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 int percpu_pagelist_fraction;
76 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
78 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
79 int pageblock_order __read_mostly;
82 static void __free_pages_ok(struct page *page, unsigned int order);
85 * results with 256, 32 in the lowmem_reserve sysctl:
86 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
87 * 1G machine -> (16M dma, 784M normal, 224M high)
88 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
89 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
90 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
92 * TBD: should special case ZONE_DMA32 machines here - in those we normally
93 * don't need any ZONE_NORMAL reservation
95 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
96 #ifdef CONFIG_ZONE_DMA
99 #ifdef CONFIG_ZONE_DMA32
102 #ifdef CONFIG_HIGHMEM
108 EXPORT_SYMBOL(totalram_pages);
110 static char * const zone_names[MAX_NR_ZONES] = {
111 #ifdef CONFIG_ZONE_DMA
114 #ifdef CONFIG_ZONE_DMA32
118 #ifdef CONFIG_HIGHMEM
124 int min_free_kbytes = 1024;
125 int min_free_order_shift = 1;
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;
238 /* Don't complain about poisoned pages */
239 if (PageHWPoison(page)) {
240 __ClearPageBuddy(page);
245 * Allow a burst of 60 reports, then keep quiet for that minute;
246 * or allow a steady drip of one report per second.
248 if (nr_shown == 60) {
249 if (time_before(jiffies, resume)) {
255 "BUG: Bad page state: %lu messages suppressed\n",
262 resume = jiffies + 60 * HZ;
264 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
265 current->comm, page_to_pfn(page));
267 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
268 page, (void *)page->flags, page_count(page),
269 page_mapcount(page), page->mapping, page->index);
273 /* Leave bad fields for debug, except PageBuddy could make trouble */
274 __ClearPageBuddy(page);
275 add_taint(TAINT_BAD_PAGE);
279 * Higher-order pages are called "compound pages". They are structured thusly:
281 * The first PAGE_SIZE page is called the "head page".
283 * The remaining PAGE_SIZE pages are called "tail pages".
285 * All pages have PG_compound set. All pages have their ->private pointing at
286 * the head page (even the head page has this).
288 * The first tail page's ->lru.next holds the address of the compound page's
289 * put_page() function. Its ->lru.prev holds the order of allocation.
290 * This usage means that zero-order pages may not be compound.
293 static void free_compound_page(struct page *page)
295 __free_pages_ok(page, compound_order(page));
298 void prep_compound_page(struct page *page, unsigned long order)
301 int nr_pages = 1 << order;
303 set_compound_page_dtor(page, free_compound_page);
304 set_compound_order(page, order);
306 for (i = 1; i < nr_pages; i++) {
307 struct page *p = page + i;
310 p->first_page = page;
314 static int destroy_compound_page(struct page *page, unsigned long order)
317 int nr_pages = 1 << order;
320 if (unlikely(compound_order(page) != order) ||
321 unlikely(!PageHead(page))) {
326 __ClearPageHead(page);
328 for (i = 1; i < nr_pages; i++) {
329 struct page *p = page + i;
331 if (unlikely(!PageTail(p) || (p->first_page != page))) {
341 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
346 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
347 * and __GFP_HIGHMEM from hard or soft interrupt context.
349 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
350 for (i = 0; i < (1 << order); i++)
351 clear_highpage(page + i);
354 static inline void set_page_order(struct page *page, int order)
356 set_page_private(page, order);
357 __SetPageBuddy(page);
360 static inline void rmv_page_order(struct page *page)
362 __ClearPageBuddy(page);
363 set_page_private(page, 0);
367 * Locate the struct page for both the matching buddy in our
368 * pair (buddy1) and the combined O(n+1) page they form (page).
370 * 1) Any buddy B1 will have an order O twin B2 which satisfies
371 * the following equation:
373 * For example, if the starting buddy (buddy2) is #8 its order
375 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
377 * 2) Any buddy B will have an order O+1 parent P which
378 * satisfies the following equation:
381 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
383 static inline struct page *
384 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
386 unsigned long buddy_idx = page_idx ^ (1 << order);
388 return page + (buddy_idx - page_idx);
391 static inline unsigned long
392 __find_combined_index(unsigned long page_idx, unsigned int order)
394 return (page_idx & ~(1 << order));
398 * This function checks whether a page is free && is the buddy
399 * we can do coalesce a page and its buddy if
400 * (a) the buddy is not in a hole &&
401 * (b) the buddy is in the buddy system &&
402 * (c) a page and its buddy have the same order &&
403 * (d) a page and its buddy are in the same zone.
405 * For recording whether a page is in the buddy system, we use PG_buddy.
406 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
408 * For recording page's order, we use page_private(page).
410 static inline int page_is_buddy(struct page *page, struct page *buddy,
413 if (!pfn_valid_within(page_to_pfn(buddy)))
416 if (page_zone_id(page) != page_zone_id(buddy))
419 if (PageBuddy(buddy) && page_order(buddy) == order) {
420 VM_BUG_ON(page_count(buddy) != 0);
427 * Freeing function for a buddy system allocator.
429 * The concept of a buddy system is to maintain direct-mapped table
430 * (containing bit values) for memory blocks of various "orders".
431 * The bottom level table contains the map for the smallest allocatable
432 * units of memory (here, pages), and each level above it describes
433 * pairs of units from the levels below, hence, "buddies".
434 * At a high level, all that happens here is marking the table entry
435 * at the bottom level available, and propagating the changes upward
436 * as necessary, plus some accounting needed to play nicely with other
437 * parts of the VM system.
438 * At each level, we keep a list of pages, which are heads of continuous
439 * free pages of length of (1 << order) and marked with PG_buddy. Page's
440 * order is recorded in page_private(page) field.
441 * So when we are allocating or freeing one, we can derive the state of the
442 * other. That is, if we allocate a small block, and both were
443 * free, the remainder of the region must be split into blocks.
444 * If a block is freed, and its buddy is also free, then this
445 * triggers coalescing into a block of larger size.
450 static inline void __free_one_page(struct page *page,
451 struct zone *zone, unsigned int order,
454 unsigned long page_idx;
456 if (unlikely(PageCompound(page)))
457 if (unlikely(destroy_compound_page(page, order)))
460 VM_BUG_ON(migratetype == -1);
462 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
464 VM_BUG_ON(page_idx & ((1 << order) - 1));
465 VM_BUG_ON(bad_range(zone, page));
467 while (order < MAX_ORDER-1) {
468 unsigned long combined_idx;
471 buddy = __page_find_buddy(page, page_idx, order);
472 if (!page_is_buddy(page, buddy, order))
475 /* Our buddy is free, merge with it and move up one order. */
476 list_del(&buddy->lru);
477 zone->free_area[order].nr_free--;
478 rmv_page_order(buddy);
479 combined_idx = __find_combined_index(page_idx, order);
480 page = page + (combined_idx - page_idx);
481 page_idx = combined_idx;
484 set_page_order(page, order);
486 &zone->free_area[order].free_list[migratetype]);
487 zone->free_area[order].nr_free++;
490 #ifdef CONFIG_HAVE_MLOCKED_PAGE_BIT
492 * free_page_mlock() -- clean up attempts to free and mlocked() page.
493 * Page should not be on lru, so no need to fix that up.
494 * free_pages_check() will verify...
496 static inline void free_page_mlock(struct page *page)
498 __dec_zone_page_state(page, NR_MLOCK);
499 __count_vm_event(UNEVICTABLE_MLOCKFREED);
502 static void free_page_mlock(struct page *page) { }
505 static inline int free_pages_check(struct page *page)
507 if (unlikely(page_mapcount(page) |
508 (page->mapping != NULL) |
509 (atomic_read(&page->_count) != 0) |
510 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) {
514 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
515 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
520 * Frees a number of pages from the PCP lists
521 * Assumes all pages on list are in same zone, and of same order.
522 * count is the number of pages to free.
524 * If the zone was previously in an "all pages pinned" state then look to
525 * see if this freeing clears that state.
527 * And clear the zone's pages_scanned counter, to hold off the "all pages are
528 * pinned" detection logic.
530 static void free_pcppages_bulk(struct zone *zone, int count,
531 struct per_cpu_pages *pcp)
536 spin_lock(&zone->lock);
537 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
538 zone->pages_scanned = 0;
540 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
543 struct list_head *list;
546 * Remove pages from lists in a round-robin fashion. A
547 * batch_free count is maintained that is incremented when an
548 * empty list is encountered. This is so more pages are freed
549 * off fuller lists instead of spinning excessively around empty
554 if (++migratetype == MIGRATE_PCPTYPES)
556 list = &pcp->lists[migratetype];
557 } while (list_empty(list));
560 page = list_entry(list->prev, struct page, lru);
561 /* must delete as __free_one_page list manipulates */
562 list_del(&page->lru);
563 __free_one_page(page, zone, 0, migratetype);
564 trace_mm_page_pcpu_drain(page, 0, migratetype);
565 } while (--count && --batch_free && !list_empty(list));
567 spin_unlock(&zone->lock);
570 static void free_one_page(struct zone *zone, struct page *page, int order,
573 spin_lock(&zone->lock);
574 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
575 zone->pages_scanned = 0;
577 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
578 __free_one_page(page, zone, order, migratetype);
579 spin_unlock(&zone->lock);
582 static void __free_pages_ok(struct page *page, unsigned int order)
587 int wasMlocked = __TestClearPageMlocked(page);
589 kmemcheck_free_shadow(page, order);
591 for (i = 0 ; i < (1 << order) ; ++i)
592 bad += free_pages_check(page + i);
596 if (!PageHighMem(page)) {
597 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
598 debug_check_no_obj_freed(page_address(page),
601 arch_free_page(page, order);
602 kernel_map_pages(page, 1 << order, 0);
604 local_irq_save(flags);
605 if (unlikely(wasMlocked))
606 free_page_mlock(page);
607 __count_vm_events(PGFREE, 1 << order);
608 free_one_page(page_zone(page), page, order,
609 get_pageblock_migratetype(page));
610 local_irq_restore(flags);
614 * permit the bootmem allocator to evade page validation on high-order frees
616 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
619 __ClearPageReserved(page);
620 set_page_count(page, 0);
621 set_page_refcounted(page);
627 for (loop = 0; loop < BITS_PER_LONG; loop++) {
628 struct page *p = &page[loop];
630 if (loop + 1 < BITS_PER_LONG)
632 __ClearPageReserved(p);
633 set_page_count(p, 0);
636 set_page_refcounted(page);
637 __free_pages(page, order);
643 * The order of subdivision here is critical for the IO subsystem.
644 * Please do not alter this order without good reasons and regression
645 * testing. Specifically, as large blocks of memory are subdivided,
646 * the order in which smaller blocks are delivered depends on the order
647 * they're subdivided in this function. This is the primary factor
648 * influencing the order in which pages are delivered to the IO
649 * subsystem according to empirical testing, and this is also justified
650 * by considering the behavior of a buddy system containing a single
651 * large block of memory acted on by a series of small allocations.
652 * This behavior is a critical factor in sglist merging's success.
656 static inline void expand(struct zone *zone, struct page *page,
657 int low, int high, struct free_area *area,
660 unsigned long size = 1 << high;
666 VM_BUG_ON(bad_range(zone, &page[size]));
667 list_add(&page[size].lru, &area->free_list[migratetype]);
669 set_page_order(&page[size], high);
674 * This page is about to be returned from the page allocator
676 static inline int check_new_page(struct page *page)
678 if (unlikely(page_mapcount(page) |
679 (page->mapping != NULL) |
680 (atomic_read(&page->_count) != 0) |
681 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) {
688 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
692 for (i = 0; i < (1 << order); i++) {
693 struct page *p = page + i;
694 if (unlikely(check_new_page(p)))
698 set_page_private(page, 0);
699 set_page_refcounted(page);
701 arch_alloc_page(page, order);
702 kernel_map_pages(page, 1 << order, 1);
704 if (gfp_flags & __GFP_ZERO)
705 prep_zero_page(page, order, gfp_flags);
707 if (order && (gfp_flags & __GFP_COMP))
708 prep_compound_page(page, order);
714 * Go through the free lists for the given migratetype and remove
715 * the smallest available page from the freelists
718 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
721 unsigned int current_order;
722 struct free_area * area;
725 /* Find a page of the appropriate size in the preferred list */
726 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
727 area = &(zone->free_area[current_order]);
728 if (list_empty(&area->free_list[migratetype]))
731 page = list_entry(area->free_list[migratetype].next,
733 list_del(&page->lru);
734 rmv_page_order(page);
736 expand(zone, page, order, current_order, area, migratetype);
745 * This array describes the order lists are fallen back to when
746 * the free lists for the desirable migrate type are depleted
748 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
749 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
750 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
751 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
752 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
756 * Move the free pages in a range to the free lists of the requested type.
757 * Note that start_page and end_pages are not aligned on a pageblock
758 * boundary. If alignment is required, use move_freepages_block()
760 static int move_freepages(struct zone *zone,
761 struct page *start_page, struct page *end_page,
768 #ifndef CONFIG_HOLES_IN_ZONE
770 * page_zone is not safe to call in this context when
771 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
772 * anyway as we check zone boundaries in move_freepages_block().
773 * Remove at a later date when no bug reports exist related to
774 * grouping pages by mobility
776 BUG_ON(page_zone(start_page) != page_zone(end_page));
779 for (page = start_page; page <= end_page;) {
780 /* Make sure we are not inadvertently changing nodes */
781 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
783 if (!pfn_valid_within(page_to_pfn(page))) {
788 if (!PageBuddy(page)) {
793 order = page_order(page);
794 list_del(&page->lru);
796 &zone->free_area[order].free_list[migratetype]);
798 pages_moved += 1 << order;
804 static int move_freepages_block(struct zone *zone, struct page *page,
807 unsigned long start_pfn, end_pfn;
808 struct page *start_page, *end_page;
810 start_pfn = page_to_pfn(page);
811 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
812 start_page = pfn_to_page(start_pfn);
813 end_page = start_page + pageblock_nr_pages - 1;
814 end_pfn = start_pfn + pageblock_nr_pages - 1;
816 /* Do not cross zone boundaries */
817 if (start_pfn < zone->zone_start_pfn)
819 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
822 return move_freepages(zone, start_page, end_page, migratetype);
825 static void change_pageblock_range(struct page *pageblock_page,
826 int start_order, int migratetype)
828 int nr_pageblocks = 1 << (start_order - pageblock_order);
830 while (nr_pageblocks--) {
831 set_pageblock_migratetype(pageblock_page, migratetype);
832 pageblock_page += pageblock_nr_pages;
836 /* Remove an element from the buddy allocator from the fallback list */
837 static inline struct page *
838 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
840 struct free_area * area;
845 /* Find the largest possible block of pages in the other list */
846 for (current_order = MAX_ORDER-1; current_order >= order;
848 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
849 migratetype = fallbacks[start_migratetype][i];
851 /* MIGRATE_RESERVE handled later if necessary */
852 if (migratetype == MIGRATE_RESERVE)
855 area = &(zone->free_area[current_order]);
856 if (list_empty(&area->free_list[migratetype]))
859 page = list_entry(area->free_list[migratetype].next,
864 * If breaking a large block of pages, move all free
865 * pages to the preferred allocation list. If falling
866 * back for a reclaimable kernel allocation, be more
867 * agressive about taking ownership of free pages
869 if (unlikely(current_order >= (pageblock_order >> 1)) ||
870 start_migratetype == MIGRATE_RECLAIMABLE ||
871 page_group_by_mobility_disabled) {
873 pages = move_freepages_block(zone, page,
876 /* Claim the whole block if over half of it is free */
877 if (pages >= (1 << (pageblock_order-1)) ||
878 page_group_by_mobility_disabled)
879 set_pageblock_migratetype(page,
882 migratetype = start_migratetype;
885 /* Remove the page from the freelists */
886 list_del(&page->lru);
887 rmv_page_order(page);
889 /* Take ownership for orders >= pageblock_order */
890 if (current_order >= pageblock_order)
891 change_pageblock_range(page, current_order,
894 expand(zone, page, order, current_order, area, migratetype);
896 trace_mm_page_alloc_extfrag(page, order, current_order,
897 start_migratetype, migratetype);
907 * Do the hard work of removing an element from the buddy allocator.
908 * Call me with the zone->lock already held.
910 static struct page *__rmqueue(struct zone *zone, unsigned int order,
916 page = __rmqueue_smallest(zone, order, migratetype);
918 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
919 page = __rmqueue_fallback(zone, order, migratetype);
922 * Use MIGRATE_RESERVE rather than fail an allocation. goto
923 * is used because __rmqueue_smallest is an inline function
924 * and we want just one call site
927 migratetype = MIGRATE_RESERVE;
932 trace_mm_page_alloc_zone_locked(page, order, migratetype);
937 * Obtain a specified number of elements from the buddy allocator, all under
938 * a single hold of the lock, for efficiency. Add them to the supplied list.
939 * Returns the number of new pages which were placed at *list.
941 static int rmqueue_bulk(struct zone *zone, unsigned int order,
942 unsigned long count, struct list_head *list,
943 int migratetype, int cold)
947 spin_lock(&zone->lock);
948 for (i = 0; i < count; ++i) {
949 struct page *page = __rmqueue(zone, order, migratetype);
950 if (unlikely(page == NULL))
954 * Split buddy pages returned by expand() are received here
955 * in physical page order. The page is added to the callers and
956 * list and the list head then moves forward. From the callers
957 * perspective, the linked list is ordered by page number in
958 * some conditions. This is useful for IO devices that can
959 * merge IO requests if the physical pages are ordered
962 if (likely(cold == 0))
963 list_add(&page->lru, list);
965 list_add_tail(&page->lru, list);
966 set_page_private(page, migratetype);
969 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
970 spin_unlock(&zone->lock);
976 * Called from the vmstat counter updater to drain pagesets of this
977 * currently executing processor on remote nodes after they have
980 * Note that this function must be called with the thread pinned to
981 * a single processor.
983 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
988 local_irq_save(flags);
989 if (pcp->count >= pcp->batch)
990 to_drain = pcp->batch;
992 to_drain = pcp->count;
993 free_pcppages_bulk(zone, to_drain, pcp);
994 pcp->count -= to_drain;
995 local_irq_restore(flags);
1000 * Drain pages of the indicated processor.
1002 * The processor must either be the current processor and the
1003 * thread pinned to the current processor or a processor that
1006 static void drain_pages(unsigned int cpu)
1008 unsigned long flags;
1011 for_each_populated_zone(zone) {
1012 struct per_cpu_pageset *pset;
1013 struct per_cpu_pages *pcp;
1015 pset = zone_pcp(zone, cpu);
1018 local_irq_save(flags);
1019 free_pcppages_bulk(zone, pcp->count, pcp);
1021 local_irq_restore(flags);
1026 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1028 void drain_local_pages(void *arg)
1030 drain_pages(smp_processor_id());
1034 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1036 void drain_all_pages(void)
1038 on_each_cpu(drain_local_pages, NULL, 1);
1041 #ifdef CONFIG_HIBERNATION
1043 void mark_free_pages(struct zone *zone)
1045 unsigned long pfn, max_zone_pfn;
1046 unsigned long flags;
1048 struct list_head *curr;
1050 if (!zone->spanned_pages)
1053 spin_lock_irqsave(&zone->lock, flags);
1055 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1056 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1057 if (pfn_valid(pfn)) {
1058 struct page *page = pfn_to_page(pfn);
1060 if (!swsusp_page_is_forbidden(page))
1061 swsusp_unset_page_free(page);
1064 for_each_migratetype_order(order, t) {
1065 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1068 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1069 for (i = 0; i < (1UL << order); i++)
1070 swsusp_set_page_free(pfn_to_page(pfn + i));
1073 spin_unlock_irqrestore(&zone->lock, flags);
1075 #endif /* CONFIG_PM */
1078 * Free a 0-order page
1080 static void free_hot_cold_page(struct page *page, int cold)
1082 struct zone *zone = page_zone(page);
1083 struct per_cpu_pages *pcp;
1084 unsigned long flags;
1086 int wasMlocked = __TestClearPageMlocked(page);
1088 kmemcheck_free_shadow(page, 0);
1091 page->mapping = NULL;
1092 if (free_pages_check(page))
1095 if (!PageHighMem(page)) {
1096 debug_check_no_locks_freed(page_address(page), PAGE_SIZE);
1097 debug_check_no_obj_freed(page_address(page), PAGE_SIZE);
1099 arch_free_page(page, 0);
1100 kernel_map_pages(page, 1, 0);
1102 pcp = &zone_pcp(zone, get_cpu())->pcp;
1103 migratetype = get_pageblock_migratetype(page);
1104 set_page_private(page, migratetype);
1105 local_irq_save(flags);
1106 if (unlikely(wasMlocked))
1107 free_page_mlock(page);
1108 __count_vm_event(PGFREE);
1111 * We only track unmovable, reclaimable and movable on pcp lists.
1112 * Free ISOLATE pages back to the allocator because they are being
1113 * offlined but treat RESERVE as movable pages so we can get those
1114 * areas back if necessary. Otherwise, we may have to free
1115 * excessively into the page allocator
1117 if (migratetype >= MIGRATE_PCPTYPES) {
1118 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1119 free_one_page(zone, page, 0, migratetype);
1122 migratetype = MIGRATE_MOVABLE;
1126 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1128 list_add(&page->lru, &pcp->lists[migratetype]);
1130 if (pcp->count >= pcp->high) {
1131 free_pcppages_bulk(zone, pcp->batch, pcp);
1132 pcp->count -= pcp->batch;
1136 local_irq_restore(flags);
1140 void free_hot_page(struct page *page)
1142 trace_mm_page_free_direct(page, 0);
1143 free_hot_cold_page(page, 0);
1147 * split_page takes a non-compound higher-order page, and splits it into
1148 * n (1<<order) sub-pages: page[0..n]
1149 * Each sub-page must be freed individually.
1151 * Note: this is probably too low level an operation for use in drivers.
1152 * Please consult with lkml before using this in your driver.
1154 void split_page(struct page *page, unsigned int order)
1158 VM_BUG_ON(PageCompound(page));
1159 VM_BUG_ON(!page_count(page));
1161 #ifdef CONFIG_KMEMCHECK
1163 * Split shadow pages too, because free(page[0]) would
1164 * otherwise free the whole shadow.
1166 if (kmemcheck_page_is_tracked(page))
1167 split_page(virt_to_page(page[0].shadow), order);
1170 for (i = 1; i < (1 << order); i++)
1171 set_page_refcounted(page + i);
1175 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1176 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1180 struct page *buffered_rmqueue(struct zone *preferred_zone,
1181 struct zone *zone, int order, gfp_t gfp_flags,
1184 unsigned long flags;
1186 int cold = !!(gfp_flags & __GFP_COLD);
1191 if (likely(order == 0)) {
1192 struct per_cpu_pages *pcp;
1193 struct list_head *list;
1195 pcp = &zone_pcp(zone, cpu)->pcp;
1196 list = &pcp->lists[migratetype];
1197 local_irq_save(flags);
1198 if (list_empty(list)) {
1199 pcp->count += rmqueue_bulk(zone, 0,
1202 if (unlikely(list_empty(list)))
1207 page = list_entry(list->prev, struct page, lru);
1209 page = list_entry(list->next, struct page, lru);
1211 list_del(&page->lru);
1214 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1216 * __GFP_NOFAIL is not to be used in new code.
1218 * All __GFP_NOFAIL callers should be fixed so that they
1219 * properly detect and handle allocation failures.
1221 * We most definitely don't want callers attempting to
1222 * allocate greater than order-1 page units with
1225 WARN_ON_ONCE(order > 1);
1227 spin_lock_irqsave(&zone->lock, flags);
1228 page = __rmqueue(zone, order, migratetype);
1229 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1230 spin_unlock(&zone->lock);
1235 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1236 zone_statistics(preferred_zone, zone);
1237 local_irq_restore(flags);
1240 VM_BUG_ON(bad_range(zone, page));
1241 if (prep_new_page(page, order, gfp_flags))
1246 local_irq_restore(flags);
1251 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1252 #define ALLOC_WMARK_MIN WMARK_MIN
1253 #define ALLOC_WMARK_LOW WMARK_LOW
1254 #define ALLOC_WMARK_HIGH WMARK_HIGH
1255 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1257 /* Mask to get the watermark bits */
1258 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1260 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1261 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1262 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1264 #ifdef CONFIG_FAIL_PAGE_ALLOC
1266 static struct fail_page_alloc_attr {
1267 struct fault_attr attr;
1269 u32 ignore_gfp_highmem;
1270 u32 ignore_gfp_wait;
1273 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1275 struct dentry *ignore_gfp_highmem_file;
1276 struct dentry *ignore_gfp_wait_file;
1277 struct dentry *min_order_file;
1279 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1281 } fail_page_alloc = {
1282 .attr = FAULT_ATTR_INITIALIZER,
1283 .ignore_gfp_wait = 1,
1284 .ignore_gfp_highmem = 1,
1288 static int __init setup_fail_page_alloc(char *str)
1290 return setup_fault_attr(&fail_page_alloc.attr, str);
1292 __setup("fail_page_alloc=", setup_fail_page_alloc);
1294 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1296 if (order < fail_page_alloc.min_order)
1298 if (gfp_mask & __GFP_NOFAIL)
1300 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1302 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1305 return should_fail(&fail_page_alloc.attr, 1 << order);
1308 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1310 static int __init fail_page_alloc_debugfs(void)
1312 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1316 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1320 dir = fail_page_alloc.attr.dentries.dir;
1322 fail_page_alloc.ignore_gfp_wait_file =
1323 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1324 &fail_page_alloc.ignore_gfp_wait);
1326 fail_page_alloc.ignore_gfp_highmem_file =
1327 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1328 &fail_page_alloc.ignore_gfp_highmem);
1329 fail_page_alloc.min_order_file =
1330 debugfs_create_u32("min-order", mode, dir,
1331 &fail_page_alloc.min_order);
1333 if (!fail_page_alloc.ignore_gfp_wait_file ||
1334 !fail_page_alloc.ignore_gfp_highmem_file ||
1335 !fail_page_alloc.min_order_file) {
1337 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1338 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1339 debugfs_remove(fail_page_alloc.min_order_file);
1340 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1346 late_initcall(fail_page_alloc_debugfs);
1348 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1350 #else /* CONFIG_FAIL_PAGE_ALLOC */
1352 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1357 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1360 * Return 1 if free pages are above 'mark'. This takes into account the order
1361 * of the allocation.
1363 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1364 int classzone_idx, int alloc_flags)
1366 /* free_pages my go negative - that's OK */
1368 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
1371 if (alloc_flags & ALLOC_HIGH)
1373 if (alloc_flags & ALLOC_HARDER)
1376 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1378 for (o = 0; o < order; o++) {
1379 /* At the next order, this order's pages become unavailable */
1380 free_pages -= z->free_area[o].nr_free << o;
1382 /* Require fewer higher order pages to be free */
1383 min >>= min_free_order_shift;
1385 if (free_pages <= min)
1393 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1394 * skip over zones that are not allowed by the cpuset, or that have
1395 * been recently (in last second) found to be nearly full. See further
1396 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1397 * that have to skip over a lot of full or unallowed zones.
1399 * If the zonelist cache is present in the passed in zonelist, then
1400 * returns a pointer to the allowed node mask (either the current
1401 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1403 * If the zonelist cache is not available for this zonelist, does
1404 * nothing and returns NULL.
1406 * If the fullzones BITMAP in the zonelist cache is stale (more than
1407 * a second since last zap'd) then we zap it out (clear its bits.)
1409 * We hold off even calling zlc_setup, until after we've checked the
1410 * first zone in the zonelist, on the theory that most allocations will
1411 * be satisfied from that first zone, so best to examine that zone as
1412 * quickly as we can.
1414 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1416 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1417 nodemask_t *allowednodes; /* zonelist_cache approximation */
1419 zlc = zonelist->zlcache_ptr;
1423 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1424 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1425 zlc->last_full_zap = jiffies;
1428 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1429 &cpuset_current_mems_allowed :
1430 &node_states[N_HIGH_MEMORY];
1431 return allowednodes;
1435 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1436 * if it is worth looking at further for free memory:
1437 * 1) Check that the zone isn't thought to be full (doesn't have its
1438 * bit set in the zonelist_cache fullzones BITMAP).
1439 * 2) Check that the zones node (obtained from the zonelist_cache
1440 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1441 * Return true (non-zero) if zone is worth looking at further, or
1442 * else return false (zero) if it is not.
1444 * This check -ignores- the distinction between various watermarks,
1445 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1446 * found to be full for any variation of these watermarks, it will
1447 * be considered full for up to one second by all requests, unless
1448 * we are so low on memory on all allowed nodes that we are forced
1449 * into the second scan of the zonelist.
1451 * In the second scan we ignore this zonelist cache and exactly
1452 * apply the watermarks to all zones, even it is slower to do so.
1453 * We are low on memory in the second scan, and should leave no stone
1454 * unturned looking for a free page.
1456 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1457 nodemask_t *allowednodes)
1459 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1460 int i; /* index of *z in zonelist zones */
1461 int n; /* node that zone *z is on */
1463 zlc = zonelist->zlcache_ptr;
1467 i = z - zonelist->_zonerefs;
1470 /* This zone is worth trying if it is allowed but not full */
1471 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1475 * Given 'z' scanning a zonelist, set the corresponding bit in
1476 * zlc->fullzones, so that subsequent attempts to allocate a page
1477 * from that zone don't waste time re-examining it.
1479 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1481 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1482 int i; /* index of *z in zonelist zones */
1484 zlc = zonelist->zlcache_ptr;
1488 i = z - zonelist->_zonerefs;
1490 set_bit(i, zlc->fullzones);
1493 #else /* CONFIG_NUMA */
1495 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1500 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1501 nodemask_t *allowednodes)
1506 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1509 #endif /* CONFIG_NUMA */
1512 * get_page_from_freelist goes through the zonelist trying to allocate
1515 static struct page *
1516 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1517 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1518 struct zone *preferred_zone, int migratetype)
1521 struct page *page = NULL;
1524 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1525 int zlc_active = 0; /* set if using zonelist_cache */
1526 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1528 classzone_idx = zone_idx(preferred_zone);
1531 * Scan zonelist, looking for a zone with enough free.
1532 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1534 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1535 high_zoneidx, nodemask) {
1536 if (NUMA_BUILD && zlc_active &&
1537 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1539 if ((alloc_flags & ALLOC_CPUSET) &&
1540 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1543 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1544 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1548 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1549 if (zone_watermark_ok(zone, order, mark,
1550 classzone_idx, alloc_flags))
1553 if (zone_reclaim_mode == 0)
1554 goto this_zone_full;
1556 ret = zone_reclaim(zone, gfp_mask, order);
1558 case ZONE_RECLAIM_NOSCAN:
1561 case ZONE_RECLAIM_FULL:
1562 /* scanned but unreclaimable */
1563 goto this_zone_full;
1565 /* did we reclaim enough */
1566 if (!zone_watermark_ok(zone, order, mark,
1567 classzone_idx, alloc_flags))
1568 goto this_zone_full;
1573 page = buffered_rmqueue(preferred_zone, zone, order,
1574 gfp_mask, migratetype);
1579 zlc_mark_zone_full(zonelist, z);
1581 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1583 * we do zlc_setup after the first zone is tried but only
1584 * if there are multiple nodes make it worthwhile
1586 allowednodes = zlc_setup(zonelist, alloc_flags);
1592 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1593 /* Disable zlc cache for second zonelist scan */
1601 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1602 unsigned long pages_reclaimed)
1604 /* Do not loop if specifically requested */
1605 if (gfp_mask & __GFP_NORETRY)
1609 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1610 * means __GFP_NOFAIL, but that may not be true in other
1613 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1617 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1618 * specified, then we retry until we no longer reclaim any pages
1619 * (above), or we've reclaimed an order of pages at least as
1620 * large as the allocation's order. In both cases, if the
1621 * allocation still fails, we stop retrying.
1623 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1627 * Don't let big-order allocations loop unless the caller
1628 * explicitly requests that.
1630 if (gfp_mask & __GFP_NOFAIL)
1636 static inline struct page *
1637 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1638 struct zonelist *zonelist, enum zone_type high_zoneidx,
1639 nodemask_t *nodemask, struct zone *preferred_zone,
1644 /* Acquire the OOM killer lock for the zones in zonelist */
1645 if (!try_set_zone_oom(zonelist, gfp_mask)) {
1646 schedule_timeout_uninterruptible(1);
1651 * Go through the zonelist yet one more time, keep very high watermark
1652 * here, this is only to catch a parallel oom killing, we must fail if
1653 * we're still under heavy pressure.
1655 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1656 order, zonelist, high_zoneidx,
1657 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1658 preferred_zone, migratetype);
1662 /* The OOM killer will not help higher order allocs */
1663 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_NOFAIL))
1666 /* Exhausted what can be done so it's blamo time */
1667 out_of_memory(zonelist, gfp_mask, order);
1670 clear_zonelist_oom(zonelist, gfp_mask);
1674 /* The really slow allocator path where we enter direct reclaim */
1675 static inline struct page *
1676 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1677 struct zonelist *zonelist, enum zone_type high_zoneidx,
1678 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1679 int migratetype, unsigned long *did_some_progress)
1681 struct page *page = NULL;
1682 struct reclaim_state reclaim_state;
1683 struct task_struct *p = current;
1687 /* We now go into synchronous reclaim */
1688 cpuset_memory_pressure_bump();
1689 p->flags |= PF_MEMALLOC;
1690 lockdep_set_current_reclaim_state(gfp_mask);
1691 reclaim_state.reclaimed_slab = 0;
1692 p->reclaim_state = &reclaim_state;
1694 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1696 p->reclaim_state = NULL;
1697 lockdep_clear_current_reclaim_state();
1698 p->flags &= ~PF_MEMALLOC;
1705 if (likely(*did_some_progress))
1706 page = get_page_from_freelist(gfp_mask, nodemask, order,
1707 zonelist, high_zoneidx,
1708 alloc_flags, preferred_zone,
1714 * This is called in the allocator slow-path if the allocation request is of
1715 * sufficient urgency to ignore watermarks and take other desperate measures
1717 static inline struct page *
1718 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1719 struct zonelist *zonelist, enum zone_type high_zoneidx,
1720 nodemask_t *nodemask, struct zone *preferred_zone,
1726 page = get_page_from_freelist(gfp_mask, nodemask, order,
1727 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
1728 preferred_zone, migratetype);
1730 if (!page && gfp_mask & __GFP_NOFAIL)
1731 congestion_wait(BLK_RW_ASYNC, HZ/50);
1732 } while (!page && (gfp_mask & __GFP_NOFAIL));
1738 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
1739 enum zone_type high_zoneidx)
1744 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1745 wakeup_kswapd(zone, order);
1749 gfp_to_alloc_flags(gfp_t gfp_mask)
1751 struct task_struct *p = current;
1752 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
1753 const gfp_t wait = gfp_mask & __GFP_WAIT;
1755 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1756 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH);
1759 * The caller may dip into page reserves a bit more if the caller
1760 * cannot run direct reclaim, or if the caller has realtime scheduling
1761 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1762 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1764 alloc_flags |= (gfp_mask & __GFP_HIGH);
1767 alloc_flags |= ALLOC_HARDER;
1769 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1770 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1772 alloc_flags &= ~ALLOC_CPUSET;
1773 } else if (unlikely(rt_task(p)) && !in_interrupt())
1774 alloc_flags |= ALLOC_HARDER;
1776 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
1777 if (!in_interrupt() &&
1778 ((p->flags & PF_MEMALLOC) ||
1779 unlikely(test_thread_flag(TIF_MEMDIE))))
1780 alloc_flags |= ALLOC_NO_WATERMARKS;
1786 static inline struct page *
1787 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
1788 struct zonelist *zonelist, enum zone_type high_zoneidx,
1789 nodemask_t *nodemask, struct zone *preferred_zone,
1792 const gfp_t wait = gfp_mask & __GFP_WAIT;
1793 struct page *page = NULL;
1795 unsigned long pages_reclaimed = 0;
1796 unsigned long did_some_progress;
1797 struct task_struct *p = current;
1800 * In the slowpath, we sanity check order to avoid ever trying to
1801 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1802 * be using allocators in order of preference for an area that is
1805 if (order >= MAX_ORDER) {
1806 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
1811 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1812 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1813 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1814 * using a larger set of nodes after it has established that the
1815 * allowed per node queues are empty and that nodes are
1818 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1822 wake_all_kswapd(order, zonelist, high_zoneidx);
1825 * OK, we're below the kswapd watermark and have kicked background
1826 * reclaim. Now things get more complex, so set up alloc_flags according
1827 * to how we want to proceed.
1829 alloc_flags = gfp_to_alloc_flags(gfp_mask);
1831 /* This is the last chance, in general, before the goto nopage. */
1832 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
1833 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
1834 preferred_zone, migratetype);
1839 /* Allocate without watermarks if the context allows */
1840 if (alloc_flags & ALLOC_NO_WATERMARKS) {
1841 page = __alloc_pages_high_priority(gfp_mask, order,
1842 zonelist, high_zoneidx, nodemask,
1843 preferred_zone, migratetype);
1848 /* Atomic allocations - we can't balance anything */
1852 /* Avoid recursion of direct reclaim */
1853 if (p->flags & PF_MEMALLOC)
1856 /* Avoid allocations with no watermarks from looping endlessly */
1857 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
1860 /* Try direct reclaim and then allocating */
1861 page = __alloc_pages_direct_reclaim(gfp_mask, order,
1862 zonelist, high_zoneidx,
1864 alloc_flags, preferred_zone,
1865 migratetype, &did_some_progress);
1870 * If we failed to make any progress reclaiming, then we are
1871 * running out of options and have to consider going OOM
1873 if (!did_some_progress) {
1874 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1875 if (oom_killer_disabled)
1877 page = __alloc_pages_may_oom(gfp_mask, order,
1878 zonelist, high_zoneidx,
1879 nodemask, preferred_zone,
1885 * The OOM killer does not trigger for high-order
1886 * ~__GFP_NOFAIL allocations so if no progress is being
1887 * made, there are no other options and retrying is
1890 if (order > PAGE_ALLOC_COSTLY_ORDER &&
1891 !(gfp_mask & __GFP_NOFAIL))
1898 /* Check if we should retry the allocation */
1899 pages_reclaimed += did_some_progress;
1900 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
1901 /* Wait for some write requests to complete then retry */
1902 congestion_wait(BLK_RW_ASYNC, HZ/50);
1907 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1908 printk(KERN_WARNING "%s: page allocation failure."
1909 " order:%d, mode:0x%x\n",
1910 p->comm, order, gfp_mask);
1916 if (kmemcheck_enabled)
1917 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
1923 * This is the 'heart' of the zoned buddy allocator.
1926 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
1927 struct zonelist *zonelist, nodemask_t *nodemask)
1929 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1930 struct zone *preferred_zone;
1932 int migratetype = allocflags_to_migratetype(gfp_mask);
1934 gfp_mask &= gfp_allowed_mask;
1936 lockdep_trace_alloc(gfp_mask);
1938 might_sleep_if(gfp_mask & __GFP_WAIT);
1940 if (should_fail_alloc_page(gfp_mask, order))
1944 * Check the zones suitable for the gfp_mask contain at least one
1945 * valid zone. It's possible to have an empty zonelist as a result
1946 * of GFP_THISNODE and a memoryless node
1948 if (unlikely(!zonelist->_zonerefs->zone))
1951 /* The preferred zone is used for statistics later */
1952 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone);
1953 if (!preferred_zone)
1956 /* First allocation attempt */
1957 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
1958 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
1959 preferred_zone, migratetype);
1960 if (unlikely(!page))
1961 page = __alloc_pages_slowpath(gfp_mask, order,
1962 zonelist, high_zoneidx, nodemask,
1963 preferred_zone, migratetype);
1965 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
1968 EXPORT_SYMBOL(__alloc_pages_nodemask);
1971 * Common helper functions.
1973 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1978 * __get_free_pages() returns a 32-bit address, which cannot represent
1981 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1983 page = alloc_pages(gfp_mask, order);
1986 return (unsigned long) page_address(page);
1988 EXPORT_SYMBOL(__get_free_pages);
1990 unsigned long get_zeroed_page(gfp_t gfp_mask)
1992 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
1994 EXPORT_SYMBOL(get_zeroed_page);
1996 void __pagevec_free(struct pagevec *pvec)
1998 int i = pagevec_count(pvec);
2001 trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
2002 free_hot_cold_page(pvec->pages[i], pvec->cold);
2006 void __free_pages(struct page *page, unsigned int order)
2008 if (put_page_testzero(page)) {
2009 trace_mm_page_free_direct(page, order);
2011 free_hot_page(page);
2013 __free_pages_ok(page, order);
2017 EXPORT_SYMBOL(__free_pages);
2019 void free_pages(unsigned long addr, unsigned int order)
2022 VM_BUG_ON(!virt_addr_valid((void *)addr));
2023 __free_pages(virt_to_page((void *)addr), order);
2027 EXPORT_SYMBOL(free_pages);
2030 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2031 * @size: the number of bytes to allocate
2032 * @gfp_mask: GFP flags for the allocation
2034 * This function is similar to alloc_pages(), except that it allocates the
2035 * minimum number of pages to satisfy the request. alloc_pages() can only
2036 * allocate memory in power-of-two pages.
2038 * This function is also limited by MAX_ORDER.
2040 * Memory allocated by this function must be released by free_pages_exact().
2042 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2044 unsigned int order = get_order(size);
2047 addr = __get_free_pages(gfp_mask, order);
2049 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2050 unsigned long used = addr + PAGE_ALIGN(size);
2052 split_page(virt_to_page((void *)addr), order);
2053 while (used < alloc_end) {
2059 return (void *)addr;
2061 EXPORT_SYMBOL(alloc_pages_exact);
2064 * free_pages_exact - release memory allocated via alloc_pages_exact()
2065 * @virt: the value returned by alloc_pages_exact.
2066 * @size: size of allocation, same value as passed to alloc_pages_exact().
2068 * Release the memory allocated by a previous call to alloc_pages_exact.
2070 void free_pages_exact(void *virt, size_t size)
2072 unsigned long addr = (unsigned long)virt;
2073 unsigned long end = addr + PAGE_ALIGN(size);
2075 while (addr < end) {
2080 EXPORT_SYMBOL(free_pages_exact);
2082 static unsigned int nr_free_zone_pages(int offset)
2087 /* Just pick one node, since fallback list is circular */
2088 unsigned int sum = 0;
2090 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2092 for_each_zone_zonelist(zone, z, zonelist, offset) {
2093 unsigned long size = zone->present_pages;
2094 unsigned long high = high_wmark_pages(zone);
2103 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2105 unsigned int nr_free_buffer_pages(void)
2107 return nr_free_zone_pages(gfp_zone(GFP_USER));
2109 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2112 * Amount of free RAM allocatable within all zones
2114 unsigned int nr_free_pagecache_pages(void)
2116 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2119 static inline void show_node(struct zone *zone)
2122 printk("Node %d ", zone_to_nid(zone));
2125 void si_meminfo(struct sysinfo *val)
2127 val->totalram = totalram_pages;
2129 val->freeram = global_page_state(NR_FREE_PAGES);
2130 val->bufferram = nr_blockdev_pages();
2131 val->totalhigh = totalhigh_pages;
2132 val->freehigh = nr_free_highpages();
2133 val->mem_unit = PAGE_SIZE;
2136 EXPORT_SYMBOL(si_meminfo);
2139 void si_meminfo_node(struct sysinfo *val, int nid)
2141 pg_data_t *pgdat = NODE_DATA(nid);
2143 val->totalram = pgdat->node_present_pages;
2144 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2145 #ifdef CONFIG_HIGHMEM
2146 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2147 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2153 val->mem_unit = PAGE_SIZE;
2157 #define K(x) ((x) << (PAGE_SHIFT-10))
2160 * Show free area list (used inside shift_scroll-lock stuff)
2161 * We also calculate the percentage fragmentation. We do this by counting the
2162 * memory on each free list with the exception of the first item on the list.
2164 void show_free_areas(void)
2169 for_each_populated_zone(zone) {
2171 printk("%s per-cpu:\n", zone->name);
2173 for_each_online_cpu(cpu) {
2174 struct per_cpu_pageset *pageset;
2176 pageset = zone_pcp(zone, cpu);
2178 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2179 cpu, pageset->pcp.high,
2180 pageset->pcp.batch, pageset->pcp.count);
2184 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2185 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2187 " dirty:%lu writeback:%lu unstable:%lu\n"
2188 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2189 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2190 global_page_state(NR_ACTIVE_ANON),
2191 global_page_state(NR_INACTIVE_ANON),
2192 global_page_state(NR_ISOLATED_ANON),
2193 global_page_state(NR_ACTIVE_FILE),
2194 global_page_state(NR_INACTIVE_FILE),
2195 global_page_state(NR_ISOLATED_FILE),
2196 global_page_state(NR_UNEVICTABLE),
2197 global_page_state(NR_FILE_DIRTY),
2198 global_page_state(NR_WRITEBACK),
2199 global_page_state(NR_UNSTABLE_NFS),
2200 global_page_state(NR_FREE_PAGES),
2201 global_page_state(NR_SLAB_RECLAIMABLE),
2202 global_page_state(NR_SLAB_UNRECLAIMABLE),
2203 global_page_state(NR_FILE_MAPPED),
2204 global_page_state(NR_SHMEM),
2205 global_page_state(NR_PAGETABLE),
2206 global_page_state(NR_BOUNCE));
2208 for_each_populated_zone(zone) {
2217 " active_anon:%lukB"
2218 " inactive_anon:%lukB"
2219 " active_file:%lukB"
2220 " inactive_file:%lukB"
2221 " unevictable:%lukB"
2222 " isolated(anon):%lukB"
2223 " isolated(file):%lukB"
2230 " slab_reclaimable:%lukB"
2231 " slab_unreclaimable:%lukB"
2232 " kernel_stack:%lukB"
2236 " writeback_tmp:%lukB"
2237 " pages_scanned:%lu"
2238 " all_unreclaimable? %s"
2241 K(zone_page_state(zone, NR_FREE_PAGES)),
2242 K(min_wmark_pages(zone)),
2243 K(low_wmark_pages(zone)),
2244 K(high_wmark_pages(zone)),
2245 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2246 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2247 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2248 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2249 K(zone_page_state(zone, NR_UNEVICTABLE)),
2250 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2251 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2252 K(zone->present_pages),
2253 K(zone_page_state(zone, NR_MLOCK)),
2254 K(zone_page_state(zone, NR_FILE_DIRTY)),
2255 K(zone_page_state(zone, NR_WRITEBACK)),
2256 K(zone_page_state(zone, NR_FILE_MAPPED)),
2257 K(zone_page_state(zone, NR_SHMEM)),
2258 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2259 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2260 zone_page_state(zone, NR_KERNEL_STACK) *
2262 K(zone_page_state(zone, NR_PAGETABLE)),
2263 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2264 K(zone_page_state(zone, NR_BOUNCE)),
2265 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2266 zone->pages_scanned,
2267 (zone_is_all_unreclaimable(zone) ? "yes" : "no")
2269 printk("lowmem_reserve[]:");
2270 for (i = 0; i < MAX_NR_ZONES; i++)
2271 printk(" %lu", zone->lowmem_reserve[i]);
2275 for_each_populated_zone(zone) {
2276 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2279 printk("%s: ", zone->name);
2281 spin_lock_irqsave(&zone->lock, flags);
2282 for (order = 0; order < MAX_ORDER; order++) {
2283 nr[order] = zone->free_area[order].nr_free;
2284 total += nr[order] << order;
2286 spin_unlock_irqrestore(&zone->lock, flags);
2287 for (order = 0; order < MAX_ORDER; order++)
2288 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2289 printk("= %lukB\n", K(total));
2292 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2294 show_swap_cache_info();
2297 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2299 zoneref->zone = zone;
2300 zoneref->zone_idx = zone_idx(zone);
2304 * Builds allocation fallback zone lists.
2306 * Add all populated zones of a node to the zonelist.
2308 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2309 int nr_zones, enum zone_type zone_type)
2313 BUG_ON(zone_type >= MAX_NR_ZONES);
2318 zone = pgdat->node_zones + zone_type;
2319 if (populated_zone(zone)) {
2320 zoneref_set_zone(zone,
2321 &zonelist->_zonerefs[nr_zones++]);
2322 check_highest_zone(zone_type);
2325 } while (zone_type);
2332 * 0 = automatic detection of better ordering.
2333 * 1 = order by ([node] distance, -zonetype)
2334 * 2 = order by (-zonetype, [node] distance)
2336 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2337 * the same zonelist. So only NUMA can configure this param.
2339 #define ZONELIST_ORDER_DEFAULT 0
2340 #define ZONELIST_ORDER_NODE 1
2341 #define ZONELIST_ORDER_ZONE 2
2343 /* zonelist order in the kernel.
2344 * set_zonelist_order() will set this to NODE or ZONE.
2346 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2347 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2351 /* The value user specified ....changed by config */
2352 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2353 /* string for sysctl */
2354 #define NUMA_ZONELIST_ORDER_LEN 16
2355 char numa_zonelist_order[16] = "default";
2358 * interface for configure zonelist ordering.
2359 * command line option "numa_zonelist_order"
2360 * = "[dD]efault - default, automatic configuration.
2361 * = "[nN]ode - order by node locality, then by zone within node
2362 * = "[zZ]one - order by zone, then by locality within zone
2365 static int __parse_numa_zonelist_order(char *s)
2367 if (*s == 'd' || *s == 'D') {
2368 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2369 } else if (*s == 'n' || *s == 'N') {
2370 user_zonelist_order = ZONELIST_ORDER_NODE;
2371 } else if (*s == 'z' || *s == 'Z') {
2372 user_zonelist_order = ZONELIST_ORDER_ZONE;
2375 "Ignoring invalid numa_zonelist_order value: "
2382 static __init int setup_numa_zonelist_order(char *s)
2385 return __parse_numa_zonelist_order(s);
2388 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2391 * sysctl handler for numa_zonelist_order
2393 int numa_zonelist_order_handler(ctl_table *table, int write,
2394 void __user *buffer, size_t *length,
2397 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2401 strncpy(saved_string, (char*)table->data,
2402 NUMA_ZONELIST_ORDER_LEN);
2403 ret = proc_dostring(table, write, buffer, length, ppos);
2407 int oldval = user_zonelist_order;
2408 if (__parse_numa_zonelist_order((char*)table->data)) {
2410 * bogus value. restore saved string
2412 strncpy((char*)table->data, saved_string,
2413 NUMA_ZONELIST_ORDER_LEN);
2414 user_zonelist_order = oldval;
2415 } else if (oldval != user_zonelist_order)
2416 build_all_zonelists();
2422 #define MAX_NODE_LOAD (nr_online_nodes)
2423 static int node_load[MAX_NUMNODES];
2426 * find_next_best_node - find the next node that should appear in a given node's fallback list
2427 * @node: node whose fallback list we're appending
2428 * @used_node_mask: nodemask_t of already used nodes
2430 * We use a number of factors to determine which is the next node that should
2431 * appear on a given node's fallback list. The node should not have appeared
2432 * already in @node's fallback list, and it should be the next closest node
2433 * according to the distance array (which contains arbitrary distance values
2434 * from each node to each node in the system), and should also prefer nodes
2435 * with no CPUs, since presumably they'll have very little allocation pressure
2436 * on them otherwise.
2437 * It returns -1 if no node is found.
2439 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2442 int min_val = INT_MAX;
2444 const struct cpumask *tmp = cpumask_of_node(0);
2446 /* Use the local node if we haven't already */
2447 if (!node_isset(node, *used_node_mask)) {
2448 node_set(node, *used_node_mask);
2452 for_each_node_state(n, N_HIGH_MEMORY) {
2454 /* Don't want a node to appear more than once */
2455 if (node_isset(n, *used_node_mask))
2458 /* Use the distance array to find the distance */
2459 val = node_distance(node, n);
2461 /* Penalize nodes under us ("prefer the next node") */
2464 /* Give preference to headless and unused nodes */
2465 tmp = cpumask_of_node(n);
2466 if (!cpumask_empty(tmp))
2467 val += PENALTY_FOR_NODE_WITH_CPUS;
2469 /* Slight preference for less loaded node */
2470 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2471 val += node_load[n];
2473 if (val < min_val) {
2480 node_set(best_node, *used_node_mask);
2487 * Build zonelists ordered by node and zones within node.
2488 * This results in maximum locality--normal zone overflows into local
2489 * DMA zone, if any--but risks exhausting DMA zone.
2491 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2494 struct zonelist *zonelist;
2496 zonelist = &pgdat->node_zonelists[0];
2497 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2499 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2501 zonelist->_zonerefs[j].zone = NULL;
2502 zonelist->_zonerefs[j].zone_idx = 0;
2506 * Build gfp_thisnode zonelists
2508 static void build_thisnode_zonelists(pg_data_t *pgdat)
2511 struct zonelist *zonelist;
2513 zonelist = &pgdat->node_zonelists[1];
2514 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2515 zonelist->_zonerefs[j].zone = NULL;
2516 zonelist->_zonerefs[j].zone_idx = 0;
2520 * Build zonelists ordered by zone and nodes within zones.
2521 * This results in conserving DMA zone[s] until all Normal memory is
2522 * exhausted, but results in overflowing to remote node while memory
2523 * may still exist in local DMA zone.
2525 static int node_order[MAX_NUMNODES];
2527 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2530 int zone_type; /* needs to be signed */
2532 struct zonelist *zonelist;
2534 zonelist = &pgdat->node_zonelists[0];
2536 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2537 for (j = 0; j < nr_nodes; j++) {
2538 node = node_order[j];
2539 z = &NODE_DATA(node)->node_zones[zone_type];
2540 if (populated_zone(z)) {
2542 &zonelist->_zonerefs[pos++]);
2543 check_highest_zone(zone_type);
2547 zonelist->_zonerefs[pos].zone = NULL;
2548 zonelist->_zonerefs[pos].zone_idx = 0;
2551 static int default_zonelist_order(void)
2554 unsigned long low_kmem_size,total_size;
2558 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2559 * If they are really small and used heavily, the system can fall
2560 * into OOM very easily.
2561 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2563 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2566 for_each_online_node(nid) {
2567 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2568 z = &NODE_DATA(nid)->node_zones[zone_type];
2569 if (populated_zone(z)) {
2570 if (zone_type < ZONE_NORMAL)
2571 low_kmem_size += z->present_pages;
2572 total_size += z->present_pages;
2576 if (!low_kmem_size || /* there are no DMA area. */
2577 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2578 return ZONELIST_ORDER_NODE;
2580 * look into each node's config.
2581 * If there is a node whose DMA/DMA32 memory is very big area on
2582 * local memory, NODE_ORDER may be suitable.
2584 average_size = total_size /
2585 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2586 for_each_online_node(nid) {
2589 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2590 z = &NODE_DATA(nid)->node_zones[zone_type];
2591 if (populated_zone(z)) {
2592 if (zone_type < ZONE_NORMAL)
2593 low_kmem_size += z->present_pages;
2594 total_size += z->present_pages;
2597 if (low_kmem_size &&
2598 total_size > average_size && /* ignore small node */
2599 low_kmem_size > total_size * 70/100)
2600 return ZONELIST_ORDER_NODE;
2602 return ZONELIST_ORDER_ZONE;
2605 static void set_zonelist_order(void)
2607 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2608 current_zonelist_order = default_zonelist_order();
2610 current_zonelist_order = user_zonelist_order;
2613 static void build_zonelists(pg_data_t *pgdat)
2617 nodemask_t used_mask;
2618 int local_node, prev_node;
2619 struct zonelist *zonelist;
2620 int order = current_zonelist_order;
2622 /* initialize zonelists */
2623 for (i = 0; i < MAX_ZONELISTS; i++) {
2624 zonelist = pgdat->node_zonelists + i;
2625 zonelist->_zonerefs[0].zone = NULL;
2626 zonelist->_zonerefs[0].zone_idx = 0;
2629 /* NUMA-aware ordering of nodes */
2630 local_node = pgdat->node_id;
2631 load = nr_online_nodes;
2632 prev_node = local_node;
2633 nodes_clear(used_mask);
2635 memset(node_order, 0, sizeof(node_order));
2638 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2639 int distance = node_distance(local_node, node);
2642 * If another node is sufficiently far away then it is better
2643 * to reclaim pages in a zone before going off node.
2645 if (distance > RECLAIM_DISTANCE)
2646 zone_reclaim_mode = 1;
2649 * We don't want to pressure a particular node.
2650 * So adding penalty to the first node in same
2651 * distance group to make it round-robin.
2653 if (distance != node_distance(local_node, prev_node))
2654 node_load[node] = load;
2658 if (order == ZONELIST_ORDER_NODE)
2659 build_zonelists_in_node_order(pgdat, node);
2661 node_order[j++] = node; /* remember order */
2664 if (order == ZONELIST_ORDER_ZONE) {
2665 /* calculate node order -- i.e., DMA last! */
2666 build_zonelists_in_zone_order(pgdat, j);
2669 build_thisnode_zonelists(pgdat);
2672 /* Construct the zonelist performance cache - see further mmzone.h */
2673 static void build_zonelist_cache(pg_data_t *pgdat)
2675 struct zonelist *zonelist;
2676 struct zonelist_cache *zlc;
2679 zonelist = &pgdat->node_zonelists[0];
2680 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2681 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2682 for (z = zonelist->_zonerefs; z->zone; z++)
2683 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2687 #else /* CONFIG_NUMA */
2689 static void set_zonelist_order(void)
2691 current_zonelist_order = ZONELIST_ORDER_ZONE;
2694 static void build_zonelists(pg_data_t *pgdat)
2696 int node, local_node;
2698 struct zonelist *zonelist;
2700 local_node = pgdat->node_id;
2702 zonelist = &pgdat->node_zonelists[0];
2703 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2706 * Now we build the zonelist so that it contains the zones
2707 * of all the other nodes.
2708 * We don't want to pressure a particular node, so when
2709 * building the zones for node N, we make sure that the
2710 * zones coming right after the local ones are those from
2711 * node N+1 (modulo N)
2713 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2714 if (!node_online(node))
2716 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2719 for (node = 0; node < local_node; node++) {
2720 if (!node_online(node))
2722 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2726 zonelist->_zonerefs[j].zone = NULL;
2727 zonelist->_zonerefs[j].zone_idx = 0;
2730 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2731 static void build_zonelist_cache(pg_data_t *pgdat)
2733 pgdat->node_zonelists[0].zlcache_ptr = NULL;
2736 #endif /* CONFIG_NUMA */
2738 /* return values int ....just for stop_machine() */
2739 static int __build_all_zonelists(void *dummy)
2744 memset(node_load, 0, sizeof(node_load));
2746 for_each_online_node(nid) {
2747 pg_data_t *pgdat = NODE_DATA(nid);
2749 build_zonelists(pgdat);
2750 build_zonelist_cache(pgdat);
2755 void build_all_zonelists(void)
2757 set_zonelist_order();
2759 if (system_state == SYSTEM_BOOTING) {
2760 __build_all_zonelists(NULL);
2761 mminit_verify_zonelist();
2762 cpuset_init_current_mems_allowed();
2764 /* we have to stop all cpus to guarantee there is no user
2766 stop_machine(__build_all_zonelists, NULL, NULL);
2767 /* cpuset refresh routine should be here */
2769 vm_total_pages = nr_free_pagecache_pages();
2771 * Disable grouping by mobility if the number of pages in the
2772 * system is too low to allow the mechanism to work. It would be
2773 * more accurate, but expensive to check per-zone. This check is
2774 * made on memory-hotadd so a system can start with mobility
2775 * disabled and enable it later
2777 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
2778 page_group_by_mobility_disabled = 1;
2780 page_group_by_mobility_disabled = 0;
2782 printk("Built %i zonelists in %s order, mobility grouping %s. "
2783 "Total pages: %ld\n",
2785 zonelist_order_name[current_zonelist_order],
2786 page_group_by_mobility_disabled ? "off" : "on",
2789 printk("Policy zone: %s\n", zone_names[policy_zone]);
2794 * Helper functions to size the waitqueue hash table.
2795 * Essentially these want to choose hash table sizes sufficiently
2796 * large so that collisions trying to wait on pages are rare.
2797 * But in fact, the number of active page waitqueues on typical
2798 * systems is ridiculously low, less than 200. So this is even
2799 * conservative, even though it seems large.
2801 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2802 * waitqueues, i.e. the size of the waitq table given the number of pages.
2804 #define PAGES_PER_WAITQUEUE 256
2806 #ifndef CONFIG_MEMORY_HOTPLUG
2807 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2809 unsigned long size = 1;
2811 pages /= PAGES_PER_WAITQUEUE;
2813 while (size < pages)
2817 * Once we have dozens or even hundreds of threads sleeping
2818 * on IO we've got bigger problems than wait queue collision.
2819 * Limit the size of the wait table to a reasonable size.
2821 size = min(size, 4096UL);
2823 return max(size, 4UL);
2827 * A zone's size might be changed by hot-add, so it is not possible to determine
2828 * a suitable size for its wait_table. So we use the maximum size now.
2830 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2832 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2833 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2834 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2836 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2837 * or more by the traditional way. (See above). It equals:
2839 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2840 * ia64(16K page size) : = ( 8G + 4M)byte.
2841 * powerpc (64K page size) : = (32G +16M)byte.
2843 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2850 * This is an integer logarithm so that shifts can be used later
2851 * to extract the more random high bits from the multiplicative
2852 * hash function before the remainder is taken.
2854 static inline unsigned long wait_table_bits(unsigned long size)
2859 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2862 * Check if a pageblock contains reserved pages
2864 static int pageblock_is_reserved(unsigned long start_pfn)
2866 unsigned long end_pfn = start_pfn + pageblock_nr_pages;
2869 for (pfn = start_pfn; pfn < end_pfn; pfn++)
2870 if (PageReserved(pfn_to_page(pfn)))
2876 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2877 * of blocks reserved is based on min_wmark_pages(zone). The memory within
2878 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
2879 * higher will lead to a bigger reserve which will get freed as contiguous
2880 * blocks as reclaim kicks in
2882 static void setup_zone_migrate_reserve(struct zone *zone)
2884 unsigned long start_pfn, pfn, end_pfn;
2886 unsigned long block_migratetype;
2889 /* Get the start pfn, end pfn and the number of blocks to reserve */
2890 start_pfn = zone->zone_start_pfn;
2891 end_pfn = start_pfn + zone->spanned_pages;
2892 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
2896 * Reserve blocks are generally in place to help high-order atomic
2897 * allocations that are short-lived. A min_free_kbytes value that
2898 * would result in more than 2 reserve blocks for atomic allocations
2899 * is assumed to be in place to help anti-fragmentation for the
2900 * future allocation of hugepages at runtime.
2902 reserve = min(2, reserve);
2904 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
2905 if (!pfn_valid(pfn))
2907 page = pfn_to_page(pfn);
2909 /* Watch out for overlapping nodes */
2910 if (page_to_nid(page) != zone_to_nid(zone))
2913 /* Blocks with reserved pages will never free, skip them. */
2914 if (pageblock_is_reserved(pfn))
2917 block_migratetype = get_pageblock_migratetype(page);
2919 /* If this block is reserved, account for it */
2920 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
2925 /* Suitable for reserving if this block is movable */
2926 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
2927 set_pageblock_migratetype(page, MIGRATE_RESERVE);
2928 move_freepages_block(zone, page, MIGRATE_RESERVE);
2934 * If the reserve is met and this is a previous reserved block,
2937 if (block_migratetype == MIGRATE_RESERVE) {
2938 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2939 move_freepages_block(zone, page, MIGRATE_MOVABLE);
2945 * Initially all pages are reserved - free ones are freed
2946 * up by free_all_bootmem() once the early boot process is
2947 * done. Non-atomic initialization, single-pass.
2949 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
2950 unsigned long start_pfn, enum memmap_context context)
2953 unsigned long end_pfn = start_pfn + size;
2957 if (highest_memmap_pfn < end_pfn - 1)
2958 highest_memmap_pfn = end_pfn - 1;
2960 z = &NODE_DATA(nid)->node_zones[zone];
2961 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
2963 * There can be holes in boot-time mem_map[]s
2964 * handed to this function. They do not
2965 * exist on hotplugged memory.
2967 if (context == MEMMAP_EARLY) {
2968 if (!early_pfn_valid(pfn))
2970 if (!early_pfn_in_nid(pfn, nid))
2973 page = pfn_to_page(pfn);
2974 set_page_links(page, zone, nid, pfn);
2975 mminit_verify_page_links(page, zone, nid, pfn);
2976 init_page_count(page);
2977 reset_page_mapcount(page);
2978 SetPageReserved(page);
2980 * Mark the block movable so that blocks are reserved for
2981 * movable at startup. This will force kernel allocations
2982 * to reserve their blocks rather than leaking throughout
2983 * the address space during boot when many long-lived
2984 * kernel allocations are made. Later some blocks near
2985 * the start are marked MIGRATE_RESERVE by
2986 * setup_zone_migrate_reserve()
2988 * bitmap is created for zone's valid pfn range. but memmap
2989 * can be created for invalid pages (for alignment)
2990 * check here not to call set_pageblock_migratetype() against
2993 if ((z->zone_start_pfn <= pfn)
2994 && (pfn < z->zone_start_pfn + z->spanned_pages)
2995 && !(pfn & (pageblock_nr_pages - 1)))
2996 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2998 INIT_LIST_HEAD(&page->lru);
2999 #ifdef WANT_PAGE_VIRTUAL
3000 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3001 if (!is_highmem_idx(zone))
3002 set_page_address(page, __va(pfn << PAGE_SHIFT));
3007 static void __meminit zone_init_free_lists(struct zone *zone)
3010 for_each_migratetype_order(order, t) {
3011 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3012 zone->free_area[order].nr_free = 0;
3016 #ifndef __HAVE_ARCH_MEMMAP_INIT
3017 #define memmap_init(size, nid, zone, start_pfn) \
3018 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3021 static int zone_batchsize(struct zone *zone)
3027 * The per-cpu-pages pools are set to around 1000th of the
3028 * size of the zone. But no more than 1/2 of a meg.
3030 * OK, so we don't know how big the cache is. So guess.
3032 batch = zone->present_pages / 1024;
3033 if (batch * PAGE_SIZE > 512 * 1024)
3034 batch = (512 * 1024) / PAGE_SIZE;
3035 batch /= 4; /* We effectively *= 4 below */
3040 * Clamp the batch to a 2^n - 1 value. Having a power
3041 * of 2 value was found to be more likely to have
3042 * suboptimal cache aliasing properties in some cases.
3044 * For example if 2 tasks are alternately allocating
3045 * batches of pages, one task can end up with a lot
3046 * of pages of one half of the possible page colors
3047 * and the other with pages of the other colors.
3049 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3054 /* The deferral and batching of frees should be suppressed under NOMMU
3057 * The problem is that NOMMU needs to be able to allocate large chunks
3058 * of contiguous memory as there's no hardware page translation to
3059 * assemble apparent contiguous memory from discontiguous pages.
3061 * Queueing large contiguous runs of pages for batching, however,
3062 * causes the pages to actually be freed in smaller chunks. As there
3063 * can be a significant delay between the individual batches being
3064 * recycled, this leads to the once large chunks of space being
3065 * fragmented and becoming unavailable for high-order allocations.
3071 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3073 struct per_cpu_pages *pcp;
3076 memset(p, 0, sizeof(*p));
3080 pcp->high = 6 * batch;
3081 pcp->batch = max(1UL, 1 * batch);
3082 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3083 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3087 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3088 * to the value high for the pageset p.
3091 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3094 struct per_cpu_pages *pcp;
3098 pcp->batch = max(1UL, high/4);
3099 if ((high/4) > (PAGE_SHIFT * 8))
3100 pcp->batch = PAGE_SHIFT * 8;
3106 * Boot pageset table. One per cpu which is going to be used for all
3107 * zones and all nodes. The parameters will be set in such a way
3108 * that an item put on a list will immediately be handed over to
3109 * the buddy list. This is safe since pageset manipulation is done
3110 * with interrupts disabled.
3112 * Some NUMA counter updates may also be caught by the boot pagesets.
3114 * The boot_pagesets must be kept even after bootup is complete for
3115 * unused processors and/or zones. They do play a role for bootstrapping
3116 * hotplugged processors.
3118 * zoneinfo_show() and maybe other functions do
3119 * not check if the processor is online before following the pageset pointer.
3120 * Other parts of the kernel may not check if the zone is available.
3122 static struct per_cpu_pageset boot_pageset[NR_CPUS];
3125 * Dynamically allocate memory for the
3126 * per cpu pageset array in struct zone.
3128 static int __cpuinit process_zones(int cpu)
3130 struct zone *zone, *dzone;
3131 int node = cpu_to_node(cpu);
3133 node_set_state(node, N_CPU); /* this node has a cpu */
3135 for_each_populated_zone(zone) {
3136 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
3138 if (!zone_pcp(zone, cpu))
3141 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
3143 if (percpu_pagelist_fraction)
3144 setup_pagelist_highmark(zone_pcp(zone, cpu),
3145 (zone->present_pages / percpu_pagelist_fraction));
3150 for_each_zone(dzone) {
3151 if (!populated_zone(dzone))
3155 kfree(zone_pcp(dzone, cpu));
3156 zone_pcp(dzone, cpu) = &boot_pageset[cpu];
3161 static inline void free_zone_pagesets(int cpu)
3165 for_each_zone(zone) {
3166 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
3168 /* Free per_cpu_pageset if it is slab allocated */
3169 if (pset != &boot_pageset[cpu])
3171 zone_pcp(zone, cpu) = &boot_pageset[cpu];
3175 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
3176 unsigned long action,
3179 int cpu = (long)hcpu;
3180 int ret = NOTIFY_OK;
3183 case CPU_UP_PREPARE:
3184 case CPU_UP_PREPARE_FROZEN:
3185 if (process_zones(cpu))
3188 case CPU_UP_CANCELED:
3189 case CPU_UP_CANCELED_FROZEN:
3191 case CPU_DEAD_FROZEN:
3192 free_zone_pagesets(cpu);
3200 static struct notifier_block __cpuinitdata pageset_notifier =
3201 { &pageset_cpuup_callback, NULL, 0 };
3203 void __init setup_per_cpu_pageset(void)
3207 /* Initialize per_cpu_pageset for cpu 0.
3208 * A cpuup callback will do this for every cpu
3209 * as it comes online
3211 err = process_zones(smp_processor_id());
3213 register_cpu_notifier(&pageset_notifier);
3218 static noinline __init_refok
3219 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3222 struct pglist_data *pgdat = zone->zone_pgdat;
3226 * The per-page waitqueue mechanism uses hashed waitqueues
3229 zone->wait_table_hash_nr_entries =
3230 wait_table_hash_nr_entries(zone_size_pages);
3231 zone->wait_table_bits =
3232 wait_table_bits(zone->wait_table_hash_nr_entries);
3233 alloc_size = zone->wait_table_hash_nr_entries
3234 * sizeof(wait_queue_head_t);
3236 if (!slab_is_available()) {
3237 zone->wait_table = (wait_queue_head_t *)
3238 alloc_bootmem_node(pgdat, alloc_size);
3241 * This case means that a zone whose size was 0 gets new memory
3242 * via memory hot-add.
3243 * But it may be the case that a new node was hot-added. In
3244 * this case vmalloc() will not be able to use this new node's
3245 * memory - this wait_table must be initialized to use this new
3246 * node itself as well.
3247 * To use this new node's memory, further consideration will be
3250 zone->wait_table = vmalloc(alloc_size);
3252 if (!zone->wait_table)
3255 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3256 init_waitqueue_head(zone->wait_table + i);
3261 static int __zone_pcp_update(void *data)
3263 struct zone *zone = data;
3265 unsigned long batch = zone_batchsize(zone), flags;
3267 for (cpu = 0; cpu < NR_CPUS; cpu++) {
3268 struct per_cpu_pageset *pset;
3269 struct per_cpu_pages *pcp;
3271 pset = zone_pcp(zone, cpu);
3274 local_irq_save(flags);
3275 free_pcppages_bulk(zone, pcp->count, pcp);
3276 setup_pageset(pset, batch);
3277 local_irq_restore(flags);
3282 void zone_pcp_update(struct zone *zone)
3284 stop_machine(__zone_pcp_update, zone, NULL);
3287 static __meminit void zone_pcp_init(struct zone *zone)
3290 unsigned long batch = zone_batchsize(zone);
3292 for (cpu = 0; cpu < NR_CPUS; cpu++) {
3294 /* Early boot. Slab allocator not functional yet */
3295 zone_pcp(zone, cpu) = &boot_pageset[cpu];
3296 setup_pageset(&boot_pageset[cpu],0);
3298 setup_pageset(zone_pcp(zone,cpu), batch);
3301 if (zone->present_pages)
3302 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
3303 zone->name, zone->present_pages, batch);
3306 __meminit int init_currently_empty_zone(struct zone *zone,
3307 unsigned long zone_start_pfn,
3309 enum memmap_context context)
3311 struct pglist_data *pgdat = zone->zone_pgdat;
3313 ret = zone_wait_table_init(zone, size);
3316 pgdat->nr_zones = zone_idx(zone) + 1;
3318 zone->zone_start_pfn = zone_start_pfn;
3320 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3321 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3323 (unsigned long)zone_idx(zone),
3324 zone_start_pfn, (zone_start_pfn + size));
3326 zone_init_free_lists(zone);
3331 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3333 * Basic iterator support. Return the first range of PFNs for a node
3334 * Note: nid == MAX_NUMNODES returns first region regardless of node
3336 static int __meminit first_active_region_index_in_nid(int nid)
3340 for (i = 0; i < nr_nodemap_entries; i++)
3341 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3348 * Basic iterator support. Return the next active range of PFNs for a node
3349 * Note: nid == MAX_NUMNODES returns next region regardless of node
3351 static int __meminit next_active_region_index_in_nid(int index, int nid)
3353 for (index = index + 1; index < nr_nodemap_entries; index++)
3354 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3360 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3362 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3363 * Architectures may implement their own version but if add_active_range()
3364 * was used and there are no special requirements, this is a convenient
3367 int __meminit __early_pfn_to_nid(unsigned long pfn)
3371 for (i = 0; i < nr_nodemap_entries; i++) {
3372 unsigned long start_pfn = early_node_map[i].start_pfn;
3373 unsigned long end_pfn = early_node_map[i].end_pfn;
3375 if (start_pfn <= pfn && pfn < end_pfn)
3376 return early_node_map[i].nid;
3378 /* This is a memory hole */
3381 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3383 int __meminit early_pfn_to_nid(unsigned long pfn)
3387 nid = __early_pfn_to_nid(pfn);
3390 /* just returns 0 */
3394 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3395 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3399 nid = __early_pfn_to_nid(pfn);
3400 if (nid >= 0 && nid != node)
3406 /* Basic iterator support to walk early_node_map[] */
3407 #define for_each_active_range_index_in_nid(i, nid) \
3408 for (i = first_active_region_index_in_nid(nid); i != -1; \
3409 i = next_active_region_index_in_nid(i, nid))
3412 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3413 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3414 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3416 * If an architecture guarantees that all ranges registered with
3417 * add_active_ranges() contain no holes and may be freed, this
3418 * this function may be used instead of calling free_bootmem() manually.
3420 void __init free_bootmem_with_active_regions(int nid,
3421 unsigned long max_low_pfn)
3425 for_each_active_range_index_in_nid(i, nid) {
3426 unsigned long size_pages = 0;
3427 unsigned long end_pfn = early_node_map[i].end_pfn;
3429 if (early_node_map[i].start_pfn >= max_low_pfn)
3432 if (end_pfn > max_low_pfn)
3433 end_pfn = max_low_pfn;
3435 size_pages = end_pfn - early_node_map[i].start_pfn;
3436 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3437 PFN_PHYS(early_node_map[i].start_pfn),
3438 size_pages << PAGE_SHIFT);
3442 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3447 for_each_active_range_index_in_nid(i, nid) {
3448 ret = work_fn(early_node_map[i].start_pfn,
3449 early_node_map[i].end_pfn, data);
3455 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3456 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3458 * If an architecture guarantees that all ranges registered with
3459 * add_active_ranges() contain no holes and may be freed, this
3460 * function may be used instead of calling memory_present() manually.
3462 void __init sparse_memory_present_with_active_regions(int nid)
3466 for_each_active_range_index_in_nid(i, nid)
3467 memory_present(early_node_map[i].nid,
3468 early_node_map[i].start_pfn,
3469 early_node_map[i].end_pfn);
3473 * get_pfn_range_for_nid - Return the start and end page frames for a node
3474 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3475 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3476 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3478 * It returns the start and end page frame of a node based on information
3479 * provided by an arch calling add_active_range(). If called for a node
3480 * with no available memory, a warning is printed and the start and end
3483 void __meminit get_pfn_range_for_nid(unsigned int nid,
3484 unsigned long *start_pfn, unsigned long *end_pfn)
3490 for_each_active_range_index_in_nid(i, nid) {
3491 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3492 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3495 if (*start_pfn == -1UL)
3500 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3501 * assumption is made that zones within a node are ordered in monotonic
3502 * increasing memory addresses so that the "highest" populated zone is used
3504 static void __init find_usable_zone_for_movable(void)
3507 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3508 if (zone_index == ZONE_MOVABLE)
3511 if (arch_zone_highest_possible_pfn[zone_index] >
3512 arch_zone_lowest_possible_pfn[zone_index])
3516 VM_BUG_ON(zone_index == -1);
3517 movable_zone = zone_index;
3521 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3522 * because it is sized independant of architecture. Unlike the other zones,
3523 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3524 * in each node depending on the size of each node and how evenly kernelcore
3525 * is distributed. This helper function adjusts the zone ranges
3526 * provided by the architecture for a given node by using the end of the
3527 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3528 * zones within a node are in order of monotonic increases memory addresses
3530 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3531 unsigned long zone_type,
3532 unsigned long node_start_pfn,
3533 unsigned long node_end_pfn,
3534 unsigned long *zone_start_pfn,
3535 unsigned long *zone_end_pfn)
3537 /* Only adjust if ZONE_MOVABLE is on this node */
3538 if (zone_movable_pfn[nid]) {
3539 /* Size ZONE_MOVABLE */
3540 if (zone_type == ZONE_MOVABLE) {
3541 *zone_start_pfn = zone_movable_pfn[nid];
3542 *zone_end_pfn = min(node_end_pfn,
3543 arch_zone_highest_possible_pfn[movable_zone]);
3545 /* Adjust for ZONE_MOVABLE starting within this range */
3546 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3547 *zone_end_pfn > zone_movable_pfn[nid]) {
3548 *zone_end_pfn = zone_movable_pfn[nid];
3550 /* Check if this whole range is within ZONE_MOVABLE */
3551 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3552 *zone_start_pfn = *zone_end_pfn;
3557 * Return the number of pages a zone spans in a node, including holes
3558 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3560 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3561 unsigned long zone_type,
3562 unsigned long *ignored)
3564 unsigned long node_start_pfn, node_end_pfn;
3565 unsigned long zone_start_pfn, zone_end_pfn;
3567 /* Get the start and end of the node and zone */
3568 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3569 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3570 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3571 adjust_zone_range_for_zone_movable(nid, zone_type,
3572 node_start_pfn, node_end_pfn,
3573 &zone_start_pfn, &zone_end_pfn);
3575 /* Check that this node has pages within the zone's required range */
3576 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3579 /* Move the zone boundaries inside the node if necessary */
3580 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3581 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3583 /* Return the spanned pages */
3584 return zone_end_pfn - zone_start_pfn;
3588 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3589 * then all holes in the requested range will be accounted for.
3591 static unsigned long __meminit __absent_pages_in_range(int nid,
3592 unsigned long range_start_pfn,
3593 unsigned long range_end_pfn)
3596 unsigned long prev_end_pfn = 0, hole_pages = 0;
3597 unsigned long start_pfn;
3599 /* Find the end_pfn of the first active range of pfns in the node */
3600 i = first_active_region_index_in_nid(nid);
3604 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3606 /* Account for ranges before physical memory on this node */
3607 if (early_node_map[i].start_pfn > range_start_pfn)
3608 hole_pages = prev_end_pfn - range_start_pfn;
3610 /* Find all holes for the zone within the node */
3611 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3613 /* No need to continue if prev_end_pfn is outside the zone */
3614 if (prev_end_pfn >= range_end_pfn)
3617 /* Make sure the end of the zone is not within the hole */
3618 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3619 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3621 /* Update the hole size cound and move on */
3622 if (start_pfn > range_start_pfn) {
3623 BUG_ON(prev_end_pfn > start_pfn);
3624 hole_pages += start_pfn - prev_end_pfn;
3626 prev_end_pfn = early_node_map[i].end_pfn;
3629 /* Account for ranges past physical memory on this node */
3630 if (range_end_pfn > prev_end_pfn)
3631 hole_pages += range_end_pfn -
3632 max(range_start_pfn, prev_end_pfn);
3638 * absent_pages_in_range - Return number of page frames in holes within a range
3639 * @start_pfn: The start PFN to start searching for holes
3640 * @end_pfn: The end PFN to stop searching for holes
3642 * It returns the number of pages frames in memory holes within a range.
3644 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3645 unsigned long end_pfn)
3647 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3650 /* Return the number of page frames in holes in a zone on a node */
3651 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3652 unsigned long zone_type,
3653 unsigned long *ignored)
3655 unsigned long node_start_pfn, node_end_pfn;
3656 unsigned long zone_start_pfn, zone_end_pfn;
3658 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3659 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3661 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3664 adjust_zone_range_for_zone_movable(nid, zone_type,
3665 node_start_pfn, node_end_pfn,
3666 &zone_start_pfn, &zone_end_pfn);
3667 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3671 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3672 unsigned long zone_type,
3673 unsigned long *zones_size)
3675 return zones_size[zone_type];
3678 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3679 unsigned long zone_type,
3680 unsigned long *zholes_size)
3685 return zholes_size[zone_type];
3690 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3691 unsigned long *zones_size, unsigned long *zholes_size)
3693 unsigned long realtotalpages, totalpages = 0;
3696 for (i = 0; i < MAX_NR_ZONES; i++)
3697 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3699 pgdat->node_spanned_pages = totalpages;
3701 realtotalpages = totalpages;
3702 for (i = 0; i < MAX_NR_ZONES; i++)
3704 zone_absent_pages_in_node(pgdat->node_id, i,
3706 pgdat->node_present_pages = realtotalpages;
3707 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3711 #ifndef CONFIG_SPARSEMEM
3713 * Calculate the size of the zone->blockflags rounded to an unsigned long
3714 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3715 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3716 * round what is now in bits to nearest long in bits, then return it in
3719 static unsigned long __init usemap_size(unsigned long zonesize)
3721 unsigned long usemapsize;
3723 usemapsize = roundup(zonesize, pageblock_nr_pages);
3724 usemapsize = usemapsize >> pageblock_order;
3725 usemapsize *= NR_PAGEBLOCK_BITS;
3726 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
3728 return usemapsize / 8;
3731 static void __init setup_usemap(struct pglist_data *pgdat,
3732 struct zone *zone, unsigned long zonesize)
3734 unsigned long usemapsize = usemap_size(zonesize);
3735 zone->pageblock_flags = NULL;
3737 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
3740 static void inline setup_usemap(struct pglist_data *pgdat,
3741 struct zone *zone, unsigned long zonesize) {}
3742 #endif /* CONFIG_SPARSEMEM */
3744 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3746 /* Return a sensible default order for the pageblock size. */
3747 static inline int pageblock_default_order(void)
3749 if (HPAGE_SHIFT > PAGE_SHIFT)
3750 return HUGETLB_PAGE_ORDER;
3755 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3756 static inline void __init set_pageblock_order(unsigned int order)
3758 /* Check that pageblock_nr_pages has not already been setup */
3759 if (pageblock_order)
3763 * Assume the largest contiguous order of interest is a huge page.
3764 * This value may be variable depending on boot parameters on IA64
3766 pageblock_order = order;
3768 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3771 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3772 * and pageblock_default_order() are unused as pageblock_order is set
3773 * at compile-time. See include/linux/pageblock-flags.h for the values of
3774 * pageblock_order based on the kernel config
3776 static inline int pageblock_default_order(unsigned int order)
3780 #define set_pageblock_order(x) do {} while (0)
3782 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3785 * Set up the zone data structures:
3786 * - mark all pages reserved
3787 * - mark all memory queues empty
3788 * - clear the memory bitmaps
3790 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
3791 unsigned long *zones_size, unsigned long *zholes_size)
3794 int nid = pgdat->node_id;
3795 unsigned long zone_start_pfn = pgdat->node_start_pfn;
3798 pgdat_resize_init(pgdat);
3799 pgdat->nr_zones = 0;
3800 init_waitqueue_head(&pgdat->kswapd_wait);
3801 pgdat->kswapd_max_order = 0;
3802 pgdat_page_cgroup_init(pgdat);
3804 for (j = 0; j < MAX_NR_ZONES; j++) {
3805 struct zone *zone = pgdat->node_zones + j;
3806 unsigned long size, realsize, memmap_pages;
3809 size = zone_spanned_pages_in_node(nid, j, zones_size);
3810 realsize = size - zone_absent_pages_in_node(nid, j,
3814 * Adjust realsize so that it accounts for how much memory
3815 * is used by this zone for memmap. This affects the watermark
3816 * and per-cpu initialisations
3819 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
3820 if (realsize >= memmap_pages) {
3821 realsize -= memmap_pages;
3824 " %s zone: %lu pages used for memmap\n",
3825 zone_names[j], memmap_pages);
3828 " %s zone: %lu pages exceeds realsize %lu\n",
3829 zone_names[j], memmap_pages, realsize);
3831 /* Account for reserved pages */
3832 if (j == 0 && realsize > dma_reserve) {
3833 realsize -= dma_reserve;
3834 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
3835 zone_names[0], dma_reserve);
3838 if (!is_highmem_idx(j))
3839 nr_kernel_pages += realsize;
3840 nr_all_pages += realsize;
3842 zone->spanned_pages = size;
3843 zone->present_pages = realsize;
3846 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
3848 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
3850 zone->name = zone_names[j];
3851 spin_lock_init(&zone->lock);
3852 spin_lock_init(&zone->lru_lock);
3853 zone_seqlock_init(zone);
3854 zone->zone_pgdat = pgdat;
3856 zone->prev_priority = DEF_PRIORITY;
3858 zone_pcp_init(zone);
3860 INIT_LIST_HEAD(&zone->lru[l].list);
3861 zone->reclaim_stat.nr_saved_scan[l] = 0;
3863 zone->reclaim_stat.recent_rotated[0] = 0;
3864 zone->reclaim_stat.recent_rotated[1] = 0;
3865 zone->reclaim_stat.recent_scanned[0] = 0;
3866 zone->reclaim_stat.recent_scanned[1] = 0;
3867 zap_zone_vm_stats(zone);
3872 set_pageblock_order(pageblock_default_order());
3873 setup_usemap(pgdat, zone, size);
3874 ret = init_currently_empty_zone(zone, zone_start_pfn,
3875 size, MEMMAP_EARLY);
3877 memmap_init(size, nid, j, zone_start_pfn);
3878 zone_start_pfn += size;
3882 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
3884 /* Skip empty nodes */
3885 if (!pgdat->node_spanned_pages)
3888 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3889 /* ia64 gets its own node_mem_map, before this, without bootmem */
3890 if (!pgdat->node_mem_map) {
3891 unsigned long size, start, end;
3895 * The zone's endpoints aren't required to be MAX_ORDER
3896 * aligned but the node_mem_map endpoints must be in order
3897 * for the buddy allocator to function correctly.
3899 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
3900 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
3901 end = ALIGN(end, MAX_ORDER_NR_PAGES);
3902 size = (end - start) * sizeof(struct page);
3903 map = alloc_remap(pgdat->node_id, size);
3905 map = alloc_bootmem_node(pgdat, size);
3906 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
3908 #ifndef CONFIG_NEED_MULTIPLE_NODES
3910 * With no DISCONTIG, the global mem_map is just set as node 0's
3912 if (pgdat == NODE_DATA(0)) {
3913 mem_map = NODE_DATA(0)->node_mem_map;
3914 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3915 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
3916 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
3917 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3920 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3923 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
3924 unsigned long node_start_pfn, unsigned long *zholes_size)
3926 pg_data_t *pgdat = NODE_DATA(nid);
3928 pgdat->node_id = nid;
3929 pgdat->node_start_pfn = node_start_pfn;
3930 calculate_node_totalpages(pgdat, zones_size, zholes_size);
3932 alloc_node_mem_map(pgdat);
3933 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3934 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3935 nid, (unsigned long)pgdat,
3936 (unsigned long)pgdat->node_mem_map);
3939 free_area_init_core(pgdat, zones_size, zholes_size);
3942 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3944 #if MAX_NUMNODES > 1
3946 * Figure out the number of possible node ids.
3948 static void __init setup_nr_node_ids(void)
3951 unsigned int highest = 0;
3953 for_each_node_mask(node, node_possible_map)
3955 nr_node_ids = highest + 1;
3958 static inline void setup_nr_node_ids(void)
3964 * add_active_range - Register a range of PFNs backed by physical memory
3965 * @nid: The node ID the range resides on
3966 * @start_pfn: The start PFN of the available physical memory
3967 * @end_pfn: The end PFN of the available physical memory
3969 * These ranges are stored in an early_node_map[] and later used by
3970 * free_area_init_nodes() to calculate zone sizes and holes. If the
3971 * range spans a memory hole, it is up to the architecture to ensure
3972 * the memory is not freed by the bootmem allocator. If possible
3973 * the range being registered will be merged with existing ranges.
3975 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
3976 unsigned long end_pfn)
3980 mminit_dprintk(MMINIT_TRACE, "memory_register",
3981 "Entering add_active_range(%d, %#lx, %#lx) "
3982 "%d entries of %d used\n",
3983 nid, start_pfn, end_pfn,
3984 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
3986 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
3988 /* Merge with existing active regions if possible */
3989 for (i = 0; i < nr_nodemap_entries; i++) {
3990 if (early_node_map[i].nid != nid)
3993 /* Skip if an existing region covers this new one */
3994 if (start_pfn >= early_node_map[i].start_pfn &&
3995 end_pfn <= early_node_map[i].end_pfn)
3998 /* Merge forward if suitable */
3999 if (start_pfn <= early_node_map[i].end_pfn &&
4000 end_pfn > early_node_map[i].end_pfn) {
4001 early_node_map[i].end_pfn = end_pfn;
4005 /* Merge backward if suitable */
4006 if (start_pfn < early_node_map[i].end_pfn &&
4007 end_pfn >= early_node_map[i].start_pfn) {
4008 early_node_map[i].start_pfn = start_pfn;
4013 /* Check that early_node_map is large enough */
4014 if (i >= MAX_ACTIVE_REGIONS) {
4015 printk(KERN_CRIT "More than %d memory regions, truncating\n",
4016 MAX_ACTIVE_REGIONS);
4020 early_node_map[i].nid = nid;
4021 early_node_map[i].start_pfn = start_pfn;
4022 early_node_map[i].end_pfn = end_pfn;
4023 nr_nodemap_entries = i + 1;
4027 * remove_active_range - Shrink an existing registered range of PFNs
4028 * @nid: The node id the range is on that should be shrunk
4029 * @start_pfn: The new PFN of the range
4030 * @end_pfn: The new PFN of the range
4032 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4033 * The map is kept near the end physical page range that has already been
4034 * registered. This function allows an arch to shrink an existing registered
4037 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4038 unsigned long end_pfn)
4043 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4044 nid, start_pfn, end_pfn);
4046 /* Find the old active region end and shrink */
4047 for_each_active_range_index_in_nid(i, nid) {
4048 if (early_node_map[i].start_pfn >= start_pfn &&
4049 early_node_map[i].end_pfn <= end_pfn) {
4051 early_node_map[i].start_pfn = 0;
4052 early_node_map[i].end_pfn = 0;
4056 if (early_node_map[i].start_pfn < start_pfn &&
4057 early_node_map[i].end_pfn > start_pfn) {
4058 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4059 early_node_map[i].end_pfn = start_pfn;
4060 if (temp_end_pfn > end_pfn)
4061 add_active_range(nid, end_pfn, temp_end_pfn);
4064 if (early_node_map[i].start_pfn >= start_pfn &&
4065 early_node_map[i].end_pfn > end_pfn &&
4066 early_node_map[i].start_pfn < end_pfn) {
4067 early_node_map[i].start_pfn = end_pfn;
4075 /* remove the blank ones */
4076 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4077 if (early_node_map[i].nid != nid)
4079 if (early_node_map[i].end_pfn)
4081 /* we found it, get rid of it */
4082 for (j = i; j < nr_nodemap_entries - 1; j++)
4083 memcpy(&early_node_map[j], &early_node_map[j+1],
4084 sizeof(early_node_map[j]));
4085 j = nr_nodemap_entries - 1;
4086 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4087 nr_nodemap_entries--;
4092 * remove_all_active_ranges - Remove all currently registered regions
4094 * During discovery, it may be found that a table like SRAT is invalid
4095 * and an alternative discovery method must be used. This function removes
4096 * all currently registered regions.
4098 void __init remove_all_active_ranges(void)
4100 memset(early_node_map, 0, sizeof(early_node_map));
4101 nr_nodemap_entries = 0;
4104 /* Compare two active node_active_regions */
4105 static int __init cmp_node_active_region(const void *a, const void *b)
4107 struct node_active_region *arange = (struct node_active_region *)a;
4108 struct node_active_region *brange = (struct node_active_region *)b;
4110 /* Done this way to avoid overflows */
4111 if (arange->start_pfn > brange->start_pfn)
4113 if (arange->start_pfn < brange->start_pfn)
4119 /* sort the node_map by start_pfn */
4120 static void __init sort_node_map(void)
4122 sort(early_node_map, (size_t)nr_nodemap_entries,
4123 sizeof(struct node_active_region),
4124 cmp_node_active_region, NULL);
4127 /* Find the lowest pfn for a node */
4128 static unsigned long __init find_min_pfn_for_node(int nid)
4131 unsigned long min_pfn = ULONG_MAX;
4133 /* Assuming a sorted map, the first range found has the starting pfn */
4134 for_each_active_range_index_in_nid(i, nid)
4135 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4137 if (min_pfn == ULONG_MAX) {
4139 "Could not find start_pfn for node %d\n", nid);
4147 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4149 * It returns the minimum PFN based on information provided via
4150 * add_active_range().
4152 unsigned long __init find_min_pfn_with_active_regions(void)
4154 return find_min_pfn_for_node(MAX_NUMNODES);
4158 * early_calculate_totalpages()
4159 * Sum pages in active regions for movable zone.
4160 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4162 static unsigned long __init early_calculate_totalpages(void)
4165 unsigned long totalpages = 0;
4167 for (i = 0; i < nr_nodemap_entries; i++) {
4168 unsigned long pages = early_node_map[i].end_pfn -
4169 early_node_map[i].start_pfn;
4170 totalpages += pages;
4172 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4178 * Find the PFN the Movable zone begins in each node. Kernel memory
4179 * is spread evenly between nodes as long as the nodes have enough
4180 * memory. When they don't, some nodes will have more kernelcore than
4183 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4186 unsigned long usable_startpfn;
4187 unsigned long kernelcore_node, kernelcore_remaining;
4188 /* save the state before borrow the nodemask */
4189 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4190 unsigned long totalpages = early_calculate_totalpages();
4191 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4194 * If movablecore was specified, calculate what size of
4195 * kernelcore that corresponds so that memory usable for
4196 * any allocation type is evenly spread. If both kernelcore
4197 * and movablecore are specified, then the value of kernelcore
4198 * will be used for required_kernelcore if it's greater than
4199 * what movablecore would have allowed.
4201 if (required_movablecore) {
4202 unsigned long corepages;
4205 * Round-up so that ZONE_MOVABLE is at least as large as what
4206 * was requested by the user
4208 required_movablecore =
4209 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4210 corepages = totalpages - required_movablecore;
4212 required_kernelcore = max(required_kernelcore, corepages);
4215 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4216 if (!required_kernelcore)
4219 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4220 find_usable_zone_for_movable();
4221 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4224 /* Spread kernelcore memory as evenly as possible throughout nodes */
4225 kernelcore_node = required_kernelcore / usable_nodes;
4226 for_each_node_state(nid, N_HIGH_MEMORY) {
4228 * Recalculate kernelcore_node if the division per node
4229 * now exceeds what is necessary to satisfy the requested
4230 * amount of memory for the kernel
4232 if (required_kernelcore < kernelcore_node)
4233 kernelcore_node = required_kernelcore / usable_nodes;
4236 * As the map is walked, we track how much memory is usable
4237 * by the kernel using kernelcore_remaining. When it is
4238 * 0, the rest of the node is usable by ZONE_MOVABLE
4240 kernelcore_remaining = kernelcore_node;
4242 /* Go through each range of PFNs within this node */
4243 for_each_active_range_index_in_nid(i, nid) {
4244 unsigned long start_pfn, end_pfn;
4245 unsigned long size_pages;
4247 start_pfn = max(early_node_map[i].start_pfn,
4248 zone_movable_pfn[nid]);
4249 end_pfn = early_node_map[i].end_pfn;
4250 if (start_pfn >= end_pfn)
4253 /* Account for what is only usable for kernelcore */
4254 if (start_pfn < usable_startpfn) {
4255 unsigned long kernel_pages;
4256 kernel_pages = min(end_pfn, usable_startpfn)
4259 kernelcore_remaining -= min(kernel_pages,
4260 kernelcore_remaining);
4261 required_kernelcore -= min(kernel_pages,
4262 required_kernelcore);
4264 /* Continue if range is now fully accounted */
4265 if (end_pfn <= usable_startpfn) {
4268 * Push zone_movable_pfn to the end so
4269 * that if we have to rebalance
4270 * kernelcore across nodes, we will
4271 * not double account here
4273 zone_movable_pfn[nid] = end_pfn;
4276 start_pfn = usable_startpfn;
4280 * The usable PFN range for ZONE_MOVABLE is from
4281 * start_pfn->end_pfn. Calculate size_pages as the
4282 * number of pages used as kernelcore
4284 size_pages = end_pfn - start_pfn;
4285 if (size_pages > kernelcore_remaining)
4286 size_pages = kernelcore_remaining;
4287 zone_movable_pfn[nid] = start_pfn + size_pages;
4290 * Some kernelcore has been met, update counts and
4291 * break if the kernelcore for this node has been
4294 required_kernelcore -= min(required_kernelcore,
4296 kernelcore_remaining -= size_pages;
4297 if (!kernelcore_remaining)
4303 * If there is still required_kernelcore, we do another pass with one
4304 * less node in the count. This will push zone_movable_pfn[nid] further
4305 * along on the nodes that still have memory until kernelcore is
4309 if (usable_nodes && required_kernelcore > usable_nodes)
4312 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4313 for (nid = 0; nid < MAX_NUMNODES; nid++)
4314 zone_movable_pfn[nid] =
4315 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4318 /* restore the node_state */
4319 node_states[N_HIGH_MEMORY] = saved_node_state;
4322 /* Any regular memory on that node ? */
4323 static void check_for_regular_memory(pg_data_t *pgdat)
4325 #ifdef CONFIG_HIGHMEM
4326 enum zone_type zone_type;
4328 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4329 struct zone *zone = &pgdat->node_zones[zone_type];
4330 if (zone->present_pages)
4331 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4337 * free_area_init_nodes - Initialise all pg_data_t and zone data
4338 * @max_zone_pfn: an array of max PFNs for each zone
4340 * This will call free_area_init_node() for each active node in the system.
4341 * Using the page ranges provided by add_active_range(), the size of each
4342 * zone in each node and their holes is calculated. If the maximum PFN
4343 * between two adjacent zones match, it is assumed that the zone is empty.
4344 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4345 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4346 * starts where the previous one ended. For example, ZONE_DMA32 starts
4347 * at arch_max_dma_pfn.
4349 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4354 /* Sort early_node_map as initialisation assumes it is sorted */
4357 /* Record where the zone boundaries are */
4358 memset(arch_zone_lowest_possible_pfn, 0,
4359 sizeof(arch_zone_lowest_possible_pfn));
4360 memset(arch_zone_highest_possible_pfn, 0,
4361 sizeof(arch_zone_highest_possible_pfn));
4362 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4363 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4364 for (i = 1; i < MAX_NR_ZONES; i++) {
4365 if (i == ZONE_MOVABLE)
4367 arch_zone_lowest_possible_pfn[i] =
4368 arch_zone_highest_possible_pfn[i-1];
4369 arch_zone_highest_possible_pfn[i] =
4370 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4372 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4373 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4375 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4376 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4377 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4379 /* Print out the zone ranges */
4380 printk("Zone PFN ranges:\n");
4381 for (i = 0; i < MAX_NR_ZONES; i++) {
4382 if (i == ZONE_MOVABLE)
4384 printk(" %-8s %0#10lx -> %0#10lx\n",
4386 arch_zone_lowest_possible_pfn[i],
4387 arch_zone_highest_possible_pfn[i]);
4390 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4391 printk("Movable zone start PFN for each node\n");
4392 for (i = 0; i < MAX_NUMNODES; i++) {
4393 if (zone_movable_pfn[i])
4394 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4397 /* Print out the early_node_map[] */
4398 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4399 for (i = 0; i < nr_nodemap_entries; i++)
4400 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4401 early_node_map[i].start_pfn,
4402 early_node_map[i].end_pfn);
4404 /* Initialise every node */
4405 mminit_verify_pageflags_layout();
4406 setup_nr_node_ids();
4407 for_each_online_node(nid) {
4408 pg_data_t *pgdat = NODE_DATA(nid);
4409 free_area_init_node(nid, NULL,
4410 find_min_pfn_for_node(nid), NULL);
4412 /* Any memory on that node */
4413 if (pgdat->node_present_pages)
4414 node_set_state(nid, N_HIGH_MEMORY);
4415 check_for_regular_memory(pgdat);
4419 static int __init cmdline_parse_core(char *p, unsigned long *core)
4421 unsigned long long coremem;
4425 coremem = memparse(p, &p);
4426 *core = coremem >> PAGE_SHIFT;
4428 /* Paranoid check that UL is enough for the coremem value */
4429 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4435 * kernelcore=size sets the amount of memory for use for allocations that
4436 * cannot be reclaimed or migrated.
4438 static int __init cmdline_parse_kernelcore(char *p)
4440 return cmdline_parse_core(p, &required_kernelcore);
4444 * movablecore=size sets the amount of memory for use for allocations that
4445 * can be reclaimed or migrated.
4447 static int __init cmdline_parse_movablecore(char *p)
4449 return cmdline_parse_core(p, &required_movablecore);
4452 early_param("kernelcore", cmdline_parse_kernelcore);
4453 early_param("movablecore", cmdline_parse_movablecore);
4455 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4458 * set_dma_reserve - set the specified number of pages reserved in the first zone
4459 * @new_dma_reserve: The number of pages to mark reserved
4461 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4462 * In the DMA zone, a significant percentage may be consumed by kernel image
4463 * and other unfreeable allocations which can skew the watermarks badly. This
4464 * function may optionally be used to account for unfreeable pages in the
4465 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4466 * smaller per-cpu batchsize.
4468 void __init set_dma_reserve(unsigned long new_dma_reserve)
4470 dma_reserve = new_dma_reserve;
4473 #ifndef CONFIG_NEED_MULTIPLE_NODES
4474 struct pglist_data __refdata contig_page_data = { .bdata = &bootmem_node_data[0] };
4475 EXPORT_SYMBOL(contig_page_data);
4478 void __init free_area_init(unsigned long *zones_size)
4480 free_area_init_node(0, zones_size,
4481 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4484 static int page_alloc_cpu_notify(struct notifier_block *self,
4485 unsigned long action, void *hcpu)
4487 int cpu = (unsigned long)hcpu;
4489 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4493 * Spill the event counters of the dead processor
4494 * into the current processors event counters.
4495 * This artificially elevates the count of the current
4498 vm_events_fold_cpu(cpu);
4501 * Zero the differential counters of the dead processor
4502 * so that the vm statistics are consistent.
4504 * This is only okay since the processor is dead and cannot
4505 * race with what we are doing.
4507 refresh_cpu_vm_stats(cpu);
4512 void __init page_alloc_init(void)
4514 hotcpu_notifier(page_alloc_cpu_notify, 0);
4518 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4519 * or min_free_kbytes changes.
4521 static void calculate_totalreserve_pages(void)
4523 struct pglist_data *pgdat;
4524 unsigned long reserve_pages = 0;
4525 enum zone_type i, j;
4527 for_each_online_pgdat(pgdat) {
4528 for (i = 0; i < MAX_NR_ZONES; i++) {
4529 struct zone *zone = pgdat->node_zones + i;
4530 unsigned long max = 0;
4532 /* Find valid and maximum lowmem_reserve in the zone */
4533 for (j = i; j < MAX_NR_ZONES; j++) {
4534 if (zone->lowmem_reserve[j] > max)
4535 max = zone->lowmem_reserve[j];
4538 /* we treat the high watermark as reserved pages. */
4539 max += high_wmark_pages(zone);
4541 if (max > zone->present_pages)
4542 max = zone->present_pages;
4543 reserve_pages += max;
4546 totalreserve_pages = reserve_pages;
4550 * setup_per_zone_lowmem_reserve - called whenever
4551 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4552 * has a correct pages reserved value, so an adequate number of
4553 * pages are left in the zone after a successful __alloc_pages().
4555 static void setup_per_zone_lowmem_reserve(void)
4557 struct pglist_data *pgdat;
4558 enum zone_type j, idx;
4560 for_each_online_pgdat(pgdat) {
4561 for (j = 0; j < MAX_NR_ZONES; j++) {
4562 struct zone *zone = pgdat->node_zones + j;
4563 unsigned long present_pages = zone->present_pages;
4565 zone->lowmem_reserve[j] = 0;
4569 struct zone *lower_zone;
4573 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4574 sysctl_lowmem_reserve_ratio[idx] = 1;
4576 lower_zone = pgdat->node_zones + idx;
4577 lower_zone->lowmem_reserve[j] = present_pages /
4578 sysctl_lowmem_reserve_ratio[idx];
4579 present_pages += lower_zone->present_pages;
4584 /* update totalreserve_pages */
4585 calculate_totalreserve_pages();
4589 * setup_per_zone_wmarks - called when min_free_kbytes changes
4590 * or when memory is hot-{added|removed}
4592 * Ensures that the watermark[min,low,high] values for each zone are set
4593 * correctly with respect to min_free_kbytes.
4595 void setup_per_zone_wmarks(void)
4597 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4598 unsigned long lowmem_pages = 0;
4600 unsigned long flags;
4602 /* Calculate total number of !ZONE_HIGHMEM pages */
4603 for_each_zone(zone) {
4604 if (!is_highmem(zone))
4605 lowmem_pages += zone->present_pages;
4608 for_each_zone(zone) {
4611 spin_lock_irqsave(&zone->lock, flags);
4612 tmp = (u64)pages_min * zone->present_pages;
4613 do_div(tmp, lowmem_pages);
4614 if (is_highmem(zone)) {
4616 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4617 * need highmem pages, so cap pages_min to a small
4620 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4621 * deltas controls asynch page reclaim, and so should
4622 * not be capped for highmem.
4626 min_pages = zone->present_pages / 1024;
4627 if (min_pages < SWAP_CLUSTER_MAX)
4628 min_pages = SWAP_CLUSTER_MAX;
4629 if (min_pages > 128)
4631 zone->watermark[WMARK_MIN] = min_pages;
4634 * If it's a lowmem zone, reserve a number of pages
4635 * proportionate to the zone's size.
4637 zone->watermark[WMARK_MIN] = tmp;
4640 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
4641 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
4642 setup_zone_migrate_reserve(zone);
4643 spin_unlock_irqrestore(&zone->lock, flags);
4646 /* update totalreserve_pages */
4647 calculate_totalreserve_pages();
4651 * The inactive anon list should be small enough that the VM never has to
4652 * do too much work, but large enough that each inactive page has a chance
4653 * to be referenced again before it is swapped out.
4655 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4656 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4657 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4658 * the anonymous pages are kept on the inactive list.
4661 * memory ratio inactive anon
4662 * -------------------------------------
4671 void calculate_zone_inactive_ratio(struct zone *zone)
4673 unsigned int gb, ratio;
4675 /* Zone size in gigabytes */
4676 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4678 ratio = int_sqrt(10 * gb);
4682 zone->inactive_ratio = ratio;
4685 static void __init setup_per_zone_inactive_ratio(void)
4690 calculate_zone_inactive_ratio(zone);
4694 * Initialise min_free_kbytes.
4696 * For small machines we want it small (128k min). For large machines
4697 * we want it large (64MB max). But it is not linear, because network
4698 * bandwidth does not increase linearly with machine size. We use
4700 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4701 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4717 static int __init init_per_zone_wmark_min(void)
4719 unsigned long lowmem_kbytes;
4721 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4723 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4724 if (min_free_kbytes < 128)
4725 min_free_kbytes = 128;
4726 if (min_free_kbytes > 65536)
4727 min_free_kbytes = 65536;
4728 setup_per_zone_wmarks();
4729 setup_per_zone_lowmem_reserve();
4730 setup_per_zone_inactive_ratio();
4733 module_init(init_per_zone_wmark_min)
4736 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4737 * that we can call two helper functions whenever min_free_kbytes
4740 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
4741 void __user *buffer, size_t *length, loff_t *ppos)
4743 proc_dointvec(table, write, buffer, length, ppos);
4745 setup_per_zone_wmarks();
4750 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
4751 void __user *buffer, size_t *length, loff_t *ppos)
4756 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
4761 zone->min_unmapped_pages = (zone->present_pages *
4762 sysctl_min_unmapped_ratio) / 100;
4766 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
4767 void __user *buffer, size_t *length, loff_t *ppos)
4772 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
4777 zone->min_slab_pages = (zone->present_pages *
4778 sysctl_min_slab_ratio) / 100;
4784 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4785 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4786 * whenever sysctl_lowmem_reserve_ratio changes.
4788 * The reserve ratio obviously has absolutely no relation with the
4789 * minimum watermarks. The lowmem reserve ratio can only make sense
4790 * if in function of the boot time zone sizes.
4792 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
4793 void __user *buffer, size_t *length, loff_t *ppos)
4795 proc_dointvec_minmax(table, write, buffer, length, ppos);
4796 setup_per_zone_lowmem_reserve();
4801 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4802 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4803 * can have before it gets flushed back to buddy allocator.
4806 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
4807 void __user *buffer, size_t *length, loff_t *ppos)
4813 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
4814 if (!write || (ret == -EINVAL))
4816 for_each_populated_zone(zone) {
4817 for_each_online_cpu(cpu) {
4819 high = zone->present_pages / percpu_pagelist_fraction;
4820 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
4826 int hashdist = HASHDIST_DEFAULT;
4829 static int __init set_hashdist(char *str)
4833 hashdist = simple_strtoul(str, &str, 0);
4836 __setup("hashdist=", set_hashdist);
4840 * allocate a large system hash table from bootmem
4841 * - it is assumed that the hash table must contain an exact power-of-2
4842 * quantity of entries
4843 * - limit is the number of hash buckets, not the total allocation size
4845 void *__init alloc_large_system_hash(const char *tablename,
4846 unsigned long bucketsize,
4847 unsigned long numentries,
4850 unsigned int *_hash_shift,
4851 unsigned int *_hash_mask,
4852 unsigned long limit)
4854 unsigned long long max = limit;
4855 unsigned long log2qty, size;
4858 /* allow the kernel cmdline to have a say */
4860 /* round applicable memory size up to nearest megabyte */
4861 numentries = nr_kernel_pages;
4862 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
4863 numentries >>= 20 - PAGE_SHIFT;
4864 numentries <<= 20 - PAGE_SHIFT;
4866 /* limit to 1 bucket per 2^scale bytes of low memory */
4867 if (scale > PAGE_SHIFT)
4868 numentries >>= (scale - PAGE_SHIFT);
4870 numentries <<= (PAGE_SHIFT - scale);
4872 /* Make sure we've got at least a 0-order allocation.. */
4873 if (unlikely(flags & HASH_SMALL)) {
4874 /* Makes no sense without HASH_EARLY */
4875 WARN_ON(!(flags & HASH_EARLY));
4876 if (!(numentries >> *_hash_shift)) {
4877 numentries = 1UL << *_hash_shift;
4878 BUG_ON(!numentries);
4880 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
4881 numentries = PAGE_SIZE / bucketsize;
4883 numentries = roundup_pow_of_two(numentries);
4885 /* limit allocation size to 1/16 total memory by default */
4887 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
4888 do_div(max, bucketsize);
4891 if (numentries > max)
4894 log2qty = ilog2(numentries);
4897 size = bucketsize << log2qty;
4898 if (flags & HASH_EARLY)
4899 table = alloc_bootmem_nopanic(size);
4901 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
4904 * If bucketsize is not a power-of-two, we may free
4905 * some pages at the end of hash table which
4906 * alloc_pages_exact() automatically does
4908 if (get_order(size) < MAX_ORDER) {
4909 table = alloc_pages_exact(size, GFP_ATOMIC);
4910 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
4913 } while (!table && size > PAGE_SIZE && --log2qty);
4916 panic("Failed to allocate %s hash table\n", tablename);
4918 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
4921 ilog2(size) - PAGE_SHIFT,
4925 *_hash_shift = log2qty;
4927 *_hash_mask = (1 << log2qty) - 1;
4932 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4933 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
4936 #ifdef CONFIG_SPARSEMEM
4937 return __pfn_to_section(pfn)->pageblock_flags;
4939 return zone->pageblock_flags;
4940 #endif /* CONFIG_SPARSEMEM */
4943 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
4945 #ifdef CONFIG_SPARSEMEM
4946 pfn &= (PAGES_PER_SECTION-1);
4947 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4949 pfn = pfn - zone->zone_start_pfn;
4950 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4951 #endif /* CONFIG_SPARSEMEM */
4955 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4956 * @page: The page within the block of interest
4957 * @start_bitidx: The first bit of interest to retrieve
4958 * @end_bitidx: The last bit of interest
4959 * returns pageblock_bits flags
4961 unsigned long get_pageblock_flags_group(struct page *page,
4962 int start_bitidx, int end_bitidx)
4965 unsigned long *bitmap;
4966 unsigned long pfn, bitidx;
4967 unsigned long flags = 0;
4968 unsigned long value = 1;
4970 zone = page_zone(page);
4971 pfn = page_to_pfn(page);
4972 bitmap = get_pageblock_bitmap(zone, pfn);
4973 bitidx = pfn_to_bitidx(zone, pfn);
4975 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4976 if (test_bit(bitidx + start_bitidx, bitmap))
4983 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4984 * @page: The page within the block of interest
4985 * @start_bitidx: The first bit of interest
4986 * @end_bitidx: The last bit of interest
4987 * @flags: The flags to set
4989 void set_pageblock_flags_group(struct page *page, unsigned long flags,
4990 int start_bitidx, int end_bitidx)
4993 unsigned long *bitmap;
4994 unsigned long pfn, bitidx;
4995 unsigned long value = 1;
4997 zone = page_zone(page);
4998 pfn = page_to_pfn(page);
4999 bitmap = get_pageblock_bitmap(zone, pfn);
5000 bitidx = pfn_to_bitidx(zone, pfn);
5001 VM_BUG_ON(pfn < zone->zone_start_pfn);
5002 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5004 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5006 __set_bit(bitidx + start_bitidx, bitmap);
5008 __clear_bit(bitidx + start_bitidx, bitmap);
5012 * This is designed as sub function...plz see page_isolation.c also.
5013 * set/clear page block's type to be ISOLATE.
5014 * page allocater never alloc memory from ISOLATE block.
5017 int set_migratetype_isolate(struct page *page)
5020 unsigned long flags;
5024 zone = page_zone(page);
5025 zone_idx = zone_idx(zone);
5026 spin_lock_irqsave(&zone->lock, flags);
5028 * In future, more migrate types will be able to be isolation target.
5030 if (get_pageblock_migratetype(page) != MIGRATE_MOVABLE &&
5031 zone_idx != ZONE_MOVABLE)
5033 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5034 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5037 spin_unlock_irqrestore(&zone->lock, flags);
5043 void unset_migratetype_isolate(struct page *page)
5046 unsigned long flags;
5047 zone = page_zone(page);
5048 spin_lock_irqsave(&zone->lock, flags);
5049 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5051 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5052 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5054 spin_unlock_irqrestore(&zone->lock, flags);
5057 #ifdef CONFIG_MEMORY_HOTREMOVE
5059 * All pages in the range must be isolated before calling this.
5062 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5068 unsigned long flags;
5069 /* find the first valid pfn */
5070 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5075 zone = page_zone(pfn_to_page(pfn));
5076 spin_lock_irqsave(&zone->lock, flags);
5078 while (pfn < end_pfn) {
5079 if (!pfn_valid(pfn)) {
5083 page = pfn_to_page(pfn);
5084 BUG_ON(page_count(page));
5085 BUG_ON(!PageBuddy(page));
5086 order = page_order(page);
5087 #ifdef CONFIG_DEBUG_VM
5088 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5089 pfn, 1 << order, end_pfn);
5091 list_del(&page->lru);
5092 rmv_page_order(page);
5093 zone->free_area[order].nr_free--;
5094 __mod_zone_page_state(zone, NR_FREE_PAGES,
5096 for (i = 0; i < (1 << order); i++)
5097 SetPageReserved((page+i));
5098 pfn += (1 << order);
5100 spin_unlock_irqrestore(&zone->lock, flags);