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/kmemcheck.h>
27 #include <linux/module.h>
28 #include <linux/suspend.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/slab.h>
32 #include <linux/oom.h>
33 #include <linux/notifier.h>
34 #include <linux/topology.h>
35 #include <linux/sysctl.h>
36 #include <linux/cpu.h>
37 #include <linux/cpuset.h>
38 #include <linux/memory_hotplug.h>
39 #include <linux/nodemask.h>
40 #include <linux/vmalloc.h>
41 #include <linux/mempolicy.h>
42 #include <linux/stop_machine.h>
43 #include <linux/sort.h>
44 #include <linux/pfn.h>
45 #include <linux/backing-dev.h>
46 #include <linux/fault-inject.h>
47 #include <linux/page-isolation.h>
48 #include <linux/page_cgroup.h>
49 #include <linux/debugobjects.h>
50 #include <linux/kmemleak.h>
51 #include <trace/events/kmem.h>
53 #include <asm/tlbflush.h>
54 #include <asm/div64.h>
58 * Array of node states.
60 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
61 [N_POSSIBLE] = NODE_MASK_ALL,
62 [N_ONLINE] = { { [0] = 1UL } },
64 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
66 [N_HIGH_MEMORY] = { { [0] = 1UL } },
68 [N_CPU] = { { [0] = 1UL } },
71 EXPORT_SYMBOL(node_states);
73 unsigned long totalram_pages __read_mostly;
74 unsigned long totalreserve_pages __read_mostly;
75 unsigned long highest_memmap_pfn __read_mostly;
76 int percpu_pagelist_fraction;
77 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
79 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
80 int pageblock_order __read_mostly;
83 static void __free_pages_ok(struct page *page, unsigned int order);
86 * results with 256, 32 in the lowmem_reserve sysctl:
87 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
88 * 1G machine -> (16M dma, 784M normal, 224M high)
89 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
90 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
91 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
93 * TBD: should special case ZONE_DMA32 machines here - in those we normally
94 * don't need any ZONE_NORMAL reservation
96 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
97 #ifdef CONFIG_ZONE_DMA
100 #ifdef CONFIG_ZONE_DMA32
103 #ifdef CONFIG_HIGHMEM
109 EXPORT_SYMBOL(totalram_pages);
111 static char * const zone_names[MAX_NR_ZONES] = {
112 #ifdef CONFIG_ZONE_DMA
115 #ifdef CONFIG_ZONE_DMA32
119 #ifdef CONFIG_HIGHMEM
125 int min_free_kbytes = 1024;
127 static unsigned long __meminitdata nr_kernel_pages;
128 static unsigned long __meminitdata nr_all_pages;
129 static unsigned long __meminitdata dma_reserve;
131 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
133 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
134 * ranges of memory (RAM) that may be registered with add_active_range().
135 * Ranges passed to add_active_range() will be merged if possible
136 * so the number of times add_active_range() can be called is
137 * related to the number of nodes and the number of holes
139 #ifdef CONFIG_MAX_ACTIVE_REGIONS
140 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
141 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
143 #if MAX_NUMNODES >= 32
144 /* If there can be many nodes, allow up to 50 holes per node */
145 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
147 /* By default, allow up to 256 distinct regions */
148 #define MAX_ACTIVE_REGIONS 256
152 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
153 static int __meminitdata nr_nodemap_entries;
154 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
155 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
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 int nr_online_nodes __read_mostly = 1;
168 EXPORT_SYMBOL(nr_node_ids);
169 EXPORT_SYMBOL(nr_online_nodes);
172 int page_group_by_mobility_disabled __read_mostly;
174 static void set_pageblock_migratetype(struct page *page, int migratetype)
177 if (unlikely(page_group_by_mobility_disabled))
178 migratetype = MIGRATE_UNMOVABLE;
180 set_pageblock_flags_group(page, (unsigned long)migratetype,
181 PB_migrate, PB_migrate_end);
184 bool oom_killer_disabled __read_mostly;
186 #ifdef CONFIG_DEBUG_VM
187 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
191 unsigned long pfn = page_to_pfn(page);
194 seq = zone_span_seqbegin(zone);
195 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
197 else if (pfn < zone->zone_start_pfn)
199 } while (zone_span_seqretry(zone, seq));
204 static int page_is_consistent(struct zone *zone, struct page *page)
206 if (!pfn_valid_within(page_to_pfn(page)))
208 if (zone != page_zone(page))
214 * Temporary debugging check for pages not lying within a given zone.
216 static int bad_range(struct zone *zone, struct page *page)
218 if (page_outside_zone_boundaries(zone, page))
220 if (!page_is_consistent(zone, page))
226 static inline int bad_range(struct zone *zone, struct page *page)
232 static void bad_page(struct page *page)
234 static unsigned long resume;
235 static unsigned long nr_shown;
236 static unsigned long nr_unshown;
239 * Allow a burst of 60 reports, then keep quiet for that minute;
240 * or allow a steady drip of one report per second.
242 if (nr_shown == 60) {
243 if (time_before(jiffies, resume)) {
249 "BUG: Bad page state: %lu messages suppressed\n",
256 resume = jiffies + 60 * HZ;
258 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
259 current->comm, page_to_pfn(page));
261 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
262 page, (void *)page->flags, page_count(page),
263 page_mapcount(page), page->mapping, page->index);
267 /* Leave bad fields for debug, except PageBuddy could make trouble */
268 __ClearPageBuddy(page);
269 add_taint(TAINT_BAD_PAGE);
273 * Higher-order pages are called "compound pages". They are structured thusly:
275 * The first PAGE_SIZE page is called the "head page".
277 * The remaining PAGE_SIZE pages are called "tail pages".
279 * All pages have PG_compound set. All pages have their ->private pointing at
280 * the head page (even the head page has this).
282 * The first tail page's ->lru.next holds the address of the compound page's
283 * put_page() function. Its ->lru.prev holds the order of allocation.
284 * This usage means that zero-order pages may not be compound.
287 static void free_compound_page(struct page *page)
289 __free_pages_ok(page, compound_order(page));
292 void prep_compound_page(struct page *page, unsigned long order)
295 int nr_pages = 1 << order;
297 set_compound_page_dtor(page, free_compound_page);
298 set_compound_order(page, order);
300 for (i = 1; i < nr_pages; i++) {
301 struct page *p = page + i;
304 p->first_page = page;
308 static int destroy_compound_page(struct page *page, unsigned long order)
311 int nr_pages = 1 << order;
314 if (unlikely(compound_order(page) != order) ||
315 unlikely(!PageHead(page))) {
320 __ClearPageHead(page);
322 for (i = 1; i < nr_pages; i++) {
323 struct page *p = page + i;
325 if (unlikely(!PageTail(p) || (p->first_page != page))) {
335 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
340 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
341 * and __GFP_HIGHMEM from hard or soft interrupt context.
343 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
344 for (i = 0; i < (1 << order); i++)
345 clear_highpage(page + i);
348 static inline void set_page_order(struct page *page, int order)
350 set_page_private(page, order);
351 __SetPageBuddy(page);
354 static inline void rmv_page_order(struct page *page)
356 __ClearPageBuddy(page);
357 set_page_private(page, 0);
361 * Locate the struct page for both the matching buddy in our
362 * pair (buddy1) and the combined O(n+1) page they form (page).
364 * 1) Any buddy B1 will have an order O twin B2 which satisfies
365 * the following equation:
367 * For example, if the starting buddy (buddy2) is #8 its order
369 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
371 * 2) Any buddy B will have an order O+1 parent P which
372 * satisfies the following equation:
375 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
377 static inline struct page *
378 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
380 unsigned long buddy_idx = page_idx ^ (1 << order);
382 return page + (buddy_idx - page_idx);
385 static inline unsigned long
386 __find_combined_index(unsigned long page_idx, unsigned int order)
388 return (page_idx & ~(1 << order));
392 * This function checks whether a page is free && is the buddy
393 * we can do coalesce a page and its buddy if
394 * (a) the buddy is not in a hole &&
395 * (b) the buddy is in the buddy system &&
396 * (c) a page and its buddy have the same order &&
397 * (d) a page and its buddy are in the same zone.
399 * For recording whether a page is in the buddy system, we use PG_buddy.
400 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
402 * For recording page's order, we use page_private(page).
404 static inline int page_is_buddy(struct page *page, struct page *buddy,
407 if (!pfn_valid_within(page_to_pfn(buddy)))
410 if (page_zone_id(page) != page_zone_id(buddy))
413 if (PageBuddy(buddy) && page_order(buddy) == order) {
414 VM_BUG_ON(page_count(buddy) != 0);
421 * Freeing function for a buddy system allocator.
423 * The concept of a buddy system is to maintain direct-mapped table
424 * (containing bit values) for memory blocks of various "orders".
425 * The bottom level table contains the map for the smallest allocatable
426 * units of memory (here, pages), and each level above it describes
427 * pairs of units from the levels below, hence, "buddies".
428 * At a high level, all that happens here is marking the table entry
429 * at the bottom level available, and propagating the changes upward
430 * as necessary, plus some accounting needed to play nicely with other
431 * parts of the VM system.
432 * At each level, we keep a list of pages, which are heads of continuous
433 * free pages of length of (1 << order) and marked with PG_buddy. Page's
434 * order is recorded in page_private(page) field.
435 * So when we are allocating or freeing one, we can derive the state of the
436 * other. That is, if we allocate a small block, and both were
437 * free, the remainder of the region must be split into blocks.
438 * If a block is freed, and its buddy is also free, then this
439 * triggers coalescing into a block of larger size.
444 static inline void __free_one_page(struct page *page,
445 struct zone *zone, unsigned int order,
448 unsigned long page_idx;
450 if (unlikely(PageCompound(page)))
451 if (unlikely(destroy_compound_page(page, order)))
454 VM_BUG_ON(migratetype == -1);
456 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
458 VM_BUG_ON(page_idx & ((1 << order) - 1));
459 VM_BUG_ON(bad_range(zone, page));
461 while (order < MAX_ORDER-1) {
462 unsigned long combined_idx;
465 buddy = __page_find_buddy(page, page_idx, order);
466 if (!page_is_buddy(page, buddy, order))
469 /* Our buddy is free, merge with it and move up one order. */
470 list_del(&buddy->lru);
471 zone->free_area[order].nr_free--;
472 rmv_page_order(buddy);
473 combined_idx = __find_combined_index(page_idx, order);
474 page = page + (combined_idx - page_idx);
475 page_idx = combined_idx;
478 set_page_order(page, order);
480 &zone->free_area[order].free_list[migratetype]);
481 zone->free_area[order].nr_free++;
484 #ifdef CONFIG_HAVE_MLOCKED_PAGE_BIT
486 * free_page_mlock() -- clean up attempts to free and mlocked() page.
487 * Page should not be on lru, so no need to fix that up.
488 * free_pages_check() will verify...
490 static inline void free_page_mlock(struct page *page)
492 __dec_zone_page_state(page, NR_MLOCK);
493 __count_vm_event(UNEVICTABLE_MLOCKFREED);
496 static void free_page_mlock(struct page *page) { }
499 static inline int free_pages_check(struct page *page)
501 if (unlikely(page_mapcount(page) |
502 (page->mapping != NULL) |
503 (atomic_read(&page->_count) != 0) |
504 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) {
508 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
509 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
514 * Frees a number of pages from the PCP lists
515 * Assumes all pages on list are in same zone, and of same order.
516 * count is the number of pages to free.
518 * If the zone was previously in an "all pages pinned" state then look to
519 * see if this freeing clears that state.
521 * And clear the zone's pages_scanned counter, to hold off the "all pages are
522 * pinned" detection logic.
524 static void free_pcppages_bulk(struct zone *zone, int count,
525 struct per_cpu_pages *pcp)
530 spin_lock(&zone->lock);
531 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
532 zone->pages_scanned = 0;
534 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
537 struct list_head *list;
540 * Remove pages from lists in a round-robin fashion. A
541 * batch_free count is maintained that is incremented when an
542 * empty list is encountered. This is so more pages are freed
543 * off fuller lists instead of spinning excessively around empty
548 if (++migratetype == MIGRATE_PCPTYPES)
550 list = &pcp->lists[migratetype];
551 } while (list_empty(list));
554 page = list_entry(list->prev, struct page, lru);
555 /* must delete as __free_one_page list manipulates */
556 list_del(&page->lru);
557 __free_one_page(page, zone, 0, migratetype);
558 trace_mm_page_pcpu_drain(page, 0, migratetype);
559 } while (--count && --batch_free && !list_empty(list));
561 spin_unlock(&zone->lock);
564 static void free_one_page(struct zone *zone, struct page *page, int order,
567 spin_lock(&zone->lock);
568 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
569 zone->pages_scanned = 0;
571 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
572 __free_one_page(page, zone, order, migratetype);
573 spin_unlock(&zone->lock);
576 static void __free_pages_ok(struct page *page, unsigned int order)
581 int wasMlocked = __TestClearPageMlocked(page);
583 kmemcheck_free_shadow(page, order);
585 for (i = 0 ; i < (1 << order) ; ++i)
586 bad += free_pages_check(page + i);
590 if (!PageHighMem(page)) {
591 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
592 debug_check_no_obj_freed(page_address(page),
595 arch_free_page(page, order);
596 kernel_map_pages(page, 1 << order, 0);
598 local_irq_save(flags);
599 if (unlikely(wasMlocked))
600 free_page_mlock(page);
601 __count_vm_events(PGFREE, 1 << order);
602 free_one_page(page_zone(page), page, order,
603 get_pageblock_migratetype(page));
604 local_irq_restore(flags);
608 * permit the bootmem allocator to evade page validation on high-order frees
610 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
613 __ClearPageReserved(page);
614 set_page_count(page, 0);
615 set_page_refcounted(page);
621 for (loop = 0; loop < BITS_PER_LONG; loop++) {
622 struct page *p = &page[loop];
624 if (loop + 1 < BITS_PER_LONG)
626 __ClearPageReserved(p);
627 set_page_count(p, 0);
630 set_page_refcounted(page);
631 __free_pages(page, order);
637 * The order of subdivision here is critical for the IO subsystem.
638 * Please do not alter this order without good reasons and regression
639 * testing. Specifically, as large blocks of memory are subdivided,
640 * the order in which smaller blocks are delivered depends on the order
641 * they're subdivided in this function. This is the primary factor
642 * influencing the order in which pages are delivered to the IO
643 * subsystem according to empirical testing, and this is also justified
644 * by considering the behavior of a buddy system containing a single
645 * large block of memory acted on by a series of small allocations.
646 * This behavior is a critical factor in sglist merging's success.
650 static inline void expand(struct zone *zone, struct page *page,
651 int low, int high, struct free_area *area,
654 unsigned long size = 1 << high;
660 VM_BUG_ON(bad_range(zone, &page[size]));
661 list_add(&page[size].lru, &area->free_list[migratetype]);
663 set_page_order(&page[size], high);
668 * This page is about to be returned from the page allocator
670 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
672 if (unlikely(page_mapcount(page) |
673 (page->mapping != NULL) |
674 (atomic_read(&page->_count) != 0) |
675 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) {
680 set_page_private(page, 0);
681 set_page_refcounted(page);
683 arch_alloc_page(page, order);
684 kernel_map_pages(page, 1 << order, 1);
686 if (gfp_flags & __GFP_ZERO)
687 prep_zero_page(page, order, gfp_flags);
689 if (order && (gfp_flags & __GFP_COMP))
690 prep_compound_page(page, order);
696 * Go through the free lists for the given migratetype and remove
697 * the smallest available page from the freelists
700 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
703 unsigned int current_order;
704 struct free_area * area;
707 /* Find a page of the appropriate size in the preferred list */
708 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
709 area = &(zone->free_area[current_order]);
710 if (list_empty(&area->free_list[migratetype]))
713 page = list_entry(area->free_list[migratetype].next,
715 list_del(&page->lru);
716 rmv_page_order(page);
718 expand(zone, page, order, current_order, area, migratetype);
727 * This array describes the order lists are fallen back to when
728 * the free lists for the desirable migrate type are depleted
730 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
731 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
732 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
733 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
734 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
738 * Move the free pages in a range to the free lists of the requested type.
739 * Note that start_page and end_pages are not aligned on a pageblock
740 * boundary. If alignment is required, use move_freepages_block()
742 static int move_freepages(struct zone *zone,
743 struct page *start_page, struct page *end_page,
750 #ifndef CONFIG_HOLES_IN_ZONE
752 * page_zone is not safe to call in this context when
753 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
754 * anyway as we check zone boundaries in move_freepages_block().
755 * Remove at a later date when no bug reports exist related to
756 * grouping pages by mobility
758 BUG_ON(page_zone(start_page) != page_zone(end_page));
761 for (page = start_page; page <= end_page;) {
762 /* Make sure we are not inadvertently changing nodes */
763 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
765 if (!pfn_valid_within(page_to_pfn(page))) {
770 if (!PageBuddy(page)) {
775 order = page_order(page);
776 list_del(&page->lru);
778 &zone->free_area[order].free_list[migratetype]);
780 pages_moved += 1 << order;
786 static int move_freepages_block(struct zone *zone, struct page *page,
789 unsigned long start_pfn, end_pfn;
790 struct page *start_page, *end_page;
792 start_pfn = page_to_pfn(page);
793 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
794 start_page = pfn_to_page(start_pfn);
795 end_page = start_page + pageblock_nr_pages - 1;
796 end_pfn = start_pfn + pageblock_nr_pages - 1;
798 /* Do not cross zone boundaries */
799 if (start_pfn < zone->zone_start_pfn)
801 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
804 return move_freepages(zone, start_page, end_page, migratetype);
807 static void change_pageblock_range(struct page *pageblock_page,
808 int start_order, int migratetype)
810 int nr_pageblocks = 1 << (start_order - pageblock_order);
812 while (nr_pageblocks--) {
813 set_pageblock_migratetype(pageblock_page, migratetype);
814 pageblock_page += pageblock_nr_pages;
818 /* Remove an element from the buddy allocator from the fallback list */
819 static inline struct page *
820 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
822 struct free_area * area;
827 /* Find the largest possible block of pages in the other list */
828 for (current_order = MAX_ORDER-1; current_order >= order;
830 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
831 migratetype = fallbacks[start_migratetype][i];
833 /* MIGRATE_RESERVE handled later if necessary */
834 if (migratetype == MIGRATE_RESERVE)
837 area = &(zone->free_area[current_order]);
838 if (list_empty(&area->free_list[migratetype]))
841 page = list_entry(area->free_list[migratetype].next,
846 * If breaking a large block of pages, move all free
847 * pages to the preferred allocation list. If falling
848 * back for a reclaimable kernel allocation, be more
849 * agressive about taking ownership of free pages
851 if (unlikely(current_order >= (pageblock_order >> 1)) ||
852 start_migratetype == MIGRATE_RECLAIMABLE ||
853 page_group_by_mobility_disabled) {
855 pages = move_freepages_block(zone, page,
858 /* Claim the whole block if over half of it is free */
859 if (pages >= (1 << (pageblock_order-1)) ||
860 page_group_by_mobility_disabled)
861 set_pageblock_migratetype(page,
864 migratetype = start_migratetype;
867 /* Remove the page from the freelists */
868 list_del(&page->lru);
869 rmv_page_order(page);
871 /* Take ownership for orders >= pageblock_order */
872 if (current_order >= pageblock_order)
873 change_pageblock_range(page, current_order,
876 expand(zone, page, order, current_order, area, migratetype);
878 trace_mm_page_alloc_extfrag(page, order, current_order,
879 start_migratetype, migratetype);
889 * Do the hard work of removing an element from the buddy allocator.
890 * Call me with the zone->lock already held.
892 static struct page *__rmqueue(struct zone *zone, unsigned int order,
898 page = __rmqueue_smallest(zone, order, migratetype);
900 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
901 page = __rmqueue_fallback(zone, order, migratetype);
904 * Use MIGRATE_RESERVE rather than fail an allocation. goto
905 * is used because __rmqueue_smallest is an inline function
906 * and we want just one call site
909 migratetype = MIGRATE_RESERVE;
914 trace_mm_page_alloc_zone_locked(page, order, migratetype);
919 * Obtain a specified number of elements from the buddy allocator, all under
920 * a single hold of the lock, for efficiency. Add them to the supplied list.
921 * Returns the number of new pages which were placed at *list.
923 static int rmqueue_bulk(struct zone *zone, unsigned int order,
924 unsigned long count, struct list_head *list,
925 int migratetype, int cold)
929 spin_lock(&zone->lock);
930 for (i = 0; i < count; ++i) {
931 struct page *page = __rmqueue(zone, order, migratetype);
932 if (unlikely(page == NULL))
936 * Split buddy pages returned by expand() are received here
937 * in physical page order. The page is added to the callers and
938 * list and the list head then moves forward. From the callers
939 * perspective, the linked list is ordered by page number in
940 * some conditions. This is useful for IO devices that can
941 * merge IO requests if the physical pages are ordered
944 if (likely(cold == 0))
945 list_add(&page->lru, list);
947 list_add_tail(&page->lru, list);
948 set_page_private(page, migratetype);
951 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
952 spin_unlock(&zone->lock);
958 * Called from the vmstat counter updater to drain pagesets of this
959 * currently executing processor on remote nodes after they have
962 * Note that this function must be called with the thread pinned to
963 * a single processor.
965 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
970 local_irq_save(flags);
971 if (pcp->count >= pcp->batch)
972 to_drain = pcp->batch;
974 to_drain = pcp->count;
975 free_pcppages_bulk(zone, to_drain, pcp);
976 pcp->count -= to_drain;
977 local_irq_restore(flags);
982 * Drain pages of the indicated processor.
984 * The processor must either be the current processor and the
985 * thread pinned to the current processor or a processor that
988 static void drain_pages(unsigned int cpu)
993 for_each_populated_zone(zone) {
994 struct per_cpu_pageset *pset;
995 struct per_cpu_pages *pcp;
997 pset = zone_pcp(zone, cpu);
1000 local_irq_save(flags);
1001 free_pcppages_bulk(zone, pcp->count, pcp);
1003 local_irq_restore(flags);
1008 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1010 void drain_local_pages(void *arg)
1012 drain_pages(smp_processor_id());
1016 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1018 void drain_all_pages(void)
1020 on_each_cpu(drain_local_pages, NULL, 1);
1023 #ifdef CONFIG_HIBERNATION
1025 void mark_free_pages(struct zone *zone)
1027 unsigned long pfn, max_zone_pfn;
1028 unsigned long flags;
1030 struct list_head *curr;
1032 if (!zone->spanned_pages)
1035 spin_lock_irqsave(&zone->lock, flags);
1037 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1038 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1039 if (pfn_valid(pfn)) {
1040 struct page *page = pfn_to_page(pfn);
1042 if (!swsusp_page_is_forbidden(page))
1043 swsusp_unset_page_free(page);
1046 for_each_migratetype_order(order, t) {
1047 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1050 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1051 for (i = 0; i < (1UL << order); i++)
1052 swsusp_set_page_free(pfn_to_page(pfn + i));
1055 spin_unlock_irqrestore(&zone->lock, flags);
1057 #endif /* CONFIG_PM */
1060 * Free a 0-order page
1062 static void free_hot_cold_page(struct page *page, int cold)
1064 struct zone *zone = page_zone(page);
1065 struct per_cpu_pages *pcp;
1066 unsigned long flags;
1068 int wasMlocked = __TestClearPageMlocked(page);
1070 kmemcheck_free_shadow(page, 0);
1073 page->mapping = NULL;
1074 if (free_pages_check(page))
1077 if (!PageHighMem(page)) {
1078 debug_check_no_locks_freed(page_address(page), PAGE_SIZE);
1079 debug_check_no_obj_freed(page_address(page), PAGE_SIZE);
1081 arch_free_page(page, 0);
1082 kernel_map_pages(page, 1, 0);
1084 pcp = &zone_pcp(zone, get_cpu())->pcp;
1085 migratetype = get_pageblock_migratetype(page);
1086 set_page_private(page, migratetype);
1087 local_irq_save(flags);
1088 if (unlikely(wasMlocked))
1089 free_page_mlock(page);
1090 __count_vm_event(PGFREE);
1093 * We only track unmovable, reclaimable and movable on pcp lists.
1094 * Free ISOLATE pages back to the allocator because they are being
1095 * offlined but treat RESERVE as movable pages so we can get those
1096 * areas back if necessary. Otherwise, we may have to free
1097 * excessively into the page allocator
1099 if (migratetype >= MIGRATE_PCPTYPES) {
1100 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1101 free_one_page(zone, page, 0, migratetype);
1104 migratetype = MIGRATE_MOVABLE;
1108 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1110 list_add(&page->lru, &pcp->lists[migratetype]);
1112 if (pcp->count >= pcp->high) {
1113 free_pcppages_bulk(zone, pcp->batch, pcp);
1114 pcp->count -= pcp->batch;
1118 local_irq_restore(flags);
1122 void free_hot_page(struct page *page)
1124 trace_mm_page_free_direct(page, 0);
1125 free_hot_cold_page(page, 0);
1129 * split_page takes a non-compound higher-order page, and splits it into
1130 * n (1<<order) sub-pages: page[0..n]
1131 * Each sub-page must be freed individually.
1133 * Note: this is probably too low level an operation for use in drivers.
1134 * Please consult with lkml before using this in your driver.
1136 void split_page(struct page *page, unsigned int order)
1140 VM_BUG_ON(PageCompound(page));
1141 VM_BUG_ON(!page_count(page));
1143 #ifdef CONFIG_KMEMCHECK
1145 * Split shadow pages too, because free(page[0]) would
1146 * otherwise free the whole shadow.
1148 if (kmemcheck_page_is_tracked(page))
1149 split_page(virt_to_page(page[0].shadow), order);
1152 for (i = 1; i < (1 << order); i++)
1153 set_page_refcounted(page + i);
1157 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1158 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1162 struct page *buffered_rmqueue(struct zone *preferred_zone,
1163 struct zone *zone, int order, gfp_t gfp_flags,
1166 unsigned long flags;
1168 int cold = !!(gfp_flags & __GFP_COLD);
1173 if (likely(order == 0)) {
1174 struct per_cpu_pages *pcp;
1175 struct list_head *list;
1177 pcp = &zone_pcp(zone, cpu)->pcp;
1178 list = &pcp->lists[migratetype];
1179 local_irq_save(flags);
1180 if (list_empty(list)) {
1181 pcp->count += rmqueue_bulk(zone, 0,
1184 if (unlikely(list_empty(list)))
1189 page = list_entry(list->prev, struct page, lru);
1191 page = list_entry(list->next, struct page, lru);
1193 list_del(&page->lru);
1196 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1198 * __GFP_NOFAIL is not to be used in new code.
1200 * All __GFP_NOFAIL callers should be fixed so that they
1201 * properly detect and handle allocation failures.
1203 * We most definitely don't want callers attempting to
1204 * allocate greater than order-1 page units with
1207 WARN_ON_ONCE(order > 1);
1209 spin_lock_irqsave(&zone->lock, flags);
1210 page = __rmqueue(zone, order, migratetype);
1211 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1212 spin_unlock(&zone->lock);
1217 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1218 zone_statistics(preferred_zone, zone);
1219 local_irq_restore(flags);
1222 VM_BUG_ON(bad_range(zone, page));
1223 if (prep_new_page(page, order, gfp_flags))
1228 local_irq_restore(flags);
1233 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1234 #define ALLOC_WMARK_MIN WMARK_MIN
1235 #define ALLOC_WMARK_LOW WMARK_LOW
1236 #define ALLOC_WMARK_HIGH WMARK_HIGH
1237 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1239 /* Mask to get the watermark bits */
1240 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1242 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1243 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1244 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1246 #ifdef CONFIG_FAIL_PAGE_ALLOC
1248 static struct fail_page_alloc_attr {
1249 struct fault_attr attr;
1251 u32 ignore_gfp_highmem;
1252 u32 ignore_gfp_wait;
1255 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1257 struct dentry *ignore_gfp_highmem_file;
1258 struct dentry *ignore_gfp_wait_file;
1259 struct dentry *min_order_file;
1261 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1263 } fail_page_alloc = {
1264 .attr = FAULT_ATTR_INITIALIZER,
1265 .ignore_gfp_wait = 1,
1266 .ignore_gfp_highmem = 1,
1270 static int __init setup_fail_page_alloc(char *str)
1272 return setup_fault_attr(&fail_page_alloc.attr, str);
1274 __setup("fail_page_alloc=", setup_fail_page_alloc);
1276 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1278 if (order < fail_page_alloc.min_order)
1280 if (gfp_mask & __GFP_NOFAIL)
1282 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1284 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1287 return should_fail(&fail_page_alloc.attr, 1 << order);
1290 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1292 static int __init fail_page_alloc_debugfs(void)
1294 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1298 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1302 dir = fail_page_alloc.attr.dentries.dir;
1304 fail_page_alloc.ignore_gfp_wait_file =
1305 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1306 &fail_page_alloc.ignore_gfp_wait);
1308 fail_page_alloc.ignore_gfp_highmem_file =
1309 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1310 &fail_page_alloc.ignore_gfp_highmem);
1311 fail_page_alloc.min_order_file =
1312 debugfs_create_u32("min-order", mode, dir,
1313 &fail_page_alloc.min_order);
1315 if (!fail_page_alloc.ignore_gfp_wait_file ||
1316 !fail_page_alloc.ignore_gfp_highmem_file ||
1317 !fail_page_alloc.min_order_file) {
1319 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1320 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1321 debugfs_remove(fail_page_alloc.min_order_file);
1322 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1328 late_initcall(fail_page_alloc_debugfs);
1330 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1332 #else /* CONFIG_FAIL_PAGE_ALLOC */
1334 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1339 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1342 * Return 1 if free pages are above 'mark'. This takes into account the order
1343 * of the allocation.
1345 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1346 int classzone_idx, int alloc_flags)
1348 /* free_pages my go negative - that's OK */
1350 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
1353 if (alloc_flags & ALLOC_HIGH)
1355 if (alloc_flags & ALLOC_HARDER)
1358 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1360 for (o = 0; o < order; o++) {
1361 /* At the next order, this order's pages become unavailable */
1362 free_pages -= z->free_area[o].nr_free << o;
1364 /* Require fewer higher order pages to be free */
1367 if (free_pages <= min)
1375 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1376 * skip over zones that are not allowed by the cpuset, or that have
1377 * been recently (in last second) found to be nearly full. See further
1378 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1379 * that have to skip over a lot of full or unallowed zones.
1381 * If the zonelist cache is present in the passed in zonelist, then
1382 * returns a pointer to the allowed node mask (either the current
1383 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1385 * If the zonelist cache is not available for this zonelist, does
1386 * nothing and returns NULL.
1388 * If the fullzones BITMAP in the zonelist cache is stale (more than
1389 * a second since last zap'd) then we zap it out (clear its bits.)
1391 * We hold off even calling zlc_setup, until after we've checked the
1392 * first zone in the zonelist, on the theory that most allocations will
1393 * be satisfied from that first zone, so best to examine that zone as
1394 * quickly as we can.
1396 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1398 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1399 nodemask_t *allowednodes; /* zonelist_cache approximation */
1401 zlc = zonelist->zlcache_ptr;
1405 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1406 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1407 zlc->last_full_zap = jiffies;
1410 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1411 &cpuset_current_mems_allowed :
1412 &node_states[N_HIGH_MEMORY];
1413 return allowednodes;
1417 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1418 * if it is worth looking at further for free memory:
1419 * 1) Check that the zone isn't thought to be full (doesn't have its
1420 * bit set in the zonelist_cache fullzones BITMAP).
1421 * 2) Check that the zones node (obtained from the zonelist_cache
1422 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1423 * Return true (non-zero) if zone is worth looking at further, or
1424 * else return false (zero) if it is not.
1426 * This check -ignores- the distinction between various watermarks,
1427 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1428 * found to be full for any variation of these watermarks, it will
1429 * be considered full for up to one second by all requests, unless
1430 * we are so low on memory on all allowed nodes that we are forced
1431 * into the second scan of the zonelist.
1433 * In the second scan we ignore this zonelist cache and exactly
1434 * apply the watermarks to all zones, even it is slower to do so.
1435 * We are low on memory in the second scan, and should leave no stone
1436 * unturned looking for a free page.
1438 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1439 nodemask_t *allowednodes)
1441 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1442 int i; /* index of *z in zonelist zones */
1443 int n; /* node that zone *z is on */
1445 zlc = zonelist->zlcache_ptr;
1449 i = z - zonelist->_zonerefs;
1452 /* This zone is worth trying if it is allowed but not full */
1453 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1457 * Given 'z' scanning a zonelist, set the corresponding bit in
1458 * zlc->fullzones, so that subsequent attempts to allocate a page
1459 * from that zone don't waste time re-examining it.
1461 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1463 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1464 int i; /* index of *z in zonelist zones */
1466 zlc = zonelist->zlcache_ptr;
1470 i = z - zonelist->_zonerefs;
1472 set_bit(i, zlc->fullzones);
1475 #else /* CONFIG_NUMA */
1477 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1482 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1483 nodemask_t *allowednodes)
1488 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1491 #endif /* CONFIG_NUMA */
1494 * get_page_from_freelist goes through the zonelist trying to allocate
1497 static struct page *
1498 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1499 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1500 struct zone *preferred_zone, int migratetype)
1503 struct page *page = NULL;
1506 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1507 int zlc_active = 0; /* set if using zonelist_cache */
1508 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1510 classzone_idx = zone_idx(preferred_zone);
1513 * Scan zonelist, looking for a zone with enough free.
1514 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1516 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1517 high_zoneidx, nodemask) {
1518 if (NUMA_BUILD && zlc_active &&
1519 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1521 if ((alloc_flags & ALLOC_CPUSET) &&
1522 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1525 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1526 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1530 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1531 if (zone_watermark_ok(zone, order, mark,
1532 classzone_idx, alloc_flags))
1535 if (zone_reclaim_mode == 0)
1536 goto this_zone_full;
1538 ret = zone_reclaim(zone, gfp_mask, order);
1540 case ZONE_RECLAIM_NOSCAN:
1543 case ZONE_RECLAIM_FULL:
1544 /* scanned but unreclaimable */
1545 goto this_zone_full;
1547 /* did we reclaim enough */
1548 if (!zone_watermark_ok(zone, order, mark,
1549 classzone_idx, alloc_flags))
1550 goto this_zone_full;
1555 page = buffered_rmqueue(preferred_zone, zone, order,
1556 gfp_mask, migratetype);
1561 zlc_mark_zone_full(zonelist, z);
1563 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1565 * we do zlc_setup after the first zone is tried but only
1566 * if there are multiple nodes make it worthwhile
1568 allowednodes = zlc_setup(zonelist, alloc_flags);
1574 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1575 /* Disable zlc cache for second zonelist scan */
1583 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1584 unsigned long pages_reclaimed)
1586 /* Do not loop if specifically requested */
1587 if (gfp_mask & __GFP_NORETRY)
1591 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1592 * means __GFP_NOFAIL, but that may not be true in other
1595 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1599 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1600 * specified, then we retry until we no longer reclaim any pages
1601 * (above), or we've reclaimed an order of pages at least as
1602 * large as the allocation's order. In both cases, if the
1603 * allocation still fails, we stop retrying.
1605 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1609 * Don't let big-order allocations loop unless the caller
1610 * explicitly requests that.
1612 if (gfp_mask & __GFP_NOFAIL)
1618 static inline struct page *
1619 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1620 struct zonelist *zonelist, enum zone_type high_zoneidx,
1621 nodemask_t *nodemask, struct zone *preferred_zone,
1626 /* Acquire the OOM killer lock for the zones in zonelist */
1627 if (!try_set_zone_oom(zonelist, gfp_mask)) {
1628 schedule_timeout_uninterruptible(1);
1633 * Go through the zonelist yet one more time, keep very high watermark
1634 * here, this is only to catch a parallel oom killing, we must fail if
1635 * we're still under heavy pressure.
1637 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1638 order, zonelist, high_zoneidx,
1639 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1640 preferred_zone, migratetype);
1644 /* The OOM killer will not help higher order allocs */
1645 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_NOFAIL))
1648 /* Exhausted what can be done so it's blamo time */
1649 out_of_memory(zonelist, gfp_mask, order);
1652 clear_zonelist_oom(zonelist, gfp_mask);
1656 /* The really slow allocator path where we enter direct reclaim */
1657 static inline struct page *
1658 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1659 struct zonelist *zonelist, enum zone_type high_zoneidx,
1660 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1661 int migratetype, unsigned long *did_some_progress)
1663 struct page *page = NULL;
1664 struct reclaim_state reclaim_state;
1665 struct task_struct *p = current;
1669 /* We now go into synchronous reclaim */
1670 cpuset_memory_pressure_bump();
1671 p->flags |= PF_MEMALLOC;
1672 lockdep_set_current_reclaim_state(gfp_mask);
1673 reclaim_state.reclaimed_slab = 0;
1674 p->reclaim_state = &reclaim_state;
1676 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1678 p->reclaim_state = NULL;
1679 lockdep_clear_current_reclaim_state();
1680 p->flags &= ~PF_MEMALLOC;
1687 if (likely(*did_some_progress))
1688 page = get_page_from_freelist(gfp_mask, nodemask, order,
1689 zonelist, high_zoneidx,
1690 alloc_flags, preferred_zone,
1696 * This is called in the allocator slow-path if the allocation request is of
1697 * sufficient urgency to ignore watermarks and take other desperate measures
1699 static inline struct page *
1700 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1701 struct zonelist *zonelist, enum zone_type high_zoneidx,
1702 nodemask_t *nodemask, struct zone *preferred_zone,
1708 page = get_page_from_freelist(gfp_mask, nodemask, order,
1709 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
1710 preferred_zone, migratetype);
1712 if (!page && gfp_mask & __GFP_NOFAIL)
1713 congestion_wait(BLK_RW_ASYNC, HZ/50);
1714 } while (!page && (gfp_mask & __GFP_NOFAIL));
1720 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
1721 enum zone_type high_zoneidx)
1726 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1727 wakeup_kswapd(zone, order);
1731 gfp_to_alloc_flags(gfp_t gfp_mask)
1733 struct task_struct *p = current;
1734 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
1735 const gfp_t wait = gfp_mask & __GFP_WAIT;
1737 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1738 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH);
1741 * The caller may dip into page reserves a bit more if the caller
1742 * cannot run direct reclaim, or if the caller has realtime scheduling
1743 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1744 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1746 alloc_flags |= (gfp_mask & __GFP_HIGH);
1749 alloc_flags |= ALLOC_HARDER;
1751 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1752 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1754 alloc_flags &= ~ALLOC_CPUSET;
1755 } else if (unlikely(rt_task(p)))
1756 alloc_flags |= ALLOC_HARDER;
1758 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
1759 if (!in_interrupt() &&
1760 ((p->flags & PF_MEMALLOC) ||
1761 unlikely(test_thread_flag(TIF_MEMDIE))))
1762 alloc_flags |= ALLOC_NO_WATERMARKS;
1768 static inline struct page *
1769 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
1770 struct zonelist *zonelist, enum zone_type high_zoneidx,
1771 nodemask_t *nodemask, struct zone *preferred_zone,
1774 const gfp_t wait = gfp_mask & __GFP_WAIT;
1775 struct page *page = NULL;
1777 unsigned long pages_reclaimed = 0;
1778 unsigned long did_some_progress;
1779 struct task_struct *p = current;
1782 * In the slowpath, we sanity check order to avoid ever trying to
1783 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1784 * be using allocators in order of preference for an area that is
1787 if (order >= MAX_ORDER) {
1788 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
1793 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1794 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1795 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1796 * using a larger set of nodes after it has established that the
1797 * allowed per node queues are empty and that nodes are
1800 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1803 wake_all_kswapd(order, zonelist, high_zoneidx);
1807 * OK, we're below the kswapd watermark and have kicked background
1808 * reclaim. Now things get more complex, so set up alloc_flags according
1809 * to how we want to proceed.
1811 alloc_flags = gfp_to_alloc_flags(gfp_mask);
1813 /* This is the last chance, in general, before the goto nopage. */
1814 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
1815 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
1816 preferred_zone, migratetype);
1821 /* Allocate without watermarks if the context allows */
1822 if (alloc_flags & ALLOC_NO_WATERMARKS) {
1823 page = __alloc_pages_high_priority(gfp_mask, order,
1824 zonelist, high_zoneidx, nodemask,
1825 preferred_zone, migratetype);
1830 /* Atomic allocations - we can't balance anything */
1834 /* Avoid recursion of direct reclaim */
1835 if (p->flags & PF_MEMALLOC)
1838 /* Avoid allocations with no watermarks from looping endlessly */
1839 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
1842 /* Try direct reclaim and then allocating */
1843 page = __alloc_pages_direct_reclaim(gfp_mask, order,
1844 zonelist, high_zoneidx,
1846 alloc_flags, preferred_zone,
1847 migratetype, &did_some_progress);
1852 * If we failed to make any progress reclaiming, then we are
1853 * running out of options and have to consider going OOM
1855 if (!did_some_progress) {
1856 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1857 if (oom_killer_disabled)
1859 page = __alloc_pages_may_oom(gfp_mask, order,
1860 zonelist, high_zoneidx,
1861 nodemask, preferred_zone,
1867 * The OOM killer does not trigger for high-order
1868 * ~__GFP_NOFAIL allocations so if no progress is being
1869 * made, there are no other options and retrying is
1872 if (order > PAGE_ALLOC_COSTLY_ORDER &&
1873 !(gfp_mask & __GFP_NOFAIL))
1880 /* Check if we should retry the allocation */
1881 pages_reclaimed += did_some_progress;
1882 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
1883 /* Wait for some write requests to complete then retry */
1884 congestion_wait(BLK_RW_ASYNC, HZ/50);
1889 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1890 printk(KERN_WARNING "%s: page allocation failure."
1891 " order:%d, mode:0x%x\n",
1892 p->comm, order, gfp_mask);
1898 if (kmemcheck_enabled)
1899 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
1905 * This is the 'heart' of the zoned buddy allocator.
1908 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
1909 struct zonelist *zonelist, nodemask_t *nodemask)
1911 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1912 struct zone *preferred_zone;
1914 int migratetype = allocflags_to_migratetype(gfp_mask);
1916 gfp_mask &= gfp_allowed_mask;
1918 lockdep_trace_alloc(gfp_mask);
1920 might_sleep_if(gfp_mask & __GFP_WAIT);
1922 if (should_fail_alloc_page(gfp_mask, order))
1926 * Check the zones suitable for the gfp_mask contain at least one
1927 * valid zone. It's possible to have an empty zonelist as a result
1928 * of GFP_THISNODE and a memoryless node
1930 if (unlikely(!zonelist->_zonerefs->zone))
1933 /* The preferred zone is used for statistics later */
1934 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone);
1935 if (!preferred_zone)
1938 /* First allocation attempt */
1939 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
1940 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
1941 preferred_zone, migratetype);
1942 if (unlikely(!page))
1943 page = __alloc_pages_slowpath(gfp_mask, order,
1944 zonelist, high_zoneidx, nodemask,
1945 preferred_zone, migratetype);
1947 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
1950 EXPORT_SYMBOL(__alloc_pages_nodemask);
1953 * Common helper functions.
1955 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1960 * __get_free_pages() returns a 32-bit address, which cannot represent
1963 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1965 page = alloc_pages(gfp_mask, order);
1968 return (unsigned long) page_address(page);
1970 EXPORT_SYMBOL(__get_free_pages);
1972 unsigned long get_zeroed_page(gfp_t gfp_mask)
1974 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
1976 EXPORT_SYMBOL(get_zeroed_page);
1978 void __pagevec_free(struct pagevec *pvec)
1980 int i = pagevec_count(pvec);
1983 trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
1984 free_hot_cold_page(pvec->pages[i], pvec->cold);
1988 void __free_pages(struct page *page, unsigned int order)
1990 if (put_page_testzero(page)) {
1991 trace_mm_page_free_direct(page, order);
1993 free_hot_page(page);
1995 __free_pages_ok(page, order);
1999 EXPORT_SYMBOL(__free_pages);
2001 void free_pages(unsigned long addr, unsigned int order)
2004 VM_BUG_ON(!virt_addr_valid((void *)addr));
2005 __free_pages(virt_to_page((void *)addr), order);
2009 EXPORT_SYMBOL(free_pages);
2012 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2013 * @size: the number of bytes to allocate
2014 * @gfp_mask: GFP flags for the allocation
2016 * This function is similar to alloc_pages(), except that it allocates the
2017 * minimum number of pages to satisfy the request. alloc_pages() can only
2018 * allocate memory in power-of-two pages.
2020 * This function is also limited by MAX_ORDER.
2022 * Memory allocated by this function must be released by free_pages_exact().
2024 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2026 unsigned int order = get_order(size);
2029 addr = __get_free_pages(gfp_mask, order);
2031 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2032 unsigned long used = addr + PAGE_ALIGN(size);
2034 split_page(virt_to_page((void *)addr), order);
2035 while (used < alloc_end) {
2041 return (void *)addr;
2043 EXPORT_SYMBOL(alloc_pages_exact);
2046 * free_pages_exact - release memory allocated via alloc_pages_exact()
2047 * @virt: the value returned by alloc_pages_exact.
2048 * @size: size of allocation, same value as passed to alloc_pages_exact().
2050 * Release the memory allocated by a previous call to alloc_pages_exact.
2052 void free_pages_exact(void *virt, size_t size)
2054 unsigned long addr = (unsigned long)virt;
2055 unsigned long end = addr + PAGE_ALIGN(size);
2057 while (addr < end) {
2062 EXPORT_SYMBOL(free_pages_exact);
2064 static unsigned int nr_free_zone_pages(int offset)
2069 /* Just pick one node, since fallback list is circular */
2070 unsigned int sum = 0;
2072 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2074 for_each_zone_zonelist(zone, z, zonelist, offset) {
2075 unsigned long size = zone->present_pages;
2076 unsigned long high = high_wmark_pages(zone);
2085 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2087 unsigned int nr_free_buffer_pages(void)
2089 return nr_free_zone_pages(gfp_zone(GFP_USER));
2091 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2094 * Amount of free RAM allocatable within all zones
2096 unsigned int nr_free_pagecache_pages(void)
2098 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2101 static inline void show_node(struct zone *zone)
2104 printk("Node %d ", zone_to_nid(zone));
2107 void si_meminfo(struct sysinfo *val)
2109 val->totalram = totalram_pages;
2111 val->freeram = global_page_state(NR_FREE_PAGES);
2112 val->bufferram = nr_blockdev_pages();
2113 val->totalhigh = totalhigh_pages;
2114 val->freehigh = nr_free_highpages();
2115 val->mem_unit = PAGE_SIZE;
2118 EXPORT_SYMBOL(si_meminfo);
2121 void si_meminfo_node(struct sysinfo *val, int nid)
2123 pg_data_t *pgdat = NODE_DATA(nid);
2125 val->totalram = pgdat->node_present_pages;
2126 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2127 #ifdef CONFIG_HIGHMEM
2128 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2129 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2135 val->mem_unit = PAGE_SIZE;
2139 #define K(x) ((x) << (PAGE_SHIFT-10))
2142 * Show free area list (used inside shift_scroll-lock stuff)
2143 * We also calculate the percentage fragmentation. We do this by counting the
2144 * memory on each free list with the exception of the first item on the list.
2146 void show_free_areas(void)
2151 for_each_populated_zone(zone) {
2153 printk("%s per-cpu:\n", zone->name);
2155 for_each_online_cpu(cpu) {
2156 struct per_cpu_pageset *pageset;
2158 pageset = zone_pcp(zone, cpu);
2160 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2161 cpu, pageset->pcp.high,
2162 pageset->pcp.batch, pageset->pcp.count);
2166 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2167 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2169 " dirty:%lu writeback:%lu unstable:%lu buffer:%lu\n"
2170 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2171 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2172 global_page_state(NR_ACTIVE_ANON),
2173 global_page_state(NR_INACTIVE_ANON),
2174 global_page_state(NR_ISOLATED_ANON),
2175 global_page_state(NR_ACTIVE_FILE),
2176 global_page_state(NR_INACTIVE_FILE),
2177 global_page_state(NR_ISOLATED_FILE),
2178 global_page_state(NR_UNEVICTABLE),
2179 global_page_state(NR_FILE_DIRTY),
2180 global_page_state(NR_WRITEBACK),
2181 global_page_state(NR_UNSTABLE_NFS),
2182 nr_blockdev_pages(),
2183 global_page_state(NR_FREE_PAGES),
2184 global_page_state(NR_SLAB_RECLAIMABLE),
2185 global_page_state(NR_SLAB_UNRECLAIMABLE),
2186 global_page_state(NR_FILE_MAPPED),
2187 global_page_state(NR_SHMEM),
2188 global_page_state(NR_PAGETABLE),
2189 global_page_state(NR_BOUNCE));
2191 for_each_populated_zone(zone) {
2200 " active_anon:%lukB"
2201 " inactive_anon:%lukB"
2202 " active_file:%lukB"
2203 " inactive_file:%lukB"
2204 " unevictable:%lukB"
2205 " isolated(anon):%lukB"
2206 " isolated(file):%lukB"
2213 " slab_reclaimable:%lukB"
2214 " slab_unreclaimable:%lukB"
2215 " kernel_stack:%lukB"
2219 " writeback_tmp:%lukB"
2220 " pages_scanned:%lu"
2221 " all_unreclaimable? %s"
2224 K(zone_page_state(zone, NR_FREE_PAGES)),
2225 K(min_wmark_pages(zone)),
2226 K(low_wmark_pages(zone)),
2227 K(high_wmark_pages(zone)),
2228 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2229 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2230 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2231 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2232 K(zone_page_state(zone, NR_UNEVICTABLE)),
2233 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2234 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2235 K(zone->present_pages),
2236 K(zone_page_state(zone, NR_MLOCK)),
2237 K(zone_page_state(zone, NR_FILE_DIRTY)),
2238 K(zone_page_state(zone, NR_WRITEBACK)),
2239 K(zone_page_state(zone, NR_FILE_MAPPED)),
2240 K(zone_page_state(zone, NR_SHMEM)),
2241 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2242 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2243 zone_page_state(zone, NR_KERNEL_STACK) *
2245 K(zone_page_state(zone, NR_PAGETABLE)),
2246 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2247 K(zone_page_state(zone, NR_BOUNCE)),
2248 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2249 zone->pages_scanned,
2250 (zone_is_all_unreclaimable(zone) ? "yes" : "no")
2252 printk("lowmem_reserve[]:");
2253 for (i = 0; i < MAX_NR_ZONES; i++)
2254 printk(" %lu", zone->lowmem_reserve[i]);
2258 for_each_populated_zone(zone) {
2259 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2262 printk("%s: ", zone->name);
2264 spin_lock_irqsave(&zone->lock, flags);
2265 for (order = 0; order < MAX_ORDER; order++) {
2266 nr[order] = zone->free_area[order].nr_free;
2267 total += nr[order] << order;
2269 spin_unlock_irqrestore(&zone->lock, flags);
2270 for (order = 0; order < MAX_ORDER; order++)
2271 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2272 printk("= %lukB\n", K(total));
2275 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2277 show_swap_cache_info();
2280 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2282 zoneref->zone = zone;
2283 zoneref->zone_idx = zone_idx(zone);
2287 * Builds allocation fallback zone lists.
2289 * Add all populated zones of a node to the zonelist.
2291 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2292 int nr_zones, enum zone_type zone_type)
2296 BUG_ON(zone_type >= MAX_NR_ZONES);
2301 zone = pgdat->node_zones + zone_type;
2302 if (populated_zone(zone)) {
2303 zoneref_set_zone(zone,
2304 &zonelist->_zonerefs[nr_zones++]);
2305 check_highest_zone(zone_type);
2308 } while (zone_type);
2315 * 0 = automatic detection of better ordering.
2316 * 1 = order by ([node] distance, -zonetype)
2317 * 2 = order by (-zonetype, [node] distance)
2319 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2320 * the same zonelist. So only NUMA can configure this param.
2322 #define ZONELIST_ORDER_DEFAULT 0
2323 #define ZONELIST_ORDER_NODE 1
2324 #define ZONELIST_ORDER_ZONE 2
2326 /* zonelist order in the kernel.
2327 * set_zonelist_order() will set this to NODE or ZONE.
2329 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2330 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2334 /* The value user specified ....changed by config */
2335 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2336 /* string for sysctl */
2337 #define NUMA_ZONELIST_ORDER_LEN 16
2338 char numa_zonelist_order[16] = "default";
2341 * interface for configure zonelist ordering.
2342 * command line option "numa_zonelist_order"
2343 * = "[dD]efault - default, automatic configuration.
2344 * = "[nN]ode - order by node locality, then by zone within node
2345 * = "[zZ]one - order by zone, then by locality within zone
2348 static int __parse_numa_zonelist_order(char *s)
2350 if (*s == 'd' || *s == 'D') {
2351 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2352 } else if (*s == 'n' || *s == 'N') {
2353 user_zonelist_order = ZONELIST_ORDER_NODE;
2354 } else if (*s == 'z' || *s == 'Z') {
2355 user_zonelist_order = ZONELIST_ORDER_ZONE;
2358 "Ignoring invalid numa_zonelist_order value: "
2365 static __init int setup_numa_zonelist_order(char *s)
2368 return __parse_numa_zonelist_order(s);
2371 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2374 * sysctl handler for numa_zonelist_order
2376 int numa_zonelist_order_handler(ctl_table *table, int write,
2377 struct file *file, void __user *buffer, size_t *length,
2380 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2384 strncpy(saved_string, (char*)table->data,
2385 NUMA_ZONELIST_ORDER_LEN);
2386 ret = proc_dostring(table, write, file, buffer, length, ppos);
2390 int oldval = user_zonelist_order;
2391 if (__parse_numa_zonelist_order((char*)table->data)) {
2393 * bogus value. restore saved string
2395 strncpy((char*)table->data, saved_string,
2396 NUMA_ZONELIST_ORDER_LEN);
2397 user_zonelist_order = oldval;
2398 } else if (oldval != user_zonelist_order)
2399 build_all_zonelists();
2405 #define MAX_NODE_LOAD (nr_online_nodes)
2406 static int node_load[MAX_NUMNODES];
2409 * find_next_best_node - find the next node that should appear in a given node's fallback list
2410 * @node: node whose fallback list we're appending
2411 * @used_node_mask: nodemask_t of already used nodes
2413 * We use a number of factors to determine which is the next node that should
2414 * appear on a given node's fallback list. The node should not have appeared
2415 * already in @node's fallback list, and it should be the next closest node
2416 * according to the distance array (which contains arbitrary distance values
2417 * from each node to each node in the system), and should also prefer nodes
2418 * with no CPUs, since presumably they'll have very little allocation pressure
2419 * on them otherwise.
2420 * It returns -1 if no node is found.
2422 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2425 int min_val = INT_MAX;
2427 const struct cpumask *tmp = cpumask_of_node(0);
2429 /* Use the local node if we haven't already */
2430 if (!node_isset(node, *used_node_mask)) {
2431 node_set(node, *used_node_mask);
2435 for_each_node_state(n, N_HIGH_MEMORY) {
2437 /* Don't want a node to appear more than once */
2438 if (node_isset(n, *used_node_mask))
2441 /* Use the distance array to find the distance */
2442 val = node_distance(node, n);
2444 /* Penalize nodes under us ("prefer the next node") */
2447 /* Give preference to headless and unused nodes */
2448 tmp = cpumask_of_node(n);
2449 if (!cpumask_empty(tmp))
2450 val += PENALTY_FOR_NODE_WITH_CPUS;
2452 /* Slight preference for less loaded node */
2453 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2454 val += node_load[n];
2456 if (val < min_val) {
2463 node_set(best_node, *used_node_mask);
2470 * Build zonelists ordered by node and zones within node.
2471 * This results in maximum locality--normal zone overflows into local
2472 * DMA zone, if any--but risks exhausting DMA zone.
2474 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2477 struct zonelist *zonelist;
2479 zonelist = &pgdat->node_zonelists[0];
2480 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2482 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2484 zonelist->_zonerefs[j].zone = NULL;
2485 zonelist->_zonerefs[j].zone_idx = 0;
2489 * Build gfp_thisnode zonelists
2491 static void build_thisnode_zonelists(pg_data_t *pgdat)
2494 struct zonelist *zonelist;
2496 zonelist = &pgdat->node_zonelists[1];
2497 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2498 zonelist->_zonerefs[j].zone = NULL;
2499 zonelist->_zonerefs[j].zone_idx = 0;
2503 * Build zonelists ordered by zone and nodes within zones.
2504 * This results in conserving DMA zone[s] until all Normal memory is
2505 * exhausted, but results in overflowing to remote node while memory
2506 * may still exist in local DMA zone.
2508 static int node_order[MAX_NUMNODES];
2510 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2513 int zone_type; /* needs to be signed */
2515 struct zonelist *zonelist;
2517 zonelist = &pgdat->node_zonelists[0];
2519 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2520 for (j = 0; j < nr_nodes; j++) {
2521 node = node_order[j];
2522 z = &NODE_DATA(node)->node_zones[zone_type];
2523 if (populated_zone(z)) {
2525 &zonelist->_zonerefs[pos++]);
2526 check_highest_zone(zone_type);
2530 zonelist->_zonerefs[pos].zone = NULL;
2531 zonelist->_zonerefs[pos].zone_idx = 0;
2534 static int default_zonelist_order(void)
2537 unsigned long low_kmem_size,total_size;
2541 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2542 * If they are really small and used heavily, the system can fall
2543 * into OOM very easily.
2544 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2546 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2549 for_each_online_node(nid) {
2550 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2551 z = &NODE_DATA(nid)->node_zones[zone_type];
2552 if (populated_zone(z)) {
2553 if (zone_type < ZONE_NORMAL)
2554 low_kmem_size += z->present_pages;
2555 total_size += z->present_pages;
2559 if (!low_kmem_size || /* there are no DMA area. */
2560 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2561 return ZONELIST_ORDER_NODE;
2563 * look into each node's config.
2564 * If there is a node whose DMA/DMA32 memory is very big area on
2565 * local memory, NODE_ORDER may be suitable.
2567 average_size = total_size /
2568 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2569 for_each_online_node(nid) {
2572 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2573 z = &NODE_DATA(nid)->node_zones[zone_type];
2574 if (populated_zone(z)) {
2575 if (zone_type < ZONE_NORMAL)
2576 low_kmem_size += z->present_pages;
2577 total_size += z->present_pages;
2580 if (low_kmem_size &&
2581 total_size > average_size && /* ignore small node */
2582 low_kmem_size > total_size * 70/100)
2583 return ZONELIST_ORDER_NODE;
2585 return ZONELIST_ORDER_ZONE;
2588 static void set_zonelist_order(void)
2590 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2591 current_zonelist_order = default_zonelist_order();
2593 current_zonelist_order = user_zonelist_order;
2596 static void build_zonelists(pg_data_t *pgdat)
2600 nodemask_t used_mask;
2601 int local_node, prev_node;
2602 struct zonelist *zonelist;
2603 int order = current_zonelist_order;
2605 /* initialize zonelists */
2606 for (i = 0; i < MAX_ZONELISTS; i++) {
2607 zonelist = pgdat->node_zonelists + i;
2608 zonelist->_zonerefs[0].zone = NULL;
2609 zonelist->_zonerefs[0].zone_idx = 0;
2612 /* NUMA-aware ordering of nodes */
2613 local_node = pgdat->node_id;
2614 load = nr_online_nodes;
2615 prev_node = local_node;
2616 nodes_clear(used_mask);
2618 memset(node_order, 0, sizeof(node_order));
2621 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2622 int distance = node_distance(local_node, node);
2625 * If another node is sufficiently far away then it is better
2626 * to reclaim pages in a zone before going off node.
2628 if (distance > RECLAIM_DISTANCE)
2629 zone_reclaim_mode = 1;
2632 * We don't want to pressure a particular node.
2633 * So adding penalty to the first node in same
2634 * distance group to make it round-robin.
2636 if (distance != node_distance(local_node, prev_node))
2637 node_load[node] = load;
2641 if (order == ZONELIST_ORDER_NODE)
2642 build_zonelists_in_node_order(pgdat, node);
2644 node_order[j++] = node; /* remember order */
2647 if (order == ZONELIST_ORDER_ZONE) {
2648 /* calculate node order -- i.e., DMA last! */
2649 build_zonelists_in_zone_order(pgdat, j);
2652 build_thisnode_zonelists(pgdat);
2655 /* Construct the zonelist performance cache - see further mmzone.h */
2656 static void build_zonelist_cache(pg_data_t *pgdat)
2658 struct zonelist *zonelist;
2659 struct zonelist_cache *zlc;
2662 zonelist = &pgdat->node_zonelists[0];
2663 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2664 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2665 for (z = zonelist->_zonerefs; z->zone; z++)
2666 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2670 #else /* CONFIG_NUMA */
2672 static void set_zonelist_order(void)
2674 current_zonelist_order = ZONELIST_ORDER_ZONE;
2677 static void build_zonelists(pg_data_t *pgdat)
2679 int node, local_node;
2681 struct zonelist *zonelist;
2683 local_node = pgdat->node_id;
2685 zonelist = &pgdat->node_zonelists[0];
2686 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2689 * Now we build the zonelist so that it contains the zones
2690 * of all the other nodes.
2691 * We don't want to pressure a particular node, so when
2692 * building the zones for node N, we make sure that the
2693 * zones coming right after the local ones are those from
2694 * node N+1 (modulo N)
2696 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2697 if (!node_online(node))
2699 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2702 for (node = 0; node < local_node; node++) {
2703 if (!node_online(node))
2705 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2709 zonelist->_zonerefs[j].zone = NULL;
2710 zonelist->_zonerefs[j].zone_idx = 0;
2713 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2714 static void build_zonelist_cache(pg_data_t *pgdat)
2716 pgdat->node_zonelists[0].zlcache_ptr = NULL;
2719 #endif /* CONFIG_NUMA */
2721 /* return values int ....just for stop_machine() */
2722 static int __build_all_zonelists(void *dummy)
2727 memset(node_load, 0, sizeof(node_load));
2729 for_each_online_node(nid) {
2730 pg_data_t *pgdat = NODE_DATA(nid);
2732 build_zonelists(pgdat);
2733 build_zonelist_cache(pgdat);
2738 void build_all_zonelists(void)
2740 set_zonelist_order();
2742 if (system_state == SYSTEM_BOOTING) {
2743 __build_all_zonelists(NULL);
2744 mminit_verify_zonelist();
2745 cpuset_init_current_mems_allowed();
2747 /* we have to stop all cpus to guarantee there is no user
2749 stop_machine(__build_all_zonelists, NULL, NULL);
2750 /* cpuset refresh routine should be here */
2752 vm_total_pages = nr_free_pagecache_pages();
2754 * Disable grouping by mobility if the number of pages in the
2755 * system is too low to allow the mechanism to work. It would be
2756 * more accurate, but expensive to check per-zone. This check is
2757 * made on memory-hotadd so a system can start with mobility
2758 * disabled and enable it later
2760 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
2761 page_group_by_mobility_disabled = 1;
2763 page_group_by_mobility_disabled = 0;
2765 printk("Built %i zonelists in %s order, mobility grouping %s. "
2766 "Total pages: %ld\n",
2768 zonelist_order_name[current_zonelist_order],
2769 page_group_by_mobility_disabled ? "off" : "on",
2772 printk("Policy zone: %s\n", zone_names[policy_zone]);
2777 * Helper functions to size the waitqueue hash table.
2778 * Essentially these want to choose hash table sizes sufficiently
2779 * large so that collisions trying to wait on pages are rare.
2780 * But in fact, the number of active page waitqueues on typical
2781 * systems is ridiculously low, less than 200. So this is even
2782 * conservative, even though it seems large.
2784 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2785 * waitqueues, i.e. the size of the waitq table given the number of pages.
2787 #define PAGES_PER_WAITQUEUE 256
2789 #ifndef CONFIG_MEMORY_HOTPLUG
2790 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2792 unsigned long size = 1;
2794 pages /= PAGES_PER_WAITQUEUE;
2796 while (size < pages)
2800 * Once we have dozens or even hundreds of threads sleeping
2801 * on IO we've got bigger problems than wait queue collision.
2802 * Limit the size of the wait table to a reasonable size.
2804 size = min(size, 4096UL);
2806 return max(size, 4UL);
2810 * A zone's size might be changed by hot-add, so it is not possible to determine
2811 * a suitable size for its wait_table. So we use the maximum size now.
2813 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2815 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2816 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2817 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2819 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2820 * or more by the traditional way. (See above). It equals:
2822 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2823 * ia64(16K page size) : = ( 8G + 4M)byte.
2824 * powerpc (64K page size) : = (32G +16M)byte.
2826 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2833 * This is an integer logarithm so that shifts can be used later
2834 * to extract the more random high bits from the multiplicative
2835 * hash function before the remainder is taken.
2837 static inline unsigned long wait_table_bits(unsigned long size)
2842 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2845 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2846 * of blocks reserved is based on min_wmark_pages(zone). The memory within
2847 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
2848 * higher will lead to a bigger reserve which will get freed as contiguous
2849 * blocks as reclaim kicks in
2851 static void setup_zone_migrate_reserve(struct zone *zone)
2853 unsigned long start_pfn, pfn, end_pfn;
2855 unsigned long block_migratetype;
2858 /* Get the start pfn, end pfn and the number of blocks to reserve */
2859 start_pfn = zone->zone_start_pfn;
2860 end_pfn = start_pfn + zone->spanned_pages;
2861 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
2865 * Reserve blocks are generally in place to help high-order atomic
2866 * allocations that are short-lived. A min_free_kbytes value that
2867 * would result in more than 2 reserve blocks for atomic allocations
2868 * is assumed to be in place to help anti-fragmentation for the
2869 * future allocation of hugepages at runtime.
2871 reserve = min(2, reserve);
2873 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
2874 if (!pfn_valid(pfn))
2876 page = pfn_to_page(pfn);
2878 /* Watch out for overlapping nodes */
2879 if (page_to_nid(page) != zone_to_nid(zone))
2882 /* Blocks with reserved pages will never free, skip them. */
2883 if (PageReserved(page))
2886 block_migratetype = get_pageblock_migratetype(page);
2888 /* If this block is reserved, account for it */
2889 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
2894 /* Suitable for reserving if this block is movable */
2895 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
2896 set_pageblock_migratetype(page, MIGRATE_RESERVE);
2897 move_freepages_block(zone, page, MIGRATE_RESERVE);
2903 * If the reserve is met and this is a previous reserved block,
2906 if (block_migratetype == MIGRATE_RESERVE) {
2907 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2908 move_freepages_block(zone, page, MIGRATE_MOVABLE);
2914 * Initially all pages are reserved - free ones are freed
2915 * up by free_all_bootmem() once the early boot process is
2916 * done. Non-atomic initialization, single-pass.
2918 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
2919 unsigned long start_pfn, enum memmap_context context)
2922 unsigned long end_pfn = start_pfn + size;
2926 if (highest_memmap_pfn < end_pfn - 1)
2927 highest_memmap_pfn = end_pfn - 1;
2929 z = &NODE_DATA(nid)->node_zones[zone];
2930 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
2932 * There can be holes in boot-time mem_map[]s
2933 * handed to this function. They do not
2934 * exist on hotplugged memory.
2936 if (context == MEMMAP_EARLY) {
2937 if (!early_pfn_valid(pfn))
2939 if (!early_pfn_in_nid(pfn, nid))
2942 page = pfn_to_page(pfn);
2943 set_page_links(page, zone, nid, pfn);
2944 mminit_verify_page_links(page, zone, nid, pfn);
2945 init_page_count(page);
2946 reset_page_mapcount(page);
2947 SetPageReserved(page);
2949 * Mark the block movable so that blocks are reserved for
2950 * movable at startup. This will force kernel allocations
2951 * to reserve their blocks rather than leaking throughout
2952 * the address space during boot when many long-lived
2953 * kernel allocations are made. Later some blocks near
2954 * the start are marked MIGRATE_RESERVE by
2955 * setup_zone_migrate_reserve()
2957 * bitmap is created for zone's valid pfn range. but memmap
2958 * can be created for invalid pages (for alignment)
2959 * check here not to call set_pageblock_migratetype() against
2962 if ((z->zone_start_pfn <= pfn)
2963 && (pfn < z->zone_start_pfn + z->spanned_pages)
2964 && !(pfn & (pageblock_nr_pages - 1)))
2965 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2967 INIT_LIST_HEAD(&page->lru);
2968 #ifdef WANT_PAGE_VIRTUAL
2969 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2970 if (!is_highmem_idx(zone))
2971 set_page_address(page, __va(pfn << PAGE_SHIFT));
2976 static void __meminit zone_init_free_lists(struct zone *zone)
2979 for_each_migratetype_order(order, t) {
2980 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
2981 zone->free_area[order].nr_free = 0;
2985 #ifndef __HAVE_ARCH_MEMMAP_INIT
2986 #define memmap_init(size, nid, zone, start_pfn) \
2987 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
2990 static int zone_batchsize(struct zone *zone)
2996 * The per-cpu-pages pools are set to around 1000th of the
2997 * size of the zone. But no more than 1/2 of a meg.
2999 * OK, so we don't know how big the cache is. So guess.
3001 batch = zone->present_pages / 1024;
3002 if (batch * PAGE_SIZE > 512 * 1024)
3003 batch = (512 * 1024) / PAGE_SIZE;
3004 batch /= 4; /* We effectively *= 4 below */
3009 * Clamp the batch to a 2^n - 1 value. Having a power
3010 * of 2 value was found to be more likely to have
3011 * suboptimal cache aliasing properties in some cases.
3013 * For example if 2 tasks are alternately allocating
3014 * batches of pages, one task can end up with a lot
3015 * of pages of one half of the possible page colors
3016 * and the other with pages of the other colors.
3018 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3023 /* The deferral and batching of frees should be suppressed under NOMMU
3026 * The problem is that NOMMU needs to be able to allocate large chunks
3027 * of contiguous memory as there's no hardware page translation to
3028 * assemble apparent contiguous memory from discontiguous pages.
3030 * Queueing large contiguous runs of pages for batching, however,
3031 * causes the pages to actually be freed in smaller chunks. As there
3032 * can be a significant delay between the individual batches being
3033 * recycled, this leads to the once large chunks of space being
3034 * fragmented and becoming unavailable for high-order allocations.
3040 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3042 struct per_cpu_pages *pcp;
3045 memset(p, 0, sizeof(*p));
3049 pcp->high = 6 * batch;
3050 pcp->batch = max(1UL, 1 * batch);
3051 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3052 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3056 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3057 * to the value high for the pageset p.
3060 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3063 struct per_cpu_pages *pcp;
3067 pcp->batch = max(1UL, high/4);
3068 if ((high/4) > (PAGE_SHIFT * 8))
3069 pcp->batch = PAGE_SHIFT * 8;
3075 * Boot pageset table. One per cpu which is going to be used for all
3076 * zones and all nodes. The parameters will be set in such a way
3077 * that an item put on a list will immediately be handed over to
3078 * the buddy list. This is safe since pageset manipulation is done
3079 * with interrupts disabled.
3081 * Some NUMA counter updates may also be caught by the boot pagesets.
3083 * The boot_pagesets must be kept even after bootup is complete for
3084 * unused processors and/or zones. They do play a role for bootstrapping
3085 * hotplugged processors.
3087 * zoneinfo_show() and maybe other functions do
3088 * not check if the processor is online before following the pageset pointer.
3089 * Other parts of the kernel may not check if the zone is available.
3091 static struct per_cpu_pageset boot_pageset[NR_CPUS];
3094 * Dynamically allocate memory for the
3095 * per cpu pageset array in struct zone.
3097 static int __cpuinit process_zones(int cpu)
3099 struct zone *zone, *dzone;
3100 int node = cpu_to_node(cpu);
3102 node_set_state(node, N_CPU); /* this node has a cpu */
3104 for_each_populated_zone(zone) {
3105 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
3107 if (!zone_pcp(zone, cpu))
3110 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
3112 if (percpu_pagelist_fraction)
3113 setup_pagelist_highmark(zone_pcp(zone, cpu),
3114 (zone->present_pages / percpu_pagelist_fraction));
3119 for_each_zone(dzone) {
3120 if (!populated_zone(dzone))
3124 kfree(zone_pcp(dzone, cpu));
3125 zone_pcp(dzone, cpu) = &boot_pageset[cpu];
3130 static inline void free_zone_pagesets(int cpu)
3134 for_each_zone(zone) {
3135 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
3137 /* Free per_cpu_pageset if it is slab allocated */
3138 if (pset != &boot_pageset[cpu])
3140 zone_pcp(zone, cpu) = &boot_pageset[cpu];
3144 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
3145 unsigned long action,
3148 int cpu = (long)hcpu;
3149 int ret = NOTIFY_OK;
3152 case CPU_UP_PREPARE:
3153 case CPU_UP_PREPARE_FROZEN:
3154 if (process_zones(cpu))
3157 case CPU_UP_CANCELED:
3158 case CPU_UP_CANCELED_FROZEN:
3160 case CPU_DEAD_FROZEN:
3161 free_zone_pagesets(cpu);
3169 static struct notifier_block __cpuinitdata pageset_notifier =
3170 { &pageset_cpuup_callback, NULL, 0 };
3172 void __init setup_per_cpu_pageset(void)
3176 /* Initialize per_cpu_pageset for cpu 0.
3177 * A cpuup callback will do this for every cpu
3178 * as it comes online
3180 err = process_zones(smp_processor_id());
3182 register_cpu_notifier(&pageset_notifier);
3187 static noinline __init_refok
3188 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3191 struct pglist_data *pgdat = zone->zone_pgdat;
3195 * The per-page waitqueue mechanism uses hashed waitqueues
3198 zone->wait_table_hash_nr_entries =
3199 wait_table_hash_nr_entries(zone_size_pages);
3200 zone->wait_table_bits =
3201 wait_table_bits(zone->wait_table_hash_nr_entries);
3202 alloc_size = zone->wait_table_hash_nr_entries
3203 * sizeof(wait_queue_head_t);
3205 if (!slab_is_available()) {
3206 zone->wait_table = (wait_queue_head_t *)
3207 alloc_bootmem_node(pgdat, alloc_size);
3210 * This case means that a zone whose size was 0 gets new memory
3211 * via memory hot-add.
3212 * But it may be the case that a new node was hot-added. In
3213 * this case vmalloc() will not be able to use this new node's
3214 * memory - this wait_table must be initialized to use this new
3215 * node itself as well.
3216 * To use this new node's memory, further consideration will be
3219 zone->wait_table = vmalloc(alloc_size);
3221 if (!zone->wait_table)
3224 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3225 init_waitqueue_head(zone->wait_table + i);
3230 static int __zone_pcp_update(void *data)
3232 struct zone *zone = data;
3234 unsigned long batch = zone_batchsize(zone), flags;
3236 for (cpu = 0; cpu < NR_CPUS; cpu++) {
3237 struct per_cpu_pageset *pset;
3238 struct per_cpu_pages *pcp;
3240 pset = zone_pcp(zone, cpu);
3243 local_irq_save(flags);
3244 free_pcppages_bulk(zone, pcp->count, pcp);
3245 setup_pageset(pset, batch);
3246 local_irq_restore(flags);
3251 void zone_pcp_update(struct zone *zone)
3253 stop_machine(__zone_pcp_update, zone, NULL);
3256 static __meminit void zone_pcp_init(struct zone *zone)
3259 unsigned long batch = zone_batchsize(zone);
3261 for (cpu = 0; cpu < NR_CPUS; cpu++) {
3263 /* Early boot. Slab allocator not functional yet */
3264 zone_pcp(zone, cpu) = &boot_pageset[cpu];
3265 setup_pageset(&boot_pageset[cpu],0);
3267 setup_pageset(zone_pcp(zone,cpu), batch);
3270 if (zone->present_pages)
3271 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
3272 zone->name, zone->present_pages, batch);
3275 __meminit int init_currently_empty_zone(struct zone *zone,
3276 unsigned long zone_start_pfn,
3278 enum memmap_context context)
3280 struct pglist_data *pgdat = zone->zone_pgdat;
3282 ret = zone_wait_table_init(zone, size);
3285 pgdat->nr_zones = zone_idx(zone) + 1;
3287 zone->zone_start_pfn = zone_start_pfn;
3289 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3290 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3292 (unsigned long)zone_idx(zone),
3293 zone_start_pfn, (zone_start_pfn + size));
3295 zone_init_free_lists(zone);
3300 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3302 * Basic iterator support. Return the first range of PFNs for a node
3303 * Note: nid == MAX_NUMNODES returns first region regardless of node
3305 static int __meminit first_active_region_index_in_nid(int nid)
3309 for (i = 0; i < nr_nodemap_entries; i++)
3310 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3317 * Basic iterator support. Return the next active range of PFNs for a node
3318 * Note: nid == MAX_NUMNODES returns next region regardless of node
3320 static int __meminit next_active_region_index_in_nid(int index, int nid)
3322 for (index = index + 1; index < nr_nodemap_entries; index++)
3323 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3329 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3331 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3332 * Architectures may implement their own version but if add_active_range()
3333 * was used and there are no special requirements, this is a convenient
3336 int __meminit __early_pfn_to_nid(unsigned long pfn)
3340 for (i = 0; i < nr_nodemap_entries; i++) {
3341 unsigned long start_pfn = early_node_map[i].start_pfn;
3342 unsigned long end_pfn = early_node_map[i].end_pfn;
3344 if (start_pfn <= pfn && pfn < end_pfn)
3345 return early_node_map[i].nid;
3347 /* This is a memory hole */
3350 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3352 int __meminit early_pfn_to_nid(unsigned long pfn)
3356 nid = __early_pfn_to_nid(pfn);
3359 /* just returns 0 */
3363 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3364 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3368 nid = __early_pfn_to_nid(pfn);
3369 if (nid >= 0 && nid != node)
3375 /* Basic iterator support to walk early_node_map[] */
3376 #define for_each_active_range_index_in_nid(i, nid) \
3377 for (i = first_active_region_index_in_nid(nid); i != -1; \
3378 i = next_active_region_index_in_nid(i, nid))
3381 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3382 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3383 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3385 * If an architecture guarantees that all ranges registered with
3386 * add_active_ranges() contain no holes and may be freed, this
3387 * this function may be used instead of calling free_bootmem() manually.
3389 void __init free_bootmem_with_active_regions(int nid,
3390 unsigned long max_low_pfn)
3394 for_each_active_range_index_in_nid(i, nid) {
3395 unsigned long size_pages = 0;
3396 unsigned long end_pfn = early_node_map[i].end_pfn;
3398 if (early_node_map[i].start_pfn >= max_low_pfn)
3401 if (end_pfn > max_low_pfn)
3402 end_pfn = max_low_pfn;
3404 size_pages = end_pfn - early_node_map[i].start_pfn;
3405 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3406 PFN_PHYS(early_node_map[i].start_pfn),
3407 size_pages << PAGE_SHIFT);
3411 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3416 for_each_active_range_index_in_nid(i, nid) {
3417 ret = work_fn(early_node_map[i].start_pfn,
3418 early_node_map[i].end_pfn, data);
3424 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3425 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3427 * If an architecture guarantees that all ranges registered with
3428 * add_active_ranges() contain no holes and may be freed, this
3429 * function may be used instead of calling memory_present() manually.
3431 void __init sparse_memory_present_with_active_regions(int nid)
3435 for_each_active_range_index_in_nid(i, nid)
3436 memory_present(early_node_map[i].nid,
3437 early_node_map[i].start_pfn,
3438 early_node_map[i].end_pfn);
3442 * get_pfn_range_for_nid - Return the start and end page frames for a node
3443 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3444 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3445 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3447 * It returns the start and end page frame of a node based on information
3448 * provided by an arch calling add_active_range(). If called for a node
3449 * with no available memory, a warning is printed and the start and end
3452 void __meminit get_pfn_range_for_nid(unsigned int nid,
3453 unsigned long *start_pfn, unsigned long *end_pfn)
3459 for_each_active_range_index_in_nid(i, nid) {
3460 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3461 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3464 if (*start_pfn == -1UL)
3469 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3470 * assumption is made that zones within a node are ordered in monotonic
3471 * increasing memory addresses so that the "highest" populated zone is used
3473 static void __init find_usable_zone_for_movable(void)
3476 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3477 if (zone_index == ZONE_MOVABLE)
3480 if (arch_zone_highest_possible_pfn[zone_index] >
3481 arch_zone_lowest_possible_pfn[zone_index])
3485 VM_BUG_ON(zone_index == -1);
3486 movable_zone = zone_index;
3490 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3491 * because it is sized independant of architecture. Unlike the other zones,
3492 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3493 * in each node depending on the size of each node and how evenly kernelcore
3494 * is distributed. This helper function adjusts the zone ranges
3495 * provided by the architecture for a given node by using the end of the
3496 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3497 * zones within a node are in order of monotonic increases memory addresses
3499 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3500 unsigned long zone_type,
3501 unsigned long node_start_pfn,
3502 unsigned long node_end_pfn,
3503 unsigned long *zone_start_pfn,
3504 unsigned long *zone_end_pfn)
3506 /* Only adjust if ZONE_MOVABLE is on this node */
3507 if (zone_movable_pfn[nid]) {
3508 /* Size ZONE_MOVABLE */
3509 if (zone_type == ZONE_MOVABLE) {
3510 *zone_start_pfn = zone_movable_pfn[nid];
3511 *zone_end_pfn = min(node_end_pfn,
3512 arch_zone_highest_possible_pfn[movable_zone]);
3514 /* Adjust for ZONE_MOVABLE starting within this range */
3515 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3516 *zone_end_pfn > zone_movable_pfn[nid]) {
3517 *zone_end_pfn = zone_movable_pfn[nid];
3519 /* Check if this whole range is within ZONE_MOVABLE */
3520 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3521 *zone_start_pfn = *zone_end_pfn;
3526 * Return the number of pages a zone spans in a node, including holes
3527 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3529 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3530 unsigned long zone_type,
3531 unsigned long *ignored)
3533 unsigned long node_start_pfn, node_end_pfn;
3534 unsigned long zone_start_pfn, zone_end_pfn;
3536 /* Get the start and end of the node and zone */
3537 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3538 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3539 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3540 adjust_zone_range_for_zone_movable(nid, zone_type,
3541 node_start_pfn, node_end_pfn,
3542 &zone_start_pfn, &zone_end_pfn);
3544 /* Check that this node has pages within the zone's required range */
3545 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3548 /* Move the zone boundaries inside the node if necessary */
3549 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3550 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3552 /* Return the spanned pages */
3553 return zone_end_pfn - zone_start_pfn;
3557 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3558 * then all holes in the requested range will be accounted for.
3560 static unsigned long __meminit __absent_pages_in_range(int nid,
3561 unsigned long range_start_pfn,
3562 unsigned long range_end_pfn)
3565 unsigned long prev_end_pfn = 0, hole_pages = 0;
3566 unsigned long start_pfn;
3568 /* Find the end_pfn of the first active range of pfns in the node */
3569 i = first_active_region_index_in_nid(nid);
3573 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3575 /* Account for ranges before physical memory on this node */
3576 if (early_node_map[i].start_pfn > range_start_pfn)
3577 hole_pages = prev_end_pfn - range_start_pfn;
3579 /* Find all holes for the zone within the node */
3580 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3582 /* No need to continue if prev_end_pfn is outside the zone */
3583 if (prev_end_pfn >= range_end_pfn)
3586 /* Make sure the end of the zone is not within the hole */
3587 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3588 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3590 /* Update the hole size cound and move on */
3591 if (start_pfn > range_start_pfn) {
3592 BUG_ON(prev_end_pfn > start_pfn);
3593 hole_pages += start_pfn - prev_end_pfn;
3595 prev_end_pfn = early_node_map[i].end_pfn;
3598 /* Account for ranges past physical memory on this node */
3599 if (range_end_pfn > prev_end_pfn)
3600 hole_pages += range_end_pfn -
3601 max(range_start_pfn, prev_end_pfn);
3607 * absent_pages_in_range - Return number of page frames in holes within a range
3608 * @start_pfn: The start PFN to start searching for holes
3609 * @end_pfn: The end PFN to stop searching for holes
3611 * It returns the number of pages frames in memory holes within a range.
3613 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3614 unsigned long end_pfn)
3616 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3619 /* Return the number of page frames in holes in a zone on a node */
3620 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3621 unsigned long zone_type,
3622 unsigned long *ignored)
3624 unsigned long node_start_pfn, node_end_pfn;
3625 unsigned long zone_start_pfn, zone_end_pfn;
3627 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3628 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3630 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3633 adjust_zone_range_for_zone_movable(nid, zone_type,
3634 node_start_pfn, node_end_pfn,
3635 &zone_start_pfn, &zone_end_pfn);
3636 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3640 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3641 unsigned long zone_type,
3642 unsigned long *zones_size)
3644 return zones_size[zone_type];
3647 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3648 unsigned long zone_type,
3649 unsigned long *zholes_size)
3654 return zholes_size[zone_type];
3659 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3660 unsigned long *zones_size, unsigned long *zholes_size)
3662 unsigned long realtotalpages, totalpages = 0;
3665 for (i = 0; i < MAX_NR_ZONES; i++)
3666 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3668 pgdat->node_spanned_pages = totalpages;
3670 realtotalpages = totalpages;
3671 for (i = 0; i < MAX_NR_ZONES; i++)
3673 zone_absent_pages_in_node(pgdat->node_id, i,
3675 pgdat->node_present_pages = realtotalpages;
3676 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3680 #ifndef CONFIG_SPARSEMEM
3682 * Calculate the size of the zone->blockflags rounded to an unsigned long
3683 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3684 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3685 * round what is now in bits to nearest long in bits, then return it in
3688 static unsigned long __init usemap_size(unsigned long zonesize)
3690 unsigned long usemapsize;
3692 usemapsize = roundup(zonesize, pageblock_nr_pages);
3693 usemapsize = usemapsize >> pageblock_order;
3694 usemapsize *= NR_PAGEBLOCK_BITS;
3695 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
3697 return usemapsize / 8;
3700 static void __init setup_usemap(struct pglist_data *pgdat,
3701 struct zone *zone, unsigned long zonesize)
3703 unsigned long usemapsize = usemap_size(zonesize);
3704 zone->pageblock_flags = NULL;
3706 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
3709 static void inline setup_usemap(struct pglist_data *pgdat,
3710 struct zone *zone, unsigned long zonesize) {}
3711 #endif /* CONFIG_SPARSEMEM */
3713 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3715 /* Return a sensible default order for the pageblock size. */
3716 static inline int pageblock_default_order(void)
3718 if (HPAGE_SHIFT > PAGE_SHIFT)
3719 return HUGETLB_PAGE_ORDER;
3724 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3725 static inline void __init set_pageblock_order(unsigned int order)
3727 /* Check that pageblock_nr_pages has not already been setup */
3728 if (pageblock_order)
3732 * Assume the largest contiguous order of interest is a huge page.
3733 * This value may be variable depending on boot parameters on IA64
3735 pageblock_order = order;
3737 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3740 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3741 * and pageblock_default_order() are unused as pageblock_order is set
3742 * at compile-time. See include/linux/pageblock-flags.h for the values of
3743 * pageblock_order based on the kernel config
3745 static inline int pageblock_default_order(unsigned int order)
3749 #define set_pageblock_order(x) do {} while (0)
3751 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3754 * Set up the zone data structures:
3755 * - mark all pages reserved
3756 * - mark all memory queues empty
3757 * - clear the memory bitmaps
3759 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
3760 unsigned long *zones_size, unsigned long *zholes_size)
3763 int nid = pgdat->node_id;
3764 unsigned long zone_start_pfn = pgdat->node_start_pfn;
3767 pgdat_resize_init(pgdat);
3768 pgdat->nr_zones = 0;
3769 init_waitqueue_head(&pgdat->kswapd_wait);
3770 pgdat->kswapd_max_order = 0;
3771 pgdat_page_cgroup_init(pgdat);
3773 for (j = 0; j < MAX_NR_ZONES; j++) {
3774 struct zone *zone = pgdat->node_zones + j;
3775 unsigned long size, realsize, memmap_pages;
3778 size = zone_spanned_pages_in_node(nid, j, zones_size);
3779 realsize = size - zone_absent_pages_in_node(nid, j,
3783 * Adjust realsize so that it accounts for how much memory
3784 * is used by this zone for memmap. This affects the watermark
3785 * and per-cpu initialisations
3788 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
3789 if (realsize >= memmap_pages) {
3790 realsize -= memmap_pages;
3793 " %s zone: %lu pages used for memmap\n",
3794 zone_names[j], memmap_pages);
3797 " %s zone: %lu pages exceeds realsize %lu\n",
3798 zone_names[j], memmap_pages, realsize);
3800 /* Account for reserved pages */
3801 if (j == 0 && realsize > dma_reserve) {
3802 realsize -= dma_reserve;
3803 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
3804 zone_names[0], dma_reserve);
3807 if (!is_highmem_idx(j))
3808 nr_kernel_pages += realsize;
3809 nr_all_pages += realsize;
3811 zone->spanned_pages = size;
3812 zone->present_pages = realsize;
3815 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
3817 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
3819 zone->name = zone_names[j];
3820 spin_lock_init(&zone->lock);
3821 spin_lock_init(&zone->lru_lock);
3822 zone_seqlock_init(zone);
3823 zone->zone_pgdat = pgdat;
3825 zone->prev_priority = DEF_PRIORITY;
3827 zone_pcp_init(zone);
3829 INIT_LIST_HEAD(&zone->lru[l].list);
3830 zone->reclaim_stat.nr_saved_scan[l] = 0;
3832 zone->reclaim_stat.recent_rotated[0] = 0;
3833 zone->reclaim_stat.recent_rotated[1] = 0;
3834 zone->reclaim_stat.recent_scanned[0] = 0;
3835 zone->reclaim_stat.recent_scanned[1] = 0;
3836 zap_zone_vm_stats(zone);
3841 set_pageblock_order(pageblock_default_order());
3842 setup_usemap(pgdat, zone, size);
3843 ret = init_currently_empty_zone(zone, zone_start_pfn,
3844 size, MEMMAP_EARLY);
3846 memmap_init(size, nid, j, zone_start_pfn);
3847 zone_start_pfn += size;
3851 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
3853 /* Skip empty nodes */
3854 if (!pgdat->node_spanned_pages)
3857 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3858 /* ia64 gets its own node_mem_map, before this, without bootmem */
3859 if (!pgdat->node_mem_map) {
3860 unsigned long size, start, end;
3864 * The zone's endpoints aren't required to be MAX_ORDER
3865 * aligned but the node_mem_map endpoints must be in order
3866 * for the buddy allocator to function correctly.
3868 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
3869 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
3870 end = ALIGN(end, MAX_ORDER_NR_PAGES);
3871 size = (end - start) * sizeof(struct page);
3872 map = alloc_remap(pgdat->node_id, size);
3874 map = alloc_bootmem_node(pgdat, size);
3875 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
3877 #ifndef CONFIG_NEED_MULTIPLE_NODES
3879 * With no DISCONTIG, the global mem_map is just set as node 0's
3881 if (pgdat == NODE_DATA(0)) {
3882 mem_map = NODE_DATA(0)->node_mem_map;
3883 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3884 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
3885 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
3886 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3889 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3892 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
3893 unsigned long node_start_pfn, unsigned long *zholes_size)
3895 pg_data_t *pgdat = NODE_DATA(nid);
3897 pgdat->node_id = nid;
3898 pgdat->node_start_pfn = node_start_pfn;
3899 calculate_node_totalpages(pgdat, zones_size, zholes_size);
3901 alloc_node_mem_map(pgdat);
3902 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3903 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3904 nid, (unsigned long)pgdat,
3905 (unsigned long)pgdat->node_mem_map);
3908 free_area_init_core(pgdat, zones_size, zholes_size);
3911 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3913 #if MAX_NUMNODES > 1
3915 * Figure out the number of possible node ids.
3917 static void __init setup_nr_node_ids(void)
3920 unsigned int highest = 0;
3922 for_each_node_mask(node, node_possible_map)
3924 nr_node_ids = highest + 1;
3927 static inline void setup_nr_node_ids(void)
3933 * add_active_range - Register a range of PFNs backed by physical memory
3934 * @nid: The node ID the range resides on
3935 * @start_pfn: The start PFN of the available physical memory
3936 * @end_pfn: The end PFN of the available physical memory
3938 * These ranges are stored in an early_node_map[] and later used by
3939 * free_area_init_nodes() to calculate zone sizes and holes. If the
3940 * range spans a memory hole, it is up to the architecture to ensure
3941 * the memory is not freed by the bootmem allocator. If possible
3942 * the range being registered will be merged with existing ranges.
3944 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
3945 unsigned long end_pfn)
3949 mminit_dprintk(MMINIT_TRACE, "memory_register",
3950 "Entering add_active_range(%d, %#lx, %#lx) "
3951 "%d entries of %d used\n",
3952 nid, start_pfn, end_pfn,
3953 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
3955 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
3957 /* Merge with existing active regions if possible */
3958 for (i = 0; i < nr_nodemap_entries; i++) {
3959 if (early_node_map[i].nid != nid)
3962 /* Skip if an existing region covers this new one */
3963 if (start_pfn >= early_node_map[i].start_pfn &&
3964 end_pfn <= early_node_map[i].end_pfn)
3967 /* Merge forward if suitable */
3968 if (start_pfn <= early_node_map[i].end_pfn &&
3969 end_pfn > early_node_map[i].end_pfn) {
3970 early_node_map[i].end_pfn = end_pfn;
3974 /* Merge backward if suitable */
3975 if (start_pfn < early_node_map[i].end_pfn &&
3976 end_pfn >= early_node_map[i].start_pfn) {
3977 early_node_map[i].start_pfn = start_pfn;
3982 /* Check that early_node_map is large enough */
3983 if (i >= MAX_ACTIVE_REGIONS) {
3984 printk(KERN_CRIT "More than %d memory regions, truncating\n",
3985 MAX_ACTIVE_REGIONS);
3989 early_node_map[i].nid = nid;
3990 early_node_map[i].start_pfn = start_pfn;
3991 early_node_map[i].end_pfn = end_pfn;
3992 nr_nodemap_entries = i + 1;
3996 * remove_active_range - Shrink an existing registered range of PFNs
3997 * @nid: The node id the range is on that should be shrunk
3998 * @start_pfn: The new PFN of the range
3999 * @end_pfn: The new PFN of the range
4001 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4002 * The map is kept near the end physical page range that has already been
4003 * registered. This function allows an arch to shrink an existing registered
4006 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4007 unsigned long end_pfn)
4012 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4013 nid, start_pfn, end_pfn);
4015 /* Find the old active region end and shrink */
4016 for_each_active_range_index_in_nid(i, nid) {
4017 if (early_node_map[i].start_pfn >= start_pfn &&
4018 early_node_map[i].end_pfn <= end_pfn) {
4020 early_node_map[i].start_pfn = 0;
4021 early_node_map[i].end_pfn = 0;
4025 if (early_node_map[i].start_pfn < start_pfn &&
4026 early_node_map[i].end_pfn > start_pfn) {
4027 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4028 early_node_map[i].end_pfn = start_pfn;
4029 if (temp_end_pfn > end_pfn)
4030 add_active_range(nid, end_pfn, temp_end_pfn);
4033 if (early_node_map[i].start_pfn >= start_pfn &&
4034 early_node_map[i].end_pfn > end_pfn &&
4035 early_node_map[i].start_pfn < end_pfn) {
4036 early_node_map[i].start_pfn = end_pfn;
4044 /* remove the blank ones */
4045 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4046 if (early_node_map[i].nid != nid)
4048 if (early_node_map[i].end_pfn)
4050 /* we found it, get rid of it */
4051 for (j = i; j < nr_nodemap_entries - 1; j++)
4052 memcpy(&early_node_map[j], &early_node_map[j+1],
4053 sizeof(early_node_map[j]));
4054 j = nr_nodemap_entries - 1;
4055 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4056 nr_nodemap_entries--;
4061 * remove_all_active_ranges - Remove all currently registered regions
4063 * During discovery, it may be found that a table like SRAT is invalid
4064 * and an alternative discovery method must be used. This function removes
4065 * all currently registered regions.
4067 void __init remove_all_active_ranges(void)
4069 memset(early_node_map, 0, sizeof(early_node_map));
4070 nr_nodemap_entries = 0;
4073 /* Compare two active node_active_regions */
4074 static int __init cmp_node_active_region(const void *a, const void *b)
4076 struct node_active_region *arange = (struct node_active_region *)a;
4077 struct node_active_region *brange = (struct node_active_region *)b;
4079 /* Done this way to avoid overflows */
4080 if (arange->start_pfn > brange->start_pfn)
4082 if (arange->start_pfn < brange->start_pfn)
4088 /* sort the node_map by start_pfn */
4089 static void __init sort_node_map(void)
4091 sort(early_node_map, (size_t)nr_nodemap_entries,
4092 sizeof(struct node_active_region),
4093 cmp_node_active_region, NULL);
4096 /* Find the lowest pfn for a node */
4097 static unsigned long __init find_min_pfn_for_node(int nid)
4100 unsigned long min_pfn = ULONG_MAX;
4102 /* Assuming a sorted map, the first range found has the starting pfn */
4103 for_each_active_range_index_in_nid(i, nid)
4104 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4106 if (min_pfn == ULONG_MAX) {
4108 "Could not find start_pfn for node %d\n", nid);
4116 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4118 * It returns the minimum PFN based on information provided via
4119 * add_active_range().
4121 unsigned long __init find_min_pfn_with_active_regions(void)
4123 return find_min_pfn_for_node(MAX_NUMNODES);
4127 * early_calculate_totalpages()
4128 * Sum pages in active regions for movable zone.
4129 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4131 static unsigned long __init early_calculate_totalpages(void)
4134 unsigned long totalpages = 0;
4136 for (i = 0; i < nr_nodemap_entries; i++) {
4137 unsigned long pages = early_node_map[i].end_pfn -
4138 early_node_map[i].start_pfn;
4139 totalpages += pages;
4141 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4147 * Find the PFN the Movable zone begins in each node. Kernel memory
4148 * is spread evenly between nodes as long as the nodes have enough
4149 * memory. When they don't, some nodes will have more kernelcore than
4152 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4155 unsigned long usable_startpfn;
4156 unsigned long kernelcore_node, kernelcore_remaining;
4157 /* save the state before borrow the nodemask */
4158 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4159 unsigned long totalpages = early_calculate_totalpages();
4160 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4163 * If movablecore was specified, calculate what size of
4164 * kernelcore that corresponds so that memory usable for
4165 * any allocation type is evenly spread. If both kernelcore
4166 * and movablecore are specified, then the value of kernelcore
4167 * will be used for required_kernelcore if it's greater than
4168 * what movablecore would have allowed.
4170 if (required_movablecore) {
4171 unsigned long corepages;
4174 * Round-up so that ZONE_MOVABLE is at least as large as what
4175 * was requested by the user
4177 required_movablecore =
4178 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4179 corepages = totalpages - required_movablecore;
4181 required_kernelcore = max(required_kernelcore, corepages);
4184 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4185 if (!required_kernelcore)
4188 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4189 find_usable_zone_for_movable();
4190 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4193 /* Spread kernelcore memory as evenly as possible throughout nodes */
4194 kernelcore_node = required_kernelcore / usable_nodes;
4195 for_each_node_state(nid, N_HIGH_MEMORY) {
4197 * Recalculate kernelcore_node if the division per node
4198 * now exceeds what is necessary to satisfy the requested
4199 * amount of memory for the kernel
4201 if (required_kernelcore < kernelcore_node)
4202 kernelcore_node = required_kernelcore / usable_nodes;
4205 * As the map is walked, we track how much memory is usable
4206 * by the kernel using kernelcore_remaining. When it is
4207 * 0, the rest of the node is usable by ZONE_MOVABLE
4209 kernelcore_remaining = kernelcore_node;
4211 /* Go through each range of PFNs within this node */
4212 for_each_active_range_index_in_nid(i, nid) {
4213 unsigned long start_pfn, end_pfn;
4214 unsigned long size_pages;
4216 start_pfn = max(early_node_map[i].start_pfn,
4217 zone_movable_pfn[nid]);
4218 end_pfn = early_node_map[i].end_pfn;
4219 if (start_pfn >= end_pfn)
4222 /* Account for what is only usable for kernelcore */
4223 if (start_pfn < usable_startpfn) {
4224 unsigned long kernel_pages;
4225 kernel_pages = min(end_pfn, usable_startpfn)
4228 kernelcore_remaining -= min(kernel_pages,
4229 kernelcore_remaining);
4230 required_kernelcore -= min(kernel_pages,
4231 required_kernelcore);
4233 /* Continue if range is now fully accounted */
4234 if (end_pfn <= usable_startpfn) {
4237 * Push zone_movable_pfn to the end so
4238 * that if we have to rebalance
4239 * kernelcore across nodes, we will
4240 * not double account here
4242 zone_movable_pfn[nid] = end_pfn;
4245 start_pfn = usable_startpfn;
4249 * The usable PFN range for ZONE_MOVABLE is from
4250 * start_pfn->end_pfn. Calculate size_pages as the
4251 * number of pages used as kernelcore
4253 size_pages = end_pfn - start_pfn;
4254 if (size_pages > kernelcore_remaining)
4255 size_pages = kernelcore_remaining;
4256 zone_movable_pfn[nid] = start_pfn + size_pages;
4259 * Some kernelcore has been met, update counts and
4260 * break if the kernelcore for this node has been
4263 required_kernelcore -= min(required_kernelcore,
4265 kernelcore_remaining -= size_pages;
4266 if (!kernelcore_remaining)
4272 * If there is still required_kernelcore, we do another pass with one
4273 * less node in the count. This will push zone_movable_pfn[nid] further
4274 * along on the nodes that still have memory until kernelcore is
4278 if (usable_nodes && required_kernelcore > usable_nodes)
4281 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4282 for (nid = 0; nid < MAX_NUMNODES; nid++)
4283 zone_movable_pfn[nid] =
4284 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4287 /* restore the node_state */
4288 node_states[N_HIGH_MEMORY] = saved_node_state;
4291 /* Any regular memory on that node ? */
4292 static void check_for_regular_memory(pg_data_t *pgdat)
4294 #ifdef CONFIG_HIGHMEM
4295 enum zone_type zone_type;
4297 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4298 struct zone *zone = &pgdat->node_zones[zone_type];
4299 if (zone->present_pages)
4300 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4306 * free_area_init_nodes - Initialise all pg_data_t and zone data
4307 * @max_zone_pfn: an array of max PFNs for each zone
4309 * This will call free_area_init_node() for each active node in the system.
4310 * Using the page ranges provided by add_active_range(), the size of each
4311 * zone in each node and their holes is calculated. If the maximum PFN
4312 * between two adjacent zones match, it is assumed that the zone is empty.
4313 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4314 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4315 * starts where the previous one ended. For example, ZONE_DMA32 starts
4316 * at arch_max_dma_pfn.
4318 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4323 /* Sort early_node_map as initialisation assumes it is sorted */
4326 /* Record where the zone boundaries are */
4327 memset(arch_zone_lowest_possible_pfn, 0,
4328 sizeof(arch_zone_lowest_possible_pfn));
4329 memset(arch_zone_highest_possible_pfn, 0,
4330 sizeof(arch_zone_highest_possible_pfn));
4331 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4332 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4333 for (i = 1; i < MAX_NR_ZONES; i++) {
4334 if (i == ZONE_MOVABLE)
4336 arch_zone_lowest_possible_pfn[i] =
4337 arch_zone_highest_possible_pfn[i-1];
4338 arch_zone_highest_possible_pfn[i] =
4339 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4341 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4342 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4344 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4345 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4346 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4348 /* Print out the zone ranges */
4349 printk("Zone PFN ranges:\n");
4350 for (i = 0; i < MAX_NR_ZONES; i++) {
4351 if (i == ZONE_MOVABLE)
4353 printk(" %-8s %0#10lx -> %0#10lx\n",
4355 arch_zone_lowest_possible_pfn[i],
4356 arch_zone_highest_possible_pfn[i]);
4359 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4360 printk("Movable zone start PFN for each node\n");
4361 for (i = 0; i < MAX_NUMNODES; i++) {
4362 if (zone_movable_pfn[i])
4363 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4366 /* Print out the early_node_map[] */
4367 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4368 for (i = 0; i < nr_nodemap_entries; i++)
4369 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4370 early_node_map[i].start_pfn,
4371 early_node_map[i].end_pfn);
4373 /* Initialise every node */
4374 mminit_verify_pageflags_layout();
4375 setup_nr_node_ids();
4376 for_each_online_node(nid) {
4377 pg_data_t *pgdat = NODE_DATA(nid);
4378 free_area_init_node(nid, NULL,
4379 find_min_pfn_for_node(nid), NULL);
4381 /* Any memory on that node */
4382 if (pgdat->node_present_pages)
4383 node_set_state(nid, N_HIGH_MEMORY);
4384 check_for_regular_memory(pgdat);
4388 static int __init cmdline_parse_core(char *p, unsigned long *core)
4390 unsigned long long coremem;
4394 coremem = memparse(p, &p);
4395 *core = coremem >> PAGE_SHIFT;
4397 /* Paranoid check that UL is enough for the coremem value */
4398 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4404 * kernelcore=size sets the amount of memory for use for allocations that
4405 * cannot be reclaimed or migrated.
4407 static int __init cmdline_parse_kernelcore(char *p)
4409 return cmdline_parse_core(p, &required_kernelcore);
4413 * movablecore=size sets the amount of memory for use for allocations that
4414 * can be reclaimed or migrated.
4416 static int __init cmdline_parse_movablecore(char *p)
4418 return cmdline_parse_core(p, &required_movablecore);
4421 early_param("kernelcore", cmdline_parse_kernelcore);
4422 early_param("movablecore", cmdline_parse_movablecore);
4424 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4427 * set_dma_reserve - set the specified number of pages reserved in the first zone
4428 * @new_dma_reserve: The number of pages to mark reserved
4430 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4431 * In the DMA zone, a significant percentage may be consumed by kernel image
4432 * and other unfreeable allocations which can skew the watermarks badly. This
4433 * function may optionally be used to account for unfreeable pages in the
4434 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4435 * smaller per-cpu batchsize.
4437 void __init set_dma_reserve(unsigned long new_dma_reserve)
4439 dma_reserve = new_dma_reserve;
4442 #ifndef CONFIG_NEED_MULTIPLE_NODES
4443 struct pglist_data __refdata contig_page_data = { .bdata = &bootmem_node_data[0] };
4444 EXPORT_SYMBOL(contig_page_data);
4447 void __init free_area_init(unsigned long *zones_size)
4449 free_area_init_node(0, zones_size,
4450 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4453 static int page_alloc_cpu_notify(struct notifier_block *self,
4454 unsigned long action, void *hcpu)
4456 int cpu = (unsigned long)hcpu;
4458 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4462 * Spill the event counters of the dead processor
4463 * into the current processors event counters.
4464 * This artificially elevates the count of the current
4467 vm_events_fold_cpu(cpu);
4470 * Zero the differential counters of the dead processor
4471 * so that the vm statistics are consistent.
4473 * This is only okay since the processor is dead and cannot
4474 * race with what we are doing.
4476 refresh_cpu_vm_stats(cpu);
4481 void __init page_alloc_init(void)
4483 hotcpu_notifier(page_alloc_cpu_notify, 0);
4487 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4488 * or min_free_kbytes changes.
4490 static void calculate_totalreserve_pages(void)
4492 struct pglist_data *pgdat;
4493 unsigned long reserve_pages = 0;
4494 enum zone_type i, j;
4496 for_each_online_pgdat(pgdat) {
4497 for (i = 0; i < MAX_NR_ZONES; i++) {
4498 struct zone *zone = pgdat->node_zones + i;
4499 unsigned long max = 0;
4501 /* Find valid and maximum lowmem_reserve in the zone */
4502 for (j = i; j < MAX_NR_ZONES; j++) {
4503 if (zone->lowmem_reserve[j] > max)
4504 max = zone->lowmem_reserve[j];
4507 /* we treat the high watermark as reserved pages. */
4508 max += high_wmark_pages(zone);
4510 if (max > zone->present_pages)
4511 max = zone->present_pages;
4512 reserve_pages += max;
4515 totalreserve_pages = reserve_pages;
4519 * setup_per_zone_lowmem_reserve - called whenever
4520 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4521 * has a correct pages reserved value, so an adequate number of
4522 * pages are left in the zone after a successful __alloc_pages().
4524 static void setup_per_zone_lowmem_reserve(void)
4526 struct pglist_data *pgdat;
4527 enum zone_type j, idx;
4529 for_each_online_pgdat(pgdat) {
4530 for (j = 0; j < MAX_NR_ZONES; j++) {
4531 struct zone *zone = pgdat->node_zones + j;
4532 unsigned long present_pages = zone->present_pages;
4534 zone->lowmem_reserve[j] = 0;
4538 struct zone *lower_zone;
4542 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4543 sysctl_lowmem_reserve_ratio[idx] = 1;
4545 lower_zone = pgdat->node_zones + idx;
4546 lower_zone->lowmem_reserve[j] = present_pages /
4547 sysctl_lowmem_reserve_ratio[idx];
4548 present_pages += lower_zone->present_pages;
4553 /* update totalreserve_pages */
4554 calculate_totalreserve_pages();
4558 * setup_per_zone_wmarks - called when min_free_kbytes changes
4559 * or when memory is hot-{added|removed}
4561 * Ensures that the watermark[min,low,high] values for each zone are set
4562 * correctly with respect to min_free_kbytes.
4564 void setup_per_zone_wmarks(void)
4566 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4567 unsigned long lowmem_pages = 0;
4569 unsigned long flags;
4571 /* Calculate total number of !ZONE_HIGHMEM pages */
4572 for_each_zone(zone) {
4573 if (!is_highmem(zone))
4574 lowmem_pages += zone->present_pages;
4577 for_each_zone(zone) {
4580 spin_lock_irqsave(&zone->lock, flags);
4581 tmp = (u64)pages_min * zone->present_pages;
4582 do_div(tmp, lowmem_pages);
4583 if (is_highmem(zone)) {
4585 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4586 * need highmem pages, so cap pages_min to a small
4589 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4590 * deltas controls asynch page reclaim, and so should
4591 * not be capped for highmem.
4595 min_pages = zone->present_pages / 1024;
4596 if (min_pages < SWAP_CLUSTER_MAX)
4597 min_pages = SWAP_CLUSTER_MAX;
4598 if (min_pages > 128)
4600 zone->watermark[WMARK_MIN] = min_pages;
4603 * If it's a lowmem zone, reserve a number of pages
4604 * proportionate to the zone's size.
4606 zone->watermark[WMARK_MIN] = tmp;
4609 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
4610 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
4611 setup_zone_migrate_reserve(zone);
4612 spin_unlock_irqrestore(&zone->lock, flags);
4615 /* update totalreserve_pages */
4616 calculate_totalreserve_pages();
4620 * The inactive anon list should be small enough that the VM never has to
4621 * do too much work, but large enough that each inactive page has a chance
4622 * to be referenced again before it is swapped out.
4624 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4625 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4626 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4627 * the anonymous pages are kept on the inactive list.
4630 * memory ratio inactive anon
4631 * -------------------------------------
4640 void calculate_zone_inactive_ratio(struct zone *zone)
4642 unsigned int gb, ratio;
4644 /* Zone size in gigabytes */
4645 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4647 ratio = int_sqrt(10 * gb);
4651 zone->inactive_ratio = ratio;
4654 static void __init setup_per_zone_inactive_ratio(void)
4659 calculate_zone_inactive_ratio(zone);
4663 * Initialise min_free_kbytes.
4665 * For small machines we want it small (128k min). For large machines
4666 * we want it large (64MB max). But it is not linear, because network
4667 * bandwidth does not increase linearly with machine size. We use
4669 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4670 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4686 static int __init init_per_zone_wmark_min(void)
4688 unsigned long lowmem_kbytes;
4690 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4692 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4693 if (min_free_kbytes < 128)
4694 min_free_kbytes = 128;
4695 if (min_free_kbytes > 65536)
4696 min_free_kbytes = 65536;
4697 setup_per_zone_wmarks();
4698 setup_per_zone_lowmem_reserve();
4699 setup_per_zone_inactive_ratio();
4702 module_init(init_per_zone_wmark_min)
4705 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4706 * that we can call two helper functions whenever min_free_kbytes
4709 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
4710 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4712 proc_dointvec(table, write, file, buffer, length, ppos);
4714 setup_per_zone_wmarks();
4719 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
4720 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4725 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4730 zone->min_unmapped_pages = (zone->present_pages *
4731 sysctl_min_unmapped_ratio) / 100;
4735 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
4736 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4741 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4746 zone->min_slab_pages = (zone->present_pages *
4747 sysctl_min_slab_ratio) / 100;
4753 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4754 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4755 * whenever sysctl_lowmem_reserve_ratio changes.
4757 * The reserve ratio obviously has absolutely no relation with the
4758 * minimum watermarks. The lowmem reserve ratio can only make sense
4759 * if in function of the boot time zone sizes.
4761 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
4762 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4764 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4765 setup_per_zone_lowmem_reserve();
4770 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4771 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4772 * can have before it gets flushed back to buddy allocator.
4775 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
4776 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4782 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4783 if (!write || (ret == -EINVAL))
4785 for_each_populated_zone(zone) {
4786 for_each_online_cpu(cpu) {
4788 high = zone->present_pages / percpu_pagelist_fraction;
4789 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
4795 int hashdist = HASHDIST_DEFAULT;
4798 static int __init set_hashdist(char *str)
4802 hashdist = simple_strtoul(str, &str, 0);
4805 __setup("hashdist=", set_hashdist);
4809 * allocate a large system hash table from bootmem
4810 * - it is assumed that the hash table must contain an exact power-of-2
4811 * quantity of entries
4812 * - limit is the number of hash buckets, not the total allocation size
4814 void *__init alloc_large_system_hash(const char *tablename,
4815 unsigned long bucketsize,
4816 unsigned long numentries,
4819 unsigned int *_hash_shift,
4820 unsigned int *_hash_mask,
4821 unsigned long limit)
4823 unsigned long long max = limit;
4824 unsigned long log2qty, size;
4827 /* allow the kernel cmdline to have a say */
4829 /* round applicable memory size up to nearest megabyte */
4830 numentries = nr_kernel_pages;
4831 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
4832 numentries >>= 20 - PAGE_SHIFT;
4833 numentries <<= 20 - PAGE_SHIFT;
4835 /* limit to 1 bucket per 2^scale bytes of low memory */
4836 if (scale > PAGE_SHIFT)
4837 numentries >>= (scale - PAGE_SHIFT);
4839 numentries <<= (PAGE_SHIFT - scale);
4841 /* Make sure we've got at least a 0-order allocation.. */
4842 if (unlikely(flags & HASH_SMALL)) {
4843 /* Makes no sense without HASH_EARLY */
4844 WARN_ON(!(flags & HASH_EARLY));
4845 if (!(numentries >> *_hash_shift)) {
4846 numentries = 1UL << *_hash_shift;
4847 BUG_ON(!numentries);
4849 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
4850 numentries = PAGE_SIZE / bucketsize;
4852 numentries = roundup_pow_of_two(numentries);
4854 /* limit allocation size to 1/16 total memory by default */
4856 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
4857 do_div(max, bucketsize);
4860 if (numentries > max)
4863 log2qty = ilog2(numentries);
4866 size = bucketsize << log2qty;
4867 if (flags & HASH_EARLY)
4868 table = alloc_bootmem_nopanic(size);
4870 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
4873 * If bucketsize is not a power-of-two, we may free
4874 * some pages at the end of hash table which
4875 * alloc_pages_exact() automatically does
4877 if (get_order(size) < MAX_ORDER) {
4878 table = alloc_pages_exact(size, GFP_ATOMIC);
4879 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
4882 } while (!table && size > PAGE_SIZE && --log2qty);
4885 panic("Failed to allocate %s hash table\n", tablename);
4887 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
4890 ilog2(size) - PAGE_SHIFT,
4894 *_hash_shift = log2qty;
4896 *_hash_mask = (1 << log2qty) - 1;
4901 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4902 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
4905 #ifdef CONFIG_SPARSEMEM
4906 return __pfn_to_section(pfn)->pageblock_flags;
4908 return zone->pageblock_flags;
4909 #endif /* CONFIG_SPARSEMEM */
4912 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
4914 #ifdef CONFIG_SPARSEMEM
4915 pfn &= (PAGES_PER_SECTION-1);
4916 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4918 pfn = pfn - zone->zone_start_pfn;
4919 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4920 #endif /* CONFIG_SPARSEMEM */
4924 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4925 * @page: The page within the block of interest
4926 * @start_bitidx: The first bit of interest to retrieve
4927 * @end_bitidx: The last bit of interest
4928 * returns pageblock_bits flags
4930 unsigned long get_pageblock_flags_group(struct page *page,
4931 int start_bitidx, int end_bitidx)
4934 unsigned long *bitmap;
4935 unsigned long pfn, bitidx;
4936 unsigned long flags = 0;
4937 unsigned long value = 1;
4939 zone = page_zone(page);
4940 pfn = page_to_pfn(page);
4941 bitmap = get_pageblock_bitmap(zone, pfn);
4942 bitidx = pfn_to_bitidx(zone, pfn);
4944 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4945 if (test_bit(bitidx + start_bitidx, bitmap))
4952 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4953 * @page: The page within the block of interest
4954 * @start_bitidx: The first bit of interest
4955 * @end_bitidx: The last bit of interest
4956 * @flags: The flags to set
4958 void set_pageblock_flags_group(struct page *page, unsigned long flags,
4959 int start_bitidx, int end_bitidx)
4962 unsigned long *bitmap;
4963 unsigned long pfn, bitidx;
4964 unsigned long value = 1;
4966 zone = page_zone(page);
4967 pfn = page_to_pfn(page);
4968 bitmap = get_pageblock_bitmap(zone, pfn);
4969 bitidx = pfn_to_bitidx(zone, pfn);
4970 VM_BUG_ON(pfn < zone->zone_start_pfn);
4971 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
4973 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4975 __set_bit(bitidx + start_bitidx, bitmap);
4977 __clear_bit(bitidx + start_bitidx, bitmap);
4981 * This is designed as sub function...plz see page_isolation.c also.
4982 * set/clear page block's type to be ISOLATE.
4983 * page allocater never alloc memory from ISOLATE block.
4986 int set_migratetype_isolate(struct page *page)
4989 unsigned long flags;
4993 zone = page_zone(page);
4994 zone_idx = zone_idx(zone);
4995 spin_lock_irqsave(&zone->lock, flags);
4997 * In future, more migrate types will be able to be isolation target.
4999 if (get_pageblock_migratetype(page) != MIGRATE_MOVABLE &&
5000 zone_idx != ZONE_MOVABLE)
5002 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5003 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5006 spin_unlock_irqrestore(&zone->lock, flags);
5012 void unset_migratetype_isolate(struct page *page)
5015 unsigned long flags;
5016 zone = page_zone(page);
5017 spin_lock_irqsave(&zone->lock, flags);
5018 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5020 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5021 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5023 spin_unlock_irqrestore(&zone->lock, flags);
5026 #ifdef CONFIG_MEMORY_HOTREMOVE
5028 * All pages in the range must be isolated before calling this.
5031 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5037 unsigned long flags;
5038 /* find the first valid pfn */
5039 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5044 zone = page_zone(pfn_to_page(pfn));
5045 spin_lock_irqsave(&zone->lock, flags);
5047 while (pfn < end_pfn) {
5048 if (!pfn_valid(pfn)) {
5052 page = pfn_to_page(pfn);
5053 BUG_ON(page_count(page));
5054 BUG_ON(!PageBuddy(page));
5055 order = page_order(page);
5056 #ifdef CONFIG_DEBUG_VM
5057 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5058 pfn, 1 << order, end_pfn);
5060 list_del(&page->lru);
5061 rmv_page_order(page);
5062 zone->free_area[order].nr_free--;
5063 __mod_zone_page_state(zone, NR_FREE_PAGES,
5065 for (i = 0; i < (1 << order); i++)
5066 SetPageReserved((page+i));
5067 pfn += (1 << order);
5069 spin_unlock_irqrestore(&zone->lock, flags);