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 <linux/memory.h>
52 #include <linux/compaction.h>
53 #include <trace/events/kmem.h>
54 #include <linux/ftrace_event.h>
56 #include <asm/tlbflush.h>
57 #include <asm/div64.h>
61 * Array of node states.
63 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
64 [N_POSSIBLE] = NODE_MASK_ALL,
65 [N_ONLINE] = { { [0] = 1UL } },
67 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
69 [N_HIGH_MEMORY] = { { [0] = 1UL } },
71 [N_CPU] = { { [0] = 1UL } },
74 EXPORT_SYMBOL(node_states);
76 unsigned long totalram_pages __read_mostly;
77 unsigned long totalreserve_pages __read_mostly;
78 int percpu_pagelist_fraction;
79 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
81 #ifdef CONFIG_PM_SLEEP
83 * The following functions are used by the suspend/hibernate code to temporarily
84 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
85 * while devices are suspended. To avoid races with the suspend/hibernate code,
86 * they should always be called with pm_mutex held (gfp_allowed_mask also should
87 * only be modified with pm_mutex held, unless the suspend/hibernate code is
88 * guaranteed not to run in parallel with that modification).
90 void set_gfp_allowed_mask(gfp_t mask)
92 WARN_ON(!mutex_is_locked(&pm_mutex));
93 gfp_allowed_mask = mask;
96 gfp_t clear_gfp_allowed_mask(gfp_t mask)
98 gfp_t ret = gfp_allowed_mask;
100 WARN_ON(!mutex_is_locked(&pm_mutex));
101 gfp_allowed_mask &= ~mask;
104 #endif /* CONFIG_PM_SLEEP */
106 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
107 int pageblock_order __read_mostly;
110 static void __free_pages_ok(struct page *page, unsigned int order);
113 * results with 256, 32 in the lowmem_reserve sysctl:
114 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
115 * 1G machine -> (16M dma, 784M normal, 224M high)
116 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
117 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
118 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
120 * TBD: should special case ZONE_DMA32 machines here - in those we normally
121 * don't need any ZONE_NORMAL reservation
123 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
124 #ifdef CONFIG_ZONE_DMA
127 #ifdef CONFIG_ZONE_DMA32
130 #ifdef CONFIG_HIGHMEM
136 EXPORT_SYMBOL(totalram_pages);
138 static char * const zone_names[MAX_NR_ZONES] = {
139 #ifdef CONFIG_ZONE_DMA
142 #ifdef CONFIG_ZONE_DMA32
146 #ifdef CONFIG_HIGHMEM
152 int min_free_kbytes = 1024;
154 static unsigned long __meminitdata nr_kernel_pages;
155 static unsigned long __meminitdata nr_all_pages;
156 static unsigned long __meminitdata dma_reserve;
158 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
160 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
161 * ranges of memory (RAM) that may be registered with add_active_range().
162 * Ranges passed to add_active_range() will be merged if possible
163 * so the number of times add_active_range() can be called is
164 * related to the number of nodes and the number of holes
166 #ifdef CONFIG_MAX_ACTIVE_REGIONS
167 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
168 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
170 #if MAX_NUMNODES >= 32
171 /* If there can be many nodes, allow up to 50 holes per node */
172 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
174 /* By default, allow up to 256 distinct regions */
175 #define MAX_ACTIVE_REGIONS 256
179 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
180 static int __meminitdata nr_nodemap_entries;
181 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
182 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
183 static unsigned long __initdata required_kernelcore;
184 static unsigned long __initdata required_movablecore;
185 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
187 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
189 EXPORT_SYMBOL(movable_zone);
190 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
193 int nr_node_ids __read_mostly = MAX_NUMNODES;
194 int nr_online_nodes __read_mostly = 1;
195 EXPORT_SYMBOL(nr_node_ids);
196 EXPORT_SYMBOL(nr_online_nodes);
199 int page_group_by_mobility_disabled __read_mostly;
201 static void set_pageblock_migratetype(struct page *page, int migratetype)
204 if (unlikely(page_group_by_mobility_disabled))
205 migratetype = MIGRATE_UNMOVABLE;
207 set_pageblock_flags_group(page, (unsigned long)migratetype,
208 PB_migrate, PB_migrate_end);
211 bool oom_killer_disabled __read_mostly;
213 #ifdef CONFIG_DEBUG_VM
214 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
218 unsigned long pfn = page_to_pfn(page);
221 seq = zone_span_seqbegin(zone);
222 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
224 else if (pfn < zone->zone_start_pfn)
226 } while (zone_span_seqretry(zone, seq));
231 static int page_is_consistent(struct zone *zone, struct page *page)
233 if (!pfn_valid_within(page_to_pfn(page)))
235 if (zone != page_zone(page))
241 * Temporary debugging check for pages not lying within a given zone.
243 static int bad_range(struct zone *zone, struct page *page)
245 if (page_outside_zone_boundaries(zone, page))
247 if (!page_is_consistent(zone, page))
253 static inline int bad_range(struct zone *zone, struct page *page)
259 static void bad_page(struct page *page)
261 static unsigned long resume;
262 static unsigned long nr_shown;
263 static unsigned long nr_unshown;
265 /* Don't complain about poisoned pages */
266 if (PageHWPoison(page)) {
267 __ClearPageBuddy(page);
272 * Allow a burst of 60 reports, then keep quiet for that minute;
273 * or allow a steady drip of one report per second.
275 if (nr_shown == 60) {
276 if (time_before(jiffies, resume)) {
282 "BUG: Bad page state: %lu messages suppressed\n",
289 resume = jiffies + 60 * HZ;
291 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
292 current->comm, page_to_pfn(page));
297 /* Leave bad fields for debug, except PageBuddy could make trouble */
298 __ClearPageBuddy(page);
299 add_taint(TAINT_BAD_PAGE);
303 * Higher-order pages are called "compound pages". They are structured thusly:
305 * The first PAGE_SIZE page is called the "head page".
307 * The remaining PAGE_SIZE pages are called "tail pages".
309 * All pages have PG_compound set. All pages have their ->private pointing at
310 * the head page (even the head page has this).
312 * The first tail page's ->lru.next holds the address of the compound page's
313 * put_page() function. Its ->lru.prev holds the order of allocation.
314 * This usage means that zero-order pages may not be compound.
317 static void free_compound_page(struct page *page)
319 __free_pages_ok(page, compound_order(page));
322 void prep_compound_page(struct page *page, unsigned long order)
325 int nr_pages = 1 << order;
327 set_compound_page_dtor(page, free_compound_page);
328 set_compound_order(page, order);
330 for (i = 1; i < nr_pages; i++) {
331 struct page *p = page + i;
334 p->first_page = page;
338 static int destroy_compound_page(struct page *page, unsigned long order)
341 int nr_pages = 1 << order;
344 if (unlikely(compound_order(page) != order) ||
345 unlikely(!PageHead(page))) {
350 __ClearPageHead(page);
352 for (i = 1; i < nr_pages; i++) {
353 struct page *p = page + i;
355 if (unlikely(!PageTail(p) || (p->first_page != page))) {
365 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
370 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
371 * and __GFP_HIGHMEM from hard or soft interrupt context.
373 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
374 for (i = 0; i < (1 << order); i++)
375 clear_highpage(page + i);
378 static inline void set_page_order(struct page *page, int order)
380 set_page_private(page, order);
381 __SetPageBuddy(page);
384 static inline void rmv_page_order(struct page *page)
386 __ClearPageBuddy(page);
387 set_page_private(page, 0);
391 * Locate the struct page for both the matching buddy in our
392 * pair (buddy1) and the combined O(n+1) page they form (page).
394 * 1) Any buddy B1 will have an order O twin B2 which satisfies
395 * the following equation:
397 * For example, if the starting buddy (buddy2) is #8 its order
399 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
401 * 2) Any buddy B will have an order O+1 parent P which
402 * satisfies the following equation:
405 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
407 static inline struct page *
408 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
410 unsigned long buddy_idx = page_idx ^ (1 << order);
412 return page + (buddy_idx - page_idx);
415 static inline unsigned long
416 __find_combined_index(unsigned long page_idx, unsigned int order)
418 return (page_idx & ~(1 << order));
422 * This function checks whether a page is free && is the buddy
423 * we can do coalesce a page and its buddy if
424 * (a) the buddy is not in a hole &&
425 * (b) the buddy is in the buddy system &&
426 * (c) a page and its buddy have the same order &&
427 * (d) a page and its buddy are in the same zone.
429 * For recording whether a page is in the buddy system, we use PG_buddy.
430 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
432 * For recording page's order, we use page_private(page).
434 static inline int page_is_buddy(struct page *page, struct page *buddy,
437 if (!pfn_valid_within(page_to_pfn(buddy)))
440 if (page_zone_id(page) != page_zone_id(buddy))
443 if (PageBuddy(buddy) && page_order(buddy) == order) {
444 VM_BUG_ON(page_count(buddy) != 0);
451 * Freeing function for a buddy system allocator.
453 * The concept of a buddy system is to maintain direct-mapped table
454 * (containing bit values) for memory blocks of various "orders".
455 * The bottom level table contains the map for the smallest allocatable
456 * units of memory (here, pages), and each level above it describes
457 * pairs of units from the levels below, hence, "buddies".
458 * At a high level, all that happens here is marking the table entry
459 * at the bottom level available, and propagating the changes upward
460 * as necessary, plus some accounting needed to play nicely with other
461 * parts of the VM system.
462 * At each level, we keep a list of pages, which are heads of continuous
463 * free pages of length of (1 << order) and marked with PG_buddy. Page's
464 * order is recorded in page_private(page) field.
465 * So when we are allocating or freeing one, we can derive the state of the
466 * other. That is, if we allocate a small block, and both were
467 * free, the remainder of the region must be split into blocks.
468 * If a block is freed, and its buddy is also free, then this
469 * triggers coalescing into a block of larger size.
474 static inline void __free_one_page(struct page *page,
475 struct zone *zone, unsigned int order,
478 unsigned long page_idx;
479 unsigned long combined_idx;
482 if (unlikely(PageCompound(page)))
483 if (unlikely(destroy_compound_page(page, order)))
486 VM_BUG_ON(migratetype == -1);
488 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
490 VM_BUG_ON(page_idx & ((1 << order) - 1));
491 VM_BUG_ON(bad_range(zone, page));
493 while (order < MAX_ORDER-1) {
494 buddy = __page_find_buddy(page, page_idx, order);
495 if (!page_is_buddy(page, buddy, order))
498 /* Our buddy is free, merge with it and move up one order. */
499 list_del(&buddy->lru);
500 zone->free_area[order].nr_free--;
501 rmv_page_order(buddy);
502 combined_idx = __find_combined_index(page_idx, order);
503 page = page + (combined_idx - page_idx);
504 page_idx = combined_idx;
507 set_page_order(page, order);
510 * If this is not the largest possible page, check if the buddy
511 * of the next-highest order is free. If it is, it's possible
512 * that pages are being freed that will coalesce soon. In case,
513 * that is happening, add the free page to the tail of the list
514 * so it's less likely to be used soon and more likely to be merged
515 * as a higher order page
517 if ((order < MAX_ORDER-1) && pfn_valid_within(page_to_pfn(buddy))) {
518 struct page *higher_page, *higher_buddy;
519 combined_idx = __find_combined_index(page_idx, order);
520 higher_page = page + combined_idx - page_idx;
521 higher_buddy = __page_find_buddy(higher_page, combined_idx, order + 1);
522 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
523 list_add_tail(&page->lru,
524 &zone->free_area[order].free_list[migratetype]);
529 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
531 zone->free_area[order].nr_free++;
535 * free_page_mlock() -- clean up attempts to free and mlocked() page.
536 * Page should not be on lru, so no need to fix that up.
537 * free_pages_check() will verify...
539 static inline void free_page_mlock(struct page *page)
541 __dec_zone_page_state(page, NR_MLOCK);
542 __count_vm_event(UNEVICTABLE_MLOCKFREED);
545 static inline int free_pages_check(struct page *page)
547 if (unlikely(page_mapcount(page) |
548 (page->mapping != NULL) |
549 (atomic_read(&page->_count) != 0) |
550 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) {
554 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
555 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
560 * Frees a number of pages from the PCP lists
561 * Assumes all pages on list are in same zone, and of same order.
562 * count is the number of pages to free.
564 * If the zone was previously in an "all pages pinned" state then look to
565 * see if this freeing clears that state.
567 * And clear the zone's pages_scanned counter, to hold off the "all pages are
568 * pinned" detection logic.
570 static void free_pcppages_bulk(struct zone *zone, int count,
571 struct per_cpu_pages *pcp)
576 spin_lock(&zone->lock);
577 zone->all_unreclaimable = 0;
578 zone->pages_scanned = 0;
580 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
583 struct list_head *list;
586 * Remove pages from lists in a round-robin fashion. A
587 * batch_free count is maintained that is incremented when an
588 * empty list is encountered. This is so more pages are freed
589 * off fuller lists instead of spinning excessively around empty
594 if (++migratetype == MIGRATE_PCPTYPES)
596 list = &pcp->lists[migratetype];
597 } while (list_empty(list));
600 page = list_entry(list->prev, struct page, lru);
601 /* must delete as __free_one_page list manipulates */
602 list_del(&page->lru);
603 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
604 __free_one_page(page, zone, 0, page_private(page));
605 trace_mm_page_pcpu_drain(page, 0, page_private(page));
606 } while (--count && --batch_free && !list_empty(list));
608 spin_unlock(&zone->lock);
611 static void free_one_page(struct zone *zone, struct page *page, int order,
614 spin_lock(&zone->lock);
615 zone->all_unreclaimable = 0;
616 zone->pages_scanned = 0;
618 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
619 __free_one_page(page, zone, order, migratetype);
620 spin_unlock(&zone->lock);
623 static void __free_pages_ok(struct page *page, unsigned int order)
628 int wasMlocked = __TestClearPageMlocked(page);
630 trace_mm_page_free_direct(page, order);
631 kmemcheck_free_shadow(page, order);
633 for (i = 0 ; i < (1 << order) ; ++i)
634 bad += free_pages_check(page + i);
638 if (!PageHighMem(page)) {
639 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
640 debug_check_no_obj_freed(page_address(page),
643 arch_free_page(page, order);
644 kernel_map_pages(page, 1 << order, 0);
646 local_irq_save(flags);
647 if (unlikely(wasMlocked))
648 free_page_mlock(page);
649 __count_vm_events(PGFREE, 1 << order);
650 free_one_page(page_zone(page), page, order,
651 get_pageblock_migratetype(page));
652 local_irq_restore(flags);
656 * permit the bootmem allocator to evade page validation on high-order frees
658 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
661 __ClearPageReserved(page);
662 set_page_count(page, 0);
663 set_page_refcounted(page);
669 for (loop = 0; loop < BITS_PER_LONG; loop++) {
670 struct page *p = &page[loop];
672 if (loop + 1 < BITS_PER_LONG)
674 __ClearPageReserved(p);
675 set_page_count(p, 0);
678 set_page_refcounted(page);
679 __free_pages(page, order);
685 * The order of subdivision here is critical for the IO subsystem.
686 * Please do not alter this order without good reasons and regression
687 * testing. Specifically, as large blocks of memory are subdivided,
688 * the order in which smaller blocks are delivered depends on the order
689 * they're subdivided in this function. This is the primary factor
690 * influencing the order in which pages are delivered to the IO
691 * subsystem according to empirical testing, and this is also justified
692 * by considering the behavior of a buddy system containing a single
693 * large block of memory acted on by a series of small allocations.
694 * This behavior is a critical factor in sglist merging's success.
698 static inline void expand(struct zone *zone, struct page *page,
699 int low, int high, struct free_area *area,
702 unsigned long size = 1 << high;
708 VM_BUG_ON(bad_range(zone, &page[size]));
709 list_add(&page[size].lru, &area->free_list[migratetype]);
711 set_page_order(&page[size], high);
716 * This page is about to be returned from the page allocator
718 static inline int check_new_page(struct page *page)
720 if (unlikely(page_mapcount(page) |
721 (page->mapping != NULL) |
722 (atomic_read(&page->_count) != 0) |
723 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) {
730 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
734 for (i = 0; i < (1 << order); i++) {
735 struct page *p = page + i;
736 if (unlikely(check_new_page(p)))
740 set_page_private(page, 0);
741 set_page_refcounted(page);
743 arch_alloc_page(page, order);
744 kernel_map_pages(page, 1 << order, 1);
746 if (gfp_flags & __GFP_ZERO)
747 prep_zero_page(page, order, gfp_flags);
749 if (order && (gfp_flags & __GFP_COMP))
750 prep_compound_page(page, order);
756 * Go through the free lists for the given migratetype and remove
757 * the smallest available page from the freelists
760 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
763 unsigned int current_order;
764 struct free_area * area;
767 /* Find a page of the appropriate size in the preferred list */
768 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
769 area = &(zone->free_area[current_order]);
770 if (list_empty(&area->free_list[migratetype]))
773 page = list_entry(area->free_list[migratetype].next,
775 list_del(&page->lru);
776 rmv_page_order(page);
778 expand(zone, page, order, current_order, area, migratetype);
787 * This array describes the order lists are fallen back to when
788 * the free lists for the desirable migrate type are depleted
790 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
791 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
792 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
793 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
794 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
798 * Move the free pages in a range to the free lists of the requested type.
799 * Note that start_page and end_pages are not aligned on a pageblock
800 * boundary. If alignment is required, use move_freepages_block()
802 static int move_freepages(struct zone *zone,
803 struct page *start_page, struct page *end_page,
810 #ifndef CONFIG_HOLES_IN_ZONE
812 * page_zone is not safe to call in this context when
813 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
814 * anyway as we check zone boundaries in move_freepages_block().
815 * Remove at a later date when no bug reports exist related to
816 * grouping pages by mobility
818 BUG_ON(page_zone(start_page) != page_zone(end_page));
821 for (page = start_page; page <= end_page;) {
822 /* Make sure we are not inadvertently changing nodes */
823 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
825 if (!pfn_valid_within(page_to_pfn(page))) {
830 if (!PageBuddy(page)) {
835 order = page_order(page);
836 list_del(&page->lru);
838 &zone->free_area[order].free_list[migratetype]);
840 pages_moved += 1 << order;
846 static int move_freepages_block(struct zone *zone, struct page *page,
849 unsigned long start_pfn, end_pfn;
850 struct page *start_page, *end_page;
852 start_pfn = page_to_pfn(page);
853 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
854 start_page = pfn_to_page(start_pfn);
855 end_page = start_page + pageblock_nr_pages - 1;
856 end_pfn = start_pfn + pageblock_nr_pages - 1;
858 /* Do not cross zone boundaries */
859 if (start_pfn < zone->zone_start_pfn)
861 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
864 return move_freepages(zone, start_page, end_page, migratetype);
867 static void change_pageblock_range(struct page *pageblock_page,
868 int start_order, int migratetype)
870 int nr_pageblocks = 1 << (start_order - pageblock_order);
872 while (nr_pageblocks--) {
873 set_pageblock_migratetype(pageblock_page, migratetype);
874 pageblock_page += pageblock_nr_pages;
878 /* Remove an element from the buddy allocator from the fallback list */
879 static inline struct page *
880 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
882 struct free_area * area;
887 /* Find the largest possible block of pages in the other list */
888 for (current_order = MAX_ORDER-1; current_order >= order;
890 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
891 migratetype = fallbacks[start_migratetype][i];
893 /* MIGRATE_RESERVE handled later if necessary */
894 if (migratetype == MIGRATE_RESERVE)
897 area = &(zone->free_area[current_order]);
898 if (list_empty(&area->free_list[migratetype]))
901 page = list_entry(area->free_list[migratetype].next,
906 * If breaking a large block of pages, move all free
907 * pages to the preferred allocation list. If falling
908 * back for a reclaimable kernel allocation, be more
909 * agressive about taking ownership of free pages
911 if (unlikely(current_order >= (pageblock_order >> 1)) ||
912 start_migratetype == MIGRATE_RECLAIMABLE ||
913 page_group_by_mobility_disabled) {
915 pages = move_freepages_block(zone, page,
918 /* Claim the whole block if over half of it is free */
919 if (pages >= (1 << (pageblock_order-1)) ||
920 page_group_by_mobility_disabled)
921 set_pageblock_migratetype(page,
924 migratetype = start_migratetype;
927 /* Remove the page from the freelists */
928 list_del(&page->lru);
929 rmv_page_order(page);
931 /* Take ownership for orders >= pageblock_order */
932 if (current_order >= pageblock_order)
933 change_pageblock_range(page, current_order,
936 expand(zone, page, order, current_order, area, migratetype);
938 trace_mm_page_alloc_extfrag(page, order, current_order,
939 start_migratetype, migratetype);
949 * Do the hard work of removing an element from the buddy allocator.
950 * Call me with the zone->lock already held.
952 static struct page *__rmqueue(struct zone *zone, unsigned int order,
958 page = __rmqueue_smallest(zone, order, migratetype);
960 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
961 page = __rmqueue_fallback(zone, order, migratetype);
964 * Use MIGRATE_RESERVE rather than fail an allocation. goto
965 * is used because __rmqueue_smallest is an inline function
966 * and we want just one call site
969 migratetype = MIGRATE_RESERVE;
974 trace_mm_page_alloc_zone_locked(page, order, migratetype);
979 * Obtain a specified number of elements from the buddy allocator, all under
980 * a single hold of the lock, for efficiency. Add them to the supplied list.
981 * Returns the number of new pages which were placed at *list.
983 static int rmqueue_bulk(struct zone *zone, unsigned int order,
984 unsigned long count, struct list_head *list,
985 int migratetype, int cold)
989 spin_lock(&zone->lock);
990 for (i = 0; i < count; ++i) {
991 struct page *page = __rmqueue(zone, order, migratetype);
992 if (unlikely(page == NULL))
996 * Split buddy pages returned by expand() are received here
997 * in physical page order. The page is added to the callers and
998 * list and the list head then moves forward. From the callers
999 * perspective, the linked list is ordered by page number in
1000 * some conditions. This is useful for IO devices that can
1001 * merge IO requests if the physical pages are ordered
1004 if (likely(cold == 0))
1005 list_add(&page->lru, list);
1007 list_add_tail(&page->lru, list);
1008 set_page_private(page, migratetype);
1011 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1012 spin_unlock(&zone->lock);
1018 * Called from the vmstat counter updater to drain pagesets of this
1019 * currently executing processor on remote nodes after they have
1022 * Note that this function must be called with the thread pinned to
1023 * a single processor.
1025 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1027 unsigned long flags;
1030 local_irq_save(flags);
1031 if (pcp->count >= pcp->batch)
1032 to_drain = pcp->batch;
1034 to_drain = pcp->count;
1035 free_pcppages_bulk(zone, to_drain, pcp);
1036 pcp->count -= to_drain;
1037 local_irq_restore(flags);
1042 * Drain pages of the indicated processor.
1044 * The processor must either be the current processor and the
1045 * thread pinned to the current processor or a processor that
1048 static void drain_pages(unsigned int cpu)
1050 unsigned long flags;
1053 for_each_populated_zone(zone) {
1054 struct per_cpu_pageset *pset;
1055 struct per_cpu_pages *pcp;
1057 local_irq_save(flags);
1058 pset = per_cpu_ptr(zone->pageset, cpu);
1061 free_pcppages_bulk(zone, pcp->count, pcp);
1063 local_irq_restore(flags);
1068 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1070 void drain_local_pages(void *arg)
1072 drain_pages(smp_processor_id());
1076 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1078 void drain_all_pages(void)
1080 on_each_cpu(drain_local_pages, NULL, 1);
1083 #ifdef CONFIG_HIBERNATION
1085 void mark_free_pages(struct zone *zone)
1087 unsigned long pfn, max_zone_pfn;
1088 unsigned long flags;
1090 struct list_head *curr;
1092 if (!zone->spanned_pages)
1095 spin_lock_irqsave(&zone->lock, flags);
1097 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1098 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1099 if (pfn_valid(pfn)) {
1100 struct page *page = pfn_to_page(pfn);
1102 if (!swsusp_page_is_forbidden(page))
1103 swsusp_unset_page_free(page);
1106 for_each_migratetype_order(order, t) {
1107 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1110 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1111 for (i = 0; i < (1UL << order); i++)
1112 swsusp_set_page_free(pfn_to_page(pfn + i));
1115 spin_unlock_irqrestore(&zone->lock, flags);
1117 #endif /* CONFIG_PM */
1120 * Free a 0-order page
1121 * cold == 1 ? free a cold page : free a hot page
1123 void free_hot_cold_page(struct page *page, int cold)
1125 struct zone *zone = page_zone(page);
1126 struct per_cpu_pages *pcp;
1127 unsigned long flags;
1129 int wasMlocked = __TestClearPageMlocked(page);
1131 trace_mm_page_free_direct(page, 0);
1132 kmemcheck_free_shadow(page, 0);
1135 page->mapping = NULL;
1136 if (free_pages_check(page))
1139 if (!PageHighMem(page)) {
1140 debug_check_no_locks_freed(page_address(page), PAGE_SIZE);
1141 debug_check_no_obj_freed(page_address(page), PAGE_SIZE);
1143 arch_free_page(page, 0);
1144 kernel_map_pages(page, 1, 0);
1146 migratetype = get_pageblock_migratetype(page);
1147 set_page_private(page, migratetype);
1148 local_irq_save(flags);
1149 if (unlikely(wasMlocked))
1150 free_page_mlock(page);
1151 __count_vm_event(PGFREE);
1154 * We only track unmovable, reclaimable and movable on pcp lists.
1155 * Free ISOLATE pages back to the allocator because they are being
1156 * offlined but treat RESERVE as movable pages so we can get those
1157 * areas back if necessary. Otherwise, we may have to free
1158 * excessively into the page allocator
1160 if (migratetype >= MIGRATE_PCPTYPES) {
1161 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1162 free_one_page(zone, page, 0, migratetype);
1165 migratetype = MIGRATE_MOVABLE;
1168 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1170 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1172 list_add(&page->lru, &pcp->lists[migratetype]);
1174 if (pcp->count >= pcp->high) {
1175 free_pcppages_bulk(zone, pcp->batch, pcp);
1176 pcp->count -= pcp->batch;
1180 local_irq_restore(flags);
1184 * split_page takes a non-compound higher-order page, and splits it into
1185 * n (1<<order) sub-pages: page[0..n]
1186 * Each sub-page must be freed individually.
1188 * Note: this is probably too low level an operation for use in drivers.
1189 * Please consult with lkml before using this in your driver.
1191 void split_page(struct page *page, unsigned int order)
1195 VM_BUG_ON(PageCompound(page));
1196 VM_BUG_ON(!page_count(page));
1198 #ifdef CONFIG_KMEMCHECK
1200 * Split shadow pages too, because free(page[0]) would
1201 * otherwise free the whole shadow.
1203 if (kmemcheck_page_is_tracked(page))
1204 split_page(virt_to_page(page[0].shadow), order);
1207 for (i = 1; i < (1 << order); i++)
1208 set_page_refcounted(page + i);
1212 * Similar to split_page except the page is already free. As this is only
1213 * being used for migration, the migratetype of the block also changes.
1214 * As this is called with interrupts disabled, the caller is responsible
1215 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1218 * Note: this is probably too low level an operation for use in drivers.
1219 * Please consult with lkml before using this in your driver.
1221 int split_free_page(struct page *page)
1224 unsigned long watermark;
1227 BUG_ON(!PageBuddy(page));
1229 zone = page_zone(page);
1230 order = page_order(page);
1232 /* Obey watermarks as if the page was being allocated */
1233 watermark = low_wmark_pages(zone) + (1 << order);
1234 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1237 /* Remove page from free list */
1238 list_del(&page->lru);
1239 zone->free_area[order].nr_free--;
1240 rmv_page_order(page);
1241 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1243 /* Split into individual pages */
1244 set_page_refcounted(page);
1245 split_page(page, order);
1247 if (order >= pageblock_order - 1) {
1248 struct page *endpage = page + (1 << order) - 1;
1249 for (; page < endpage; page += pageblock_nr_pages)
1250 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1257 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1258 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1262 struct page *buffered_rmqueue(struct zone *preferred_zone,
1263 struct zone *zone, int order, gfp_t gfp_flags,
1266 unsigned long flags;
1268 int cold = !!(gfp_flags & __GFP_COLD);
1271 if (likely(order == 0)) {
1272 struct per_cpu_pages *pcp;
1273 struct list_head *list;
1275 local_irq_save(flags);
1276 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1277 list = &pcp->lists[migratetype];
1278 if (list_empty(list)) {
1279 pcp->count += rmqueue_bulk(zone, 0,
1282 if (unlikely(list_empty(list)))
1287 page = list_entry(list->prev, struct page, lru);
1289 page = list_entry(list->next, struct page, lru);
1291 list_del(&page->lru);
1294 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1296 * __GFP_NOFAIL is not to be used in new code.
1298 * All __GFP_NOFAIL callers should be fixed so that they
1299 * properly detect and handle allocation failures.
1301 * We most definitely don't want callers attempting to
1302 * allocate greater than order-1 page units with
1305 WARN_ON_ONCE(order > 1);
1307 spin_lock_irqsave(&zone->lock, flags);
1308 page = __rmqueue(zone, order, migratetype);
1309 spin_unlock(&zone->lock);
1312 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1315 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1316 zone_statistics(preferred_zone, zone);
1317 local_irq_restore(flags);
1319 VM_BUG_ON(bad_range(zone, page));
1320 if (prep_new_page(page, order, gfp_flags))
1325 local_irq_restore(flags);
1329 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1330 #define ALLOC_WMARK_MIN WMARK_MIN
1331 #define ALLOC_WMARK_LOW WMARK_LOW
1332 #define ALLOC_WMARK_HIGH WMARK_HIGH
1333 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1335 /* Mask to get the watermark bits */
1336 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1338 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1339 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1340 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1342 #ifdef CONFIG_FAIL_PAGE_ALLOC
1344 static struct fail_page_alloc_attr {
1345 struct fault_attr attr;
1347 u32 ignore_gfp_highmem;
1348 u32 ignore_gfp_wait;
1351 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1353 struct dentry *ignore_gfp_highmem_file;
1354 struct dentry *ignore_gfp_wait_file;
1355 struct dentry *min_order_file;
1357 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1359 } fail_page_alloc = {
1360 .attr = FAULT_ATTR_INITIALIZER,
1361 .ignore_gfp_wait = 1,
1362 .ignore_gfp_highmem = 1,
1366 static int __init setup_fail_page_alloc(char *str)
1368 return setup_fault_attr(&fail_page_alloc.attr, str);
1370 __setup("fail_page_alloc=", setup_fail_page_alloc);
1372 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1374 if (order < fail_page_alloc.min_order)
1376 if (gfp_mask & __GFP_NOFAIL)
1378 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1380 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1383 return should_fail(&fail_page_alloc.attr, 1 << order);
1386 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1388 static int __init fail_page_alloc_debugfs(void)
1390 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1394 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1398 dir = fail_page_alloc.attr.dentries.dir;
1400 fail_page_alloc.ignore_gfp_wait_file =
1401 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1402 &fail_page_alloc.ignore_gfp_wait);
1404 fail_page_alloc.ignore_gfp_highmem_file =
1405 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1406 &fail_page_alloc.ignore_gfp_highmem);
1407 fail_page_alloc.min_order_file =
1408 debugfs_create_u32("min-order", mode, dir,
1409 &fail_page_alloc.min_order);
1411 if (!fail_page_alloc.ignore_gfp_wait_file ||
1412 !fail_page_alloc.ignore_gfp_highmem_file ||
1413 !fail_page_alloc.min_order_file) {
1415 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1416 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1417 debugfs_remove(fail_page_alloc.min_order_file);
1418 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1424 late_initcall(fail_page_alloc_debugfs);
1426 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1428 #else /* CONFIG_FAIL_PAGE_ALLOC */
1430 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1435 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1438 * Return 1 if free pages are above 'mark'. This takes into account the order
1439 * of the allocation.
1441 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1442 int classzone_idx, int alloc_flags)
1444 /* free_pages my go negative - that's OK */
1446 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
1449 if (alloc_flags & ALLOC_HIGH)
1451 if (alloc_flags & ALLOC_HARDER)
1454 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1456 for (o = 0; o < order; o++) {
1457 /* At the next order, this order's pages become unavailable */
1458 free_pages -= z->free_area[o].nr_free << o;
1460 /* Require fewer higher order pages to be free */
1463 if (free_pages <= min)
1471 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1472 * skip over zones that are not allowed by the cpuset, or that have
1473 * been recently (in last second) found to be nearly full. See further
1474 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1475 * that have to skip over a lot of full or unallowed zones.
1477 * If the zonelist cache is present in the passed in zonelist, then
1478 * returns a pointer to the allowed node mask (either the current
1479 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1481 * If the zonelist cache is not available for this zonelist, does
1482 * nothing and returns NULL.
1484 * If the fullzones BITMAP in the zonelist cache is stale (more than
1485 * a second since last zap'd) then we zap it out (clear its bits.)
1487 * We hold off even calling zlc_setup, until after we've checked the
1488 * first zone in the zonelist, on the theory that most allocations will
1489 * be satisfied from that first zone, so best to examine that zone as
1490 * quickly as we can.
1492 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1494 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1495 nodemask_t *allowednodes; /* zonelist_cache approximation */
1497 zlc = zonelist->zlcache_ptr;
1501 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1502 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1503 zlc->last_full_zap = jiffies;
1506 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1507 &cpuset_current_mems_allowed :
1508 &node_states[N_HIGH_MEMORY];
1509 return allowednodes;
1513 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1514 * if it is worth looking at further for free memory:
1515 * 1) Check that the zone isn't thought to be full (doesn't have its
1516 * bit set in the zonelist_cache fullzones BITMAP).
1517 * 2) Check that the zones node (obtained from the zonelist_cache
1518 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1519 * Return true (non-zero) if zone is worth looking at further, or
1520 * else return false (zero) if it is not.
1522 * This check -ignores- the distinction between various watermarks,
1523 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1524 * found to be full for any variation of these watermarks, it will
1525 * be considered full for up to one second by all requests, unless
1526 * we are so low on memory on all allowed nodes that we are forced
1527 * into the second scan of the zonelist.
1529 * In the second scan we ignore this zonelist cache and exactly
1530 * apply the watermarks to all zones, even it is slower to do so.
1531 * We are low on memory in the second scan, and should leave no stone
1532 * unturned looking for a free page.
1534 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1535 nodemask_t *allowednodes)
1537 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1538 int i; /* index of *z in zonelist zones */
1539 int n; /* node that zone *z is on */
1541 zlc = zonelist->zlcache_ptr;
1545 i = z - zonelist->_zonerefs;
1548 /* This zone is worth trying if it is allowed but not full */
1549 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1553 * Given 'z' scanning a zonelist, set the corresponding bit in
1554 * zlc->fullzones, so that subsequent attempts to allocate a page
1555 * from that zone don't waste time re-examining it.
1557 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1559 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1560 int i; /* index of *z in zonelist zones */
1562 zlc = zonelist->zlcache_ptr;
1566 i = z - zonelist->_zonerefs;
1568 set_bit(i, zlc->fullzones);
1571 #else /* CONFIG_NUMA */
1573 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1578 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1579 nodemask_t *allowednodes)
1584 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1587 #endif /* CONFIG_NUMA */
1590 * get_page_from_freelist goes through the zonelist trying to allocate
1593 static struct page *
1594 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1595 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1596 struct zone *preferred_zone, int migratetype)
1599 struct page *page = NULL;
1602 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1603 int zlc_active = 0; /* set if using zonelist_cache */
1604 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1606 classzone_idx = zone_idx(preferred_zone);
1609 * Scan zonelist, looking for a zone with enough free.
1610 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1612 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1613 high_zoneidx, nodemask) {
1614 if (NUMA_BUILD && zlc_active &&
1615 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1617 if ((alloc_flags & ALLOC_CPUSET) &&
1618 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1621 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1622 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1626 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1627 if (zone_watermark_ok(zone, order, mark,
1628 classzone_idx, alloc_flags))
1631 if (zone_reclaim_mode == 0)
1632 goto this_zone_full;
1634 ret = zone_reclaim(zone, gfp_mask, order);
1636 case ZONE_RECLAIM_NOSCAN:
1639 case ZONE_RECLAIM_FULL:
1640 /* scanned but unreclaimable */
1641 goto this_zone_full;
1643 /* did we reclaim enough */
1644 if (!zone_watermark_ok(zone, order, mark,
1645 classzone_idx, alloc_flags))
1646 goto this_zone_full;
1651 page = buffered_rmqueue(preferred_zone, zone, order,
1652 gfp_mask, migratetype);
1657 zlc_mark_zone_full(zonelist, z);
1659 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1661 * we do zlc_setup after the first zone is tried but only
1662 * if there are multiple nodes make it worthwhile
1664 allowednodes = zlc_setup(zonelist, alloc_flags);
1670 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1671 /* Disable zlc cache for second zonelist scan */
1679 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1680 unsigned long pages_reclaimed)
1682 /* Do not loop if specifically requested */
1683 if (gfp_mask & __GFP_NORETRY)
1687 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1688 * means __GFP_NOFAIL, but that may not be true in other
1691 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1695 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1696 * specified, then we retry until we no longer reclaim any pages
1697 * (above), or we've reclaimed an order of pages at least as
1698 * large as the allocation's order. In both cases, if the
1699 * allocation still fails, we stop retrying.
1701 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1705 * Don't let big-order allocations loop unless the caller
1706 * explicitly requests that.
1708 if (gfp_mask & __GFP_NOFAIL)
1714 static inline struct page *
1715 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1716 struct zonelist *zonelist, enum zone_type high_zoneidx,
1717 nodemask_t *nodemask, struct zone *preferred_zone,
1722 /* Acquire the OOM killer lock for the zones in zonelist */
1723 if (!try_set_zone_oom(zonelist, gfp_mask)) {
1724 schedule_timeout_uninterruptible(1);
1729 * Go through the zonelist yet one more time, keep very high watermark
1730 * here, this is only to catch a parallel oom killing, we must fail if
1731 * we're still under heavy pressure.
1733 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1734 order, zonelist, high_zoneidx,
1735 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1736 preferred_zone, migratetype);
1740 if (!(gfp_mask & __GFP_NOFAIL)) {
1741 /* The OOM killer will not help higher order allocs */
1742 if (order > PAGE_ALLOC_COSTLY_ORDER)
1745 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1746 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1747 * The caller should handle page allocation failure by itself if
1748 * it specifies __GFP_THISNODE.
1749 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1751 if (gfp_mask & __GFP_THISNODE)
1754 /* Exhausted what can be done so it's blamo time */
1755 out_of_memory(zonelist, gfp_mask, order, nodemask);
1758 clear_zonelist_oom(zonelist, gfp_mask);
1762 #ifdef CONFIG_COMPACTION
1763 /* Try memory compaction for high-order allocations before reclaim */
1764 static struct page *
1765 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1766 struct zonelist *zonelist, enum zone_type high_zoneidx,
1767 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1768 int migratetype, unsigned long *did_some_progress)
1775 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1777 if (*did_some_progress != COMPACT_SKIPPED) {
1779 /* Page migration frees to the PCP lists but we want merging */
1780 drain_pages(get_cpu());
1783 page = get_page_from_freelist(gfp_mask, nodemask,
1784 order, zonelist, high_zoneidx,
1785 alloc_flags, preferred_zone,
1788 count_vm_event(COMPACTSUCCESS);
1793 * It's bad if compaction run occurs and fails.
1794 * The most likely reason is that pages exist,
1795 * but not enough to satisfy watermarks.
1797 count_vm_event(COMPACTFAIL);
1805 static inline struct page *
1806 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1807 struct zonelist *zonelist, enum zone_type high_zoneidx,
1808 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1809 int migratetype, unsigned long *did_some_progress)
1813 #endif /* CONFIG_COMPACTION */
1815 /* The really slow allocator path where we enter direct reclaim */
1816 static inline struct page *
1817 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1818 struct zonelist *zonelist, enum zone_type high_zoneidx,
1819 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1820 int migratetype, unsigned long *did_some_progress)
1822 struct page *page = NULL;
1823 struct reclaim_state reclaim_state;
1824 struct task_struct *p = current;
1828 /* We now go into synchronous reclaim */
1829 cpuset_memory_pressure_bump();
1830 p->flags |= PF_MEMALLOC;
1831 lockdep_set_current_reclaim_state(gfp_mask);
1832 reclaim_state.reclaimed_slab = 0;
1833 p->reclaim_state = &reclaim_state;
1835 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1837 p->reclaim_state = NULL;
1838 lockdep_clear_current_reclaim_state();
1839 p->flags &= ~PF_MEMALLOC;
1846 if (likely(*did_some_progress))
1847 page = get_page_from_freelist(gfp_mask, nodemask, order,
1848 zonelist, high_zoneidx,
1849 alloc_flags, preferred_zone,
1855 * This is called in the allocator slow-path if the allocation request is of
1856 * sufficient urgency to ignore watermarks and take other desperate measures
1858 static inline struct page *
1859 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1860 struct zonelist *zonelist, enum zone_type high_zoneidx,
1861 nodemask_t *nodemask, struct zone *preferred_zone,
1867 page = get_page_from_freelist(gfp_mask, nodemask, order,
1868 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
1869 preferred_zone, migratetype);
1871 if (!page && gfp_mask & __GFP_NOFAIL)
1872 congestion_wait(BLK_RW_ASYNC, HZ/50);
1873 } while (!page && (gfp_mask & __GFP_NOFAIL));
1879 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
1880 enum zone_type high_zoneidx)
1885 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1886 wakeup_kswapd(zone, order);
1890 gfp_to_alloc_flags(gfp_t gfp_mask)
1892 struct task_struct *p = current;
1893 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
1894 const gfp_t wait = gfp_mask & __GFP_WAIT;
1896 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1897 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH);
1900 * The caller may dip into page reserves a bit more if the caller
1901 * cannot run direct reclaim, or if the caller has realtime scheduling
1902 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1903 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1905 alloc_flags |= (gfp_mask & __GFP_HIGH);
1908 alloc_flags |= ALLOC_HARDER;
1910 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1911 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1913 alloc_flags &= ~ALLOC_CPUSET;
1914 } else if (unlikely(rt_task(p)) && !in_interrupt())
1915 alloc_flags |= ALLOC_HARDER;
1917 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
1918 if (!in_interrupt() &&
1919 ((p->flags & PF_MEMALLOC) ||
1920 unlikely(test_thread_flag(TIF_MEMDIE))))
1921 alloc_flags |= ALLOC_NO_WATERMARKS;
1927 static inline struct page *
1928 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
1929 struct zonelist *zonelist, enum zone_type high_zoneidx,
1930 nodemask_t *nodemask, struct zone *preferred_zone,
1933 const gfp_t wait = gfp_mask & __GFP_WAIT;
1934 struct page *page = NULL;
1936 unsigned long pages_reclaimed = 0;
1937 unsigned long did_some_progress;
1938 struct task_struct *p = current;
1941 * In the slowpath, we sanity check order to avoid ever trying to
1942 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1943 * be using allocators in order of preference for an area that is
1946 if (order >= MAX_ORDER) {
1947 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
1952 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1953 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1954 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1955 * using a larger set of nodes after it has established that the
1956 * allowed per node queues are empty and that nodes are
1959 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1963 wake_all_kswapd(order, zonelist, high_zoneidx);
1966 * OK, we're below the kswapd watermark and have kicked background
1967 * reclaim. Now things get more complex, so set up alloc_flags according
1968 * to how we want to proceed.
1970 alloc_flags = gfp_to_alloc_flags(gfp_mask);
1972 /* This is the last chance, in general, before the goto nopage. */
1973 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
1974 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
1975 preferred_zone, migratetype);
1980 /* Allocate without watermarks if the context allows */
1981 if (alloc_flags & ALLOC_NO_WATERMARKS) {
1982 page = __alloc_pages_high_priority(gfp_mask, order,
1983 zonelist, high_zoneidx, nodemask,
1984 preferred_zone, migratetype);
1989 /* Atomic allocations - we can't balance anything */
1993 /* Avoid recursion of direct reclaim */
1994 if (p->flags & PF_MEMALLOC)
1997 /* Avoid allocations with no watermarks from looping endlessly */
1998 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2001 /* Try direct compaction */
2002 page = __alloc_pages_direct_compact(gfp_mask, order,
2003 zonelist, high_zoneidx,
2005 alloc_flags, preferred_zone,
2006 migratetype, &did_some_progress);
2010 /* Try direct reclaim and then allocating */
2011 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2012 zonelist, high_zoneidx,
2014 alloc_flags, preferred_zone,
2015 migratetype, &did_some_progress);
2020 * If we failed to make any progress reclaiming, then we are
2021 * running out of options and have to consider going OOM
2023 if (!did_some_progress) {
2024 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2025 if (oom_killer_disabled)
2027 page = __alloc_pages_may_oom(gfp_mask, order,
2028 zonelist, high_zoneidx,
2029 nodemask, preferred_zone,
2035 * The OOM killer does not trigger for high-order
2036 * ~__GFP_NOFAIL allocations so if no progress is being
2037 * made, there are no other options and retrying is
2040 if (order > PAGE_ALLOC_COSTLY_ORDER &&
2041 !(gfp_mask & __GFP_NOFAIL))
2048 /* Check if we should retry the allocation */
2049 pages_reclaimed += did_some_progress;
2050 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
2051 /* Wait for some write requests to complete then retry */
2052 congestion_wait(BLK_RW_ASYNC, HZ/50);
2057 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
2058 printk(KERN_WARNING "%s: page allocation failure."
2059 " order:%d, mode:0x%x\n",
2060 p->comm, order, gfp_mask);
2066 if (kmemcheck_enabled)
2067 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2073 * This is the 'heart' of the zoned buddy allocator.
2076 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2077 struct zonelist *zonelist, nodemask_t *nodemask)
2079 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2080 struct zone *preferred_zone;
2082 int migratetype = allocflags_to_migratetype(gfp_mask);
2084 gfp_mask &= gfp_allowed_mask;
2086 lockdep_trace_alloc(gfp_mask);
2088 might_sleep_if(gfp_mask & __GFP_WAIT);
2090 if (should_fail_alloc_page(gfp_mask, order))
2094 * Check the zones suitable for the gfp_mask contain at least one
2095 * valid zone. It's possible to have an empty zonelist as a result
2096 * of GFP_THISNODE and a memoryless node
2098 if (unlikely(!zonelist->_zonerefs->zone))
2102 /* The preferred zone is used for statistics later */
2103 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone);
2104 if (!preferred_zone) {
2109 /* First allocation attempt */
2110 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2111 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2112 preferred_zone, migratetype);
2113 if (unlikely(!page))
2114 page = __alloc_pages_slowpath(gfp_mask, order,
2115 zonelist, high_zoneidx, nodemask,
2116 preferred_zone, migratetype);
2119 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2122 EXPORT_SYMBOL(__alloc_pages_nodemask);
2125 * Common helper functions.
2127 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2132 * __get_free_pages() returns a 32-bit address, which cannot represent
2135 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2137 page = alloc_pages(gfp_mask, order);
2140 return (unsigned long) page_address(page);
2142 EXPORT_SYMBOL(__get_free_pages);
2144 unsigned long get_zeroed_page(gfp_t gfp_mask)
2146 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2148 EXPORT_SYMBOL(get_zeroed_page);
2150 void __pagevec_free(struct pagevec *pvec)
2152 int i = pagevec_count(pvec);
2155 trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
2156 free_hot_cold_page(pvec->pages[i], pvec->cold);
2160 void __free_pages(struct page *page, unsigned int order)
2162 if (put_page_testzero(page)) {
2164 free_hot_cold_page(page, 0);
2166 __free_pages_ok(page, order);
2170 EXPORT_SYMBOL(__free_pages);
2172 void free_pages(unsigned long addr, unsigned int order)
2175 VM_BUG_ON(!virt_addr_valid((void *)addr));
2176 __free_pages(virt_to_page((void *)addr), order);
2180 EXPORT_SYMBOL(free_pages);
2183 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2184 * @size: the number of bytes to allocate
2185 * @gfp_mask: GFP flags for the allocation
2187 * This function is similar to alloc_pages(), except that it allocates the
2188 * minimum number of pages to satisfy the request. alloc_pages() can only
2189 * allocate memory in power-of-two pages.
2191 * This function is also limited by MAX_ORDER.
2193 * Memory allocated by this function must be released by free_pages_exact().
2195 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2197 unsigned int order = get_order(size);
2200 addr = __get_free_pages(gfp_mask, order);
2202 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2203 unsigned long used = addr + PAGE_ALIGN(size);
2205 split_page(virt_to_page((void *)addr), order);
2206 while (used < alloc_end) {
2212 return (void *)addr;
2214 EXPORT_SYMBOL(alloc_pages_exact);
2217 * free_pages_exact - release memory allocated via alloc_pages_exact()
2218 * @virt: the value returned by alloc_pages_exact.
2219 * @size: size of allocation, same value as passed to alloc_pages_exact().
2221 * Release the memory allocated by a previous call to alloc_pages_exact.
2223 void free_pages_exact(void *virt, size_t size)
2225 unsigned long addr = (unsigned long)virt;
2226 unsigned long end = addr + PAGE_ALIGN(size);
2228 while (addr < end) {
2233 EXPORT_SYMBOL(free_pages_exact);
2235 static unsigned int nr_free_zone_pages(int offset)
2240 /* Just pick one node, since fallback list is circular */
2241 unsigned int sum = 0;
2243 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2245 for_each_zone_zonelist(zone, z, zonelist, offset) {
2246 unsigned long size = zone->present_pages;
2247 unsigned long high = high_wmark_pages(zone);
2256 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2258 unsigned int nr_free_buffer_pages(void)
2260 return nr_free_zone_pages(gfp_zone(GFP_USER));
2262 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2265 * Amount of free RAM allocatable within all zones
2267 unsigned int nr_free_pagecache_pages(void)
2269 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2272 static inline void show_node(struct zone *zone)
2275 printk("Node %d ", zone_to_nid(zone));
2278 void si_meminfo(struct sysinfo *val)
2280 val->totalram = totalram_pages;
2282 val->freeram = global_page_state(NR_FREE_PAGES);
2283 val->bufferram = nr_blockdev_pages();
2284 val->totalhigh = totalhigh_pages;
2285 val->freehigh = nr_free_highpages();
2286 val->mem_unit = PAGE_SIZE;
2289 EXPORT_SYMBOL(si_meminfo);
2292 void si_meminfo_node(struct sysinfo *val, int nid)
2294 pg_data_t *pgdat = NODE_DATA(nid);
2296 val->totalram = pgdat->node_present_pages;
2297 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2298 #ifdef CONFIG_HIGHMEM
2299 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2300 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2306 val->mem_unit = PAGE_SIZE;
2310 #define K(x) ((x) << (PAGE_SHIFT-10))
2313 * Show free area list (used inside shift_scroll-lock stuff)
2314 * We also calculate the percentage fragmentation. We do this by counting the
2315 * memory on each free list with the exception of the first item on the list.
2317 void show_free_areas(void)
2322 for_each_populated_zone(zone) {
2324 printk("%s per-cpu:\n", zone->name);
2326 for_each_online_cpu(cpu) {
2327 struct per_cpu_pageset *pageset;
2329 pageset = per_cpu_ptr(zone->pageset, cpu);
2331 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2332 cpu, pageset->pcp.high,
2333 pageset->pcp.batch, pageset->pcp.count);
2337 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2338 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2340 " dirty:%lu writeback:%lu unstable:%lu\n"
2341 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2342 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2343 global_page_state(NR_ACTIVE_ANON),
2344 global_page_state(NR_INACTIVE_ANON),
2345 global_page_state(NR_ISOLATED_ANON),
2346 global_page_state(NR_ACTIVE_FILE),
2347 global_page_state(NR_INACTIVE_FILE),
2348 global_page_state(NR_ISOLATED_FILE),
2349 global_page_state(NR_UNEVICTABLE),
2350 global_page_state(NR_FILE_DIRTY),
2351 global_page_state(NR_WRITEBACK),
2352 global_page_state(NR_UNSTABLE_NFS),
2353 global_page_state(NR_FREE_PAGES),
2354 global_page_state(NR_SLAB_RECLAIMABLE),
2355 global_page_state(NR_SLAB_UNRECLAIMABLE),
2356 global_page_state(NR_FILE_MAPPED),
2357 global_page_state(NR_SHMEM),
2358 global_page_state(NR_PAGETABLE),
2359 global_page_state(NR_BOUNCE));
2361 for_each_populated_zone(zone) {
2370 " active_anon:%lukB"
2371 " inactive_anon:%lukB"
2372 " active_file:%lukB"
2373 " inactive_file:%lukB"
2374 " unevictable:%lukB"
2375 " isolated(anon):%lukB"
2376 " isolated(file):%lukB"
2383 " slab_reclaimable:%lukB"
2384 " slab_unreclaimable:%lukB"
2385 " kernel_stack:%lukB"
2389 " writeback_tmp:%lukB"
2390 " pages_scanned:%lu"
2391 " all_unreclaimable? %s"
2394 K(zone_page_state(zone, NR_FREE_PAGES)),
2395 K(min_wmark_pages(zone)),
2396 K(low_wmark_pages(zone)),
2397 K(high_wmark_pages(zone)),
2398 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2399 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2400 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2401 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2402 K(zone_page_state(zone, NR_UNEVICTABLE)),
2403 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2404 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2405 K(zone->present_pages),
2406 K(zone_page_state(zone, NR_MLOCK)),
2407 K(zone_page_state(zone, NR_FILE_DIRTY)),
2408 K(zone_page_state(zone, NR_WRITEBACK)),
2409 K(zone_page_state(zone, NR_FILE_MAPPED)),
2410 K(zone_page_state(zone, NR_SHMEM)),
2411 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2412 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2413 zone_page_state(zone, NR_KERNEL_STACK) *
2415 K(zone_page_state(zone, NR_PAGETABLE)),
2416 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2417 K(zone_page_state(zone, NR_BOUNCE)),
2418 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2419 zone->pages_scanned,
2420 (zone->all_unreclaimable ? "yes" : "no")
2422 printk("lowmem_reserve[]:");
2423 for (i = 0; i < MAX_NR_ZONES; i++)
2424 printk(" %lu", zone->lowmem_reserve[i]);
2428 for_each_populated_zone(zone) {
2429 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2432 printk("%s: ", zone->name);
2434 spin_lock_irqsave(&zone->lock, flags);
2435 for (order = 0; order < MAX_ORDER; order++) {
2436 nr[order] = zone->free_area[order].nr_free;
2437 total += nr[order] << order;
2439 spin_unlock_irqrestore(&zone->lock, flags);
2440 for (order = 0; order < MAX_ORDER; order++)
2441 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2442 printk("= %lukB\n", K(total));
2445 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2447 show_swap_cache_info();
2450 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2452 zoneref->zone = zone;
2453 zoneref->zone_idx = zone_idx(zone);
2457 * Builds allocation fallback zone lists.
2459 * Add all populated zones of a node to the zonelist.
2461 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2462 int nr_zones, enum zone_type zone_type)
2466 BUG_ON(zone_type >= MAX_NR_ZONES);
2471 zone = pgdat->node_zones + zone_type;
2472 if (populated_zone(zone)) {
2473 zoneref_set_zone(zone,
2474 &zonelist->_zonerefs[nr_zones++]);
2475 check_highest_zone(zone_type);
2478 } while (zone_type);
2485 * 0 = automatic detection of better ordering.
2486 * 1 = order by ([node] distance, -zonetype)
2487 * 2 = order by (-zonetype, [node] distance)
2489 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2490 * the same zonelist. So only NUMA can configure this param.
2492 #define ZONELIST_ORDER_DEFAULT 0
2493 #define ZONELIST_ORDER_NODE 1
2494 #define ZONELIST_ORDER_ZONE 2
2496 /* zonelist order in the kernel.
2497 * set_zonelist_order() will set this to NODE or ZONE.
2499 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2500 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2504 /* The value user specified ....changed by config */
2505 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2506 /* string for sysctl */
2507 #define NUMA_ZONELIST_ORDER_LEN 16
2508 char numa_zonelist_order[16] = "default";
2511 * interface for configure zonelist ordering.
2512 * command line option "numa_zonelist_order"
2513 * = "[dD]efault - default, automatic configuration.
2514 * = "[nN]ode - order by node locality, then by zone within node
2515 * = "[zZ]one - order by zone, then by locality within zone
2518 static int __parse_numa_zonelist_order(char *s)
2520 if (*s == 'd' || *s == 'D') {
2521 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2522 } else if (*s == 'n' || *s == 'N') {
2523 user_zonelist_order = ZONELIST_ORDER_NODE;
2524 } else if (*s == 'z' || *s == 'Z') {
2525 user_zonelist_order = ZONELIST_ORDER_ZONE;
2528 "Ignoring invalid numa_zonelist_order value: "
2535 static __init int setup_numa_zonelist_order(char *s)
2538 return __parse_numa_zonelist_order(s);
2541 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2544 * sysctl handler for numa_zonelist_order
2546 int numa_zonelist_order_handler(ctl_table *table, int write,
2547 void __user *buffer, size_t *length,
2550 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2552 static DEFINE_MUTEX(zl_order_mutex);
2554 mutex_lock(&zl_order_mutex);
2556 strcpy(saved_string, (char*)table->data);
2557 ret = proc_dostring(table, write, buffer, length, ppos);
2561 int oldval = user_zonelist_order;
2562 if (__parse_numa_zonelist_order((char*)table->data)) {
2564 * bogus value. restore saved string
2566 strncpy((char*)table->data, saved_string,
2567 NUMA_ZONELIST_ORDER_LEN);
2568 user_zonelist_order = oldval;
2569 } else if (oldval != user_zonelist_order)
2570 build_all_zonelists();
2573 mutex_unlock(&zl_order_mutex);
2578 #define MAX_NODE_LOAD (nr_online_nodes)
2579 static int node_load[MAX_NUMNODES];
2582 * find_next_best_node - find the next node that should appear in a given node's fallback list
2583 * @node: node whose fallback list we're appending
2584 * @used_node_mask: nodemask_t of already used nodes
2586 * We use a number of factors to determine which is the next node that should
2587 * appear on a given node's fallback list. The node should not have appeared
2588 * already in @node's fallback list, and it should be the next closest node
2589 * according to the distance array (which contains arbitrary distance values
2590 * from each node to each node in the system), and should also prefer nodes
2591 * with no CPUs, since presumably they'll have very little allocation pressure
2592 * on them otherwise.
2593 * It returns -1 if no node is found.
2595 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2598 int min_val = INT_MAX;
2600 const struct cpumask *tmp = cpumask_of_node(0);
2602 /* Use the local node if we haven't already */
2603 if (!node_isset(node, *used_node_mask)) {
2604 node_set(node, *used_node_mask);
2608 for_each_node_state(n, N_HIGH_MEMORY) {
2610 /* Don't want a node to appear more than once */
2611 if (node_isset(n, *used_node_mask))
2614 /* Use the distance array to find the distance */
2615 val = node_distance(node, n);
2617 /* Penalize nodes under us ("prefer the next node") */
2620 /* Give preference to headless and unused nodes */
2621 tmp = cpumask_of_node(n);
2622 if (!cpumask_empty(tmp))
2623 val += PENALTY_FOR_NODE_WITH_CPUS;
2625 /* Slight preference for less loaded node */
2626 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2627 val += node_load[n];
2629 if (val < min_val) {
2636 node_set(best_node, *used_node_mask);
2643 * Build zonelists ordered by node and zones within node.
2644 * This results in maximum locality--normal zone overflows into local
2645 * DMA zone, if any--but risks exhausting DMA zone.
2647 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2650 struct zonelist *zonelist;
2652 zonelist = &pgdat->node_zonelists[0];
2653 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2655 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2657 zonelist->_zonerefs[j].zone = NULL;
2658 zonelist->_zonerefs[j].zone_idx = 0;
2662 * Build gfp_thisnode zonelists
2664 static void build_thisnode_zonelists(pg_data_t *pgdat)
2667 struct zonelist *zonelist;
2669 zonelist = &pgdat->node_zonelists[1];
2670 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2671 zonelist->_zonerefs[j].zone = NULL;
2672 zonelist->_zonerefs[j].zone_idx = 0;
2676 * Build zonelists ordered by zone and nodes within zones.
2677 * This results in conserving DMA zone[s] until all Normal memory is
2678 * exhausted, but results in overflowing to remote node while memory
2679 * may still exist in local DMA zone.
2681 static int node_order[MAX_NUMNODES];
2683 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2686 int zone_type; /* needs to be signed */
2688 struct zonelist *zonelist;
2690 zonelist = &pgdat->node_zonelists[0];
2692 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2693 for (j = 0; j < nr_nodes; j++) {
2694 node = node_order[j];
2695 z = &NODE_DATA(node)->node_zones[zone_type];
2696 if (populated_zone(z)) {
2698 &zonelist->_zonerefs[pos++]);
2699 check_highest_zone(zone_type);
2703 zonelist->_zonerefs[pos].zone = NULL;
2704 zonelist->_zonerefs[pos].zone_idx = 0;
2707 static int default_zonelist_order(void)
2710 unsigned long low_kmem_size,total_size;
2714 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2715 * If they are really small and used heavily, the system can fall
2716 * into OOM very easily.
2717 * This function detect ZONE_DMA/DMA32 size and configures zone order.
2719 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2722 for_each_online_node(nid) {
2723 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2724 z = &NODE_DATA(nid)->node_zones[zone_type];
2725 if (populated_zone(z)) {
2726 if (zone_type < ZONE_NORMAL)
2727 low_kmem_size += z->present_pages;
2728 total_size += z->present_pages;
2729 } else if (zone_type == ZONE_NORMAL) {
2731 * If any node has only lowmem, then node order
2732 * is preferred to allow kernel allocations
2733 * locally; otherwise, they can easily infringe
2734 * on other nodes when there is an abundance of
2735 * lowmem available to allocate from.
2737 return ZONELIST_ORDER_NODE;
2741 if (!low_kmem_size || /* there are no DMA area. */
2742 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2743 return ZONELIST_ORDER_NODE;
2745 * look into each node's config.
2746 * If there is a node whose DMA/DMA32 memory is very big area on
2747 * local memory, NODE_ORDER may be suitable.
2749 average_size = total_size /
2750 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2751 for_each_online_node(nid) {
2754 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2755 z = &NODE_DATA(nid)->node_zones[zone_type];
2756 if (populated_zone(z)) {
2757 if (zone_type < ZONE_NORMAL)
2758 low_kmem_size += z->present_pages;
2759 total_size += z->present_pages;
2762 if (low_kmem_size &&
2763 total_size > average_size && /* ignore small node */
2764 low_kmem_size > total_size * 70/100)
2765 return ZONELIST_ORDER_NODE;
2767 return ZONELIST_ORDER_ZONE;
2770 static void set_zonelist_order(void)
2772 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2773 current_zonelist_order = default_zonelist_order();
2775 current_zonelist_order = user_zonelist_order;
2778 static void build_zonelists(pg_data_t *pgdat)
2782 nodemask_t used_mask;
2783 int local_node, prev_node;
2784 struct zonelist *zonelist;
2785 int order = current_zonelist_order;
2787 /* initialize zonelists */
2788 for (i = 0; i < MAX_ZONELISTS; i++) {
2789 zonelist = pgdat->node_zonelists + i;
2790 zonelist->_zonerefs[0].zone = NULL;
2791 zonelist->_zonerefs[0].zone_idx = 0;
2794 /* NUMA-aware ordering of nodes */
2795 local_node = pgdat->node_id;
2796 load = nr_online_nodes;
2797 prev_node = local_node;
2798 nodes_clear(used_mask);
2800 memset(node_order, 0, sizeof(node_order));
2803 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2804 int distance = node_distance(local_node, node);
2807 * If another node is sufficiently far away then it is better
2808 * to reclaim pages in a zone before going off node.
2810 if (distance > RECLAIM_DISTANCE)
2811 zone_reclaim_mode = 1;
2814 * We don't want to pressure a particular node.
2815 * So adding penalty to the first node in same
2816 * distance group to make it round-robin.
2818 if (distance != node_distance(local_node, prev_node))
2819 node_load[node] = load;
2823 if (order == ZONELIST_ORDER_NODE)
2824 build_zonelists_in_node_order(pgdat, node);
2826 node_order[j++] = node; /* remember order */
2829 if (order == ZONELIST_ORDER_ZONE) {
2830 /* calculate node order -- i.e., DMA last! */
2831 build_zonelists_in_zone_order(pgdat, j);
2834 build_thisnode_zonelists(pgdat);
2837 /* Construct the zonelist performance cache - see further mmzone.h */
2838 static void build_zonelist_cache(pg_data_t *pgdat)
2840 struct zonelist *zonelist;
2841 struct zonelist_cache *zlc;
2844 zonelist = &pgdat->node_zonelists[0];
2845 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2846 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2847 for (z = zonelist->_zonerefs; z->zone; z++)
2848 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2852 #else /* CONFIG_NUMA */
2854 static void set_zonelist_order(void)
2856 current_zonelist_order = ZONELIST_ORDER_ZONE;
2859 static void build_zonelists(pg_data_t *pgdat)
2861 int node, local_node;
2863 struct zonelist *zonelist;
2865 local_node = pgdat->node_id;
2867 zonelist = &pgdat->node_zonelists[0];
2868 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2871 * Now we build the zonelist so that it contains the zones
2872 * of all the other nodes.
2873 * We don't want to pressure a particular node, so when
2874 * building the zones for node N, we make sure that the
2875 * zones coming right after the local ones are those from
2876 * node N+1 (modulo N)
2878 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2879 if (!node_online(node))
2881 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2884 for (node = 0; node < local_node; node++) {
2885 if (!node_online(node))
2887 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2891 zonelist->_zonerefs[j].zone = NULL;
2892 zonelist->_zonerefs[j].zone_idx = 0;
2895 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2896 static void build_zonelist_cache(pg_data_t *pgdat)
2898 pgdat->node_zonelists[0].zlcache_ptr = NULL;
2901 #endif /* CONFIG_NUMA */
2904 * Boot pageset table. One per cpu which is going to be used for all
2905 * zones and all nodes. The parameters will be set in such a way
2906 * that an item put on a list will immediately be handed over to
2907 * the buddy list. This is safe since pageset manipulation is done
2908 * with interrupts disabled.
2910 * The boot_pagesets must be kept even after bootup is complete for
2911 * unused processors and/or zones. They do play a role for bootstrapping
2912 * hotplugged processors.
2914 * zoneinfo_show() and maybe other functions do
2915 * not check if the processor is online before following the pageset pointer.
2916 * Other parts of the kernel may not check if the zone is available.
2918 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
2919 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
2921 /* return values int ....just for stop_machine() */
2922 static int __build_all_zonelists(void *dummy)
2928 memset(node_load, 0, sizeof(node_load));
2930 for_each_online_node(nid) {
2931 pg_data_t *pgdat = NODE_DATA(nid);
2933 build_zonelists(pgdat);
2934 build_zonelist_cache(pgdat);
2938 * Initialize the boot_pagesets that are going to be used
2939 * for bootstrapping processors. The real pagesets for
2940 * each zone will be allocated later when the per cpu
2941 * allocator is available.
2943 * boot_pagesets are used also for bootstrapping offline
2944 * cpus if the system is already booted because the pagesets
2945 * are needed to initialize allocators on a specific cpu too.
2946 * F.e. the percpu allocator needs the page allocator which
2947 * needs the percpu allocator in order to allocate its pagesets
2948 * (a chicken-egg dilemma).
2950 for_each_possible_cpu(cpu)
2951 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
2956 void build_all_zonelists(void)
2958 set_zonelist_order();
2960 if (system_state == SYSTEM_BOOTING) {
2961 __build_all_zonelists(NULL);
2962 mminit_verify_zonelist();
2963 cpuset_init_current_mems_allowed();
2965 /* we have to stop all cpus to guarantee there is no user
2967 stop_machine(__build_all_zonelists, NULL, NULL);
2968 /* cpuset refresh routine should be here */
2970 vm_total_pages = nr_free_pagecache_pages();
2972 * Disable grouping by mobility if the number of pages in the
2973 * system is too low to allow the mechanism to work. It would be
2974 * more accurate, but expensive to check per-zone. This check is
2975 * made on memory-hotadd so a system can start with mobility
2976 * disabled and enable it later
2978 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
2979 page_group_by_mobility_disabled = 1;
2981 page_group_by_mobility_disabled = 0;
2983 printk("Built %i zonelists in %s order, mobility grouping %s. "
2984 "Total pages: %ld\n",
2986 zonelist_order_name[current_zonelist_order],
2987 page_group_by_mobility_disabled ? "off" : "on",
2990 printk("Policy zone: %s\n", zone_names[policy_zone]);
2995 * Helper functions to size the waitqueue hash table.
2996 * Essentially these want to choose hash table sizes sufficiently
2997 * large so that collisions trying to wait on pages are rare.
2998 * But in fact, the number of active page waitqueues on typical
2999 * systems is ridiculously low, less than 200. So this is even
3000 * conservative, even though it seems large.
3002 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3003 * waitqueues, i.e. the size of the waitq table given the number of pages.
3005 #define PAGES_PER_WAITQUEUE 256
3007 #ifndef CONFIG_MEMORY_HOTPLUG
3008 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3010 unsigned long size = 1;
3012 pages /= PAGES_PER_WAITQUEUE;
3014 while (size < pages)
3018 * Once we have dozens or even hundreds of threads sleeping
3019 * on IO we've got bigger problems than wait queue collision.
3020 * Limit the size of the wait table to a reasonable size.
3022 size = min(size, 4096UL);
3024 return max(size, 4UL);
3028 * A zone's size might be changed by hot-add, so it is not possible to determine
3029 * a suitable size for its wait_table. So we use the maximum size now.
3031 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3033 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3034 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3035 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3037 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3038 * or more by the traditional way. (See above). It equals:
3040 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3041 * ia64(16K page size) : = ( 8G + 4M)byte.
3042 * powerpc (64K page size) : = (32G +16M)byte.
3044 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3051 * This is an integer logarithm so that shifts can be used later
3052 * to extract the more random high bits from the multiplicative
3053 * hash function before the remainder is taken.
3055 static inline unsigned long wait_table_bits(unsigned long size)
3060 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3063 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3064 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3065 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3066 * higher will lead to a bigger reserve which will get freed as contiguous
3067 * blocks as reclaim kicks in
3069 static void setup_zone_migrate_reserve(struct zone *zone)
3071 unsigned long start_pfn, pfn, end_pfn;
3073 unsigned long block_migratetype;
3076 /* Get the start pfn, end pfn and the number of blocks to reserve */
3077 start_pfn = zone->zone_start_pfn;
3078 end_pfn = start_pfn + zone->spanned_pages;
3079 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3083 * Reserve blocks are generally in place to help high-order atomic
3084 * allocations that are short-lived. A min_free_kbytes value that
3085 * would result in more than 2 reserve blocks for atomic allocations
3086 * is assumed to be in place to help anti-fragmentation for the
3087 * future allocation of hugepages at runtime.
3089 reserve = min(2, reserve);
3091 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3092 if (!pfn_valid(pfn))
3094 page = pfn_to_page(pfn);
3096 /* Watch out for overlapping nodes */
3097 if (page_to_nid(page) != zone_to_nid(zone))
3100 /* Blocks with reserved pages will never free, skip them. */
3101 if (PageReserved(page))
3104 block_migratetype = get_pageblock_migratetype(page);
3106 /* If this block is reserved, account for it */
3107 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
3112 /* Suitable for reserving if this block is movable */
3113 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
3114 set_pageblock_migratetype(page, MIGRATE_RESERVE);
3115 move_freepages_block(zone, page, MIGRATE_RESERVE);
3121 * If the reserve is met and this is a previous reserved block,
3124 if (block_migratetype == MIGRATE_RESERVE) {
3125 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3126 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3132 * Initially all pages are reserved - free ones are freed
3133 * up by free_all_bootmem() once the early boot process is
3134 * done. Non-atomic initialization, single-pass.
3136 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3137 unsigned long start_pfn, enum memmap_context context)
3140 unsigned long end_pfn = start_pfn + size;
3144 if (highest_memmap_pfn < end_pfn - 1)
3145 highest_memmap_pfn = end_pfn - 1;
3147 z = &NODE_DATA(nid)->node_zones[zone];
3148 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3150 * There can be holes in boot-time mem_map[]s
3151 * handed to this function. They do not
3152 * exist on hotplugged memory.
3154 if (context == MEMMAP_EARLY) {
3155 if (!early_pfn_valid(pfn))
3157 if (!early_pfn_in_nid(pfn, nid))
3160 page = pfn_to_page(pfn);
3161 set_page_links(page, zone, nid, pfn);
3162 mminit_verify_page_links(page, zone, nid, pfn);
3163 init_page_count(page);
3164 reset_page_mapcount(page);
3165 SetPageReserved(page);
3167 * Mark the block movable so that blocks are reserved for
3168 * movable at startup. This will force kernel allocations
3169 * to reserve their blocks rather than leaking throughout
3170 * the address space during boot when many long-lived
3171 * kernel allocations are made. Later some blocks near
3172 * the start are marked MIGRATE_RESERVE by
3173 * setup_zone_migrate_reserve()
3175 * bitmap is created for zone's valid pfn range. but memmap
3176 * can be created for invalid pages (for alignment)
3177 * check here not to call set_pageblock_migratetype() against
3180 if ((z->zone_start_pfn <= pfn)
3181 && (pfn < z->zone_start_pfn + z->spanned_pages)
3182 && !(pfn & (pageblock_nr_pages - 1)))
3183 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3185 INIT_LIST_HEAD(&page->lru);
3186 #ifdef WANT_PAGE_VIRTUAL
3187 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3188 if (!is_highmem_idx(zone))
3189 set_page_address(page, __va(pfn << PAGE_SHIFT));
3194 static void __meminit zone_init_free_lists(struct zone *zone)
3197 for_each_migratetype_order(order, t) {
3198 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3199 zone->free_area[order].nr_free = 0;
3203 #ifndef __HAVE_ARCH_MEMMAP_INIT
3204 #define memmap_init(size, nid, zone, start_pfn) \
3205 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3208 static int zone_batchsize(struct zone *zone)
3214 * The per-cpu-pages pools are set to around 1000th of the
3215 * size of the zone. But no more than 1/2 of a meg.
3217 * OK, so we don't know how big the cache is. So guess.
3219 batch = zone->present_pages / 1024;
3220 if (batch * PAGE_SIZE > 512 * 1024)
3221 batch = (512 * 1024) / PAGE_SIZE;
3222 batch /= 4; /* We effectively *= 4 below */
3227 * Clamp the batch to a 2^n - 1 value. Having a power
3228 * of 2 value was found to be more likely to have
3229 * suboptimal cache aliasing properties in some cases.
3231 * For example if 2 tasks are alternately allocating
3232 * batches of pages, one task can end up with a lot
3233 * of pages of one half of the possible page colors
3234 * and the other with pages of the other colors.
3236 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3241 /* The deferral and batching of frees should be suppressed under NOMMU
3244 * The problem is that NOMMU needs to be able to allocate large chunks
3245 * of contiguous memory as there's no hardware page translation to
3246 * assemble apparent contiguous memory from discontiguous pages.
3248 * Queueing large contiguous runs of pages for batching, however,
3249 * causes the pages to actually be freed in smaller chunks. As there
3250 * can be a significant delay between the individual batches being
3251 * recycled, this leads to the once large chunks of space being
3252 * fragmented and becoming unavailable for high-order allocations.
3258 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3260 struct per_cpu_pages *pcp;
3263 memset(p, 0, sizeof(*p));
3267 pcp->high = 6 * batch;
3268 pcp->batch = max(1UL, 1 * batch);
3269 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3270 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3274 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3275 * to the value high for the pageset p.
3278 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3281 struct per_cpu_pages *pcp;
3285 pcp->batch = max(1UL, high/4);
3286 if ((high/4) > (PAGE_SHIFT * 8))
3287 pcp->batch = PAGE_SHIFT * 8;
3291 * Allocate per cpu pagesets and initialize them.
3292 * Before this call only boot pagesets were available.
3293 * Boot pagesets will no longer be used by this processorr
3294 * after setup_per_cpu_pageset().
3296 void __init setup_per_cpu_pageset(void)
3301 for_each_populated_zone(zone) {
3302 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3304 for_each_possible_cpu(cpu) {
3305 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3307 setup_pageset(pcp, zone_batchsize(zone));
3309 if (percpu_pagelist_fraction)
3310 setup_pagelist_highmark(pcp,
3311 (zone->present_pages /
3312 percpu_pagelist_fraction));
3317 static noinline __init_refok
3318 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3321 struct pglist_data *pgdat = zone->zone_pgdat;
3325 * The per-page waitqueue mechanism uses hashed waitqueues
3328 zone->wait_table_hash_nr_entries =
3329 wait_table_hash_nr_entries(zone_size_pages);
3330 zone->wait_table_bits =
3331 wait_table_bits(zone->wait_table_hash_nr_entries);
3332 alloc_size = zone->wait_table_hash_nr_entries
3333 * sizeof(wait_queue_head_t);
3335 if (!slab_is_available()) {
3336 zone->wait_table = (wait_queue_head_t *)
3337 alloc_bootmem_node(pgdat, alloc_size);
3340 * This case means that a zone whose size was 0 gets new memory
3341 * via memory hot-add.
3342 * But it may be the case that a new node was hot-added. In
3343 * this case vmalloc() will not be able to use this new node's
3344 * memory - this wait_table must be initialized to use this new
3345 * node itself as well.
3346 * To use this new node's memory, further consideration will be
3349 zone->wait_table = vmalloc(alloc_size);
3351 if (!zone->wait_table)
3354 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3355 init_waitqueue_head(zone->wait_table + i);
3360 static int __zone_pcp_update(void *data)
3362 struct zone *zone = data;
3364 unsigned long batch = zone_batchsize(zone), flags;
3366 for_each_possible_cpu(cpu) {
3367 struct per_cpu_pageset *pset;
3368 struct per_cpu_pages *pcp;
3370 pset = per_cpu_ptr(zone->pageset, cpu);
3373 local_irq_save(flags);
3374 free_pcppages_bulk(zone, pcp->count, pcp);
3375 setup_pageset(pset, batch);
3376 local_irq_restore(flags);
3381 void zone_pcp_update(struct zone *zone)
3383 stop_machine(__zone_pcp_update, zone, NULL);
3386 static __meminit void zone_pcp_init(struct zone *zone)
3389 * per cpu subsystem is not up at this point. The following code
3390 * relies on the ability of the linker to provide the
3391 * offset of a (static) per cpu variable into the per cpu area.
3393 zone->pageset = &boot_pageset;
3395 if (zone->present_pages)
3396 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3397 zone->name, zone->present_pages,
3398 zone_batchsize(zone));
3401 __meminit int init_currently_empty_zone(struct zone *zone,
3402 unsigned long zone_start_pfn,
3404 enum memmap_context context)
3406 struct pglist_data *pgdat = zone->zone_pgdat;
3408 ret = zone_wait_table_init(zone, size);
3411 pgdat->nr_zones = zone_idx(zone) + 1;
3413 zone->zone_start_pfn = zone_start_pfn;
3415 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3416 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3418 (unsigned long)zone_idx(zone),
3419 zone_start_pfn, (zone_start_pfn + size));
3421 zone_init_free_lists(zone);
3426 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3428 * Basic iterator support. Return the first range of PFNs for a node
3429 * Note: nid == MAX_NUMNODES returns first region regardless of node
3431 static int __meminit first_active_region_index_in_nid(int nid)
3435 for (i = 0; i < nr_nodemap_entries; i++)
3436 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3443 * Basic iterator support. Return the next active range of PFNs for a node
3444 * Note: nid == MAX_NUMNODES returns next region regardless of node
3446 static int __meminit next_active_region_index_in_nid(int index, int nid)
3448 for (index = index + 1; index < nr_nodemap_entries; index++)
3449 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3455 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3457 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3458 * Architectures may implement their own version but if add_active_range()
3459 * was used and there are no special requirements, this is a convenient
3462 int __meminit __early_pfn_to_nid(unsigned long pfn)
3466 for (i = 0; i < nr_nodemap_entries; i++) {
3467 unsigned long start_pfn = early_node_map[i].start_pfn;
3468 unsigned long end_pfn = early_node_map[i].end_pfn;
3470 if (start_pfn <= pfn && pfn < end_pfn)
3471 return early_node_map[i].nid;
3473 /* This is a memory hole */
3476 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3478 int __meminit early_pfn_to_nid(unsigned long pfn)
3482 nid = __early_pfn_to_nid(pfn);
3485 /* just returns 0 */
3489 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3490 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3494 nid = __early_pfn_to_nid(pfn);
3495 if (nid >= 0 && nid != node)
3501 /* Basic iterator support to walk early_node_map[] */
3502 #define for_each_active_range_index_in_nid(i, nid) \
3503 for (i = first_active_region_index_in_nid(nid); i != -1; \
3504 i = next_active_region_index_in_nid(i, nid))
3507 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3508 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3509 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3511 * If an architecture guarantees that all ranges registered with
3512 * add_active_ranges() contain no holes and may be freed, this
3513 * this function may be used instead of calling free_bootmem() manually.
3515 void __init free_bootmem_with_active_regions(int nid,
3516 unsigned long max_low_pfn)
3520 for_each_active_range_index_in_nid(i, nid) {
3521 unsigned long size_pages = 0;
3522 unsigned long end_pfn = early_node_map[i].end_pfn;
3524 if (early_node_map[i].start_pfn >= max_low_pfn)
3527 if (end_pfn > max_low_pfn)
3528 end_pfn = max_low_pfn;
3530 size_pages = end_pfn - early_node_map[i].start_pfn;
3531 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3532 PFN_PHYS(early_node_map[i].start_pfn),
3533 size_pages << PAGE_SHIFT);
3537 int __init add_from_early_node_map(struct range *range, int az,
3538 int nr_range, int nid)
3543 /* need to go over early_node_map to find out good range for node */
3544 for_each_active_range_index_in_nid(i, nid) {
3545 start = early_node_map[i].start_pfn;
3546 end = early_node_map[i].end_pfn;
3547 nr_range = add_range(range, az, nr_range, start, end);
3552 #ifdef CONFIG_NO_BOOTMEM
3553 void * __init __alloc_memory_core_early(int nid, u64 size, u64 align,
3554 u64 goal, u64 limit)
3559 /* need to go over early_node_map to find out good range for node */
3560 for_each_active_range_index_in_nid(i, nid) {
3562 u64 ei_start, ei_last;
3564 ei_last = early_node_map[i].end_pfn;
3565 ei_last <<= PAGE_SHIFT;
3566 ei_start = early_node_map[i].start_pfn;
3567 ei_start <<= PAGE_SHIFT;
3568 addr = find_early_area(ei_start, ei_last,
3569 goal, limit, size, align);
3575 printk(KERN_DEBUG "alloc (nid=%d %llx - %llx) (%llx - %llx) %llx %llx => %llx\n",
3577 ei_start, ei_last, goal, limit, size,
3581 ptr = phys_to_virt(addr);
3582 memset(ptr, 0, size);
3583 reserve_early_without_check(addr, addr + size, "BOOTMEM");
3592 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3597 for_each_active_range_index_in_nid(i, nid) {
3598 ret = work_fn(early_node_map[i].start_pfn,
3599 early_node_map[i].end_pfn, data);
3605 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3606 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3608 * If an architecture guarantees that all ranges registered with
3609 * add_active_ranges() contain no holes and may be freed, this
3610 * function may be used instead of calling memory_present() manually.
3612 void __init sparse_memory_present_with_active_regions(int nid)
3616 for_each_active_range_index_in_nid(i, nid)
3617 memory_present(early_node_map[i].nid,
3618 early_node_map[i].start_pfn,
3619 early_node_map[i].end_pfn);
3623 * get_pfn_range_for_nid - Return the start and end page frames for a node
3624 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3625 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3626 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3628 * It returns the start and end page frame of a node based on information
3629 * provided by an arch calling add_active_range(). If called for a node
3630 * with no available memory, a warning is printed and the start and end
3633 void __meminit get_pfn_range_for_nid(unsigned int nid,
3634 unsigned long *start_pfn, unsigned long *end_pfn)
3640 for_each_active_range_index_in_nid(i, nid) {
3641 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3642 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3645 if (*start_pfn == -1UL)
3650 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3651 * assumption is made that zones within a node are ordered in monotonic
3652 * increasing memory addresses so that the "highest" populated zone is used
3654 static void __init find_usable_zone_for_movable(void)
3657 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3658 if (zone_index == ZONE_MOVABLE)
3661 if (arch_zone_highest_possible_pfn[zone_index] >
3662 arch_zone_lowest_possible_pfn[zone_index])
3666 VM_BUG_ON(zone_index == -1);
3667 movable_zone = zone_index;
3671 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3672 * because it is sized independant of architecture. Unlike the other zones,
3673 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3674 * in each node depending on the size of each node and how evenly kernelcore
3675 * is distributed. This helper function adjusts the zone ranges
3676 * provided by the architecture for a given node by using the end of the
3677 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3678 * zones within a node are in order of monotonic increases memory addresses
3680 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3681 unsigned long zone_type,
3682 unsigned long node_start_pfn,
3683 unsigned long node_end_pfn,
3684 unsigned long *zone_start_pfn,
3685 unsigned long *zone_end_pfn)
3687 /* Only adjust if ZONE_MOVABLE is on this node */
3688 if (zone_movable_pfn[nid]) {
3689 /* Size ZONE_MOVABLE */
3690 if (zone_type == ZONE_MOVABLE) {
3691 *zone_start_pfn = zone_movable_pfn[nid];
3692 *zone_end_pfn = min(node_end_pfn,
3693 arch_zone_highest_possible_pfn[movable_zone]);
3695 /* Adjust for ZONE_MOVABLE starting within this range */
3696 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3697 *zone_end_pfn > zone_movable_pfn[nid]) {
3698 *zone_end_pfn = zone_movable_pfn[nid];
3700 /* Check if this whole range is within ZONE_MOVABLE */
3701 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3702 *zone_start_pfn = *zone_end_pfn;
3707 * Return the number of pages a zone spans in a node, including holes
3708 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3710 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3711 unsigned long zone_type,
3712 unsigned long *ignored)
3714 unsigned long node_start_pfn, node_end_pfn;
3715 unsigned long zone_start_pfn, zone_end_pfn;
3717 /* Get the start and end of the node and zone */
3718 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3719 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3720 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3721 adjust_zone_range_for_zone_movable(nid, zone_type,
3722 node_start_pfn, node_end_pfn,
3723 &zone_start_pfn, &zone_end_pfn);
3725 /* Check that this node has pages within the zone's required range */
3726 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3729 /* Move the zone boundaries inside the node if necessary */
3730 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3731 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3733 /* Return the spanned pages */
3734 return zone_end_pfn - zone_start_pfn;
3738 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3739 * then all holes in the requested range will be accounted for.
3741 unsigned long __meminit __absent_pages_in_range(int nid,
3742 unsigned long range_start_pfn,
3743 unsigned long range_end_pfn)
3746 unsigned long prev_end_pfn = 0, hole_pages = 0;
3747 unsigned long start_pfn;
3749 /* Find the end_pfn of the first active range of pfns in the node */
3750 i = first_active_region_index_in_nid(nid);
3754 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3756 /* Account for ranges before physical memory on this node */
3757 if (early_node_map[i].start_pfn > range_start_pfn)
3758 hole_pages = prev_end_pfn - range_start_pfn;
3760 /* Find all holes for the zone within the node */
3761 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3763 /* No need to continue if prev_end_pfn is outside the zone */
3764 if (prev_end_pfn >= range_end_pfn)
3767 /* Make sure the end of the zone is not within the hole */
3768 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3769 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3771 /* Update the hole size cound and move on */
3772 if (start_pfn > range_start_pfn) {
3773 BUG_ON(prev_end_pfn > start_pfn);
3774 hole_pages += start_pfn - prev_end_pfn;
3776 prev_end_pfn = early_node_map[i].end_pfn;
3779 /* Account for ranges past physical memory on this node */
3780 if (range_end_pfn > prev_end_pfn)
3781 hole_pages += range_end_pfn -
3782 max(range_start_pfn, prev_end_pfn);
3788 * absent_pages_in_range - Return number of page frames in holes within a range
3789 * @start_pfn: The start PFN to start searching for holes
3790 * @end_pfn: The end PFN to stop searching for holes
3792 * It returns the number of pages frames in memory holes within a range.
3794 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3795 unsigned long end_pfn)
3797 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3800 /* Return the number of page frames in holes in a zone on a node */
3801 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3802 unsigned long zone_type,
3803 unsigned long *ignored)
3805 unsigned long node_start_pfn, node_end_pfn;
3806 unsigned long zone_start_pfn, zone_end_pfn;
3808 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3809 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3811 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3814 adjust_zone_range_for_zone_movable(nid, zone_type,
3815 node_start_pfn, node_end_pfn,
3816 &zone_start_pfn, &zone_end_pfn);
3817 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3821 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3822 unsigned long zone_type,
3823 unsigned long *zones_size)
3825 return zones_size[zone_type];
3828 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3829 unsigned long zone_type,
3830 unsigned long *zholes_size)
3835 return zholes_size[zone_type];
3840 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3841 unsigned long *zones_size, unsigned long *zholes_size)
3843 unsigned long realtotalpages, totalpages = 0;
3846 for (i = 0; i < MAX_NR_ZONES; i++)
3847 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3849 pgdat->node_spanned_pages = totalpages;
3851 realtotalpages = totalpages;
3852 for (i = 0; i < MAX_NR_ZONES; i++)
3854 zone_absent_pages_in_node(pgdat->node_id, i,
3856 pgdat->node_present_pages = realtotalpages;
3857 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3861 #ifndef CONFIG_SPARSEMEM
3863 * Calculate the size of the zone->blockflags rounded to an unsigned long
3864 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3865 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3866 * round what is now in bits to nearest long in bits, then return it in
3869 static unsigned long __init usemap_size(unsigned long zonesize)
3871 unsigned long usemapsize;
3873 usemapsize = roundup(zonesize, pageblock_nr_pages);
3874 usemapsize = usemapsize >> pageblock_order;
3875 usemapsize *= NR_PAGEBLOCK_BITS;
3876 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
3878 return usemapsize / 8;
3881 static void __init setup_usemap(struct pglist_data *pgdat,
3882 struct zone *zone, unsigned long zonesize)
3884 unsigned long usemapsize = usemap_size(zonesize);
3885 zone->pageblock_flags = NULL;
3887 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
3890 static void inline setup_usemap(struct pglist_data *pgdat,
3891 struct zone *zone, unsigned long zonesize) {}
3892 #endif /* CONFIG_SPARSEMEM */
3894 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3896 /* Return a sensible default order for the pageblock size. */
3897 static inline int pageblock_default_order(void)
3899 if (HPAGE_SHIFT > PAGE_SHIFT)
3900 return HUGETLB_PAGE_ORDER;
3905 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3906 static inline void __init set_pageblock_order(unsigned int order)
3908 /* Check that pageblock_nr_pages has not already been setup */
3909 if (pageblock_order)
3913 * Assume the largest contiguous order of interest is a huge page.
3914 * This value may be variable depending on boot parameters on IA64
3916 pageblock_order = order;
3918 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3921 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3922 * and pageblock_default_order() are unused as pageblock_order is set
3923 * at compile-time. See include/linux/pageblock-flags.h for the values of
3924 * pageblock_order based on the kernel config
3926 static inline int pageblock_default_order(unsigned int order)
3930 #define set_pageblock_order(x) do {} while (0)
3932 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3935 * Set up the zone data structures:
3936 * - mark all pages reserved
3937 * - mark all memory queues empty
3938 * - clear the memory bitmaps
3940 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
3941 unsigned long *zones_size, unsigned long *zholes_size)
3944 int nid = pgdat->node_id;
3945 unsigned long zone_start_pfn = pgdat->node_start_pfn;
3948 pgdat_resize_init(pgdat);
3949 pgdat->nr_zones = 0;
3950 init_waitqueue_head(&pgdat->kswapd_wait);
3951 pgdat->kswapd_max_order = 0;
3952 pgdat_page_cgroup_init(pgdat);
3954 for (j = 0; j < MAX_NR_ZONES; j++) {
3955 struct zone *zone = pgdat->node_zones + j;
3956 unsigned long size, realsize, memmap_pages;
3959 size = zone_spanned_pages_in_node(nid, j, zones_size);
3960 realsize = size - zone_absent_pages_in_node(nid, j,
3964 * Adjust realsize so that it accounts for how much memory
3965 * is used by this zone for memmap. This affects the watermark
3966 * and per-cpu initialisations
3969 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
3970 if (realsize >= memmap_pages) {
3971 realsize -= memmap_pages;
3974 " %s zone: %lu pages used for memmap\n",
3975 zone_names[j], memmap_pages);
3978 " %s zone: %lu pages exceeds realsize %lu\n",
3979 zone_names[j], memmap_pages, realsize);
3981 /* Account for reserved pages */
3982 if (j == 0 && realsize > dma_reserve) {
3983 realsize -= dma_reserve;
3984 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
3985 zone_names[0], dma_reserve);
3988 if (!is_highmem_idx(j))
3989 nr_kernel_pages += realsize;
3990 nr_all_pages += realsize;
3992 zone->spanned_pages = size;
3993 zone->present_pages = realsize;
3996 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
3998 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4000 zone->name = zone_names[j];
4001 spin_lock_init(&zone->lock);
4002 spin_lock_init(&zone->lru_lock);
4003 zone_seqlock_init(zone);
4004 zone->zone_pgdat = pgdat;
4006 zone->prev_priority = DEF_PRIORITY;
4008 zone_pcp_init(zone);
4010 INIT_LIST_HEAD(&zone->lru[l].list);
4011 zone->reclaim_stat.nr_saved_scan[l] = 0;
4013 zone->reclaim_stat.recent_rotated[0] = 0;
4014 zone->reclaim_stat.recent_rotated[1] = 0;
4015 zone->reclaim_stat.recent_scanned[0] = 0;
4016 zone->reclaim_stat.recent_scanned[1] = 0;
4017 zap_zone_vm_stats(zone);
4022 set_pageblock_order(pageblock_default_order());
4023 setup_usemap(pgdat, zone, size);
4024 ret = init_currently_empty_zone(zone, zone_start_pfn,
4025 size, MEMMAP_EARLY);
4027 memmap_init(size, nid, j, zone_start_pfn);
4028 zone_start_pfn += size;
4032 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4034 /* Skip empty nodes */
4035 if (!pgdat->node_spanned_pages)
4038 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4039 /* ia64 gets its own node_mem_map, before this, without bootmem */
4040 if (!pgdat->node_mem_map) {
4041 unsigned long size, start, end;
4045 * The zone's endpoints aren't required to be MAX_ORDER
4046 * aligned but the node_mem_map endpoints must be in order
4047 * for the buddy allocator to function correctly.
4049 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4050 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4051 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4052 size = (end - start) * sizeof(struct page);
4053 map = alloc_remap(pgdat->node_id, size);
4055 map = alloc_bootmem_node(pgdat, size);
4056 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4058 #ifndef CONFIG_NEED_MULTIPLE_NODES
4060 * With no DISCONTIG, the global mem_map is just set as node 0's
4062 if (pgdat == NODE_DATA(0)) {
4063 mem_map = NODE_DATA(0)->node_mem_map;
4064 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4065 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4066 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4067 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4070 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4073 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4074 unsigned long node_start_pfn, unsigned long *zholes_size)
4076 pg_data_t *pgdat = NODE_DATA(nid);
4078 pgdat->node_id = nid;
4079 pgdat->node_start_pfn = node_start_pfn;
4080 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4082 alloc_node_mem_map(pgdat);
4083 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4084 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4085 nid, (unsigned long)pgdat,
4086 (unsigned long)pgdat->node_mem_map);
4089 free_area_init_core(pgdat, zones_size, zholes_size);
4092 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4094 #if MAX_NUMNODES > 1
4096 * Figure out the number of possible node ids.
4098 static void __init setup_nr_node_ids(void)
4101 unsigned int highest = 0;
4103 for_each_node_mask(node, node_possible_map)
4105 nr_node_ids = highest + 1;
4108 static inline void setup_nr_node_ids(void)
4114 * add_active_range - Register a range of PFNs backed by physical memory
4115 * @nid: The node ID the range resides on
4116 * @start_pfn: The start PFN of the available physical memory
4117 * @end_pfn: The end PFN of the available physical memory
4119 * These ranges are stored in an early_node_map[] and later used by
4120 * free_area_init_nodes() to calculate zone sizes and holes. If the
4121 * range spans a memory hole, it is up to the architecture to ensure
4122 * the memory is not freed by the bootmem allocator. If possible
4123 * the range being registered will be merged with existing ranges.
4125 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4126 unsigned long end_pfn)
4130 mminit_dprintk(MMINIT_TRACE, "memory_register",
4131 "Entering add_active_range(%d, %#lx, %#lx) "
4132 "%d entries of %d used\n",
4133 nid, start_pfn, end_pfn,
4134 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4136 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4138 /* Merge with existing active regions if possible */
4139 for (i = 0; i < nr_nodemap_entries; i++) {
4140 if (early_node_map[i].nid != nid)
4143 /* Skip if an existing region covers this new one */
4144 if (start_pfn >= early_node_map[i].start_pfn &&
4145 end_pfn <= early_node_map[i].end_pfn)
4148 /* Merge forward if suitable */
4149 if (start_pfn <= early_node_map[i].end_pfn &&
4150 end_pfn > early_node_map[i].end_pfn) {
4151 early_node_map[i].end_pfn = end_pfn;
4155 /* Merge backward if suitable */
4156 if (start_pfn < early_node_map[i].start_pfn &&
4157 end_pfn >= early_node_map[i].start_pfn) {
4158 early_node_map[i].start_pfn = start_pfn;
4163 /* Check that early_node_map is large enough */
4164 if (i >= MAX_ACTIVE_REGIONS) {
4165 printk(KERN_CRIT "More than %d memory regions, truncating\n",
4166 MAX_ACTIVE_REGIONS);
4170 early_node_map[i].nid = nid;
4171 early_node_map[i].start_pfn = start_pfn;
4172 early_node_map[i].end_pfn = end_pfn;
4173 nr_nodemap_entries = i + 1;
4177 * remove_active_range - Shrink an existing registered range of PFNs
4178 * @nid: The node id the range is on that should be shrunk
4179 * @start_pfn: The new PFN of the range
4180 * @end_pfn: The new PFN of the range
4182 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4183 * The map is kept near the end physical page range that has already been
4184 * registered. This function allows an arch to shrink an existing registered
4187 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4188 unsigned long end_pfn)
4193 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4194 nid, start_pfn, end_pfn);
4196 /* Find the old active region end and shrink */
4197 for_each_active_range_index_in_nid(i, nid) {
4198 if (early_node_map[i].start_pfn >= start_pfn &&
4199 early_node_map[i].end_pfn <= end_pfn) {
4201 early_node_map[i].start_pfn = 0;
4202 early_node_map[i].end_pfn = 0;
4206 if (early_node_map[i].start_pfn < start_pfn &&
4207 early_node_map[i].end_pfn > start_pfn) {
4208 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4209 early_node_map[i].end_pfn = start_pfn;
4210 if (temp_end_pfn > end_pfn)
4211 add_active_range(nid, end_pfn, temp_end_pfn);
4214 if (early_node_map[i].start_pfn >= start_pfn &&
4215 early_node_map[i].end_pfn > end_pfn &&
4216 early_node_map[i].start_pfn < end_pfn) {
4217 early_node_map[i].start_pfn = end_pfn;
4225 /* remove the blank ones */
4226 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4227 if (early_node_map[i].nid != nid)
4229 if (early_node_map[i].end_pfn)
4231 /* we found it, get rid of it */
4232 for (j = i; j < nr_nodemap_entries - 1; j++)
4233 memcpy(&early_node_map[j], &early_node_map[j+1],
4234 sizeof(early_node_map[j]));
4235 j = nr_nodemap_entries - 1;
4236 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4237 nr_nodemap_entries--;
4242 * remove_all_active_ranges - Remove all currently registered regions
4244 * During discovery, it may be found that a table like SRAT is invalid
4245 * and an alternative discovery method must be used. This function removes
4246 * all currently registered regions.
4248 void __init remove_all_active_ranges(void)
4250 memset(early_node_map, 0, sizeof(early_node_map));
4251 nr_nodemap_entries = 0;
4254 /* Compare two active node_active_regions */
4255 static int __init cmp_node_active_region(const void *a, const void *b)
4257 struct node_active_region *arange = (struct node_active_region *)a;
4258 struct node_active_region *brange = (struct node_active_region *)b;
4260 /* Done this way to avoid overflows */
4261 if (arange->start_pfn > brange->start_pfn)
4263 if (arange->start_pfn < brange->start_pfn)
4269 /* sort the node_map by start_pfn */
4270 void __init sort_node_map(void)
4272 sort(early_node_map, (size_t)nr_nodemap_entries,
4273 sizeof(struct node_active_region),
4274 cmp_node_active_region, NULL);
4277 /* Find the lowest pfn for a node */
4278 static unsigned long __init find_min_pfn_for_node(int nid)
4281 unsigned long min_pfn = ULONG_MAX;
4283 /* Assuming a sorted map, the first range found has the starting pfn */
4284 for_each_active_range_index_in_nid(i, nid)
4285 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4287 if (min_pfn == ULONG_MAX) {
4289 "Could not find start_pfn for node %d\n", nid);
4297 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4299 * It returns the minimum PFN based on information provided via
4300 * add_active_range().
4302 unsigned long __init find_min_pfn_with_active_regions(void)
4304 return find_min_pfn_for_node(MAX_NUMNODES);
4308 * early_calculate_totalpages()
4309 * Sum pages in active regions for movable zone.
4310 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4312 static unsigned long __init early_calculate_totalpages(void)
4315 unsigned long totalpages = 0;
4317 for (i = 0; i < nr_nodemap_entries; i++) {
4318 unsigned long pages = early_node_map[i].end_pfn -
4319 early_node_map[i].start_pfn;
4320 totalpages += pages;
4322 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4328 * Find the PFN the Movable zone begins in each node. Kernel memory
4329 * is spread evenly between nodes as long as the nodes have enough
4330 * memory. When they don't, some nodes will have more kernelcore than
4333 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4336 unsigned long usable_startpfn;
4337 unsigned long kernelcore_node, kernelcore_remaining;
4338 /* save the state before borrow the nodemask */
4339 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4340 unsigned long totalpages = early_calculate_totalpages();
4341 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4344 * If movablecore was specified, calculate what size of
4345 * kernelcore that corresponds so that memory usable for
4346 * any allocation type is evenly spread. If both kernelcore
4347 * and movablecore are specified, then the value of kernelcore
4348 * will be used for required_kernelcore if it's greater than
4349 * what movablecore would have allowed.
4351 if (required_movablecore) {
4352 unsigned long corepages;
4355 * Round-up so that ZONE_MOVABLE is at least as large as what
4356 * was requested by the user
4358 required_movablecore =
4359 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4360 corepages = totalpages - required_movablecore;
4362 required_kernelcore = max(required_kernelcore, corepages);
4365 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4366 if (!required_kernelcore)
4369 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4370 find_usable_zone_for_movable();
4371 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4374 /* Spread kernelcore memory as evenly as possible throughout nodes */
4375 kernelcore_node = required_kernelcore / usable_nodes;
4376 for_each_node_state(nid, N_HIGH_MEMORY) {
4378 * Recalculate kernelcore_node if the division per node
4379 * now exceeds what is necessary to satisfy the requested
4380 * amount of memory for the kernel
4382 if (required_kernelcore < kernelcore_node)
4383 kernelcore_node = required_kernelcore / usable_nodes;
4386 * As the map is walked, we track how much memory is usable
4387 * by the kernel using kernelcore_remaining. When it is
4388 * 0, the rest of the node is usable by ZONE_MOVABLE
4390 kernelcore_remaining = kernelcore_node;
4392 /* Go through each range of PFNs within this node */
4393 for_each_active_range_index_in_nid(i, nid) {
4394 unsigned long start_pfn, end_pfn;
4395 unsigned long size_pages;
4397 start_pfn = max(early_node_map[i].start_pfn,
4398 zone_movable_pfn[nid]);
4399 end_pfn = early_node_map[i].end_pfn;
4400 if (start_pfn >= end_pfn)
4403 /* Account for what is only usable for kernelcore */
4404 if (start_pfn < usable_startpfn) {
4405 unsigned long kernel_pages;
4406 kernel_pages = min(end_pfn, usable_startpfn)
4409 kernelcore_remaining -= min(kernel_pages,
4410 kernelcore_remaining);
4411 required_kernelcore -= min(kernel_pages,
4412 required_kernelcore);
4414 /* Continue if range is now fully accounted */
4415 if (end_pfn <= usable_startpfn) {
4418 * Push zone_movable_pfn to the end so
4419 * that if we have to rebalance
4420 * kernelcore across nodes, we will
4421 * not double account here
4423 zone_movable_pfn[nid] = end_pfn;
4426 start_pfn = usable_startpfn;
4430 * The usable PFN range for ZONE_MOVABLE is from
4431 * start_pfn->end_pfn. Calculate size_pages as the
4432 * number of pages used as kernelcore
4434 size_pages = end_pfn - start_pfn;
4435 if (size_pages > kernelcore_remaining)
4436 size_pages = kernelcore_remaining;
4437 zone_movable_pfn[nid] = start_pfn + size_pages;
4440 * Some kernelcore has been met, update counts and
4441 * break if the kernelcore for this node has been
4444 required_kernelcore -= min(required_kernelcore,
4446 kernelcore_remaining -= size_pages;
4447 if (!kernelcore_remaining)
4453 * If there is still required_kernelcore, we do another pass with one
4454 * less node in the count. This will push zone_movable_pfn[nid] further
4455 * along on the nodes that still have memory until kernelcore is
4459 if (usable_nodes && required_kernelcore > usable_nodes)
4462 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4463 for (nid = 0; nid < MAX_NUMNODES; nid++)
4464 zone_movable_pfn[nid] =
4465 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4468 /* restore the node_state */
4469 node_states[N_HIGH_MEMORY] = saved_node_state;
4472 /* Any regular memory on that node ? */
4473 static void check_for_regular_memory(pg_data_t *pgdat)
4475 #ifdef CONFIG_HIGHMEM
4476 enum zone_type zone_type;
4478 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4479 struct zone *zone = &pgdat->node_zones[zone_type];
4480 if (zone->present_pages)
4481 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4487 * free_area_init_nodes - Initialise all pg_data_t and zone data
4488 * @max_zone_pfn: an array of max PFNs for each zone
4490 * This will call free_area_init_node() for each active node in the system.
4491 * Using the page ranges provided by add_active_range(), the size of each
4492 * zone in each node and their holes is calculated. If the maximum PFN
4493 * between two adjacent zones match, it is assumed that the zone is empty.
4494 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4495 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4496 * starts where the previous one ended. For example, ZONE_DMA32 starts
4497 * at arch_max_dma_pfn.
4499 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4504 /* Sort early_node_map as initialisation assumes it is sorted */
4507 /* Record where the zone boundaries are */
4508 memset(arch_zone_lowest_possible_pfn, 0,
4509 sizeof(arch_zone_lowest_possible_pfn));
4510 memset(arch_zone_highest_possible_pfn, 0,
4511 sizeof(arch_zone_highest_possible_pfn));
4512 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4513 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4514 for (i = 1; i < MAX_NR_ZONES; i++) {
4515 if (i == ZONE_MOVABLE)
4517 arch_zone_lowest_possible_pfn[i] =
4518 arch_zone_highest_possible_pfn[i-1];
4519 arch_zone_highest_possible_pfn[i] =
4520 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4522 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4523 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4525 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4526 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4527 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4529 /* Print out the zone ranges */
4530 printk("Zone PFN ranges:\n");
4531 for (i = 0; i < MAX_NR_ZONES; i++) {
4532 if (i == ZONE_MOVABLE)
4534 printk(" %-8s ", zone_names[i]);
4535 if (arch_zone_lowest_possible_pfn[i] ==
4536 arch_zone_highest_possible_pfn[i])
4539 printk("%0#10lx -> %0#10lx\n",
4540 arch_zone_lowest_possible_pfn[i],
4541 arch_zone_highest_possible_pfn[i]);
4544 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4545 printk("Movable zone start PFN for each node\n");
4546 for (i = 0; i < MAX_NUMNODES; i++) {
4547 if (zone_movable_pfn[i])
4548 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4551 /* Print out the early_node_map[] */
4552 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4553 for (i = 0; i < nr_nodemap_entries; i++)
4554 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4555 early_node_map[i].start_pfn,
4556 early_node_map[i].end_pfn);
4558 /* Initialise every node */
4559 mminit_verify_pageflags_layout();
4560 setup_nr_node_ids();
4561 for_each_online_node(nid) {
4562 pg_data_t *pgdat = NODE_DATA(nid);
4563 free_area_init_node(nid, NULL,
4564 find_min_pfn_for_node(nid), NULL);
4566 /* Any memory on that node */
4567 if (pgdat->node_present_pages)
4568 node_set_state(nid, N_HIGH_MEMORY);
4569 check_for_regular_memory(pgdat);
4573 static int __init cmdline_parse_core(char *p, unsigned long *core)
4575 unsigned long long coremem;
4579 coremem = memparse(p, &p);
4580 *core = coremem >> PAGE_SHIFT;
4582 /* Paranoid check that UL is enough for the coremem value */
4583 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4589 * kernelcore=size sets the amount of memory for use for allocations that
4590 * cannot be reclaimed or migrated.
4592 static int __init cmdline_parse_kernelcore(char *p)
4594 return cmdline_parse_core(p, &required_kernelcore);
4598 * movablecore=size sets the amount of memory for use for allocations that
4599 * can be reclaimed or migrated.
4601 static int __init cmdline_parse_movablecore(char *p)
4603 return cmdline_parse_core(p, &required_movablecore);
4606 early_param("kernelcore", cmdline_parse_kernelcore);
4607 early_param("movablecore", cmdline_parse_movablecore);
4609 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4612 * set_dma_reserve - set the specified number of pages reserved in the first zone
4613 * @new_dma_reserve: The number of pages to mark reserved
4615 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4616 * In the DMA zone, a significant percentage may be consumed by kernel image
4617 * and other unfreeable allocations which can skew the watermarks badly. This
4618 * function may optionally be used to account for unfreeable pages in the
4619 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4620 * smaller per-cpu batchsize.
4622 void __init set_dma_reserve(unsigned long new_dma_reserve)
4624 dma_reserve = new_dma_reserve;
4627 #ifndef CONFIG_NEED_MULTIPLE_NODES
4628 struct pglist_data __refdata contig_page_data = {
4629 #ifndef CONFIG_NO_BOOTMEM
4630 .bdata = &bootmem_node_data[0]
4633 EXPORT_SYMBOL(contig_page_data);
4636 void __init free_area_init(unsigned long *zones_size)
4638 free_area_init_node(0, zones_size,
4639 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4642 static int page_alloc_cpu_notify(struct notifier_block *self,
4643 unsigned long action, void *hcpu)
4645 int cpu = (unsigned long)hcpu;
4647 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4651 * Spill the event counters of the dead processor
4652 * into the current processors event counters.
4653 * This artificially elevates the count of the current
4656 vm_events_fold_cpu(cpu);
4659 * Zero the differential counters of the dead processor
4660 * so that the vm statistics are consistent.
4662 * This is only okay since the processor is dead and cannot
4663 * race with what we are doing.
4665 refresh_cpu_vm_stats(cpu);
4670 void __init page_alloc_init(void)
4672 hotcpu_notifier(page_alloc_cpu_notify, 0);
4676 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4677 * or min_free_kbytes changes.
4679 static void calculate_totalreserve_pages(void)
4681 struct pglist_data *pgdat;
4682 unsigned long reserve_pages = 0;
4683 enum zone_type i, j;
4685 for_each_online_pgdat(pgdat) {
4686 for (i = 0; i < MAX_NR_ZONES; i++) {
4687 struct zone *zone = pgdat->node_zones + i;
4688 unsigned long max = 0;
4690 /* Find valid and maximum lowmem_reserve in the zone */
4691 for (j = i; j < MAX_NR_ZONES; j++) {
4692 if (zone->lowmem_reserve[j] > max)
4693 max = zone->lowmem_reserve[j];
4696 /* we treat the high watermark as reserved pages. */
4697 max += high_wmark_pages(zone);
4699 if (max > zone->present_pages)
4700 max = zone->present_pages;
4701 reserve_pages += max;
4704 totalreserve_pages = reserve_pages;
4708 * setup_per_zone_lowmem_reserve - called whenever
4709 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4710 * has a correct pages reserved value, so an adequate number of
4711 * pages are left in the zone after a successful __alloc_pages().
4713 static void setup_per_zone_lowmem_reserve(void)
4715 struct pglist_data *pgdat;
4716 enum zone_type j, idx;
4718 for_each_online_pgdat(pgdat) {
4719 for (j = 0; j < MAX_NR_ZONES; j++) {
4720 struct zone *zone = pgdat->node_zones + j;
4721 unsigned long present_pages = zone->present_pages;
4723 zone->lowmem_reserve[j] = 0;
4727 struct zone *lower_zone;
4731 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4732 sysctl_lowmem_reserve_ratio[idx] = 1;
4734 lower_zone = pgdat->node_zones + idx;
4735 lower_zone->lowmem_reserve[j] = present_pages /
4736 sysctl_lowmem_reserve_ratio[idx];
4737 present_pages += lower_zone->present_pages;
4742 /* update totalreserve_pages */
4743 calculate_totalreserve_pages();
4747 * setup_per_zone_wmarks - called when min_free_kbytes changes
4748 * or when memory is hot-{added|removed}
4750 * Ensures that the watermark[min,low,high] values for each zone are set
4751 * correctly with respect to min_free_kbytes.
4753 void setup_per_zone_wmarks(void)
4755 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4756 unsigned long lowmem_pages = 0;
4758 unsigned long flags;
4760 /* Calculate total number of !ZONE_HIGHMEM pages */
4761 for_each_zone(zone) {
4762 if (!is_highmem(zone))
4763 lowmem_pages += zone->present_pages;
4766 for_each_zone(zone) {
4769 spin_lock_irqsave(&zone->lock, flags);
4770 tmp = (u64)pages_min * zone->present_pages;
4771 do_div(tmp, lowmem_pages);
4772 if (is_highmem(zone)) {
4774 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4775 * need highmem pages, so cap pages_min to a small
4778 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4779 * deltas controls asynch page reclaim, and so should
4780 * not be capped for highmem.
4784 min_pages = zone->present_pages / 1024;
4785 if (min_pages < SWAP_CLUSTER_MAX)
4786 min_pages = SWAP_CLUSTER_MAX;
4787 if (min_pages > 128)
4789 zone->watermark[WMARK_MIN] = min_pages;
4792 * If it's a lowmem zone, reserve a number of pages
4793 * proportionate to the zone's size.
4795 zone->watermark[WMARK_MIN] = tmp;
4798 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
4799 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
4800 setup_zone_migrate_reserve(zone);
4801 spin_unlock_irqrestore(&zone->lock, flags);
4804 /* update totalreserve_pages */
4805 calculate_totalreserve_pages();
4809 * The inactive anon list should be small enough that the VM never has to
4810 * do too much work, but large enough that each inactive page has a chance
4811 * to be referenced again before it is swapped out.
4813 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4814 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4815 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4816 * the anonymous pages are kept on the inactive list.
4819 * memory ratio inactive anon
4820 * -------------------------------------
4829 void calculate_zone_inactive_ratio(struct zone *zone)
4831 unsigned int gb, ratio;
4833 /* Zone size in gigabytes */
4834 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4836 ratio = int_sqrt(10 * gb);
4840 zone->inactive_ratio = ratio;
4843 static void __init setup_per_zone_inactive_ratio(void)
4848 calculate_zone_inactive_ratio(zone);
4852 * Initialise min_free_kbytes.
4854 * For small machines we want it small (128k min). For large machines
4855 * we want it large (64MB max). But it is not linear, because network
4856 * bandwidth does not increase linearly with machine size. We use
4858 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4859 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4875 static int __init init_per_zone_wmark_min(void)
4877 unsigned long lowmem_kbytes;
4879 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4881 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4882 if (min_free_kbytes < 128)
4883 min_free_kbytes = 128;
4884 if (min_free_kbytes > 65536)
4885 min_free_kbytes = 65536;
4886 setup_per_zone_wmarks();
4887 setup_per_zone_lowmem_reserve();
4888 setup_per_zone_inactive_ratio();
4891 module_init(init_per_zone_wmark_min)
4894 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4895 * that we can call two helper functions whenever min_free_kbytes
4898 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
4899 void __user *buffer, size_t *length, loff_t *ppos)
4901 proc_dointvec(table, write, buffer, length, ppos);
4903 setup_per_zone_wmarks();
4908 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
4909 void __user *buffer, size_t *length, loff_t *ppos)
4914 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
4919 zone->min_unmapped_pages = (zone->present_pages *
4920 sysctl_min_unmapped_ratio) / 100;
4924 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
4925 void __user *buffer, size_t *length, loff_t *ppos)
4930 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
4935 zone->min_slab_pages = (zone->present_pages *
4936 sysctl_min_slab_ratio) / 100;
4942 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4943 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4944 * whenever sysctl_lowmem_reserve_ratio changes.
4946 * The reserve ratio obviously has absolutely no relation with the
4947 * minimum watermarks. The lowmem reserve ratio can only make sense
4948 * if in function of the boot time zone sizes.
4950 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
4951 void __user *buffer, size_t *length, loff_t *ppos)
4953 proc_dointvec_minmax(table, write, buffer, length, ppos);
4954 setup_per_zone_lowmem_reserve();
4959 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4960 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4961 * can have before it gets flushed back to buddy allocator.
4964 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
4965 void __user *buffer, size_t *length, loff_t *ppos)
4971 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
4972 if (!write || (ret == -EINVAL))
4974 for_each_populated_zone(zone) {
4975 for_each_possible_cpu(cpu) {
4977 high = zone->present_pages / percpu_pagelist_fraction;
4978 setup_pagelist_highmark(
4979 per_cpu_ptr(zone->pageset, cpu), high);
4985 int hashdist = HASHDIST_DEFAULT;
4988 static int __init set_hashdist(char *str)
4992 hashdist = simple_strtoul(str, &str, 0);
4995 __setup("hashdist=", set_hashdist);
4999 * allocate a large system hash table from bootmem
5000 * - it is assumed that the hash table must contain an exact power-of-2
5001 * quantity of entries
5002 * - limit is the number of hash buckets, not the total allocation size
5004 void *__init alloc_large_system_hash(const char *tablename,
5005 unsigned long bucketsize,
5006 unsigned long numentries,
5009 unsigned int *_hash_shift,
5010 unsigned int *_hash_mask,
5011 unsigned long limit)
5013 unsigned long long max = limit;
5014 unsigned long log2qty, size;
5017 /* allow the kernel cmdline to have a say */
5019 /* round applicable memory size up to nearest megabyte */
5020 numentries = nr_kernel_pages;
5021 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5022 numentries >>= 20 - PAGE_SHIFT;
5023 numentries <<= 20 - PAGE_SHIFT;
5025 /* limit to 1 bucket per 2^scale bytes of low memory */
5026 if (scale > PAGE_SHIFT)
5027 numentries >>= (scale - PAGE_SHIFT);
5029 numentries <<= (PAGE_SHIFT - scale);
5031 /* Make sure we've got at least a 0-order allocation.. */
5032 if (unlikely(flags & HASH_SMALL)) {
5033 /* Makes no sense without HASH_EARLY */
5034 WARN_ON(!(flags & HASH_EARLY));
5035 if (!(numentries >> *_hash_shift)) {
5036 numentries = 1UL << *_hash_shift;
5037 BUG_ON(!numentries);
5039 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5040 numentries = PAGE_SIZE / bucketsize;
5042 numentries = roundup_pow_of_two(numentries);
5044 /* limit allocation size to 1/16 total memory by default */
5046 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5047 do_div(max, bucketsize);
5050 if (numentries > max)
5053 log2qty = ilog2(numentries);
5056 size = bucketsize << log2qty;
5057 if (flags & HASH_EARLY)
5058 table = alloc_bootmem_nopanic(size);
5060 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5063 * If bucketsize is not a power-of-two, we may free
5064 * some pages at the end of hash table which
5065 * alloc_pages_exact() automatically does
5067 if (get_order(size) < MAX_ORDER) {
5068 table = alloc_pages_exact(size, GFP_ATOMIC);
5069 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5072 } while (!table && size > PAGE_SIZE && --log2qty);
5075 panic("Failed to allocate %s hash table\n", tablename);
5077 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
5080 ilog2(size) - PAGE_SHIFT,
5084 *_hash_shift = log2qty;
5086 *_hash_mask = (1 << log2qty) - 1;
5091 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5092 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5095 #ifdef CONFIG_SPARSEMEM
5096 return __pfn_to_section(pfn)->pageblock_flags;
5098 return zone->pageblock_flags;
5099 #endif /* CONFIG_SPARSEMEM */
5102 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5104 #ifdef CONFIG_SPARSEMEM
5105 pfn &= (PAGES_PER_SECTION-1);
5106 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5108 pfn = pfn - zone->zone_start_pfn;
5109 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5110 #endif /* CONFIG_SPARSEMEM */
5114 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5115 * @page: The page within the block of interest
5116 * @start_bitidx: The first bit of interest to retrieve
5117 * @end_bitidx: The last bit of interest
5118 * returns pageblock_bits flags
5120 unsigned long get_pageblock_flags_group(struct page *page,
5121 int start_bitidx, int end_bitidx)
5124 unsigned long *bitmap;
5125 unsigned long pfn, bitidx;
5126 unsigned long flags = 0;
5127 unsigned long value = 1;
5129 zone = page_zone(page);
5130 pfn = page_to_pfn(page);
5131 bitmap = get_pageblock_bitmap(zone, pfn);
5132 bitidx = pfn_to_bitidx(zone, pfn);
5134 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5135 if (test_bit(bitidx + start_bitidx, bitmap))
5142 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5143 * @page: The page within the block of interest
5144 * @start_bitidx: The first bit of interest
5145 * @end_bitidx: The last bit of interest
5146 * @flags: The flags to set
5148 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5149 int start_bitidx, int end_bitidx)
5152 unsigned long *bitmap;
5153 unsigned long pfn, bitidx;
5154 unsigned long value = 1;
5156 zone = page_zone(page);
5157 pfn = page_to_pfn(page);
5158 bitmap = get_pageblock_bitmap(zone, pfn);
5159 bitidx = pfn_to_bitidx(zone, pfn);
5160 VM_BUG_ON(pfn < zone->zone_start_pfn);
5161 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5163 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5165 __set_bit(bitidx + start_bitidx, bitmap);
5167 __clear_bit(bitidx + start_bitidx, bitmap);
5171 * This is designed as sub function...plz see page_isolation.c also.
5172 * set/clear page block's type to be ISOLATE.
5173 * page allocater never alloc memory from ISOLATE block.
5176 int set_migratetype_isolate(struct page *page)
5179 struct page *curr_page;
5180 unsigned long flags, pfn, iter;
5181 unsigned long immobile = 0;
5182 struct memory_isolate_notify arg;
5187 zone = page_zone(page);
5188 zone_idx = zone_idx(zone);
5190 spin_lock_irqsave(&zone->lock, flags);
5191 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE ||
5192 zone_idx == ZONE_MOVABLE) {
5197 pfn = page_to_pfn(page);
5198 arg.start_pfn = pfn;
5199 arg.nr_pages = pageblock_nr_pages;
5200 arg.pages_found = 0;
5203 * It may be possible to isolate a pageblock even if the
5204 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5205 * notifier chain is used by balloon drivers to return the
5206 * number of pages in a range that are held by the balloon
5207 * driver to shrink memory. If all the pages are accounted for
5208 * by balloons, are free, or on the LRU, isolation can continue.
5209 * Later, for example, when memory hotplug notifier runs, these
5210 * pages reported as "can be isolated" should be isolated(freed)
5211 * by the balloon driver through the memory notifier chain.
5213 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5214 notifier_ret = notifier_to_errno(notifier_ret);
5215 if (notifier_ret || !arg.pages_found)
5218 for (iter = pfn; iter < (pfn + pageblock_nr_pages); iter++) {
5219 if (!pfn_valid_within(pfn))
5222 curr_page = pfn_to_page(iter);
5223 if (!page_count(curr_page) || PageLRU(curr_page))
5229 if (arg.pages_found == immobile)
5234 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5235 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5238 spin_unlock_irqrestore(&zone->lock, flags);
5244 void unset_migratetype_isolate(struct page *page)
5247 unsigned long flags;
5248 zone = page_zone(page);
5249 spin_lock_irqsave(&zone->lock, flags);
5250 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5252 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5253 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5255 spin_unlock_irqrestore(&zone->lock, flags);
5258 #ifdef CONFIG_MEMORY_HOTREMOVE
5260 * All pages in the range must be isolated before calling this.
5263 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5269 unsigned long flags;
5270 /* find the first valid pfn */
5271 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5276 zone = page_zone(pfn_to_page(pfn));
5277 spin_lock_irqsave(&zone->lock, flags);
5279 while (pfn < end_pfn) {
5280 if (!pfn_valid(pfn)) {
5284 page = pfn_to_page(pfn);
5285 BUG_ON(page_count(page));
5286 BUG_ON(!PageBuddy(page));
5287 order = page_order(page);
5288 #ifdef CONFIG_DEBUG_VM
5289 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5290 pfn, 1 << order, end_pfn);
5292 list_del(&page->lru);
5293 rmv_page_order(page);
5294 zone->free_area[order].nr_free--;
5295 __mod_zone_page_state(zone, NR_FREE_PAGES,
5297 for (i = 0; i < (1 << order); i++)
5298 SetPageReserved((page+i));
5299 pfn += (1 << order);
5301 spin_unlock_irqrestore(&zone->lock, flags);
5305 #ifdef CONFIG_MEMORY_FAILURE
5306 bool is_free_buddy_page(struct page *page)
5308 struct zone *zone = page_zone(page);
5309 unsigned long pfn = page_to_pfn(page);
5310 unsigned long flags;
5313 spin_lock_irqsave(&zone->lock, flags);
5314 for (order = 0; order < MAX_ORDER; order++) {
5315 struct page *page_head = page - (pfn & ((1 << order) - 1));
5317 if (PageBuddy(page_head) && page_order(page_head) >= order)
5320 spin_unlock_irqrestore(&zone->lock, flags);
5322 return order < MAX_ORDER;
5326 static struct trace_print_flags pageflag_names[] = {
5327 {1UL << PG_locked, "locked" },
5328 {1UL << PG_error, "error" },
5329 {1UL << PG_referenced, "referenced" },
5330 {1UL << PG_uptodate, "uptodate" },
5331 {1UL << PG_dirty, "dirty" },
5332 {1UL << PG_lru, "lru" },
5333 {1UL << PG_active, "active" },
5334 {1UL << PG_slab, "slab" },
5335 {1UL << PG_owner_priv_1, "owner_priv_1" },
5336 {1UL << PG_arch_1, "arch_1" },
5337 {1UL << PG_reserved, "reserved" },
5338 {1UL << PG_private, "private" },
5339 {1UL << PG_private_2, "private_2" },
5340 {1UL << PG_writeback, "writeback" },
5341 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5342 {1UL << PG_head, "head" },
5343 {1UL << PG_tail, "tail" },
5345 {1UL << PG_compound, "compound" },
5347 {1UL << PG_swapcache, "swapcache" },
5348 {1UL << PG_mappedtodisk, "mappedtodisk" },
5349 {1UL << PG_reclaim, "reclaim" },
5350 {1UL << PG_buddy, "buddy" },
5351 {1UL << PG_swapbacked, "swapbacked" },
5352 {1UL << PG_unevictable, "unevictable" },
5354 {1UL << PG_mlocked, "mlocked" },
5356 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5357 {1UL << PG_uncached, "uncached" },
5359 #ifdef CONFIG_MEMORY_FAILURE
5360 {1UL << PG_hwpoison, "hwpoison" },
5365 static void dump_page_flags(unsigned long flags)
5367 const char *delim = "";
5371 printk(KERN_ALERT "page flags: %#lx(", flags);
5373 /* remove zone id */
5374 flags &= (1UL << NR_PAGEFLAGS) - 1;
5376 for (i = 0; pageflag_names[i].name && flags; i++) {
5378 mask = pageflag_names[i].mask;
5379 if ((flags & mask) != mask)
5383 printk("%s%s", delim, pageflag_names[i].name);
5387 /* check for left over flags */
5389 printk("%s%#lx", delim, flags);
5394 void dump_page(struct page *page)
5397 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5398 page, page_count(page), page_mapcount(page),
5399 page->mapping, page->index);
5400 dump_page_flags(page->flags);