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 bool free_pages_prepare(struct page *page, unsigned int order)
628 trace_mm_page_free_direct(page, order);
629 kmemcheck_free_shadow(page, order);
631 for (i = 0; i < (1 << order); i++) {
632 struct page *pg = page + i;
636 bad += free_pages_check(pg);
641 if (!PageHighMem(page)) {
642 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
643 debug_check_no_obj_freed(page_address(page),
646 arch_free_page(page, order);
647 kernel_map_pages(page, 1 << order, 0);
652 static void __free_pages_ok(struct page *page, unsigned int order)
655 int wasMlocked = __TestClearPageMlocked(page);
657 if (!free_pages_prepare(page, order))
660 local_irq_save(flags);
661 if (unlikely(wasMlocked))
662 free_page_mlock(page);
663 __count_vm_events(PGFREE, 1 << order);
664 free_one_page(page_zone(page), page, order,
665 get_pageblock_migratetype(page));
666 local_irq_restore(flags);
670 * permit the bootmem allocator to evade page validation on high-order frees
672 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
675 __ClearPageReserved(page);
676 set_page_count(page, 0);
677 set_page_refcounted(page);
683 for (loop = 0; loop < BITS_PER_LONG; loop++) {
684 struct page *p = &page[loop];
686 if (loop + 1 < BITS_PER_LONG)
688 __ClearPageReserved(p);
689 set_page_count(p, 0);
692 set_page_refcounted(page);
693 __free_pages(page, order);
699 * The order of subdivision here is critical for the IO subsystem.
700 * Please do not alter this order without good reasons and regression
701 * testing. Specifically, as large blocks of memory are subdivided,
702 * the order in which smaller blocks are delivered depends on the order
703 * they're subdivided in this function. This is the primary factor
704 * influencing the order in which pages are delivered to the IO
705 * subsystem according to empirical testing, and this is also justified
706 * by considering the behavior of a buddy system containing a single
707 * large block of memory acted on by a series of small allocations.
708 * This behavior is a critical factor in sglist merging's success.
712 static inline void expand(struct zone *zone, struct page *page,
713 int low, int high, struct free_area *area,
716 unsigned long size = 1 << high;
722 VM_BUG_ON(bad_range(zone, &page[size]));
723 list_add(&page[size].lru, &area->free_list[migratetype]);
725 set_page_order(&page[size], high);
730 * This page is about to be returned from the page allocator
732 static inline int check_new_page(struct page *page)
734 if (unlikely(page_mapcount(page) |
735 (page->mapping != NULL) |
736 (atomic_read(&page->_count) != 0) |
737 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) {
744 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
748 for (i = 0; i < (1 << order); i++) {
749 struct page *p = page + i;
750 if (unlikely(check_new_page(p)))
754 set_page_private(page, 0);
755 set_page_refcounted(page);
757 arch_alloc_page(page, order);
758 kernel_map_pages(page, 1 << order, 1);
760 if (gfp_flags & __GFP_ZERO)
761 prep_zero_page(page, order, gfp_flags);
763 if (order && (gfp_flags & __GFP_COMP))
764 prep_compound_page(page, order);
770 * Go through the free lists for the given migratetype and remove
771 * the smallest available page from the freelists
774 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
777 unsigned int current_order;
778 struct free_area * area;
781 /* Find a page of the appropriate size in the preferred list */
782 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
783 area = &(zone->free_area[current_order]);
784 if (list_empty(&area->free_list[migratetype]))
787 page = list_entry(area->free_list[migratetype].next,
789 list_del(&page->lru);
790 rmv_page_order(page);
792 expand(zone, page, order, current_order, area, migratetype);
801 * This array describes the order lists are fallen back to when
802 * the free lists for the desirable migrate type are depleted
804 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
805 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
806 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
807 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
808 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
812 * Move the free pages in a range to the free lists of the requested type.
813 * Note that start_page and end_pages are not aligned on a pageblock
814 * boundary. If alignment is required, use move_freepages_block()
816 static int move_freepages(struct zone *zone,
817 struct page *start_page, struct page *end_page,
824 #ifndef CONFIG_HOLES_IN_ZONE
826 * page_zone is not safe to call in this context when
827 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
828 * anyway as we check zone boundaries in move_freepages_block().
829 * Remove at a later date when no bug reports exist related to
830 * grouping pages by mobility
832 BUG_ON(page_zone(start_page) != page_zone(end_page));
835 for (page = start_page; page <= end_page;) {
836 /* Make sure we are not inadvertently changing nodes */
837 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
839 if (!pfn_valid_within(page_to_pfn(page))) {
844 if (!PageBuddy(page)) {
849 order = page_order(page);
850 list_del(&page->lru);
852 &zone->free_area[order].free_list[migratetype]);
854 pages_moved += 1 << order;
860 static int move_freepages_block(struct zone *zone, struct page *page,
863 unsigned long start_pfn, end_pfn;
864 struct page *start_page, *end_page;
866 start_pfn = page_to_pfn(page);
867 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
868 start_page = pfn_to_page(start_pfn);
869 end_page = start_page + pageblock_nr_pages - 1;
870 end_pfn = start_pfn + pageblock_nr_pages - 1;
872 /* Do not cross zone boundaries */
873 if (start_pfn < zone->zone_start_pfn)
875 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
878 return move_freepages(zone, start_page, end_page, migratetype);
881 static void change_pageblock_range(struct page *pageblock_page,
882 int start_order, int migratetype)
884 int nr_pageblocks = 1 << (start_order - pageblock_order);
886 while (nr_pageblocks--) {
887 set_pageblock_migratetype(pageblock_page, migratetype);
888 pageblock_page += pageblock_nr_pages;
892 /* Remove an element from the buddy allocator from the fallback list */
893 static inline struct page *
894 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
896 struct free_area * area;
901 /* Find the largest possible block of pages in the other list */
902 for (current_order = MAX_ORDER-1; current_order >= order;
904 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
905 migratetype = fallbacks[start_migratetype][i];
907 /* MIGRATE_RESERVE handled later if necessary */
908 if (migratetype == MIGRATE_RESERVE)
911 area = &(zone->free_area[current_order]);
912 if (list_empty(&area->free_list[migratetype]))
915 page = list_entry(area->free_list[migratetype].next,
920 * If breaking a large block of pages, move all free
921 * pages to the preferred allocation list. If falling
922 * back for a reclaimable kernel allocation, be more
923 * agressive about taking ownership of free pages
925 if (unlikely(current_order >= (pageblock_order >> 1)) ||
926 start_migratetype == MIGRATE_RECLAIMABLE ||
927 page_group_by_mobility_disabled) {
929 pages = move_freepages_block(zone, page,
932 /* Claim the whole block if over half of it is free */
933 if (pages >= (1 << (pageblock_order-1)) ||
934 page_group_by_mobility_disabled)
935 set_pageblock_migratetype(page,
938 migratetype = start_migratetype;
941 /* Remove the page from the freelists */
942 list_del(&page->lru);
943 rmv_page_order(page);
945 /* Take ownership for orders >= pageblock_order */
946 if (current_order >= pageblock_order)
947 change_pageblock_range(page, current_order,
950 expand(zone, page, order, current_order, area, migratetype);
952 trace_mm_page_alloc_extfrag(page, order, current_order,
953 start_migratetype, migratetype);
963 * Do the hard work of removing an element from the buddy allocator.
964 * Call me with the zone->lock already held.
966 static struct page *__rmqueue(struct zone *zone, unsigned int order,
972 page = __rmqueue_smallest(zone, order, migratetype);
974 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
975 page = __rmqueue_fallback(zone, order, migratetype);
978 * Use MIGRATE_RESERVE rather than fail an allocation. goto
979 * is used because __rmqueue_smallest is an inline function
980 * and we want just one call site
983 migratetype = MIGRATE_RESERVE;
988 trace_mm_page_alloc_zone_locked(page, order, migratetype);
993 * Obtain a specified number of elements from the buddy allocator, all under
994 * a single hold of the lock, for efficiency. Add them to the supplied list.
995 * Returns the number of new pages which were placed at *list.
997 static int rmqueue_bulk(struct zone *zone, unsigned int order,
998 unsigned long count, struct list_head *list,
999 int migratetype, int cold)
1003 spin_lock(&zone->lock);
1004 for (i = 0; i < count; ++i) {
1005 struct page *page = __rmqueue(zone, order, migratetype);
1006 if (unlikely(page == NULL))
1010 * Split buddy pages returned by expand() are received here
1011 * in physical page order. The page is added to the callers and
1012 * list and the list head then moves forward. From the callers
1013 * perspective, the linked list is ordered by page number in
1014 * some conditions. This is useful for IO devices that can
1015 * merge IO requests if the physical pages are ordered
1018 if (likely(cold == 0))
1019 list_add(&page->lru, list);
1021 list_add_tail(&page->lru, list);
1022 set_page_private(page, migratetype);
1025 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1026 spin_unlock(&zone->lock);
1032 * Called from the vmstat counter updater to drain pagesets of this
1033 * currently executing processor on remote nodes after they have
1036 * Note that this function must be called with the thread pinned to
1037 * a single processor.
1039 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1041 unsigned long flags;
1044 local_irq_save(flags);
1045 if (pcp->count >= pcp->batch)
1046 to_drain = pcp->batch;
1048 to_drain = pcp->count;
1049 free_pcppages_bulk(zone, to_drain, pcp);
1050 pcp->count -= to_drain;
1051 local_irq_restore(flags);
1056 * Drain pages of the indicated processor.
1058 * The processor must either be the current processor and the
1059 * thread pinned to the current processor or a processor that
1062 static void drain_pages(unsigned int cpu)
1064 unsigned long flags;
1067 for_each_populated_zone(zone) {
1068 struct per_cpu_pageset *pset;
1069 struct per_cpu_pages *pcp;
1071 local_irq_save(flags);
1072 pset = per_cpu_ptr(zone->pageset, cpu);
1075 free_pcppages_bulk(zone, pcp->count, pcp);
1077 local_irq_restore(flags);
1082 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1084 void drain_local_pages(void *arg)
1086 drain_pages(smp_processor_id());
1090 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1092 void drain_all_pages(void)
1094 on_each_cpu(drain_local_pages, NULL, 1);
1097 #ifdef CONFIG_HIBERNATION
1099 void mark_free_pages(struct zone *zone)
1101 unsigned long pfn, max_zone_pfn;
1102 unsigned long flags;
1104 struct list_head *curr;
1106 if (!zone->spanned_pages)
1109 spin_lock_irqsave(&zone->lock, flags);
1111 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1112 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1113 if (pfn_valid(pfn)) {
1114 struct page *page = pfn_to_page(pfn);
1116 if (!swsusp_page_is_forbidden(page))
1117 swsusp_unset_page_free(page);
1120 for_each_migratetype_order(order, t) {
1121 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1124 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1125 for (i = 0; i < (1UL << order); i++)
1126 swsusp_set_page_free(pfn_to_page(pfn + i));
1129 spin_unlock_irqrestore(&zone->lock, flags);
1131 #endif /* CONFIG_PM */
1134 * Free a 0-order page
1135 * cold == 1 ? free a cold page : free a hot page
1137 void free_hot_cold_page(struct page *page, int cold)
1139 struct zone *zone = page_zone(page);
1140 struct per_cpu_pages *pcp;
1141 unsigned long flags;
1143 int wasMlocked = __TestClearPageMlocked(page);
1145 if (!free_pages_prepare(page, 0))
1148 migratetype = get_pageblock_migratetype(page);
1149 set_page_private(page, migratetype);
1150 local_irq_save(flags);
1151 if (unlikely(wasMlocked))
1152 free_page_mlock(page);
1153 __count_vm_event(PGFREE);
1156 * We only track unmovable, reclaimable and movable on pcp lists.
1157 * Free ISOLATE pages back to the allocator because they are being
1158 * offlined but treat RESERVE as movable pages so we can get those
1159 * areas back if necessary. Otherwise, we may have to free
1160 * excessively into the page allocator
1162 if (migratetype >= MIGRATE_PCPTYPES) {
1163 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1164 free_one_page(zone, page, 0, migratetype);
1167 migratetype = MIGRATE_MOVABLE;
1170 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1172 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1174 list_add(&page->lru, &pcp->lists[migratetype]);
1176 if (pcp->count >= pcp->high) {
1177 free_pcppages_bulk(zone, pcp->batch, pcp);
1178 pcp->count -= pcp->batch;
1182 local_irq_restore(flags);
1186 * split_page takes a non-compound higher-order page, and splits it into
1187 * n (1<<order) sub-pages: page[0..n]
1188 * Each sub-page must be freed individually.
1190 * Note: this is probably too low level an operation for use in drivers.
1191 * Please consult with lkml before using this in your driver.
1193 void split_page(struct page *page, unsigned int order)
1197 VM_BUG_ON(PageCompound(page));
1198 VM_BUG_ON(!page_count(page));
1200 #ifdef CONFIG_KMEMCHECK
1202 * Split shadow pages too, because free(page[0]) would
1203 * otherwise free the whole shadow.
1205 if (kmemcheck_page_is_tracked(page))
1206 split_page(virt_to_page(page[0].shadow), order);
1209 for (i = 1; i < (1 << order); i++)
1210 set_page_refcounted(page + i);
1214 * Similar to split_page except the page is already free. As this is only
1215 * being used for migration, the migratetype of the block also changes.
1216 * As this is called with interrupts disabled, the caller is responsible
1217 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1220 * Note: this is probably too low level an operation for use in drivers.
1221 * Please consult with lkml before using this in your driver.
1223 int split_free_page(struct page *page)
1226 unsigned long watermark;
1229 BUG_ON(!PageBuddy(page));
1231 zone = page_zone(page);
1232 order = page_order(page);
1234 /* Obey watermarks as if the page was being allocated */
1235 watermark = low_wmark_pages(zone) + (1 << order);
1236 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1239 /* Remove page from free list */
1240 list_del(&page->lru);
1241 zone->free_area[order].nr_free--;
1242 rmv_page_order(page);
1243 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1245 /* Split into individual pages */
1246 set_page_refcounted(page);
1247 split_page(page, order);
1249 if (order >= pageblock_order - 1) {
1250 struct page *endpage = page + (1 << order) - 1;
1251 for (; page < endpage; page += pageblock_nr_pages)
1252 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1259 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1260 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1264 struct page *buffered_rmqueue(struct zone *preferred_zone,
1265 struct zone *zone, int order, gfp_t gfp_flags,
1268 unsigned long flags;
1270 int cold = !!(gfp_flags & __GFP_COLD);
1273 if (likely(order == 0)) {
1274 struct per_cpu_pages *pcp;
1275 struct list_head *list;
1277 local_irq_save(flags);
1278 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1279 list = &pcp->lists[migratetype];
1280 if (list_empty(list)) {
1281 pcp->count += rmqueue_bulk(zone, 0,
1284 if (unlikely(list_empty(list)))
1289 page = list_entry(list->prev, struct page, lru);
1291 page = list_entry(list->next, struct page, lru);
1293 list_del(&page->lru);
1296 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1298 * __GFP_NOFAIL is not to be used in new code.
1300 * All __GFP_NOFAIL callers should be fixed so that they
1301 * properly detect and handle allocation failures.
1303 * We most definitely don't want callers attempting to
1304 * allocate greater than order-1 page units with
1307 WARN_ON_ONCE(order > 1);
1309 spin_lock_irqsave(&zone->lock, flags);
1310 page = __rmqueue(zone, order, migratetype);
1311 spin_unlock(&zone->lock);
1314 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1317 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1318 zone_statistics(preferred_zone, zone);
1319 local_irq_restore(flags);
1321 VM_BUG_ON(bad_range(zone, page));
1322 if (prep_new_page(page, order, gfp_flags))
1327 local_irq_restore(flags);
1331 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1332 #define ALLOC_WMARK_MIN WMARK_MIN
1333 #define ALLOC_WMARK_LOW WMARK_LOW
1334 #define ALLOC_WMARK_HIGH WMARK_HIGH
1335 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1337 /* Mask to get the watermark bits */
1338 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1340 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1341 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1342 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1344 #ifdef CONFIG_FAIL_PAGE_ALLOC
1346 static struct fail_page_alloc_attr {
1347 struct fault_attr attr;
1349 u32 ignore_gfp_highmem;
1350 u32 ignore_gfp_wait;
1353 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1355 struct dentry *ignore_gfp_highmem_file;
1356 struct dentry *ignore_gfp_wait_file;
1357 struct dentry *min_order_file;
1359 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1361 } fail_page_alloc = {
1362 .attr = FAULT_ATTR_INITIALIZER,
1363 .ignore_gfp_wait = 1,
1364 .ignore_gfp_highmem = 1,
1368 static int __init setup_fail_page_alloc(char *str)
1370 return setup_fault_attr(&fail_page_alloc.attr, str);
1372 __setup("fail_page_alloc=", setup_fail_page_alloc);
1374 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1376 if (order < fail_page_alloc.min_order)
1378 if (gfp_mask & __GFP_NOFAIL)
1380 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1382 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1385 return should_fail(&fail_page_alloc.attr, 1 << order);
1388 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1390 static int __init fail_page_alloc_debugfs(void)
1392 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1396 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1400 dir = fail_page_alloc.attr.dentries.dir;
1402 fail_page_alloc.ignore_gfp_wait_file =
1403 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1404 &fail_page_alloc.ignore_gfp_wait);
1406 fail_page_alloc.ignore_gfp_highmem_file =
1407 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1408 &fail_page_alloc.ignore_gfp_highmem);
1409 fail_page_alloc.min_order_file =
1410 debugfs_create_u32("min-order", mode, dir,
1411 &fail_page_alloc.min_order);
1413 if (!fail_page_alloc.ignore_gfp_wait_file ||
1414 !fail_page_alloc.ignore_gfp_highmem_file ||
1415 !fail_page_alloc.min_order_file) {
1417 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1418 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1419 debugfs_remove(fail_page_alloc.min_order_file);
1420 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1426 late_initcall(fail_page_alloc_debugfs);
1428 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1430 #else /* CONFIG_FAIL_PAGE_ALLOC */
1432 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1437 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1440 * Return 1 if free pages are above 'mark'. This takes into account the order
1441 * of the allocation.
1443 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1444 int classzone_idx, int alloc_flags)
1446 /* free_pages my go negative - that's OK */
1448 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
1451 if (alloc_flags & ALLOC_HIGH)
1453 if (alloc_flags & ALLOC_HARDER)
1456 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1458 for (o = 0; o < order; o++) {
1459 /* At the next order, this order's pages become unavailable */
1460 free_pages -= z->free_area[o].nr_free << o;
1462 /* Require fewer higher order pages to be free */
1465 if (free_pages <= min)
1473 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1474 * skip over zones that are not allowed by the cpuset, or that have
1475 * been recently (in last second) found to be nearly full. See further
1476 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1477 * that have to skip over a lot of full or unallowed zones.
1479 * If the zonelist cache is present in the passed in zonelist, then
1480 * returns a pointer to the allowed node mask (either the current
1481 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1483 * If the zonelist cache is not available for this zonelist, does
1484 * nothing and returns NULL.
1486 * If the fullzones BITMAP in the zonelist cache is stale (more than
1487 * a second since last zap'd) then we zap it out (clear its bits.)
1489 * We hold off even calling zlc_setup, until after we've checked the
1490 * first zone in the zonelist, on the theory that most allocations will
1491 * be satisfied from that first zone, so best to examine that zone as
1492 * quickly as we can.
1494 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1496 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1497 nodemask_t *allowednodes; /* zonelist_cache approximation */
1499 zlc = zonelist->zlcache_ptr;
1503 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1504 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1505 zlc->last_full_zap = jiffies;
1508 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1509 &cpuset_current_mems_allowed :
1510 &node_states[N_HIGH_MEMORY];
1511 return allowednodes;
1515 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1516 * if it is worth looking at further for free memory:
1517 * 1) Check that the zone isn't thought to be full (doesn't have its
1518 * bit set in the zonelist_cache fullzones BITMAP).
1519 * 2) Check that the zones node (obtained from the zonelist_cache
1520 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1521 * Return true (non-zero) if zone is worth looking at further, or
1522 * else return false (zero) if it is not.
1524 * This check -ignores- the distinction between various watermarks,
1525 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1526 * found to be full for any variation of these watermarks, it will
1527 * be considered full for up to one second by all requests, unless
1528 * we are so low on memory on all allowed nodes that we are forced
1529 * into the second scan of the zonelist.
1531 * In the second scan we ignore this zonelist cache and exactly
1532 * apply the watermarks to all zones, even it is slower to do so.
1533 * We are low on memory in the second scan, and should leave no stone
1534 * unturned looking for a free page.
1536 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1537 nodemask_t *allowednodes)
1539 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1540 int i; /* index of *z in zonelist zones */
1541 int n; /* node that zone *z is on */
1543 zlc = zonelist->zlcache_ptr;
1547 i = z - zonelist->_zonerefs;
1550 /* This zone is worth trying if it is allowed but not full */
1551 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1555 * Given 'z' scanning a zonelist, set the corresponding bit in
1556 * zlc->fullzones, so that subsequent attempts to allocate a page
1557 * from that zone don't waste time re-examining it.
1559 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1561 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1562 int i; /* index of *z in zonelist zones */
1564 zlc = zonelist->zlcache_ptr;
1568 i = z - zonelist->_zonerefs;
1570 set_bit(i, zlc->fullzones);
1573 #else /* CONFIG_NUMA */
1575 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1580 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1581 nodemask_t *allowednodes)
1586 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1589 #endif /* CONFIG_NUMA */
1592 * get_page_from_freelist goes through the zonelist trying to allocate
1595 static struct page *
1596 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1597 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1598 struct zone *preferred_zone, int migratetype)
1601 struct page *page = NULL;
1604 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1605 int zlc_active = 0; /* set if using zonelist_cache */
1606 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1608 classzone_idx = zone_idx(preferred_zone);
1611 * Scan zonelist, looking for a zone with enough free.
1612 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1614 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1615 high_zoneidx, nodemask) {
1616 if (NUMA_BUILD && zlc_active &&
1617 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1619 if ((alloc_flags & ALLOC_CPUSET) &&
1620 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1623 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1624 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1628 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1629 if (zone_watermark_ok(zone, order, mark,
1630 classzone_idx, alloc_flags))
1633 if (zone_reclaim_mode == 0)
1634 goto this_zone_full;
1636 ret = zone_reclaim(zone, gfp_mask, order);
1638 case ZONE_RECLAIM_NOSCAN:
1641 case ZONE_RECLAIM_FULL:
1642 /* scanned but unreclaimable */
1643 goto this_zone_full;
1645 /* did we reclaim enough */
1646 if (!zone_watermark_ok(zone, order, mark,
1647 classzone_idx, alloc_flags))
1648 goto this_zone_full;
1653 page = buffered_rmqueue(preferred_zone, zone, order,
1654 gfp_mask, migratetype);
1659 zlc_mark_zone_full(zonelist, z);
1661 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1663 * we do zlc_setup after the first zone is tried but only
1664 * if there are multiple nodes make it worthwhile
1666 allowednodes = zlc_setup(zonelist, alloc_flags);
1672 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1673 /* Disable zlc cache for second zonelist scan */
1681 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1682 unsigned long pages_reclaimed)
1684 /* Do not loop if specifically requested */
1685 if (gfp_mask & __GFP_NORETRY)
1689 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1690 * means __GFP_NOFAIL, but that may not be true in other
1693 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1697 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1698 * specified, then we retry until we no longer reclaim any pages
1699 * (above), or we've reclaimed an order of pages at least as
1700 * large as the allocation's order. In both cases, if the
1701 * allocation still fails, we stop retrying.
1703 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1707 * Don't let big-order allocations loop unless the caller
1708 * explicitly requests that.
1710 if (gfp_mask & __GFP_NOFAIL)
1716 static inline struct page *
1717 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1718 struct zonelist *zonelist, enum zone_type high_zoneidx,
1719 nodemask_t *nodemask, struct zone *preferred_zone,
1724 /* Acquire the OOM killer lock for the zones in zonelist */
1725 if (!try_set_zone_oom(zonelist, gfp_mask)) {
1726 schedule_timeout_uninterruptible(1);
1731 * Go through the zonelist yet one more time, keep very high watermark
1732 * here, this is only to catch a parallel oom killing, we must fail if
1733 * we're still under heavy pressure.
1735 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1736 order, zonelist, high_zoneidx,
1737 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1738 preferred_zone, migratetype);
1742 if (!(gfp_mask & __GFP_NOFAIL)) {
1743 /* The OOM killer will not help higher order allocs */
1744 if (order > PAGE_ALLOC_COSTLY_ORDER)
1747 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1748 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1749 * The caller should handle page allocation failure by itself if
1750 * it specifies __GFP_THISNODE.
1751 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1753 if (gfp_mask & __GFP_THISNODE)
1756 /* Exhausted what can be done so it's blamo time */
1757 out_of_memory(zonelist, gfp_mask, order, nodemask);
1760 clear_zonelist_oom(zonelist, gfp_mask);
1764 #ifdef CONFIG_COMPACTION
1765 /* Try memory compaction for high-order allocations before reclaim */
1766 static struct page *
1767 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1768 struct zonelist *zonelist, enum zone_type high_zoneidx,
1769 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1770 int migratetype, unsigned long *did_some_progress)
1774 if (!order || compaction_deferred(preferred_zone))
1777 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1779 if (*did_some_progress != COMPACT_SKIPPED) {
1781 /* Page migration frees to the PCP lists but we want merging */
1782 drain_pages(get_cpu());
1785 page = get_page_from_freelist(gfp_mask, nodemask,
1786 order, zonelist, high_zoneidx,
1787 alloc_flags, preferred_zone,
1790 preferred_zone->compact_considered = 0;
1791 preferred_zone->compact_defer_shift = 0;
1792 count_vm_event(COMPACTSUCCESS);
1797 * It's bad if compaction run occurs and fails.
1798 * The most likely reason is that pages exist,
1799 * but not enough to satisfy watermarks.
1801 count_vm_event(COMPACTFAIL);
1802 defer_compaction(preferred_zone);
1810 static inline struct page *
1811 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1812 struct zonelist *zonelist, enum zone_type high_zoneidx,
1813 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1814 int migratetype, unsigned long *did_some_progress)
1818 #endif /* CONFIG_COMPACTION */
1820 /* The really slow allocator path where we enter direct reclaim */
1821 static inline struct page *
1822 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1823 struct zonelist *zonelist, enum zone_type high_zoneidx,
1824 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1825 int migratetype, unsigned long *did_some_progress)
1827 struct page *page = NULL;
1828 struct reclaim_state reclaim_state;
1829 struct task_struct *p = current;
1833 /* We now go into synchronous reclaim */
1834 cpuset_memory_pressure_bump();
1835 p->flags |= PF_MEMALLOC;
1836 lockdep_set_current_reclaim_state(gfp_mask);
1837 reclaim_state.reclaimed_slab = 0;
1838 p->reclaim_state = &reclaim_state;
1840 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1842 p->reclaim_state = NULL;
1843 lockdep_clear_current_reclaim_state();
1844 p->flags &= ~PF_MEMALLOC;
1851 if (likely(*did_some_progress))
1852 page = get_page_from_freelist(gfp_mask, nodemask, order,
1853 zonelist, high_zoneidx,
1854 alloc_flags, preferred_zone,
1860 * This is called in the allocator slow-path if the allocation request is of
1861 * sufficient urgency to ignore watermarks and take other desperate measures
1863 static inline struct page *
1864 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1865 struct zonelist *zonelist, enum zone_type high_zoneidx,
1866 nodemask_t *nodemask, struct zone *preferred_zone,
1872 page = get_page_from_freelist(gfp_mask, nodemask, order,
1873 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
1874 preferred_zone, migratetype);
1876 if (!page && gfp_mask & __GFP_NOFAIL)
1877 congestion_wait(BLK_RW_ASYNC, HZ/50);
1878 } while (!page && (gfp_mask & __GFP_NOFAIL));
1884 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
1885 enum zone_type high_zoneidx)
1890 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1891 wakeup_kswapd(zone, order);
1895 gfp_to_alloc_flags(gfp_t gfp_mask)
1897 struct task_struct *p = current;
1898 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
1899 const gfp_t wait = gfp_mask & __GFP_WAIT;
1901 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1902 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH);
1905 * The caller may dip into page reserves a bit more if the caller
1906 * cannot run direct reclaim, or if the caller has realtime scheduling
1907 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1908 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1910 alloc_flags |= (gfp_mask & __GFP_HIGH);
1913 alloc_flags |= ALLOC_HARDER;
1915 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1916 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1918 alloc_flags &= ~ALLOC_CPUSET;
1919 } else if (unlikely(rt_task(p)) && !in_interrupt())
1920 alloc_flags |= ALLOC_HARDER;
1922 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
1923 if (!in_interrupt() &&
1924 ((p->flags & PF_MEMALLOC) ||
1925 unlikely(test_thread_flag(TIF_MEMDIE))))
1926 alloc_flags |= ALLOC_NO_WATERMARKS;
1932 static inline struct page *
1933 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
1934 struct zonelist *zonelist, enum zone_type high_zoneidx,
1935 nodemask_t *nodemask, struct zone *preferred_zone,
1938 const gfp_t wait = gfp_mask & __GFP_WAIT;
1939 struct page *page = NULL;
1941 unsigned long pages_reclaimed = 0;
1942 unsigned long did_some_progress;
1943 struct task_struct *p = current;
1946 * In the slowpath, we sanity check order to avoid ever trying to
1947 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1948 * be using allocators in order of preference for an area that is
1951 if (order >= MAX_ORDER) {
1952 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
1957 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1958 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1959 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1960 * using a larger set of nodes after it has established that the
1961 * allowed per node queues are empty and that nodes are
1964 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1968 wake_all_kswapd(order, zonelist, high_zoneidx);
1971 * OK, we're below the kswapd watermark and have kicked background
1972 * reclaim. Now things get more complex, so set up alloc_flags according
1973 * to how we want to proceed.
1975 alloc_flags = gfp_to_alloc_flags(gfp_mask);
1977 /* This is the last chance, in general, before the goto nopage. */
1978 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
1979 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
1980 preferred_zone, migratetype);
1985 /* Allocate without watermarks if the context allows */
1986 if (alloc_flags & ALLOC_NO_WATERMARKS) {
1987 page = __alloc_pages_high_priority(gfp_mask, order,
1988 zonelist, high_zoneidx, nodemask,
1989 preferred_zone, migratetype);
1994 /* Atomic allocations - we can't balance anything */
1998 /* Avoid recursion of direct reclaim */
1999 if (p->flags & PF_MEMALLOC)
2002 /* Avoid allocations with no watermarks from looping endlessly */
2003 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2006 /* Try direct compaction */
2007 page = __alloc_pages_direct_compact(gfp_mask, order,
2008 zonelist, high_zoneidx,
2010 alloc_flags, preferred_zone,
2011 migratetype, &did_some_progress);
2015 /* Try direct reclaim and then allocating */
2016 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2017 zonelist, high_zoneidx,
2019 alloc_flags, preferred_zone,
2020 migratetype, &did_some_progress);
2025 * If we failed to make any progress reclaiming, then we are
2026 * running out of options and have to consider going OOM
2028 if (!did_some_progress) {
2029 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2030 if (oom_killer_disabled)
2032 page = __alloc_pages_may_oom(gfp_mask, order,
2033 zonelist, high_zoneidx,
2034 nodemask, preferred_zone,
2040 * The OOM killer does not trigger for high-order
2041 * ~__GFP_NOFAIL allocations so if no progress is being
2042 * made, there are no other options and retrying is
2045 if (order > PAGE_ALLOC_COSTLY_ORDER &&
2046 !(gfp_mask & __GFP_NOFAIL))
2053 /* Check if we should retry the allocation */
2054 pages_reclaimed += did_some_progress;
2055 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
2056 /* Wait for some write requests to complete then retry */
2057 congestion_wait(BLK_RW_ASYNC, HZ/50);
2062 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
2063 printk(KERN_WARNING "%s: page allocation failure."
2064 " order:%d, mode:0x%x\n",
2065 p->comm, order, gfp_mask);
2071 if (kmemcheck_enabled)
2072 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2078 * This is the 'heart' of the zoned buddy allocator.
2081 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2082 struct zonelist *zonelist, nodemask_t *nodemask)
2084 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2085 struct zone *preferred_zone;
2087 int migratetype = allocflags_to_migratetype(gfp_mask);
2089 gfp_mask &= gfp_allowed_mask;
2091 lockdep_trace_alloc(gfp_mask);
2093 might_sleep_if(gfp_mask & __GFP_WAIT);
2095 if (should_fail_alloc_page(gfp_mask, order))
2099 * Check the zones suitable for the gfp_mask contain at least one
2100 * valid zone. It's possible to have an empty zonelist as a result
2101 * of GFP_THISNODE and a memoryless node
2103 if (unlikely(!zonelist->_zonerefs->zone))
2107 /* The preferred zone is used for statistics later */
2108 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone);
2109 if (!preferred_zone) {
2114 /* First allocation attempt */
2115 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2116 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2117 preferred_zone, migratetype);
2118 if (unlikely(!page))
2119 page = __alloc_pages_slowpath(gfp_mask, order,
2120 zonelist, high_zoneidx, nodemask,
2121 preferred_zone, migratetype);
2124 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2127 EXPORT_SYMBOL(__alloc_pages_nodemask);
2130 * Common helper functions.
2132 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2137 * __get_free_pages() returns a 32-bit address, which cannot represent
2140 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2142 page = alloc_pages(gfp_mask, order);
2145 return (unsigned long) page_address(page);
2147 EXPORT_SYMBOL(__get_free_pages);
2149 unsigned long get_zeroed_page(gfp_t gfp_mask)
2151 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2153 EXPORT_SYMBOL(get_zeroed_page);
2155 void __pagevec_free(struct pagevec *pvec)
2157 int i = pagevec_count(pvec);
2160 trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
2161 free_hot_cold_page(pvec->pages[i], pvec->cold);
2165 void __free_pages(struct page *page, unsigned int order)
2167 if (put_page_testzero(page)) {
2169 free_hot_cold_page(page, 0);
2171 __free_pages_ok(page, order);
2175 EXPORT_SYMBOL(__free_pages);
2177 void free_pages(unsigned long addr, unsigned int order)
2180 VM_BUG_ON(!virt_addr_valid((void *)addr));
2181 __free_pages(virt_to_page((void *)addr), order);
2185 EXPORT_SYMBOL(free_pages);
2188 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2189 * @size: the number of bytes to allocate
2190 * @gfp_mask: GFP flags for the allocation
2192 * This function is similar to alloc_pages(), except that it allocates the
2193 * minimum number of pages to satisfy the request. alloc_pages() can only
2194 * allocate memory in power-of-two pages.
2196 * This function is also limited by MAX_ORDER.
2198 * Memory allocated by this function must be released by free_pages_exact().
2200 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2202 unsigned int order = get_order(size);
2205 addr = __get_free_pages(gfp_mask, order);
2207 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2208 unsigned long used = addr + PAGE_ALIGN(size);
2210 split_page(virt_to_page((void *)addr), order);
2211 while (used < alloc_end) {
2217 return (void *)addr;
2219 EXPORT_SYMBOL(alloc_pages_exact);
2222 * free_pages_exact - release memory allocated via alloc_pages_exact()
2223 * @virt: the value returned by alloc_pages_exact.
2224 * @size: size of allocation, same value as passed to alloc_pages_exact().
2226 * Release the memory allocated by a previous call to alloc_pages_exact.
2228 void free_pages_exact(void *virt, size_t size)
2230 unsigned long addr = (unsigned long)virt;
2231 unsigned long end = addr + PAGE_ALIGN(size);
2233 while (addr < end) {
2238 EXPORT_SYMBOL(free_pages_exact);
2240 static unsigned int nr_free_zone_pages(int offset)
2245 /* Just pick one node, since fallback list is circular */
2246 unsigned int sum = 0;
2248 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2250 for_each_zone_zonelist(zone, z, zonelist, offset) {
2251 unsigned long size = zone->present_pages;
2252 unsigned long high = high_wmark_pages(zone);
2261 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2263 unsigned int nr_free_buffer_pages(void)
2265 return nr_free_zone_pages(gfp_zone(GFP_USER));
2267 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2270 * Amount of free RAM allocatable within all zones
2272 unsigned int nr_free_pagecache_pages(void)
2274 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2277 static inline void show_node(struct zone *zone)
2280 printk("Node %d ", zone_to_nid(zone));
2283 void si_meminfo(struct sysinfo *val)
2285 val->totalram = totalram_pages;
2287 val->freeram = global_page_state(NR_FREE_PAGES);
2288 val->bufferram = nr_blockdev_pages();
2289 val->totalhigh = totalhigh_pages;
2290 val->freehigh = nr_free_highpages();
2291 val->mem_unit = PAGE_SIZE;
2294 EXPORT_SYMBOL(si_meminfo);
2297 void si_meminfo_node(struct sysinfo *val, int nid)
2299 pg_data_t *pgdat = NODE_DATA(nid);
2301 val->totalram = pgdat->node_present_pages;
2302 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2303 #ifdef CONFIG_HIGHMEM
2304 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2305 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2311 val->mem_unit = PAGE_SIZE;
2315 #define K(x) ((x) << (PAGE_SHIFT-10))
2318 * Show free area list (used inside shift_scroll-lock stuff)
2319 * We also calculate the percentage fragmentation. We do this by counting the
2320 * memory on each free list with the exception of the first item on the list.
2322 void show_free_areas(void)
2327 for_each_populated_zone(zone) {
2329 printk("%s per-cpu:\n", zone->name);
2331 for_each_online_cpu(cpu) {
2332 struct per_cpu_pageset *pageset;
2334 pageset = per_cpu_ptr(zone->pageset, cpu);
2336 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2337 cpu, pageset->pcp.high,
2338 pageset->pcp.batch, pageset->pcp.count);
2342 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2343 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2345 " dirty:%lu writeback:%lu unstable:%lu\n"
2346 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2347 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2348 global_page_state(NR_ACTIVE_ANON),
2349 global_page_state(NR_INACTIVE_ANON),
2350 global_page_state(NR_ISOLATED_ANON),
2351 global_page_state(NR_ACTIVE_FILE),
2352 global_page_state(NR_INACTIVE_FILE),
2353 global_page_state(NR_ISOLATED_FILE),
2354 global_page_state(NR_UNEVICTABLE),
2355 global_page_state(NR_FILE_DIRTY),
2356 global_page_state(NR_WRITEBACK),
2357 global_page_state(NR_UNSTABLE_NFS),
2358 global_page_state(NR_FREE_PAGES),
2359 global_page_state(NR_SLAB_RECLAIMABLE),
2360 global_page_state(NR_SLAB_UNRECLAIMABLE),
2361 global_page_state(NR_FILE_MAPPED),
2362 global_page_state(NR_SHMEM),
2363 global_page_state(NR_PAGETABLE),
2364 global_page_state(NR_BOUNCE));
2366 for_each_populated_zone(zone) {
2375 " active_anon:%lukB"
2376 " inactive_anon:%lukB"
2377 " active_file:%lukB"
2378 " inactive_file:%lukB"
2379 " unevictable:%lukB"
2380 " isolated(anon):%lukB"
2381 " isolated(file):%lukB"
2388 " slab_reclaimable:%lukB"
2389 " slab_unreclaimable:%lukB"
2390 " kernel_stack:%lukB"
2394 " writeback_tmp:%lukB"
2395 " pages_scanned:%lu"
2396 " all_unreclaimable? %s"
2399 K(zone_page_state(zone, NR_FREE_PAGES)),
2400 K(min_wmark_pages(zone)),
2401 K(low_wmark_pages(zone)),
2402 K(high_wmark_pages(zone)),
2403 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2404 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2405 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2406 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2407 K(zone_page_state(zone, NR_UNEVICTABLE)),
2408 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2409 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2410 K(zone->present_pages),
2411 K(zone_page_state(zone, NR_MLOCK)),
2412 K(zone_page_state(zone, NR_FILE_DIRTY)),
2413 K(zone_page_state(zone, NR_WRITEBACK)),
2414 K(zone_page_state(zone, NR_FILE_MAPPED)),
2415 K(zone_page_state(zone, NR_SHMEM)),
2416 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2417 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2418 zone_page_state(zone, NR_KERNEL_STACK) *
2420 K(zone_page_state(zone, NR_PAGETABLE)),
2421 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2422 K(zone_page_state(zone, NR_BOUNCE)),
2423 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2424 zone->pages_scanned,
2425 (zone->all_unreclaimable ? "yes" : "no")
2427 printk("lowmem_reserve[]:");
2428 for (i = 0; i < MAX_NR_ZONES; i++)
2429 printk(" %lu", zone->lowmem_reserve[i]);
2433 for_each_populated_zone(zone) {
2434 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2437 printk("%s: ", zone->name);
2439 spin_lock_irqsave(&zone->lock, flags);
2440 for (order = 0; order < MAX_ORDER; order++) {
2441 nr[order] = zone->free_area[order].nr_free;
2442 total += nr[order] << order;
2444 spin_unlock_irqrestore(&zone->lock, flags);
2445 for (order = 0; order < MAX_ORDER; order++)
2446 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2447 printk("= %lukB\n", K(total));
2450 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2452 show_swap_cache_info();
2455 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2457 zoneref->zone = zone;
2458 zoneref->zone_idx = zone_idx(zone);
2462 * Builds allocation fallback zone lists.
2464 * Add all populated zones of a node to the zonelist.
2466 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2467 int nr_zones, enum zone_type zone_type)
2471 BUG_ON(zone_type >= MAX_NR_ZONES);
2476 zone = pgdat->node_zones + zone_type;
2477 if (populated_zone(zone)) {
2478 zoneref_set_zone(zone,
2479 &zonelist->_zonerefs[nr_zones++]);
2480 check_highest_zone(zone_type);
2483 } while (zone_type);
2490 * 0 = automatic detection of better ordering.
2491 * 1 = order by ([node] distance, -zonetype)
2492 * 2 = order by (-zonetype, [node] distance)
2494 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2495 * the same zonelist. So only NUMA can configure this param.
2497 #define ZONELIST_ORDER_DEFAULT 0
2498 #define ZONELIST_ORDER_NODE 1
2499 #define ZONELIST_ORDER_ZONE 2
2501 /* zonelist order in the kernel.
2502 * set_zonelist_order() will set this to NODE or ZONE.
2504 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2505 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2509 /* The value user specified ....changed by config */
2510 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2511 /* string for sysctl */
2512 #define NUMA_ZONELIST_ORDER_LEN 16
2513 char numa_zonelist_order[16] = "default";
2516 * interface for configure zonelist ordering.
2517 * command line option "numa_zonelist_order"
2518 * = "[dD]efault - default, automatic configuration.
2519 * = "[nN]ode - order by node locality, then by zone within node
2520 * = "[zZ]one - order by zone, then by locality within zone
2523 static int __parse_numa_zonelist_order(char *s)
2525 if (*s == 'd' || *s == 'D') {
2526 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2527 } else if (*s == 'n' || *s == 'N') {
2528 user_zonelist_order = ZONELIST_ORDER_NODE;
2529 } else if (*s == 'z' || *s == 'Z') {
2530 user_zonelist_order = ZONELIST_ORDER_ZONE;
2533 "Ignoring invalid numa_zonelist_order value: "
2540 static __init int setup_numa_zonelist_order(char *s)
2543 return __parse_numa_zonelist_order(s);
2546 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2549 * sysctl handler for numa_zonelist_order
2551 int numa_zonelist_order_handler(ctl_table *table, int write,
2552 void __user *buffer, size_t *length,
2555 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2557 static DEFINE_MUTEX(zl_order_mutex);
2559 mutex_lock(&zl_order_mutex);
2561 strcpy(saved_string, (char*)table->data);
2562 ret = proc_dostring(table, write, buffer, length, ppos);
2566 int oldval = user_zonelist_order;
2567 if (__parse_numa_zonelist_order((char*)table->data)) {
2569 * bogus value. restore saved string
2571 strncpy((char*)table->data, saved_string,
2572 NUMA_ZONELIST_ORDER_LEN);
2573 user_zonelist_order = oldval;
2574 } else if (oldval != user_zonelist_order)
2575 build_all_zonelists(NULL);
2578 mutex_unlock(&zl_order_mutex);
2583 #define MAX_NODE_LOAD (nr_online_nodes)
2584 static int node_load[MAX_NUMNODES];
2587 * find_next_best_node - find the next node that should appear in a given node's fallback list
2588 * @node: node whose fallback list we're appending
2589 * @used_node_mask: nodemask_t of already used nodes
2591 * We use a number of factors to determine which is the next node that should
2592 * appear on a given node's fallback list. The node should not have appeared
2593 * already in @node's fallback list, and it should be the next closest node
2594 * according to the distance array (which contains arbitrary distance values
2595 * from each node to each node in the system), and should also prefer nodes
2596 * with no CPUs, since presumably they'll have very little allocation pressure
2597 * on them otherwise.
2598 * It returns -1 if no node is found.
2600 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2603 int min_val = INT_MAX;
2605 const struct cpumask *tmp = cpumask_of_node(0);
2607 /* Use the local node if we haven't already */
2608 if (!node_isset(node, *used_node_mask)) {
2609 node_set(node, *used_node_mask);
2613 for_each_node_state(n, N_HIGH_MEMORY) {
2615 /* Don't want a node to appear more than once */
2616 if (node_isset(n, *used_node_mask))
2619 /* Use the distance array to find the distance */
2620 val = node_distance(node, n);
2622 /* Penalize nodes under us ("prefer the next node") */
2625 /* Give preference to headless and unused nodes */
2626 tmp = cpumask_of_node(n);
2627 if (!cpumask_empty(tmp))
2628 val += PENALTY_FOR_NODE_WITH_CPUS;
2630 /* Slight preference for less loaded node */
2631 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2632 val += node_load[n];
2634 if (val < min_val) {
2641 node_set(best_node, *used_node_mask);
2648 * Build zonelists ordered by node and zones within node.
2649 * This results in maximum locality--normal zone overflows into local
2650 * DMA zone, if any--but risks exhausting DMA zone.
2652 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2655 struct zonelist *zonelist;
2657 zonelist = &pgdat->node_zonelists[0];
2658 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2660 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2662 zonelist->_zonerefs[j].zone = NULL;
2663 zonelist->_zonerefs[j].zone_idx = 0;
2667 * Build gfp_thisnode zonelists
2669 static void build_thisnode_zonelists(pg_data_t *pgdat)
2672 struct zonelist *zonelist;
2674 zonelist = &pgdat->node_zonelists[1];
2675 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2676 zonelist->_zonerefs[j].zone = NULL;
2677 zonelist->_zonerefs[j].zone_idx = 0;
2681 * Build zonelists ordered by zone and nodes within zones.
2682 * This results in conserving DMA zone[s] until all Normal memory is
2683 * exhausted, but results in overflowing to remote node while memory
2684 * may still exist in local DMA zone.
2686 static int node_order[MAX_NUMNODES];
2688 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2691 int zone_type; /* needs to be signed */
2693 struct zonelist *zonelist;
2695 zonelist = &pgdat->node_zonelists[0];
2697 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2698 for (j = 0; j < nr_nodes; j++) {
2699 node = node_order[j];
2700 z = &NODE_DATA(node)->node_zones[zone_type];
2701 if (populated_zone(z)) {
2703 &zonelist->_zonerefs[pos++]);
2704 check_highest_zone(zone_type);
2708 zonelist->_zonerefs[pos].zone = NULL;
2709 zonelist->_zonerefs[pos].zone_idx = 0;
2712 static int default_zonelist_order(void)
2715 unsigned long low_kmem_size,total_size;
2719 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2720 * If they are really small and used heavily, the system can fall
2721 * into OOM very easily.
2722 * This function detect ZONE_DMA/DMA32 size and configures zone order.
2724 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2727 for_each_online_node(nid) {
2728 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2729 z = &NODE_DATA(nid)->node_zones[zone_type];
2730 if (populated_zone(z)) {
2731 if (zone_type < ZONE_NORMAL)
2732 low_kmem_size += z->present_pages;
2733 total_size += z->present_pages;
2734 } else if (zone_type == ZONE_NORMAL) {
2736 * If any node has only lowmem, then node order
2737 * is preferred to allow kernel allocations
2738 * locally; otherwise, they can easily infringe
2739 * on other nodes when there is an abundance of
2740 * lowmem available to allocate from.
2742 return ZONELIST_ORDER_NODE;
2746 if (!low_kmem_size || /* there are no DMA area. */
2747 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2748 return ZONELIST_ORDER_NODE;
2750 * look into each node's config.
2751 * If there is a node whose DMA/DMA32 memory is very big area on
2752 * local memory, NODE_ORDER may be suitable.
2754 average_size = total_size /
2755 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2756 for_each_online_node(nid) {
2759 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2760 z = &NODE_DATA(nid)->node_zones[zone_type];
2761 if (populated_zone(z)) {
2762 if (zone_type < ZONE_NORMAL)
2763 low_kmem_size += z->present_pages;
2764 total_size += z->present_pages;
2767 if (low_kmem_size &&
2768 total_size > average_size && /* ignore small node */
2769 low_kmem_size > total_size * 70/100)
2770 return ZONELIST_ORDER_NODE;
2772 return ZONELIST_ORDER_ZONE;
2775 static void set_zonelist_order(void)
2777 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2778 current_zonelist_order = default_zonelist_order();
2780 current_zonelist_order = user_zonelist_order;
2783 static void build_zonelists(pg_data_t *pgdat)
2787 nodemask_t used_mask;
2788 int local_node, prev_node;
2789 struct zonelist *zonelist;
2790 int order = current_zonelist_order;
2792 /* initialize zonelists */
2793 for (i = 0; i < MAX_ZONELISTS; i++) {
2794 zonelist = pgdat->node_zonelists + i;
2795 zonelist->_zonerefs[0].zone = NULL;
2796 zonelist->_zonerefs[0].zone_idx = 0;
2799 /* NUMA-aware ordering of nodes */
2800 local_node = pgdat->node_id;
2801 load = nr_online_nodes;
2802 prev_node = local_node;
2803 nodes_clear(used_mask);
2805 memset(node_order, 0, sizeof(node_order));
2808 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2809 int distance = node_distance(local_node, node);
2812 * If another node is sufficiently far away then it is better
2813 * to reclaim pages in a zone before going off node.
2815 if (distance > RECLAIM_DISTANCE)
2816 zone_reclaim_mode = 1;
2819 * We don't want to pressure a particular node.
2820 * So adding penalty to the first node in same
2821 * distance group to make it round-robin.
2823 if (distance != node_distance(local_node, prev_node))
2824 node_load[node] = load;
2828 if (order == ZONELIST_ORDER_NODE)
2829 build_zonelists_in_node_order(pgdat, node);
2831 node_order[j++] = node; /* remember order */
2834 if (order == ZONELIST_ORDER_ZONE) {
2835 /* calculate node order -- i.e., DMA last! */
2836 build_zonelists_in_zone_order(pgdat, j);
2839 build_thisnode_zonelists(pgdat);
2842 /* Construct the zonelist performance cache - see further mmzone.h */
2843 static void build_zonelist_cache(pg_data_t *pgdat)
2845 struct zonelist *zonelist;
2846 struct zonelist_cache *zlc;
2849 zonelist = &pgdat->node_zonelists[0];
2850 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2851 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2852 for (z = zonelist->_zonerefs; z->zone; z++)
2853 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2857 #else /* CONFIG_NUMA */
2859 static void set_zonelist_order(void)
2861 current_zonelist_order = ZONELIST_ORDER_ZONE;
2864 static void build_zonelists(pg_data_t *pgdat)
2866 int node, local_node;
2868 struct zonelist *zonelist;
2870 local_node = pgdat->node_id;
2872 zonelist = &pgdat->node_zonelists[0];
2873 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2876 * Now we build the zonelist so that it contains the zones
2877 * of all the other nodes.
2878 * We don't want to pressure a particular node, so when
2879 * building the zones for node N, we make sure that the
2880 * zones coming right after the local ones are those from
2881 * node N+1 (modulo N)
2883 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2884 if (!node_online(node))
2886 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2889 for (node = 0; node < local_node; node++) {
2890 if (!node_online(node))
2892 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2896 zonelist->_zonerefs[j].zone = NULL;
2897 zonelist->_zonerefs[j].zone_idx = 0;
2900 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2901 static void build_zonelist_cache(pg_data_t *pgdat)
2903 pgdat->node_zonelists[0].zlcache_ptr = NULL;
2906 #endif /* CONFIG_NUMA */
2909 * Boot pageset table. One per cpu which is going to be used for all
2910 * zones and all nodes. The parameters will be set in such a way
2911 * that an item put on a list will immediately be handed over to
2912 * the buddy list. This is safe since pageset manipulation is done
2913 * with interrupts disabled.
2915 * The boot_pagesets must be kept even after bootup is complete for
2916 * unused processors and/or zones. They do play a role for bootstrapping
2917 * hotplugged processors.
2919 * zoneinfo_show() and maybe other functions do
2920 * not check if the processor is online before following the pageset pointer.
2921 * Other parts of the kernel may not check if the zone is available.
2923 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
2924 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
2925 static void setup_zone_pageset(struct zone *zone);
2927 /* return values int ....just for stop_machine() */
2928 static __init_refok int __build_all_zonelists(void *data)
2934 memset(node_load, 0, sizeof(node_load));
2936 for_each_online_node(nid) {
2937 pg_data_t *pgdat = NODE_DATA(nid);
2939 build_zonelists(pgdat);
2940 build_zonelist_cache(pgdat);
2943 #ifdef CONFIG_MEMORY_HOTPLUG
2944 /* Setup real pagesets for the new zone */
2946 struct zone *zone = data;
2947 setup_zone_pageset(zone);
2952 * Initialize the boot_pagesets that are going to be used
2953 * for bootstrapping processors. The real pagesets for
2954 * each zone will be allocated later when the per cpu
2955 * allocator is available.
2957 * boot_pagesets are used also for bootstrapping offline
2958 * cpus if the system is already booted because the pagesets
2959 * are needed to initialize allocators on a specific cpu too.
2960 * F.e. the percpu allocator needs the page allocator which
2961 * needs the percpu allocator in order to allocate its pagesets
2962 * (a chicken-egg dilemma).
2964 for_each_possible_cpu(cpu)
2965 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
2970 void build_all_zonelists(void *data)
2972 set_zonelist_order();
2974 if (system_state == SYSTEM_BOOTING) {
2975 __build_all_zonelists(NULL);
2976 mminit_verify_zonelist();
2977 cpuset_init_current_mems_allowed();
2979 /* we have to stop all cpus to guarantee there is no user
2981 stop_machine(__build_all_zonelists, data, NULL);
2982 /* cpuset refresh routine should be here */
2984 vm_total_pages = nr_free_pagecache_pages();
2986 * Disable grouping by mobility if the number of pages in the
2987 * system is too low to allow the mechanism to work. It would be
2988 * more accurate, but expensive to check per-zone. This check is
2989 * made on memory-hotadd so a system can start with mobility
2990 * disabled and enable it later
2992 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
2993 page_group_by_mobility_disabled = 1;
2995 page_group_by_mobility_disabled = 0;
2997 printk("Built %i zonelists in %s order, mobility grouping %s. "
2998 "Total pages: %ld\n",
3000 zonelist_order_name[current_zonelist_order],
3001 page_group_by_mobility_disabled ? "off" : "on",
3004 printk("Policy zone: %s\n", zone_names[policy_zone]);
3009 * Helper functions to size the waitqueue hash table.
3010 * Essentially these want to choose hash table sizes sufficiently
3011 * large so that collisions trying to wait on pages are rare.
3012 * But in fact, the number of active page waitqueues on typical
3013 * systems is ridiculously low, less than 200. So this is even
3014 * conservative, even though it seems large.
3016 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3017 * waitqueues, i.e. the size of the waitq table given the number of pages.
3019 #define PAGES_PER_WAITQUEUE 256
3021 #ifndef CONFIG_MEMORY_HOTPLUG
3022 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3024 unsigned long size = 1;
3026 pages /= PAGES_PER_WAITQUEUE;
3028 while (size < pages)
3032 * Once we have dozens or even hundreds of threads sleeping
3033 * on IO we've got bigger problems than wait queue collision.
3034 * Limit the size of the wait table to a reasonable size.
3036 size = min(size, 4096UL);
3038 return max(size, 4UL);
3042 * A zone's size might be changed by hot-add, so it is not possible to determine
3043 * a suitable size for its wait_table. So we use the maximum size now.
3045 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3047 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3048 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3049 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3051 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3052 * or more by the traditional way. (See above). It equals:
3054 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3055 * ia64(16K page size) : = ( 8G + 4M)byte.
3056 * powerpc (64K page size) : = (32G +16M)byte.
3058 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3065 * This is an integer logarithm so that shifts can be used later
3066 * to extract the more random high bits from the multiplicative
3067 * hash function before the remainder is taken.
3069 static inline unsigned long wait_table_bits(unsigned long size)
3074 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3077 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3078 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3079 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3080 * higher will lead to a bigger reserve which will get freed as contiguous
3081 * blocks as reclaim kicks in
3083 static void setup_zone_migrate_reserve(struct zone *zone)
3085 unsigned long start_pfn, pfn, end_pfn;
3087 unsigned long block_migratetype;
3090 /* Get the start pfn, end pfn and the number of blocks to reserve */
3091 start_pfn = zone->zone_start_pfn;
3092 end_pfn = start_pfn + zone->spanned_pages;
3093 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3097 * Reserve blocks are generally in place to help high-order atomic
3098 * allocations that are short-lived. A min_free_kbytes value that
3099 * would result in more than 2 reserve blocks for atomic allocations
3100 * is assumed to be in place to help anti-fragmentation for the
3101 * future allocation of hugepages at runtime.
3103 reserve = min(2, reserve);
3105 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3106 if (!pfn_valid(pfn))
3108 page = pfn_to_page(pfn);
3110 /* Watch out for overlapping nodes */
3111 if (page_to_nid(page) != zone_to_nid(zone))
3114 /* Blocks with reserved pages will never free, skip them. */
3115 if (PageReserved(page))
3118 block_migratetype = get_pageblock_migratetype(page);
3120 /* If this block is reserved, account for it */
3121 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
3126 /* Suitable for reserving if this block is movable */
3127 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
3128 set_pageblock_migratetype(page, MIGRATE_RESERVE);
3129 move_freepages_block(zone, page, MIGRATE_RESERVE);
3135 * If the reserve is met and this is a previous reserved block,
3138 if (block_migratetype == MIGRATE_RESERVE) {
3139 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3140 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3146 * Initially all pages are reserved - free ones are freed
3147 * up by free_all_bootmem() once the early boot process is
3148 * done. Non-atomic initialization, single-pass.
3150 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3151 unsigned long start_pfn, enum memmap_context context)
3154 unsigned long end_pfn = start_pfn + size;
3158 if (highest_memmap_pfn < end_pfn - 1)
3159 highest_memmap_pfn = end_pfn - 1;
3161 z = &NODE_DATA(nid)->node_zones[zone];
3162 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3164 * There can be holes in boot-time mem_map[]s
3165 * handed to this function. They do not
3166 * exist on hotplugged memory.
3168 if (context == MEMMAP_EARLY) {
3169 if (!early_pfn_valid(pfn))
3171 if (!early_pfn_in_nid(pfn, nid))
3174 page = pfn_to_page(pfn);
3175 set_page_links(page, zone, nid, pfn);
3176 mminit_verify_page_links(page, zone, nid, pfn);
3177 init_page_count(page);
3178 reset_page_mapcount(page);
3179 SetPageReserved(page);
3181 * Mark the block movable so that blocks are reserved for
3182 * movable at startup. This will force kernel allocations
3183 * to reserve their blocks rather than leaking throughout
3184 * the address space during boot when many long-lived
3185 * kernel allocations are made. Later some blocks near
3186 * the start are marked MIGRATE_RESERVE by
3187 * setup_zone_migrate_reserve()
3189 * bitmap is created for zone's valid pfn range. but memmap
3190 * can be created for invalid pages (for alignment)
3191 * check here not to call set_pageblock_migratetype() against
3194 if ((z->zone_start_pfn <= pfn)
3195 && (pfn < z->zone_start_pfn + z->spanned_pages)
3196 && !(pfn & (pageblock_nr_pages - 1)))
3197 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3199 INIT_LIST_HEAD(&page->lru);
3200 #ifdef WANT_PAGE_VIRTUAL
3201 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3202 if (!is_highmem_idx(zone))
3203 set_page_address(page, __va(pfn << PAGE_SHIFT));
3208 static void __meminit zone_init_free_lists(struct zone *zone)
3211 for_each_migratetype_order(order, t) {
3212 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3213 zone->free_area[order].nr_free = 0;
3217 #ifndef __HAVE_ARCH_MEMMAP_INIT
3218 #define memmap_init(size, nid, zone, start_pfn) \
3219 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3222 static int zone_batchsize(struct zone *zone)
3228 * The per-cpu-pages pools are set to around 1000th of the
3229 * size of the zone. But no more than 1/2 of a meg.
3231 * OK, so we don't know how big the cache is. So guess.
3233 batch = zone->present_pages / 1024;
3234 if (batch * PAGE_SIZE > 512 * 1024)
3235 batch = (512 * 1024) / PAGE_SIZE;
3236 batch /= 4; /* We effectively *= 4 below */
3241 * Clamp the batch to a 2^n - 1 value. Having a power
3242 * of 2 value was found to be more likely to have
3243 * suboptimal cache aliasing properties in some cases.
3245 * For example if 2 tasks are alternately allocating
3246 * batches of pages, one task can end up with a lot
3247 * of pages of one half of the possible page colors
3248 * and the other with pages of the other colors.
3250 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3255 /* The deferral and batching of frees should be suppressed under NOMMU
3258 * The problem is that NOMMU needs to be able to allocate large chunks
3259 * of contiguous memory as there's no hardware page translation to
3260 * assemble apparent contiguous memory from discontiguous pages.
3262 * Queueing large contiguous runs of pages for batching, however,
3263 * causes the pages to actually be freed in smaller chunks. As there
3264 * can be a significant delay between the individual batches being
3265 * recycled, this leads to the once large chunks of space being
3266 * fragmented and becoming unavailable for high-order allocations.
3272 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3274 struct per_cpu_pages *pcp;
3277 memset(p, 0, sizeof(*p));
3281 pcp->high = 6 * batch;
3282 pcp->batch = max(1UL, 1 * batch);
3283 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3284 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3288 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3289 * to the value high for the pageset p.
3292 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3295 struct per_cpu_pages *pcp;
3299 pcp->batch = max(1UL, high/4);
3300 if ((high/4) > (PAGE_SHIFT * 8))
3301 pcp->batch = PAGE_SHIFT * 8;
3304 static __meminit void setup_zone_pageset(struct zone *zone)
3308 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3310 for_each_possible_cpu(cpu) {
3311 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3313 setup_pageset(pcp, zone_batchsize(zone));
3315 if (percpu_pagelist_fraction)
3316 setup_pagelist_highmark(pcp,
3317 (zone->present_pages /
3318 percpu_pagelist_fraction));
3323 * Allocate per cpu pagesets and initialize them.
3324 * Before this call only boot pagesets were available.
3326 void __init setup_per_cpu_pageset(void)
3330 for_each_populated_zone(zone)
3331 setup_zone_pageset(zone);
3334 static noinline __init_refok
3335 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3338 struct pglist_data *pgdat = zone->zone_pgdat;
3342 * The per-page waitqueue mechanism uses hashed waitqueues
3345 zone->wait_table_hash_nr_entries =
3346 wait_table_hash_nr_entries(zone_size_pages);
3347 zone->wait_table_bits =
3348 wait_table_bits(zone->wait_table_hash_nr_entries);
3349 alloc_size = zone->wait_table_hash_nr_entries
3350 * sizeof(wait_queue_head_t);
3352 if (!slab_is_available()) {
3353 zone->wait_table = (wait_queue_head_t *)
3354 alloc_bootmem_node(pgdat, alloc_size);
3357 * This case means that a zone whose size was 0 gets new memory
3358 * via memory hot-add.
3359 * But it may be the case that a new node was hot-added. In
3360 * this case vmalloc() will not be able to use this new node's
3361 * memory - this wait_table must be initialized to use this new
3362 * node itself as well.
3363 * To use this new node's memory, further consideration will be
3366 zone->wait_table = vmalloc(alloc_size);
3368 if (!zone->wait_table)
3371 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3372 init_waitqueue_head(zone->wait_table + i);
3377 static int __zone_pcp_update(void *data)
3379 struct zone *zone = data;
3381 unsigned long batch = zone_batchsize(zone), flags;
3383 for_each_possible_cpu(cpu) {
3384 struct per_cpu_pageset *pset;
3385 struct per_cpu_pages *pcp;
3387 pset = per_cpu_ptr(zone->pageset, cpu);
3390 local_irq_save(flags);
3391 free_pcppages_bulk(zone, pcp->count, pcp);
3392 setup_pageset(pset, batch);
3393 local_irq_restore(flags);
3398 void zone_pcp_update(struct zone *zone)
3400 stop_machine(__zone_pcp_update, zone, NULL);
3403 static __meminit void zone_pcp_init(struct zone *zone)
3406 * per cpu subsystem is not up at this point. The following code
3407 * relies on the ability of the linker to provide the
3408 * offset of a (static) per cpu variable into the per cpu area.
3410 zone->pageset = &boot_pageset;
3412 if (zone->present_pages)
3413 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3414 zone->name, zone->present_pages,
3415 zone_batchsize(zone));
3418 __meminit int init_currently_empty_zone(struct zone *zone,
3419 unsigned long zone_start_pfn,
3421 enum memmap_context context)
3423 struct pglist_data *pgdat = zone->zone_pgdat;
3425 ret = zone_wait_table_init(zone, size);
3428 pgdat->nr_zones = zone_idx(zone) + 1;
3430 zone->zone_start_pfn = zone_start_pfn;
3432 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3433 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3435 (unsigned long)zone_idx(zone),
3436 zone_start_pfn, (zone_start_pfn + size));
3438 zone_init_free_lists(zone);
3443 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3445 * Basic iterator support. Return the first range of PFNs for a node
3446 * Note: nid == MAX_NUMNODES returns first region regardless of node
3448 static int __meminit first_active_region_index_in_nid(int nid)
3452 for (i = 0; i < nr_nodemap_entries; i++)
3453 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3460 * Basic iterator support. Return the next active range of PFNs for a node
3461 * Note: nid == MAX_NUMNODES returns next region regardless of node
3463 static int __meminit next_active_region_index_in_nid(int index, int nid)
3465 for (index = index + 1; index < nr_nodemap_entries; index++)
3466 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3472 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3474 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3475 * Architectures may implement their own version but if add_active_range()
3476 * was used and there are no special requirements, this is a convenient
3479 int __meminit __early_pfn_to_nid(unsigned long pfn)
3483 for (i = 0; i < nr_nodemap_entries; i++) {
3484 unsigned long start_pfn = early_node_map[i].start_pfn;
3485 unsigned long end_pfn = early_node_map[i].end_pfn;
3487 if (start_pfn <= pfn && pfn < end_pfn)
3488 return early_node_map[i].nid;
3490 /* This is a memory hole */
3493 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3495 int __meminit early_pfn_to_nid(unsigned long pfn)
3499 nid = __early_pfn_to_nid(pfn);
3502 /* just returns 0 */
3506 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3507 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3511 nid = __early_pfn_to_nid(pfn);
3512 if (nid >= 0 && nid != node)
3518 /* Basic iterator support to walk early_node_map[] */
3519 #define for_each_active_range_index_in_nid(i, nid) \
3520 for (i = first_active_region_index_in_nid(nid); i != -1; \
3521 i = next_active_region_index_in_nid(i, nid))
3524 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3525 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3526 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3528 * If an architecture guarantees that all ranges registered with
3529 * add_active_ranges() contain no holes and may be freed, this
3530 * this function may be used instead of calling free_bootmem() manually.
3532 void __init free_bootmem_with_active_regions(int nid,
3533 unsigned long max_low_pfn)
3537 for_each_active_range_index_in_nid(i, nid) {
3538 unsigned long size_pages = 0;
3539 unsigned long end_pfn = early_node_map[i].end_pfn;
3541 if (early_node_map[i].start_pfn >= max_low_pfn)
3544 if (end_pfn > max_low_pfn)
3545 end_pfn = max_low_pfn;
3547 size_pages = end_pfn - early_node_map[i].start_pfn;
3548 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3549 PFN_PHYS(early_node_map[i].start_pfn),
3550 size_pages << PAGE_SHIFT);
3554 int __init add_from_early_node_map(struct range *range, int az,
3555 int nr_range, int nid)
3560 /* need to go over early_node_map to find out good range for node */
3561 for_each_active_range_index_in_nid(i, nid) {
3562 start = early_node_map[i].start_pfn;
3563 end = early_node_map[i].end_pfn;
3564 nr_range = add_range(range, az, nr_range, start, end);
3569 #ifdef CONFIG_NO_BOOTMEM
3570 void * __init __alloc_memory_core_early(int nid, u64 size, u64 align,
3571 u64 goal, u64 limit)
3576 /* need to go over early_node_map to find out good range for node */
3577 for_each_active_range_index_in_nid(i, nid) {
3579 u64 ei_start, ei_last;
3581 ei_last = early_node_map[i].end_pfn;
3582 ei_last <<= PAGE_SHIFT;
3583 ei_start = early_node_map[i].start_pfn;
3584 ei_start <<= PAGE_SHIFT;
3585 addr = find_early_area(ei_start, ei_last,
3586 goal, limit, size, align);
3592 printk(KERN_DEBUG "alloc (nid=%d %llx - %llx) (%llx - %llx) %llx %llx => %llx\n",
3594 ei_start, ei_last, goal, limit, size,
3598 ptr = phys_to_virt(addr);
3599 memset(ptr, 0, size);
3600 reserve_early_without_check(addr, addr + size, "BOOTMEM");
3609 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3614 for_each_active_range_index_in_nid(i, nid) {
3615 ret = work_fn(early_node_map[i].start_pfn,
3616 early_node_map[i].end_pfn, data);
3622 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3623 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3625 * If an architecture guarantees that all ranges registered with
3626 * add_active_ranges() contain no holes and may be freed, this
3627 * function may be used instead of calling memory_present() manually.
3629 void __init sparse_memory_present_with_active_regions(int nid)
3633 for_each_active_range_index_in_nid(i, nid)
3634 memory_present(early_node_map[i].nid,
3635 early_node_map[i].start_pfn,
3636 early_node_map[i].end_pfn);
3640 * get_pfn_range_for_nid - Return the start and end page frames for a node
3641 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3642 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3643 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3645 * It returns the start and end page frame of a node based on information
3646 * provided by an arch calling add_active_range(). If called for a node
3647 * with no available memory, a warning is printed and the start and end
3650 void __meminit get_pfn_range_for_nid(unsigned int nid,
3651 unsigned long *start_pfn, unsigned long *end_pfn)
3657 for_each_active_range_index_in_nid(i, nid) {
3658 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3659 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3662 if (*start_pfn == -1UL)
3667 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3668 * assumption is made that zones within a node are ordered in monotonic
3669 * increasing memory addresses so that the "highest" populated zone is used
3671 static void __init find_usable_zone_for_movable(void)
3674 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3675 if (zone_index == ZONE_MOVABLE)
3678 if (arch_zone_highest_possible_pfn[zone_index] >
3679 arch_zone_lowest_possible_pfn[zone_index])
3683 VM_BUG_ON(zone_index == -1);
3684 movable_zone = zone_index;
3688 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3689 * because it is sized independant of architecture. Unlike the other zones,
3690 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3691 * in each node depending on the size of each node and how evenly kernelcore
3692 * is distributed. This helper function adjusts the zone ranges
3693 * provided by the architecture for a given node by using the end of the
3694 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3695 * zones within a node are in order of monotonic increases memory addresses
3697 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3698 unsigned long zone_type,
3699 unsigned long node_start_pfn,
3700 unsigned long node_end_pfn,
3701 unsigned long *zone_start_pfn,
3702 unsigned long *zone_end_pfn)
3704 /* Only adjust if ZONE_MOVABLE is on this node */
3705 if (zone_movable_pfn[nid]) {
3706 /* Size ZONE_MOVABLE */
3707 if (zone_type == ZONE_MOVABLE) {
3708 *zone_start_pfn = zone_movable_pfn[nid];
3709 *zone_end_pfn = min(node_end_pfn,
3710 arch_zone_highest_possible_pfn[movable_zone]);
3712 /* Adjust for ZONE_MOVABLE starting within this range */
3713 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3714 *zone_end_pfn > zone_movable_pfn[nid]) {
3715 *zone_end_pfn = zone_movable_pfn[nid];
3717 /* Check if this whole range is within ZONE_MOVABLE */
3718 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3719 *zone_start_pfn = *zone_end_pfn;
3724 * Return the number of pages a zone spans in a node, including holes
3725 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3727 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3728 unsigned long zone_type,
3729 unsigned long *ignored)
3731 unsigned long node_start_pfn, node_end_pfn;
3732 unsigned long zone_start_pfn, zone_end_pfn;
3734 /* Get the start and end of the node and zone */
3735 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3736 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3737 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3738 adjust_zone_range_for_zone_movable(nid, zone_type,
3739 node_start_pfn, node_end_pfn,
3740 &zone_start_pfn, &zone_end_pfn);
3742 /* Check that this node has pages within the zone's required range */
3743 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3746 /* Move the zone boundaries inside the node if necessary */
3747 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3748 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3750 /* Return the spanned pages */
3751 return zone_end_pfn - zone_start_pfn;
3755 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3756 * then all holes in the requested range will be accounted for.
3758 unsigned long __meminit __absent_pages_in_range(int nid,
3759 unsigned long range_start_pfn,
3760 unsigned long range_end_pfn)
3763 unsigned long prev_end_pfn = 0, hole_pages = 0;
3764 unsigned long start_pfn;
3766 /* Find the end_pfn of the first active range of pfns in the node */
3767 i = first_active_region_index_in_nid(nid);
3771 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3773 /* Account for ranges before physical memory on this node */
3774 if (early_node_map[i].start_pfn > range_start_pfn)
3775 hole_pages = prev_end_pfn - range_start_pfn;
3777 /* Find all holes for the zone within the node */
3778 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3780 /* No need to continue if prev_end_pfn is outside the zone */
3781 if (prev_end_pfn >= range_end_pfn)
3784 /* Make sure the end of the zone is not within the hole */
3785 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3786 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3788 /* Update the hole size cound and move on */
3789 if (start_pfn > range_start_pfn) {
3790 BUG_ON(prev_end_pfn > start_pfn);
3791 hole_pages += start_pfn - prev_end_pfn;
3793 prev_end_pfn = early_node_map[i].end_pfn;
3796 /* Account for ranges past physical memory on this node */
3797 if (range_end_pfn > prev_end_pfn)
3798 hole_pages += range_end_pfn -
3799 max(range_start_pfn, prev_end_pfn);
3805 * absent_pages_in_range - Return number of page frames in holes within a range
3806 * @start_pfn: The start PFN to start searching for holes
3807 * @end_pfn: The end PFN to stop searching for holes
3809 * It returns the number of pages frames in memory holes within a range.
3811 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3812 unsigned long end_pfn)
3814 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3817 /* Return the number of page frames in holes in a zone on a node */
3818 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3819 unsigned long zone_type,
3820 unsigned long *ignored)
3822 unsigned long node_start_pfn, node_end_pfn;
3823 unsigned long zone_start_pfn, zone_end_pfn;
3825 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3826 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3828 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3831 adjust_zone_range_for_zone_movable(nid, zone_type,
3832 node_start_pfn, node_end_pfn,
3833 &zone_start_pfn, &zone_end_pfn);
3834 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3838 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3839 unsigned long zone_type,
3840 unsigned long *zones_size)
3842 return zones_size[zone_type];
3845 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3846 unsigned long zone_type,
3847 unsigned long *zholes_size)
3852 return zholes_size[zone_type];
3857 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3858 unsigned long *zones_size, unsigned long *zholes_size)
3860 unsigned long realtotalpages, totalpages = 0;
3863 for (i = 0; i < MAX_NR_ZONES; i++)
3864 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3866 pgdat->node_spanned_pages = totalpages;
3868 realtotalpages = totalpages;
3869 for (i = 0; i < MAX_NR_ZONES; i++)
3871 zone_absent_pages_in_node(pgdat->node_id, i,
3873 pgdat->node_present_pages = realtotalpages;
3874 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3878 #ifndef CONFIG_SPARSEMEM
3880 * Calculate the size of the zone->blockflags rounded to an unsigned long
3881 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3882 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3883 * round what is now in bits to nearest long in bits, then return it in
3886 static unsigned long __init usemap_size(unsigned long zonesize)
3888 unsigned long usemapsize;
3890 usemapsize = roundup(zonesize, pageblock_nr_pages);
3891 usemapsize = usemapsize >> pageblock_order;
3892 usemapsize *= NR_PAGEBLOCK_BITS;
3893 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
3895 return usemapsize / 8;
3898 static void __init setup_usemap(struct pglist_data *pgdat,
3899 struct zone *zone, unsigned long zonesize)
3901 unsigned long usemapsize = usemap_size(zonesize);
3902 zone->pageblock_flags = NULL;
3904 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
3907 static void inline setup_usemap(struct pglist_data *pgdat,
3908 struct zone *zone, unsigned long zonesize) {}
3909 #endif /* CONFIG_SPARSEMEM */
3911 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3913 /* Return a sensible default order for the pageblock size. */
3914 static inline int pageblock_default_order(void)
3916 if (HPAGE_SHIFT > PAGE_SHIFT)
3917 return HUGETLB_PAGE_ORDER;
3922 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3923 static inline void __init set_pageblock_order(unsigned int order)
3925 /* Check that pageblock_nr_pages has not already been setup */
3926 if (pageblock_order)
3930 * Assume the largest contiguous order of interest is a huge page.
3931 * This value may be variable depending on boot parameters on IA64
3933 pageblock_order = order;
3935 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3938 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3939 * and pageblock_default_order() are unused as pageblock_order is set
3940 * at compile-time. See include/linux/pageblock-flags.h for the values of
3941 * pageblock_order based on the kernel config
3943 static inline int pageblock_default_order(unsigned int order)
3947 #define set_pageblock_order(x) do {} while (0)
3949 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3952 * Set up the zone data structures:
3953 * - mark all pages reserved
3954 * - mark all memory queues empty
3955 * - clear the memory bitmaps
3957 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
3958 unsigned long *zones_size, unsigned long *zholes_size)
3961 int nid = pgdat->node_id;
3962 unsigned long zone_start_pfn = pgdat->node_start_pfn;
3965 pgdat_resize_init(pgdat);
3966 pgdat->nr_zones = 0;
3967 init_waitqueue_head(&pgdat->kswapd_wait);
3968 pgdat->kswapd_max_order = 0;
3969 pgdat_page_cgroup_init(pgdat);
3971 for (j = 0; j < MAX_NR_ZONES; j++) {
3972 struct zone *zone = pgdat->node_zones + j;
3973 unsigned long size, realsize, memmap_pages;
3976 size = zone_spanned_pages_in_node(nid, j, zones_size);
3977 realsize = size - zone_absent_pages_in_node(nid, j,
3981 * Adjust realsize so that it accounts for how much memory
3982 * is used by this zone for memmap. This affects the watermark
3983 * and per-cpu initialisations
3986 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
3987 if (realsize >= memmap_pages) {
3988 realsize -= memmap_pages;
3991 " %s zone: %lu pages used for memmap\n",
3992 zone_names[j], memmap_pages);
3995 " %s zone: %lu pages exceeds realsize %lu\n",
3996 zone_names[j], memmap_pages, realsize);
3998 /* Account for reserved pages */
3999 if (j == 0 && realsize > dma_reserve) {
4000 realsize -= dma_reserve;
4001 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4002 zone_names[0], dma_reserve);
4005 if (!is_highmem_idx(j))
4006 nr_kernel_pages += realsize;
4007 nr_all_pages += realsize;
4009 zone->spanned_pages = size;
4010 zone->present_pages = realsize;
4013 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4015 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4017 zone->name = zone_names[j];
4018 spin_lock_init(&zone->lock);
4019 spin_lock_init(&zone->lru_lock);
4020 zone_seqlock_init(zone);
4021 zone->zone_pgdat = pgdat;
4023 zone->prev_priority = DEF_PRIORITY;
4025 zone_pcp_init(zone);
4027 INIT_LIST_HEAD(&zone->lru[l].list);
4028 zone->reclaim_stat.nr_saved_scan[l] = 0;
4030 zone->reclaim_stat.recent_rotated[0] = 0;
4031 zone->reclaim_stat.recent_rotated[1] = 0;
4032 zone->reclaim_stat.recent_scanned[0] = 0;
4033 zone->reclaim_stat.recent_scanned[1] = 0;
4034 zap_zone_vm_stats(zone);
4039 set_pageblock_order(pageblock_default_order());
4040 setup_usemap(pgdat, zone, size);
4041 ret = init_currently_empty_zone(zone, zone_start_pfn,
4042 size, MEMMAP_EARLY);
4044 memmap_init(size, nid, j, zone_start_pfn);
4045 zone_start_pfn += size;
4049 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4051 /* Skip empty nodes */
4052 if (!pgdat->node_spanned_pages)
4055 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4056 /* ia64 gets its own node_mem_map, before this, without bootmem */
4057 if (!pgdat->node_mem_map) {
4058 unsigned long size, start, end;
4062 * The zone's endpoints aren't required to be MAX_ORDER
4063 * aligned but the node_mem_map endpoints must be in order
4064 * for the buddy allocator to function correctly.
4066 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4067 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4068 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4069 size = (end - start) * sizeof(struct page);
4070 map = alloc_remap(pgdat->node_id, size);
4072 map = alloc_bootmem_node(pgdat, size);
4073 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4075 #ifndef CONFIG_NEED_MULTIPLE_NODES
4077 * With no DISCONTIG, the global mem_map is just set as node 0's
4079 if (pgdat == NODE_DATA(0)) {
4080 mem_map = NODE_DATA(0)->node_mem_map;
4081 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4082 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4083 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4084 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4087 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4090 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4091 unsigned long node_start_pfn, unsigned long *zholes_size)
4093 pg_data_t *pgdat = NODE_DATA(nid);
4095 pgdat->node_id = nid;
4096 pgdat->node_start_pfn = node_start_pfn;
4097 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4099 alloc_node_mem_map(pgdat);
4100 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4101 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4102 nid, (unsigned long)pgdat,
4103 (unsigned long)pgdat->node_mem_map);
4106 free_area_init_core(pgdat, zones_size, zholes_size);
4109 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4111 #if MAX_NUMNODES > 1
4113 * Figure out the number of possible node ids.
4115 static void __init setup_nr_node_ids(void)
4118 unsigned int highest = 0;
4120 for_each_node_mask(node, node_possible_map)
4122 nr_node_ids = highest + 1;
4125 static inline void setup_nr_node_ids(void)
4131 * add_active_range - Register a range of PFNs backed by physical memory
4132 * @nid: The node ID the range resides on
4133 * @start_pfn: The start PFN of the available physical memory
4134 * @end_pfn: The end PFN of the available physical memory
4136 * These ranges are stored in an early_node_map[] and later used by
4137 * free_area_init_nodes() to calculate zone sizes and holes. If the
4138 * range spans a memory hole, it is up to the architecture to ensure
4139 * the memory is not freed by the bootmem allocator. If possible
4140 * the range being registered will be merged with existing ranges.
4142 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4143 unsigned long end_pfn)
4147 mminit_dprintk(MMINIT_TRACE, "memory_register",
4148 "Entering add_active_range(%d, %#lx, %#lx) "
4149 "%d entries of %d used\n",
4150 nid, start_pfn, end_pfn,
4151 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4153 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4155 /* Merge with existing active regions if possible */
4156 for (i = 0; i < nr_nodemap_entries; i++) {
4157 if (early_node_map[i].nid != nid)
4160 /* Skip if an existing region covers this new one */
4161 if (start_pfn >= early_node_map[i].start_pfn &&
4162 end_pfn <= early_node_map[i].end_pfn)
4165 /* Merge forward if suitable */
4166 if (start_pfn <= early_node_map[i].end_pfn &&
4167 end_pfn > early_node_map[i].end_pfn) {
4168 early_node_map[i].end_pfn = end_pfn;
4172 /* Merge backward if suitable */
4173 if (start_pfn < early_node_map[i].start_pfn &&
4174 end_pfn >= early_node_map[i].start_pfn) {
4175 early_node_map[i].start_pfn = start_pfn;
4180 /* Check that early_node_map is large enough */
4181 if (i >= MAX_ACTIVE_REGIONS) {
4182 printk(KERN_CRIT "More than %d memory regions, truncating\n",
4183 MAX_ACTIVE_REGIONS);
4187 early_node_map[i].nid = nid;
4188 early_node_map[i].start_pfn = start_pfn;
4189 early_node_map[i].end_pfn = end_pfn;
4190 nr_nodemap_entries = i + 1;
4194 * remove_active_range - Shrink an existing registered range of PFNs
4195 * @nid: The node id the range is on that should be shrunk
4196 * @start_pfn: The new PFN of the range
4197 * @end_pfn: The new PFN of the range
4199 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4200 * The map is kept near the end physical page range that has already been
4201 * registered. This function allows an arch to shrink an existing registered
4204 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4205 unsigned long end_pfn)
4210 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4211 nid, start_pfn, end_pfn);
4213 /* Find the old active region end and shrink */
4214 for_each_active_range_index_in_nid(i, nid) {
4215 if (early_node_map[i].start_pfn >= start_pfn &&
4216 early_node_map[i].end_pfn <= end_pfn) {
4218 early_node_map[i].start_pfn = 0;
4219 early_node_map[i].end_pfn = 0;
4223 if (early_node_map[i].start_pfn < start_pfn &&
4224 early_node_map[i].end_pfn > start_pfn) {
4225 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4226 early_node_map[i].end_pfn = start_pfn;
4227 if (temp_end_pfn > end_pfn)
4228 add_active_range(nid, end_pfn, temp_end_pfn);
4231 if (early_node_map[i].start_pfn >= start_pfn &&
4232 early_node_map[i].end_pfn > end_pfn &&
4233 early_node_map[i].start_pfn < end_pfn) {
4234 early_node_map[i].start_pfn = end_pfn;
4242 /* remove the blank ones */
4243 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4244 if (early_node_map[i].nid != nid)
4246 if (early_node_map[i].end_pfn)
4248 /* we found it, get rid of it */
4249 for (j = i; j < nr_nodemap_entries - 1; j++)
4250 memcpy(&early_node_map[j], &early_node_map[j+1],
4251 sizeof(early_node_map[j]));
4252 j = nr_nodemap_entries - 1;
4253 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4254 nr_nodemap_entries--;
4259 * remove_all_active_ranges - Remove all currently registered regions
4261 * During discovery, it may be found that a table like SRAT is invalid
4262 * and an alternative discovery method must be used. This function removes
4263 * all currently registered regions.
4265 void __init remove_all_active_ranges(void)
4267 memset(early_node_map, 0, sizeof(early_node_map));
4268 nr_nodemap_entries = 0;
4271 /* Compare two active node_active_regions */
4272 static int __init cmp_node_active_region(const void *a, const void *b)
4274 struct node_active_region *arange = (struct node_active_region *)a;
4275 struct node_active_region *brange = (struct node_active_region *)b;
4277 /* Done this way to avoid overflows */
4278 if (arange->start_pfn > brange->start_pfn)
4280 if (arange->start_pfn < brange->start_pfn)
4286 /* sort the node_map by start_pfn */
4287 void __init sort_node_map(void)
4289 sort(early_node_map, (size_t)nr_nodemap_entries,
4290 sizeof(struct node_active_region),
4291 cmp_node_active_region, NULL);
4294 /* Find the lowest pfn for a node */
4295 static unsigned long __init find_min_pfn_for_node(int nid)
4298 unsigned long min_pfn = ULONG_MAX;
4300 /* Assuming a sorted map, the first range found has the starting pfn */
4301 for_each_active_range_index_in_nid(i, nid)
4302 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4304 if (min_pfn == ULONG_MAX) {
4306 "Could not find start_pfn for node %d\n", nid);
4314 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4316 * It returns the minimum PFN based on information provided via
4317 * add_active_range().
4319 unsigned long __init find_min_pfn_with_active_regions(void)
4321 return find_min_pfn_for_node(MAX_NUMNODES);
4325 * early_calculate_totalpages()
4326 * Sum pages in active regions for movable zone.
4327 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4329 static unsigned long __init early_calculate_totalpages(void)
4332 unsigned long totalpages = 0;
4334 for (i = 0; i < nr_nodemap_entries; i++) {
4335 unsigned long pages = early_node_map[i].end_pfn -
4336 early_node_map[i].start_pfn;
4337 totalpages += pages;
4339 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4345 * Find the PFN the Movable zone begins in each node. Kernel memory
4346 * is spread evenly between nodes as long as the nodes have enough
4347 * memory. When they don't, some nodes will have more kernelcore than
4350 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4353 unsigned long usable_startpfn;
4354 unsigned long kernelcore_node, kernelcore_remaining;
4355 /* save the state before borrow the nodemask */
4356 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4357 unsigned long totalpages = early_calculate_totalpages();
4358 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4361 * If movablecore was specified, calculate what size of
4362 * kernelcore that corresponds so that memory usable for
4363 * any allocation type is evenly spread. If both kernelcore
4364 * and movablecore are specified, then the value of kernelcore
4365 * will be used for required_kernelcore if it's greater than
4366 * what movablecore would have allowed.
4368 if (required_movablecore) {
4369 unsigned long corepages;
4372 * Round-up so that ZONE_MOVABLE is at least as large as what
4373 * was requested by the user
4375 required_movablecore =
4376 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4377 corepages = totalpages - required_movablecore;
4379 required_kernelcore = max(required_kernelcore, corepages);
4382 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4383 if (!required_kernelcore)
4386 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4387 find_usable_zone_for_movable();
4388 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4391 /* Spread kernelcore memory as evenly as possible throughout nodes */
4392 kernelcore_node = required_kernelcore / usable_nodes;
4393 for_each_node_state(nid, N_HIGH_MEMORY) {
4395 * Recalculate kernelcore_node if the division per node
4396 * now exceeds what is necessary to satisfy the requested
4397 * amount of memory for the kernel
4399 if (required_kernelcore < kernelcore_node)
4400 kernelcore_node = required_kernelcore / usable_nodes;
4403 * As the map is walked, we track how much memory is usable
4404 * by the kernel using kernelcore_remaining. When it is
4405 * 0, the rest of the node is usable by ZONE_MOVABLE
4407 kernelcore_remaining = kernelcore_node;
4409 /* Go through each range of PFNs within this node */
4410 for_each_active_range_index_in_nid(i, nid) {
4411 unsigned long start_pfn, end_pfn;
4412 unsigned long size_pages;
4414 start_pfn = max(early_node_map[i].start_pfn,
4415 zone_movable_pfn[nid]);
4416 end_pfn = early_node_map[i].end_pfn;
4417 if (start_pfn >= end_pfn)
4420 /* Account for what is only usable for kernelcore */
4421 if (start_pfn < usable_startpfn) {
4422 unsigned long kernel_pages;
4423 kernel_pages = min(end_pfn, usable_startpfn)
4426 kernelcore_remaining -= min(kernel_pages,
4427 kernelcore_remaining);
4428 required_kernelcore -= min(kernel_pages,
4429 required_kernelcore);
4431 /* Continue if range is now fully accounted */
4432 if (end_pfn <= usable_startpfn) {
4435 * Push zone_movable_pfn to the end so
4436 * that if we have to rebalance
4437 * kernelcore across nodes, we will
4438 * not double account here
4440 zone_movable_pfn[nid] = end_pfn;
4443 start_pfn = usable_startpfn;
4447 * The usable PFN range for ZONE_MOVABLE is from
4448 * start_pfn->end_pfn. Calculate size_pages as the
4449 * number of pages used as kernelcore
4451 size_pages = end_pfn - start_pfn;
4452 if (size_pages > kernelcore_remaining)
4453 size_pages = kernelcore_remaining;
4454 zone_movable_pfn[nid] = start_pfn + size_pages;
4457 * Some kernelcore has been met, update counts and
4458 * break if the kernelcore for this node has been
4461 required_kernelcore -= min(required_kernelcore,
4463 kernelcore_remaining -= size_pages;
4464 if (!kernelcore_remaining)
4470 * If there is still required_kernelcore, we do another pass with one
4471 * less node in the count. This will push zone_movable_pfn[nid] further
4472 * along on the nodes that still have memory until kernelcore is
4476 if (usable_nodes && required_kernelcore > usable_nodes)
4479 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4480 for (nid = 0; nid < MAX_NUMNODES; nid++)
4481 zone_movable_pfn[nid] =
4482 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4485 /* restore the node_state */
4486 node_states[N_HIGH_MEMORY] = saved_node_state;
4489 /* Any regular memory on that node ? */
4490 static void check_for_regular_memory(pg_data_t *pgdat)
4492 #ifdef CONFIG_HIGHMEM
4493 enum zone_type zone_type;
4495 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4496 struct zone *zone = &pgdat->node_zones[zone_type];
4497 if (zone->present_pages)
4498 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4504 * free_area_init_nodes - Initialise all pg_data_t and zone data
4505 * @max_zone_pfn: an array of max PFNs for each zone
4507 * This will call free_area_init_node() for each active node in the system.
4508 * Using the page ranges provided by add_active_range(), the size of each
4509 * zone in each node and their holes is calculated. If the maximum PFN
4510 * between two adjacent zones match, it is assumed that the zone is empty.
4511 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4512 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4513 * starts where the previous one ended. For example, ZONE_DMA32 starts
4514 * at arch_max_dma_pfn.
4516 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4521 /* Sort early_node_map as initialisation assumes it is sorted */
4524 /* Record where the zone boundaries are */
4525 memset(arch_zone_lowest_possible_pfn, 0,
4526 sizeof(arch_zone_lowest_possible_pfn));
4527 memset(arch_zone_highest_possible_pfn, 0,
4528 sizeof(arch_zone_highest_possible_pfn));
4529 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4530 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4531 for (i = 1; i < MAX_NR_ZONES; i++) {
4532 if (i == ZONE_MOVABLE)
4534 arch_zone_lowest_possible_pfn[i] =
4535 arch_zone_highest_possible_pfn[i-1];
4536 arch_zone_highest_possible_pfn[i] =
4537 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4539 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4540 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4542 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4543 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4544 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4546 /* Print out the zone ranges */
4547 printk("Zone PFN ranges:\n");
4548 for (i = 0; i < MAX_NR_ZONES; i++) {
4549 if (i == ZONE_MOVABLE)
4551 printk(" %-8s ", zone_names[i]);
4552 if (arch_zone_lowest_possible_pfn[i] ==
4553 arch_zone_highest_possible_pfn[i])
4556 printk("%0#10lx -> %0#10lx\n",
4557 arch_zone_lowest_possible_pfn[i],
4558 arch_zone_highest_possible_pfn[i]);
4561 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4562 printk("Movable zone start PFN for each node\n");
4563 for (i = 0; i < MAX_NUMNODES; i++) {
4564 if (zone_movable_pfn[i])
4565 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4568 /* Print out the early_node_map[] */
4569 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4570 for (i = 0; i < nr_nodemap_entries; i++)
4571 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4572 early_node_map[i].start_pfn,
4573 early_node_map[i].end_pfn);
4575 /* Initialise every node */
4576 mminit_verify_pageflags_layout();
4577 setup_nr_node_ids();
4578 for_each_online_node(nid) {
4579 pg_data_t *pgdat = NODE_DATA(nid);
4580 free_area_init_node(nid, NULL,
4581 find_min_pfn_for_node(nid), NULL);
4583 /* Any memory on that node */
4584 if (pgdat->node_present_pages)
4585 node_set_state(nid, N_HIGH_MEMORY);
4586 check_for_regular_memory(pgdat);
4590 static int __init cmdline_parse_core(char *p, unsigned long *core)
4592 unsigned long long coremem;
4596 coremem = memparse(p, &p);
4597 *core = coremem >> PAGE_SHIFT;
4599 /* Paranoid check that UL is enough for the coremem value */
4600 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4606 * kernelcore=size sets the amount of memory for use for allocations that
4607 * cannot be reclaimed or migrated.
4609 static int __init cmdline_parse_kernelcore(char *p)
4611 return cmdline_parse_core(p, &required_kernelcore);
4615 * movablecore=size sets the amount of memory for use for allocations that
4616 * can be reclaimed or migrated.
4618 static int __init cmdline_parse_movablecore(char *p)
4620 return cmdline_parse_core(p, &required_movablecore);
4623 early_param("kernelcore", cmdline_parse_kernelcore);
4624 early_param("movablecore", cmdline_parse_movablecore);
4626 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4629 * set_dma_reserve - set the specified number of pages reserved in the first zone
4630 * @new_dma_reserve: The number of pages to mark reserved
4632 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4633 * In the DMA zone, a significant percentage may be consumed by kernel image
4634 * and other unfreeable allocations which can skew the watermarks badly. This
4635 * function may optionally be used to account for unfreeable pages in the
4636 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4637 * smaller per-cpu batchsize.
4639 void __init set_dma_reserve(unsigned long new_dma_reserve)
4641 dma_reserve = new_dma_reserve;
4644 #ifndef CONFIG_NEED_MULTIPLE_NODES
4645 struct pglist_data __refdata contig_page_data = {
4646 #ifndef CONFIG_NO_BOOTMEM
4647 .bdata = &bootmem_node_data[0]
4650 EXPORT_SYMBOL(contig_page_data);
4653 void __init free_area_init(unsigned long *zones_size)
4655 free_area_init_node(0, zones_size,
4656 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4659 static int page_alloc_cpu_notify(struct notifier_block *self,
4660 unsigned long action, void *hcpu)
4662 int cpu = (unsigned long)hcpu;
4664 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4668 * Spill the event counters of the dead processor
4669 * into the current processors event counters.
4670 * This artificially elevates the count of the current
4673 vm_events_fold_cpu(cpu);
4676 * Zero the differential counters of the dead processor
4677 * so that the vm statistics are consistent.
4679 * This is only okay since the processor is dead and cannot
4680 * race with what we are doing.
4682 refresh_cpu_vm_stats(cpu);
4687 void __init page_alloc_init(void)
4689 hotcpu_notifier(page_alloc_cpu_notify, 0);
4693 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4694 * or min_free_kbytes changes.
4696 static void calculate_totalreserve_pages(void)
4698 struct pglist_data *pgdat;
4699 unsigned long reserve_pages = 0;
4700 enum zone_type i, j;
4702 for_each_online_pgdat(pgdat) {
4703 for (i = 0; i < MAX_NR_ZONES; i++) {
4704 struct zone *zone = pgdat->node_zones + i;
4705 unsigned long max = 0;
4707 /* Find valid and maximum lowmem_reserve in the zone */
4708 for (j = i; j < MAX_NR_ZONES; j++) {
4709 if (zone->lowmem_reserve[j] > max)
4710 max = zone->lowmem_reserve[j];
4713 /* we treat the high watermark as reserved pages. */
4714 max += high_wmark_pages(zone);
4716 if (max > zone->present_pages)
4717 max = zone->present_pages;
4718 reserve_pages += max;
4721 totalreserve_pages = reserve_pages;
4725 * setup_per_zone_lowmem_reserve - called whenever
4726 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4727 * has a correct pages reserved value, so an adequate number of
4728 * pages are left in the zone after a successful __alloc_pages().
4730 static void setup_per_zone_lowmem_reserve(void)
4732 struct pglist_data *pgdat;
4733 enum zone_type j, idx;
4735 for_each_online_pgdat(pgdat) {
4736 for (j = 0; j < MAX_NR_ZONES; j++) {
4737 struct zone *zone = pgdat->node_zones + j;
4738 unsigned long present_pages = zone->present_pages;
4740 zone->lowmem_reserve[j] = 0;
4744 struct zone *lower_zone;
4748 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4749 sysctl_lowmem_reserve_ratio[idx] = 1;
4751 lower_zone = pgdat->node_zones + idx;
4752 lower_zone->lowmem_reserve[j] = present_pages /
4753 sysctl_lowmem_reserve_ratio[idx];
4754 present_pages += lower_zone->present_pages;
4759 /* update totalreserve_pages */
4760 calculate_totalreserve_pages();
4764 * setup_per_zone_wmarks - called when min_free_kbytes changes
4765 * or when memory is hot-{added|removed}
4767 * Ensures that the watermark[min,low,high] values for each zone are set
4768 * correctly with respect to min_free_kbytes.
4770 void setup_per_zone_wmarks(void)
4772 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4773 unsigned long lowmem_pages = 0;
4775 unsigned long flags;
4777 /* Calculate total number of !ZONE_HIGHMEM pages */
4778 for_each_zone(zone) {
4779 if (!is_highmem(zone))
4780 lowmem_pages += zone->present_pages;
4783 for_each_zone(zone) {
4786 spin_lock_irqsave(&zone->lock, flags);
4787 tmp = (u64)pages_min * zone->present_pages;
4788 do_div(tmp, lowmem_pages);
4789 if (is_highmem(zone)) {
4791 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4792 * need highmem pages, so cap pages_min to a small
4795 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4796 * deltas controls asynch page reclaim, and so should
4797 * not be capped for highmem.
4801 min_pages = zone->present_pages / 1024;
4802 if (min_pages < SWAP_CLUSTER_MAX)
4803 min_pages = SWAP_CLUSTER_MAX;
4804 if (min_pages > 128)
4806 zone->watermark[WMARK_MIN] = min_pages;
4809 * If it's a lowmem zone, reserve a number of pages
4810 * proportionate to the zone's size.
4812 zone->watermark[WMARK_MIN] = tmp;
4815 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
4816 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
4817 setup_zone_migrate_reserve(zone);
4818 spin_unlock_irqrestore(&zone->lock, flags);
4821 /* update totalreserve_pages */
4822 calculate_totalreserve_pages();
4826 * The inactive anon list should be small enough that the VM never has to
4827 * do too much work, but large enough that each inactive page has a chance
4828 * to be referenced again before it is swapped out.
4830 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4831 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4832 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4833 * the anonymous pages are kept on the inactive list.
4836 * memory ratio inactive anon
4837 * -------------------------------------
4846 void calculate_zone_inactive_ratio(struct zone *zone)
4848 unsigned int gb, ratio;
4850 /* Zone size in gigabytes */
4851 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4853 ratio = int_sqrt(10 * gb);
4857 zone->inactive_ratio = ratio;
4860 static void __init setup_per_zone_inactive_ratio(void)
4865 calculate_zone_inactive_ratio(zone);
4869 * Initialise min_free_kbytes.
4871 * For small machines we want it small (128k min). For large machines
4872 * we want it large (64MB max). But it is not linear, because network
4873 * bandwidth does not increase linearly with machine size. We use
4875 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4876 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4892 static int __init init_per_zone_wmark_min(void)
4894 unsigned long lowmem_kbytes;
4896 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4898 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4899 if (min_free_kbytes < 128)
4900 min_free_kbytes = 128;
4901 if (min_free_kbytes > 65536)
4902 min_free_kbytes = 65536;
4903 setup_per_zone_wmarks();
4904 setup_per_zone_lowmem_reserve();
4905 setup_per_zone_inactive_ratio();
4908 module_init(init_per_zone_wmark_min)
4911 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4912 * that we can call two helper functions whenever min_free_kbytes
4915 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
4916 void __user *buffer, size_t *length, loff_t *ppos)
4918 proc_dointvec(table, write, buffer, length, ppos);
4920 setup_per_zone_wmarks();
4925 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
4926 void __user *buffer, size_t *length, loff_t *ppos)
4931 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
4936 zone->min_unmapped_pages = (zone->present_pages *
4937 sysctl_min_unmapped_ratio) / 100;
4941 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
4942 void __user *buffer, size_t *length, loff_t *ppos)
4947 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
4952 zone->min_slab_pages = (zone->present_pages *
4953 sysctl_min_slab_ratio) / 100;
4959 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4960 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4961 * whenever sysctl_lowmem_reserve_ratio changes.
4963 * The reserve ratio obviously has absolutely no relation with the
4964 * minimum watermarks. The lowmem reserve ratio can only make sense
4965 * if in function of the boot time zone sizes.
4967 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
4968 void __user *buffer, size_t *length, loff_t *ppos)
4970 proc_dointvec_minmax(table, write, buffer, length, ppos);
4971 setup_per_zone_lowmem_reserve();
4976 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4977 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4978 * can have before it gets flushed back to buddy allocator.
4981 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
4982 void __user *buffer, size_t *length, loff_t *ppos)
4988 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
4989 if (!write || (ret == -EINVAL))
4991 for_each_populated_zone(zone) {
4992 for_each_possible_cpu(cpu) {
4994 high = zone->present_pages / percpu_pagelist_fraction;
4995 setup_pagelist_highmark(
4996 per_cpu_ptr(zone->pageset, cpu), high);
5002 int hashdist = HASHDIST_DEFAULT;
5005 static int __init set_hashdist(char *str)
5009 hashdist = simple_strtoul(str, &str, 0);
5012 __setup("hashdist=", set_hashdist);
5016 * allocate a large system hash table from bootmem
5017 * - it is assumed that the hash table must contain an exact power-of-2
5018 * quantity of entries
5019 * - limit is the number of hash buckets, not the total allocation size
5021 void *__init alloc_large_system_hash(const char *tablename,
5022 unsigned long bucketsize,
5023 unsigned long numentries,
5026 unsigned int *_hash_shift,
5027 unsigned int *_hash_mask,
5028 unsigned long limit)
5030 unsigned long long max = limit;
5031 unsigned long log2qty, size;
5034 /* allow the kernel cmdline to have a say */
5036 /* round applicable memory size up to nearest megabyte */
5037 numentries = nr_kernel_pages;
5038 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5039 numentries >>= 20 - PAGE_SHIFT;
5040 numentries <<= 20 - PAGE_SHIFT;
5042 /* limit to 1 bucket per 2^scale bytes of low memory */
5043 if (scale > PAGE_SHIFT)
5044 numentries >>= (scale - PAGE_SHIFT);
5046 numentries <<= (PAGE_SHIFT - scale);
5048 /* Make sure we've got at least a 0-order allocation.. */
5049 if (unlikely(flags & HASH_SMALL)) {
5050 /* Makes no sense without HASH_EARLY */
5051 WARN_ON(!(flags & HASH_EARLY));
5052 if (!(numentries >> *_hash_shift)) {
5053 numentries = 1UL << *_hash_shift;
5054 BUG_ON(!numentries);
5056 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5057 numentries = PAGE_SIZE / bucketsize;
5059 numentries = roundup_pow_of_two(numentries);
5061 /* limit allocation size to 1/16 total memory by default */
5063 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5064 do_div(max, bucketsize);
5067 if (numentries > max)
5070 log2qty = ilog2(numentries);
5073 size = bucketsize << log2qty;
5074 if (flags & HASH_EARLY)
5075 table = alloc_bootmem_nopanic(size);
5077 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5080 * If bucketsize is not a power-of-two, we may free
5081 * some pages at the end of hash table which
5082 * alloc_pages_exact() automatically does
5084 if (get_order(size) < MAX_ORDER) {
5085 table = alloc_pages_exact(size, GFP_ATOMIC);
5086 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5089 } while (!table && size > PAGE_SIZE && --log2qty);
5092 panic("Failed to allocate %s hash table\n", tablename);
5094 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
5097 ilog2(size) - PAGE_SHIFT,
5101 *_hash_shift = log2qty;
5103 *_hash_mask = (1 << log2qty) - 1;
5108 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5109 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5112 #ifdef CONFIG_SPARSEMEM
5113 return __pfn_to_section(pfn)->pageblock_flags;
5115 return zone->pageblock_flags;
5116 #endif /* CONFIG_SPARSEMEM */
5119 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5121 #ifdef CONFIG_SPARSEMEM
5122 pfn &= (PAGES_PER_SECTION-1);
5123 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5125 pfn = pfn - zone->zone_start_pfn;
5126 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5127 #endif /* CONFIG_SPARSEMEM */
5131 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5132 * @page: The page within the block of interest
5133 * @start_bitidx: The first bit of interest to retrieve
5134 * @end_bitidx: The last bit of interest
5135 * returns pageblock_bits flags
5137 unsigned long get_pageblock_flags_group(struct page *page,
5138 int start_bitidx, int end_bitidx)
5141 unsigned long *bitmap;
5142 unsigned long pfn, bitidx;
5143 unsigned long flags = 0;
5144 unsigned long value = 1;
5146 zone = page_zone(page);
5147 pfn = page_to_pfn(page);
5148 bitmap = get_pageblock_bitmap(zone, pfn);
5149 bitidx = pfn_to_bitidx(zone, pfn);
5151 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5152 if (test_bit(bitidx + start_bitidx, bitmap))
5159 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5160 * @page: The page within the block of interest
5161 * @start_bitidx: The first bit of interest
5162 * @end_bitidx: The last bit of interest
5163 * @flags: The flags to set
5165 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5166 int start_bitidx, int end_bitidx)
5169 unsigned long *bitmap;
5170 unsigned long pfn, bitidx;
5171 unsigned long value = 1;
5173 zone = page_zone(page);
5174 pfn = page_to_pfn(page);
5175 bitmap = get_pageblock_bitmap(zone, pfn);
5176 bitidx = pfn_to_bitidx(zone, pfn);
5177 VM_BUG_ON(pfn < zone->zone_start_pfn);
5178 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5180 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5182 __set_bit(bitidx + start_bitidx, bitmap);
5184 __clear_bit(bitidx + start_bitidx, bitmap);
5188 * This is designed as sub function...plz see page_isolation.c also.
5189 * set/clear page block's type to be ISOLATE.
5190 * page allocater never alloc memory from ISOLATE block.
5193 int set_migratetype_isolate(struct page *page)
5196 struct page *curr_page;
5197 unsigned long flags, pfn, iter;
5198 unsigned long immobile = 0;
5199 struct memory_isolate_notify arg;
5204 zone = page_zone(page);
5205 zone_idx = zone_idx(zone);
5207 spin_lock_irqsave(&zone->lock, flags);
5208 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE ||
5209 zone_idx == ZONE_MOVABLE) {
5214 pfn = page_to_pfn(page);
5215 arg.start_pfn = pfn;
5216 arg.nr_pages = pageblock_nr_pages;
5217 arg.pages_found = 0;
5220 * It may be possible to isolate a pageblock even if the
5221 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5222 * notifier chain is used by balloon drivers to return the
5223 * number of pages in a range that are held by the balloon
5224 * driver to shrink memory. If all the pages are accounted for
5225 * by balloons, are free, or on the LRU, isolation can continue.
5226 * Later, for example, when memory hotplug notifier runs, these
5227 * pages reported as "can be isolated" should be isolated(freed)
5228 * by the balloon driver through the memory notifier chain.
5230 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5231 notifier_ret = notifier_to_errno(notifier_ret);
5232 if (notifier_ret || !arg.pages_found)
5235 for (iter = pfn; iter < (pfn + pageblock_nr_pages); iter++) {
5236 if (!pfn_valid_within(pfn))
5239 curr_page = pfn_to_page(iter);
5240 if (!page_count(curr_page) || PageLRU(curr_page))
5246 if (arg.pages_found == immobile)
5251 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5252 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5255 spin_unlock_irqrestore(&zone->lock, flags);
5261 void unset_migratetype_isolate(struct page *page)
5264 unsigned long flags;
5265 zone = page_zone(page);
5266 spin_lock_irqsave(&zone->lock, flags);
5267 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5269 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5270 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5272 spin_unlock_irqrestore(&zone->lock, flags);
5275 #ifdef CONFIG_MEMORY_HOTREMOVE
5277 * All pages in the range must be isolated before calling this.
5280 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5286 unsigned long flags;
5287 /* find the first valid pfn */
5288 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5293 zone = page_zone(pfn_to_page(pfn));
5294 spin_lock_irqsave(&zone->lock, flags);
5296 while (pfn < end_pfn) {
5297 if (!pfn_valid(pfn)) {
5301 page = pfn_to_page(pfn);
5302 BUG_ON(page_count(page));
5303 BUG_ON(!PageBuddy(page));
5304 order = page_order(page);
5305 #ifdef CONFIG_DEBUG_VM
5306 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5307 pfn, 1 << order, end_pfn);
5309 list_del(&page->lru);
5310 rmv_page_order(page);
5311 zone->free_area[order].nr_free--;
5312 __mod_zone_page_state(zone, NR_FREE_PAGES,
5314 for (i = 0; i < (1 << order); i++)
5315 SetPageReserved((page+i));
5316 pfn += (1 << order);
5318 spin_unlock_irqrestore(&zone->lock, flags);
5322 #ifdef CONFIG_MEMORY_FAILURE
5323 bool is_free_buddy_page(struct page *page)
5325 struct zone *zone = page_zone(page);
5326 unsigned long pfn = page_to_pfn(page);
5327 unsigned long flags;
5330 spin_lock_irqsave(&zone->lock, flags);
5331 for (order = 0; order < MAX_ORDER; order++) {
5332 struct page *page_head = page - (pfn & ((1 << order) - 1));
5334 if (PageBuddy(page_head) && page_order(page_head) >= order)
5337 spin_unlock_irqrestore(&zone->lock, flags);
5339 return order < MAX_ORDER;
5343 static struct trace_print_flags pageflag_names[] = {
5344 {1UL << PG_locked, "locked" },
5345 {1UL << PG_error, "error" },
5346 {1UL << PG_referenced, "referenced" },
5347 {1UL << PG_uptodate, "uptodate" },
5348 {1UL << PG_dirty, "dirty" },
5349 {1UL << PG_lru, "lru" },
5350 {1UL << PG_active, "active" },
5351 {1UL << PG_slab, "slab" },
5352 {1UL << PG_owner_priv_1, "owner_priv_1" },
5353 {1UL << PG_arch_1, "arch_1" },
5354 {1UL << PG_reserved, "reserved" },
5355 {1UL << PG_private, "private" },
5356 {1UL << PG_private_2, "private_2" },
5357 {1UL << PG_writeback, "writeback" },
5358 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5359 {1UL << PG_head, "head" },
5360 {1UL << PG_tail, "tail" },
5362 {1UL << PG_compound, "compound" },
5364 {1UL << PG_swapcache, "swapcache" },
5365 {1UL << PG_mappedtodisk, "mappedtodisk" },
5366 {1UL << PG_reclaim, "reclaim" },
5367 {1UL << PG_buddy, "buddy" },
5368 {1UL << PG_swapbacked, "swapbacked" },
5369 {1UL << PG_unevictable, "unevictable" },
5371 {1UL << PG_mlocked, "mlocked" },
5373 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5374 {1UL << PG_uncached, "uncached" },
5376 #ifdef CONFIG_MEMORY_FAILURE
5377 {1UL << PG_hwpoison, "hwpoison" },
5382 static void dump_page_flags(unsigned long flags)
5384 const char *delim = "";
5388 printk(KERN_ALERT "page flags: %#lx(", flags);
5390 /* remove zone id */
5391 flags &= (1UL << NR_PAGEFLAGS) - 1;
5393 for (i = 0; pageflag_names[i].name && flags; i++) {
5395 mask = pageflag_names[i].mask;
5396 if ((flags & mask) != mask)
5400 printk("%s%s", delim, pageflag_names[i].name);
5404 /* check for left over flags */
5406 printk("%s%#lx", delim, flags);
5411 void dump_page(struct page *page)
5414 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5415 page, page_count(page), page_mapcount(page),
5416 page->mapping, page->index);
5417 dump_page_flags(page->flags);