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/bootmem.h>
23 #include <linux/compiler.h>
24 #include <linux/kernel.h>
25 #include <linux/module.h>
26 #include <linux/suspend.h>
27 #include <linux/pagevec.h>
28 #include <linux/blkdev.h>
29 #include <linux/slab.h>
30 #include <linux/notifier.h>
31 #include <linux/topology.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/memory_hotplug.h>
36 #include <linux/nodemask.h>
37 #include <linux/vmalloc.h>
38 #include <linux/mempolicy.h>
39 #include <linux/stop_machine.h>
40 #include <linux/sort.h>
41 #include <linux/pfn.h>
42 #include <linux/backing-dev.h>
43 #include <linux/fault-inject.h>
45 #include <asm/tlbflush.h>
46 #include <asm/div64.h>
50 * Array of node states.
52 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
53 [N_POSSIBLE] = NODE_MASK_ALL,
54 [N_ONLINE] = { { [0] = 1UL } },
56 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
58 [N_HIGH_MEMORY] = { { [0] = 1UL } },
60 [N_CPU] = { { [0] = 1UL } },
63 EXPORT_SYMBOL(node_states);
65 unsigned long totalram_pages __read_mostly;
66 unsigned long totalreserve_pages __read_mostly;
68 int percpu_pagelist_fraction;
70 static void __free_pages_ok(struct page *page, unsigned int order);
73 * results with 256, 32 in the lowmem_reserve sysctl:
74 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
75 * 1G machine -> (16M dma, 784M normal, 224M high)
76 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
77 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
78 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
80 * TBD: should special case ZONE_DMA32 machines here - in those we normally
81 * don't need any ZONE_NORMAL reservation
83 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
84 #ifdef CONFIG_ZONE_DMA
87 #ifdef CONFIG_ZONE_DMA32
96 EXPORT_SYMBOL(totalram_pages);
98 static char * const zone_names[MAX_NR_ZONES] = {
99 #ifdef CONFIG_ZONE_DMA
102 #ifdef CONFIG_ZONE_DMA32
106 #ifdef CONFIG_HIGHMEM
112 int min_free_kbytes = 1024;
114 unsigned long __meminitdata nr_kernel_pages;
115 unsigned long __meminitdata nr_all_pages;
116 static unsigned long __meminitdata dma_reserve;
118 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
120 * MAX_ACTIVE_REGIONS determines the maxmimum number of distinct
121 * ranges of memory (RAM) that may be registered with add_active_range().
122 * Ranges passed to add_active_range() will be merged if possible
123 * so the number of times add_active_range() can be called is
124 * related to the number of nodes and the number of holes
126 #ifdef CONFIG_MAX_ACTIVE_REGIONS
127 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
128 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
130 #if MAX_NUMNODES >= 32
131 /* If there can be many nodes, allow up to 50 holes per node */
132 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
134 /* By default, allow up to 256 distinct regions */
135 #define MAX_ACTIVE_REGIONS 256
139 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
140 static int __meminitdata nr_nodemap_entries;
141 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
142 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
143 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
144 static unsigned long __meminitdata node_boundary_start_pfn[MAX_NUMNODES];
145 static unsigned long __meminitdata node_boundary_end_pfn[MAX_NUMNODES];
146 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
147 unsigned long __initdata required_kernelcore;
148 unsigned long __initdata required_movablecore;
149 unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
151 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
153 EXPORT_SYMBOL(movable_zone);
154 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
157 int nr_node_ids __read_mostly = MAX_NUMNODES;
158 EXPORT_SYMBOL(nr_node_ids);
161 int page_group_by_mobility_disabled __read_mostly;
163 static inline int get_pageblock_migratetype(struct page *page)
165 if (unlikely(page_group_by_mobility_disabled))
166 return MIGRATE_UNMOVABLE;
168 return get_pageblock_flags_group(page, PB_migrate, PB_migrate_end);
171 static void set_pageblock_migratetype(struct page *page, int migratetype)
173 set_pageblock_flags_group(page, (unsigned long)migratetype,
174 PB_migrate, PB_migrate_end);
177 static inline int allocflags_to_migratetype(gfp_t gfp_flags, int order)
179 WARN_ON((gfp_flags & GFP_MOVABLE_MASK) == GFP_MOVABLE_MASK);
181 if (unlikely(page_group_by_mobility_disabled))
182 return MIGRATE_UNMOVABLE;
184 /* Cluster high-order atomic allocations together */
185 if (unlikely(order > 0) &&
186 (!(gfp_flags & __GFP_WAIT) || in_interrupt()))
187 return MIGRATE_HIGHATOMIC;
189 /* Cluster based on mobility */
190 return (((gfp_flags & __GFP_MOVABLE) != 0) << 1) |
191 ((gfp_flags & __GFP_RECLAIMABLE) != 0);
194 #ifdef CONFIG_DEBUG_VM
195 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
199 unsigned long pfn = page_to_pfn(page);
202 seq = zone_span_seqbegin(zone);
203 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
205 else if (pfn < zone->zone_start_pfn)
207 } while (zone_span_seqretry(zone, seq));
212 static int page_is_consistent(struct zone *zone, struct page *page)
214 if (!pfn_valid_within(page_to_pfn(page)))
216 if (zone != page_zone(page))
222 * Temporary debugging check for pages not lying within a given zone.
224 static int bad_range(struct zone *zone, struct page *page)
226 if (page_outside_zone_boundaries(zone, page))
228 if (!page_is_consistent(zone, page))
234 static inline int bad_range(struct zone *zone, struct page *page)
240 static void bad_page(struct page *page)
242 printk(KERN_EMERG "Bad page state in process '%s'\n"
243 KERN_EMERG "page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d\n"
244 KERN_EMERG "Trying to fix it up, but a reboot is needed\n"
245 KERN_EMERG "Backtrace:\n",
246 current->comm, page, (int)(2*sizeof(unsigned long)),
247 (unsigned long)page->flags, page->mapping,
248 page_mapcount(page), page_count(page));
250 page->flags &= ~(1 << PG_lru |
260 set_page_count(page, 0);
261 reset_page_mapcount(page);
262 page->mapping = NULL;
263 add_taint(TAINT_BAD_PAGE);
267 * Higher-order pages are called "compound pages". They are structured thusly:
269 * The first PAGE_SIZE page is called the "head page".
271 * The remaining PAGE_SIZE pages are called "tail pages".
273 * All pages have PG_compound set. All pages have their ->private pointing at
274 * the head page (even the head page has this).
276 * The first tail page's ->lru.next holds the address of the compound page's
277 * put_page() function. Its ->lru.prev holds the order of allocation.
278 * This usage means that zero-order pages may not be compound.
281 static void free_compound_page(struct page *page)
283 __free_pages_ok(page, compound_order(page));
286 static void prep_compound_page(struct page *page, unsigned long order)
289 int nr_pages = 1 << order;
291 set_compound_page_dtor(page, free_compound_page);
292 set_compound_order(page, order);
294 for (i = 1; i < nr_pages; i++) {
295 struct page *p = page + i;
298 p->first_page = page;
302 static void destroy_compound_page(struct page *page, unsigned long order)
305 int nr_pages = 1 << order;
307 if (unlikely(compound_order(page) != order))
310 if (unlikely(!PageHead(page)))
312 __ClearPageHead(page);
313 for (i = 1; i < nr_pages; i++) {
314 struct page *p = page + i;
316 if (unlikely(!PageTail(p) |
317 (p->first_page != page)))
323 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
327 VM_BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
329 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
330 * and __GFP_HIGHMEM from hard or soft interrupt context.
332 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
333 for (i = 0; i < (1 << order); i++)
334 clear_highpage(page + i);
338 * function for dealing with page's order in buddy system.
339 * zone->lock is already acquired when we use these.
340 * So, we don't need atomic page->flags operations here.
342 static inline unsigned long page_order(struct page *page)
344 return page_private(page);
347 static inline void set_page_order(struct page *page, int order)
349 set_page_private(page, order);
350 __SetPageBuddy(page);
353 static inline void rmv_page_order(struct page *page)
355 __ClearPageBuddy(page);
356 set_page_private(page, 0);
360 * Locate the struct page for both the matching buddy in our
361 * pair (buddy1) and the combined O(n+1) page they form (page).
363 * 1) Any buddy B1 will have an order O twin B2 which satisfies
364 * the following equation:
366 * For example, if the starting buddy (buddy2) is #8 its order
368 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
370 * 2) Any buddy B will have an order O+1 parent P which
371 * satisfies the following equation:
374 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
376 static inline struct page *
377 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
379 unsigned long buddy_idx = page_idx ^ (1 << order);
381 return page + (buddy_idx - page_idx);
384 static inline unsigned long
385 __find_combined_index(unsigned long page_idx, unsigned int order)
387 return (page_idx & ~(1 << order));
391 * This function checks whether a page is free && is the buddy
392 * we can do coalesce a page and its buddy if
393 * (a) the buddy is not in a hole &&
394 * (b) the buddy is in the buddy system &&
395 * (c) a page and its buddy have the same order &&
396 * (d) a page and its buddy are in the same zone.
398 * For recording whether a page is in the buddy system, we use PG_buddy.
399 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
401 * For recording page's order, we use page_private(page).
403 static inline int page_is_buddy(struct page *page, struct page *buddy,
406 if (!pfn_valid_within(page_to_pfn(buddy)))
409 if (page_zone_id(page) != page_zone_id(buddy))
412 if (PageBuddy(buddy) && page_order(buddy) == order) {
413 BUG_ON(page_count(buddy) != 0);
420 * Freeing function for a buddy system allocator.
422 * The concept of a buddy system is to maintain direct-mapped table
423 * (containing bit values) for memory blocks of various "orders".
424 * The bottom level table contains the map for the smallest allocatable
425 * units of memory (here, pages), and each level above it describes
426 * pairs of units from the levels below, hence, "buddies".
427 * At a high level, all that happens here is marking the table entry
428 * at the bottom level available, and propagating the changes upward
429 * as necessary, plus some accounting needed to play nicely with other
430 * parts of the VM system.
431 * At each level, we keep a list of pages, which are heads of continuous
432 * free pages of length of (1 << order) and marked with PG_buddy. Page's
433 * order is recorded in page_private(page) field.
434 * So when we are allocating or freeing one, we can derive the state of the
435 * other. That is, if we allocate a small block, and both were
436 * free, the remainder of the region must be split into blocks.
437 * If a block is freed, and its buddy is also free, then this
438 * triggers coalescing into a block of larger size.
443 static inline void __free_one_page(struct page *page,
444 struct zone *zone, unsigned int order)
446 unsigned long page_idx;
447 int order_size = 1 << order;
448 int migratetype = get_pageblock_migratetype(page);
450 if (unlikely(PageCompound(page)))
451 destroy_compound_page(page, order);
453 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
455 VM_BUG_ON(page_idx & (order_size - 1));
456 VM_BUG_ON(bad_range(zone, page));
458 __mod_zone_page_state(zone, NR_FREE_PAGES, order_size);
459 while (order < MAX_ORDER-1) {
460 unsigned long combined_idx;
463 buddy = __page_find_buddy(page, page_idx, order);
464 if (!page_is_buddy(page, buddy, order))
465 break; /* Move the buddy up one level. */
467 list_del(&buddy->lru);
468 zone->free_area[order].nr_free--;
469 rmv_page_order(buddy);
470 combined_idx = __find_combined_index(page_idx, order);
471 page = page + (combined_idx - page_idx);
472 page_idx = combined_idx;
475 set_page_order(page, order);
477 &zone->free_area[order].free_list[migratetype]);
478 zone->free_area[order].nr_free++;
481 static inline int free_pages_check(struct page *page)
483 if (unlikely(page_mapcount(page) |
484 (page->mapping != NULL) |
485 (page_count(page) != 0) |
498 __ClearPageDirty(page);
500 * For now, we report if PG_reserved was found set, but do not
501 * clear it, and do not free the page. But we shall soon need
502 * to do more, for when the ZERO_PAGE count wraps negative.
504 return PageReserved(page);
508 * Frees a list of pages.
509 * Assumes all pages on list are in same zone, and of same order.
510 * count is the number of pages to free.
512 * If the zone was previously in an "all pages pinned" state then look to
513 * see if this freeing clears that state.
515 * And clear the zone's pages_scanned counter, to hold off the "all pages are
516 * pinned" detection logic.
518 static void free_pages_bulk(struct zone *zone, int count,
519 struct list_head *list, int order)
521 spin_lock(&zone->lock);
522 zone->all_unreclaimable = 0;
523 zone->pages_scanned = 0;
527 VM_BUG_ON(list_empty(list));
528 page = list_entry(list->prev, struct page, lru);
529 /* have to delete it as __free_one_page list manipulates */
530 list_del(&page->lru);
531 __free_one_page(page, zone, order);
533 spin_unlock(&zone->lock);
536 static void free_one_page(struct zone *zone, struct page *page, int order)
538 spin_lock(&zone->lock);
539 zone->all_unreclaimable = 0;
540 zone->pages_scanned = 0;
541 __free_one_page(page, zone, order);
542 spin_unlock(&zone->lock);
545 static void __free_pages_ok(struct page *page, unsigned int order)
551 for (i = 0 ; i < (1 << order) ; ++i)
552 reserved += free_pages_check(page + i);
556 if (!PageHighMem(page))
557 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
558 arch_free_page(page, order);
559 kernel_map_pages(page, 1 << order, 0);
561 local_irq_save(flags);
562 __count_vm_events(PGFREE, 1 << order);
563 free_one_page(page_zone(page), page, order);
564 local_irq_restore(flags);
568 * permit the bootmem allocator to evade page validation on high-order frees
570 void fastcall __init __free_pages_bootmem(struct page *page, unsigned int order)
573 __ClearPageReserved(page);
574 set_page_count(page, 0);
575 set_page_refcounted(page);
581 for (loop = 0; loop < BITS_PER_LONG; loop++) {
582 struct page *p = &page[loop];
584 if (loop + 1 < BITS_PER_LONG)
586 __ClearPageReserved(p);
587 set_page_count(p, 0);
590 set_page_refcounted(page);
591 __free_pages(page, order);
597 * The order of subdivision here is critical for the IO subsystem.
598 * Please do not alter this order without good reasons and regression
599 * testing. Specifically, as large blocks of memory are subdivided,
600 * the order in which smaller blocks are delivered depends on the order
601 * they're subdivided in this function. This is the primary factor
602 * influencing the order in which pages are delivered to the IO
603 * subsystem according to empirical testing, and this is also justified
604 * by considering the behavior of a buddy system containing a single
605 * large block of memory acted on by a series of small allocations.
606 * This behavior is a critical factor in sglist merging's success.
610 static inline void expand(struct zone *zone, struct page *page,
611 int low, int high, struct free_area *area,
614 unsigned long size = 1 << high;
620 VM_BUG_ON(bad_range(zone, &page[size]));
621 list_add(&page[size].lru, &area->free_list[migratetype]);
623 set_page_order(&page[size], high);
628 * This page is about to be returned from the page allocator
630 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
632 if (unlikely(page_mapcount(page) |
633 (page->mapping != NULL) |
634 (page_count(page) != 0) |
649 * For now, we report if PG_reserved was found set, but do not
650 * clear it, and do not allocate the page: as a safety net.
652 if (PageReserved(page))
655 page->flags &= ~(1 << PG_uptodate | 1 << PG_error | 1 << PG_readahead |
656 1 << PG_referenced | 1 << PG_arch_1 |
657 1 << PG_owner_priv_1 | 1 << PG_mappedtodisk);
658 set_page_private(page, 0);
659 set_page_refcounted(page);
661 arch_alloc_page(page, order);
662 kernel_map_pages(page, 1 << order, 1);
664 if (gfp_flags & __GFP_ZERO)
665 prep_zero_page(page, order, gfp_flags);
667 if (order && (gfp_flags & __GFP_COMP))
668 prep_compound_page(page, order);
674 * Go through the free lists for the given migratetype and remove
675 * the smallest available page from the freelists
677 static struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
680 unsigned int current_order;
681 struct free_area * area;
684 /* Find a page of the appropriate size in the preferred list */
685 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
686 area = &(zone->free_area[current_order]);
687 if (list_empty(&area->free_list[migratetype]))
690 page = list_entry(area->free_list[migratetype].next,
692 list_del(&page->lru);
693 rmv_page_order(page);
695 __mod_zone_page_state(zone, NR_FREE_PAGES, - (1UL << order));
696 expand(zone, page, order, current_order, area, migratetype);
705 * This array describes the order lists are fallen back to when
706 * the free lists for the desirable migrate type are depleted
708 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
709 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_HIGHATOMIC, MIGRATE_RESERVE },
710 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_HIGHATOMIC, MIGRATE_RESERVE },
711 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_HIGHATOMIC, MIGRATE_RESERVE },
712 [MIGRATE_HIGHATOMIC] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
713 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
717 * Move the free pages in a range to the free lists of the requested type.
718 * Note that start_page and end_pages are not aligned in a MAX_ORDER_NR_PAGES
719 * boundary. If alignment is required, use move_freepages_block()
721 int move_freepages(struct zone *zone,
722 struct page *start_page, struct page *end_page,
727 int blocks_moved = 0;
729 #ifndef CONFIG_HOLES_IN_ZONE
731 * page_zone is not safe to call in this context when
732 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
733 * anyway as we check zone boundaries in move_freepages_block().
734 * Remove at a later date when no bug reports exist related to
735 * grouping pages by mobility
737 BUG_ON(page_zone(start_page) != page_zone(end_page));
740 for (page = start_page; page <= end_page;) {
741 if (!pfn_valid_within(page_to_pfn(page))) {
746 if (!PageBuddy(page)) {
751 order = page_order(page);
752 list_del(&page->lru);
754 &zone->free_area[order].free_list[migratetype]);
762 int move_freepages_block(struct zone *zone, struct page *page, int migratetype)
764 unsigned long start_pfn, end_pfn;
765 struct page *start_page, *end_page;
767 start_pfn = page_to_pfn(page);
768 start_pfn = start_pfn & ~(MAX_ORDER_NR_PAGES-1);
769 start_page = pfn_to_page(start_pfn);
770 end_page = start_page + MAX_ORDER_NR_PAGES - 1;
771 end_pfn = start_pfn + MAX_ORDER_NR_PAGES - 1;
773 /* Do not cross zone boundaries */
774 if (start_pfn < zone->zone_start_pfn)
776 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
779 return move_freepages(zone, start_page, end_page, migratetype);
782 /* Return the page with the lowest PFN in the list */
783 static struct page *min_page(struct list_head *list)
785 unsigned long min_pfn = -1UL;
786 struct page *min_page = NULL, *page;;
788 list_for_each_entry(page, list, lru) {
789 unsigned long pfn = page_to_pfn(page);
799 /* Remove an element from the buddy allocator from the fallback list */
800 static struct page *__rmqueue_fallback(struct zone *zone, int order,
801 int start_migratetype)
803 struct free_area * area;
807 int nonatomic_fallback_atomic = 0;
810 /* Find the largest possible block of pages in the other list */
811 for (current_order = MAX_ORDER-1; current_order >= order;
813 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
814 migratetype = fallbacks[start_migratetype][i];
816 /* MIGRATE_RESERVE handled later if necessary */
817 if (migratetype == MIGRATE_RESERVE)
820 * Make it hard to fallback to blocks used for
821 * high-order atomic allocations
823 if (migratetype == MIGRATE_HIGHATOMIC &&
824 start_migratetype != MIGRATE_UNMOVABLE &&
825 !nonatomic_fallback_atomic)
828 area = &(zone->free_area[current_order]);
829 if (list_empty(&area->free_list[migratetype]))
832 /* Bias kernel allocations towards low pfns */
833 page = list_entry(area->free_list[migratetype].next,
835 if (unlikely(start_migratetype != MIGRATE_MOVABLE))
836 page = min_page(&area->free_list[migratetype]);
840 * If breaking a large block of pages, move all free
841 * pages to the preferred allocation list. If falling
842 * back for a reclaimable kernel allocation, be more
843 * agressive about taking ownership of free pages
845 if (unlikely(current_order >= MAX_ORDER / 2) ||
846 start_migratetype == MIGRATE_RECLAIMABLE) {
848 pages = move_freepages_block(zone, page,
851 /* Claim the whole block if over half of it is free */
852 if ((pages << current_order) >= (1 << (MAX_ORDER-2)) &&
853 migratetype != MIGRATE_HIGHATOMIC)
854 set_pageblock_migratetype(page,
857 migratetype = start_migratetype;
860 /* Remove the page from the freelists */
861 list_del(&page->lru);
862 rmv_page_order(page);
863 __mod_zone_page_state(zone, NR_FREE_PAGES,
866 if (current_order == MAX_ORDER - 1)
867 set_pageblock_migratetype(page,
870 expand(zone, page, order, current_order, area, migratetype);
875 /* Allow fallback to high-order atomic blocks if memory is that low */
876 if (!nonatomic_fallback_atomic) {
877 nonatomic_fallback_atomic = 1;
881 /* Use MIGRATE_RESERVE rather than fail an allocation */
882 return __rmqueue_smallest(zone, order, MIGRATE_RESERVE);
886 * Do the hard work of removing an element from the buddy allocator.
887 * Call me with the zone->lock already held.
889 static struct page *__rmqueue(struct zone *zone, unsigned int order,
894 page = __rmqueue_smallest(zone, order, migratetype);
897 page = __rmqueue_fallback(zone, order, migratetype);
903 * Obtain a specified number of elements from the buddy allocator, all under
904 * a single hold of the lock, for efficiency. Add them to the supplied list.
905 * Returns the number of new pages which were placed at *list.
907 static int rmqueue_bulk(struct zone *zone, unsigned int order,
908 unsigned long count, struct list_head *list,
913 spin_lock(&zone->lock);
914 for (i = 0; i < count; ++i) {
915 struct page *page = __rmqueue(zone, order, migratetype);
916 if (unlikely(page == NULL))
918 list_add(&page->lru, list);
919 set_page_private(page, migratetype);
921 spin_unlock(&zone->lock);
927 * Called from the vmstat counter updater to drain pagesets of this
928 * currently executing processor on remote nodes after they have
931 * Note that this function must be called with the thread pinned to
932 * a single processor.
934 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
939 local_irq_save(flags);
940 if (pcp->count >= pcp->batch)
941 to_drain = pcp->batch;
943 to_drain = pcp->count;
944 free_pages_bulk(zone, to_drain, &pcp->list, 0);
945 pcp->count -= to_drain;
946 local_irq_restore(flags);
950 static void __drain_pages(unsigned int cpu)
956 for_each_zone(zone) {
957 struct per_cpu_pageset *pset;
959 if (!populated_zone(zone))
962 pset = zone_pcp(zone, cpu);
963 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
964 struct per_cpu_pages *pcp;
967 local_irq_save(flags);
968 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
970 local_irq_restore(flags);
975 #ifdef CONFIG_HIBERNATION
977 void mark_free_pages(struct zone *zone)
979 unsigned long pfn, max_zone_pfn;
982 struct list_head *curr;
984 if (!zone->spanned_pages)
987 spin_lock_irqsave(&zone->lock, flags);
989 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
990 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
991 if (pfn_valid(pfn)) {
992 struct page *page = pfn_to_page(pfn);
994 if (!swsusp_page_is_forbidden(page))
995 swsusp_unset_page_free(page);
998 for_each_migratetype_order(order, t) {
999 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1002 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1003 for (i = 0; i < (1UL << order); i++)
1004 swsusp_set_page_free(pfn_to_page(pfn + i));
1007 spin_unlock_irqrestore(&zone->lock, flags);
1009 #endif /* CONFIG_PM */
1012 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1014 void drain_local_pages(void)
1016 unsigned long flags;
1018 local_irq_save(flags);
1019 __drain_pages(smp_processor_id());
1020 local_irq_restore(flags);
1023 void smp_drain_local_pages(void *arg)
1025 drain_local_pages();
1029 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1031 void drain_all_local_pages(void)
1033 unsigned long flags;
1035 local_irq_save(flags);
1036 __drain_pages(smp_processor_id());
1037 local_irq_restore(flags);
1039 smp_call_function(smp_drain_local_pages, NULL, 0, 1);
1043 * Free a 0-order page
1045 static void fastcall free_hot_cold_page(struct page *page, int cold)
1047 struct zone *zone = page_zone(page);
1048 struct per_cpu_pages *pcp;
1049 unsigned long flags;
1052 page->mapping = NULL;
1053 if (free_pages_check(page))
1056 if (!PageHighMem(page))
1057 debug_check_no_locks_freed(page_address(page), PAGE_SIZE);
1058 arch_free_page(page, 0);
1059 kernel_map_pages(page, 1, 0);
1061 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
1062 local_irq_save(flags);
1063 __count_vm_event(PGFREE);
1064 list_add(&page->lru, &pcp->list);
1065 set_page_private(page, get_pageblock_migratetype(page));
1067 if (pcp->count >= pcp->high) {
1068 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
1069 pcp->count -= pcp->batch;
1071 local_irq_restore(flags);
1075 void fastcall free_hot_page(struct page *page)
1077 free_hot_cold_page(page, 0);
1080 void fastcall free_cold_page(struct page *page)
1082 free_hot_cold_page(page, 1);
1086 * split_page takes a non-compound higher-order page, and splits it into
1087 * n (1<<order) sub-pages: page[0..n]
1088 * Each sub-page must be freed individually.
1090 * Note: this is probably too low level an operation for use in drivers.
1091 * Please consult with lkml before using this in your driver.
1093 void split_page(struct page *page, unsigned int order)
1097 VM_BUG_ON(PageCompound(page));
1098 VM_BUG_ON(!page_count(page));
1099 for (i = 1; i < (1 << order); i++)
1100 set_page_refcounted(page + i);
1104 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1105 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1108 static struct page *buffered_rmqueue(struct zonelist *zonelist,
1109 struct zone *zone, int order, gfp_t gfp_flags)
1111 unsigned long flags;
1113 int cold = !!(gfp_flags & __GFP_COLD);
1115 int migratetype = allocflags_to_migratetype(gfp_flags, order);
1119 if (likely(order == 0)) {
1120 struct per_cpu_pages *pcp;
1122 pcp = &zone_pcp(zone, cpu)->pcp[cold];
1123 local_irq_save(flags);
1125 pcp->count = rmqueue_bulk(zone, 0,
1126 pcp->batch, &pcp->list, migratetype);
1127 if (unlikely(!pcp->count))
1131 /* Find a page of the appropriate migrate type */
1132 list_for_each_entry(page, &pcp->list, lru)
1133 if (page_private(page) == migratetype)
1136 /* Allocate more to the pcp list if necessary */
1137 if (unlikely(&page->lru == &pcp->list)) {
1138 pcp->count += rmqueue_bulk(zone, 0,
1139 pcp->batch, &pcp->list, migratetype);
1140 page = list_entry(pcp->list.next, struct page, lru);
1143 list_del(&page->lru);
1146 spin_lock_irqsave(&zone->lock, flags);
1147 page = __rmqueue(zone, order, migratetype);
1148 spin_unlock(&zone->lock);
1153 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1154 zone_statistics(zonelist, zone);
1155 local_irq_restore(flags);
1158 VM_BUG_ON(bad_range(zone, page));
1159 if (prep_new_page(page, order, gfp_flags))
1164 local_irq_restore(flags);
1169 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
1170 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
1171 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
1172 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
1173 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1174 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1175 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1177 #ifdef CONFIG_FAIL_PAGE_ALLOC
1179 static struct fail_page_alloc_attr {
1180 struct fault_attr attr;
1182 u32 ignore_gfp_highmem;
1183 u32 ignore_gfp_wait;
1186 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1188 struct dentry *ignore_gfp_highmem_file;
1189 struct dentry *ignore_gfp_wait_file;
1190 struct dentry *min_order_file;
1192 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1194 } fail_page_alloc = {
1195 .attr = FAULT_ATTR_INITIALIZER,
1196 .ignore_gfp_wait = 1,
1197 .ignore_gfp_highmem = 1,
1201 static int __init setup_fail_page_alloc(char *str)
1203 return setup_fault_attr(&fail_page_alloc.attr, str);
1205 __setup("fail_page_alloc=", setup_fail_page_alloc);
1207 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1209 if (order < fail_page_alloc.min_order)
1211 if (gfp_mask & __GFP_NOFAIL)
1213 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1215 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1218 return should_fail(&fail_page_alloc.attr, 1 << order);
1221 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1223 static int __init fail_page_alloc_debugfs(void)
1225 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1229 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1233 dir = fail_page_alloc.attr.dentries.dir;
1235 fail_page_alloc.ignore_gfp_wait_file =
1236 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1237 &fail_page_alloc.ignore_gfp_wait);
1239 fail_page_alloc.ignore_gfp_highmem_file =
1240 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1241 &fail_page_alloc.ignore_gfp_highmem);
1242 fail_page_alloc.min_order_file =
1243 debugfs_create_u32("min-order", mode, dir,
1244 &fail_page_alloc.min_order);
1246 if (!fail_page_alloc.ignore_gfp_wait_file ||
1247 !fail_page_alloc.ignore_gfp_highmem_file ||
1248 !fail_page_alloc.min_order_file) {
1250 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1251 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1252 debugfs_remove(fail_page_alloc.min_order_file);
1253 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1259 late_initcall(fail_page_alloc_debugfs);
1261 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1263 #else /* CONFIG_FAIL_PAGE_ALLOC */
1265 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1270 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1273 * Return 1 if free pages are above 'mark'. This takes into account the order
1274 * of the allocation.
1276 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1277 int classzone_idx, int alloc_flags)
1279 /* free_pages my go negative - that's OK */
1281 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
1284 if (alloc_flags & ALLOC_HIGH)
1286 if (alloc_flags & ALLOC_HARDER)
1289 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1291 for (o = 0; o < order; o++) {
1292 /* At the next order, this order's pages become unavailable */
1293 free_pages -= z->free_area[o].nr_free << o;
1295 /* Require fewer higher order pages to be free */
1298 if (free_pages <= min)
1306 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1307 * skip over zones that are not allowed by the cpuset, or that have
1308 * been recently (in last second) found to be nearly full. See further
1309 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1310 * that have to skip over alot of full or unallowed zones.
1312 * If the zonelist cache is present in the passed in zonelist, then
1313 * returns a pointer to the allowed node mask (either the current
1314 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1316 * If the zonelist cache is not available for this zonelist, does
1317 * nothing and returns NULL.
1319 * If the fullzones BITMAP in the zonelist cache is stale (more than
1320 * a second since last zap'd) then we zap it out (clear its bits.)
1322 * We hold off even calling zlc_setup, until after we've checked the
1323 * first zone in the zonelist, on the theory that most allocations will
1324 * be satisfied from that first zone, so best to examine that zone as
1325 * quickly as we can.
1327 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1329 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1330 nodemask_t *allowednodes; /* zonelist_cache approximation */
1332 zlc = zonelist->zlcache_ptr;
1336 if (jiffies - zlc->last_full_zap > 1 * HZ) {
1337 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1338 zlc->last_full_zap = jiffies;
1341 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1342 &cpuset_current_mems_allowed :
1343 &node_states[N_HIGH_MEMORY];
1344 return allowednodes;
1348 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1349 * if it is worth looking at further for free memory:
1350 * 1) Check that the zone isn't thought to be full (doesn't have its
1351 * bit set in the zonelist_cache fullzones BITMAP).
1352 * 2) Check that the zones node (obtained from the zonelist_cache
1353 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1354 * Return true (non-zero) if zone is worth looking at further, or
1355 * else return false (zero) if it is not.
1357 * This check -ignores- the distinction between various watermarks,
1358 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1359 * found to be full for any variation of these watermarks, it will
1360 * be considered full for up to one second by all requests, unless
1361 * we are so low on memory on all allowed nodes that we are forced
1362 * into the second scan of the zonelist.
1364 * In the second scan we ignore this zonelist cache and exactly
1365 * apply the watermarks to all zones, even it is slower to do so.
1366 * We are low on memory in the second scan, and should leave no stone
1367 * unturned looking for a free page.
1369 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zone **z,
1370 nodemask_t *allowednodes)
1372 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1373 int i; /* index of *z in zonelist zones */
1374 int n; /* node that zone *z is on */
1376 zlc = zonelist->zlcache_ptr;
1380 i = z - zonelist->zones;
1383 /* This zone is worth trying if it is allowed but not full */
1384 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1388 * Given 'z' scanning a zonelist, set the corresponding bit in
1389 * zlc->fullzones, so that subsequent attempts to allocate a page
1390 * from that zone don't waste time re-examining it.
1392 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zone **z)
1394 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1395 int i; /* index of *z in zonelist zones */
1397 zlc = zonelist->zlcache_ptr;
1401 i = z - zonelist->zones;
1403 set_bit(i, zlc->fullzones);
1406 #else /* CONFIG_NUMA */
1408 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1413 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zone **z,
1414 nodemask_t *allowednodes)
1419 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zone **z)
1422 #endif /* CONFIG_NUMA */
1425 * get_page_from_freelist goes through the zonelist trying to allocate
1428 static struct page *
1429 get_page_from_freelist(gfp_t gfp_mask, unsigned int order,
1430 struct zonelist *zonelist, int alloc_flags)
1433 struct page *page = NULL;
1434 int classzone_idx = zone_idx(zonelist->zones[0]);
1436 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1437 int zlc_active = 0; /* set if using zonelist_cache */
1438 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1439 enum zone_type highest_zoneidx = -1; /* Gets set for policy zonelists */
1443 * Scan zonelist, looking for a zone with enough free.
1444 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1446 z = zonelist->zones;
1450 * In NUMA, this could be a policy zonelist which contains
1451 * zones that may not be allowed by the current gfp_mask.
1452 * Check the zone is allowed by the current flags
1454 if (unlikely(alloc_should_filter_zonelist(zonelist))) {
1455 if (highest_zoneidx == -1)
1456 highest_zoneidx = gfp_zone(gfp_mask);
1457 if (zone_idx(*z) > highest_zoneidx)
1461 if (NUMA_BUILD && zlc_active &&
1462 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1465 if ((alloc_flags & ALLOC_CPUSET) &&
1466 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1469 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1471 if (alloc_flags & ALLOC_WMARK_MIN)
1472 mark = zone->pages_min;
1473 else if (alloc_flags & ALLOC_WMARK_LOW)
1474 mark = zone->pages_low;
1476 mark = zone->pages_high;
1477 if (!zone_watermark_ok(zone, order, mark,
1478 classzone_idx, alloc_flags)) {
1479 if (!zone_reclaim_mode ||
1480 !zone_reclaim(zone, gfp_mask, order))
1481 goto this_zone_full;
1485 page = buffered_rmqueue(zonelist, zone, order, gfp_mask);
1490 zlc_mark_zone_full(zonelist, z);
1492 if (NUMA_BUILD && !did_zlc_setup) {
1493 /* we do zlc_setup after the first zone is tried */
1494 allowednodes = zlc_setup(zonelist, alloc_flags);
1498 } while (*(++z) != NULL);
1500 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1501 /* Disable zlc cache for second zonelist scan */
1509 * This is the 'heart' of the zoned buddy allocator.
1511 struct page * fastcall
1512 __alloc_pages(gfp_t gfp_mask, unsigned int order,
1513 struct zonelist *zonelist)
1515 const gfp_t wait = gfp_mask & __GFP_WAIT;
1518 struct reclaim_state reclaim_state;
1519 struct task_struct *p = current;
1522 int did_some_progress;
1524 might_sleep_if(wait);
1526 if (should_fail_alloc_page(gfp_mask, order))
1530 z = zonelist->zones; /* the list of zones suitable for gfp_mask */
1532 if (unlikely(*z == NULL)) {
1534 * Happens if we have an empty zonelist as a result of
1535 * GFP_THISNODE being used on a memoryless node
1540 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1541 zonelist, ALLOC_WMARK_LOW|ALLOC_CPUSET);
1546 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1547 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1548 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1549 * using a larger set of nodes after it has established that the
1550 * allowed per node queues are empty and that nodes are
1553 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1556 for (z = zonelist->zones; *z; z++)
1557 wakeup_kswapd(*z, order);
1560 * OK, we're below the kswapd watermark and have kicked background
1561 * reclaim. Now things get more complex, so set up alloc_flags according
1562 * to how we want to proceed.
1564 * The caller may dip into page reserves a bit more if the caller
1565 * cannot run direct reclaim, or if the caller has realtime scheduling
1566 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1567 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1569 alloc_flags = ALLOC_WMARK_MIN;
1570 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
1571 alloc_flags |= ALLOC_HARDER;
1572 if (gfp_mask & __GFP_HIGH)
1573 alloc_flags |= ALLOC_HIGH;
1575 alloc_flags |= ALLOC_CPUSET;
1578 * Go through the zonelist again. Let __GFP_HIGH and allocations
1579 * coming from realtime tasks go deeper into reserves.
1581 * This is the last chance, in general, before the goto nopage.
1582 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1583 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1585 page = get_page_from_freelist(gfp_mask, order, zonelist, alloc_flags);
1589 /* This allocation should allow future memory freeing. */
1592 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
1593 && !in_interrupt()) {
1594 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
1596 /* go through the zonelist yet again, ignoring mins */
1597 page = get_page_from_freelist(gfp_mask, order,
1598 zonelist, ALLOC_NO_WATERMARKS);
1601 if (gfp_mask & __GFP_NOFAIL) {
1602 congestion_wait(WRITE, HZ/50);
1609 /* Atomic allocations - we can't balance anything */
1615 /* We now go into synchronous reclaim */
1616 cpuset_memory_pressure_bump();
1617 p->flags |= PF_MEMALLOC;
1618 reclaim_state.reclaimed_slab = 0;
1619 p->reclaim_state = &reclaim_state;
1621 did_some_progress = try_to_free_pages(zonelist->zones, order, gfp_mask);
1623 p->reclaim_state = NULL;
1624 p->flags &= ~PF_MEMALLOC;
1629 drain_all_local_pages();
1631 if (likely(did_some_progress)) {
1632 page = get_page_from_freelist(gfp_mask, order,
1633 zonelist, alloc_flags);
1636 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1638 * Go through the zonelist yet one more time, keep
1639 * very high watermark here, this is only to catch
1640 * a parallel oom killing, we must fail if we're still
1641 * under heavy pressure.
1643 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1644 zonelist, ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1648 /* The OOM killer will not help higher order allocs so fail */
1649 if (order > PAGE_ALLOC_COSTLY_ORDER)
1652 out_of_memory(zonelist, gfp_mask, order);
1657 * Don't let big-order allocations loop unless the caller explicitly
1658 * requests that. Wait for some write requests to complete then retry.
1660 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
1661 * <= 3, but that may not be true in other implementations.
1664 if (!(gfp_mask & __GFP_NORETRY)) {
1665 if ((order <= PAGE_ALLOC_COSTLY_ORDER) ||
1666 (gfp_mask & __GFP_REPEAT))
1668 if (gfp_mask & __GFP_NOFAIL)
1672 congestion_wait(WRITE, HZ/50);
1677 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1678 printk(KERN_WARNING "%s: page allocation failure."
1679 " order:%d, mode:0x%x\n",
1680 p->comm, order, gfp_mask);
1688 EXPORT_SYMBOL(__alloc_pages);
1691 * Common helper functions.
1693 fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1696 page = alloc_pages(gfp_mask, order);
1699 return (unsigned long) page_address(page);
1702 EXPORT_SYMBOL(__get_free_pages);
1704 fastcall unsigned long get_zeroed_page(gfp_t gfp_mask)
1709 * get_zeroed_page() returns a 32-bit address, which cannot represent
1712 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1714 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1716 return (unsigned long) page_address(page);
1720 EXPORT_SYMBOL(get_zeroed_page);
1722 void __pagevec_free(struct pagevec *pvec)
1724 int i = pagevec_count(pvec);
1727 free_hot_cold_page(pvec->pages[i], pvec->cold);
1730 fastcall void __free_pages(struct page *page, unsigned int order)
1732 if (put_page_testzero(page)) {
1734 free_hot_page(page);
1736 __free_pages_ok(page, order);
1740 EXPORT_SYMBOL(__free_pages);
1742 fastcall void free_pages(unsigned long addr, unsigned int order)
1745 VM_BUG_ON(!virt_addr_valid((void *)addr));
1746 __free_pages(virt_to_page((void *)addr), order);
1750 EXPORT_SYMBOL(free_pages);
1752 static unsigned int nr_free_zone_pages(int offset)
1754 /* Just pick one node, since fallback list is circular */
1755 pg_data_t *pgdat = NODE_DATA(numa_node_id());
1756 unsigned int sum = 0;
1758 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1759 struct zone **zonep = zonelist->zones;
1762 for (zone = *zonep++; zone; zone = *zonep++) {
1763 unsigned long size = zone->present_pages;
1764 unsigned long high = zone->pages_high;
1773 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1775 unsigned int nr_free_buffer_pages(void)
1777 return nr_free_zone_pages(gfp_zone(GFP_USER));
1779 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
1782 * Amount of free RAM allocatable within all zones
1784 unsigned int nr_free_pagecache_pages(void)
1786 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
1789 static inline void show_node(struct zone *zone)
1792 printk("Node %d ", zone_to_nid(zone));
1795 void si_meminfo(struct sysinfo *val)
1797 val->totalram = totalram_pages;
1799 val->freeram = global_page_state(NR_FREE_PAGES);
1800 val->bufferram = nr_blockdev_pages();
1801 val->totalhigh = totalhigh_pages;
1802 val->freehigh = nr_free_highpages();
1803 val->mem_unit = PAGE_SIZE;
1806 EXPORT_SYMBOL(si_meminfo);
1809 void si_meminfo_node(struct sysinfo *val, int nid)
1811 pg_data_t *pgdat = NODE_DATA(nid);
1813 val->totalram = pgdat->node_present_pages;
1814 val->freeram = node_page_state(nid, NR_FREE_PAGES);
1815 #ifdef CONFIG_HIGHMEM
1816 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1817 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
1823 val->mem_unit = PAGE_SIZE;
1827 #define K(x) ((x) << (PAGE_SHIFT-10))
1830 * Show free area list (used inside shift_scroll-lock stuff)
1831 * We also calculate the percentage fragmentation. We do this by counting the
1832 * memory on each free list with the exception of the first item on the list.
1834 void show_free_areas(void)
1839 for_each_zone(zone) {
1840 if (!populated_zone(zone))
1844 printk("%s per-cpu:\n", zone->name);
1846 for_each_online_cpu(cpu) {
1847 struct per_cpu_pageset *pageset;
1849 pageset = zone_pcp(zone, cpu);
1851 printk("CPU %4d: Hot: hi:%5d, btch:%4d usd:%4d "
1852 "Cold: hi:%5d, btch:%4d usd:%4d\n",
1853 cpu, pageset->pcp[0].high,
1854 pageset->pcp[0].batch, pageset->pcp[0].count,
1855 pageset->pcp[1].high, pageset->pcp[1].batch,
1856 pageset->pcp[1].count);
1860 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu unstable:%lu\n"
1861 " free:%lu slab:%lu mapped:%lu pagetables:%lu bounce:%lu\n",
1862 global_page_state(NR_ACTIVE),
1863 global_page_state(NR_INACTIVE),
1864 global_page_state(NR_FILE_DIRTY),
1865 global_page_state(NR_WRITEBACK),
1866 global_page_state(NR_UNSTABLE_NFS),
1867 global_page_state(NR_FREE_PAGES),
1868 global_page_state(NR_SLAB_RECLAIMABLE) +
1869 global_page_state(NR_SLAB_UNRECLAIMABLE),
1870 global_page_state(NR_FILE_MAPPED),
1871 global_page_state(NR_PAGETABLE),
1872 global_page_state(NR_BOUNCE));
1874 for_each_zone(zone) {
1877 if (!populated_zone(zone))
1889 " pages_scanned:%lu"
1890 " all_unreclaimable? %s"
1893 K(zone_page_state(zone, NR_FREE_PAGES)),
1896 K(zone->pages_high),
1897 K(zone_page_state(zone, NR_ACTIVE)),
1898 K(zone_page_state(zone, NR_INACTIVE)),
1899 K(zone->present_pages),
1900 zone->pages_scanned,
1901 (zone->all_unreclaimable ? "yes" : "no")
1903 printk("lowmem_reserve[]:");
1904 for (i = 0; i < MAX_NR_ZONES; i++)
1905 printk(" %lu", zone->lowmem_reserve[i]);
1909 for_each_zone(zone) {
1910 unsigned long nr[MAX_ORDER], flags, order, total = 0;
1912 if (!populated_zone(zone))
1916 printk("%s: ", zone->name);
1918 spin_lock_irqsave(&zone->lock, flags);
1919 for (order = 0; order < MAX_ORDER; order++) {
1920 nr[order] = zone->free_area[order].nr_free;
1921 total += nr[order] << order;
1923 spin_unlock_irqrestore(&zone->lock, flags);
1924 for (order = 0; order < MAX_ORDER; order++)
1925 printk("%lu*%lukB ", nr[order], K(1UL) << order);
1926 printk("= %lukB\n", K(total));
1929 show_swap_cache_info();
1933 * Builds allocation fallback zone lists.
1935 * Add all populated zones of a node to the zonelist.
1937 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
1938 int nr_zones, enum zone_type zone_type)
1942 BUG_ON(zone_type >= MAX_NR_ZONES);
1947 zone = pgdat->node_zones + zone_type;
1948 if (populated_zone(zone)) {
1949 zonelist->zones[nr_zones++] = zone;
1950 check_highest_zone(zone_type);
1953 } while (zone_type);
1960 * 0 = automatic detection of better ordering.
1961 * 1 = order by ([node] distance, -zonetype)
1962 * 2 = order by (-zonetype, [node] distance)
1964 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
1965 * the same zonelist. So only NUMA can configure this param.
1967 #define ZONELIST_ORDER_DEFAULT 0
1968 #define ZONELIST_ORDER_NODE 1
1969 #define ZONELIST_ORDER_ZONE 2
1971 /* zonelist order in the kernel.
1972 * set_zonelist_order() will set this to NODE or ZONE.
1974 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
1975 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
1979 /* The value user specified ....changed by config */
1980 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
1981 /* string for sysctl */
1982 #define NUMA_ZONELIST_ORDER_LEN 16
1983 char numa_zonelist_order[16] = "default";
1986 * interface for configure zonelist ordering.
1987 * command line option "numa_zonelist_order"
1988 * = "[dD]efault - default, automatic configuration.
1989 * = "[nN]ode - order by node locality, then by zone within node
1990 * = "[zZ]one - order by zone, then by locality within zone
1993 static int __parse_numa_zonelist_order(char *s)
1995 if (*s == 'd' || *s == 'D') {
1996 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
1997 } else if (*s == 'n' || *s == 'N') {
1998 user_zonelist_order = ZONELIST_ORDER_NODE;
1999 } else if (*s == 'z' || *s == 'Z') {
2000 user_zonelist_order = ZONELIST_ORDER_ZONE;
2003 "Ignoring invalid numa_zonelist_order value: "
2010 static __init int setup_numa_zonelist_order(char *s)
2013 return __parse_numa_zonelist_order(s);
2016 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2019 * sysctl handler for numa_zonelist_order
2021 int numa_zonelist_order_handler(ctl_table *table, int write,
2022 struct file *file, void __user *buffer, size_t *length,
2025 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2029 strncpy(saved_string, (char*)table->data,
2030 NUMA_ZONELIST_ORDER_LEN);
2031 ret = proc_dostring(table, write, file, buffer, length, ppos);
2035 int oldval = user_zonelist_order;
2036 if (__parse_numa_zonelist_order((char*)table->data)) {
2038 * bogus value. restore saved string
2040 strncpy((char*)table->data, saved_string,
2041 NUMA_ZONELIST_ORDER_LEN);
2042 user_zonelist_order = oldval;
2043 } else if (oldval != user_zonelist_order)
2044 build_all_zonelists();
2050 #define MAX_NODE_LOAD (num_online_nodes())
2051 static int node_load[MAX_NUMNODES];
2054 * find_next_best_node - find the next node that should appear in a given node's fallback list
2055 * @node: node whose fallback list we're appending
2056 * @used_node_mask: nodemask_t of already used nodes
2058 * We use a number of factors to determine which is the next node that should
2059 * appear on a given node's fallback list. The node should not have appeared
2060 * already in @node's fallback list, and it should be the next closest node
2061 * according to the distance array (which contains arbitrary distance values
2062 * from each node to each node in the system), and should also prefer nodes
2063 * with no CPUs, since presumably they'll have very little allocation pressure
2064 * on them otherwise.
2065 * It returns -1 if no node is found.
2067 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2070 int min_val = INT_MAX;
2073 /* Use the local node if we haven't already */
2074 if (!node_isset(node, *used_node_mask)) {
2075 node_set(node, *used_node_mask);
2079 for_each_node_state(n, N_HIGH_MEMORY) {
2082 /* Don't want a node to appear more than once */
2083 if (node_isset(n, *used_node_mask))
2086 /* Use the distance array to find the distance */
2087 val = node_distance(node, n);
2089 /* Penalize nodes under us ("prefer the next node") */
2092 /* Give preference to headless and unused nodes */
2093 tmp = node_to_cpumask(n);
2094 if (!cpus_empty(tmp))
2095 val += PENALTY_FOR_NODE_WITH_CPUS;
2097 /* Slight preference for less loaded node */
2098 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2099 val += node_load[n];
2101 if (val < min_val) {
2108 node_set(best_node, *used_node_mask);
2115 * Build zonelists ordered by node and zones within node.
2116 * This results in maximum locality--normal zone overflows into local
2117 * DMA zone, if any--but risks exhausting DMA zone.
2119 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2123 struct zonelist *zonelist;
2125 for (i = 0; i < MAX_NR_ZONES; i++) {
2126 zonelist = pgdat->node_zonelists + i;
2127 for (j = 0; zonelist->zones[j] != NULL; j++)
2129 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
2130 zonelist->zones[j] = NULL;
2135 * Build gfp_thisnode zonelists
2137 static void build_thisnode_zonelists(pg_data_t *pgdat)
2141 struct zonelist *zonelist;
2143 for (i = 0; i < MAX_NR_ZONES; i++) {
2144 zonelist = pgdat->node_zonelists + MAX_NR_ZONES + i;
2145 j = build_zonelists_node(pgdat, zonelist, 0, i);
2146 zonelist->zones[j] = NULL;
2151 * Build zonelists ordered by zone and nodes within zones.
2152 * This results in conserving DMA zone[s] until all Normal memory is
2153 * exhausted, but results in overflowing to remote node while memory
2154 * may still exist in local DMA zone.
2156 static int node_order[MAX_NUMNODES];
2158 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2162 int zone_type; /* needs to be signed */
2164 struct zonelist *zonelist;
2166 for (i = 0; i < MAX_NR_ZONES; i++) {
2167 zonelist = pgdat->node_zonelists + i;
2169 for (zone_type = i; zone_type >= 0; zone_type--) {
2170 for (j = 0; j < nr_nodes; j++) {
2171 node = node_order[j];
2172 z = &NODE_DATA(node)->node_zones[zone_type];
2173 if (populated_zone(z)) {
2174 zonelist->zones[pos++] = z;
2175 check_highest_zone(zone_type);
2179 zonelist->zones[pos] = NULL;
2183 static int default_zonelist_order(void)
2186 unsigned long low_kmem_size,total_size;
2190 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2191 * If they are really small and used heavily, the system can fall
2192 * into OOM very easily.
2193 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2195 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2198 for_each_online_node(nid) {
2199 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2200 z = &NODE_DATA(nid)->node_zones[zone_type];
2201 if (populated_zone(z)) {
2202 if (zone_type < ZONE_NORMAL)
2203 low_kmem_size += z->present_pages;
2204 total_size += z->present_pages;
2208 if (!low_kmem_size || /* there are no DMA area. */
2209 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2210 return ZONELIST_ORDER_NODE;
2212 * look into each node's config.
2213 * If there is a node whose DMA/DMA32 memory is very big area on
2214 * local memory, NODE_ORDER may be suitable.
2216 average_size = total_size /
2217 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2218 for_each_online_node(nid) {
2221 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2222 z = &NODE_DATA(nid)->node_zones[zone_type];
2223 if (populated_zone(z)) {
2224 if (zone_type < ZONE_NORMAL)
2225 low_kmem_size += z->present_pages;
2226 total_size += z->present_pages;
2229 if (low_kmem_size &&
2230 total_size > average_size && /* ignore small node */
2231 low_kmem_size > total_size * 70/100)
2232 return ZONELIST_ORDER_NODE;
2234 return ZONELIST_ORDER_ZONE;
2237 static void set_zonelist_order(void)
2239 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2240 current_zonelist_order = default_zonelist_order();
2242 current_zonelist_order = user_zonelist_order;
2245 static void build_zonelists(pg_data_t *pgdat)
2249 nodemask_t used_mask;
2250 int local_node, prev_node;
2251 struct zonelist *zonelist;
2252 int order = current_zonelist_order;
2254 /* initialize zonelists */
2255 for (i = 0; i < MAX_ZONELISTS; i++) {
2256 zonelist = pgdat->node_zonelists + i;
2257 zonelist->zones[0] = NULL;
2260 /* NUMA-aware ordering of nodes */
2261 local_node = pgdat->node_id;
2262 load = num_online_nodes();
2263 prev_node = local_node;
2264 nodes_clear(used_mask);
2266 memset(node_load, 0, sizeof(node_load));
2267 memset(node_order, 0, sizeof(node_order));
2270 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2271 int distance = node_distance(local_node, node);
2274 * If another node is sufficiently far away then it is better
2275 * to reclaim pages in a zone before going off node.
2277 if (distance > RECLAIM_DISTANCE)
2278 zone_reclaim_mode = 1;
2281 * We don't want to pressure a particular node.
2282 * So adding penalty to the first node in same
2283 * distance group to make it round-robin.
2285 if (distance != node_distance(local_node, prev_node))
2286 node_load[node] = load;
2290 if (order == ZONELIST_ORDER_NODE)
2291 build_zonelists_in_node_order(pgdat, node);
2293 node_order[j++] = node; /* remember order */
2296 if (order == ZONELIST_ORDER_ZONE) {
2297 /* calculate node order -- i.e., DMA last! */
2298 build_zonelists_in_zone_order(pgdat, j);
2301 build_thisnode_zonelists(pgdat);
2304 /* Construct the zonelist performance cache - see further mmzone.h */
2305 static void build_zonelist_cache(pg_data_t *pgdat)
2309 for (i = 0; i < MAX_NR_ZONES; i++) {
2310 struct zonelist *zonelist;
2311 struct zonelist_cache *zlc;
2314 zonelist = pgdat->node_zonelists + i;
2315 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2316 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2317 for (z = zonelist->zones; *z; z++)
2318 zlc->z_to_n[z - zonelist->zones] = zone_to_nid(*z);
2323 #else /* CONFIG_NUMA */
2325 static void set_zonelist_order(void)
2327 current_zonelist_order = ZONELIST_ORDER_ZONE;
2330 static void build_zonelists(pg_data_t *pgdat)
2332 int node, local_node;
2335 local_node = pgdat->node_id;
2336 for (i = 0; i < MAX_NR_ZONES; i++) {
2337 struct zonelist *zonelist;
2339 zonelist = pgdat->node_zonelists + i;
2341 j = build_zonelists_node(pgdat, zonelist, 0, i);
2343 * Now we build the zonelist so that it contains the zones
2344 * of all the other nodes.
2345 * We don't want to pressure a particular node, so when
2346 * building the zones for node N, we make sure that the
2347 * zones coming right after the local ones are those from
2348 * node N+1 (modulo N)
2350 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2351 if (!node_online(node))
2353 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
2355 for (node = 0; node < local_node; node++) {
2356 if (!node_online(node))
2358 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
2361 zonelist->zones[j] = NULL;
2365 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2366 static void build_zonelist_cache(pg_data_t *pgdat)
2370 for (i = 0; i < MAX_NR_ZONES; i++)
2371 pgdat->node_zonelists[i].zlcache_ptr = NULL;
2374 #endif /* CONFIG_NUMA */
2376 /* return values int ....just for stop_machine_run() */
2377 static int __build_all_zonelists(void *dummy)
2381 for_each_online_node(nid) {
2382 pg_data_t *pgdat = NODE_DATA(nid);
2384 build_zonelists(pgdat);
2385 build_zonelist_cache(pgdat);
2390 void build_all_zonelists(void)
2392 set_zonelist_order();
2394 if (system_state == SYSTEM_BOOTING) {
2395 __build_all_zonelists(NULL);
2396 cpuset_init_current_mems_allowed();
2398 /* we have to stop all cpus to guaranntee there is no user
2400 stop_machine_run(__build_all_zonelists, NULL, NR_CPUS);
2401 /* cpuset refresh routine should be here */
2403 vm_total_pages = nr_free_pagecache_pages();
2405 * Disable grouping by mobility if the number of pages in the
2406 * system is too low to allow the mechanism to work. It would be
2407 * more accurate, but expensive to check per-zone. This check is
2408 * made on memory-hotadd so a system can start with mobility
2409 * disabled and enable it later
2411 if (vm_total_pages < (MAX_ORDER_NR_PAGES * MIGRATE_TYPES))
2412 page_group_by_mobility_disabled = 1;
2414 page_group_by_mobility_disabled = 0;
2416 printk("Built %i zonelists in %s order, mobility grouping %s. "
2417 "Total pages: %ld\n",
2419 zonelist_order_name[current_zonelist_order],
2420 page_group_by_mobility_disabled ? "off" : "on",
2423 printk("Policy zone: %s\n", zone_names[policy_zone]);
2428 * Helper functions to size the waitqueue hash table.
2429 * Essentially these want to choose hash table sizes sufficiently
2430 * large so that collisions trying to wait on pages are rare.
2431 * But in fact, the number of active page waitqueues on typical
2432 * systems is ridiculously low, less than 200. So this is even
2433 * conservative, even though it seems large.
2435 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2436 * waitqueues, i.e. the size of the waitq table given the number of pages.
2438 #define PAGES_PER_WAITQUEUE 256
2440 #ifndef CONFIG_MEMORY_HOTPLUG
2441 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2443 unsigned long size = 1;
2445 pages /= PAGES_PER_WAITQUEUE;
2447 while (size < pages)
2451 * Once we have dozens or even hundreds of threads sleeping
2452 * on IO we've got bigger problems than wait queue collision.
2453 * Limit the size of the wait table to a reasonable size.
2455 size = min(size, 4096UL);
2457 return max(size, 4UL);
2461 * A zone's size might be changed by hot-add, so it is not possible to determine
2462 * a suitable size for its wait_table. So we use the maximum size now.
2464 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2466 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2467 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2468 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2470 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2471 * or more by the traditional way. (See above). It equals:
2473 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2474 * ia64(16K page size) : = ( 8G + 4M)byte.
2475 * powerpc (64K page size) : = (32G +16M)byte.
2477 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2484 * This is an integer logarithm so that shifts can be used later
2485 * to extract the more random high bits from the multiplicative
2486 * hash function before the remainder is taken.
2488 static inline unsigned long wait_table_bits(unsigned long size)
2493 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2496 * Mark a number of MAX_ORDER_NR_PAGES blocks as MIGRATE_RESERVE. The number
2497 * of blocks reserved is based on zone->pages_min. The memory within the
2498 * reserve will tend to store contiguous free pages. Setting min_free_kbytes
2499 * higher will lead to a bigger reserve which will get freed as contiguous
2500 * blocks as reclaim kicks in
2502 static void setup_zone_migrate_reserve(struct zone *zone)
2504 unsigned long start_pfn, pfn, end_pfn;
2506 unsigned long reserve, block_migratetype;
2508 /* Get the start pfn, end pfn and the number of blocks to reserve */
2509 start_pfn = zone->zone_start_pfn;
2510 end_pfn = start_pfn + zone->spanned_pages;
2511 reserve = roundup(zone->pages_min, MAX_ORDER_NR_PAGES) >> (MAX_ORDER-1);
2513 for (pfn = start_pfn; pfn < end_pfn; pfn += MAX_ORDER_NR_PAGES) {
2514 if (!pfn_valid(pfn))
2516 page = pfn_to_page(pfn);
2518 /* Blocks with reserved pages will never free, skip them. */
2519 if (PageReserved(page))
2522 block_migratetype = get_pageblock_migratetype(page);
2524 /* If this block is reserved, account for it */
2525 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
2530 /* Suitable for reserving if this block is movable */
2531 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
2532 set_pageblock_migratetype(page, MIGRATE_RESERVE);
2533 move_freepages_block(zone, page, MIGRATE_RESERVE);
2539 * If the reserve is met and this is a previous reserved block,
2542 if (block_migratetype == MIGRATE_RESERVE) {
2543 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2544 move_freepages_block(zone, page, MIGRATE_MOVABLE);
2550 * Initially all pages are reserved - free ones are freed
2551 * up by free_all_bootmem() once the early boot process is
2552 * done. Non-atomic initialization, single-pass.
2554 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
2555 unsigned long start_pfn, enum memmap_context context)
2558 unsigned long end_pfn = start_pfn + size;
2561 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
2563 * There can be holes in boot-time mem_map[]s
2564 * handed to this function. They do not
2565 * exist on hotplugged memory.
2567 if (context == MEMMAP_EARLY) {
2568 if (!early_pfn_valid(pfn))
2570 if (!early_pfn_in_nid(pfn, nid))
2573 page = pfn_to_page(pfn);
2574 set_page_links(page, zone, nid, pfn);
2575 init_page_count(page);
2576 reset_page_mapcount(page);
2577 SetPageReserved(page);
2580 * Mark the block movable so that blocks are reserved for
2581 * movable at startup. This will force kernel allocations
2582 * to reserve their blocks rather than leaking throughout
2583 * the address space during boot when many long-lived
2584 * kernel allocations are made. Later some blocks near
2585 * the start are marked MIGRATE_RESERVE by
2586 * setup_zone_migrate_reserve()
2588 if ((pfn & (MAX_ORDER_NR_PAGES-1)))
2589 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2591 INIT_LIST_HEAD(&page->lru);
2592 #ifdef WANT_PAGE_VIRTUAL
2593 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2594 if (!is_highmem_idx(zone))
2595 set_page_address(page, __va(pfn << PAGE_SHIFT));
2600 static void __meminit zone_init_free_lists(struct pglist_data *pgdat,
2601 struct zone *zone, unsigned long size)
2604 for_each_migratetype_order(order, t) {
2605 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
2606 zone->free_area[order].nr_free = 0;
2610 #ifndef __HAVE_ARCH_MEMMAP_INIT
2611 #define memmap_init(size, nid, zone, start_pfn) \
2612 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
2615 static int __devinit zone_batchsize(struct zone *zone)
2620 * The per-cpu-pages pools are set to around 1000th of the
2621 * size of the zone. But no more than 1/2 of a meg.
2623 * OK, so we don't know how big the cache is. So guess.
2625 batch = zone->present_pages / 1024;
2626 if (batch * PAGE_SIZE > 512 * 1024)
2627 batch = (512 * 1024) / PAGE_SIZE;
2628 batch /= 4; /* We effectively *= 4 below */
2633 * Clamp the batch to a 2^n - 1 value. Having a power
2634 * of 2 value was found to be more likely to have
2635 * suboptimal cache aliasing properties in some cases.
2637 * For example if 2 tasks are alternately allocating
2638 * batches of pages, one task can end up with a lot
2639 * of pages of one half of the possible page colors
2640 * and the other with pages of the other colors.
2642 batch = (1 << (fls(batch + batch/2)-1)) - 1;
2647 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
2649 struct per_cpu_pages *pcp;
2651 memset(p, 0, sizeof(*p));
2653 pcp = &p->pcp[0]; /* hot */
2655 pcp->high = 6 * batch;
2656 pcp->batch = max(1UL, 1 * batch);
2657 INIT_LIST_HEAD(&pcp->list);
2659 pcp = &p->pcp[1]; /* cold*/
2661 pcp->high = 2 * batch;
2662 pcp->batch = max(1UL, batch/2);
2663 INIT_LIST_HEAD(&pcp->list);
2667 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
2668 * to the value high for the pageset p.
2671 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
2674 struct per_cpu_pages *pcp;
2676 pcp = &p->pcp[0]; /* hot list */
2678 pcp->batch = max(1UL, high/4);
2679 if ((high/4) > (PAGE_SHIFT * 8))
2680 pcp->batch = PAGE_SHIFT * 8;
2686 * Boot pageset table. One per cpu which is going to be used for all
2687 * zones and all nodes. The parameters will be set in such a way
2688 * that an item put on a list will immediately be handed over to
2689 * the buddy list. This is safe since pageset manipulation is done
2690 * with interrupts disabled.
2692 * Some NUMA counter updates may also be caught by the boot pagesets.
2694 * The boot_pagesets must be kept even after bootup is complete for
2695 * unused processors and/or zones. They do play a role for bootstrapping
2696 * hotplugged processors.
2698 * zoneinfo_show() and maybe other functions do
2699 * not check if the processor is online before following the pageset pointer.
2700 * Other parts of the kernel may not check if the zone is available.
2702 static struct per_cpu_pageset boot_pageset[NR_CPUS];
2705 * Dynamically allocate memory for the
2706 * per cpu pageset array in struct zone.
2708 static int __cpuinit process_zones(int cpu)
2710 struct zone *zone, *dzone;
2711 int node = cpu_to_node(cpu);
2713 node_set_state(node, N_CPU); /* this node has a cpu */
2715 for_each_zone(zone) {
2717 if (!populated_zone(zone))
2720 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
2722 if (!zone_pcp(zone, cpu))
2725 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
2727 if (percpu_pagelist_fraction)
2728 setup_pagelist_highmark(zone_pcp(zone, cpu),
2729 (zone->present_pages / percpu_pagelist_fraction));
2734 for_each_zone(dzone) {
2735 if (!populated_zone(dzone))
2739 kfree(zone_pcp(dzone, cpu));
2740 zone_pcp(dzone, cpu) = NULL;
2745 static inline void free_zone_pagesets(int cpu)
2749 for_each_zone(zone) {
2750 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
2752 /* Free per_cpu_pageset if it is slab allocated */
2753 if (pset != &boot_pageset[cpu])
2755 zone_pcp(zone, cpu) = NULL;
2759 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
2760 unsigned long action,
2763 int cpu = (long)hcpu;
2764 int ret = NOTIFY_OK;
2767 case CPU_UP_PREPARE:
2768 case CPU_UP_PREPARE_FROZEN:
2769 if (process_zones(cpu))
2772 case CPU_UP_CANCELED:
2773 case CPU_UP_CANCELED_FROZEN:
2775 case CPU_DEAD_FROZEN:
2776 free_zone_pagesets(cpu);
2784 static struct notifier_block __cpuinitdata pageset_notifier =
2785 { &pageset_cpuup_callback, NULL, 0 };
2787 void __init setup_per_cpu_pageset(void)
2791 /* Initialize per_cpu_pageset for cpu 0.
2792 * A cpuup callback will do this for every cpu
2793 * as it comes online
2795 err = process_zones(smp_processor_id());
2797 register_cpu_notifier(&pageset_notifier);
2802 static noinline __init_refok
2803 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
2806 struct pglist_data *pgdat = zone->zone_pgdat;
2810 * The per-page waitqueue mechanism uses hashed waitqueues
2813 zone->wait_table_hash_nr_entries =
2814 wait_table_hash_nr_entries(zone_size_pages);
2815 zone->wait_table_bits =
2816 wait_table_bits(zone->wait_table_hash_nr_entries);
2817 alloc_size = zone->wait_table_hash_nr_entries
2818 * sizeof(wait_queue_head_t);
2820 if (system_state == SYSTEM_BOOTING) {
2821 zone->wait_table = (wait_queue_head_t *)
2822 alloc_bootmem_node(pgdat, alloc_size);
2825 * This case means that a zone whose size was 0 gets new memory
2826 * via memory hot-add.
2827 * But it may be the case that a new node was hot-added. In
2828 * this case vmalloc() will not be able to use this new node's
2829 * memory - this wait_table must be initialized to use this new
2830 * node itself as well.
2831 * To use this new node's memory, further consideration will be
2834 zone->wait_table = vmalloc(alloc_size);
2836 if (!zone->wait_table)
2839 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
2840 init_waitqueue_head(zone->wait_table + i);
2845 static __meminit void zone_pcp_init(struct zone *zone)
2848 unsigned long batch = zone_batchsize(zone);
2850 for (cpu = 0; cpu < NR_CPUS; cpu++) {
2852 /* Early boot. Slab allocator not functional yet */
2853 zone_pcp(zone, cpu) = &boot_pageset[cpu];
2854 setup_pageset(&boot_pageset[cpu],0);
2856 setup_pageset(zone_pcp(zone,cpu), batch);
2859 if (zone->present_pages)
2860 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
2861 zone->name, zone->present_pages, batch);
2864 __meminit int init_currently_empty_zone(struct zone *zone,
2865 unsigned long zone_start_pfn,
2867 enum memmap_context context)
2869 struct pglist_data *pgdat = zone->zone_pgdat;
2871 ret = zone_wait_table_init(zone, size);
2874 pgdat->nr_zones = zone_idx(zone) + 1;
2876 zone->zone_start_pfn = zone_start_pfn;
2878 memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn);
2880 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
2885 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
2887 * Basic iterator support. Return the first range of PFNs for a node
2888 * Note: nid == MAX_NUMNODES returns first region regardless of node
2890 static int __meminit first_active_region_index_in_nid(int nid)
2894 for (i = 0; i < nr_nodemap_entries; i++)
2895 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
2902 * Basic iterator support. Return the next active range of PFNs for a node
2903 * Note: nid == MAX_NUMNODES returns next region regardles of node
2905 static int __meminit next_active_region_index_in_nid(int index, int nid)
2907 for (index = index + 1; index < nr_nodemap_entries; index++)
2908 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
2914 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
2916 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
2917 * Architectures may implement their own version but if add_active_range()
2918 * was used and there are no special requirements, this is a convenient
2921 int __meminit early_pfn_to_nid(unsigned long pfn)
2925 for (i = 0; i < nr_nodemap_entries; i++) {
2926 unsigned long start_pfn = early_node_map[i].start_pfn;
2927 unsigned long end_pfn = early_node_map[i].end_pfn;
2929 if (start_pfn <= pfn && pfn < end_pfn)
2930 return early_node_map[i].nid;
2935 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
2937 /* Basic iterator support to walk early_node_map[] */
2938 #define for_each_active_range_index_in_nid(i, nid) \
2939 for (i = first_active_region_index_in_nid(nid); i != -1; \
2940 i = next_active_region_index_in_nid(i, nid))
2943 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
2944 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
2945 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
2947 * If an architecture guarantees that all ranges registered with
2948 * add_active_ranges() contain no holes and may be freed, this
2949 * this function may be used instead of calling free_bootmem() manually.
2951 void __init free_bootmem_with_active_regions(int nid,
2952 unsigned long max_low_pfn)
2956 for_each_active_range_index_in_nid(i, nid) {
2957 unsigned long size_pages = 0;
2958 unsigned long end_pfn = early_node_map[i].end_pfn;
2960 if (early_node_map[i].start_pfn >= max_low_pfn)
2963 if (end_pfn > max_low_pfn)
2964 end_pfn = max_low_pfn;
2966 size_pages = end_pfn - early_node_map[i].start_pfn;
2967 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
2968 PFN_PHYS(early_node_map[i].start_pfn),
2969 size_pages << PAGE_SHIFT);
2974 * sparse_memory_present_with_active_regions - Call memory_present for each active range
2975 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
2977 * If an architecture guarantees that all ranges registered with
2978 * add_active_ranges() contain no holes and may be freed, this
2979 * function may be used instead of calling memory_present() manually.
2981 void __init sparse_memory_present_with_active_regions(int nid)
2985 for_each_active_range_index_in_nid(i, nid)
2986 memory_present(early_node_map[i].nid,
2987 early_node_map[i].start_pfn,
2988 early_node_map[i].end_pfn);
2992 * push_node_boundaries - Push node boundaries to at least the requested boundary
2993 * @nid: The nid of the node to push the boundary for
2994 * @start_pfn: The start pfn of the node
2995 * @end_pfn: The end pfn of the node
2997 * In reserve-based hot-add, mem_map is allocated that is unused until hotadd
2998 * time. Specifically, on x86_64, SRAT will report ranges that can potentially
2999 * be hotplugged even though no physical memory exists. This function allows
3000 * an arch to push out the node boundaries so mem_map is allocated that can
3003 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
3004 void __init push_node_boundaries(unsigned int nid,
3005 unsigned long start_pfn, unsigned long end_pfn)
3007 printk(KERN_DEBUG "Entering push_node_boundaries(%u, %lu, %lu)\n",
3008 nid, start_pfn, end_pfn);
3010 /* Initialise the boundary for this node if necessary */
3011 if (node_boundary_end_pfn[nid] == 0)
3012 node_boundary_start_pfn[nid] = -1UL;
3014 /* Update the boundaries */
3015 if (node_boundary_start_pfn[nid] > start_pfn)
3016 node_boundary_start_pfn[nid] = start_pfn;
3017 if (node_boundary_end_pfn[nid] < end_pfn)
3018 node_boundary_end_pfn[nid] = end_pfn;
3021 /* If necessary, push the node boundary out for reserve hotadd */
3022 static void __meminit account_node_boundary(unsigned int nid,
3023 unsigned long *start_pfn, unsigned long *end_pfn)
3025 printk(KERN_DEBUG "Entering account_node_boundary(%u, %lu, %lu)\n",
3026 nid, *start_pfn, *end_pfn);
3028 /* Return if boundary information has not been provided */
3029 if (node_boundary_end_pfn[nid] == 0)
3032 /* Check the boundaries and update if necessary */
3033 if (node_boundary_start_pfn[nid] < *start_pfn)
3034 *start_pfn = node_boundary_start_pfn[nid];
3035 if (node_boundary_end_pfn[nid] > *end_pfn)
3036 *end_pfn = node_boundary_end_pfn[nid];
3039 void __init push_node_boundaries(unsigned int nid,
3040 unsigned long start_pfn, unsigned long end_pfn) {}
3042 static void __meminit account_node_boundary(unsigned int nid,
3043 unsigned long *start_pfn, unsigned long *end_pfn) {}
3048 * get_pfn_range_for_nid - Return the start and end page frames for a node
3049 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3050 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3051 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3053 * It returns the start and end page frame of a node based on information
3054 * provided by an arch calling add_active_range(). If called for a node
3055 * with no available memory, a warning is printed and the start and end
3058 void __meminit get_pfn_range_for_nid(unsigned int nid,
3059 unsigned long *start_pfn, unsigned long *end_pfn)
3065 for_each_active_range_index_in_nid(i, nid) {
3066 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3067 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3070 if (*start_pfn == -1UL)
3073 /* Push the node boundaries out if requested */
3074 account_node_boundary(nid, start_pfn, end_pfn);
3078 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3079 * assumption is made that zones within a node are ordered in monotonic
3080 * increasing memory addresses so that the "highest" populated zone is used
3082 void __init find_usable_zone_for_movable(void)
3085 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3086 if (zone_index == ZONE_MOVABLE)
3089 if (arch_zone_highest_possible_pfn[zone_index] >
3090 arch_zone_lowest_possible_pfn[zone_index])
3094 VM_BUG_ON(zone_index == -1);
3095 movable_zone = zone_index;
3099 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3100 * because it is sized independant of architecture. Unlike the other zones,
3101 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3102 * in each node depending on the size of each node and how evenly kernelcore
3103 * is distributed. This helper function adjusts the zone ranges
3104 * provided by the architecture for a given node by using the end of the
3105 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3106 * zones within a node are in order of monotonic increases memory addresses
3108 void __meminit adjust_zone_range_for_zone_movable(int nid,
3109 unsigned long zone_type,
3110 unsigned long node_start_pfn,
3111 unsigned long node_end_pfn,
3112 unsigned long *zone_start_pfn,
3113 unsigned long *zone_end_pfn)
3115 /* Only adjust if ZONE_MOVABLE is on this node */
3116 if (zone_movable_pfn[nid]) {
3117 /* Size ZONE_MOVABLE */
3118 if (zone_type == ZONE_MOVABLE) {
3119 *zone_start_pfn = zone_movable_pfn[nid];
3120 *zone_end_pfn = min(node_end_pfn,
3121 arch_zone_highest_possible_pfn[movable_zone]);
3123 /* Adjust for ZONE_MOVABLE starting within this range */
3124 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3125 *zone_end_pfn > zone_movable_pfn[nid]) {
3126 *zone_end_pfn = zone_movable_pfn[nid];
3128 /* Check if this whole range is within ZONE_MOVABLE */
3129 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3130 *zone_start_pfn = *zone_end_pfn;
3135 * Return the number of pages a zone spans in a node, including holes
3136 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3138 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3139 unsigned long zone_type,
3140 unsigned long *ignored)
3142 unsigned long node_start_pfn, node_end_pfn;
3143 unsigned long zone_start_pfn, zone_end_pfn;
3145 /* Get the start and end of the node and zone */
3146 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3147 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3148 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3149 adjust_zone_range_for_zone_movable(nid, zone_type,
3150 node_start_pfn, node_end_pfn,
3151 &zone_start_pfn, &zone_end_pfn);
3153 /* Check that this node has pages within the zone's required range */
3154 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3157 /* Move the zone boundaries inside the node if necessary */
3158 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3159 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3161 /* Return the spanned pages */
3162 return zone_end_pfn - zone_start_pfn;
3166 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3167 * then all holes in the requested range will be accounted for.
3169 unsigned long __meminit __absent_pages_in_range(int nid,
3170 unsigned long range_start_pfn,
3171 unsigned long range_end_pfn)
3174 unsigned long prev_end_pfn = 0, hole_pages = 0;
3175 unsigned long start_pfn;
3177 /* Find the end_pfn of the first active range of pfns in the node */
3178 i = first_active_region_index_in_nid(nid);
3182 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3184 /* Account for ranges before physical memory on this node */
3185 if (early_node_map[i].start_pfn > range_start_pfn)
3186 hole_pages = prev_end_pfn - range_start_pfn;
3188 /* Find all holes for the zone within the node */
3189 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3191 /* No need to continue if prev_end_pfn is outside the zone */
3192 if (prev_end_pfn >= range_end_pfn)
3195 /* Make sure the end of the zone is not within the hole */
3196 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3197 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3199 /* Update the hole size cound and move on */
3200 if (start_pfn > range_start_pfn) {
3201 BUG_ON(prev_end_pfn > start_pfn);
3202 hole_pages += start_pfn - prev_end_pfn;
3204 prev_end_pfn = early_node_map[i].end_pfn;
3207 /* Account for ranges past physical memory on this node */
3208 if (range_end_pfn > prev_end_pfn)
3209 hole_pages += range_end_pfn -
3210 max(range_start_pfn, prev_end_pfn);
3216 * absent_pages_in_range - Return number of page frames in holes within a range
3217 * @start_pfn: The start PFN to start searching for holes
3218 * @end_pfn: The end PFN to stop searching for holes
3220 * It returns the number of pages frames in memory holes within a range.
3222 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3223 unsigned long end_pfn)
3225 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3228 /* Return the number of page frames in holes in a zone on a node */
3229 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3230 unsigned long zone_type,
3231 unsigned long *ignored)
3233 unsigned long node_start_pfn, node_end_pfn;
3234 unsigned long zone_start_pfn, zone_end_pfn;
3236 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3237 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3239 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3242 adjust_zone_range_for_zone_movable(nid, zone_type,
3243 node_start_pfn, node_end_pfn,
3244 &zone_start_pfn, &zone_end_pfn);
3245 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3249 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3250 unsigned long zone_type,
3251 unsigned long *zones_size)
3253 return zones_size[zone_type];
3256 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3257 unsigned long zone_type,
3258 unsigned long *zholes_size)
3263 return zholes_size[zone_type];
3268 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3269 unsigned long *zones_size, unsigned long *zholes_size)
3271 unsigned long realtotalpages, totalpages = 0;
3274 for (i = 0; i < MAX_NR_ZONES; i++)
3275 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3277 pgdat->node_spanned_pages = totalpages;
3279 realtotalpages = totalpages;
3280 for (i = 0; i < MAX_NR_ZONES; i++)
3282 zone_absent_pages_in_node(pgdat->node_id, i,
3284 pgdat->node_present_pages = realtotalpages;
3285 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3289 #ifndef CONFIG_SPARSEMEM
3291 * Calculate the size of the zone->blockflags rounded to an unsigned long
3292 * Start by making sure zonesize is a multiple of MAX_ORDER-1 by rounding up
3293 * Then figure 1 NR_PAGEBLOCK_BITS worth of bits per MAX_ORDER-1, finally
3294 * round what is now in bits to nearest long in bits, then return it in
3297 static unsigned long __init usemap_size(unsigned long zonesize)
3299 unsigned long usemapsize;
3301 usemapsize = roundup(zonesize, MAX_ORDER_NR_PAGES);
3302 usemapsize = usemapsize >> (MAX_ORDER-1);
3303 usemapsize *= NR_PAGEBLOCK_BITS;
3304 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
3306 return usemapsize / 8;
3309 static void __init setup_usemap(struct pglist_data *pgdat,
3310 struct zone *zone, unsigned long zonesize)
3312 unsigned long usemapsize = usemap_size(zonesize);
3313 zone->pageblock_flags = NULL;
3315 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
3316 memset(zone->pageblock_flags, 0, usemapsize);
3320 static void inline setup_usemap(struct pglist_data *pgdat,
3321 struct zone *zone, unsigned long zonesize) {}
3322 #endif /* CONFIG_SPARSEMEM */
3325 * Set up the zone data structures:
3326 * - mark all pages reserved
3327 * - mark all memory queues empty
3328 * - clear the memory bitmaps
3330 static void __meminit free_area_init_core(struct pglist_data *pgdat,
3331 unsigned long *zones_size, unsigned long *zholes_size)
3334 int nid = pgdat->node_id;
3335 unsigned long zone_start_pfn = pgdat->node_start_pfn;
3338 pgdat_resize_init(pgdat);
3339 pgdat->nr_zones = 0;
3340 init_waitqueue_head(&pgdat->kswapd_wait);
3341 pgdat->kswapd_max_order = 0;
3343 for (j = 0; j < MAX_NR_ZONES; j++) {
3344 struct zone *zone = pgdat->node_zones + j;
3345 unsigned long size, realsize, memmap_pages;
3347 size = zone_spanned_pages_in_node(nid, j, zones_size);
3348 realsize = size - zone_absent_pages_in_node(nid, j,
3352 * Adjust realsize so that it accounts for how much memory
3353 * is used by this zone for memmap. This affects the watermark
3354 * and per-cpu initialisations
3356 memmap_pages = (size * sizeof(struct page)) >> PAGE_SHIFT;
3357 if (realsize >= memmap_pages) {
3358 realsize -= memmap_pages;
3360 " %s zone: %lu pages used for memmap\n",
3361 zone_names[j], memmap_pages);
3364 " %s zone: %lu pages exceeds realsize %lu\n",
3365 zone_names[j], memmap_pages, realsize);
3367 /* Account for reserved pages */
3368 if (j == 0 && realsize > dma_reserve) {
3369 realsize -= dma_reserve;
3370 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
3371 zone_names[0], dma_reserve);
3374 if (!is_highmem_idx(j))
3375 nr_kernel_pages += realsize;
3376 nr_all_pages += realsize;
3378 zone->spanned_pages = size;
3379 zone->present_pages = realsize;
3382 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
3384 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
3386 zone->name = zone_names[j];
3387 spin_lock_init(&zone->lock);
3388 spin_lock_init(&zone->lru_lock);
3389 zone_seqlock_init(zone);
3390 zone->zone_pgdat = pgdat;
3392 zone->prev_priority = DEF_PRIORITY;
3394 zone_pcp_init(zone);
3395 INIT_LIST_HEAD(&zone->active_list);
3396 INIT_LIST_HEAD(&zone->inactive_list);
3397 zone->nr_scan_active = 0;
3398 zone->nr_scan_inactive = 0;
3399 zap_zone_vm_stats(zone);
3400 atomic_set(&zone->reclaim_in_progress, 0);
3404 setup_usemap(pgdat, zone, size);
3405 ret = init_currently_empty_zone(zone, zone_start_pfn,
3406 size, MEMMAP_EARLY);
3408 zone_start_pfn += size;
3412 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
3414 /* Skip empty nodes */
3415 if (!pgdat->node_spanned_pages)
3418 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3419 /* ia64 gets its own node_mem_map, before this, without bootmem */
3420 if (!pgdat->node_mem_map) {
3421 unsigned long size, start, end;
3425 * The zone's endpoints aren't required to be MAX_ORDER
3426 * aligned but the node_mem_map endpoints must be in order
3427 * for the buddy allocator to function correctly.
3429 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
3430 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
3431 end = ALIGN(end, MAX_ORDER_NR_PAGES);
3432 size = (end - start) * sizeof(struct page);
3433 map = alloc_remap(pgdat->node_id, size);
3435 map = alloc_bootmem_node(pgdat, size);
3436 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
3438 #ifndef CONFIG_NEED_MULTIPLE_NODES
3440 * With no DISCONTIG, the global mem_map is just set as node 0's
3442 if (pgdat == NODE_DATA(0)) {
3443 mem_map = NODE_DATA(0)->node_mem_map;
3444 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3445 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
3446 mem_map -= pgdat->node_start_pfn;
3447 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3450 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3453 void __meminit free_area_init_node(int nid, struct pglist_data *pgdat,
3454 unsigned long *zones_size, unsigned long node_start_pfn,
3455 unsigned long *zholes_size)
3457 pgdat->node_id = nid;
3458 pgdat->node_start_pfn = node_start_pfn;
3459 calculate_node_totalpages(pgdat, zones_size, zholes_size);
3461 alloc_node_mem_map(pgdat);
3463 free_area_init_core(pgdat, zones_size, zholes_size);
3466 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3468 #if MAX_NUMNODES > 1
3470 * Figure out the number of possible node ids.
3472 static void __init setup_nr_node_ids(void)
3475 unsigned int highest = 0;
3477 for_each_node_mask(node, node_possible_map)
3479 nr_node_ids = highest + 1;
3482 static inline void setup_nr_node_ids(void)
3488 * add_active_range - Register a range of PFNs backed by physical memory
3489 * @nid: The node ID the range resides on
3490 * @start_pfn: The start PFN of the available physical memory
3491 * @end_pfn: The end PFN of the available physical memory
3493 * These ranges are stored in an early_node_map[] and later used by
3494 * free_area_init_nodes() to calculate zone sizes and holes. If the
3495 * range spans a memory hole, it is up to the architecture to ensure
3496 * the memory is not freed by the bootmem allocator. If possible
3497 * the range being registered will be merged with existing ranges.
3499 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
3500 unsigned long end_pfn)
3504 printk(KERN_DEBUG "Entering add_active_range(%d, %lu, %lu) "
3505 "%d entries of %d used\n",
3506 nid, start_pfn, end_pfn,
3507 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
3509 /* Merge with existing active regions if possible */
3510 for (i = 0; i < nr_nodemap_entries; i++) {
3511 if (early_node_map[i].nid != nid)
3514 /* Skip if an existing region covers this new one */
3515 if (start_pfn >= early_node_map[i].start_pfn &&
3516 end_pfn <= early_node_map[i].end_pfn)
3519 /* Merge forward if suitable */
3520 if (start_pfn <= early_node_map[i].end_pfn &&
3521 end_pfn > early_node_map[i].end_pfn) {
3522 early_node_map[i].end_pfn = end_pfn;
3526 /* Merge backward if suitable */
3527 if (start_pfn < early_node_map[i].end_pfn &&
3528 end_pfn >= early_node_map[i].start_pfn) {
3529 early_node_map[i].start_pfn = start_pfn;
3534 /* Check that early_node_map is large enough */
3535 if (i >= MAX_ACTIVE_REGIONS) {
3536 printk(KERN_CRIT "More than %d memory regions, truncating\n",
3537 MAX_ACTIVE_REGIONS);
3541 early_node_map[i].nid = nid;
3542 early_node_map[i].start_pfn = start_pfn;
3543 early_node_map[i].end_pfn = end_pfn;
3544 nr_nodemap_entries = i + 1;
3548 * shrink_active_range - Shrink an existing registered range of PFNs
3549 * @nid: The node id the range is on that should be shrunk
3550 * @old_end_pfn: The old end PFN of the range
3551 * @new_end_pfn: The new PFN of the range
3553 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
3554 * The map is kept at the end physical page range that has already been
3555 * registered with add_active_range(). This function allows an arch to shrink
3556 * an existing registered range.
3558 void __init shrink_active_range(unsigned int nid, unsigned long old_end_pfn,
3559 unsigned long new_end_pfn)
3563 /* Find the old active region end and shrink */
3564 for_each_active_range_index_in_nid(i, nid)
3565 if (early_node_map[i].end_pfn == old_end_pfn) {
3566 early_node_map[i].end_pfn = new_end_pfn;
3572 * remove_all_active_ranges - Remove all currently registered regions
3574 * During discovery, it may be found that a table like SRAT is invalid
3575 * and an alternative discovery method must be used. This function removes
3576 * all currently registered regions.
3578 void __init remove_all_active_ranges(void)
3580 memset(early_node_map, 0, sizeof(early_node_map));
3581 nr_nodemap_entries = 0;
3582 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
3583 memset(node_boundary_start_pfn, 0, sizeof(node_boundary_start_pfn));
3584 memset(node_boundary_end_pfn, 0, sizeof(node_boundary_end_pfn));
3585 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
3588 /* Compare two active node_active_regions */
3589 static int __init cmp_node_active_region(const void *a, const void *b)
3591 struct node_active_region *arange = (struct node_active_region *)a;
3592 struct node_active_region *brange = (struct node_active_region *)b;
3594 /* Done this way to avoid overflows */
3595 if (arange->start_pfn > brange->start_pfn)
3597 if (arange->start_pfn < brange->start_pfn)
3603 /* sort the node_map by start_pfn */
3604 static void __init sort_node_map(void)
3606 sort(early_node_map, (size_t)nr_nodemap_entries,
3607 sizeof(struct node_active_region),
3608 cmp_node_active_region, NULL);
3611 /* Find the lowest pfn for a node */
3612 unsigned long __init find_min_pfn_for_node(unsigned long nid)
3615 unsigned long min_pfn = ULONG_MAX;
3617 /* Assuming a sorted map, the first range found has the starting pfn */
3618 for_each_active_range_index_in_nid(i, nid)
3619 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
3621 if (min_pfn == ULONG_MAX) {
3623 "Could not find start_pfn for node %lu\n", nid);
3631 * find_min_pfn_with_active_regions - Find the minimum PFN registered
3633 * It returns the minimum PFN based on information provided via
3634 * add_active_range().
3636 unsigned long __init find_min_pfn_with_active_regions(void)
3638 return find_min_pfn_for_node(MAX_NUMNODES);
3642 * find_max_pfn_with_active_regions - Find the maximum PFN registered
3644 * It returns the maximum PFN based on information provided via
3645 * add_active_range().
3647 unsigned long __init find_max_pfn_with_active_regions(void)
3650 unsigned long max_pfn = 0;
3652 for (i = 0; i < nr_nodemap_entries; i++)
3653 max_pfn = max(max_pfn, early_node_map[i].end_pfn);
3659 * early_calculate_totalpages()
3660 * Sum pages in active regions for movable zone.
3661 * Populate N_HIGH_MEMORY for calculating usable_nodes.
3663 unsigned long __init early_calculate_totalpages(void)
3666 unsigned long totalpages = 0;
3668 for (i = 0; i < nr_nodemap_entries; i++) {
3669 unsigned long pages = early_node_map[i].end_pfn -
3670 early_node_map[i].start_pfn;
3671 totalpages += pages;
3673 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
3679 * Find the PFN the Movable zone begins in each node. Kernel memory
3680 * is spread evenly between nodes as long as the nodes have enough
3681 * memory. When they don't, some nodes will have more kernelcore than
3684 void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
3687 unsigned long usable_startpfn;
3688 unsigned long kernelcore_node, kernelcore_remaining;
3689 unsigned long totalpages = early_calculate_totalpages();
3690 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
3693 * If movablecore was specified, calculate what size of
3694 * kernelcore that corresponds so that memory usable for
3695 * any allocation type is evenly spread. If both kernelcore
3696 * and movablecore are specified, then the value of kernelcore
3697 * will be used for required_kernelcore if it's greater than
3698 * what movablecore would have allowed.
3700 if (required_movablecore) {
3701 unsigned long corepages;
3704 * Round-up so that ZONE_MOVABLE is at least as large as what
3705 * was requested by the user
3707 required_movablecore =
3708 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
3709 corepages = totalpages - required_movablecore;
3711 required_kernelcore = max(required_kernelcore, corepages);
3714 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
3715 if (!required_kernelcore)
3718 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
3719 find_usable_zone_for_movable();
3720 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
3723 /* Spread kernelcore memory as evenly as possible throughout nodes */
3724 kernelcore_node = required_kernelcore / usable_nodes;
3725 for_each_node_state(nid, N_HIGH_MEMORY) {
3727 * Recalculate kernelcore_node if the division per node
3728 * now exceeds what is necessary to satisfy the requested
3729 * amount of memory for the kernel
3731 if (required_kernelcore < kernelcore_node)
3732 kernelcore_node = required_kernelcore / usable_nodes;
3735 * As the map is walked, we track how much memory is usable
3736 * by the kernel using kernelcore_remaining. When it is
3737 * 0, the rest of the node is usable by ZONE_MOVABLE
3739 kernelcore_remaining = kernelcore_node;
3741 /* Go through each range of PFNs within this node */
3742 for_each_active_range_index_in_nid(i, nid) {
3743 unsigned long start_pfn, end_pfn;
3744 unsigned long size_pages;
3746 start_pfn = max(early_node_map[i].start_pfn,
3747 zone_movable_pfn[nid]);
3748 end_pfn = early_node_map[i].end_pfn;
3749 if (start_pfn >= end_pfn)
3752 /* Account for what is only usable for kernelcore */
3753 if (start_pfn < usable_startpfn) {
3754 unsigned long kernel_pages;
3755 kernel_pages = min(end_pfn, usable_startpfn)
3758 kernelcore_remaining -= min(kernel_pages,
3759 kernelcore_remaining);
3760 required_kernelcore -= min(kernel_pages,
3761 required_kernelcore);
3763 /* Continue if range is now fully accounted */
3764 if (end_pfn <= usable_startpfn) {
3767 * Push zone_movable_pfn to the end so
3768 * that if we have to rebalance
3769 * kernelcore across nodes, we will
3770 * not double account here
3772 zone_movable_pfn[nid] = end_pfn;
3775 start_pfn = usable_startpfn;
3779 * The usable PFN range for ZONE_MOVABLE is from
3780 * start_pfn->end_pfn. Calculate size_pages as the
3781 * number of pages used as kernelcore
3783 size_pages = end_pfn - start_pfn;
3784 if (size_pages > kernelcore_remaining)
3785 size_pages = kernelcore_remaining;
3786 zone_movable_pfn[nid] = start_pfn + size_pages;
3789 * Some kernelcore has been met, update counts and
3790 * break if the kernelcore for this node has been
3793 required_kernelcore -= min(required_kernelcore,
3795 kernelcore_remaining -= size_pages;
3796 if (!kernelcore_remaining)
3802 * If there is still required_kernelcore, we do another pass with one
3803 * less node in the count. This will push zone_movable_pfn[nid] further
3804 * along on the nodes that still have memory until kernelcore is
3808 if (usable_nodes && required_kernelcore > usable_nodes)
3811 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
3812 for (nid = 0; nid < MAX_NUMNODES; nid++)
3813 zone_movable_pfn[nid] =
3814 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
3817 /* Any regular memory on that node ? */
3818 static void check_for_regular_memory(pg_data_t *pgdat)
3820 #ifdef CONFIG_HIGHMEM
3821 enum zone_type zone_type;
3823 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
3824 struct zone *zone = &pgdat->node_zones[zone_type];
3825 if (zone->present_pages)
3826 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
3832 * free_area_init_nodes - Initialise all pg_data_t and zone data
3833 * @max_zone_pfn: an array of max PFNs for each zone
3835 * This will call free_area_init_node() for each active node in the system.
3836 * Using the page ranges provided by add_active_range(), the size of each
3837 * zone in each node and their holes is calculated. If the maximum PFN
3838 * between two adjacent zones match, it is assumed that the zone is empty.
3839 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
3840 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
3841 * starts where the previous one ended. For example, ZONE_DMA32 starts
3842 * at arch_max_dma_pfn.
3844 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
3849 /* Sort early_node_map as initialisation assumes it is sorted */
3852 /* Record where the zone boundaries are */
3853 memset(arch_zone_lowest_possible_pfn, 0,
3854 sizeof(arch_zone_lowest_possible_pfn));
3855 memset(arch_zone_highest_possible_pfn, 0,
3856 sizeof(arch_zone_highest_possible_pfn));
3857 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
3858 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
3859 for (i = 1; i < MAX_NR_ZONES; i++) {
3860 if (i == ZONE_MOVABLE)
3862 arch_zone_lowest_possible_pfn[i] =
3863 arch_zone_highest_possible_pfn[i-1];
3864 arch_zone_highest_possible_pfn[i] =
3865 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
3867 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
3868 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
3870 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
3871 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
3872 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
3874 /* Print out the zone ranges */
3875 printk("Zone PFN ranges:\n");
3876 for (i = 0; i < MAX_NR_ZONES; i++) {
3877 if (i == ZONE_MOVABLE)
3879 printk(" %-8s %8lu -> %8lu\n",
3881 arch_zone_lowest_possible_pfn[i],
3882 arch_zone_highest_possible_pfn[i]);
3885 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
3886 printk("Movable zone start PFN for each node\n");
3887 for (i = 0; i < MAX_NUMNODES; i++) {
3888 if (zone_movable_pfn[i])
3889 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
3892 /* Print out the early_node_map[] */
3893 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
3894 for (i = 0; i < nr_nodemap_entries; i++)
3895 printk(" %3d: %8lu -> %8lu\n", early_node_map[i].nid,
3896 early_node_map[i].start_pfn,
3897 early_node_map[i].end_pfn);
3899 /* Initialise every node */
3900 setup_nr_node_ids();
3901 for_each_online_node(nid) {
3902 pg_data_t *pgdat = NODE_DATA(nid);
3903 free_area_init_node(nid, pgdat, NULL,
3904 find_min_pfn_for_node(nid), NULL);
3906 /* Any memory on that node */
3907 if (pgdat->node_present_pages)
3908 node_set_state(nid, N_HIGH_MEMORY);
3909 check_for_regular_memory(pgdat);
3913 static int __init cmdline_parse_core(char *p, unsigned long *core)
3915 unsigned long long coremem;
3919 coremem = memparse(p, &p);
3920 *core = coremem >> PAGE_SHIFT;
3922 /* Paranoid check that UL is enough for the coremem value */
3923 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
3929 * kernelcore=size sets the amount of memory for use for allocations that
3930 * cannot be reclaimed or migrated.
3932 static int __init cmdline_parse_kernelcore(char *p)
3934 return cmdline_parse_core(p, &required_kernelcore);
3938 * movablecore=size sets the amount of memory for use for allocations that
3939 * can be reclaimed or migrated.
3941 static int __init cmdline_parse_movablecore(char *p)
3943 return cmdline_parse_core(p, &required_movablecore);
3946 early_param("kernelcore", cmdline_parse_kernelcore);
3947 early_param("movablecore", cmdline_parse_movablecore);
3949 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3952 * set_dma_reserve - set the specified number of pages reserved in the first zone
3953 * @new_dma_reserve: The number of pages to mark reserved
3955 * The per-cpu batchsize and zone watermarks are determined by present_pages.
3956 * In the DMA zone, a significant percentage may be consumed by kernel image
3957 * and other unfreeable allocations which can skew the watermarks badly. This
3958 * function may optionally be used to account for unfreeable pages in the
3959 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
3960 * smaller per-cpu batchsize.
3962 void __init set_dma_reserve(unsigned long new_dma_reserve)
3964 dma_reserve = new_dma_reserve;
3967 #ifndef CONFIG_NEED_MULTIPLE_NODES
3968 static bootmem_data_t contig_bootmem_data;
3969 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
3971 EXPORT_SYMBOL(contig_page_data);
3974 void __init free_area_init(unsigned long *zones_size)
3976 free_area_init_node(0, NODE_DATA(0), zones_size,
3977 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
3980 static int page_alloc_cpu_notify(struct notifier_block *self,
3981 unsigned long action, void *hcpu)
3983 int cpu = (unsigned long)hcpu;
3985 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
3986 local_irq_disable();
3988 vm_events_fold_cpu(cpu);
3990 refresh_cpu_vm_stats(cpu);
3995 void __init page_alloc_init(void)
3997 hotcpu_notifier(page_alloc_cpu_notify, 0);
4001 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4002 * or min_free_kbytes changes.
4004 static void calculate_totalreserve_pages(void)
4006 struct pglist_data *pgdat;
4007 unsigned long reserve_pages = 0;
4008 enum zone_type i, j;
4010 for_each_online_pgdat(pgdat) {
4011 for (i = 0; i < MAX_NR_ZONES; i++) {
4012 struct zone *zone = pgdat->node_zones + i;
4013 unsigned long max = 0;
4015 /* Find valid and maximum lowmem_reserve in the zone */
4016 for (j = i; j < MAX_NR_ZONES; j++) {
4017 if (zone->lowmem_reserve[j] > max)
4018 max = zone->lowmem_reserve[j];
4021 /* we treat pages_high as reserved pages. */
4022 max += zone->pages_high;
4024 if (max > zone->present_pages)
4025 max = zone->present_pages;
4026 reserve_pages += max;
4029 totalreserve_pages = reserve_pages;
4033 * setup_per_zone_lowmem_reserve - called whenever
4034 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4035 * has a correct pages reserved value, so an adequate number of
4036 * pages are left in the zone after a successful __alloc_pages().
4038 static void setup_per_zone_lowmem_reserve(void)
4040 struct pglist_data *pgdat;
4041 enum zone_type j, idx;
4043 for_each_online_pgdat(pgdat) {
4044 for (j = 0; j < MAX_NR_ZONES; j++) {
4045 struct zone *zone = pgdat->node_zones + j;
4046 unsigned long present_pages = zone->present_pages;
4048 zone->lowmem_reserve[j] = 0;
4052 struct zone *lower_zone;
4056 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4057 sysctl_lowmem_reserve_ratio[idx] = 1;
4059 lower_zone = pgdat->node_zones + idx;
4060 lower_zone->lowmem_reserve[j] = present_pages /
4061 sysctl_lowmem_reserve_ratio[idx];
4062 present_pages += lower_zone->present_pages;
4067 /* update totalreserve_pages */
4068 calculate_totalreserve_pages();
4072 * setup_per_zone_pages_min - called when min_free_kbytes changes.
4074 * Ensures that the pages_{min,low,high} values for each zone are set correctly
4075 * with respect to min_free_kbytes.
4077 void setup_per_zone_pages_min(void)
4079 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4080 unsigned long lowmem_pages = 0;
4082 unsigned long flags;
4084 /* Calculate total number of !ZONE_HIGHMEM pages */
4085 for_each_zone(zone) {
4086 if (!is_highmem(zone))
4087 lowmem_pages += zone->present_pages;
4090 for_each_zone(zone) {
4093 spin_lock_irqsave(&zone->lru_lock, flags);
4094 tmp = (u64)pages_min * zone->present_pages;
4095 do_div(tmp, lowmem_pages);
4096 if (is_highmem(zone)) {
4098 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4099 * need highmem pages, so cap pages_min to a small
4102 * The (pages_high-pages_low) and (pages_low-pages_min)
4103 * deltas controls asynch page reclaim, and so should
4104 * not be capped for highmem.
4108 min_pages = zone->present_pages / 1024;
4109 if (min_pages < SWAP_CLUSTER_MAX)
4110 min_pages = SWAP_CLUSTER_MAX;
4111 if (min_pages > 128)
4113 zone->pages_min = min_pages;
4116 * If it's a lowmem zone, reserve a number of pages
4117 * proportionate to the zone's size.
4119 zone->pages_min = tmp;
4122 zone->pages_low = zone->pages_min + (tmp >> 2);
4123 zone->pages_high = zone->pages_min + (tmp >> 1);
4124 setup_zone_migrate_reserve(zone);
4125 spin_unlock_irqrestore(&zone->lru_lock, flags);
4128 /* update totalreserve_pages */
4129 calculate_totalreserve_pages();
4133 * Initialise min_free_kbytes.
4135 * For small machines we want it small (128k min). For large machines
4136 * we want it large (64MB max). But it is not linear, because network
4137 * bandwidth does not increase linearly with machine size. We use
4139 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4140 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4156 static int __init init_per_zone_pages_min(void)
4158 unsigned long lowmem_kbytes;
4160 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4162 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4163 if (min_free_kbytes < 128)
4164 min_free_kbytes = 128;
4165 if (min_free_kbytes > 65536)
4166 min_free_kbytes = 65536;
4167 setup_per_zone_pages_min();
4168 setup_per_zone_lowmem_reserve();
4171 module_init(init_per_zone_pages_min)
4174 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4175 * that we can call two helper functions whenever min_free_kbytes
4178 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
4179 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4181 proc_dointvec(table, write, file, buffer, length, ppos);
4183 setup_per_zone_pages_min();
4188 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
4189 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4194 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4199 zone->min_unmapped_pages = (zone->present_pages *
4200 sysctl_min_unmapped_ratio) / 100;
4204 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
4205 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4210 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4215 zone->min_slab_pages = (zone->present_pages *
4216 sysctl_min_slab_ratio) / 100;
4222 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4223 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4224 * whenever sysctl_lowmem_reserve_ratio changes.
4226 * The reserve ratio obviously has absolutely no relation with the
4227 * pages_min watermarks. The lowmem reserve ratio can only make sense
4228 * if in function of the boot time zone sizes.
4230 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
4231 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4233 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4234 setup_per_zone_lowmem_reserve();
4239 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4240 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4241 * can have before it gets flushed back to buddy allocator.
4244 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
4245 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4251 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4252 if (!write || (ret == -EINVAL))
4254 for_each_zone(zone) {
4255 for_each_online_cpu(cpu) {
4257 high = zone->present_pages / percpu_pagelist_fraction;
4258 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
4264 int hashdist = HASHDIST_DEFAULT;
4267 static int __init set_hashdist(char *str)
4271 hashdist = simple_strtoul(str, &str, 0);
4274 __setup("hashdist=", set_hashdist);
4278 * allocate a large system hash table from bootmem
4279 * - it is assumed that the hash table must contain an exact power-of-2
4280 * quantity of entries
4281 * - limit is the number of hash buckets, not the total allocation size
4283 void *__init alloc_large_system_hash(const char *tablename,
4284 unsigned long bucketsize,
4285 unsigned long numentries,
4288 unsigned int *_hash_shift,
4289 unsigned int *_hash_mask,
4290 unsigned long limit)
4292 unsigned long long max = limit;
4293 unsigned long log2qty, size;
4296 /* allow the kernel cmdline to have a say */
4298 /* round applicable memory size up to nearest megabyte */
4299 numentries = nr_kernel_pages;
4300 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
4301 numentries >>= 20 - PAGE_SHIFT;
4302 numentries <<= 20 - PAGE_SHIFT;
4304 /* limit to 1 bucket per 2^scale bytes of low memory */
4305 if (scale > PAGE_SHIFT)
4306 numentries >>= (scale - PAGE_SHIFT);
4308 numentries <<= (PAGE_SHIFT - scale);
4310 /* Make sure we've got at least a 0-order allocation.. */
4311 if (unlikely((numentries * bucketsize) < PAGE_SIZE))
4312 numentries = PAGE_SIZE / bucketsize;
4314 numentries = roundup_pow_of_two(numentries);
4316 /* limit allocation size to 1/16 total memory by default */
4318 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
4319 do_div(max, bucketsize);
4322 if (numentries > max)
4325 log2qty = ilog2(numentries);
4328 size = bucketsize << log2qty;
4329 if (flags & HASH_EARLY)
4330 table = alloc_bootmem(size);
4332 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
4334 unsigned long order;
4335 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
4337 table = (void*) __get_free_pages(GFP_ATOMIC, order);
4339 * If bucketsize is not a power-of-two, we may free
4340 * some pages at the end of hash table.
4343 unsigned long alloc_end = (unsigned long)table +
4344 (PAGE_SIZE << order);
4345 unsigned long used = (unsigned long)table +
4347 split_page(virt_to_page(table), order);
4348 while (used < alloc_end) {
4354 } while (!table && size > PAGE_SIZE && --log2qty);
4357 panic("Failed to allocate %s hash table\n", tablename);
4359 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
4362 ilog2(size) - PAGE_SHIFT,
4366 *_hash_shift = log2qty;
4368 *_hash_mask = (1 << log2qty) - 1;
4373 #ifdef CONFIG_OUT_OF_LINE_PFN_TO_PAGE
4374 struct page *pfn_to_page(unsigned long pfn)
4376 return __pfn_to_page(pfn);
4378 unsigned long page_to_pfn(struct page *page)
4380 return __page_to_pfn(page);
4382 EXPORT_SYMBOL(pfn_to_page);
4383 EXPORT_SYMBOL(page_to_pfn);
4384 #endif /* CONFIG_OUT_OF_LINE_PFN_TO_PAGE */
4386 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4387 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
4390 #ifdef CONFIG_SPARSEMEM
4391 return __pfn_to_section(pfn)->pageblock_flags;
4393 return zone->pageblock_flags;
4394 #endif /* CONFIG_SPARSEMEM */
4397 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
4399 #ifdef CONFIG_SPARSEMEM
4400 pfn &= (PAGES_PER_SECTION-1);
4401 return (pfn >> (MAX_ORDER-1)) * NR_PAGEBLOCK_BITS;
4403 pfn = pfn - zone->zone_start_pfn;
4404 return (pfn >> (MAX_ORDER-1)) * NR_PAGEBLOCK_BITS;
4405 #endif /* CONFIG_SPARSEMEM */
4409 * get_pageblock_flags_group - Return the requested group of flags for the MAX_ORDER_NR_PAGES block of pages
4410 * @page: The page within the block of interest
4411 * @start_bitidx: The first bit of interest to retrieve
4412 * @end_bitidx: The last bit of interest
4413 * returns pageblock_bits flags
4415 unsigned long get_pageblock_flags_group(struct page *page,
4416 int start_bitidx, int end_bitidx)
4419 unsigned long *bitmap;
4420 unsigned long pfn, bitidx;
4421 unsigned long flags = 0;
4422 unsigned long value = 1;
4424 zone = page_zone(page);
4425 pfn = page_to_pfn(page);
4426 bitmap = get_pageblock_bitmap(zone, pfn);
4427 bitidx = pfn_to_bitidx(zone, pfn);
4429 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4430 if (test_bit(bitidx + start_bitidx, bitmap))
4437 * set_pageblock_flags_group - Set the requested group of flags for a MAX_ORDER_NR_PAGES block of pages
4438 * @page: The page within the block of interest
4439 * @start_bitidx: The first bit of interest
4440 * @end_bitidx: The last bit of interest
4441 * @flags: The flags to set
4443 void set_pageblock_flags_group(struct page *page, unsigned long flags,
4444 int start_bitidx, int end_bitidx)
4447 unsigned long *bitmap;
4448 unsigned long pfn, bitidx;
4449 unsigned long value = 1;
4451 zone = page_zone(page);
4452 pfn = page_to_pfn(page);
4453 bitmap = get_pageblock_bitmap(zone, pfn);
4454 bitidx = pfn_to_bitidx(zone, pfn);
4456 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4458 __set_bit(bitidx + start_bitidx, bitmap);
4460 __clear_bit(bitidx + start_bitidx, bitmap);