2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/compiler.h>
25 #include <linux/kernel.h>
26 #include <linux/module.h>
27 #include <linux/suspend.h>
28 #include <linux/pagevec.h>
29 #include <linux/blkdev.h>
30 #include <linux/slab.h>
31 #include <linux/oom.h>
32 #include <linux/notifier.h>
33 #include <linux/topology.h>
34 #include <linux/sysctl.h>
35 #include <linux/cpu.h>
36 #include <linux/cpuset.h>
37 #include <linux/memory_hotplug.h>
38 #include <linux/nodemask.h>
39 #include <linux/vmalloc.h>
40 #include <linux/mempolicy.h>
41 #include <linux/stop_machine.h>
42 #include <linux/sort.h>
43 #include <linux/pfn.h>
44 #include <linux/backing-dev.h>
45 #include <linux/fault-inject.h>
46 #include <linux/page-isolation.h>
47 #include <linux/memcontrol.h>
48 #include <linux/debugobjects.h>
50 #include <asm/tlbflush.h>
51 #include <asm/div64.h>
55 * Array of node states.
57 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
58 [N_POSSIBLE] = NODE_MASK_ALL,
59 [N_ONLINE] = { { [0] = 1UL } },
61 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
63 [N_HIGH_MEMORY] = { { [0] = 1UL } },
65 [N_CPU] = { { [0] = 1UL } },
68 EXPORT_SYMBOL(node_states);
70 unsigned long totalram_pages __read_mostly;
71 unsigned long totalreserve_pages __read_mostly;
73 int percpu_pagelist_fraction;
75 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
76 int pageblock_order __read_mostly;
79 static void __free_pages_ok(struct page *page, unsigned int order);
82 * results with 256, 32 in the lowmem_reserve sysctl:
83 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
84 * 1G machine -> (16M dma, 784M normal, 224M high)
85 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
86 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
87 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
89 * TBD: should special case ZONE_DMA32 machines here - in those we normally
90 * don't need any ZONE_NORMAL reservation
92 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
93 #ifdef CONFIG_ZONE_DMA
96 #ifdef CONFIG_ZONE_DMA32
105 EXPORT_SYMBOL(totalram_pages);
107 static char * const zone_names[MAX_NR_ZONES] = {
108 #ifdef CONFIG_ZONE_DMA
111 #ifdef CONFIG_ZONE_DMA32
115 #ifdef CONFIG_HIGHMEM
121 int min_free_kbytes = 1024;
123 unsigned long __meminitdata nr_kernel_pages;
124 unsigned long __meminitdata nr_all_pages;
125 static unsigned long __meminitdata dma_reserve;
127 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
129 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
130 * ranges of memory (RAM) that may be registered with add_active_range().
131 * Ranges passed to add_active_range() will be merged if possible
132 * so the number of times add_active_range() can be called is
133 * related to the number of nodes and the number of holes
135 #ifdef CONFIG_MAX_ACTIVE_REGIONS
136 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
137 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
139 #if MAX_NUMNODES >= 32
140 /* If there can be many nodes, allow up to 50 holes per node */
141 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
143 /* By default, allow up to 256 distinct regions */
144 #define MAX_ACTIVE_REGIONS 256
148 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
149 static int __meminitdata nr_nodemap_entries;
150 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
151 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
152 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
153 static unsigned long __meminitdata node_boundary_start_pfn[MAX_NUMNODES];
154 static unsigned long __meminitdata node_boundary_end_pfn[MAX_NUMNODES];
155 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
156 static unsigned long __initdata required_kernelcore;
157 static unsigned long __initdata required_movablecore;
158 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
160 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
162 EXPORT_SYMBOL(movable_zone);
163 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
166 int nr_node_ids __read_mostly = MAX_NUMNODES;
167 EXPORT_SYMBOL(nr_node_ids);
170 int page_group_by_mobility_disabled __read_mostly;
172 static void set_pageblock_migratetype(struct page *page, int migratetype)
174 set_pageblock_flags_group(page, (unsigned long)migratetype,
175 PB_migrate, PB_migrate_end);
178 #ifdef CONFIG_DEBUG_VM
179 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
183 unsigned long pfn = page_to_pfn(page);
186 seq = zone_span_seqbegin(zone);
187 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
189 else if (pfn < zone->zone_start_pfn)
191 } while (zone_span_seqretry(zone, seq));
196 static int page_is_consistent(struct zone *zone, struct page *page)
198 if (!pfn_valid_within(page_to_pfn(page)))
200 if (zone != page_zone(page))
206 * Temporary debugging check for pages not lying within a given zone.
208 static int bad_range(struct zone *zone, struct page *page)
210 if (page_outside_zone_boundaries(zone, page))
212 if (!page_is_consistent(zone, page))
218 static inline int bad_range(struct zone *zone, struct page *page)
224 static void bad_page(struct page *page)
226 void *pc = page_get_page_cgroup(page);
228 printk(KERN_EMERG "Bad page state in process '%s'\n" KERN_EMERG
229 "page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d\n",
230 current->comm, page, (int)(2*sizeof(unsigned long)),
231 (unsigned long)page->flags, page->mapping,
232 page_mapcount(page), page_count(page));
234 printk(KERN_EMERG "cgroup:%p\n", pc);
235 page_reset_bad_cgroup(page);
237 printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n"
238 KERN_EMERG "Backtrace:\n");
240 page->flags &= ~PAGE_FLAGS_CLEAR_WHEN_BAD;
241 set_page_count(page, 0);
242 reset_page_mapcount(page);
243 page->mapping = NULL;
244 add_taint(TAINT_BAD_PAGE);
248 * Higher-order pages are called "compound pages". They are structured thusly:
250 * The first PAGE_SIZE page is called the "head page".
252 * The remaining PAGE_SIZE pages are called "tail pages".
254 * All pages have PG_compound set. All pages have their ->private pointing at
255 * the head page (even the head page has this).
257 * The first tail page's ->lru.next holds the address of the compound page's
258 * put_page() function. Its ->lru.prev holds the order of allocation.
259 * This usage means that zero-order pages may not be compound.
262 static void free_compound_page(struct page *page)
264 __free_pages_ok(page, compound_order(page));
267 void prep_compound_page(struct page *page, unsigned long order)
270 int nr_pages = 1 << order;
271 struct page *p = page + 1;
273 set_compound_page_dtor(page, free_compound_page);
274 set_compound_order(page, order);
276 for (i = 1; i < nr_pages; i++, p++) {
277 if (unlikely((i & (MAX_ORDER_NR_PAGES - 1)) == 0))
278 p = pfn_to_page(page_to_pfn(page) + i);
280 p->first_page = page;
284 static void destroy_compound_page(struct page *page, unsigned long order)
287 int nr_pages = 1 << order;
288 struct page *p = page + 1;
290 if (unlikely(compound_order(page) != order))
293 if (unlikely(!PageHead(page)))
295 __ClearPageHead(page);
296 for (i = 1; i < nr_pages; i++, p++) {
297 if (unlikely((i & (MAX_ORDER_NR_PAGES - 1)) == 0))
298 p = pfn_to_page(page_to_pfn(page) + i);
300 if (unlikely(!PageTail(p) |
301 (p->first_page != page)))
307 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
312 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
313 * and __GFP_HIGHMEM from hard or soft interrupt context.
315 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
316 for (i = 0; i < (1 << order); i++)
317 clear_highpage(page + i);
320 static inline void set_page_order(struct page *page, int order)
322 set_page_private(page, order);
323 __SetPageBuddy(page);
326 static inline void rmv_page_order(struct page *page)
328 __ClearPageBuddy(page);
329 set_page_private(page, 0);
333 * Locate the struct page for both the matching buddy in our
334 * pair (buddy1) and the combined O(n+1) page they form (page).
336 * 1) Any buddy B1 will have an order O twin B2 which satisfies
337 * the following equation:
339 * For example, if the starting buddy (buddy2) is #8 its order
341 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
343 * 2) Any buddy B will have an order O+1 parent P which
344 * satisfies the following equation:
347 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
349 static inline struct page *
350 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
352 unsigned long buddy_idx = page_idx ^ (1 << order);
354 return page + (buddy_idx - page_idx);
357 static inline unsigned long
358 __find_combined_index(unsigned long page_idx, unsigned int order)
360 return (page_idx & ~(1 << order));
364 * This function checks whether a page is free && is the buddy
365 * we can do coalesce a page and its buddy if
366 * (a) the buddy is not in a hole &&
367 * (b) the buddy is in the buddy system &&
368 * (c) a page and its buddy have the same order &&
369 * (d) a page and its buddy are in the same zone.
371 * For recording whether a page is in the buddy system, we use PG_buddy.
372 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
374 * For recording page's order, we use page_private(page).
376 static inline int page_is_buddy(struct page *page, struct page *buddy,
379 if (!pfn_valid_within(page_to_pfn(buddy)))
382 if (page_zone_id(page) != page_zone_id(buddy))
385 if (PageBuddy(buddy) && page_order(buddy) == order) {
386 BUG_ON(page_count(buddy) != 0);
393 * Freeing function for a buddy system allocator.
395 * The concept of a buddy system is to maintain direct-mapped table
396 * (containing bit values) for memory blocks of various "orders".
397 * The bottom level table contains the map for the smallest allocatable
398 * units of memory (here, pages), and each level above it describes
399 * pairs of units from the levels below, hence, "buddies".
400 * At a high level, all that happens here is marking the table entry
401 * at the bottom level available, and propagating the changes upward
402 * as necessary, plus some accounting needed to play nicely with other
403 * parts of the VM system.
404 * At each level, we keep a list of pages, which are heads of continuous
405 * free pages of length of (1 << order) and marked with PG_buddy. Page's
406 * order is recorded in page_private(page) field.
407 * So when we are allocating or freeing one, we can derive the state of the
408 * other. That is, if we allocate a small block, and both were
409 * free, the remainder of the region must be split into blocks.
410 * If a block is freed, and its buddy is also free, then this
411 * triggers coalescing into a block of larger size.
416 static inline void __free_one_page(struct page *page,
417 struct zone *zone, unsigned int order)
419 unsigned long page_idx;
420 int order_size = 1 << order;
421 int migratetype = get_pageblock_migratetype(page);
423 if (unlikely(PageCompound(page)))
424 destroy_compound_page(page, order);
426 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
428 VM_BUG_ON(page_idx & (order_size - 1));
429 VM_BUG_ON(bad_range(zone, page));
431 __mod_zone_page_state(zone, NR_FREE_PAGES, order_size);
432 while (order < MAX_ORDER-1) {
433 unsigned long combined_idx;
436 buddy = __page_find_buddy(page, page_idx, order);
437 if (!page_is_buddy(page, buddy, order))
440 /* Our buddy is free, merge with it and move up one order. */
441 list_del(&buddy->lru);
442 zone->free_area[order].nr_free--;
443 rmv_page_order(buddy);
444 combined_idx = __find_combined_index(page_idx, order);
445 page = page + (combined_idx - page_idx);
446 page_idx = combined_idx;
449 set_page_order(page, order);
451 &zone->free_area[order].free_list[migratetype]);
452 zone->free_area[order].nr_free++;
455 static inline int free_pages_check(struct page *page)
457 if (unlikely(page_mapcount(page) |
458 (page->mapping != NULL) |
459 (page_get_page_cgroup(page) != NULL) |
460 (page_count(page) != 0) |
461 (page->flags & PAGE_FLAGS_CHECK_AT_FREE)))
464 __ClearPageDirty(page);
465 if (PageSwapBacked(page))
466 __ClearPageSwapBacked(page);
468 * For now, we report if PG_reserved was found set, but do not
469 * clear it, and do not free the page. But we shall soon need
470 * to do more, for when the ZERO_PAGE count wraps negative.
472 return PageReserved(page);
476 * Frees a list of pages.
477 * Assumes all pages on list are in same zone, and of same order.
478 * count is the number of pages to free.
480 * If the zone was previously in an "all pages pinned" state then look to
481 * see if this freeing clears that state.
483 * And clear the zone's pages_scanned counter, to hold off the "all pages are
484 * pinned" detection logic.
486 static void free_pages_bulk(struct zone *zone, int count,
487 struct list_head *list, int order)
489 spin_lock(&zone->lock);
490 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
491 zone->pages_scanned = 0;
495 VM_BUG_ON(list_empty(list));
496 page = list_entry(list->prev, struct page, lru);
497 /* have to delete it as __free_one_page list manipulates */
498 list_del(&page->lru);
499 __free_one_page(page, zone, order);
501 spin_unlock(&zone->lock);
504 static void free_one_page(struct zone *zone, struct page *page, int order)
506 spin_lock(&zone->lock);
507 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
508 zone->pages_scanned = 0;
509 __free_one_page(page, zone, order);
510 spin_unlock(&zone->lock);
513 static void __free_pages_ok(struct page *page, unsigned int order)
519 for (i = 0 ; i < (1 << order) ; ++i)
520 reserved += free_pages_check(page + i);
524 if (!PageHighMem(page)) {
525 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
526 debug_check_no_obj_freed(page_address(page),
529 arch_free_page(page, order);
530 kernel_map_pages(page, 1 << order, 0);
532 local_irq_save(flags);
533 __count_vm_events(PGFREE, 1 << order);
534 free_one_page(page_zone(page), page, order);
535 local_irq_restore(flags);
539 * permit the bootmem allocator to evade page validation on high-order frees
541 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
544 __ClearPageReserved(page);
545 set_page_count(page, 0);
546 set_page_refcounted(page);
552 for (loop = 0; loop < BITS_PER_LONG; loop++) {
553 struct page *p = &page[loop];
555 if (loop + 1 < BITS_PER_LONG)
557 __ClearPageReserved(p);
558 set_page_count(p, 0);
561 set_page_refcounted(page);
562 __free_pages(page, order);
568 * The order of subdivision here is critical for the IO subsystem.
569 * Please do not alter this order without good reasons and regression
570 * testing. Specifically, as large blocks of memory are subdivided,
571 * the order in which smaller blocks are delivered depends on the order
572 * they're subdivided in this function. This is the primary factor
573 * influencing the order in which pages are delivered to the IO
574 * subsystem according to empirical testing, and this is also justified
575 * by considering the behavior of a buddy system containing a single
576 * large block of memory acted on by a series of small allocations.
577 * This behavior is a critical factor in sglist merging's success.
581 static inline void expand(struct zone *zone, struct page *page,
582 int low, int high, struct free_area *area,
585 unsigned long size = 1 << high;
591 VM_BUG_ON(bad_range(zone, &page[size]));
592 list_add(&page[size].lru, &area->free_list[migratetype]);
594 set_page_order(&page[size], high);
599 * This page is about to be returned from the page allocator
601 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
603 if (unlikely(page_mapcount(page) |
604 (page->mapping != NULL) |
605 (page_get_page_cgroup(page) != NULL) |
606 (page_count(page) != 0) |
607 (page->flags & PAGE_FLAGS_CHECK_AT_PREP)))
611 * For now, we report if PG_reserved was found set, but do not
612 * clear it, and do not allocate the page: as a safety net.
614 if (PageReserved(page))
617 page->flags &= ~(1 << PG_uptodate | 1 << PG_error | 1 << PG_reclaim |
618 1 << PG_referenced | 1 << PG_arch_1 |
619 1 << PG_owner_priv_1 | 1 << PG_mappedtodisk);
620 set_page_private(page, 0);
621 set_page_refcounted(page);
623 arch_alloc_page(page, order);
624 kernel_map_pages(page, 1 << order, 1);
626 if (gfp_flags & __GFP_ZERO)
627 prep_zero_page(page, order, gfp_flags);
629 if (order && (gfp_flags & __GFP_COMP))
630 prep_compound_page(page, order);
636 * Go through the free lists for the given migratetype and remove
637 * the smallest available page from the freelists
639 static struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
642 unsigned int current_order;
643 struct free_area * area;
646 /* Find a page of the appropriate size in the preferred list */
647 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
648 area = &(zone->free_area[current_order]);
649 if (list_empty(&area->free_list[migratetype]))
652 page = list_entry(area->free_list[migratetype].next,
654 list_del(&page->lru);
655 rmv_page_order(page);
657 __mod_zone_page_state(zone, NR_FREE_PAGES, - (1UL << order));
658 expand(zone, page, order, current_order, area, migratetype);
667 * This array describes the order lists are fallen back to when
668 * the free lists for the desirable migrate type are depleted
670 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
671 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
672 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
673 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
674 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
678 * Move the free pages in a range to the free lists of the requested type.
679 * Note that start_page and end_pages are not aligned on a pageblock
680 * boundary. If alignment is required, use move_freepages_block()
682 static int move_freepages(struct zone *zone,
683 struct page *start_page, struct page *end_page,
690 #ifndef CONFIG_HOLES_IN_ZONE
692 * page_zone is not safe to call in this context when
693 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
694 * anyway as we check zone boundaries in move_freepages_block().
695 * Remove at a later date when no bug reports exist related to
696 * grouping pages by mobility
698 BUG_ON(page_zone(start_page) != page_zone(end_page));
701 for (page = start_page; page <= end_page;) {
702 /* Make sure we are not inadvertently changing nodes */
703 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
705 if (!pfn_valid_within(page_to_pfn(page))) {
710 if (!PageBuddy(page)) {
715 order = page_order(page);
716 list_del(&page->lru);
718 &zone->free_area[order].free_list[migratetype]);
720 pages_moved += 1 << order;
726 static int move_freepages_block(struct zone *zone, struct page *page,
729 unsigned long start_pfn, end_pfn;
730 struct page *start_page, *end_page;
732 start_pfn = page_to_pfn(page);
733 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
734 start_page = pfn_to_page(start_pfn);
735 end_page = start_page + pageblock_nr_pages - 1;
736 end_pfn = start_pfn + pageblock_nr_pages - 1;
738 /* Do not cross zone boundaries */
739 if (start_pfn < zone->zone_start_pfn)
741 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
744 return move_freepages(zone, start_page, end_page, migratetype);
747 /* Remove an element from the buddy allocator from the fallback list */
748 static struct page *__rmqueue_fallback(struct zone *zone, int order,
749 int start_migratetype)
751 struct free_area * area;
756 /* Find the largest possible block of pages in the other list */
757 for (current_order = MAX_ORDER-1; current_order >= order;
759 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
760 migratetype = fallbacks[start_migratetype][i];
762 /* MIGRATE_RESERVE handled later if necessary */
763 if (migratetype == MIGRATE_RESERVE)
766 area = &(zone->free_area[current_order]);
767 if (list_empty(&area->free_list[migratetype]))
770 page = list_entry(area->free_list[migratetype].next,
775 * If breaking a large block of pages, move all free
776 * pages to the preferred allocation list. If falling
777 * back for a reclaimable kernel allocation, be more
778 * agressive about taking ownership of free pages
780 if (unlikely(current_order >= (pageblock_order >> 1)) ||
781 start_migratetype == MIGRATE_RECLAIMABLE) {
783 pages = move_freepages_block(zone, page,
786 /* Claim the whole block if over half of it is free */
787 if (pages >= (1 << (pageblock_order-1)))
788 set_pageblock_migratetype(page,
791 migratetype = start_migratetype;
794 /* Remove the page from the freelists */
795 list_del(&page->lru);
796 rmv_page_order(page);
797 __mod_zone_page_state(zone, NR_FREE_PAGES,
800 if (current_order == pageblock_order)
801 set_pageblock_migratetype(page,
804 expand(zone, page, order, current_order, area, migratetype);
809 /* Use MIGRATE_RESERVE rather than fail an allocation */
810 return __rmqueue_smallest(zone, order, MIGRATE_RESERVE);
814 * Do the hard work of removing an element from the buddy allocator.
815 * Call me with the zone->lock already held.
817 static struct page *__rmqueue(struct zone *zone, unsigned int order,
822 page = __rmqueue_smallest(zone, order, migratetype);
825 page = __rmqueue_fallback(zone, order, migratetype);
831 * Obtain a specified number of elements from the buddy allocator, all under
832 * a single hold of the lock, for efficiency. Add them to the supplied list.
833 * Returns the number of new pages which were placed at *list.
835 static int rmqueue_bulk(struct zone *zone, unsigned int order,
836 unsigned long count, struct list_head *list,
841 spin_lock(&zone->lock);
842 for (i = 0; i < count; ++i) {
843 struct page *page = __rmqueue(zone, order, migratetype);
844 if (unlikely(page == NULL))
848 * Split buddy pages returned by expand() are received here
849 * in physical page order. The page is added to the callers and
850 * list and the list head then moves forward. From the callers
851 * perspective, the linked list is ordered by page number in
852 * some conditions. This is useful for IO devices that can
853 * merge IO requests if the physical pages are ordered
856 list_add(&page->lru, list);
857 set_page_private(page, migratetype);
860 spin_unlock(&zone->lock);
866 * Called from the vmstat counter updater to drain pagesets of this
867 * currently executing processor on remote nodes after they have
870 * Note that this function must be called with the thread pinned to
871 * a single processor.
873 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
878 local_irq_save(flags);
879 if (pcp->count >= pcp->batch)
880 to_drain = pcp->batch;
882 to_drain = pcp->count;
883 free_pages_bulk(zone, to_drain, &pcp->list, 0);
884 pcp->count -= to_drain;
885 local_irq_restore(flags);
890 * Drain pages of the indicated processor.
892 * The processor must either be the current processor and the
893 * thread pinned to the current processor or a processor that
896 static void drain_pages(unsigned int cpu)
901 for_each_zone(zone) {
902 struct per_cpu_pageset *pset;
903 struct per_cpu_pages *pcp;
905 if (!populated_zone(zone))
908 pset = zone_pcp(zone, cpu);
911 local_irq_save(flags);
912 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
914 local_irq_restore(flags);
919 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
921 void drain_local_pages(void *arg)
923 drain_pages(smp_processor_id());
927 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
929 void drain_all_pages(void)
931 on_each_cpu(drain_local_pages, NULL, 1);
934 #ifdef CONFIG_HIBERNATION
936 void mark_free_pages(struct zone *zone)
938 unsigned long pfn, max_zone_pfn;
941 struct list_head *curr;
943 if (!zone->spanned_pages)
946 spin_lock_irqsave(&zone->lock, flags);
948 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
949 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
950 if (pfn_valid(pfn)) {
951 struct page *page = pfn_to_page(pfn);
953 if (!swsusp_page_is_forbidden(page))
954 swsusp_unset_page_free(page);
957 for_each_migratetype_order(order, t) {
958 list_for_each(curr, &zone->free_area[order].free_list[t]) {
961 pfn = page_to_pfn(list_entry(curr, struct page, lru));
962 for (i = 0; i < (1UL << order); i++)
963 swsusp_set_page_free(pfn_to_page(pfn + i));
966 spin_unlock_irqrestore(&zone->lock, flags);
968 #endif /* CONFIG_PM */
971 * Free a 0-order page
973 static void free_hot_cold_page(struct page *page, int cold)
975 struct zone *zone = page_zone(page);
976 struct per_cpu_pages *pcp;
980 page->mapping = NULL;
981 if (free_pages_check(page))
984 if (!PageHighMem(page)) {
985 debug_check_no_locks_freed(page_address(page), PAGE_SIZE);
986 debug_check_no_obj_freed(page_address(page), PAGE_SIZE);
988 arch_free_page(page, 0);
989 kernel_map_pages(page, 1, 0);
991 pcp = &zone_pcp(zone, get_cpu())->pcp;
992 local_irq_save(flags);
993 __count_vm_event(PGFREE);
995 list_add_tail(&page->lru, &pcp->list);
997 list_add(&page->lru, &pcp->list);
998 set_page_private(page, get_pageblock_migratetype(page));
1000 if (pcp->count >= pcp->high) {
1001 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
1002 pcp->count -= pcp->batch;
1004 local_irq_restore(flags);
1008 void free_hot_page(struct page *page)
1010 free_hot_cold_page(page, 0);
1013 void free_cold_page(struct page *page)
1015 free_hot_cold_page(page, 1);
1019 * split_page takes a non-compound higher-order page, and splits it into
1020 * n (1<<order) sub-pages: page[0..n]
1021 * Each sub-page must be freed individually.
1023 * Note: this is probably too low level an operation for use in drivers.
1024 * Please consult with lkml before using this in your driver.
1026 void split_page(struct page *page, unsigned int order)
1030 VM_BUG_ON(PageCompound(page));
1031 VM_BUG_ON(!page_count(page));
1032 for (i = 1; i < (1 << order); i++)
1033 set_page_refcounted(page + i);
1037 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1038 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1041 static struct page *buffered_rmqueue(struct zone *preferred_zone,
1042 struct zone *zone, int order, gfp_t gfp_flags)
1044 unsigned long flags;
1046 int cold = !!(gfp_flags & __GFP_COLD);
1048 int migratetype = allocflags_to_migratetype(gfp_flags);
1052 if (likely(order == 0)) {
1053 struct per_cpu_pages *pcp;
1055 pcp = &zone_pcp(zone, cpu)->pcp;
1056 local_irq_save(flags);
1058 pcp->count = rmqueue_bulk(zone, 0,
1059 pcp->batch, &pcp->list, migratetype);
1060 if (unlikely(!pcp->count))
1064 /* Find a page of the appropriate migrate type */
1066 list_for_each_entry_reverse(page, &pcp->list, lru)
1067 if (page_private(page) == migratetype)
1070 list_for_each_entry(page, &pcp->list, lru)
1071 if (page_private(page) == migratetype)
1075 /* Allocate more to the pcp list if necessary */
1076 if (unlikely(&page->lru == &pcp->list)) {
1077 pcp->count += rmqueue_bulk(zone, 0,
1078 pcp->batch, &pcp->list, migratetype);
1079 page = list_entry(pcp->list.next, struct page, lru);
1082 list_del(&page->lru);
1085 spin_lock_irqsave(&zone->lock, flags);
1086 page = __rmqueue(zone, order, migratetype);
1087 spin_unlock(&zone->lock);
1092 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1093 zone_statistics(preferred_zone, zone);
1094 local_irq_restore(flags);
1097 VM_BUG_ON(bad_range(zone, page));
1098 if (prep_new_page(page, order, gfp_flags))
1103 local_irq_restore(flags);
1108 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
1109 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
1110 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
1111 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
1112 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1113 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1114 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1116 #ifdef CONFIG_FAIL_PAGE_ALLOC
1118 static struct fail_page_alloc_attr {
1119 struct fault_attr attr;
1121 u32 ignore_gfp_highmem;
1122 u32 ignore_gfp_wait;
1125 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1127 struct dentry *ignore_gfp_highmem_file;
1128 struct dentry *ignore_gfp_wait_file;
1129 struct dentry *min_order_file;
1131 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1133 } fail_page_alloc = {
1134 .attr = FAULT_ATTR_INITIALIZER,
1135 .ignore_gfp_wait = 1,
1136 .ignore_gfp_highmem = 1,
1140 static int __init setup_fail_page_alloc(char *str)
1142 return setup_fault_attr(&fail_page_alloc.attr, str);
1144 __setup("fail_page_alloc=", setup_fail_page_alloc);
1146 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1148 if (order < fail_page_alloc.min_order)
1150 if (gfp_mask & __GFP_NOFAIL)
1152 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1154 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1157 return should_fail(&fail_page_alloc.attr, 1 << order);
1160 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1162 static int __init fail_page_alloc_debugfs(void)
1164 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1168 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1172 dir = fail_page_alloc.attr.dentries.dir;
1174 fail_page_alloc.ignore_gfp_wait_file =
1175 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1176 &fail_page_alloc.ignore_gfp_wait);
1178 fail_page_alloc.ignore_gfp_highmem_file =
1179 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1180 &fail_page_alloc.ignore_gfp_highmem);
1181 fail_page_alloc.min_order_file =
1182 debugfs_create_u32("min-order", mode, dir,
1183 &fail_page_alloc.min_order);
1185 if (!fail_page_alloc.ignore_gfp_wait_file ||
1186 !fail_page_alloc.ignore_gfp_highmem_file ||
1187 !fail_page_alloc.min_order_file) {
1189 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1190 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1191 debugfs_remove(fail_page_alloc.min_order_file);
1192 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1198 late_initcall(fail_page_alloc_debugfs);
1200 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1202 #else /* CONFIG_FAIL_PAGE_ALLOC */
1204 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1209 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1212 * Return 1 if free pages are above 'mark'. This takes into account the order
1213 * of the allocation.
1215 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1216 int classzone_idx, int alloc_flags)
1218 /* free_pages my go negative - that's OK */
1220 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
1223 if (alloc_flags & ALLOC_HIGH)
1225 if (alloc_flags & ALLOC_HARDER)
1228 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1230 for (o = 0; o < order; o++) {
1231 /* At the next order, this order's pages become unavailable */
1232 free_pages -= z->free_area[o].nr_free << o;
1234 /* Require fewer higher order pages to be free */
1237 if (free_pages <= min)
1245 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1246 * skip over zones that are not allowed by the cpuset, or that have
1247 * been recently (in last second) found to be nearly full. See further
1248 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1249 * that have to skip over a lot of full or unallowed zones.
1251 * If the zonelist cache is present in the passed in zonelist, then
1252 * returns a pointer to the allowed node mask (either the current
1253 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1255 * If the zonelist cache is not available for this zonelist, does
1256 * nothing and returns NULL.
1258 * If the fullzones BITMAP in the zonelist cache is stale (more than
1259 * a second since last zap'd) then we zap it out (clear its bits.)
1261 * We hold off even calling zlc_setup, until after we've checked the
1262 * first zone in the zonelist, on the theory that most allocations will
1263 * be satisfied from that first zone, so best to examine that zone as
1264 * quickly as we can.
1266 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1268 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1269 nodemask_t *allowednodes; /* zonelist_cache approximation */
1271 zlc = zonelist->zlcache_ptr;
1275 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1276 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1277 zlc->last_full_zap = jiffies;
1280 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1281 &cpuset_current_mems_allowed :
1282 &node_states[N_HIGH_MEMORY];
1283 return allowednodes;
1287 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1288 * if it is worth looking at further for free memory:
1289 * 1) Check that the zone isn't thought to be full (doesn't have its
1290 * bit set in the zonelist_cache fullzones BITMAP).
1291 * 2) Check that the zones node (obtained from the zonelist_cache
1292 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1293 * Return true (non-zero) if zone is worth looking at further, or
1294 * else return false (zero) if it is not.
1296 * This check -ignores- the distinction between various watermarks,
1297 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1298 * found to be full for any variation of these watermarks, it will
1299 * be considered full for up to one second by all requests, unless
1300 * we are so low on memory on all allowed nodes that we are forced
1301 * into the second scan of the zonelist.
1303 * In the second scan we ignore this zonelist cache and exactly
1304 * apply the watermarks to all zones, even it is slower to do so.
1305 * We are low on memory in the second scan, and should leave no stone
1306 * unturned looking for a free page.
1308 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1309 nodemask_t *allowednodes)
1311 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1312 int i; /* index of *z in zonelist zones */
1313 int n; /* node that zone *z is on */
1315 zlc = zonelist->zlcache_ptr;
1319 i = z - zonelist->_zonerefs;
1322 /* This zone is worth trying if it is allowed but not full */
1323 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1327 * Given 'z' scanning a zonelist, set the corresponding bit in
1328 * zlc->fullzones, so that subsequent attempts to allocate a page
1329 * from that zone don't waste time re-examining it.
1331 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1333 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1334 int i; /* index of *z in zonelist zones */
1336 zlc = zonelist->zlcache_ptr;
1340 i = z - zonelist->_zonerefs;
1342 set_bit(i, zlc->fullzones);
1345 #else /* CONFIG_NUMA */
1347 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1352 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1353 nodemask_t *allowednodes)
1358 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1361 #endif /* CONFIG_NUMA */
1364 * get_page_from_freelist goes through the zonelist trying to allocate
1367 static struct page *
1368 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1369 struct zonelist *zonelist, int high_zoneidx, int alloc_flags)
1372 struct page *page = NULL;
1374 struct zone *zone, *preferred_zone;
1375 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1376 int zlc_active = 0; /* set if using zonelist_cache */
1377 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1379 (void)first_zones_zonelist(zonelist, high_zoneidx, nodemask,
1381 if (!preferred_zone)
1384 classzone_idx = zone_idx(preferred_zone);
1388 * Scan zonelist, looking for a zone with enough free.
1389 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1391 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1392 high_zoneidx, nodemask) {
1393 if (NUMA_BUILD && zlc_active &&
1394 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1396 if ((alloc_flags & ALLOC_CPUSET) &&
1397 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1400 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1402 if (alloc_flags & ALLOC_WMARK_MIN)
1403 mark = zone->pages_min;
1404 else if (alloc_flags & ALLOC_WMARK_LOW)
1405 mark = zone->pages_low;
1407 mark = zone->pages_high;
1408 if (!zone_watermark_ok(zone, order, mark,
1409 classzone_idx, alloc_flags)) {
1410 if (!zone_reclaim_mode ||
1411 !zone_reclaim(zone, gfp_mask, order))
1412 goto this_zone_full;
1416 page = buffered_rmqueue(preferred_zone, zone, order, gfp_mask);
1421 zlc_mark_zone_full(zonelist, z);
1423 if (NUMA_BUILD && !did_zlc_setup) {
1424 /* we do zlc_setup after the first zone is tried */
1425 allowednodes = zlc_setup(zonelist, alloc_flags);
1431 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1432 /* Disable zlc cache for second zonelist scan */
1440 * This is the 'heart' of the zoned buddy allocator.
1443 __alloc_pages_internal(gfp_t gfp_mask, unsigned int order,
1444 struct zonelist *zonelist, nodemask_t *nodemask)
1446 const gfp_t wait = gfp_mask & __GFP_WAIT;
1447 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1451 struct reclaim_state reclaim_state;
1452 struct task_struct *p = current;
1455 unsigned long did_some_progress;
1456 unsigned long pages_reclaimed = 0;
1458 might_sleep_if(wait);
1460 if (should_fail_alloc_page(gfp_mask, order))
1464 z = zonelist->_zonerefs; /* the list of zones suitable for gfp_mask */
1466 if (unlikely(!z->zone)) {
1468 * Happens if we have an empty zonelist as a result of
1469 * GFP_THISNODE being used on a memoryless node
1474 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
1475 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET);
1480 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1481 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1482 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1483 * using a larger set of nodes after it has established that the
1484 * allowed per node queues are empty and that nodes are
1487 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1490 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1491 wakeup_kswapd(zone, order);
1494 * OK, we're below the kswapd watermark and have kicked background
1495 * reclaim. Now things get more complex, so set up alloc_flags according
1496 * to how we want to proceed.
1498 * The caller may dip into page reserves a bit more if the caller
1499 * cannot run direct reclaim, or if the caller has realtime scheduling
1500 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1501 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1503 alloc_flags = ALLOC_WMARK_MIN;
1504 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
1505 alloc_flags |= ALLOC_HARDER;
1506 if (gfp_mask & __GFP_HIGH)
1507 alloc_flags |= ALLOC_HIGH;
1509 alloc_flags |= ALLOC_CPUSET;
1512 * Go through the zonelist again. Let __GFP_HIGH and allocations
1513 * coming from realtime tasks go deeper into reserves.
1515 * This is the last chance, in general, before the goto nopage.
1516 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1517 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1519 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
1520 high_zoneidx, alloc_flags);
1524 /* This allocation should allow future memory freeing. */
1527 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
1528 && !in_interrupt()) {
1529 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
1531 /* go through the zonelist yet again, ignoring mins */
1532 page = get_page_from_freelist(gfp_mask, nodemask, order,
1533 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS);
1536 if (gfp_mask & __GFP_NOFAIL) {
1537 congestion_wait(WRITE, HZ/50);
1544 /* Atomic allocations - we can't balance anything */
1550 /* We now go into synchronous reclaim */
1551 cpuset_memory_pressure_bump();
1552 p->flags |= PF_MEMALLOC;
1553 reclaim_state.reclaimed_slab = 0;
1554 p->reclaim_state = &reclaim_state;
1556 did_some_progress = try_to_free_pages(zonelist, order, gfp_mask);
1558 p->reclaim_state = NULL;
1559 p->flags &= ~PF_MEMALLOC;
1566 if (likely(did_some_progress)) {
1567 page = get_page_from_freelist(gfp_mask, nodemask, order,
1568 zonelist, high_zoneidx, alloc_flags);
1571 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1572 if (!try_set_zone_oom(zonelist, gfp_mask)) {
1573 schedule_timeout_uninterruptible(1);
1578 * Go through the zonelist yet one more time, keep
1579 * very high watermark here, this is only to catch
1580 * a parallel oom killing, we must fail if we're still
1581 * under heavy pressure.
1583 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1584 order, zonelist, high_zoneidx,
1585 ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1587 clear_zonelist_oom(zonelist, gfp_mask);
1591 /* The OOM killer will not help higher order allocs so fail */
1592 if (order > PAGE_ALLOC_COSTLY_ORDER) {
1593 clear_zonelist_oom(zonelist, gfp_mask);
1597 out_of_memory(zonelist, gfp_mask, order);
1598 clear_zonelist_oom(zonelist, gfp_mask);
1603 * Don't let big-order allocations loop unless the caller explicitly
1604 * requests that. Wait for some write requests to complete then retry.
1606 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1607 * means __GFP_NOFAIL, but that may not be true in other
1610 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1611 * specified, then we retry until we no longer reclaim any pages
1612 * (above), or we've reclaimed an order of pages at least as
1613 * large as the allocation's order. In both cases, if the
1614 * allocation still fails, we stop retrying.
1616 pages_reclaimed += did_some_progress;
1618 if (!(gfp_mask & __GFP_NORETRY)) {
1619 if (order <= PAGE_ALLOC_COSTLY_ORDER) {
1622 if (gfp_mask & __GFP_REPEAT &&
1623 pages_reclaimed < (1 << order))
1626 if (gfp_mask & __GFP_NOFAIL)
1630 congestion_wait(WRITE, HZ/50);
1635 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1636 printk(KERN_WARNING "%s: page allocation failure."
1637 " order:%d, mode:0x%x\n",
1638 p->comm, order, gfp_mask);
1645 EXPORT_SYMBOL(__alloc_pages_internal);
1648 * Common helper functions.
1650 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1653 page = alloc_pages(gfp_mask, order);
1656 return (unsigned long) page_address(page);
1659 EXPORT_SYMBOL(__get_free_pages);
1661 unsigned long get_zeroed_page(gfp_t gfp_mask)
1666 * get_zeroed_page() returns a 32-bit address, which cannot represent
1669 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1671 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1673 return (unsigned long) page_address(page);
1677 EXPORT_SYMBOL(get_zeroed_page);
1679 void __pagevec_free(struct pagevec *pvec)
1681 int i = pagevec_count(pvec);
1684 free_hot_cold_page(pvec->pages[i], pvec->cold);
1687 void __free_pages(struct page *page, unsigned int order)
1689 if (put_page_testzero(page)) {
1691 free_hot_page(page);
1693 __free_pages_ok(page, order);
1697 EXPORT_SYMBOL(__free_pages);
1699 void free_pages(unsigned long addr, unsigned int order)
1702 VM_BUG_ON(!virt_addr_valid((void *)addr));
1703 __free_pages(virt_to_page((void *)addr), order);
1707 EXPORT_SYMBOL(free_pages);
1710 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
1711 * @size: the number of bytes to allocate
1712 * @gfp_mask: GFP flags for the allocation
1714 * This function is similar to alloc_pages(), except that it allocates the
1715 * minimum number of pages to satisfy the request. alloc_pages() can only
1716 * allocate memory in power-of-two pages.
1718 * This function is also limited by MAX_ORDER.
1720 * Memory allocated by this function must be released by free_pages_exact().
1722 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
1724 unsigned int order = get_order(size);
1727 addr = __get_free_pages(gfp_mask, order);
1729 unsigned long alloc_end = addr + (PAGE_SIZE << order);
1730 unsigned long used = addr + PAGE_ALIGN(size);
1732 split_page(virt_to_page(addr), order);
1733 while (used < alloc_end) {
1739 return (void *)addr;
1741 EXPORT_SYMBOL(alloc_pages_exact);
1744 * free_pages_exact - release memory allocated via alloc_pages_exact()
1745 * @virt: the value returned by alloc_pages_exact.
1746 * @size: size of allocation, same value as passed to alloc_pages_exact().
1748 * Release the memory allocated by a previous call to alloc_pages_exact.
1750 void free_pages_exact(void *virt, size_t size)
1752 unsigned long addr = (unsigned long)virt;
1753 unsigned long end = addr + PAGE_ALIGN(size);
1755 while (addr < end) {
1760 EXPORT_SYMBOL(free_pages_exact);
1762 static unsigned int nr_free_zone_pages(int offset)
1767 /* Just pick one node, since fallback list is circular */
1768 unsigned int sum = 0;
1770 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
1772 for_each_zone_zonelist(zone, z, zonelist, offset) {
1773 unsigned long size = zone->present_pages;
1774 unsigned long high = zone->pages_high;
1783 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1785 unsigned int nr_free_buffer_pages(void)
1787 return nr_free_zone_pages(gfp_zone(GFP_USER));
1789 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
1792 * Amount of free RAM allocatable within all zones
1794 unsigned int nr_free_pagecache_pages(void)
1796 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
1799 static inline void show_node(struct zone *zone)
1802 printk("Node %d ", zone_to_nid(zone));
1805 void si_meminfo(struct sysinfo *val)
1807 val->totalram = totalram_pages;
1809 val->freeram = global_page_state(NR_FREE_PAGES);
1810 val->bufferram = nr_blockdev_pages();
1811 val->totalhigh = totalhigh_pages;
1812 val->freehigh = nr_free_highpages();
1813 val->mem_unit = PAGE_SIZE;
1816 EXPORT_SYMBOL(si_meminfo);
1819 void si_meminfo_node(struct sysinfo *val, int nid)
1821 pg_data_t *pgdat = NODE_DATA(nid);
1823 val->totalram = pgdat->node_present_pages;
1824 val->freeram = node_page_state(nid, NR_FREE_PAGES);
1825 #ifdef CONFIG_HIGHMEM
1826 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1827 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
1833 val->mem_unit = PAGE_SIZE;
1837 #define K(x) ((x) << (PAGE_SHIFT-10))
1840 * Show free area list (used inside shift_scroll-lock stuff)
1841 * We also calculate the percentage fragmentation. We do this by counting the
1842 * memory on each free list with the exception of the first item on the list.
1844 void show_free_areas(void)
1849 for_each_zone(zone) {
1850 if (!populated_zone(zone))
1854 printk("%s per-cpu:\n", zone->name);
1856 for_each_online_cpu(cpu) {
1857 struct per_cpu_pageset *pageset;
1859 pageset = zone_pcp(zone, cpu);
1861 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
1862 cpu, pageset->pcp.high,
1863 pageset->pcp.batch, pageset->pcp.count);
1867 printk("Active_anon:%lu active_file:%lu inactive_anon:%lu\n"
1868 " inactive_file:%lu"
1869 //TODO: check/adjust line lengths
1870 #ifdef CONFIG_UNEVICTABLE_LRU
1873 " dirty:%lu writeback:%lu unstable:%lu\n"
1874 " free:%lu slab:%lu mapped:%lu pagetables:%lu bounce:%lu\n",
1875 global_page_state(NR_ACTIVE_ANON),
1876 global_page_state(NR_ACTIVE_FILE),
1877 global_page_state(NR_INACTIVE_ANON),
1878 global_page_state(NR_INACTIVE_FILE),
1879 #ifdef CONFIG_UNEVICTABLE_LRU
1880 global_page_state(NR_UNEVICTABLE),
1882 global_page_state(NR_FILE_DIRTY),
1883 global_page_state(NR_WRITEBACK),
1884 global_page_state(NR_UNSTABLE_NFS),
1885 global_page_state(NR_FREE_PAGES),
1886 global_page_state(NR_SLAB_RECLAIMABLE) +
1887 global_page_state(NR_SLAB_UNRECLAIMABLE),
1888 global_page_state(NR_FILE_MAPPED),
1889 global_page_state(NR_PAGETABLE),
1890 global_page_state(NR_BOUNCE));
1892 for_each_zone(zone) {
1895 if (!populated_zone(zone))
1904 " active_anon:%lukB"
1905 " inactive_anon:%lukB"
1906 " active_file:%lukB"
1907 " inactive_file:%lukB"
1908 #ifdef CONFIG_UNEVICTABLE_LRU
1909 " unevictable:%lukB"
1912 " pages_scanned:%lu"
1913 " all_unreclaimable? %s"
1916 K(zone_page_state(zone, NR_FREE_PAGES)),
1919 K(zone->pages_high),
1920 K(zone_page_state(zone, NR_ACTIVE_ANON)),
1921 K(zone_page_state(zone, NR_INACTIVE_ANON)),
1922 K(zone_page_state(zone, NR_ACTIVE_FILE)),
1923 K(zone_page_state(zone, NR_INACTIVE_FILE)),
1924 #ifdef CONFIG_UNEVICTABLE_LRU
1925 K(zone_page_state(zone, NR_UNEVICTABLE)),
1927 K(zone->present_pages),
1928 zone->pages_scanned,
1929 (zone_is_all_unreclaimable(zone) ? "yes" : "no")
1931 printk("lowmem_reserve[]:");
1932 for (i = 0; i < MAX_NR_ZONES; i++)
1933 printk(" %lu", zone->lowmem_reserve[i]);
1937 for_each_zone(zone) {
1938 unsigned long nr[MAX_ORDER], flags, order, total = 0;
1940 if (!populated_zone(zone))
1944 printk("%s: ", zone->name);
1946 spin_lock_irqsave(&zone->lock, flags);
1947 for (order = 0; order < MAX_ORDER; order++) {
1948 nr[order] = zone->free_area[order].nr_free;
1949 total += nr[order] << order;
1951 spin_unlock_irqrestore(&zone->lock, flags);
1952 for (order = 0; order < MAX_ORDER; order++)
1953 printk("%lu*%lukB ", nr[order], K(1UL) << order);
1954 printk("= %lukB\n", K(total));
1957 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
1959 show_swap_cache_info();
1962 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
1964 zoneref->zone = zone;
1965 zoneref->zone_idx = zone_idx(zone);
1969 * Builds allocation fallback zone lists.
1971 * Add all populated zones of a node to the zonelist.
1973 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
1974 int nr_zones, enum zone_type zone_type)
1978 BUG_ON(zone_type >= MAX_NR_ZONES);
1983 zone = pgdat->node_zones + zone_type;
1984 if (populated_zone(zone)) {
1985 zoneref_set_zone(zone,
1986 &zonelist->_zonerefs[nr_zones++]);
1987 check_highest_zone(zone_type);
1990 } while (zone_type);
1997 * 0 = automatic detection of better ordering.
1998 * 1 = order by ([node] distance, -zonetype)
1999 * 2 = order by (-zonetype, [node] distance)
2001 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2002 * the same zonelist. So only NUMA can configure this param.
2004 #define ZONELIST_ORDER_DEFAULT 0
2005 #define ZONELIST_ORDER_NODE 1
2006 #define ZONELIST_ORDER_ZONE 2
2008 /* zonelist order in the kernel.
2009 * set_zonelist_order() will set this to NODE or ZONE.
2011 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2012 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2016 /* The value user specified ....changed by config */
2017 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2018 /* string for sysctl */
2019 #define NUMA_ZONELIST_ORDER_LEN 16
2020 char numa_zonelist_order[16] = "default";
2023 * interface for configure zonelist ordering.
2024 * command line option "numa_zonelist_order"
2025 * = "[dD]efault - default, automatic configuration.
2026 * = "[nN]ode - order by node locality, then by zone within node
2027 * = "[zZ]one - order by zone, then by locality within zone
2030 static int __parse_numa_zonelist_order(char *s)
2032 if (*s == 'd' || *s == 'D') {
2033 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2034 } else if (*s == 'n' || *s == 'N') {
2035 user_zonelist_order = ZONELIST_ORDER_NODE;
2036 } else if (*s == 'z' || *s == 'Z') {
2037 user_zonelist_order = ZONELIST_ORDER_ZONE;
2040 "Ignoring invalid numa_zonelist_order value: "
2047 static __init int setup_numa_zonelist_order(char *s)
2050 return __parse_numa_zonelist_order(s);
2053 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2056 * sysctl handler for numa_zonelist_order
2058 int numa_zonelist_order_handler(ctl_table *table, int write,
2059 struct file *file, void __user *buffer, size_t *length,
2062 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2066 strncpy(saved_string, (char*)table->data,
2067 NUMA_ZONELIST_ORDER_LEN);
2068 ret = proc_dostring(table, write, file, buffer, length, ppos);
2072 int oldval = user_zonelist_order;
2073 if (__parse_numa_zonelist_order((char*)table->data)) {
2075 * bogus value. restore saved string
2077 strncpy((char*)table->data, saved_string,
2078 NUMA_ZONELIST_ORDER_LEN);
2079 user_zonelist_order = oldval;
2080 } else if (oldval != user_zonelist_order)
2081 build_all_zonelists();
2087 #define MAX_NODE_LOAD (num_online_nodes())
2088 static int node_load[MAX_NUMNODES];
2091 * find_next_best_node - find the next node that should appear in a given node's fallback list
2092 * @node: node whose fallback list we're appending
2093 * @used_node_mask: nodemask_t of already used nodes
2095 * We use a number of factors to determine which is the next node that should
2096 * appear on a given node's fallback list. The node should not have appeared
2097 * already in @node's fallback list, and it should be the next closest node
2098 * according to the distance array (which contains arbitrary distance values
2099 * from each node to each node in the system), and should also prefer nodes
2100 * with no CPUs, since presumably they'll have very little allocation pressure
2101 * on them otherwise.
2102 * It returns -1 if no node is found.
2104 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2107 int min_val = INT_MAX;
2109 node_to_cpumask_ptr(tmp, 0);
2111 /* Use the local node if we haven't already */
2112 if (!node_isset(node, *used_node_mask)) {
2113 node_set(node, *used_node_mask);
2117 for_each_node_state(n, N_HIGH_MEMORY) {
2119 /* Don't want a node to appear more than once */
2120 if (node_isset(n, *used_node_mask))
2123 /* Use the distance array to find the distance */
2124 val = node_distance(node, n);
2126 /* Penalize nodes under us ("prefer the next node") */
2129 /* Give preference to headless and unused nodes */
2130 node_to_cpumask_ptr_next(tmp, n);
2131 if (!cpus_empty(*tmp))
2132 val += PENALTY_FOR_NODE_WITH_CPUS;
2134 /* Slight preference for less loaded node */
2135 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2136 val += node_load[n];
2138 if (val < min_val) {
2145 node_set(best_node, *used_node_mask);
2152 * Build zonelists ordered by node and zones within node.
2153 * This results in maximum locality--normal zone overflows into local
2154 * DMA zone, if any--but risks exhausting DMA zone.
2156 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2159 struct zonelist *zonelist;
2161 zonelist = &pgdat->node_zonelists[0];
2162 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2164 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2166 zonelist->_zonerefs[j].zone = NULL;
2167 zonelist->_zonerefs[j].zone_idx = 0;
2171 * Build gfp_thisnode zonelists
2173 static void build_thisnode_zonelists(pg_data_t *pgdat)
2176 struct zonelist *zonelist;
2178 zonelist = &pgdat->node_zonelists[1];
2179 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2180 zonelist->_zonerefs[j].zone = NULL;
2181 zonelist->_zonerefs[j].zone_idx = 0;
2185 * Build zonelists ordered by zone and nodes within zones.
2186 * This results in conserving DMA zone[s] until all Normal memory is
2187 * exhausted, but results in overflowing to remote node while memory
2188 * may still exist in local DMA zone.
2190 static int node_order[MAX_NUMNODES];
2192 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2195 int zone_type; /* needs to be signed */
2197 struct zonelist *zonelist;
2199 zonelist = &pgdat->node_zonelists[0];
2201 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2202 for (j = 0; j < nr_nodes; j++) {
2203 node = node_order[j];
2204 z = &NODE_DATA(node)->node_zones[zone_type];
2205 if (populated_zone(z)) {
2207 &zonelist->_zonerefs[pos++]);
2208 check_highest_zone(zone_type);
2212 zonelist->_zonerefs[pos].zone = NULL;
2213 zonelist->_zonerefs[pos].zone_idx = 0;
2216 static int default_zonelist_order(void)
2219 unsigned long low_kmem_size,total_size;
2223 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2224 * If they are really small and used heavily, the system can fall
2225 * into OOM very easily.
2226 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2228 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2231 for_each_online_node(nid) {
2232 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2233 z = &NODE_DATA(nid)->node_zones[zone_type];
2234 if (populated_zone(z)) {
2235 if (zone_type < ZONE_NORMAL)
2236 low_kmem_size += z->present_pages;
2237 total_size += z->present_pages;
2241 if (!low_kmem_size || /* there are no DMA area. */
2242 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2243 return ZONELIST_ORDER_NODE;
2245 * look into each node's config.
2246 * If there is a node whose DMA/DMA32 memory is very big area on
2247 * local memory, NODE_ORDER may be suitable.
2249 average_size = total_size /
2250 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2251 for_each_online_node(nid) {
2254 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2255 z = &NODE_DATA(nid)->node_zones[zone_type];
2256 if (populated_zone(z)) {
2257 if (zone_type < ZONE_NORMAL)
2258 low_kmem_size += z->present_pages;
2259 total_size += z->present_pages;
2262 if (low_kmem_size &&
2263 total_size > average_size && /* ignore small node */
2264 low_kmem_size > total_size * 70/100)
2265 return ZONELIST_ORDER_NODE;
2267 return ZONELIST_ORDER_ZONE;
2270 static void set_zonelist_order(void)
2272 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2273 current_zonelist_order = default_zonelist_order();
2275 current_zonelist_order = user_zonelist_order;
2278 static void build_zonelists(pg_data_t *pgdat)
2282 nodemask_t used_mask;
2283 int local_node, prev_node;
2284 struct zonelist *zonelist;
2285 int order = current_zonelist_order;
2287 /* initialize zonelists */
2288 for (i = 0; i < MAX_ZONELISTS; i++) {
2289 zonelist = pgdat->node_zonelists + i;
2290 zonelist->_zonerefs[0].zone = NULL;
2291 zonelist->_zonerefs[0].zone_idx = 0;
2294 /* NUMA-aware ordering of nodes */
2295 local_node = pgdat->node_id;
2296 load = num_online_nodes();
2297 prev_node = local_node;
2298 nodes_clear(used_mask);
2300 memset(node_load, 0, sizeof(node_load));
2301 memset(node_order, 0, sizeof(node_order));
2304 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2305 int distance = node_distance(local_node, node);
2308 * If another node is sufficiently far away then it is better
2309 * to reclaim pages in a zone before going off node.
2311 if (distance > RECLAIM_DISTANCE)
2312 zone_reclaim_mode = 1;
2315 * We don't want to pressure a particular node.
2316 * So adding penalty to the first node in same
2317 * distance group to make it round-robin.
2319 if (distance != node_distance(local_node, prev_node))
2320 node_load[node] = load;
2324 if (order == ZONELIST_ORDER_NODE)
2325 build_zonelists_in_node_order(pgdat, node);
2327 node_order[j++] = node; /* remember order */
2330 if (order == ZONELIST_ORDER_ZONE) {
2331 /* calculate node order -- i.e., DMA last! */
2332 build_zonelists_in_zone_order(pgdat, j);
2335 build_thisnode_zonelists(pgdat);
2338 /* Construct the zonelist performance cache - see further mmzone.h */
2339 static void build_zonelist_cache(pg_data_t *pgdat)
2341 struct zonelist *zonelist;
2342 struct zonelist_cache *zlc;
2345 zonelist = &pgdat->node_zonelists[0];
2346 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2347 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2348 for (z = zonelist->_zonerefs; z->zone; z++)
2349 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2353 #else /* CONFIG_NUMA */
2355 static void set_zonelist_order(void)
2357 current_zonelist_order = ZONELIST_ORDER_ZONE;
2360 static void build_zonelists(pg_data_t *pgdat)
2362 int node, local_node;
2364 struct zonelist *zonelist;
2366 local_node = pgdat->node_id;
2368 zonelist = &pgdat->node_zonelists[0];
2369 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2372 * Now we build the zonelist so that it contains the zones
2373 * of all the other nodes.
2374 * We don't want to pressure a particular node, so when
2375 * building the zones for node N, we make sure that the
2376 * zones coming right after the local ones are those from
2377 * node N+1 (modulo N)
2379 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2380 if (!node_online(node))
2382 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2385 for (node = 0; node < local_node; node++) {
2386 if (!node_online(node))
2388 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2392 zonelist->_zonerefs[j].zone = NULL;
2393 zonelist->_zonerefs[j].zone_idx = 0;
2396 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2397 static void build_zonelist_cache(pg_data_t *pgdat)
2399 pgdat->node_zonelists[0].zlcache_ptr = NULL;
2402 #endif /* CONFIG_NUMA */
2404 /* return values int ....just for stop_machine() */
2405 static int __build_all_zonelists(void *dummy)
2409 for_each_online_node(nid) {
2410 pg_data_t *pgdat = NODE_DATA(nid);
2412 build_zonelists(pgdat);
2413 build_zonelist_cache(pgdat);
2418 void build_all_zonelists(void)
2420 set_zonelist_order();
2422 if (system_state == SYSTEM_BOOTING) {
2423 __build_all_zonelists(NULL);
2424 mminit_verify_zonelist();
2425 cpuset_init_current_mems_allowed();
2427 /* we have to stop all cpus to guarantee there is no user
2429 stop_machine(__build_all_zonelists, NULL, NULL);
2430 /* cpuset refresh routine should be here */
2432 vm_total_pages = nr_free_pagecache_pages();
2434 * Disable grouping by mobility if the number of pages in the
2435 * system is too low to allow the mechanism to work. It would be
2436 * more accurate, but expensive to check per-zone. This check is
2437 * made on memory-hotadd so a system can start with mobility
2438 * disabled and enable it later
2440 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
2441 page_group_by_mobility_disabled = 1;
2443 page_group_by_mobility_disabled = 0;
2445 printk("Built %i zonelists in %s order, mobility grouping %s. "
2446 "Total pages: %ld\n",
2448 zonelist_order_name[current_zonelist_order],
2449 page_group_by_mobility_disabled ? "off" : "on",
2452 printk("Policy zone: %s\n", zone_names[policy_zone]);
2457 * Helper functions to size the waitqueue hash table.
2458 * Essentially these want to choose hash table sizes sufficiently
2459 * large so that collisions trying to wait on pages are rare.
2460 * But in fact, the number of active page waitqueues on typical
2461 * systems is ridiculously low, less than 200. So this is even
2462 * conservative, even though it seems large.
2464 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2465 * waitqueues, i.e. the size of the waitq table given the number of pages.
2467 #define PAGES_PER_WAITQUEUE 256
2469 #ifndef CONFIG_MEMORY_HOTPLUG
2470 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2472 unsigned long size = 1;
2474 pages /= PAGES_PER_WAITQUEUE;
2476 while (size < pages)
2480 * Once we have dozens or even hundreds of threads sleeping
2481 * on IO we've got bigger problems than wait queue collision.
2482 * Limit the size of the wait table to a reasonable size.
2484 size = min(size, 4096UL);
2486 return max(size, 4UL);
2490 * A zone's size might be changed by hot-add, so it is not possible to determine
2491 * a suitable size for its wait_table. So we use the maximum size now.
2493 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2495 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2496 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2497 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2499 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2500 * or more by the traditional way. (See above). It equals:
2502 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2503 * ia64(16K page size) : = ( 8G + 4M)byte.
2504 * powerpc (64K page size) : = (32G +16M)byte.
2506 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2513 * This is an integer logarithm so that shifts can be used later
2514 * to extract the more random high bits from the multiplicative
2515 * hash function before the remainder is taken.
2517 static inline unsigned long wait_table_bits(unsigned long size)
2522 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2525 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2526 * of blocks reserved is based on zone->pages_min. The memory within the
2527 * reserve will tend to store contiguous free pages. Setting min_free_kbytes
2528 * higher will lead to a bigger reserve which will get freed as contiguous
2529 * blocks as reclaim kicks in
2531 static void setup_zone_migrate_reserve(struct zone *zone)
2533 unsigned long start_pfn, pfn, end_pfn;
2535 unsigned long reserve, block_migratetype;
2537 /* Get the start pfn, end pfn and the number of blocks to reserve */
2538 start_pfn = zone->zone_start_pfn;
2539 end_pfn = start_pfn + zone->spanned_pages;
2540 reserve = roundup(zone->pages_min, pageblock_nr_pages) >>
2543 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
2544 if (!pfn_valid(pfn))
2546 page = pfn_to_page(pfn);
2548 /* Watch out for overlapping nodes */
2549 if (page_to_nid(page) != zone_to_nid(zone))
2552 /* Blocks with reserved pages will never free, skip them. */
2553 if (PageReserved(page))
2556 block_migratetype = get_pageblock_migratetype(page);
2558 /* If this block is reserved, account for it */
2559 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
2564 /* Suitable for reserving if this block is movable */
2565 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
2566 set_pageblock_migratetype(page, MIGRATE_RESERVE);
2567 move_freepages_block(zone, page, MIGRATE_RESERVE);
2573 * If the reserve is met and this is a previous reserved block,
2576 if (block_migratetype == MIGRATE_RESERVE) {
2577 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2578 move_freepages_block(zone, page, MIGRATE_MOVABLE);
2584 * Initially all pages are reserved - free ones are freed
2585 * up by free_all_bootmem() once the early boot process is
2586 * done. Non-atomic initialization, single-pass.
2588 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
2589 unsigned long start_pfn, enum memmap_context context)
2592 unsigned long end_pfn = start_pfn + size;
2596 z = &NODE_DATA(nid)->node_zones[zone];
2597 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
2599 * There can be holes in boot-time mem_map[]s
2600 * handed to this function. They do not
2601 * exist on hotplugged memory.
2603 if (context == MEMMAP_EARLY) {
2604 if (!early_pfn_valid(pfn))
2606 if (!early_pfn_in_nid(pfn, nid))
2609 page = pfn_to_page(pfn);
2610 set_page_links(page, zone, nid, pfn);
2611 mminit_verify_page_links(page, zone, nid, pfn);
2612 init_page_count(page);
2613 reset_page_mapcount(page);
2614 SetPageReserved(page);
2616 * Mark the block movable so that blocks are reserved for
2617 * movable at startup. This will force kernel allocations
2618 * to reserve their blocks rather than leaking throughout
2619 * the address space during boot when many long-lived
2620 * kernel allocations are made. Later some blocks near
2621 * the start are marked MIGRATE_RESERVE by
2622 * setup_zone_migrate_reserve()
2624 * bitmap is created for zone's valid pfn range. but memmap
2625 * can be created for invalid pages (for alignment)
2626 * check here not to call set_pageblock_migratetype() against
2629 if ((z->zone_start_pfn <= pfn)
2630 && (pfn < z->zone_start_pfn + z->spanned_pages)
2631 && !(pfn & (pageblock_nr_pages - 1)))
2632 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2634 INIT_LIST_HEAD(&page->lru);
2635 #ifdef WANT_PAGE_VIRTUAL
2636 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2637 if (!is_highmem_idx(zone))
2638 set_page_address(page, __va(pfn << PAGE_SHIFT));
2643 static void __meminit zone_init_free_lists(struct zone *zone)
2646 for_each_migratetype_order(order, t) {
2647 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
2648 zone->free_area[order].nr_free = 0;
2652 #ifndef __HAVE_ARCH_MEMMAP_INIT
2653 #define memmap_init(size, nid, zone, start_pfn) \
2654 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
2657 static int zone_batchsize(struct zone *zone)
2662 * The per-cpu-pages pools are set to around 1000th of the
2663 * size of the zone. But no more than 1/2 of a meg.
2665 * OK, so we don't know how big the cache is. So guess.
2667 batch = zone->present_pages / 1024;
2668 if (batch * PAGE_SIZE > 512 * 1024)
2669 batch = (512 * 1024) / PAGE_SIZE;
2670 batch /= 4; /* We effectively *= 4 below */
2675 * Clamp the batch to a 2^n - 1 value. Having a power
2676 * of 2 value was found to be more likely to have
2677 * suboptimal cache aliasing properties in some cases.
2679 * For example if 2 tasks are alternately allocating
2680 * batches of pages, one task can end up with a lot
2681 * of pages of one half of the possible page colors
2682 * and the other with pages of the other colors.
2684 batch = (1 << (fls(batch + batch/2)-1)) - 1;
2689 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
2691 struct per_cpu_pages *pcp;
2693 memset(p, 0, sizeof(*p));
2697 pcp->high = 6 * batch;
2698 pcp->batch = max(1UL, 1 * batch);
2699 INIT_LIST_HEAD(&pcp->list);
2703 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
2704 * to the value high for the pageset p.
2707 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
2710 struct per_cpu_pages *pcp;
2714 pcp->batch = max(1UL, high/4);
2715 if ((high/4) > (PAGE_SHIFT * 8))
2716 pcp->batch = PAGE_SHIFT * 8;
2722 * Boot pageset table. One per cpu which is going to be used for all
2723 * zones and all nodes. The parameters will be set in such a way
2724 * that an item put on a list will immediately be handed over to
2725 * the buddy list. This is safe since pageset manipulation is done
2726 * with interrupts disabled.
2728 * Some NUMA counter updates may also be caught by the boot pagesets.
2730 * The boot_pagesets must be kept even after bootup is complete for
2731 * unused processors and/or zones. They do play a role for bootstrapping
2732 * hotplugged processors.
2734 * zoneinfo_show() and maybe other functions do
2735 * not check if the processor is online before following the pageset pointer.
2736 * Other parts of the kernel may not check if the zone is available.
2738 static struct per_cpu_pageset boot_pageset[NR_CPUS];
2741 * Dynamically allocate memory for the
2742 * per cpu pageset array in struct zone.
2744 static int __cpuinit process_zones(int cpu)
2746 struct zone *zone, *dzone;
2747 int node = cpu_to_node(cpu);
2749 node_set_state(node, N_CPU); /* this node has a cpu */
2751 for_each_zone(zone) {
2753 if (!populated_zone(zone))
2756 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
2758 if (!zone_pcp(zone, cpu))
2761 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
2763 if (percpu_pagelist_fraction)
2764 setup_pagelist_highmark(zone_pcp(zone, cpu),
2765 (zone->present_pages / percpu_pagelist_fraction));
2770 for_each_zone(dzone) {
2771 if (!populated_zone(dzone))
2775 kfree(zone_pcp(dzone, cpu));
2776 zone_pcp(dzone, cpu) = NULL;
2781 static inline void free_zone_pagesets(int cpu)
2785 for_each_zone(zone) {
2786 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
2788 /* Free per_cpu_pageset if it is slab allocated */
2789 if (pset != &boot_pageset[cpu])
2791 zone_pcp(zone, cpu) = NULL;
2795 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
2796 unsigned long action,
2799 int cpu = (long)hcpu;
2800 int ret = NOTIFY_OK;
2803 case CPU_UP_PREPARE:
2804 case CPU_UP_PREPARE_FROZEN:
2805 if (process_zones(cpu))
2808 case CPU_UP_CANCELED:
2809 case CPU_UP_CANCELED_FROZEN:
2811 case CPU_DEAD_FROZEN:
2812 free_zone_pagesets(cpu);
2820 static struct notifier_block __cpuinitdata pageset_notifier =
2821 { &pageset_cpuup_callback, NULL, 0 };
2823 void __init setup_per_cpu_pageset(void)
2827 /* Initialize per_cpu_pageset for cpu 0.
2828 * A cpuup callback will do this for every cpu
2829 * as it comes online
2831 err = process_zones(smp_processor_id());
2833 register_cpu_notifier(&pageset_notifier);
2838 static noinline __init_refok
2839 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
2842 struct pglist_data *pgdat = zone->zone_pgdat;
2846 * The per-page waitqueue mechanism uses hashed waitqueues
2849 zone->wait_table_hash_nr_entries =
2850 wait_table_hash_nr_entries(zone_size_pages);
2851 zone->wait_table_bits =
2852 wait_table_bits(zone->wait_table_hash_nr_entries);
2853 alloc_size = zone->wait_table_hash_nr_entries
2854 * sizeof(wait_queue_head_t);
2856 if (!slab_is_available()) {
2857 zone->wait_table = (wait_queue_head_t *)
2858 alloc_bootmem_node(pgdat, alloc_size);
2861 * This case means that a zone whose size was 0 gets new memory
2862 * via memory hot-add.
2863 * But it may be the case that a new node was hot-added. In
2864 * this case vmalloc() will not be able to use this new node's
2865 * memory - this wait_table must be initialized to use this new
2866 * node itself as well.
2867 * To use this new node's memory, further consideration will be
2870 zone->wait_table = vmalloc(alloc_size);
2872 if (!zone->wait_table)
2875 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
2876 init_waitqueue_head(zone->wait_table + i);
2881 static __meminit void zone_pcp_init(struct zone *zone)
2884 unsigned long batch = zone_batchsize(zone);
2886 for (cpu = 0; cpu < NR_CPUS; cpu++) {
2888 /* Early boot. Slab allocator not functional yet */
2889 zone_pcp(zone, cpu) = &boot_pageset[cpu];
2890 setup_pageset(&boot_pageset[cpu],0);
2892 setup_pageset(zone_pcp(zone,cpu), batch);
2895 if (zone->present_pages)
2896 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
2897 zone->name, zone->present_pages, batch);
2900 __meminit int init_currently_empty_zone(struct zone *zone,
2901 unsigned long zone_start_pfn,
2903 enum memmap_context context)
2905 struct pglist_data *pgdat = zone->zone_pgdat;
2907 ret = zone_wait_table_init(zone, size);
2910 pgdat->nr_zones = zone_idx(zone) + 1;
2912 zone->zone_start_pfn = zone_start_pfn;
2914 mminit_dprintk(MMINIT_TRACE, "memmap_init",
2915 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
2917 (unsigned long)zone_idx(zone),
2918 zone_start_pfn, (zone_start_pfn + size));
2920 zone_init_free_lists(zone);
2925 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
2927 * Basic iterator support. Return the first range of PFNs for a node
2928 * Note: nid == MAX_NUMNODES returns first region regardless of node
2930 static int __meminit first_active_region_index_in_nid(int nid)
2934 for (i = 0; i < nr_nodemap_entries; i++)
2935 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
2942 * Basic iterator support. Return the next active range of PFNs for a node
2943 * Note: nid == MAX_NUMNODES returns next region regardless of node
2945 static int __meminit next_active_region_index_in_nid(int index, int nid)
2947 for (index = index + 1; index < nr_nodemap_entries; index++)
2948 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
2954 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
2956 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
2957 * Architectures may implement their own version but if add_active_range()
2958 * was used and there are no special requirements, this is a convenient
2961 int __meminit early_pfn_to_nid(unsigned long pfn)
2965 for (i = 0; i < nr_nodemap_entries; i++) {
2966 unsigned long start_pfn = early_node_map[i].start_pfn;
2967 unsigned long end_pfn = early_node_map[i].end_pfn;
2969 if (start_pfn <= pfn && pfn < end_pfn)
2970 return early_node_map[i].nid;
2975 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
2977 /* Basic iterator support to walk early_node_map[] */
2978 #define for_each_active_range_index_in_nid(i, nid) \
2979 for (i = first_active_region_index_in_nid(nid); i != -1; \
2980 i = next_active_region_index_in_nid(i, nid))
2983 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
2984 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
2985 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
2987 * If an architecture guarantees that all ranges registered with
2988 * add_active_ranges() contain no holes and may be freed, this
2989 * this function may be used instead of calling free_bootmem() manually.
2991 void __init free_bootmem_with_active_regions(int nid,
2992 unsigned long max_low_pfn)
2996 for_each_active_range_index_in_nid(i, nid) {
2997 unsigned long size_pages = 0;
2998 unsigned long end_pfn = early_node_map[i].end_pfn;
3000 if (early_node_map[i].start_pfn >= max_low_pfn)
3003 if (end_pfn > max_low_pfn)
3004 end_pfn = max_low_pfn;
3006 size_pages = end_pfn - early_node_map[i].start_pfn;
3007 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3008 PFN_PHYS(early_node_map[i].start_pfn),
3009 size_pages << PAGE_SHIFT);
3013 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3018 for_each_active_range_index_in_nid(i, nid) {
3019 ret = work_fn(early_node_map[i].start_pfn,
3020 early_node_map[i].end_pfn, data);
3026 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3027 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3029 * If an architecture guarantees that all ranges registered with
3030 * add_active_ranges() contain no holes and may be freed, this
3031 * function may be used instead of calling memory_present() manually.
3033 void __init sparse_memory_present_with_active_regions(int nid)
3037 for_each_active_range_index_in_nid(i, nid)
3038 memory_present(early_node_map[i].nid,
3039 early_node_map[i].start_pfn,
3040 early_node_map[i].end_pfn);
3044 * push_node_boundaries - Push node boundaries to at least the requested boundary
3045 * @nid: The nid of the node to push the boundary for
3046 * @start_pfn: The start pfn of the node
3047 * @end_pfn: The end pfn of the node
3049 * In reserve-based hot-add, mem_map is allocated that is unused until hotadd
3050 * time. Specifically, on x86_64, SRAT will report ranges that can potentially
3051 * be hotplugged even though no physical memory exists. This function allows
3052 * an arch to push out the node boundaries so mem_map is allocated that can
3055 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
3056 void __init push_node_boundaries(unsigned int nid,
3057 unsigned long start_pfn, unsigned long end_pfn)
3059 mminit_dprintk(MMINIT_TRACE, "zoneboundary",
3060 "Entering push_node_boundaries(%u, %lu, %lu)\n",
3061 nid, start_pfn, end_pfn);
3063 /* Initialise the boundary for this node if necessary */
3064 if (node_boundary_end_pfn[nid] == 0)
3065 node_boundary_start_pfn[nid] = -1UL;
3067 /* Update the boundaries */
3068 if (node_boundary_start_pfn[nid] > start_pfn)
3069 node_boundary_start_pfn[nid] = start_pfn;
3070 if (node_boundary_end_pfn[nid] < end_pfn)
3071 node_boundary_end_pfn[nid] = end_pfn;
3074 /* If necessary, push the node boundary out for reserve hotadd */
3075 static void __meminit account_node_boundary(unsigned int nid,
3076 unsigned long *start_pfn, unsigned long *end_pfn)
3078 mminit_dprintk(MMINIT_TRACE, "zoneboundary",
3079 "Entering account_node_boundary(%u, %lu, %lu)\n",
3080 nid, *start_pfn, *end_pfn);
3082 /* Return if boundary information has not been provided */
3083 if (node_boundary_end_pfn[nid] == 0)
3086 /* Check the boundaries and update if necessary */
3087 if (node_boundary_start_pfn[nid] < *start_pfn)
3088 *start_pfn = node_boundary_start_pfn[nid];
3089 if (node_boundary_end_pfn[nid] > *end_pfn)
3090 *end_pfn = node_boundary_end_pfn[nid];
3093 void __init push_node_boundaries(unsigned int nid,
3094 unsigned long start_pfn, unsigned long end_pfn) {}
3096 static void __meminit account_node_boundary(unsigned int nid,
3097 unsigned long *start_pfn, unsigned long *end_pfn) {}
3102 * get_pfn_range_for_nid - Return the start and end page frames for a node
3103 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3104 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3105 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3107 * It returns the start and end page frame of a node based on information
3108 * provided by an arch calling add_active_range(). If called for a node
3109 * with no available memory, a warning is printed and the start and end
3112 void __meminit get_pfn_range_for_nid(unsigned int nid,
3113 unsigned long *start_pfn, unsigned long *end_pfn)
3119 for_each_active_range_index_in_nid(i, nid) {
3120 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3121 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3124 if (*start_pfn == -1UL)
3127 /* Push the node boundaries out if requested */
3128 account_node_boundary(nid, start_pfn, end_pfn);
3132 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3133 * assumption is made that zones within a node are ordered in monotonic
3134 * increasing memory addresses so that the "highest" populated zone is used
3136 static void __init find_usable_zone_for_movable(void)
3139 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3140 if (zone_index == ZONE_MOVABLE)
3143 if (arch_zone_highest_possible_pfn[zone_index] >
3144 arch_zone_lowest_possible_pfn[zone_index])
3148 VM_BUG_ON(zone_index == -1);
3149 movable_zone = zone_index;
3153 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3154 * because it is sized independant of architecture. Unlike the other zones,
3155 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3156 * in each node depending on the size of each node and how evenly kernelcore
3157 * is distributed. This helper function adjusts the zone ranges
3158 * provided by the architecture for a given node by using the end of the
3159 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3160 * zones within a node are in order of monotonic increases memory addresses
3162 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3163 unsigned long zone_type,
3164 unsigned long node_start_pfn,
3165 unsigned long node_end_pfn,
3166 unsigned long *zone_start_pfn,
3167 unsigned long *zone_end_pfn)
3169 /* Only adjust if ZONE_MOVABLE is on this node */
3170 if (zone_movable_pfn[nid]) {
3171 /* Size ZONE_MOVABLE */
3172 if (zone_type == ZONE_MOVABLE) {
3173 *zone_start_pfn = zone_movable_pfn[nid];
3174 *zone_end_pfn = min(node_end_pfn,
3175 arch_zone_highest_possible_pfn[movable_zone]);
3177 /* Adjust for ZONE_MOVABLE starting within this range */
3178 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3179 *zone_end_pfn > zone_movable_pfn[nid]) {
3180 *zone_end_pfn = zone_movable_pfn[nid];
3182 /* Check if this whole range is within ZONE_MOVABLE */
3183 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3184 *zone_start_pfn = *zone_end_pfn;
3189 * Return the number of pages a zone spans in a node, including holes
3190 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3192 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3193 unsigned long zone_type,
3194 unsigned long *ignored)
3196 unsigned long node_start_pfn, node_end_pfn;
3197 unsigned long zone_start_pfn, zone_end_pfn;
3199 /* Get the start and end of the node and zone */
3200 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3201 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3202 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3203 adjust_zone_range_for_zone_movable(nid, zone_type,
3204 node_start_pfn, node_end_pfn,
3205 &zone_start_pfn, &zone_end_pfn);
3207 /* Check that this node has pages within the zone's required range */
3208 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3211 /* Move the zone boundaries inside the node if necessary */
3212 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3213 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3215 /* Return the spanned pages */
3216 return zone_end_pfn - zone_start_pfn;
3220 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3221 * then all holes in the requested range will be accounted for.
3223 static unsigned long __meminit __absent_pages_in_range(int nid,
3224 unsigned long range_start_pfn,
3225 unsigned long range_end_pfn)
3228 unsigned long prev_end_pfn = 0, hole_pages = 0;
3229 unsigned long start_pfn;
3231 /* Find the end_pfn of the first active range of pfns in the node */
3232 i = first_active_region_index_in_nid(nid);
3236 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3238 /* Account for ranges before physical memory on this node */
3239 if (early_node_map[i].start_pfn > range_start_pfn)
3240 hole_pages = prev_end_pfn - range_start_pfn;
3242 /* Find all holes for the zone within the node */
3243 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3245 /* No need to continue if prev_end_pfn is outside the zone */
3246 if (prev_end_pfn >= range_end_pfn)
3249 /* Make sure the end of the zone is not within the hole */
3250 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3251 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3253 /* Update the hole size cound and move on */
3254 if (start_pfn > range_start_pfn) {
3255 BUG_ON(prev_end_pfn > start_pfn);
3256 hole_pages += start_pfn - prev_end_pfn;
3258 prev_end_pfn = early_node_map[i].end_pfn;
3261 /* Account for ranges past physical memory on this node */
3262 if (range_end_pfn > prev_end_pfn)
3263 hole_pages += range_end_pfn -
3264 max(range_start_pfn, prev_end_pfn);
3270 * absent_pages_in_range - Return number of page frames in holes within a range
3271 * @start_pfn: The start PFN to start searching for holes
3272 * @end_pfn: The end PFN to stop searching for holes
3274 * It returns the number of pages frames in memory holes within a range.
3276 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3277 unsigned long end_pfn)
3279 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3282 /* Return the number of page frames in holes in a zone on a node */
3283 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3284 unsigned long zone_type,
3285 unsigned long *ignored)
3287 unsigned long node_start_pfn, node_end_pfn;
3288 unsigned long zone_start_pfn, zone_end_pfn;
3290 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3291 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3293 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3296 adjust_zone_range_for_zone_movable(nid, zone_type,
3297 node_start_pfn, node_end_pfn,
3298 &zone_start_pfn, &zone_end_pfn);
3299 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3303 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3304 unsigned long zone_type,
3305 unsigned long *zones_size)
3307 return zones_size[zone_type];
3310 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3311 unsigned long zone_type,
3312 unsigned long *zholes_size)
3317 return zholes_size[zone_type];
3322 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3323 unsigned long *zones_size, unsigned long *zholes_size)
3325 unsigned long realtotalpages, totalpages = 0;
3328 for (i = 0; i < MAX_NR_ZONES; i++)
3329 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3331 pgdat->node_spanned_pages = totalpages;
3333 realtotalpages = totalpages;
3334 for (i = 0; i < MAX_NR_ZONES; i++)
3336 zone_absent_pages_in_node(pgdat->node_id, i,
3338 pgdat->node_present_pages = realtotalpages;
3339 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3343 #ifndef CONFIG_SPARSEMEM
3345 * Calculate the size of the zone->blockflags rounded to an unsigned long
3346 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3347 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3348 * round what is now in bits to nearest long in bits, then return it in
3351 static unsigned long __init usemap_size(unsigned long zonesize)
3353 unsigned long usemapsize;
3355 usemapsize = roundup(zonesize, pageblock_nr_pages);
3356 usemapsize = usemapsize >> pageblock_order;
3357 usemapsize *= NR_PAGEBLOCK_BITS;
3358 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
3360 return usemapsize / 8;
3363 static void __init setup_usemap(struct pglist_data *pgdat,
3364 struct zone *zone, unsigned long zonesize)
3366 unsigned long usemapsize = usemap_size(zonesize);
3367 zone->pageblock_flags = NULL;
3369 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
3370 memset(zone->pageblock_flags, 0, usemapsize);
3374 static void inline setup_usemap(struct pglist_data *pgdat,
3375 struct zone *zone, unsigned long zonesize) {}
3376 #endif /* CONFIG_SPARSEMEM */
3378 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3380 /* Return a sensible default order for the pageblock size. */
3381 static inline int pageblock_default_order(void)
3383 if (HPAGE_SHIFT > PAGE_SHIFT)
3384 return HUGETLB_PAGE_ORDER;
3389 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3390 static inline void __init set_pageblock_order(unsigned int order)
3392 /* Check that pageblock_nr_pages has not already been setup */
3393 if (pageblock_order)
3397 * Assume the largest contiguous order of interest is a huge page.
3398 * This value may be variable depending on boot parameters on IA64
3400 pageblock_order = order;
3402 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3405 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3406 * and pageblock_default_order() are unused as pageblock_order is set
3407 * at compile-time. See include/linux/pageblock-flags.h for the values of
3408 * pageblock_order based on the kernel config
3410 static inline int pageblock_default_order(unsigned int order)
3414 #define set_pageblock_order(x) do {} while (0)
3416 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3419 * Set up the zone data structures:
3420 * - mark all pages reserved
3421 * - mark all memory queues empty
3422 * - clear the memory bitmaps
3424 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
3425 unsigned long *zones_size, unsigned long *zholes_size)
3428 int nid = pgdat->node_id;
3429 unsigned long zone_start_pfn = pgdat->node_start_pfn;
3432 pgdat_resize_init(pgdat);
3433 pgdat->nr_zones = 0;
3434 init_waitqueue_head(&pgdat->kswapd_wait);
3435 pgdat->kswapd_max_order = 0;
3437 for (j = 0; j < MAX_NR_ZONES; j++) {
3438 struct zone *zone = pgdat->node_zones + j;
3439 unsigned long size, realsize, memmap_pages;
3442 size = zone_spanned_pages_in_node(nid, j, zones_size);
3443 realsize = size - zone_absent_pages_in_node(nid, j,
3447 * Adjust realsize so that it accounts for how much memory
3448 * is used by this zone for memmap. This affects the watermark
3449 * and per-cpu initialisations
3452 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
3453 if (realsize >= memmap_pages) {
3454 realsize -= memmap_pages;
3455 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3456 "%s zone: %lu pages used for memmap\n",
3457 zone_names[j], memmap_pages);
3460 " %s zone: %lu pages exceeds realsize %lu\n",
3461 zone_names[j], memmap_pages, realsize);
3463 /* Account for reserved pages */
3464 if (j == 0 && realsize > dma_reserve) {
3465 realsize -= dma_reserve;
3466 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3467 "%s zone: %lu pages reserved\n",
3468 zone_names[0], dma_reserve);
3471 if (!is_highmem_idx(j))
3472 nr_kernel_pages += realsize;
3473 nr_all_pages += realsize;
3475 zone->spanned_pages = size;
3476 zone->present_pages = realsize;
3479 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
3481 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
3483 zone->name = zone_names[j];
3484 spin_lock_init(&zone->lock);
3485 spin_lock_init(&zone->lru_lock);
3486 zone_seqlock_init(zone);
3487 zone->zone_pgdat = pgdat;
3489 zone->prev_priority = DEF_PRIORITY;
3491 zone_pcp_init(zone);
3493 INIT_LIST_HEAD(&zone->lru[l].list);
3494 zone->lru[l].nr_scan = 0;
3496 zone->recent_rotated[0] = 0;
3497 zone->recent_rotated[1] = 0;
3498 zone->recent_scanned[0] = 0;
3499 zone->recent_scanned[1] = 0;
3500 zap_zone_vm_stats(zone);
3505 set_pageblock_order(pageblock_default_order());
3506 setup_usemap(pgdat, zone, size);
3507 ret = init_currently_empty_zone(zone, zone_start_pfn,
3508 size, MEMMAP_EARLY);
3510 memmap_init(size, nid, j, zone_start_pfn);
3511 zone_start_pfn += size;
3515 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
3517 /* Skip empty nodes */
3518 if (!pgdat->node_spanned_pages)
3521 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3522 /* ia64 gets its own node_mem_map, before this, without bootmem */
3523 if (!pgdat->node_mem_map) {
3524 unsigned long size, start, end;
3528 * The zone's endpoints aren't required to be MAX_ORDER
3529 * aligned but the node_mem_map endpoints must be in order
3530 * for the buddy allocator to function correctly.
3532 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
3533 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
3534 end = ALIGN(end, MAX_ORDER_NR_PAGES);
3535 size = (end - start) * sizeof(struct page);
3536 map = alloc_remap(pgdat->node_id, size);
3538 map = alloc_bootmem_node(pgdat, size);
3539 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
3541 #ifndef CONFIG_NEED_MULTIPLE_NODES
3543 * With no DISCONTIG, the global mem_map is just set as node 0's
3545 if (pgdat == NODE_DATA(0)) {
3546 mem_map = NODE_DATA(0)->node_mem_map;
3547 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3548 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
3549 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
3550 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3553 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3556 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
3557 unsigned long node_start_pfn, unsigned long *zholes_size)
3559 pg_data_t *pgdat = NODE_DATA(nid);
3561 pgdat->node_id = nid;
3562 pgdat->node_start_pfn = node_start_pfn;
3563 calculate_node_totalpages(pgdat, zones_size, zholes_size);
3565 alloc_node_mem_map(pgdat);
3566 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3567 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3568 nid, (unsigned long)pgdat,
3569 (unsigned long)pgdat->node_mem_map);
3572 free_area_init_core(pgdat, zones_size, zholes_size);
3575 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3577 #if MAX_NUMNODES > 1
3579 * Figure out the number of possible node ids.
3581 static void __init setup_nr_node_ids(void)
3584 unsigned int highest = 0;
3586 for_each_node_mask(node, node_possible_map)
3588 nr_node_ids = highest + 1;
3591 static inline void setup_nr_node_ids(void)
3597 * add_active_range - Register a range of PFNs backed by physical memory
3598 * @nid: The node ID the range resides on
3599 * @start_pfn: The start PFN of the available physical memory
3600 * @end_pfn: The end PFN of the available physical memory
3602 * These ranges are stored in an early_node_map[] and later used by
3603 * free_area_init_nodes() to calculate zone sizes and holes. If the
3604 * range spans a memory hole, it is up to the architecture to ensure
3605 * the memory is not freed by the bootmem allocator. If possible
3606 * the range being registered will be merged with existing ranges.
3608 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
3609 unsigned long end_pfn)
3613 mminit_dprintk(MMINIT_TRACE, "memory_register",
3614 "Entering add_active_range(%d, %#lx, %#lx) "
3615 "%d entries of %d used\n",
3616 nid, start_pfn, end_pfn,
3617 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
3619 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
3621 /* Merge with existing active regions if possible */
3622 for (i = 0; i < nr_nodemap_entries; i++) {
3623 if (early_node_map[i].nid != nid)
3626 /* Skip if an existing region covers this new one */
3627 if (start_pfn >= early_node_map[i].start_pfn &&
3628 end_pfn <= early_node_map[i].end_pfn)
3631 /* Merge forward if suitable */
3632 if (start_pfn <= early_node_map[i].end_pfn &&
3633 end_pfn > early_node_map[i].end_pfn) {
3634 early_node_map[i].end_pfn = end_pfn;
3638 /* Merge backward if suitable */
3639 if (start_pfn < early_node_map[i].end_pfn &&
3640 end_pfn >= early_node_map[i].start_pfn) {
3641 early_node_map[i].start_pfn = start_pfn;
3646 /* Check that early_node_map is large enough */
3647 if (i >= MAX_ACTIVE_REGIONS) {
3648 printk(KERN_CRIT "More than %d memory regions, truncating\n",
3649 MAX_ACTIVE_REGIONS);
3653 early_node_map[i].nid = nid;
3654 early_node_map[i].start_pfn = start_pfn;
3655 early_node_map[i].end_pfn = end_pfn;
3656 nr_nodemap_entries = i + 1;
3660 * remove_active_range - Shrink an existing registered range of PFNs
3661 * @nid: The node id the range is on that should be shrunk
3662 * @start_pfn: The new PFN of the range
3663 * @end_pfn: The new PFN of the range
3665 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
3666 * The map is kept near the end physical page range that has already been
3667 * registered. This function allows an arch to shrink an existing registered
3670 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
3671 unsigned long end_pfn)
3676 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
3677 nid, start_pfn, end_pfn);
3679 /* Find the old active region end and shrink */
3680 for_each_active_range_index_in_nid(i, nid) {
3681 if (early_node_map[i].start_pfn >= start_pfn &&
3682 early_node_map[i].end_pfn <= end_pfn) {
3684 early_node_map[i].start_pfn = 0;
3685 early_node_map[i].end_pfn = 0;
3689 if (early_node_map[i].start_pfn < start_pfn &&
3690 early_node_map[i].end_pfn > start_pfn) {
3691 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
3692 early_node_map[i].end_pfn = start_pfn;
3693 if (temp_end_pfn > end_pfn)
3694 add_active_range(nid, end_pfn, temp_end_pfn);
3697 if (early_node_map[i].start_pfn >= start_pfn &&
3698 early_node_map[i].end_pfn > end_pfn &&
3699 early_node_map[i].start_pfn < end_pfn) {
3700 early_node_map[i].start_pfn = end_pfn;
3708 /* remove the blank ones */
3709 for (i = nr_nodemap_entries - 1; i > 0; i--) {
3710 if (early_node_map[i].nid != nid)
3712 if (early_node_map[i].end_pfn)
3714 /* we found it, get rid of it */
3715 for (j = i; j < nr_nodemap_entries - 1; j++)
3716 memcpy(&early_node_map[j], &early_node_map[j+1],
3717 sizeof(early_node_map[j]));
3718 j = nr_nodemap_entries - 1;
3719 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
3720 nr_nodemap_entries--;
3725 * remove_all_active_ranges - Remove all currently registered regions
3727 * During discovery, it may be found that a table like SRAT is invalid
3728 * and an alternative discovery method must be used. This function removes
3729 * all currently registered regions.
3731 void __init remove_all_active_ranges(void)
3733 memset(early_node_map, 0, sizeof(early_node_map));
3734 nr_nodemap_entries = 0;
3735 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
3736 memset(node_boundary_start_pfn, 0, sizeof(node_boundary_start_pfn));
3737 memset(node_boundary_end_pfn, 0, sizeof(node_boundary_end_pfn));
3738 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
3741 /* Compare two active node_active_regions */
3742 static int __init cmp_node_active_region(const void *a, const void *b)
3744 struct node_active_region *arange = (struct node_active_region *)a;
3745 struct node_active_region *brange = (struct node_active_region *)b;
3747 /* Done this way to avoid overflows */
3748 if (arange->start_pfn > brange->start_pfn)
3750 if (arange->start_pfn < brange->start_pfn)
3756 /* sort the node_map by start_pfn */
3757 static void __init sort_node_map(void)
3759 sort(early_node_map, (size_t)nr_nodemap_entries,
3760 sizeof(struct node_active_region),
3761 cmp_node_active_region, NULL);
3764 /* Find the lowest pfn for a node */
3765 static unsigned long __init find_min_pfn_for_node(int nid)
3768 unsigned long min_pfn = ULONG_MAX;
3770 /* Assuming a sorted map, the first range found has the starting pfn */
3771 for_each_active_range_index_in_nid(i, nid)
3772 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
3774 if (min_pfn == ULONG_MAX) {
3776 "Could not find start_pfn for node %d\n", nid);
3784 * find_min_pfn_with_active_regions - Find the minimum PFN registered
3786 * It returns the minimum PFN based on information provided via
3787 * add_active_range().
3789 unsigned long __init find_min_pfn_with_active_regions(void)
3791 return find_min_pfn_for_node(MAX_NUMNODES);
3795 * early_calculate_totalpages()
3796 * Sum pages in active regions for movable zone.
3797 * Populate N_HIGH_MEMORY for calculating usable_nodes.
3799 static unsigned long __init early_calculate_totalpages(void)
3802 unsigned long totalpages = 0;
3804 for (i = 0; i < nr_nodemap_entries; i++) {
3805 unsigned long pages = early_node_map[i].end_pfn -
3806 early_node_map[i].start_pfn;
3807 totalpages += pages;
3809 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
3815 * Find the PFN the Movable zone begins in each node. Kernel memory
3816 * is spread evenly between nodes as long as the nodes have enough
3817 * memory. When they don't, some nodes will have more kernelcore than
3820 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
3823 unsigned long usable_startpfn;
3824 unsigned long kernelcore_node, kernelcore_remaining;
3825 unsigned long totalpages = early_calculate_totalpages();
3826 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
3829 * If movablecore was specified, calculate what size of
3830 * kernelcore that corresponds so that memory usable for
3831 * any allocation type is evenly spread. If both kernelcore
3832 * and movablecore are specified, then the value of kernelcore
3833 * will be used for required_kernelcore if it's greater than
3834 * what movablecore would have allowed.
3836 if (required_movablecore) {
3837 unsigned long corepages;
3840 * Round-up so that ZONE_MOVABLE is at least as large as what
3841 * was requested by the user
3843 required_movablecore =
3844 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
3845 corepages = totalpages - required_movablecore;
3847 required_kernelcore = max(required_kernelcore, corepages);
3850 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
3851 if (!required_kernelcore)
3854 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
3855 find_usable_zone_for_movable();
3856 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
3859 /* Spread kernelcore memory as evenly as possible throughout nodes */
3860 kernelcore_node = required_kernelcore / usable_nodes;
3861 for_each_node_state(nid, N_HIGH_MEMORY) {
3863 * Recalculate kernelcore_node if the division per node
3864 * now exceeds what is necessary to satisfy the requested
3865 * amount of memory for the kernel
3867 if (required_kernelcore < kernelcore_node)
3868 kernelcore_node = required_kernelcore / usable_nodes;
3871 * As the map is walked, we track how much memory is usable
3872 * by the kernel using kernelcore_remaining. When it is
3873 * 0, the rest of the node is usable by ZONE_MOVABLE
3875 kernelcore_remaining = kernelcore_node;
3877 /* Go through each range of PFNs within this node */
3878 for_each_active_range_index_in_nid(i, nid) {
3879 unsigned long start_pfn, end_pfn;
3880 unsigned long size_pages;
3882 start_pfn = max(early_node_map[i].start_pfn,
3883 zone_movable_pfn[nid]);
3884 end_pfn = early_node_map[i].end_pfn;
3885 if (start_pfn >= end_pfn)
3888 /* Account for what is only usable for kernelcore */
3889 if (start_pfn < usable_startpfn) {
3890 unsigned long kernel_pages;
3891 kernel_pages = min(end_pfn, usable_startpfn)
3894 kernelcore_remaining -= min(kernel_pages,
3895 kernelcore_remaining);
3896 required_kernelcore -= min(kernel_pages,
3897 required_kernelcore);
3899 /* Continue if range is now fully accounted */
3900 if (end_pfn <= usable_startpfn) {
3903 * Push zone_movable_pfn to the end so
3904 * that if we have to rebalance
3905 * kernelcore across nodes, we will
3906 * not double account here
3908 zone_movable_pfn[nid] = end_pfn;
3911 start_pfn = usable_startpfn;
3915 * The usable PFN range for ZONE_MOVABLE is from
3916 * start_pfn->end_pfn. Calculate size_pages as the
3917 * number of pages used as kernelcore
3919 size_pages = end_pfn - start_pfn;
3920 if (size_pages > kernelcore_remaining)
3921 size_pages = kernelcore_remaining;
3922 zone_movable_pfn[nid] = start_pfn + size_pages;
3925 * Some kernelcore has been met, update counts and
3926 * break if the kernelcore for this node has been
3929 required_kernelcore -= min(required_kernelcore,
3931 kernelcore_remaining -= size_pages;
3932 if (!kernelcore_remaining)
3938 * If there is still required_kernelcore, we do another pass with one
3939 * less node in the count. This will push zone_movable_pfn[nid] further
3940 * along on the nodes that still have memory until kernelcore is
3944 if (usable_nodes && required_kernelcore > usable_nodes)
3947 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
3948 for (nid = 0; nid < MAX_NUMNODES; nid++)
3949 zone_movable_pfn[nid] =
3950 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
3953 /* Any regular memory on that node ? */
3954 static void check_for_regular_memory(pg_data_t *pgdat)
3956 #ifdef CONFIG_HIGHMEM
3957 enum zone_type zone_type;
3959 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
3960 struct zone *zone = &pgdat->node_zones[zone_type];
3961 if (zone->present_pages)
3962 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
3968 * free_area_init_nodes - Initialise all pg_data_t and zone data
3969 * @max_zone_pfn: an array of max PFNs for each zone
3971 * This will call free_area_init_node() for each active node in the system.
3972 * Using the page ranges provided by add_active_range(), the size of each
3973 * zone in each node and their holes is calculated. If the maximum PFN
3974 * between two adjacent zones match, it is assumed that the zone is empty.
3975 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
3976 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
3977 * starts where the previous one ended. For example, ZONE_DMA32 starts
3978 * at arch_max_dma_pfn.
3980 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
3985 /* Sort early_node_map as initialisation assumes it is sorted */
3988 /* Record where the zone boundaries are */
3989 memset(arch_zone_lowest_possible_pfn, 0,
3990 sizeof(arch_zone_lowest_possible_pfn));
3991 memset(arch_zone_highest_possible_pfn, 0,
3992 sizeof(arch_zone_highest_possible_pfn));
3993 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
3994 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
3995 for (i = 1; i < MAX_NR_ZONES; i++) {
3996 if (i == ZONE_MOVABLE)
3998 arch_zone_lowest_possible_pfn[i] =
3999 arch_zone_highest_possible_pfn[i-1];
4000 arch_zone_highest_possible_pfn[i] =
4001 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4003 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4004 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4006 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4007 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4008 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4010 /* Print out the zone ranges */
4011 printk("Zone PFN ranges:\n");
4012 for (i = 0; i < MAX_NR_ZONES; i++) {
4013 if (i == ZONE_MOVABLE)
4015 printk(" %-8s %0#10lx -> %0#10lx\n",
4017 arch_zone_lowest_possible_pfn[i],
4018 arch_zone_highest_possible_pfn[i]);
4021 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4022 printk("Movable zone start PFN for each node\n");
4023 for (i = 0; i < MAX_NUMNODES; i++) {
4024 if (zone_movable_pfn[i])
4025 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4028 /* Print out the early_node_map[] */
4029 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4030 for (i = 0; i < nr_nodemap_entries; i++)
4031 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4032 early_node_map[i].start_pfn,
4033 early_node_map[i].end_pfn);
4035 /* Initialise every node */
4036 mminit_verify_pageflags_layout();
4037 setup_nr_node_ids();
4038 for_each_online_node(nid) {
4039 pg_data_t *pgdat = NODE_DATA(nid);
4040 free_area_init_node(nid, NULL,
4041 find_min_pfn_for_node(nid), NULL);
4043 /* Any memory on that node */
4044 if (pgdat->node_present_pages)
4045 node_set_state(nid, N_HIGH_MEMORY);
4046 check_for_regular_memory(pgdat);
4050 static int __init cmdline_parse_core(char *p, unsigned long *core)
4052 unsigned long long coremem;
4056 coremem = memparse(p, &p);
4057 *core = coremem >> PAGE_SHIFT;
4059 /* Paranoid check that UL is enough for the coremem value */
4060 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4066 * kernelcore=size sets the amount of memory for use for allocations that
4067 * cannot be reclaimed or migrated.
4069 static int __init cmdline_parse_kernelcore(char *p)
4071 return cmdline_parse_core(p, &required_kernelcore);
4075 * movablecore=size sets the amount of memory for use for allocations that
4076 * can be reclaimed or migrated.
4078 static int __init cmdline_parse_movablecore(char *p)
4080 return cmdline_parse_core(p, &required_movablecore);
4083 early_param("kernelcore", cmdline_parse_kernelcore);
4084 early_param("movablecore", cmdline_parse_movablecore);
4086 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4089 * set_dma_reserve - set the specified number of pages reserved in the first zone
4090 * @new_dma_reserve: The number of pages to mark reserved
4092 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4093 * In the DMA zone, a significant percentage may be consumed by kernel image
4094 * and other unfreeable allocations which can skew the watermarks badly. This
4095 * function may optionally be used to account for unfreeable pages in the
4096 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4097 * smaller per-cpu batchsize.
4099 void __init set_dma_reserve(unsigned long new_dma_reserve)
4101 dma_reserve = new_dma_reserve;
4104 #ifndef CONFIG_NEED_MULTIPLE_NODES
4105 struct pglist_data __refdata contig_page_data = { .bdata = &bootmem_node_data[0] };
4106 EXPORT_SYMBOL(contig_page_data);
4109 void __init free_area_init(unsigned long *zones_size)
4111 free_area_init_node(0, zones_size,
4112 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4115 static int page_alloc_cpu_notify(struct notifier_block *self,
4116 unsigned long action, void *hcpu)
4118 int cpu = (unsigned long)hcpu;
4120 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4124 * Spill the event counters of the dead processor
4125 * into the current processors event counters.
4126 * This artificially elevates the count of the current
4129 vm_events_fold_cpu(cpu);
4132 * Zero the differential counters of the dead processor
4133 * so that the vm statistics are consistent.
4135 * This is only okay since the processor is dead and cannot
4136 * race with what we are doing.
4138 refresh_cpu_vm_stats(cpu);
4143 void __init page_alloc_init(void)
4145 hotcpu_notifier(page_alloc_cpu_notify, 0);
4149 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4150 * or min_free_kbytes changes.
4152 static void calculate_totalreserve_pages(void)
4154 struct pglist_data *pgdat;
4155 unsigned long reserve_pages = 0;
4156 enum zone_type i, j;
4158 for_each_online_pgdat(pgdat) {
4159 for (i = 0; i < MAX_NR_ZONES; i++) {
4160 struct zone *zone = pgdat->node_zones + i;
4161 unsigned long max = 0;
4163 /* Find valid and maximum lowmem_reserve in the zone */
4164 for (j = i; j < MAX_NR_ZONES; j++) {
4165 if (zone->lowmem_reserve[j] > max)
4166 max = zone->lowmem_reserve[j];
4169 /* we treat pages_high as reserved pages. */
4170 max += zone->pages_high;
4172 if (max > zone->present_pages)
4173 max = zone->present_pages;
4174 reserve_pages += max;
4177 totalreserve_pages = reserve_pages;
4181 * setup_per_zone_lowmem_reserve - called whenever
4182 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4183 * has a correct pages reserved value, so an adequate number of
4184 * pages are left in the zone after a successful __alloc_pages().
4186 static void setup_per_zone_lowmem_reserve(void)
4188 struct pglist_data *pgdat;
4189 enum zone_type j, idx;
4191 for_each_online_pgdat(pgdat) {
4192 for (j = 0; j < MAX_NR_ZONES; j++) {
4193 struct zone *zone = pgdat->node_zones + j;
4194 unsigned long present_pages = zone->present_pages;
4196 zone->lowmem_reserve[j] = 0;
4200 struct zone *lower_zone;
4204 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4205 sysctl_lowmem_reserve_ratio[idx] = 1;
4207 lower_zone = pgdat->node_zones + idx;
4208 lower_zone->lowmem_reserve[j] = present_pages /
4209 sysctl_lowmem_reserve_ratio[idx];
4210 present_pages += lower_zone->present_pages;
4215 /* update totalreserve_pages */
4216 calculate_totalreserve_pages();
4220 * setup_per_zone_pages_min - called when min_free_kbytes changes.
4222 * Ensures that the pages_{min,low,high} values for each zone are set correctly
4223 * with respect to min_free_kbytes.
4225 void setup_per_zone_pages_min(void)
4227 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4228 unsigned long lowmem_pages = 0;
4230 unsigned long flags;
4232 /* Calculate total number of !ZONE_HIGHMEM pages */
4233 for_each_zone(zone) {
4234 if (!is_highmem(zone))
4235 lowmem_pages += zone->present_pages;
4238 for_each_zone(zone) {
4241 spin_lock_irqsave(&zone->lru_lock, flags);
4242 tmp = (u64)pages_min * zone->present_pages;
4243 do_div(tmp, lowmem_pages);
4244 if (is_highmem(zone)) {
4246 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4247 * need highmem pages, so cap pages_min to a small
4250 * The (pages_high-pages_low) and (pages_low-pages_min)
4251 * deltas controls asynch page reclaim, and so should
4252 * not be capped for highmem.
4256 min_pages = zone->present_pages / 1024;
4257 if (min_pages < SWAP_CLUSTER_MAX)
4258 min_pages = SWAP_CLUSTER_MAX;
4259 if (min_pages > 128)
4261 zone->pages_min = min_pages;
4264 * If it's a lowmem zone, reserve a number of pages
4265 * proportionate to the zone's size.
4267 zone->pages_min = tmp;
4270 zone->pages_low = zone->pages_min + (tmp >> 2);
4271 zone->pages_high = zone->pages_min + (tmp >> 1);
4272 setup_zone_migrate_reserve(zone);
4273 spin_unlock_irqrestore(&zone->lru_lock, flags);
4276 /* update totalreserve_pages */
4277 calculate_totalreserve_pages();
4281 * setup_per_zone_inactive_ratio - called when min_free_kbytes changes.
4283 * The inactive anon list should be small enough that the VM never has to
4284 * do too much work, but large enough that each inactive page has a chance
4285 * to be referenced again before it is swapped out.
4287 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4288 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4289 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4290 * the anonymous pages are kept on the inactive list.
4293 * memory ratio inactive anon
4294 * -------------------------------------
4303 void setup_per_zone_inactive_ratio(void)
4307 for_each_zone(zone) {
4308 unsigned int gb, ratio;
4310 /* Zone size in gigabytes */
4311 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4312 ratio = int_sqrt(10 * gb);
4316 zone->inactive_ratio = ratio;
4321 * Initialise min_free_kbytes.
4323 * For small machines we want it small (128k min). For large machines
4324 * we want it large (64MB max). But it is not linear, because network
4325 * bandwidth does not increase linearly with machine size. We use
4327 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4328 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4344 static int __init init_per_zone_pages_min(void)
4346 unsigned long lowmem_kbytes;
4348 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4350 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4351 if (min_free_kbytes < 128)
4352 min_free_kbytes = 128;
4353 if (min_free_kbytes > 65536)
4354 min_free_kbytes = 65536;
4355 setup_per_zone_pages_min();
4356 setup_per_zone_lowmem_reserve();
4357 setup_per_zone_inactive_ratio();
4360 module_init(init_per_zone_pages_min)
4363 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4364 * that we can call two helper functions whenever min_free_kbytes
4367 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
4368 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4370 proc_dointvec(table, write, file, buffer, length, ppos);
4372 setup_per_zone_pages_min();
4377 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
4378 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4383 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4388 zone->min_unmapped_pages = (zone->present_pages *
4389 sysctl_min_unmapped_ratio) / 100;
4393 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
4394 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4399 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4404 zone->min_slab_pages = (zone->present_pages *
4405 sysctl_min_slab_ratio) / 100;
4411 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4412 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4413 * whenever sysctl_lowmem_reserve_ratio changes.
4415 * The reserve ratio obviously has absolutely no relation with the
4416 * pages_min watermarks. The lowmem reserve ratio can only make sense
4417 * if in function of the boot time zone sizes.
4419 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
4420 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4422 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4423 setup_per_zone_lowmem_reserve();
4428 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4429 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4430 * can have before it gets flushed back to buddy allocator.
4433 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
4434 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4440 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4441 if (!write || (ret == -EINVAL))
4443 for_each_zone(zone) {
4444 for_each_online_cpu(cpu) {
4446 high = zone->present_pages / percpu_pagelist_fraction;
4447 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
4453 int hashdist = HASHDIST_DEFAULT;
4456 static int __init set_hashdist(char *str)
4460 hashdist = simple_strtoul(str, &str, 0);
4463 __setup("hashdist=", set_hashdist);
4467 * allocate a large system hash table from bootmem
4468 * - it is assumed that the hash table must contain an exact power-of-2
4469 * quantity of entries
4470 * - limit is the number of hash buckets, not the total allocation size
4472 void *__init alloc_large_system_hash(const char *tablename,
4473 unsigned long bucketsize,
4474 unsigned long numentries,
4477 unsigned int *_hash_shift,
4478 unsigned int *_hash_mask,
4479 unsigned long limit)
4481 unsigned long long max = limit;
4482 unsigned long log2qty, size;
4485 /* allow the kernel cmdline to have a say */
4487 /* round applicable memory size up to nearest megabyte */
4488 numentries = nr_kernel_pages;
4489 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
4490 numentries >>= 20 - PAGE_SHIFT;
4491 numentries <<= 20 - PAGE_SHIFT;
4493 /* limit to 1 bucket per 2^scale bytes of low memory */
4494 if (scale > PAGE_SHIFT)
4495 numentries >>= (scale - PAGE_SHIFT);
4497 numentries <<= (PAGE_SHIFT - scale);
4499 /* Make sure we've got at least a 0-order allocation.. */
4500 if (unlikely((numentries * bucketsize) < PAGE_SIZE))
4501 numentries = PAGE_SIZE / bucketsize;
4503 numentries = roundup_pow_of_two(numentries);
4505 /* limit allocation size to 1/16 total memory by default */
4507 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
4508 do_div(max, bucketsize);
4511 if (numentries > max)
4514 log2qty = ilog2(numentries);
4517 size = bucketsize << log2qty;
4518 if (flags & HASH_EARLY)
4519 table = alloc_bootmem_nopanic(size);
4521 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
4523 unsigned long order = get_order(size);
4524 table = (void*) __get_free_pages(GFP_ATOMIC, order);
4526 * If bucketsize is not a power-of-two, we may free
4527 * some pages at the end of hash table.
4530 unsigned long alloc_end = (unsigned long)table +
4531 (PAGE_SIZE << order);
4532 unsigned long used = (unsigned long)table +
4534 split_page(virt_to_page(table), order);
4535 while (used < alloc_end) {
4541 } while (!table && size > PAGE_SIZE && --log2qty);
4544 panic("Failed to allocate %s hash table\n", tablename);
4546 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
4549 ilog2(size) - PAGE_SHIFT,
4553 *_hash_shift = log2qty;
4555 *_hash_mask = (1 << log2qty) - 1;
4560 #ifdef CONFIG_OUT_OF_LINE_PFN_TO_PAGE
4561 struct page *pfn_to_page(unsigned long pfn)
4563 return __pfn_to_page(pfn);
4565 unsigned long page_to_pfn(struct page *page)
4567 return __page_to_pfn(page);
4569 EXPORT_SYMBOL(pfn_to_page);
4570 EXPORT_SYMBOL(page_to_pfn);
4571 #endif /* CONFIG_OUT_OF_LINE_PFN_TO_PAGE */
4573 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4574 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
4577 #ifdef CONFIG_SPARSEMEM
4578 return __pfn_to_section(pfn)->pageblock_flags;
4580 return zone->pageblock_flags;
4581 #endif /* CONFIG_SPARSEMEM */
4584 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
4586 #ifdef CONFIG_SPARSEMEM
4587 pfn &= (PAGES_PER_SECTION-1);
4588 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4590 pfn = pfn - zone->zone_start_pfn;
4591 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4592 #endif /* CONFIG_SPARSEMEM */
4596 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4597 * @page: The page within the block of interest
4598 * @start_bitidx: The first bit of interest to retrieve
4599 * @end_bitidx: The last bit of interest
4600 * returns pageblock_bits flags
4602 unsigned long get_pageblock_flags_group(struct page *page,
4603 int start_bitidx, int end_bitidx)
4606 unsigned long *bitmap;
4607 unsigned long pfn, bitidx;
4608 unsigned long flags = 0;
4609 unsigned long value = 1;
4611 zone = page_zone(page);
4612 pfn = page_to_pfn(page);
4613 bitmap = get_pageblock_bitmap(zone, pfn);
4614 bitidx = pfn_to_bitidx(zone, pfn);
4616 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4617 if (test_bit(bitidx + start_bitidx, bitmap))
4624 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4625 * @page: The page within the block of interest
4626 * @start_bitidx: The first bit of interest
4627 * @end_bitidx: The last bit of interest
4628 * @flags: The flags to set
4630 void set_pageblock_flags_group(struct page *page, unsigned long flags,
4631 int start_bitidx, int end_bitidx)
4634 unsigned long *bitmap;
4635 unsigned long pfn, bitidx;
4636 unsigned long value = 1;
4638 zone = page_zone(page);
4639 pfn = page_to_pfn(page);
4640 bitmap = get_pageblock_bitmap(zone, pfn);
4641 bitidx = pfn_to_bitidx(zone, pfn);
4642 VM_BUG_ON(pfn < zone->zone_start_pfn);
4643 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
4645 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4647 __set_bit(bitidx + start_bitidx, bitmap);
4649 __clear_bit(bitidx + start_bitidx, bitmap);
4653 * This is designed as sub function...plz see page_isolation.c also.
4654 * set/clear page block's type to be ISOLATE.
4655 * page allocater never alloc memory from ISOLATE block.
4658 int set_migratetype_isolate(struct page *page)
4661 unsigned long flags;
4664 zone = page_zone(page);
4665 spin_lock_irqsave(&zone->lock, flags);
4667 * In future, more migrate types will be able to be isolation target.
4669 if (get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
4671 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
4672 move_freepages_block(zone, page, MIGRATE_ISOLATE);
4675 spin_unlock_irqrestore(&zone->lock, flags);
4681 void unset_migratetype_isolate(struct page *page)
4684 unsigned long flags;
4685 zone = page_zone(page);
4686 spin_lock_irqsave(&zone->lock, flags);
4687 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
4689 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4690 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4692 spin_unlock_irqrestore(&zone->lock, flags);
4695 #ifdef CONFIG_MEMORY_HOTREMOVE
4697 * All pages in the range must be isolated before calling this.
4700 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
4706 unsigned long flags;
4707 /* find the first valid pfn */
4708 for (pfn = start_pfn; pfn < end_pfn; pfn++)
4713 zone = page_zone(pfn_to_page(pfn));
4714 spin_lock_irqsave(&zone->lock, flags);
4716 while (pfn < end_pfn) {
4717 if (!pfn_valid(pfn)) {
4721 page = pfn_to_page(pfn);
4722 BUG_ON(page_count(page));
4723 BUG_ON(!PageBuddy(page));
4724 order = page_order(page);
4725 #ifdef CONFIG_DEBUG_VM
4726 printk(KERN_INFO "remove from free list %lx %d %lx\n",
4727 pfn, 1 << order, end_pfn);
4729 list_del(&page->lru);
4730 rmv_page_order(page);
4731 zone->free_area[order].nr_free--;
4732 __mod_zone_page_state(zone, NR_FREE_PAGES,
4734 for (i = 0; i < (1 << order); i++)
4735 SetPageReserved((page+i));
4736 pfn += (1 << order);
4738 spin_unlock_irqrestore(&zone->lock, flags);