online
2862 err = process_zones(smp_processor_id());
2864 register_cpu_notifier(&pageset_notifier);
2869 static noinline __init_refok
2870 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
2873 struct pglist_data *pgdat = zone->zone_pgdat;
2877 * The per-page waitqueue mechanism uses hashed waitqueues
2880 zone->wait_table_hash_nr_entries =
2881 wait_table_hash_nr_entries(zone_size_pages);
2882 zone->wait_table_bits =
2883 wait_table_bits(zone->wait_table_hash_nr_entries);
2884 alloc_size = zone->wait_table_hash_nr_entries
2885 * sizeof(wait_queue_head_t);
2887 if (!slab_is_available()) {
2888 zone->wait_table = (wait_queue_head_t *)
2889 alloc_bootmem_node(pgdat, alloc_size);
2892 * This case means that a zone whose size was 0 gets new memory
2893 * via memory hot-add.
2894 * But it may be the case that a new node was hot-added. In
2895 * this case vmalloc() will not be able to use this new node's
2896 * memory - this wait_table must be initialized to use this new
2897 * node itself as well.
2898 * To use this new node's memory, further consideration will be
2901 zone->wait_table = vmalloc(alloc_size);
2903 if (!zone->wait_table)
2906 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
2907 init_waitqueue_head(zone->wait_table + i);
2912 static __meminit void zone_pcp_init(struct zone *zone)
2915 unsigned long batch = zone_batchsize(zone);
2917 for (cpu = 0; cpu < NR_CPUS; cpu++) {
2919 /* Early boot. Slab allocator not functional yet */
2920 zone_pcp(zone, cpu) = &boot_pageset[cpu];
2921 setup_pageset(&boot_pageset[cpu],0);
2923 setup_pageset(zone_pcp(zone,cpu), batch);
2926 if (zone->present_pages)
2927 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
2928 zone->name, zone->present_pages, batch);
2931 __meminit int init_currently_empty_zone(struct zone *zone,
2932 unsigned long zone_start_pfn,
2934 enum memmap_context context)
2936 struct pglist_data *pgdat = zone->zone_pgdat;
2938 ret = zone_wait_table_init(zone, size);
2941 pgdat->nr_zones = zone_idx(zone) + 1;
2943 zone->zone_start_pfn = zone_start_pfn;
2945 mminit_dprintk(MMINIT_TRACE, "memmap_init",
2946 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
2948 (unsigned long)zone_idx(zone),
2949 zone_start_pfn, (zone_start_pfn + size));
2951 zone_init_free_lists(zone);
2956 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
2958 * Basic iterator support. Return the first range of PFNs for a node
2959 * Note: nid == MAX_NUMNODES returns first region regardless of node
2961 static int __meminit first_active_region_index_in_nid(int nid)
2965 for (i = 0; i < nr_nodemap_entries; i++)
2966 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
2973 * Basic iterator support. Return the next active range of PFNs for a node
2974 * Note: nid == MAX_NUMNODES returns next region regardless of node
2976 static int __meminit next_active_region_index_in_nid(int index, int nid)
2978 for (index = index + 1; index < nr_nodemap_entries; index++)
2979 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
2985 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
2987 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
2988 * Architectures may implement their own version but if add_active_range()
2989 * was used and there are no special requirements, this is a convenient
2992 int __meminit early_pfn_to_nid(unsigned long pfn)
2996 for (i = 0; i < nr_nodemap_entries; i++) {
2997 unsigned long start_pfn = early_node_map[i].start_pfn;
2998 unsigned long end_pfn = early_node_map[i].end_pfn;
3000 if (start_pfn <= pfn && pfn < end_pfn)
3001 return early_node_map[i].nid;
3006 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3008 /* Basic iterator support to walk early_node_map[] */
3009 #define for_each_active_range_index_in_nid(i, nid) \
3010 for (i = first_active_region_index_in_nid(nid); i != -1; \
3011 i = next_active_region_index_in_nid(i, nid))
3014 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3015 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3016 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3018 * If an architecture guarantees that all ranges registered with
3019 * add_active_ranges() contain no holes and may be freed, this
3020 * this function may be used instead of calling free_bootmem() manually.
3022 void __init free_bootmem_with_active_regions(int nid,
3023 unsigned long max_low_pfn)
3027 for_each_active_range_index_in_nid(i, nid) {
3028 unsigned long size_pages = 0;
3029 unsigned long end_pfn = early_node_map[i].end_pfn;
3031 if (early_node_map[i].start_pfn >= max_low_pfn)
3034 if (end_pfn > max_low_pfn)
3035 end_pfn = max_low_pfn;
3037 size_pages = end_pfn - early_node_map[i].start_pfn;
3038 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3039 PFN_PHYS(early_node_map[i].start_pfn),
3040 size_pages << PAGE_SHIFT);
3044 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3049 for_each_active_range_index_in_nid(i, nid) {
3050 ret = work_fn(early_node_map[i].start_pfn,
3051 early_node_map[i].end_pfn, data);
3057 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3058 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3060 * If an architecture guarantees that all ranges registered with
3061 * add_active_ranges() contain no holes and may be freed, this
3062 * function may be used instead of calling memory_present() manually.
3064 void __init sparse_memory_present_with_active_regions(int nid)
3068 for_each_active_range_index_in_nid(i, nid)
3069 memory_present(early_node_map[i].nid,
3070 early_node_map[i].start_pfn,
3071 early_node_map[i].end_pfn);
3075 * push_node_boundaries - Push node boundaries to at least the requested boundary
3076 * @nid: The nid of the node to push the boundary for
3077 * @start_pfn: The start pfn of the node
3078 * @end_pfn: The end pfn of the node
3080 * In reserve-based hot-add, mem_map is allocated that is unused until hotadd
3081 * time. Specifically, on x86_64, SRAT will report ranges that can potentially
3082 * be hotplugged even though no physical memory exists. This function allows
3083 * an arch to push out the node boundaries so mem_map is allocated that can
3086 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
3087 void __init push_node_boundaries(unsigned int nid,
3088 unsigned long start_pfn, unsigned long end_pfn)
3090 mminit_dprintk(MMINIT_TRACE, "zoneboundary",
3091 "Entering push_node_boundaries(%u, %lu, %lu)\n",
3092 nid, start_pfn, end_pfn);
3094 /* Initialise the boundary for this node if necessary */
3095 if (node_boundary_end_pfn[nid] == 0)
3096 node_boundary_start_pfn[nid] = -1UL;
3098 /* Update the boundaries */
3099 if (node_boundary_start_pfn[nid] > start_pfn)
3100 node_boundary_start_pfn[nid] = start_pfn;
3101 if (node_boundary_end_pfn[nid] < end_pfn)
3102 node_boundary_end_pfn[nid] = end_pfn;
3105 /* If necessary, push the node boundary out for reserve hotadd */
3106 static void __meminit account_node_boundary(unsigned int nid,
3107 unsigned long *start_pfn, unsigned long *end_pfn)
3109 mminit_dprintk(MMINIT_TRACE, "zoneboundary",
3110 "Entering account_node_boundary(%u, %lu, %lu)\n",
3111 nid, *start_pfn, *end_pfn);
3113 /* Return if boundary information has not been provided */
3114 if (node_boundary_end_pfn[nid] == 0)
3117 /* Check the boundaries and update if necessary */
3118 if (node_boundary_start_pfn[nid] < *start_pfn)
3119 *start_pfn = node_boundary_start_pfn[nid];
3120 if (node_boundary_end_pfn[nid] > *end_pfn)
3121 *end_pfn = node_boundary_end_pfn[nid];
3124 void __init push_node_boundaries(unsigned int nid,
3125 unsigned long start_pfn, unsigned long end_pfn) {}
3127 static void __meminit account_node_boundary(unsigned int nid,
3128 unsigned long *start_pfn, unsigned long *end_pfn) {}
3133 * get_pfn_range_for_nid - Return the start and end page frames for a node
3134 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3135 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3136 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3138 * It returns the start and end page frame of a node based on information
3139 * provided by an arch calling add_active_range(). If called for a node
3140 * with no available memory, a warning is printed and the start and end
3143 void __meminit get_pfn_range_for_nid(unsigned int nid,
3144 unsigned long *start_pfn, unsigned long *end_pfn)
3150 for_each_active_range_index_in_nid(i, nid) {
3151 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3152 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3155 if (*start_pfn == -1UL)
3158 /* Push the node boundaries out if requested */
3159 account_node_boundary(nid, start_pfn, end_pfn);
3163 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3164 * assumption is made that zones within a node are ordered in monotonic
3165 * increasing memory addresses so that the "highest" populated zone is used
3167 static void __init find_usable_zone_for_movable(void)
3170 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3171 if (zone_index == ZONE_MOVABLE)
3174 if (arch_zone_highest_possible_pfn[zone_index] >
3175 arch_zone_lowest_possible_pfn[zone_index])
3179 VM_BUG_ON(zone_index == -1);
3180 movable_zone = zone_index;
3184 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3185 * because it is sized independant of architecture. Unlike the other zones,
3186 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3187 * in each node depending on the size of each node and how evenly kernelcore
3188 * is distributed. This helper function adjusts the zone ranges
3189 * provided by the architecture for a given node by using the end of the
3190 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3191 * zones within a node are in order of monotonic increases memory addresses
3193 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3194 unsigned long zone_type,
3195 unsigned long node_start_pfn,
3196 unsigned long node_end_pfn,
3197 unsigned long *zone_start_pfn,
3198 unsigned long *zone_end_pfn)
3200 /* Only adjust if ZONE_MOVABLE is on this node */
3201 if (zone_movable_pfn[nid]) {
3202 /* Size ZONE_MOVABLE */
3203 if (zone_type == ZONE_MOVABLE) {
3204 *zone_start_pfn = zone_movable_pfn[nid];
3205 *zone_end_pfn = min(node_end_pfn,
3206 arch_zone_highest_possible_pfn[movable_zone]);
3208 /* Adjust for ZONE_MOVABLE starting within this range */
3209 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3210 *zone_end_pfn > zone_movable_pfn[nid]) {
3211 *zone_end_pfn = zone_movable_pfn[nid];
3213 /* Check if this whole range is within ZONE_MOVABLE */
3214 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3215 *zone_start_pfn = *zone_end_pfn;
3220 * Return the number of pages a zone spans in a node, including holes
3221 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3223 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3224 unsigned long zone_type,
3225 unsigned long *ignored)
3227 unsigned long node_start_pfn, node_end_pfn;
3228 unsigned long zone_start_pfn, zone_end_pfn;
3230 /* Get the start and end of the node and zone */
3231 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3232 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3233 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3234 adjust_zone_range_for_zone_movable(nid, zone_type,
3235 node_start_pfn, node_end_pfn,
3236 &zone_start_pfn, &zone_end_pfn);
3238 /* Check that this node has pages within the zone's required range */
3239 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3242 /* Move the zone boundaries inside the node if necessary */
3243 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3244 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3246 /* Return the spanned pages */
3247 return zone_end_pfn - zone_start_pfn;
3251 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3252 * then all holes in the requested range will be accounted for.
3254 static unsigned long __meminit __absent_pages_in_range(int nid,
3255 unsigned long range_start_pfn,
3256 unsigned long range_end_pfn)
3259 unsigned long prev_end_pfn = 0, hole_pages = 0;
3260 unsigned long start_pfn;
3262 /* Find the end_pfn of the first active range of pfns in the node */
3263 i = first_active_region_index_in_nid(nid);
3267 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3269 /* Account for ranges before physical memory on this node */
3270 if (early_node_map[i].start_pfn > range_start_pfn)
3271 hole_pages = prev_end_pfn - range_start_pfn;
3273 /* Find all holes for the zone within the node */
3274 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3276 /* No need to continue if prev_end_pfn is outside the zone */
3277 if (prev_end_pfn >= range_end_pfn)
3280 /* Make sure the end of the zone is not within the hole */
3281 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3282 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3284 /* Update the hole size cound and move on */
3285 if (start_pfn > range_start_pfn) {
3286 BUG_ON(prev_end_pfn > start_pfn);
3287 hole_pages += start_pfn - prev_end_pfn;
3289 prev_end_pfn = early_node_map[i].end_pfn;
3292 /* Account for ranges past physical memory on this node */
3293 if (range_end_pfn > prev_end_pfn)
3294 hole_pages += range_end_pfn -
3295 max(range_start_pfn, prev_end_pfn);
3301 * absent_pages_in_range - Return number of page frames in holes within a range
3302 * @start_pfn: The start PFN to start searching for holes
3303 * @end_pfn: The end PFN to stop searching for holes
3305 * It returns the number of pages frames in memory holes within a range.
3307 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3308 unsigned long end_pfn)
3310 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3313 /* Return the number of page frames in holes in a zone on a node */
3314 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3315 unsigned long zone_type,
3316 unsigned long *ignored)
3318 unsigned long node_start_pfn, node_end_pfn;
3319 unsigned long zone_start_pfn, zone_end_pfn;
3321 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3322 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3324 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3327 adjust_zone_range_for_zone_movable(nid, zone_type,
3328 node_start_pfn, node_end_pfn,
3329 &zone_start_pfn, &zone_end_pfn);
3330 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3334 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3335 unsigned long zone_type,
3336 unsigned long *zones_size)
3338 return zones_size[zone_type];
3341 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3342 unsigned long zone_type,
3343 unsigned long *zholes_size)
3348 return zholes_size[zone_type];
3353 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3354 unsigned long *zones_size, unsigned long *zholes_size)
3356 unsigned long realtotalpages, totalpages = 0;
3359 for (i = 0; i < MAX_NR_ZONES; i++)
3360 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3362 pgdat->node_spanned_pages = totalpages;
3364 realtotalpages = totalpages;
3365 for (i = 0; i < MAX_NR_ZONES; i++)
3367 zone_absent_pages_in_node(pgdat->node_id, i,
3369 pgdat->node_present_pages = realtotalpages;
3370 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3374 #ifndef CONFIG_SPARSEMEM
3376 * Calculate the size of the zone->blockflags rounded to an unsigned long
3377 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3378 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3379 * round what is now in bits to nearest long in bits, then return it in
3382 static unsigned long __init usemap_size(unsigned long zonesize)
3384 unsigned long usemapsize;
3386 usemapsize = roundup(zonesize, pageblock_nr_pages);
3387 usemapsize = usemapsize >> pageblock_order;
3388 usemapsize *= NR_PAGEBLOCK_BITS;
3389 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
3391 return usemapsize / 8;
3394 static void __init setup_usemap(struct pglist_data *pgdat,
3395 struct zone *zone, unsigned long zonesize)
3397 unsigned long usemapsize = usemap_size(zonesize);
3398 zone->pageblock_flags = NULL;
3400 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
3403 static void inline setup_usemap(struct pglist_data *pgdat,
3404 struct zone *zone, unsigned long zonesize) {}
3405 #endif /* CONFIG_SPARSEMEM */
3407 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3409 /* Return a sensible default order for the pageblock size. */
3410 static inline int pageblock_default_order(void)
3412 if (HPAGE_SHIFT > PAGE_SHIFT)
3413 return HUGETLB_PAGE_ORDER;
3418 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3419 static inline void __init set_pageblock_order(unsigned int order)
3421 /* Check that pageblock_nr_pages has not already been setup */
3422 if (pageblock_order)
3426 * Assume the largest contiguous order of interest is a huge page.
3427 * This value may be variable depending on boot parameters on IA64
3429 pageblock_order = order;
3431 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3434 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3435 * and pageblock_default_order() are unused as pageblock_order is set
3436 * at compile-time. See include/linux/pageblock-flags.h for the values of
3437 * pageblock_order based on the kernel config
3439 static inline int pageblock_default_order(unsigned int order)
3443 #define set_pageblock_order(x) do {} while (0)
3445 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3448 * Set up the zone data structures:
3449 * - mark all pages reserved
3450 * - mark all memory queues empty
3451 * - clear the memory bitmaps
3453 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
3454 unsigned long *zones_size, unsigned long *zholes_size)
3457 int nid = pgdat->node_id;
3458 unsigned long zone_start_pfn = pgdat->node_start_pfn;
3461 pgdat_resize_init(pgdat);
3462 pgdat->nr_zones = 0;
3463 init_waitqueue_head(&pgdat->kswapd_wait);
3464 pgdat->kswapd_max_order = 0;
3465 pgdat_page_cgroup_init(pgdat);
3467 for (j = 0; j < MAX_NR_ZONES; j++) {
3468 struct zone *zone = pgdat->node_zones + j;
3469 unsigned long size, realsize, memmap_pages;
3472 size = zone_spanned_pages_in_node(nid, j, zones_size);
3473 realsize = size - zone_absent_pages_in_node(nid, j,
3477 * Adjust realsize so that it accounts for how much memory
3478 * is used by this zone for memmap. This affects the watermark
3479 * and per-cpu initialisations
3482 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
3483 if (realsize >= memmap_pages) {
3484 realsize -= memmap_pages;
3487 " %s zone: %lu pages used for memmap\n",
3488 zone_names[j], memmap_pages);
3491 " %s zone: %lu pages exceeds realsize %lu\n",
3492 zone_names[j], memmap_pages, realsize);
3494 /* Account for reserved pages */
3495 if (j == 0 && realsize > dma_reserve) {
3496 realsize -= dma_reserve;
3497 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
3498 zone_names[0], dma_reserve);
3501 if (!is_highmem_idx(j))
3502 nr_kernel_pages += realsize;
3503 nr_all_pages += realsize;
3505 zone->spanned_pages = size;
3506 zone->present_pages = realsize;
3509 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
3511 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
3513 zone->name = zone_names[j];
3514 spin_lock_init(&zone->lock);
3515 spin_lock_init(&zone->lru_lock);
3516 zone_seqlock_init(zone);
3517 zone->zone_pgdat = pgdat;
3519 zone->prev_priority = DEF_PRIORITY;
3521 zone_pcp_init(zone);
3523 INIT_LIST_HEAD(&zone->lru[l].list);
3524 zone->lru[l].nr_scan = 0;
3526 zone->recent_rotated[0] = 0;
3527 zone->recent_rotated[1] = 0;
3528 zone->recent_scanned[0] = 0;
3529 zone->recent_scanned[1] = 0;
3530 zap_zone_vm_stats(zone);
3535 set_pageblock_order(pageblock_default_order());
3536 setup_usemap(pgdat, zone, size);
3537 ret = init_currently_empty_zone(zone, zone_start_pfn,
3538 size, MEMMAP_EARLY);
3540 memmap_init(size, nid, j, zone_start_pfn);
3541 zone_start_pfn += size;
3545 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
3547 /* Skip empty nodes */
3548 if (!pgdat->node_spanned_pages)
3551 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3552 /* ia64 gets its own node_mem_map, before this, without bootmem */
3553 if (!pgdat->node_mem_map) {
3554 unsigned long size, start, end;
3558 * The zone's endpoints aren't required to be MAX_ORDER
3559 * aligned but the node_mem_map endpoints must be in order
3560 * for the buddy allocator to function correctly.
3562 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
3563 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
3564 end = ALIGN(end, MAX_ORDER_NR_PAGES);
3565 size = (end - start) * sizeof(struct page);
3566 map = alloc_remap(pgdat->node_id, size);
3568 map = alloc_bootmem_node(pgdat, size);
3569 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
3571 #ifndef CONFIG_NEED_MULTIPLE_NODES
3573 * With no DISCONTIG, the global mem_map is just set as node 0's
3575 if (pgdat == NODE_DATA(0)) {
3576 mem_map = NODE_DATA(0)->node_mem_map;
3577 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3578 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
3579 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
3580 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3583 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3586 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
3587 unsigned long node_start_pfn, unsigned long *zholes_size)
3589 pg_data_t *pgdat = NODE_DATA(nid);
3591 pgdat->node_id = nid;
3592 pgdat->node_start_pfn = node_start_pfn;
3593 calculate_node_totalpages(pgdat, zones_size, zholes_size);
3595 alloc_node_mem_map(pgdat);
3596 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3597 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3598 nid, (unsigned long)pgdat,
3599 (unsigned long)pgdat->node_mem_map);
3602 free_area_init_core(pgdat, zones_size, zholes_size);
3605 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3607 #if MAX_NUMNODES > 1
3609 * Figure out the number of possible node ids.
3611 static void __init setup_nr_node_ids(void)
3614 unsigned int highest = 0;
3616 for_each_node_mask(node, node_possible_map)
3618 nr_node_ids = highest + 1;
3621 static inline void setup_nr_node_ids(void)
3627 * add_active_range - Register a range of PFNs backed by physical memory
3628 * @nid: The node ID the range resides on
3629 * @start_pfn: The start PFN of the available physical memory
3630 * @end_pfn: The end PFN of the available physical memory
3632 * These ranges are stored in an early_node_map[] and later used by
3633 * free_area_init_nodes() to calculate zone sizes and holes. If the
3634 * range spans a memory hole, it is up to the architecture to ensure
3635 * the memory is not freed by the bootmem allocator. If possible
3636 * the range being registered will be merged with existing ranges.
3638 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
3639 unsigned long end_pfn)
3643 mminit_dprintk(MMINIT_TRACE, "memory_register",
3644 "Entering add_active_range(%d, %#lx, %#lx) "
3645 "%d entries of %d used\n",
3646 nid, start_pfn, end_pfn,
3647 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
3649 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
3651 /* Merge with existing active regions if possible */
3652 for (i = 0; i < nr_nodemap_entries; i++) {
3653 if (early_node_map[i].nid != nid)
3656 /* Skip if an existing region covers this new one */
3657 if (start_pfn >= early_node_map[i].start_pfn &&
3658 end_pfn <= early_node_map[i].end_pfn)
3661 /* Merge forward if suitable */
3662 if (start_pfn <= early_node_map[i].end_pfn &&
3663 end_pfn > early_node_map[i].end_pfn) {
3664 early_node_map[i].end_pfn = end_pfn;
3668 /* Merge backward if suitable */
3669 if (start_pfn < early_node_map[i].end_pfn &&
3670 end_pfn >= early_node_map[i].start_pfn) {
3671 early_node_map[i].start_pfn = start_pfn;
3676 /* Check that early_node_map is large enough */
3677 if (i >= MAX_ACTIVE_REGIONS) {
3678 printk(KERN_CRIT "More than %d memory regions, truncating\n",
3679 MAX_ACTIVE_REGIONS);
3683 early_node_map[i].nid = nid;
3684 early_node_map[i].start_pfn = start_pfn;
3685 early_node_map[i].end_pfn = end_pfn;
3686 nr_nodemap_entries = i + 1;
3690 * remove_active_range - Shrink an existing registered range of PFNs
3691 * @nid: The node id the range is on that should be shrunk
3692 * @start_pfn: The new PFN of the range
3693 * @end_pfn: The new PFN of the range
3695 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
3696 * The map is kept near the end physical page range that has already been
3697 * registered. This function allows an arch to shrink an existing registered
3700 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
3701 unsigned long end_pfn)
3706 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
3707 nid, start_pfn, end_pfn);
3709 /* Find the old active region end and shrink */
3710 for_each_active_range_index_in_nid(i, nid) {
3711 if (early_node_map[i].start_pfn >= start_pfn &&
3712 early_node_map[i].end_pfn <= end_pfn) {
3714 early_node_map[i].start_pfn = 0;
3715 early_node_map[i].end_pfn = 0;
3719 if (early_node_map[i].start_pfn < start_pfn &&
3720 early_node_map[i].end_pfn > start_pfn) {
3721 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
3722 early_node_map[i].end_pfn = start_pfn;
3723 if (temp_end_pfn > end_pfn)
3724 add_active_range(nid, end_pfn, temp_end_pfn);
3727 if (early_node_map[i].start_pfn >= start_pfn &&
3728 early_node_map[i].end_pfn > end_pfn &&
3729 early_node_map[i].start_pfn < end_pfn) {
3730 early_node_map[i].start_pfn = end_pfn;
3738 /* remove the blank ones */
3739 for (i = nr_nodemap_entries - 1; i > 0; i--) {
3740 if (early_node_map[i].nid != nid)
3742 if (early_node_map[i].end_pfn)
3744 /* we found it, get rid of it */
3745 for (j = i; j < nr_nodemap_entries - 1; j++)
3746 memcpy(&early_node_map[j], &early_node_map[j+1],
3747 sizeof(early_node_map[j]));
3748 j = nr_nodemap_entries - 1;
3749 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
3750 nr_nodemap_entries--;
3755 * remove_all_active_ranges - Remove all currently registered regions
3757 * During discovery, it may be found that a table like SRAT is invalid
3758 * and an alternative discovery method must be used. This function removes
3759 * all currently registered regions.
3761 void __init remove_all_active_ranges(void)
3763 memset(early_node_map, 0, sizeof(early_node_map));
3764 nr_nodemap_entries = 0;
3765 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
3766 memset(node_boundary_start_pfn, 0, sizeof(node_boundary_start_pfn));
3767 memset(node_boundary_end_pfn, 0, sizeof(node_boundary_end_pfn));
3768 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
3771 /* Compare two active node_active_regions */
3772 static int __init cmp_node_active_region(const void *a, const void *b)
3774 struct node_active_region *arange = (struct node_active_region *)a;
3775 struct node_active_region *brange = (struct node_active_region *)b;
3777 /* Done this way to avoid overflows */
3778 if (arange->start_pfn > brange->start_pfn)
3780 if (arange->start_pfn < brange->start_pfn)
3786 /* sort the node_map by start_pfn */
3787 static void __init sort_node_map(void)
3789 sort(early_node_map, (size_t)nr_nodemap_entries,
3790 sizeof(struct node_active_region),
3791 cmp_node_active_region, NULL);
3794 /* Find the lowest pfn for a node */
3795 static unsigned long __init find_min_pfn_for_node(int nid)
3798 unsigned long min_pfn = ULONG_MAX;
3800 /* Assuming a sorted map, the first range found has the starting pfn */
3801 for_each_active_range_index_in_nid(i, nid)
3802 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
3804 if (min_pfn == ULONG_MAX) {
3806 "Could not find start_pfn for node %d\n", nid);
3814 * find_min_pfn_with_active_regions - Find the minimum PFN registered
3816 * It returns the minimum PFN based on information provided via
3817 * add_active_range().
3819 unsigned long __init find_min_pfn_with_active_regions(void)
3821 return find_min_pfn_for_node(MAX_NUMNODES);
3825 * early_calculate_totalpages()
3826 * Sum pages in active regions for movable zone.
3827 * Populate N_HIGH_MEMORY for calculating usable_nodes.
3829 static unsigned long __init early_calculate_totalpages(void)
3832 unsigned long totalpages = 0;
3834 for (i = 0; i < nr_nodemap_entries; i++) {
3835 unsigned long pages = early_node_map[i].end_pfn -
3836 early_node_map[i].start_pfn;
3837 totalpages += pages;
3839 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
3845 * Find the PFN the Movable zone begins in each node. Kernel memory
3846 * is spread evenly between nodes as long as the nodes have enough
3847 * memory. When they don't, some nodes will have more kernelcore than
3850 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
3853 unsigned long usable_startpfn;
3854 unsigned long kernelcore_node, kernelcore_remaining;
3855 unsigned long totalpages = early_calculate_totalpages();
3856 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
3859 * If movablecore was specified, calculate what size of
3860 * kernelcore that corresponds so that memory usable for
3861 * any allocation type is evenly spread. If both kernelcore
3862 * and movablecore are specified, then the value of kernelcore
3863 * will be used for required_kernelcore if it's greater than
3864 * what movablecore would have allowed.
3866 if (required_movablecore) {
3867 unsigned long corepages;
3870 * Round-up so that ZONE_MOVABLE is at least as large as what
3871 * was requested by the user
3873 required_movablecore =
3874 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
3875 corepages = totalpages - required_movablecore;
3877 required_kernelcore = max(required_kernelcore, corepages);
3880 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
3881 if (!required_kernelcore)
3884 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
3885 find_usable_zone_for_movable();
3886 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
3889 /* Spread kernelcore memory as evenly as possible throughout nodes */
3890 kernelcore_node = required_kernelcore / usable_nodes;
3891 for_each_node_state(nid, N_HIGH_MEMORY) {
3893 * Recalculate kernelcore_node if the division per node
3894 * now exceeds what is necessary to satisfy the requested
3895 * amount of memory for the kernel
3897 if (required_kernelcore < kernelcore_node)
3898 kernelcore_node = required_kernelcore / usable_nodes;
3901 * As the map is walked, we track how much memory is usable
3902 * by the kernel using kernelcore_remaining. When it is
3903 * 0, the rest of the node is usable by ZONE_MOVABLE
3905 kernelcore_remaining = kernelcore_node;
3907 /* Go through each range of PFNs within this node */
3908 for_each_active_range_index_in_nid(i, nid) {
3909 unsigned long start_pfn, end_pfn;
3910 unsigned long size_pages;
3912 start_pfn = max(early_node_map[i].start_pfn,
3913 zone_movable_pfn[nid]);
3914 end_pfn = early_node_map[i].end_pfn;
3915 if (start_pfn >= end_pfn)
3918 /* Account for what is only usable for kernelcore */
3919 if (start_pfn < usable_startpfn) {
3920 unsigned long kernel_pages;
3921 kernel_pages = min(end_pfn, usable_startpfn)
3924 kernelcore_remaining -= min(kernel_pages,
3925 kernelcore_remaining);
3926 required_kernelcore -= min(kernel_pages,
3927 required_kernelcore);
3929 /* Continue if range is now fully accounted */
3930 if (end_pfn <= usable_startpfn) {
3933 * Push zone_movable_pfn to the end so
3934 * that if we have to rebalance
3935 * kernelcore across nodes, we will
3936 * not double account here
3938 zone_movable_pfn[nid] = end_pfn;
3941 start_pfn = usable_startpfn;
3945 * The usable PFN range for ZONE_MOVABLE is from
3946 * start_pfn->end_pfn. Calculate size_pages as the
3947 * number of pages used as kernelcore
3949 size_pages = end_pfn - start_pfn;
3950 if (size_pages > kernelcore_remaining)
3951 size_pages = kernelcore_remaining;
3952 zone_movable_pfn[nid] = start_pfn + size_pages;
3955 * Some kernelcore has been met, update counts and
3956 * break if the kernelcore for this node has been
3959 required_kernelcore -= min(required_kernelcore,
3961 kernelcore_remaining -= size_pages;
3962 if (!kernelcore_remaining)
3968 * If there is still required_kernelcore, we do another pass with one
3969 * less node in the count. This will push zone_movable_pfn[nid] further
3970 * along on the nodes that still have memory until kernelcore is
3974 if (usable_nodes && required_kernelcore > usable_nodes)
3977 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
3978 for (nid = 0; nid < MAX_NUMNODES; nid++)
3979 zone_movable_pfn[nid] =
3980 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
3983 /* Any regular memory on that node ? */
3984 static void check_for_regular_memory(pg_data_t *pgdat)
3986 #ifdef CONFIG_HIGHMEM
3987 enum zone_type zone_type;
3989 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
3990 struct zone *zone = &pgdat->node_zones[zone_type];
3991 if (zone->present_pages)
3992 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
3998 * free_area_init_nodes - Initialise all pg_data_t and zone data
3999 * @max_zone_pfn: an array of max PFNs for each zone
4001 * This will call free_area_init_node() for each active node in the system.
4002 * Using the page ranges provided by add_active_range(), the size of each
4003 * zone in each node and their holes is calculated. If the maximum PFN
4004 * between two adjacent zones match, it is assumed that the zone is empty.
4005 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4006 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4007 * starts where the previous one ended. For example, ZONE_DMA32 starts
4008 * at arch_max_dma_pfn.
4010 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4015 /* Sort early_node_map as initialisation assumes it is sorted */
4018 /* Record where the zone boundaries are */
4019 memset(arch_zone_lowest_possible_pfn, 0,
4020 sizeof(arch_zone_lowest_possible_pfn));
4021 memset(arch_zone_highest_possible_pfn, 0,
4022 sizeof(arch_zone_highest_possible_pfn));
4023 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4024 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4025 for (i = 1; i < MAX_NR_ZONES; i++) {
4026 if (i == ZONE_MOVABLE)
4028 arch_zone_lowest_possible_pfn[i] =
4029 arch_zone_highest_possible_pfn[i-1];
4030 arch_zone_highest_possible_pfn[i] =
4031 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4033 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4034 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4036 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4037 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4038 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4040 /* Print out the zone ranges */
4041 printk("Zone PFN ranges:\n");
4042 for (i = 0; i < MAX_NR_ZONES; i++) {
4043 if (i == ZONE_MOVABLE)
4045 printk(" %-8s %0#10lx -> %0#10lx\n",
4047 arch_zone_lowest_possible_pfn[i],
4048 arch_zone_highest_possible_pfn[i]);
4051 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4052 printk("Movable zone start PFN for each node\n");
4053 for (i = 0; i < MAX_NUMNODES; i++) {
4054 if (zone_movable_pfn[i])
4055 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4058 /* Print out the early_node_map[] */
4059 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4060 for (i = 0; i < nr_nodemap_entries; i++)
4061 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4062 early_node_map[i].start_pfn,
4063 early_node_map[i].end_pfn);
4065 /* Initialise every node */
4066 mminit_verify_pageflags_layout();
4067 setup_nr_node_ids();
4068 for_each_online_node(nid) {
4069 pg_data_t *pgdat = NODE_DATA(nid);
4070 free_area_init_node(nid, NULL,
4071 find_min_pfn_for_node(nid), NULL);
4073 /* Any memory on that node */
4074 if (pgdat->node_present_pages)
4075 node_set_state(nid, N_HIGH_MEMORY);
4076 check_for_regular_memory(pgdat);
4080 static int __init cmdline_parse_core(char *p, unsigned long *core)
4082 unsigned long long coremem;
4086 coremem = memparse(p, &p);
4087 *core = coremem >> PAGE_SHIFT;
4089 /* Paranoid check that UL is enough for the coremem value */
4090 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4096 * kernelcore=size sets the amount of memory for use for allocations that
4097 * cannot be reclaimed or migrated.
4099 static int __init cmdline_parse_kernelcore(char *p)
4101 return cmdline_parse_core(p, &required_kernelcore);
4105 * movablecore=size sets the amount of memory for use for allocations that
4106 * can be reclaimed or migrated.
4108 static int __init cmdline_parse_movablecore(char *p)
4110 return cmdline_parse_core(p, &required_movablecore);
4113 early_param("kernelcore", cmdline_parse_kernelcore);
4114 early_param("movablecore", cmdline_parse_movablecore);
4116 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4119 * set_dma_reserve - set the specified number of pages reserved in the first zone
4120 * @new_dma_reserve: The number of pages to mark reserved
4122 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4123 * In the DMA zone, a significant percentage may be consumed by kernel image
4124 * and other unfreeable allocations which can skew the watermarks badly. This
4125 * function may optionally be used to account for unfreeable pages in the
4126 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4127 * smaller per-cpu batchsize.
4129 void __init set_dma_reserve(unsigned long new_dma_reserve)
4131 dma_reserve = new_dma_reserve;
4134 #ifndef CONFIG_NEED_MULTIPLE_NODES
4135 struct pglist_data __refdata contig_page_data = { .bdata = &bootmem_node_data[0] };
4136 EXPORT_SYMBOL(contig_page_data);
4139 void __init free_area_init(unsigned long *zones_size)
4141 free_area_init_node(0, zones_size,
4142 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4145 static int page_alloc_cpu_notify(struct notifier_block *self,
4146 unsigned long action, void *hcpu)
4148 int cpu = (unsigned long)hcpu;
4150 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4154 * Spill the event counters of the dead processor
4155 * into the current processors event counters.
4156 * This artificially elevates the count of the current
4159 vm_events_fold_cpu(cpu);
4162 * Zero the differential counters of the dead processor
4163 * so that the vm statistics are consistent.
4165 * This is only okay since the processor is dead and cannot
4166 * race with what we are doing.
4168 refresh_cpu_vm_stats(cpu);
4173 void __init page_alloc_init(void)
4175 hotcpu_notifier(page_alloc_cpu_notify, 0);
4179 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4180 * or min_free_kbytes changes.
4182 static void calculate_totalreserve_pages(void)
4184 struct pglist_data *pgdat;
4185 unsigned long reserve_pages = 0;
4186 enum zone_type i, j;
4188 for_each_online_pgdat(pgdat) {
4189 for (i = 0; i < MAX_NR_ZONES; i++) {
4190 struct zone *zone = pgdat->node_zones + i;
4191 unsigned long max = 0;
4193 /* Find valid and maximum lowmem_reserve in the zone */
4194 for (j = i; j < MAX_NR_ZONES; j++) {
4195 if (zone->lowmem_reserve[j] > max)
4196 max = zone->lowmem_reserve[j];
4199 /* we treat pages_high as reserved pages. */
4200 max += zone->pages_high;
4202 if (max > zone->present_pages)
4203 max = zone->present_pages;
4204 reserve_pages += max;
4207 totalreserve_pages = reserve_pages;
4211 * setup_per_zone_lowmem_reserve - called whenever
4212 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4213 * has a correct pages reserved value, so an adequate number of
4214 * pages are left in the zone after a successful __alloc_pages().
4216 static void setup_per_zone_lowmem_reserve(void)
4218 struct pglist_data *pgdat;
4219 enum zone_type j, idx;
4221 for_each_online_pgdat(pgdat) {
4222 for (j = 0; j < MAX_NR_ZONES; j++) {
4223 struct zone *zone = pgdat->node_zones + j;
4224 unsigned long present_pages = zone->present_pages;
4226 zone->lowmem_reserve[j] = 0;
4230 struct zone *lower_zone;
4234 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4235 sysctl_lowmem_reserve_ratio[idx] = 1;
4237 lower_zone = pgdat->node_zones + idx;
4238 lower_zone->lowmem_reserve[j] = present_pages /
4239 sysctl_lowmem_reserve_ratio[idx];
4240 present_pages += lower_zone->present_pages;
4245 /* update totalreserve_pages */
4246 calculate_totalreserve_pages();
4250 * setup_per_zone_pages_min - called when min_free_kbytes changes.
4252 * Ensures that the pages_{min,low,high} values for each zone are set correctly
4253 * with respect to min_free_kbytes.
4255 void setup_per_zone_pages_min(void)
4257 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4258 unsigned long lowmem_pages = 0;
4260 unsigned long flags;
4262 /* Calculate total number of !ZONE_HIGHMEM pages */
4263 for_each_zone(zone) {
4264 if (!is_highmem(zone))
4265 lowmem_pages += zone->present_pages;
4268 for_each_zone(zone) {
4271 spin_lock_irqsave(&zone->lock, flags);
4272 tmp = (u64)pages_min * zone->present_pages;
4273 do_div(tmp, lowmem_pages);
4274 if (is_highmem(zone)) {
4276 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4277 * need highmem pages, so cap pages_min to a small
4280 * The (pages_high-pages_low) and (pages_low-pages_min)
4281 * deltas controls asynch page reclaim, and so should
4282 * not be capped for highmem.
4286 min_pages = zone->present_pages / 1024;
4287 if (min_pages < SWAP_CLUSTER_MAX)
4288 min_pages = SWAP_CLUSTER_MAX;
4289 if (min_pages > 128)
4291 zone->pages_min = min_pages;
4294 * If it's a lowmem zone, reserve a number of pages
4295 * proportionate to the zone's size.
4297 zone->pages_min = tmp;
4300 zone->pages_low = zone->pages_min + (tmp >> 2);
4301 zone->pages_high = zone->pages_min + (tmp >> 1);
4302 setup_zone_migrate_reserve(zone);
4303 spin_unlock_irqrestore(&zone->lock, flags);
4306 /* update totalreserve_pages */
4307 calculate_totalreserve_pages();
4311 * setup_per_zone_inactive_ratio - called when min_free_kbytes changes.
4313 * The inactive anon list should be small enough that the VM never has to
4314 * do too much work, but large enough that each inactive page has a chance
4315 * to be referenced again before it is swapped out.
4317 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4318 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4319 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4320 * the anonymous pages are kept on the inactive list.
4323 * memory ratio inactive anon
4324 * -------------------------------------
4333 static void setup_per_zone_inactive_ratio(void)
4337 for_each_zone(zone) {
4338 unsigned int gb, ratio;
4340 /* Zone size in gigabytes */
4341 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4342 ratio = int_sqrt(10 * gb);
4346 zone->inactive_ratio = ratio;
4351 * Initialise min_free_kbytes.
4353 * For small machines we want it small (128k min). For large machines
4354 * we want it large (64MB max). But it is not linear, because network
4355 * bandwidth does not increase linearly with machine size. We use
4357 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4358 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4374 static int __init init_per_zone_pages_min(void)
4376 unsigned long lowmem_kbytes;
4378 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4380 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4381 if (min_free_kbytes < 128)
4382 min_free_kbytes = 128;
4383 if (min_free_kbytes > 65536)
4384 min_free_kbytes = 65536;
4385 setup_per_zone_pages_min();
4386 setup_per_zone_lowmem_reserve();
4387 setup_per_zone_inactive_ratio();
4390 module_init(init_per_zone_pages_min)
4393 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4394 * that we can call two helper functions whenever min_free_kbytes
4397 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
4398 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4400 proc_dointvec(table, write, file, buffer, length, ppos);
4402 setup_per_zone_pages_min();
4407 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
4408 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4413 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4418 zone->min_unmapped_pages = (zone->present_pages *
4419 sysctl_min_unmapped_ratio) / 100;
4423 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
4424 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4429 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4434 zone->min_slab_pages = (zone->present_pages *
4435 sysctl_min_slab_ratio) / 100;
4441 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4442 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4443 * whenever sysctl_lowmem_reserve_ratio changes.
4445 * The reserve ratio obviously has absolutely no relation with the
4446 * pages_min watermarks. The lowmem reserve ratio can only make sense
4447 * if in function of the boot time zone sizes.
4449 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
4450 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4452 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4453 setup_per_zone_lowmem_reserve();
4458 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4459 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4460 * can have before it gets flushed back to buddy allocator.
4463 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
4464 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4470 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4471 if (!write || (ret == -EINVAL))
4473 for_each_zone(zone) {
4474 for_each_online_cpu(cpu) {
4476 high = zone->present_pages / percpu_pagelist_fraction;
4477 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
4483 int hashdist = HASHDIST_DEFAULT;
4486 static int __init set_hashdist(char *str)
4490 hashdist = simple_strtoul(str, &str, 0);
4493 __setup("hashdist=", set_hashdist);
4497 * allocate a large system hash table from bootmem
4498 * - it is assumed that the hash table must contain an exact power-of-2
4499 * quantity of entries
4500 * - limit is the number of hash buckets, not the total allocation size
4502 void *__init alloc_large_system_hash(const char *tablename,
4503 unsigned long bucketsize,
4504 unsigned long numentries,
4507 unsigned int *_hash_shift,
4508 unsigned int *_hash_mask,
4509 unsigned long limit)
4511 unsigned long long max = limit;
4512 unsigned long log2qty, size;
4515 /* allow the kernel cmdline to have a say */
4517 /* round applicable memory size up to nearest megabyte */
4518 numentries = nr_kernel_pages;
4519 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
4520 numentries >>= 20 - PAGE_SHIFT;
4521 numentries <<= 20 - PAGE_SHIFT;
4523 /* limit to 1 bucket per 2^scale bytes of low memory */
4524 if (scale > PAGE_SHIFT)
4525 numentries >>= (scale - PAGE_SHIFT);
4527 numentries <<= (PAGE_SHIFT - scale);
4529 /* Make sure we've got at least a 0-order allocation.. */
4530 if (unlikely((numentries * bucketsize) < PAGE_SIZE))
4531 numentries = PAGE_SIZE / bucketsize;
4533 numentries = roundup_pow_of_two(numentries);
4535 /* limit allocation size to 1/16 total memory by default */
4537 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
4538 do_div(max, bucketsize);
4541 if (numentries > max)
4544 log2qty = ilog2(numentries);
4547 size = bucketsize << log2qty;
4548 if (flags & HASH_EARLY)
4549 table = alloc_bootmem_nopanic(size);
4551 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
4553 unsigned long order = get_order(size);
4554 table = (void*) __get_free_pages(GFP_ATOMIC, order);
4556 * If bucketsize is not a power-of-two, we may free
4557 * some pages at the end of hash table.
4560 unsigned long alloc_end = (unsigned long)table +
4561 (PAGE_SIZE << order);
4562 unsigned long used = (unsigned long)table +
4564 split_page(virt_to_page(table), order);
4565 while (used < alloc_end) {
4571 } while (!table && size > PAGE_SIZE && --log2qty);
4574 panic("Failed to allocate %s hash table\n", tablename);
4576 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
4579 ilog2(size) - PAGE_SHIFT,
4583 *_hash_shift = log2qty;
4585 *_hash_mask = (1 << log2qty) - 1;
4590 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4591 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
4594 #ifdef CONFIG_SPARSEMEM
4595 return __pfn_to_section(pfn)->pageblock_flags;
4597 return zone->pageblock_flags;
4598 #endif /* CONFIG_SPARSEMEM */
4601 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
4603 #ifdef CONFIG_SPARSEMEM
4604 pfn &= (PAGES_PER_SECTION-1);
4605 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4607 pfn = pfn - zone->zone_start_pfn;
4608 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4609 #endif /* CONFIG_SPARSEMEM */
4613 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4614 * @page: The page within the block of interest
4615 * @start_bitidx: The first bit of interest to retrieve
4616 * @end_bitidx: The last bit of interest
4617 * returns pageblock_bits flags
4619 unsigned long get_pageblock_flags_group(struct page *page,
4620 int start_bitidx, int end_bitidx)
4623 unsigned long *bitmap;
4624 unsigned long pfn, bitidx;
4625 unsigned long flags = 0;
4626 unsigned long value = 1;
4628 zone = page_zone(page);
4629 pfn = page_to_pfn(page);
4630 bitmap = get_pageblock_bitmap(zone, pfn);
4631 bitidx = pfn_to_bitidx(zone, pfn);
4633 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4634 if (test_bit(bitidx + start_bitidx, bitmap))
4641 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4642 * @page: The page within the block of interest
4643 * @start_bitidx: The first bit of interest
4644 * @end_bitidx: The last bit of interest
4645 * @flags: The flags to set
4647 void set_pageblock_flags_group(struct page *page, unsigned long flags,
4648 int start_bitidx, int end_bitidx)
4651 unsigned long *bitmap;
4652 unsigned long pfn, bitidx;
4653 unsigned long value = 1;
4655 zone = page_zone(page);
4656 pfn = page_to_pfn(page);
4657 bitmap = get_pageblock_bitmap(zone, pfn);
4658 bitidx = pfn_to_bitidx(zone, pfn);
4659 VM_BUG_ON(pfn < zone->zone_start_pfn);
4660 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
4662 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4664 __set_bit(bitidx + start_bitidx, bitmap);
4666 __clear_bit(bitidx + start_bitidx, bitmap);
4670 * This is designed as sub function...plz see page_isolation.c also.
4671 * set/clear page block's type to be ISOLATE.
4672 * page allocater never alloc memory from ISOLATE block.
4675 int set_migratetype_isolate(struct page *page)
4678 unsigned long flags;
4681 zone = page_zone(page);
4682 spin_lock_irqsave(&zone->lock, flags);
4684 * In future, more migrate types will be able to be isolation target.
4686 if (get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
4688 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
4689 move_freepages_block(zone, page, MIGRATE_ISOLATE);
4692 spin_unlock_irqrestore(&zone->lock, flags);
4698 void unset_migratetype_isolate(struct page *page)
4701 unsigned long flags;
4702 zone = page_zone(page);
4703 spin_lock_irqsave(&zone->lock, flags);
4704 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
4706 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4707 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4709 spin_unlock_irqrestore(&zone->lock, flags);
4712 #ifdef CONFIG_MEMORY_HOTREMOVE
4714 * All pages in the range must be isolated before calling this.
4717 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
4723 unsigned long flags;
4724 /* find the first valid pfn */
4725 for (pfn = start_pfn; pfn < end_pfn; pfn++)
4730 zone = page_zone(pfn_to_page(pfn));
4731 spin_lock_irqsave(&zone->lock, flags);
4733 while (pfn < end_pfn) {
4734 if (!pfn_valid(pfn)) {
4738 page = pfn_to_page(pfn);
4739 BUG_ON(page_count(page));
4740 BUG_ON(!PageBuddy(page));
4741 order = page_order(page);
4742 #ifdef CONFIG_DEBUG_VM
4743 printk(KERN_INFO "remove from free list %lx %d %lx\n",
4744 pfn, 1 << order, end_pfn);
4746 list_del(&page->lru);
4747 rmv_page_order(page);
4748 zone->free_area[order].nr_free--;
4749 __mod_zone_page_state(zone, NR_FREE_PAGES,
4751 for (i = 0; i < (1 << order); i++)
4752 SetPageReserved((page+i));
4753 pfn += (1 << order);
4755 spin_unlock_irqrestore(&zone->lock, flags);