2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/compiler.h>
25 #include <linux/kernel.h>
26 #include <linux/kmemcheck.h>
27 #include <linux/module.h>
28 #include <linux/suspend.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/slab.h>
32 #include <linux/oom.h>
33 #include <linux/notifier.h>
34 #include <linux/topology.h>
35 #include <linux/sysctl.h>
36 #include <linux/cpu.h>
37 #include <linux/cpuset.h>
38 #include <linux/memory_hotplug.h>
39 #include <linux/nodemask.h>
40 #include <linux/vmalloc.h>
41 #include <linux/mempolicy.h>
42 #include <linux/stop_machine.h>
43 #include <linux/sort.h>
44 #include <linux/pfn.h>
45 #include <linux/backing-dev.h>
46 #include <linux/fault-inject.h>
47 #include <linux/page-isolation.h>
48 #include <linux/page_cgroup.h>
49 #include <linux/debugobjects.h>
50 #include <linux/kmemleak.h>
51 #include <trace/events/kmem.h>
53 #include <asm/tlbflush.h>
54 #include <asm/div64.h>
58 * Array of node states.
60 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
61 [N_POSSIBLE] = NODE_MASK_ALL,
62 [N_ONLINE] = { { [0] = 1UL } },
64 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
66 [N_HIGH_MEMORY] = { { [0] = 1UL } },
68 [N_CPU] = { { [0] = 1UL } },
71 EXPORT_SYMBOL(node_states);
73 unsigned long totalram_pages __read_mostly;
74 unsigned long totalreserve_pages __read_mostly;
75 int percpu_pagelist_fraction;
76 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
78 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
79 int pageblock_order __read_mostly;
82 static void __free_pages_ok(struct page *page, unsigned int order);
85 * results with 256, 32 in the lowmem_reserve sysctl:
86 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
87 * 1G machine -> (16M dma, 784M normal, 224M high)
88 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
89 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
90 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
92 * TBD: should special case ZONE_DMA32 machines here - in those we normally
93 * don't need any ZONE_NORMAL reservation
95 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
96 #ifdef CONFIG_ZONE_DMA
99 #ifdef CONFIG_ZONE_DMA32
102 #ifdef CONFIG_HIGHMEM
108 EXPORT_SYMBOL(totalram_pages);
110 static char * const zone_names[MAX_NR_ZONES] = {
111 #ifdef CONFIG_ZONE_DMA
114 #ifdef CONFIG_ZONE_DMA32
118 #ifdef CONFIG_HIGHMEM
124 int min_free_kbytes = 1024;
125 int min_free_order_shift = 1;
127 static unsigned long __meminitdata nr_kernel_pages;
128 static unsigned long __meminitdata nr_all_pages;
129 static unsigned long __meminitdata dma_reserve;
131 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
133 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
134 * ranges of memory (RAM) that may be registered with add_active_range().
135 * Ranges passed to add_active_range() will be merged if possible
136 * so the number of times add_active_range() can be called is
137 * related to the number of nodes and the number of holes
139 #ifdef CONFIG_MAX_ACTIVE_REGIONS
140 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
141 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
143 #if MAX_NUMNODES >= 32
144 /* If there can be many nodes, allow up to 50 holes per node */
145 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
147 /* By default, allow up to 256 distinct regions */
148 #define MAX_ACTIVE_REGIONS 256
152 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
153 static int __meminitdata nr_nodemap_entries;
154 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
155 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
156 static unsigned long __initdata required_kernelcore;
157 static unsigned long __initdata required_movablecore;
158 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
160 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
162 EXPORT_SYMBOL(movable_zone);
163 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
166 int nr_node_ids __read_mostly = MAX_NUMNODES;
167 int nr_online_nodes __read_mostly = 1;
168 EXPORT_SYMBOL(nr_node_ids);
169 EXPORT_SYMBOL(nr_online_nodes);
172 int page_group_by_mobility_disabled __read_mostly;
174 static void set_pageblock_migratetype(struct page *page, int migratetype)
177 if (unlikely(page_group_by_mobility_disabled))
178 migratetype = MIGRATE_UNMOVABLE;
180 set_pageblock_flags_group(page, (unsigned long)migratetype,
181 PB_migrate, PB_migrate_end);
184 bool oom_killer_disabled __read_mostly;
186 #ifdef CONFIG_DEBUG_VM
187 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
191 unsigned long pfn = page_to_pfn(page);
194 seq = zone_span_seqbegin(zone);
195 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
197 else if (pfn < zone->zone_start_pfn)
199 } while (zone_span_seqretry(zone, seq));
204 static int page_is_consistent(struct zone *zone, struct page *page)
206 if (!pfn_valid_within(page_to_pfn(page)))
208 if (zone != page_zone(page))
214 * Temporary debugging check for pages not lying within a given zone.
216 static int bad_range(struct zone *zone, struct page *page)
218 if (page_outside_zone_boundaries(zone, page))
220 if (!page_is_consistent(zone, page))
226 static inline int bad_range(struct zone *zone, struct page *page)
232 static void bad_page(struct page *page)
234 static unsigned long resume;
235 static unsigned long nr_shown;
236 static unsigned long nr_unshown;
238 /* Don't complain about poisoned pages */
239 if (PageHWPoison(page)) {
240 __ClearPageBuddy(page);
245 * Allow a burst of 60 reports, then keep quiet for that minute;
246 * or allow a steady drip of one report per second.
248 if (nr_shown == 60) {
249 if (time_before(jiffies, resume)) {
255 "BUG: Bad page state: %lu messages suppressed\n",
262 resume = jiffies + 60 * HZ;
264 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
265 current->comm, page_to_pfn(page));
267 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
268 page, (void *)page->flags, page_count(page),
269 page_mapcount(page), page->mapping, page->index);
273 /* Leave bad fields for debug, except PageBuddy could make trouble */
274 __ClearPageBuddy(page);
275 add_taint(TAINT_BAD_PAGE);
279 * Higher-order pages are called "compound pages". They are structured thusly:
281 * The first PAGE_SIZE page is called the "head page".
283 * The remaining PAGE_SIZE pages are called "tail pages".
285 * All pages have PG_compound set. All pages have their ->private pointing at
286 * the head page (even the head page has this).
288 * The first tail page's ->lru.next holds the address of the compound page's
289 * put_page() function. Its ->lru.prev holds the order of allocation.
290 * This usage means that zero-order pages may not be compound.
293 static void free_compound_page(struct page *page)
295 __free_pages_ok(page, compound_order(page));
298 void prep_compound_page(struct page *page, unsigned long order)
301 int nr_pages = 1 << order;
303 set_compound_page_dtor(page, free_compound_page);
304 set_compound_order(page, order);
306 for (i = 1; i < nr_pages; i++) {
307 struct page *p = page + i;
310 p->first_page = page;
314 static int destroy_compound_page(struct page *page, unsigned long order)
317 int nr_pages = 1 << order;
320 if (unlikely(compound_order(page) != order) ||
321 unlikely(!PageHead(page))) {
326 __ClearPageHead(page);
328 for (i = 1; i < nr_pages; i++) {
329 struct page *p = page + i;
331 if (unlikely(!PageTail(p) || (p->first_page != page))) {
341 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
346 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
347 * and __GFP_HIGHMEM from hard or soft interrupt context.
349 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
350 for (i = 0; i < (1 << order); i++)
351 clear_highpage(page + i);
354 static inline void set_page_order(struct page *page, int order)
356 set_page_private(page, order);
357 __SetPageBuddy(page);
360 static inline void rmv_page_order(struct page *page)
362 __ClearPageBuddy(page);
363 set_page_private(page, 0);
367 * Locate the struct page for both the matching buddy in our
368 * pair (buddy1) and the combined O(n+1) page they form (page).
370 * 1) Any buddy B1 will have an order O twin B2 which satisfies
371 * the following equation:
373 * For example, if the starting buddy (buddy2) is #8 its order
375 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
377 * 2) Any buddy B will have an order O+1 parent P which
378 * satisfies the following equation:
381 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
383 static inline struct page *
384 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
386 unsigned long buddy_idx = page_idx ^ (1 << order);
388 return page + (buddy_idx - page_idx);
391 static inline unsigned long
392 __find_combined_index(unsigned long page_idx, unsigned int order)
394 return (page_idx & ~(1 << order));
398 * This function checks whether a page is free && is the buddy
399 * we can do coalesce a page and its buddy if
400 * (a) the buddy is not in a hole &&
401 * (b) the buddy is in the buddy system &&
402 * (c) a page and its buddy have the same order &&
403 * (d) a page and its buddy are in the same zone.
405 * For recording whether a page is in the buddy system, we use PG_buddy.
406 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
408 * For recording page's order, we use page_private(page).
410 static inline int page_is_buddy(struct page *page, struct page *buddy,
413 if (!pfn_valid_within(page_to_pfn(buddy)))
416 if (page_zone_id(page) != page_zone_id(buddy))
419 if (PageBuddy(buddy) && page_order(buddy) == order) {
420 VM_BUG_ON(page_count(buddy) != 0);
427 * Freeing function for a buddy system allocator.
429 * The concept of a buddy system is to maintain direct-mapped table
430 * (containing bit values) for memory blocks of various "orders".
431 * The bottom level table contains the map for the smallest allocatable
432 * units of memory (here, pages), and each level above it describes
433 * pairs of units from the levels below, hence, "buddies".
434 * At a high level, all that happens here is marking the table entry
435 * at the bottom level available, and propagating the changes upward
436 * as necessary, plus some accounting needed to play nicely with other
437 * parts of the VM system.
438 * At each level, we keep a list of pages, which are heads of continuous
439 * free pages of length of (1 << order) and marked with PG_buddy. Page's
440 * order is recorded in page_private(page) field.
441 * So when we are allocating or freeing one, we can derive the state of the
442 * other. That is, if we allocate a small block, and both were
443 * free, the remainder of the region must be split into blocks.
444 * If a block is freed, and its buddy is also free, then this
445 * triggers coalescing into a block of larger size.
450 static inline void __free_one_page(struct page *page,
451 struct zone *zone, unsigned int order,
454 unsigned long page_idx;
456 if (unlikely(PageCompound(page)))
457 if (unlikely(destroy_compound_page(page, order)))
460 VM_BUG_ON(migratetype == -1);
462 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
464 VM_BUG_ON(page_idx & ((1 << order) - 1));
465 VM_BUG_ON(bad_range(zone, page));
467 while (order < MAX_ORDER-1) {
468 unsigned long combined_idx;
471 buddy = __page_find_buddy(page, page_idx, order);
472 if (!page_is_buddy(page, buddy, order))
475 /* Our buddy is free, merge with it and move up one order. */
476 list_del(&buddy->lru);
477 zone->free_area[order].nr_free--;
478 rmv_page_order(buddy);
479 combined_idx = __find_combined_index(page_idx, order);
480 page = page + (combined_idx - page_idx);
481 page_idx = combined_idx;
484 set_page_order(page, order);
486 &zone->free_area[order].free_list[migratetype]);
487 zone->free_area[order].nr_free++;
490 #ifdef CONFIG_HAVE_MLOCKED_PAGE_BIT
492 * free_page_mlock() -- clean up attempts to free and mlocked() page.
493 * Page should not be on lru, so no need to fix that up.
494 * free_pages_check() will verify...
496 static inline void free_page_mlock(struct page *page)
498 __dec_zone_page_state(page, NR_MLOCK);
499 __count_vm_event(UNEVICTABLE_MLOCKFREED);
502 static void free_page_mlock(struct page *page) { }
505 static inline int free_pages_check(struct page *page)
507 if (unlikely(page_mapcount(page) |
508 (page->mapping != NULL) |
509 (atomic_read(&page->_count) != 0) |
510 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) {
514 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
515 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
520 * Frees a number of pages from the PCP lists
521 * Assumes all pages on list are in same zone, and of same order.
522 * count is the number of pages to free.
524 * If the zone was previously in an "all pages pinned" state then look to
525 * see if this freeing clears that state.
527 * And clear the zone's pages_scanned counter, to hold off the "all pages are
528 * pinned" detection logic.
530 static void free_pcppages_bulk(struct zone *zone, int count,
531 struct per_cpu_pages *pcp)
537 spin_lock(&zone->lock);
538 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
539 zone->pages_scanned = 0;
543 struct list_head *list;
546 * Remove pages from lists in a round-robin fashion. A
547 * batch_free count is maintained that is incremented when an
548 * empty list is encountered. This is so more pages are freed
549 * off fuller lists instead of spinning excessively around empty
554 if (++migratetype == MIGRATE_PCPTYPES)
556 list = &pcp->lists[migratetype];
557 } while (list_empty(list));
560 page = list_entry(list->prev, struct page, lru);
561 /* must delete as __free_one_page list manipulates */
562 list_del(&page->lru);
563 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
564 __free_one_page(page, zone, 0, page_private(page));
565 trace_mm_page_pcpu_drain(page, 0, page_private(page));
566 } while (--to_free && --batch_free && !list_empty(list));
568 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
569 spin_unlock(&zone->lock);
572 static void free_one_page(struct zone *zone, struct page *page, int order,
575 spin_lock(&zone->lock);
576 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
577 zone->pages_scanned = 0;
579 __free_one_page(page, zone, order, migratetype);
580 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
581 spin_unlock(&zone->lock);
584 static void __free_pages_ok(struct page *page, unsigned int order)
589 int wasMlocked = __TestClearPageMlocked(page);
591 kmemcheck_free_shadow(page, order);
593 for (i = 0 ; i < (1 << order) ; ++i)
594 bad += free_pages_check(page + i);
598 if (!PageHighMem(page)) {
599 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
600 debug_check_no_obj_freed(page_address(page),
603 arch_free_page(page, order);
604 kernel_map_pages(page, 1 << order, 0);
606 local_irq_save(flags);
607 if (unlikely(wasMlocked))
608 free_page_mlock(page);
609 __count_vm_events(PGFREE, 1 << order);
610 free_one_page(page_zone(page), page, order,
611 get_pageblock_migratetype(page));
612 local_irq_restore(flags);
616 * permit the bootmem allocator to evade page validation on high-order frees
618 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
621 __ClearPageReserved(page);
622 set_page_count(page, 0);
623 set_page_refcounted(page);
629 for (loop = 0; loop < BITS_PER_LONG; loop++) {
630 struct page *p = &page[loop];
632 if (loop + 1 < BITS_PER_LONG)
634 __ClearPageReserved(p);
635 set_page_count(p, 0);
638 set_page_refcounted(page);
639 __free_pages(page, order);
645 * The order of subdivision here is critical for the IO subsystem.
646 * Please do not alter this order without good reasons and regression
647 * testing. Specifically, as large blocks of memory are subdivided,
648 * the order in which smaller blocks are delivered depends on the order
649 * they're subdivided in this function. This is the primary factor
650 * influencing the order in which pages are delivered to the IO
651 * subsystem according to empirical testing, and this is also justified
652 * by considering the behavior of a buddy system containing a single
653 * large block of memory acted on by a series of small allocations.
654 * This behavior is a critical factor in sglist merging's success.
658 static inline void expand(struct zone *zone, struct page *page,
659 int low, int high, struct free_area *area,
662 unsigned long size = 1 << high;
668 VM_BUG_ON(bad_range(zone, &page[size]));
669 list_add(&page[size].lru, &area->free_list[migratetype]);
671 set_page_order(&page[size], high);
676 * This page is about to be returned from the page allocator
678 static inline int check_new_page(struct page *page)
680 if (unlikely(page_mapcount(page) |
681 (page->mapping != NULL) |
682 (atomic_read(&page->_count) != 0) |
683 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) {
690 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
694 for (i = 0; i < (1 << order); i++) {
695 struct page *p = page + i;
696 if (unlikely(check_new_page(p)))
700 set_page_private(page, 0);
701 set_page_refcounted(page);
703 arch_alloc_page(page, order);
704 kernel_map_pages(page, 1 << order, 1);
706 if (gfp_flags & __GFP_ZERO)
707 prep_zero_page(page, order, gfp_flags);
709 if (order && (gfp_flags & __GFP_COMP))
710 prep_compound_page(page, order);
716 * Go through the free lists for the given migratetype and remove
717 * the smallest available page from the freelists
720 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
723 unsigned int current_order;
724 struct free_area * area;
727 /* Find a page of the appropriate size in the preferred list */
728 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
729 area = &(zone->free_area[current_order]);
730 if (list_empty(&area->free_list[migratetype]))
733 page = list_entry(area->free_list[migratetype].next,
735 list_del(&page->lru);
736 rmv_page_order(page);
738 expand(zone, page, order, current_order, area, migratetype);
747 * This array describes the order lists are fallen back to when
748 * the free lists for the desirable migrate type are depleted
750 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
751 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
752 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
753 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
754 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
758 * Move the free pages in a range to the free lists of the requested type.
759 * Note that start_page and end_pages are not aligned on a pageblock
760 * boundary. If alignment is required, use move_freepages_block()
762 static int move_freepages(struct zone *zone,
763 struct page *start_page, struct page *end_page,
770 #ifndef CONFIG_HOLES_IN_ZONE
772 * page_zone is not safe to call in this context when
773 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
774 * anyway as we check zone boundaries in move_freepages_block().
775 * Remove at a later date when no bug reports exist related to
776 * grouping pages by mobility
778 BUG_ON(page_zone(start_page) != page_zone(end_page));
781 for (page = start_page; page <= end_page;) {
782 /* Make sure we are not inadvertently changing nodes */
783 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
785 if (!pfn_valid_within(page_to_pfn(page))) {
790 if (!PageBuddy(page)) {
795 order = page_order(page);
796 list_del(&page->lru);
798 &zone->free_area[order].free_list[migratetype]);
800 pages_moved += 1 << order;
806 static int move_freepages_block(struct zone *zone, struct page *page,
809 unsigned long start_pfn, end_pfn;
810 struct page *start_page, *end_page;
812 start_pfn = page_to_pfn(page);
813 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
814 start_page = pfn_to_page(start_pfn);
815 end_page = start_page + pageblock_nr_pages - 1;
816 end_pfn = start_pfn + pageblock_nr_pages - 1;
818 /* Do not cross zone boundaries */
819 if (start_pfn < zone->zone_start_pfn)
821 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
824 return move_freepages(zone, start_page, end_page, migratetype);
827 static void change_pageblock_range(struct page *pageblock_page,
828 int start_order, int migratetype)
830 int nr_pageblocks = 1 << (start_order - pageblock_order);
832 while (nr_pageblocks--) {
833 set_pageblock_migratetype(pageblock_page, migratetype);
834 pageblock_page += pageblock_nr_pages;
838 /* Remove an element from the buddy allocator from the fallback list */
839 static inline struct page *
840 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
842 struct free_area * area;
847 /* Find the largest possible block of pages in the other list */
848 for (current_order = MAX_ORDER-1; current_order >= order;
850 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
851 migratetype = fallbacks[start_migratetype][i];
853 /* MIGRATE_RESERVE handled later if necessary */
854 if (migratetype == MIGRATE_RESERVE)
857 area = &(zone->free_area[current_order]);
858 if (list_empty(&area->free_list[migratetype]))
861 page = list_entry(area->free_list[migratetype].next,
866 * If breaking a large block of pages, move all free
867 * pages to the preferred allocation list. If falling
868 * back for a reclaimable kernel allocation, be more
869 * agressive about taking ownership of free pages
871 if (unlikely(current_order >= (pageblock_order >> 1)) ||
872 start_migratetype == MIGRATE_RECLAIMABLE ||
873 page_group_by_mobility_disabled) {
875 pages = move_freepages_block(zone, page,
878 /* Claim the whole block if over half of it is free */
879 if (pages >= (1 << (pageblock_order-1)) ||
880 page_group_by_mobility_disabled)
881 set_pageblock_migratetype(page,
884 migratetype = start_migratetype;
887 /* Remove the page from the freelists */
888 list_del(&page->lru);
889 rmv_page_order(page);
891 /* Take ownership for orders >= pageblock_order */
892 if (current_order >= pageblock_order)
893 change_pageblock_range(page, current_order,
896 expand(zone, page, order, current_order, area, migratetype);
898 trace_mm_page_alloc_extfrag(page, order, current_order,
899 start_migratetype, migratetype);
909 * Do the hard work of removing an element from the buddy allocator.
910 * Call me with the zone->lock already held.
912 static struct page *__rmqueue(struct zone *zone, unsigned int order,
918 page = __rmqueue_smallest(zone, order, migratetype);
920 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
921 page = __rmqueue_fallback(zone, order, migratetype);
924 * Use MIGRATE_RESERVE rather than fail an allocation. goto
925 * is used because __rmqueue_smallest is an inline function
926 * and we want just one call site
929 migratetype = MIGRATE_RESERVE;
934 trace_mm_page_alloc_zone_locked(page, order, migratetype);
939 * Obtain a specified number of elements from the buddy allocator, all under
940 * a single hold of the lock, for efficiency. Add them to the supplied list.
941 * Returns the number of new pages which were placed at *list.
943 static int rmqueue_bulk(struct zone *zone, unsigned int order,
944 unsigned long count, struct list_head *list,
945 int migratetype, int cold)
949 spin_lock(&zone->lock);
950 for (i = 0; i < count; ++i) {
951 struct page *page = __rmqueue(zone, order, migratetype);
952 if (unlikely(page == NULL))
956 * Split buddy pages returned by expand() are received here
957 * in physical page order. The page is added to the callers and
958 * list and the list head then moves forward. From the callers
959 * perspective, the linked list is ordered by page number in
960 * some conditions. This is useful for IO devices that can
961 * merge IO requests if the physical pages are ordered
964 if (likely(cold == 0))
965 list_add(&page->lru, list);
967 list_add_tail(&page->lru, list);
968 set_page_private(page, migratetype);
971 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
972 spin_unlock(&zone->lock);
978 * Called from the vmstat counter updater to drain pagesets of this
979 * currently executing processor on remote nodes after they have
982 * Note that this function must be called with the thread pinned to
983 * a single processor.
985 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
990 local_irq_save(flags);
991 if (pcp->count >= pcp->batch)
992 to_drain = pcp->batch;
994 to_drain = pcp->count;
995 free_pcppages_bulk(zone, to_drain, pcp);
996 pcp->count -= to_drain;
997 local_irq_restore(flags);
1002 * Drain pages of the indicated processor.
1004 * The processor must either be the current processor and the
1005 * thread pinned to the current processor or a processor that
1008 static void drain_pages(unsigned int cpu)
1010 unsigned long flags;
1013 for_each_populated_zone(zone) {
1014 struct per_cpu_pageset *pset;
1015 struct per_cpu_pages *pcp;
1017 pset = zone_pcp(zone, cpu);
1020 local_irq_save(flags);
1021 free_pcppages_bulk(zone, pcp->count, pcp);
1023 local_irq_restore(flags);
1028 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1030 void drain_local_pages(void *arg)
1032 drain_pages(smp_processor_id());
1036 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1038 void drain_all_pages(void)
1040 on_each_cpu(drain_local_pages, NULL, 1);
1043 #ifdef CONFIG_HIBERNATION
1045 void mark_free_pages(struct zone *zone)
1047 unsigned long pfn, max_zone_pfn;
1048 unsigned long flags;
1050 struct list_head *curr;
1052 if (!zone->spanned_pages)
1055 spin_lock_irqsave(&zone->lock, flags);
1057 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1058 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1059 if (pfn_valid(pfn)) {
1060 struct page *page = pfn_to_page(pfn);
1062 if (!swsusp_page_is_forbidden(page))
1063 swsusp_unset_page_free(page);
1066 for_each_migratetype_order(order, t) {
1067 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1070 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1071 for (i = 0; i < (1UL << order); i++)
1072 swsusp_set_page_free(pfn_to_page(pfn + i));
1075 spin_unlock_irqrestore(&zone->lock, flags);
1077 #endif /* CONFIG_PM */
1080 * Free a 0-order page
1082 static void free_hot_cold_page(struct page *page, int cold)
1084 struct zone *zone = page_zone(page);
1085 struct per_cpu_pages *pcp;
1086 unsigned long flags;
1088 int wasMlocked = __TestClearPageMlocked(page);
1090 kmemcheck_free_shadow(page, 0);
1093 page->mapping = NULL;
1094 if (free_pages_check(page))
1097 if (!PageHighMem(page)) {
1098 debug_check_no_locks_freed(page_address(page), PAGE_SIZE);
1099 debug_check_no_obj_freed(page_address(page), PAGE_SIZE);
1101 arch_free_page(page, 0);
1102 kernel_map_pages(page, 1, 0);
1104 pcp = &zone_pcp(zone, get_cpu())->pcp;
1105 migratetype = get_pageblock_migratetype(page);
1106 set_page_private(page, migratetype);
1107 local_irq_save(flags);
1108 if (unlikely(wasMlocked))
1109 free_page_mlock(page);
1110 __count_vm_event(PGFREE);
1113 * We only track unmovable, reclaimable and movable on pcp lists.
1114 * Free ISOLATE pages back to the allocator because they are being
1115 * offlined but treat RESERVE as movable pages so we can get those
1116 * areas back if necessary. Otherwise, we may have to free
1117 * excessively into the page allocator
1119 if (migratetype >= MIGRATE_PCPTYPES) {
1120 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1121 free_one_page(zone, page, 0, migratetype);
1124 migratetype = MIGRATE_MOVABLE;
1128 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1130 list_add(&page->lru, &pcp->lists[migratetype]);
1132 if (pcp->count >= pcp->high) {
1133 free_pcppages_bulk(zone, pcp->batch, pcp);
1134 pcp->count -= pcp->batch;
1138 local_irq_restore(flags);
1142 void free_hot_page(struct page *page)
1144 trace_mm_page_free_direct(page, 0);
1145 free_hot_cold_page(page, 0);
1149 * split_page takes a non-compound higher-order page, and splits it into
1150 * n (1<<order) sub-pages: page[0..n]
1151 * Each sub-page must be freed individually.
1153 * Note: this is probably too low level an operation for use in drivers.
1154 * Please consult with lkml before using this in your driver.
1156 void split_page(struct page *page, unsigned int order)
1160 VM_BUG_ON(PageCompound(page));
1161 VM_BUG_ON(!page_count(page));
1163 #ifdef CONFIG_KMEMCHECK
1165 * Split shadow pages too, because free(page[0]) would
1166 * otherwise free the whole shadow.
1168 if (kmemcheck_page_is_tracked(page))
1169 split_page(virt_to_page(page[0].shadow), order);
1172 for (i = 1; i < (1 << order); i++)
1173 set_page_refcounted(page + i);
1177 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1178 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1182 struct page *buffered_rmqueue(struct zone *preferred_zone,
1183 struct zone *zone, int order, gfp_t gfp_flags,
1186 unsigned long flags;
1188 int cold = !!(gfp_flags & __GFP_COLD);
1193 if (likely(order == 0)) {
1194 struct per_cpu_pages *pcp;
1195 struct list_head *list;
1197 pcp = &zone_pcp(zone, cpu)->pcp;
1198 list = &pcp->lists[migratetype];
1199 local_irq_save(flags);
1200 if (list_empty(list)) {
1201 pcp->count += rmqueue_bulk(zone, 0,
1204 if (unlikely(list_empty(list)))
1209 page = list_entry(list->prev, struct page, lru);
1211 page = list_entry(list->next, struct page, lru);
1213 list_del(&page->lru);
1216 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1218 * __GFP_NOFAIL is not to be used in new code.
1220 * All __GFP_NOFAIL callers should be fixed so that they
1221 * properly detect and handle allocation failures.
1223 * We most definitely don't want callers attempting to
1224 * allocate greater than order-1 page units with
1227 WARN_ON_ONCE(order > 1);
1229 spin_lock_irqsave(&zone->lock, flags);
1230 page = __rmqueue(zone, order, migratetype);
1231 spin_unlock(&zone->lock);
1234 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1237 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1238 zone_statistics(preferred_zone, zone);
1239 local_irq_restore(flags);
1242 VM_BUG_ON(bad_range(zone, page));
1243 if (prep_new_page(page, order, gfp_flags))
1248 local_irq_restore(flags);
1253 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1254 #define ALLOC_WMARK_MIN WMARK_MIN
1255 #define ALLOC_WMARK_LOW WMARK_LOW
1256 #define ALLOC_WMARK_HIGH WMARK_HIGH
1257 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1259 /* Mask to get the watermark bits */
1260 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1262 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1263 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1264 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1266 #ifdef CONFIG_FAIL_PAGE_ALLOC
1268 static struct fail_page_alloc_attr {
1269 struct fault_attr attr;
1271 u32 ignore_gfp_highmem;
1272 u32 ignore_gfp_wait;
1275 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1277 struct dentry *ignore_gfp_highmem_file;
1278 struct dentry *ignore_gfp_wait_file;
1279 struct dentry *min_order_file;
1281 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1283 } fail_page_alloc = {
1284 .attr = FAULT_ATTR_INITIALIZER,
1285 .ignore_gfp_wait = 1,
1286 .ignore_gfp_highmem = 1,
1290 static int __init setup_fail_page_alloc(char *str)
1292 return setup_fault_attr(&fail_page_alloc.attr, str);
1294 __setup("fail_page_alloc=", setup_fail_page_alloc);
1296 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1298 if (order < fail_page_alloc.min_order)
1300 if (gfp_mask & __GFP_NOFAIL)
1302 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1304 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1307 return should_fail(&fail_page_alloc.attr, 1 << order);
1310 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1312 static int __init fail_page_alloc_debugfs(void)
1314 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1318 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1322 dir = fail_page_alloc.attr.dentries.dir;
1324 fail_page_alloc.ignore_gfp_wait_file =
1325 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1326 &fail_page_alloc.ignore_gfp_wait);
1328 fail_page_alloc.ignore_gfp_highmem_file =
1329 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1330 &fail_page_alloc.ignore_gfp_highmem);
1331 fail_page_alloc.min_order_file =
1332 debugfs_create_u32("min-order", mode, dir,
1333 &fail_page_alloc.min_order);
1335 if (!fail_page_alloc.ignore_gfp_wait_file ||
1336 !fail_page_alloc.ignore_gfp_highmem_file ||
1337 !fail_page_alloc.min_order_file) {
1339 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1340 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1341 debugfs_remove(fail_page_alloc.min_order_file);
1342 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1348 late_initcall(fail_page_alloc_debugfs);
1350 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1352 #else /* CONFIG_FAIL_PAGE_ALLOC */
1354 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1359 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1362 * Return 1 if free pages are above 'mark'. This takes into account the order
1363 * of the allocation.
1365 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1366 int classzone_idx, int alloc_flags)
1368 /* free_pages my go negative - that's OK */
1370 long free_pages = zone_nr_free_pages(z) - (1 << order) + 1;
1373 if (alloc_flags & ALLOC_HIGH)
1375 if (alloc_flags & ALLOC_HARDER)
1378 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1380 for (o = 0; o < order; o++) {
1381 /* At the next order, this order's pages become unavailable */
1382 free_pages -= z->free_area[o].nr_free << o;
1384 /* Require fewer higher order pages to be free */
1385 min >>= min_free_order_shift;
1387 if (free_pages <= min)
1395 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1396 * skip over zones that are not allowed by the cpuset, or that have
1397 * been recently (in last second) found to be nearly full. See further
1398 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1399 * that have to skip over a lot of full or unallowed zones.
1401 * If the zonelist cache is present in the passed in zonelist, then
1402 * returns a pointer to the allowed node mask (either the current
1403 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1405 * If the zonelist cache is not available for this zonelist, does
1406 * nothing and returns NULL.
1408 * If the fullzones BITMAP in the zonelist cache is stale (more than
1409 * a second since last zap'd) then we zap it out (clear its bits.)
1411 * We hold off even calling zlc_setup, until after we've checked the
1412 * first zone in the zonelist, on the theory that most allocations will
1413 * be satisfied from that first zone, so best to examine that zone as
1414 * quickly as we can.
1416 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1418 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1419 nodemask_t *allowednodes; /* zonelist_cache approximation */
1421 zlc = zonelist->zlcache_ptr;
1425 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1426 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1427 zlc->last_full_zap = jiffies;
1430 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1431 &cpuset_current_mems_allowed :
1432 &node_states[N_HIGH_MEMORY];
1433 return allowednodes;
1437 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1438 * if it is worth looking at further for free memory:
1439 * 1) Check that the zone isn't thought to be full (doesn't have its
1440 * bit set in the zonelist_cache fullzones BITMAP).
1441 * 2) Check that the zones node (obtained from the zonelist_cache
1442 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1443 * Return true (non-zero) if zone is worth looking at further, or
1444 * else return false (zero) if it is not.
1446 * This check -ignores- the distinction between various watermarks,
1447 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1448 * found to be full for any variation of these watermarks, it will
1449 * be considered full for up to one second by all requests, unless
1450 * we are so low on memory on all allowed nodes that we are forced
1451 * into the second scan of the zonelist.
1453 * In the second scan we ignore this zonelist cache and exactly
1454 * apply the watermarks to all zones, even it is slower to do so.
1455 * We are low on memory in the second scan, and should leave no stone
1456 * unturned looking for a free page.
1458 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1459 nodemask_t *allowednodes)
1461 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1462 int i; /* index of *z in zonelist zones */
1463 int n; /* node that zone *z is on */
1465 zlc = zonelist->zlcache_ptr;
1469 i = z - zonelist->_zonerefs;
1472 /* This zone is worth trying if it is allowed but not full */
1473 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1477 * Given 'z' scanning a zonelist, set the corresponding bit in
1478 * zlc->fullzones, so that subsequent attempts to allocate a page
1479 * from that zone don't waste time re-examining it.
1481 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1483 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1484 int i; /* index of *z in zonelist zones */
1486 zlc = zonelist->zlcache_ptr;
1490 i = z - zonelist->_zonerefs;
1492 set_bit(i, zlc->fullzones);
1495 #else /* CONFIG_NUMA */
1497 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1502 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1503 nodemask_t *allowednodes)
1508 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1511 #endif /* CONFIG_NUMA */
1514 * get_page_from_freelist goes through the zonelist trying to allocate
1517 static struct page *
1518 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1519 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1520 struct zone *preferred_zone, int migratetype)
1523 struct page *page = NULL;
1526 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1527 int zlc_active = 0; /* set if using zonelist_cache */
1528 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1530 classzone_idx = zone_idx(preferred_zone);
1533 * Scan zonelist, looking for a zone with enough free.
1534 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1536 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1537 high_zoneidx, nodemask) {
1538 if (NUMA_BUILD && zlc_active &&
1539 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1541 if ((alloc_flags & ALLOC_CPUSET) &&
1542 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1545 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1546 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1550 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1551 if (zone_watermark_ok(zone, order, mark,
1552 classzone_idx, alloc_flags))
1555 if (zone_reclaim_mode == 0)
1556 goto this_zone_full;
1558 ret = zone_reclaim(zone, gfp_mask, order);
1560 case ZONE_RECLAIM_NOSCAN:
1563 case ZONE_RECLAIM_FULL:
1564 /* scanned but unreclaimable */
1565 goto this_zone_full;
1567 /* did we reclaim enough */
1568 if (!zone_watermark_ok(zone, order, mark,
1569 classzone_idx, alloc_flags))
1570 goto this_zone_full;
1575 page = buffered_rmqueue(preferred_zone, zone, order,
1576 gfp_mask, migratetype);
1581 zlc_mark_zone_full(zonelist, z);
1583 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1585 * we do zlc_setup after the first zone is tried but only
1586 * if there are multiple nodes make it worthwhile
1588 allowednodes = zlc_setup(zonelist, alloc_flags);
1594 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1595 /* Disable zlc cache for second zonelist scan */
1603 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1604 unsigned long pages_reclaimed)
1606 /* Do not loop if specifically requested */
1607 if (gfp_mask & __GFP_NORETRY)
1611 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1612 * means __GFP_NOFAIL, but that may not be true in other
1615 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1619 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1620 * specified, then we retry until we no longer reclaim any pages
1621 * (above), or we've reclaimed an order of pages at least as
1622 * large as the allocation's order. In both cases, if the
1623 * allocation still fails, we stop retrying.
1625 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1629 * Don't let big-order allocations loop unless the caller
1630 * explicitly requests that.
1632 if (gfp_mask & __GFP_NOFAIL)
1638 static inline struct page *
1639 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1640 struct zonelist *zonelist, enum zone_type high_zoneidx,
1641 nodemask_t *nodemask, struct zone *preferred_zone,
1646 /* Acquire the OOM killer lock for the zones in zonelist */
1647 if (!try_set_zone_oom(zonelist, gfp_mask)) {
1648 schedule_timeout_uninterruptible(1);
1653 * Go through the zonelist yet one more time, keep very high watermark
1654 * here, this is only to catch a parallel oom killing, we must fail if
1655 * we're still under heavy pressure.
1657 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1658 order, zonelist, high_zoneidx,
1659 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1660 preferred_zone, migratetype);
1664 /* The OOM killer will not help higher order allocs */
1665 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_NOFAIL))
1668 /* Exhausted what can be done so it's blamo time */
1669 out_of_memory(zonelist, gfp_mask, order);
1672 clear_zonelist_oom(zonelist, gfp_mask);
1676 /* The really slow allocator path where we enter direct reclaim */
1677 static inline struct page *
1678 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1679 struct zonelist *zonelist, enum zone_type high_zoneidx,
1680 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1681 int migratetype, unsigned long *did_some_progress)
1683 struct page *page = NULL;
1684 struct reclaim_state reclaim_state;
1685 struct task_struct *p = current;
1686 bool drained = false;
1690 /* We now go into synchronous reclaim */
1691 cpuset_memory_pressure_bump();
1692 p->flags |= PF_MEMALLOC;
1693 lockdep_set_current_reclaim_state(gfp_mask);
1694 reclaim_state.reclaimed_slab = 0;
1695 p->reclaim_state = &reclaim_state;
1697 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1699 p->reclaim_state = NULL;
1700 lockdep_clear_current_reclaim_state();
1701 p->flags &= ~PF_MEMALLOC;
1705 if (unlikely(!(*did_some_progress)))
1709 page = get_page_from_freelist(gfp_mask, nodemask, order,
1710 zonelist, high_zoneidx,
1711 alloc_flags, preferred_zone,
1715 * If an allocation failed after direct reclaim, it could be because
1716 * pages are pinned on the per-cpu lists. Drain them and try again
1718 if (!page && !drained) {
1728 * This is called in the allocator slow-path if the allocation request is of
1729 * sufficient urgency to ignore watermarks and take other desperate measures
1731 static inline struct page *
1732 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1733 struct zonelist *zonelist, enum zone_type high_zoneidx,
1734 nodemask_t *nodemask, struct zone *preferred_zone,
1740 page = get_page_from_freelist(gfp_mask, nodemask, order,
1741 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
1742 preferred_zone, migratetype);
1744 if (!page && gfp_mask & __GFP_NOFAIL)
1745 congestion_wait(BLK_RW_ASYNC, HZ/50);
1746 } while (!page && (gfp_mask & __GFP_NOFAIL));
1752 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
1753 enum zone_type high_zoneidx)
1758 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1759 wakeup_kswapd(zone, order);
1763 gfp_to_alloc_flags(gfp_t gfp_mask)
1765 struct task_struct *p = current;
1766 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
1767 const gfp_t wait = gfp_mask & __GFP_WAIT;
1769 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1770 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH);
1773 * The caller may dip into page reserves a bit more if the caller
1774 * cannot run direct reclaim, or if the caller has realtime scheduling
1775 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1776 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1778 alloc_flags |= (gfp_mask & __GFP_HIGH);
1781 alloc_flags |= ALLOC_HARDER;
1783 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1784 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1786 alloc_flags &= ~ALLOC_CPUSET;
1787 } else if (unlikely(rt_task(p)) && !in_interrupt())
1788 alloc_flags |= ALLOC_HARDER;
1790 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
1791 if (!in_interrupt() &&
1792 ((p->flags & PF_MEMALLOC) ||
1793 unlikely(test_thread_flag(TIF_MEMDIE))))
1794 alloc_flags |= ALLOC_NO_WATERMARKS;
1800 static inline struct page *
1801 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
1802 struct zonelist *zonelist, enum zone_type high_zoneidx,
1803 nodemask_t *nodemask, struct zone *preferred_zone,
1806 const gfp_t wait = gfp_mask & __GFP_WAIT;
1807 struct page *page = NULL;
1809 unsigned long pages_reclaimed = 0;
1810 unsigned long did_some_progress;
1811 struct task_struct *p = current;
1814 * In the slowpath, we sanity check order to avoid ever trying to
1815 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1816 * be using allocators in order of preference for an area that is
1819 if (order >= MAX_ORDER) {
1820 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
1825 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1826 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1827 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1828 * using a larger set of nodes after it has established that the
1829 * allowed per node queues are empty and that nodes are
1832 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1836 wake_all_kswapd(order, zonelist, high_zoneidx);
1839 * OK, we're below the kswapd watermark and have kicked background
1840 * reclaim. Now things get more complex, so set up alloc_flags according
1841 * to how we want to proceed.
1843 alloc_flags = gfp_to_alloc_flags(gfp_mask);
1846 /* This is the last chance, in general, before the goto nopage. */
1847 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
1848 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
1849 preferred_zone, migratetype);
1853 /* Allocate without watermarks if the context allows */
1854 if (alloc_flags & ALLOC_NO_WATERMARKS) {
1855 page = __alloc_pages_high_priority(gfp_mask, order,
1856 zonelist, high_zoneidx, nodemask,
1857 preferred_zone, migratetype);
1862 /* Atomic allocations - we can't balance anything */
1866 /* Avoid recursion of direct reclaim */
1867 if (p->flags & PF_MEMALLOC)
1870 /* Avoid allocations with no watermarks from looping endlessly */
1871 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
1874 /* Try direct reclaim and then allocating */
1875 page = __alloc_pages_direct_reclaim(gfp_mask, order,
1876 zonelist, high_zoneidx,
1878 alloc_flags, preferred_zone,
1879 migratetype, &did_some_progress);
1884 * If we failed to make any progress reclaiming, then we are
1885 * running out of options and have to consider going OOM
1887 if (!did_some_progress) {
1888 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1889 if (oom_killer_disabled)
1891 page = __alloc_pages_may_oom(gfp_mask, order,
1892 zonelist, high_zoneidx,
1893 nodemask, preferred_zone,
1899 * The OOM killer does not trigger for high-order
1900 * ~__GFP_NOFAIL allocations so if no progress is being
1901 * made, there are no other options and retrying is
1904 if (order > PAGE_ALLOC_COSTLY_ORDER &&
1905 !(gfp_mask & __GFP_NOFAIL))
1912 /* Check if we should retry the allocation */
1913 pages_reclaimed += did_some_progress;
1914 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
1915 /* Wait for some write requests to complete then retry */
1916 congestion_wait(BLK_RW_ASYNC, HZ/50);
1921 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1922 printk(KERN_WARNING "%s: page allocation failure."
1923 " order:%d, mode:0x%x\n",
1924 p->comm, order, gfp_mask);
1930 if (kmemcheck_enabled)
1931 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
1937 * This is the 'heart' of the zoned buddy allocator.
1940 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
1941 struct zonelist *zonelist, nodemask_t *nodemask)
1943 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1944 struct zone *preferred_zone;
1946 int migratetype = allocflags_to_migratetype(gfp_mask);
1948 gfp_mask &= gfp_allowed_mask;
1950 lockdep_trace_alloc(gfp_mask);
1952 might_sleep_if(gfp_mask & __GFP_WAIT);
1954 if (should_fail_alloc_page(gfp_mask, order))
1958 * Check the zones suitable for the gfp_mask contain at least one
1959 * valid zone. It's possible to have an empty zonelist as a result
1960 * of GFP_THISNODE and a memoryless node
1962 if (unlikely(!zonelist->_zonerefs->zone))
1965 /* The preferred zone is used for statistics later */
1966 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone);
1967 if (!preferred_zone)
1970 /* First allocation attempt */
1971 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
1972 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
1973 preferred_zone, migratetype);
1974 if (unlikely(!page))
1975 page = __alloc_pages_slowpath(gfp_mask, order,
1976 zonelist, high_zoneidx, nodemask,
1977 preferred_zone, migratetype);
1979 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
1982 EXPORT_SYMBOL(__alloc_pages_nodemask);
1985 * Common helper functions.
1987 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1992 * __get_free_pages() returns a 32-bit address, which cannot represent
1995 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1997 page = alloc_pages(gfp_mask, order);
2000 return (unsigned long) page_address(page);
2002 EXPORT_SYMBOL(__get_free_pages);
2004 unsigned long get_zeroed_page(gfp_t gfp_mask)
2006 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2008 EXPORT_SYMBOL(get_zeroed_page);
2010 void __pagevec_free(struct pagevec *pvec)
2012 int i = pagevec_count(pvec);
2015 trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
2016 free_hot_cold_page(pvec->pages[i], pvec->cold);
2020 void __free_pages(struct page *page, unsigned int order)
2022 if (put_page_testzero(page)) {
2023 trace_mm_page_free_direct(page, order);
2025 free_hot_page(page);
2027 __free_pages_ok(page, order);
2031 EXPORT_SYMBOL(__free_pages);
2033 void free_pages(unsigned long addr, unsigned int order)
2036 VM_BUG_ON(!virt_addr_valid((void *)addr));
2037 __free_pages(virt_to_page((void *)addr), order);
2041 EXPORT_SYMBOL(free_pages);
2044 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2045 * @size: the number of bytes to allocate
2046 * @gfp_mask: GFP flags for the allocation
2048 * This function is similar to alloc_pages(), except that it allocates the
2049 * minimum number of pages to satisfy the request. alloc_pages() can only
2050 * allocate memory in power-of-two pages.
2052 * This function is also limited by MAX_ORDER.
2054 * Memory allocated by this function must be released by free_pages_exact().
2056 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2058 unsigned int order = get_order(size);
2061 addr = __get_free_pages(gfp_mask, order);
2063 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2064 unsigned long used = addr + PAGE_ALIGN(size);
2066 split_page(virt_to_page((void *)addr), order);
2067 while (used < alloc_end) {
2073 return (void *)addr;
2075 EXPORT_SYMBOL(alloc_pages_exact);
2078 * free_pages_exact - release memory allocated via alloc_pages_exact()
2079 * @virt: the value returned by alloc_pages_exact.
2080 * @size: size of allocation, same value as passed to alloc_pages_exact().
2082 * Release the memory allocated by a previous call to alloc_pages_exact.
2084 void free_pages_exact(void *virt, size_t size)
2086 unsigned long addr = (unsigned long)virt;
2087 unsigned long end = addr + PAGE_ALIGN(size);
2089 while (addr < end) {
2094 EXPORT_SYMBOL(free_pages_exact);
2096 static unsigned int nr_free_zone_pages(int offset)
2101 /* Just pick one node, since fallback list is circular */
2102 unsigned int sum = 0;
2104 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2106 for_each_zone_zonelist(zone, z, zonelist, offset) {
2107 unsigned long size = zone->present_pages;
2108 unsigned long high = high_wmark_pages(zone);
2117 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2119 unsigned int nr_free_buffer_pages(void)
2121 return nr_free_zone_pages(gfp_zone(GFP_USER));
2123 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2126 * Amount of free RAM allocatable within all zones
2128 unsigned int nr_free_pagecache_pages(void)
2130 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2133 static inline void show_node(struct zone *zone)
2136 printk("Node %d ", zone_to_nid(zone));
2139 void si_meminfo(struct sysinfo *val)
2141 val->totalram = totalram_pages;
2143 val->freeram = global_page_state(NR_FREE_PAGES);
2144 val->bufferram = nr_blockdev_pages();
2145 val->totalhigh = totalhigh_pages;
2146 val->freehigh = nr_free_highpages();
2147 val->mem_unit = PAGE_SIZE;
2150 EXPORT_SYMBOL(si_meminfo);
2153 void si_meminfo_node(struct sysinfo *val, int nid)
2155 pg_data_t *pgdat = NODE_DATA(nid);
2157 val->totalram = pgdat->node_present_pages;
2158 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2159 #ifdef CONFIG_HIGHMEM
2160 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2161 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2167 val->mem_unit = PAGE_SIZE;
2171 #define K(x) ((x) << (PAGE_SHIFT-10))
2174 * Show free area list (used inside shift_scroll-lock stuff)
2175 * We also calculate the percentage fragmentation. We do this by counting the
2176 * memory on each free list with the exception of the first item on the list.
2178 void show_free_areas(void)
2183 for_each_populated_zone(zone) {
2185 printk("%s per-cpu:\n", zone->name);
2187 for_each_online_cpu(cpu) {
2188 struct per_cpu_pageset *pageset;
2190 pageset = zone_pcp(zone, cpu);
2192 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2193 cpu, pageset->pcp.high,
2194 pageset->pcp.batch, pageset->pcp.count);
2198 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2199 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2201 " dirty:%lu writeback:%lu unstable:%lu\n"
2202 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2203 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2204 global_page_state(NR_ACTIVE_ANON),
2205 global_page_state(NR_INACTIVE_ANON),
2206 global_page_state(NR_ISOLATED_ANON),
2207 global_page_state(NR_ACTIVE_FILE),
2208 global_page_state(NR_INACTIVE_FILE),
2209 global_page_state(NR_ISOLATED_FILE),
2210 global_page_state(NR_UNEVICTABLE),
2211 global_page_state(NR_FILE_DIRTY),
2212 global_page_state(NR_WRITEBACK),
2213 global_page_state(NR_UNSTABLE_NFS),
2214 global_page_state(NR_FREE_PAGES),
2215 global_page_state(NR_SLAB_RECLAIMABLE),
2216 global_page_state(NR_SLAB_UNRECLAIMABLE),
2217 global_page_state(NR_FILE_MAPPED),
2218 global_page_state(NR_SHMEM),
2219 global_page_state(NR_PAGETABLE),
2220 global_page_state(NR_BOUNCE));
2222 for_each_populated_zone(zone) {
2231 " active_anon:%lukB"
2232 " inactive_anon:%lukB"
2233 " active_file:%lukB"
2234 " inactive_file:%lukB"
2235 " unevictable:%lukB"
2236 " isolated(anon):%lukB"
2237 " isolated(file):%lukB"
2244 " slab_reclaimable:%lukB"
2245 " slab_unreclaimable:%lukB"
2246 " kernel_stack:%lukB"
2250 " writeback_tmp:%lukB"
2251 " pages_scanned:%lu"
2252 " all_unreclaimable? %s"
2255 K(zone_nr_free_pages(zone)),
2256 K(min_wmark_pages(zone)),
2257 K(low_wmark_pages(zone)),
2258 K(high_wmark_pages(zone)),
2259 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2260 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2261 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2262 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2263 K(zone_page_state(zone, NR_UNEVICTABLE)),
2264 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2265 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2266 K(zone->present_pages),
2267 K(zone_page_state(zone, NR_MLOCK)),
2268 K(zone_page_state(zone, NR_FILE_DIRTY)),
2269 K(zone_page_state(zone, NR_WRITEBACK)),
2270 K(zone_page_state(zone, NR_FILE_MAPPED)),
2271 K(zone_page_state(zone, NR_SHMEM)),
2272 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2273 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2274 zone_page_state(zone, NR_KERNEL_STACK) *
2276 K(zone_page_state(zone, NR_PAGETABLE)),
2277 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2278 K(zone_page_state(zone, NR_BOUNCE)),
2279 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2280 zone->pages_scanned,
2281 (zone_is_all_unreclaimable(zone) ? "yes" : "no")
2283 printk("lowmem_reserve[]:");
2284 for (i = 0; i < MAX_NR_ZONES; i++)
2285 printk(" %lu", zone->lowmem_reserve[i]);
2289 for_each_populated_zone(zone) {
2290 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2293 printk("%s: ", zone->name);
2295 spin_lock_irqsave(&zone->lock, flags);
2296 for (order = 0; order < MAX_ORDER; order++) {
2297 nr[order] = zone->free_area[order].nr_free;
2298 total += nr[order] << order;
2300 spin_unlock_irqrestore(&zone->lock, flags);
2301 for (order = 0; order < MAX_ORDER; order++)
2302 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2303 printk("= %lukB\n", K(total));
2306 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2308 show_swap_cache_info();
2311 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2313 zoneref->zone = zone;
2314 zoneref->zone_idx = zone_idx(zone);
2318 * Builds allocation fallback zone lists.
2320 * Add all populated zones of a node to the zonelist.
2322 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2323 int nr_zones, enum zone_type zone_type)
2327 BUG_ON(zone_type >= MAX_NR_ZONES);
2332 zone = pgdat->node_zones + zone_type;
2333 if (populated_zone(zone)) {
2334 zoneref_set_zone(zone,
2335 &zonelist->_zonerefs[nr_zones++]);
2336 check_highest_zone(zone_type);
2339 } while (zone_type);
2346 * 0 = automatic detection of better ordering.
2347 * 1 = order by ([node] distance, -zonetype)
2348 * 2 = order by (-zonetype, [node] distance)
2350 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2351 * the same zonelist. So only NUMA can configure this param.
2353 #define ZONELIST_ORDER_DEFAULT 0
2354 #define ZONELIST_ORDER_NODE 1
2355 #define ZONELIST_ORDER_ZONE 2
2357 /* zonelist order in the kernel.
2358 * set_zonelist_order() will set this to NODE or ZONE.
2360 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2361 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2365 /* The value user specified ....changed by config */
2366 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2367 /* string for sysctl */
2368 #define NUMA_ZONELIST_ORDER_LEN 16
2369 char numa_zonelist_order[16] = "default";
2372 * interface for configure zonelist ordering.
2373 * command line option "numa_zonelist_order"
2374 * = "[dD]efault - default, automatic configuration.
2375 * = "[nN]ode - order by node locality, then by zone within node
2376 * = "[zZ]one - order by zone, then by locality within zone
2379 static int __parse_numa_zonelist_order(char *s)
2381 if (*s == 'd' || *s == 'D') {
2382 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2383 } else if (*s == 'n' || *s == 'N') {
2384 user_zonelist_order = ZONELIST_ORDER_NODE;
2385 } else if (*s == 'z' || *s == 'Z') {
2386 user_zonelist_order = ZONELIST_ORDER_ZONE;
2389 "Ignoring invalid numa_zonelist_order value: "
2396 static __init int setup_numa_zonelist_order(char *s)
2399 return __parse_numa_zonelist_order(s);
2402 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2405 * sysctl handler for numa_zonelist_order
2407 int numa_zonelist_order_handler(ctl_table *table, int write,
2408 void __user *buffer, size_t *length,
2411 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2415 strncpy(saved_string, (char*)table->data,
2416 NUMA_ZONELIST_ORDER_LEN);
2417 ret = proc_dostring(table, write, buffer, length, ppos);
2421 int oldval = user_zonelist_order;
2422 if (__parse_numa_zonelist_order((char*)table->data)) {
2424 * bogus value. restore saved string
2426 strncpy((char*)table->data, saved_string,
2427 NUMA_ZONELIST_ORDER_LEN);
2428 user_zonelist_order = oldval;
2429 } else if (oldval != user_zonelist_order)
2430 build_all_zonelists();
2436 #define MAX_NODE_LOAD (nr_online_nodes)
2437 static int node_load[MAX_NUMNODES];
2440 * find_next_best_node - find the next node that should appear in a given node's fallback list
2441 * @node: node whose fallback list we're appending
2442 * @used_node_mask: nodemask_t of already used nodes
2444 * We use a number of factors to determine which is the next node that should
2445 * appear on a given node's fallback list. The node should not have appeared
2446 * already in @node's fallback list, and it should be the next closest node
2447 * according to the distance array (which contains arbitrary distance values
2448 * from each node to each node in the system), and should also prefer nodes
2449 * with no CPUs, since presumably they'll have very little allocation pressure
2450 * on them otherwise.
2451 * It returns -1 if no node is found.
2453 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2456 int min_val = INT_MAX;
2458 const struct cpumask *tmp = cpumask_of_node(0);
2460 /* Use the local node if we haven't already */
2461 if (!node_isset(node, *used_node_mask)) {
2462 node_set(node, *used_node_mask);
2466 for_each_node_state(n, N_HIGH_MEMORY) {
2468 /* Don't want a node to appear more than once */
2469 if (node_isset(n, *used_node_mask))
2472 /* Use the distance array to find the distance */
2473 val = node_distance(node, n);
2475 /* Penalize nodes under us ("prefer the next node") */
2478 /* Give preference to headless and unused nodes */
2479 tmp = cpumask_of_node(n);
2480 if (!cpumask_empty(tmp))
2481 val += PENALTY_FOR_NODE_WITH_CPUS;
2483 /* Slight preference for less loaded node */
2484 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2485 val += node_load[n];
2487 if (val < min_val) {
2494 node_set(best_node, *used_node_mask);
2501 * Build zonelists ordered by node and zones within node.
2502 * This results in maximum locality--normal zone overflows into local
2503 * DMA zone, if any--but risks exhausting DMA zone.
2505 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2508 struct zonelist *zonelist;
2510 zonelist = &pgdat->node_zonelists[0];
2511 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2513 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2515 zonelist->_zonerefs[j].zone = NULL;
2516 zonelist->_zonerefs[j].zone_idx = 0;
2520 * Build gfp_thisnode zonelists
2522 static void build_thisnode_zonelists(pg_data_t *pgdat)
2525 struct zonelist *zonelist;
2527 zonelist = &pgdat->node_zonelists[1];
2528 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2529 zonelist->_zonerefs[j].zone = NULL;
2530 zonelist->_zonerefs[j].zone_idx = 0;
2534 * Build zonelists ordered by zone and nodes within zones.
2535 * This results in conserving DMA zone[s] until all Normal memory is
2536 * exhausted, but results in overflowing to remote node while memory
2537 * may still exist in local DMA zone.
2539 static int node_order[MAX_NUMNODES];
2541 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2544 int zone_type; /* needs to be signed */
2546 struct zonelist *zonelist;
2548 zonelist = &pgdat->node_zonelists[0];
2550 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2551 for (j = 0; j < nr_nodes; j++) {
2552 node = node_order[j];
2553 z = &NODE_DATA(node)->node_zones[zone_type];
2554 if (populated_zone(z)) {
2556 &zonelist->_zonerefs[pos++]);
2557 check_highest_zone(zone_type);
2561 zonelist->_zonerefs[pos].zone = NULL;
2562 zonelist->_zonerefs[pos].zone_idx = 0;
2565 static int default_zonelist_order(void)
2568 unsigned long low_kmem_size,total_size;
2572 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2573 * If they are really small and used heavily, the system can fall
2574 * into OOM very easily.
2575 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2577 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2580 for_each_online_node(nid) {
2581 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2582 z = &NODE_DATA(nid)->node_zones[zone_type];
2583 if (populated_zone(z)) {
2584 if (zone_type < ZONE_NORMAL)
2585 low_kmem_size += z->present_pages;
2586 total_size += z->present_pages;
2590 if (!low_kmem_size || /* there are no DMA area. */
2591 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2592 return ZONELIST_ORDER_NODE;
2594 * look into each node's config.
2595 * If there is a node whose DMA/DMA32 memory is very big area on
2596 * local memory, NODE_ORDER may be suitable.
2598 average_size = total_size /
2599 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2600 for_each_online_node(nid) {
2603 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2604 z = &NODE_DATA(nid)->node_zones[zone_type];
2605 if (populated_zone(z)) {
2606 if (zone_type < ZONE_NORMAL)
2607 low_kmem_size += z->present_pages;
2608 total_size += z->present_pages;
2611 if (low_kmem_size &&
2612 total_size > average_size && /* ignore small node */
2613 low_kmem_size > total_size * 70/100)
2614 return ZONELIST_ORDER_NODE;
2616 return ZONELIST_ORDER_ZONE;
2619 static void set_zonelist_order(void)
2621 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2622 current_zonelist_order = default_zonelist_order();
2624 current_zonelist_order = user_zonelist_order;
2627 static void build_zonelists(pg_data_t *pgdat)
2631 nodemask_t used_mask;
2632 int local_node, prev_node;
2633 struct zonelist *zonelist;
2634 int order = current_zonelist_order;
2636 /* initialize zonelists */
2637 for (i = 0; i < MAX_ZONELISTS; i++) {
2638 zonelist = pgdat->node_zonelists + i;
2639 zonelist->_zonerefs[0].zone = NULL;
2640 zonelist->_zonerefs[0].zone_idx = 0;
2643 /* NUMA-aware ordering of nodes */
2644 local_node = pgdat->node_id;
2645 load = nr_online_nodes;
2646 prev_node = local_node;
2647 nodes_clear(used_mask);
2649 memset(node_order, 0, sizeof(node_order));
2652 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2653 int distance = node_distance(local_node, node);
2656 * If another node is sufficiently far away then it is better
2657 * to reclaim pages in a zone before going off node.
2659 if (distance > RECLAIM_DISTANCE)
2660 zone_reclaim_mode = 1;
2663 * We don't want to pressure a particular node.
2664 * So adding penalty to the first node in same
2665 * distance group to make it round-robin.
2667 if (distance != node_distance(local_node, prev_node))
2668 node_load[node] = load;
2672 if (order == ZONELIST_ORDER_NODE)
2673 build_zonelists_in_node_order(pgdat, node);
2675 node_order[j++] = node; /* remember order */
2678 if (order == ZONELIST_ORDER_ZONE) {
2679 /* calculate node order -- i.e., DMA last! */
2680 build_zonelists_in_zone_order(pgdat, j);
2683 build_thisnode_zonelists(pgdat);
2686 /* Construct the zonelist performance cache - see further mmzone.h */
2687 static void build_zonelist_cache(pg_data_t *pgdat)
2689 struct zonelist *zonelist;
2690 struct zonelist_cache *zlc;
2693 zonelist = &pgdat->node_zonelists[0];
2694 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2695 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2696 for (z = zonelist->_zonerefs; z->zone; z++)
2697 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2701 #else /* CONFIG_NUMA */
2703 static void set_zonelist_order(void)
2705 current_zonelist_order = ZONELIST_ORDER_ZONE;
2708 static void build_zonelists(pg_data_t *pgdat)
2710 int node, local_node;
2712 struct zonelist *zonelist;
2714 local_node = pgdat->node_id;
2716 zonelist = &pgdat->node_zonelists[0];
2717 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2720 * Now we build the zonelist so that it contains the zones
2721 * of all the other nodes.
2722 * We don't want to pressure a particular node, so when
2723 * building the zones for node N, we make sure that the
2724 * zones coming right after the local ones are those from
2725 * node N+1 (modulo N)
2727 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2728 if (!node_online(node))
2730 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2733 for (node = 0; node < local_node; node++) {
2734 if (!node_online(node))
2736 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2740 zonelist->_zonerefs[j].zone = NULL;
2741 zonelist->_zonerefs[j].zone_idx = 0;
2744 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2745 static void build_zonelist_cache(pg_data_t *pgdat)
2747 pgdat->node_zonelists[0].zlcache_ptr = NULL;
2750 #endif /* CONFIG_NUMA */
2752 /* return values int ....just for stop_machine() */
2753 static int __build_all_zonelists(void *dummy)
2758 memset(node_load, 0, sizeof(node_load));
2760 for_each_online_node(nid) {
2761 pg_data_t *pgdat = NODE_DATA(nid);
2763 build_zonelists(pgdat);
2764 build_zonelist_cache(pgdat);
2769 void build_all_zonelists(void)
2771 set_zonelist_order();
2773 if (system_state == SYSTEM_BOOTING) {
2774 __build_all_zonelists(NULL);
2775 mminit_verify_zonelist();
2776 cpuset_init_current_mems_allowed();
2778 /* we have to stop all cpus to guarantee there is no user
2780 stop_machine(__build_all_zonelists, NULL, NULL);
2781 /* cpuset refresh routine should be here */
2783 vm_total_pages = nr_free_pagecache_pages();
2785 * Disable grouping by mobility if the number of pages in the
2786 * system is too low to allow the mechanism to work. It would be
2787 * more accurate, but expensive to check per-zone. This check is
2788 * made on memory-hotadd so a system can start with mobility
2789 * disabled and enable it later
2791 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
2792 page_group_by_mobility_disabled = 1;
2794 page_group_by_mobility_disabled = 0;
2796 printk("Built %i zonelists in %s order, mobility grouping %s. "
2797 "Total pages: %ld\n",
2799 zonelist_order_name[current_zonelist_order],
2800 page_group_by_mobility_disabled ? "off" : "on",
2803 printk("Policy zone: %s\n", zone_names[policy_zone]);
2808 * Helper functions to size the waitqueue hash table.
2809 * Essentially these want to choose hash table sizes sufficiently
2810 * large so that collisions trying to wait on pages are rare.
2811 * But in fact, the number of active page waitqueues on typical
2812 * systems is ridiculously low, less than 200. So this is even
2813 * conservative, even though it seems large.
2815 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2816 * waitqueues, i.e. the size of the waitq table given the number of pages.
2818 #define PAGES_PER_WAITQUEUE 256
2820 #ifndef CONFIG_MEMORY_HOTPLUG
2821 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2823 unsigned long size = 1;
2825 pages /= PAGES_PER_WAITQUEUE;
2827 while (size < pages)
2831 * Once we have dozens or even hundreds of threads sleeping
2832 * on IO we've got bigger problems than wait queue collision.
2833 * Limit the size of the wait table to a reasonable size.
2835 size = min(size, 4096UL);
2837 return max(size, 4UL);
2841 * A zone's size might be changed by hot-add, so it is not possible to determine
2842 * a suitable size for its wait_table. So we use the maximum size now.
2844 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2846 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2847 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2848 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2850 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2851 * or more by the traditional way. (See above). It equals:
2853 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2854 * ia64(16K page size) : = ( 8G + 4M)byte.
2855 * powerpc (64K page size) : = (32G +16M)byte.
2857 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2864 * This is an integer logarithm so that shifts can be used later
2865 * to extract the more random high bits from the multiplicative
2866 * hash function before the remainder is taken.
2868 static inline unsigned long wait_table_bits(unsigned long size)
2873 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2876 * Check if a pageblock contains reserved pages
2878 static int pageblock_is_reserved(unsigned long start_pfn)
2880 unsigned long end_pfn = start_pfn + pageblock_nr_pages;
2883 for (pfn = start_pfn; pfn < end_pfn; pfn++)
2884 if (PageReserved(pfn_to_page(pfn)))
2890 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2891 * of blocks reserved is based on min_wmark_pages(zone). The memory within
2892 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
2893 * higher will lead to a bigger reserve which will get freed as contiguous
2894 * blocks as reclaim kicks in
2896 static void setup_zone_migrate_reserve(struct zone *zone)
2898 unsigned long start_pfn, pfn, end_pfn;
2900 unsigned long block_migratetype;
2903 /* Get the start pfn, end pfn and the number of blocks to reserve */
2904 start_pfn = zone->zone_start_pfn;
2905 end_pfn = start_pfn + zone->spanned_pages;
2906 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
2910 * Reserve blocks are generally in place to help high-order atomic
2911 * allocations that are short-lived. A min_free_kbytes value that
2912 * would result in more than 2 reserve blocks for atomic allocations
2913 * is assumed to be in place to help anti-fragmentation for the
2914 * future allocation of hugepages at runtime.
2916 reserve = min(2, reserve);
2918 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
2919 if (!pfn_valid(pfn))
2921 page = pfn_to_page(pfn);
2923 /* Watch out for overlapping nodes */
2924 if (page_to_nid(page) != zone_to_nid(zone))
2927 /* Blocks with reserved pages will never free, skip them. */
2928 if (pageblock_is_reserved(pfn))
2931 block_migratetype = get_pageblock_migratetype(page);
2933 /* If this block is reserved, account for it */
2934 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
2939 /* Suitable for reserving if this block is movable */
2940 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
2941 set_pageblock_migratetype(page, MIGRATE_RESERVE);
2942 move_freepages_block(zone, page, MIGRATE_RESERVE);
2948 * If the reserve is met and this is a previous reserved block,
2951 if (block_migratetype == MIGRATE_RESERVE) {
2952 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2953 move_freepages_block(zone, page, MIGRATE_MOVABLE);
2959 * Initially all pages are reserved - free ones are freed
2960 * up by free_all_bootmem() once the early boot process is
2961 * done. Non-atomic initialization, single-pass.
2963 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
2964 unsigned long start_pfn, enum memmap_context context)
2967 unsigned long end_pfn = start_pfn + size;
2971 if (highest_memmap_pfn < end_pfn - 1)
2972 highest_memmap_pfn = end_pfn - 1;
2974 z = &NODE_DATA(nid)->node_zones[zone];
2975 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
2977 * There can be holes in boot-time mem_map[]s
2978 * handed to this function. They do not
2979 * exist on hotplugged memory.
2981 if (context == MEMMAP_EARLY) {
2982 if (!early_pfn_valid(pfn))
2984 if (!early_pfn_in_nid(pfn, nid))
2987 page = pfn_to_page(pfn);
2988 set_page_links(page, zone, nid, pfn);
2989 mminit_verify_page_links(page, zone, nid, pfn);
2990 init_page_count(page);
2991 reset_page_mapcount(page);
2992 SetPageReserved(page);
2994 * Mark the block movable so that blocks are reserved for
2995 * movable at startup. This will force kernel allocations
2996 * to reserve their blocks rather than leaking throughout
2997 * the address space during boot when many long-lived
2998 * kernel allocations are made. Later some blocks near
2999 * the start are marked MIGRATE_RESERVE by
3000 * setup_zone_migrate_reserve()
3002 * bitmap is created for zone's valid pfn range. but memmap
3003 * can be created for invalid pages (for alignment)
3004 * check here not to call set_pageblock_migratetype() against
3007 if ((z->zone_start_pfn <= pfn)
3008 && (pfn < z->zone_start_pfn + z->spanned_pages)
3009 && !(pfn & (pageblock_nr_pages - 1)))
3010 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3012 INIT_LIST_HEAD(&page->lru);
3013 #ifdef WANT_PAGE_VIRTUAL
3014 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3015 if (!is_highmem_idx(zone))
3016 set_page_address(page, __va(pfn << PAGE_SHIFT));
3021 static void __meminit zone_init_free_lists(struct zone *zone)
3024 for_each_migratetype_order(order, t) {
3025 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3026 zone->free_area[order].nr_free = 0;
3030 #ifndef __HAVE_ARCH_MEMMAP_INIT
3031 #define memmap_init(size, nid, zone, start_pfn) \
3032 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3035 static int zone_batchsize(struct zone *zone)
3041 * The per-cpu-pages pools are set to around 1000th of the
3042 * size of the zone. But no more than 1/2 of a meg.
3044 * OK, so we don't know how big the cache is. So guess.
3046 batch = zone->present_pages / 1024;
3047 if (batch * PAGE_SIZE > 512 * 1024)
3048 batch = (512 * 1024) / PAGE_SIZE;
3049 batch /= 4; /* We effectively *= 4 below */
3054 * Clamp the batch to a 2^n - 1 value. Having a power
3055 * of 2 value was found to be more likely to have
3056 * suboptimal cache aliasing properties in some cases.
3058 * For example if 2 tasks are alternately allocating
3059 * batches of pages, one task can end up with a lot
3060 * of pages of one half of the possible page colors
3061 * and the other with pages of the other colors.
3063 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3068 /* The deferral and batching of frees should be suppressed under NOMMU
3071 * The problem is that NOMMU needs to be able to allocate large chunks
3072 * of contiguous memory as there's no hardware page translation to
3073 * assemble apparent contiguous memory from discontiguous pages.
3075 * Queueing large contiguous runs of pages for batching, however,
3076 * causes the pages to actually be freed in smaller chunks. As there
3077 * can be a significant delay between the individual batches being
3078 * recycled, this leads to the once large chunks of space being
3079 * fragmented and becoming unavailable for high-order allocations.
3085 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3087 struct per_cpu_pages *pcp;
3090 memset(p, 0, sizeof(*p));
3094 pcp->high = 6 * batch;
3095 pcp->batch = max(1UL, 1 * batch);
3096 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3097 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3101 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3102 * to the value high for the pageset p.
3105 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3108 struct per_cpu_pages *pcp;
3112 pcp->batch = max(1UL, high/4);
3113 if ((high/4) > (PAGE_SHIFT * 8))
3114 pcp->batch = PAGE_SHIFT * 8;
3120 * Boot pageset table. One per cpu which is going to be used for all
3121 * zones and all nodes. The parameters will be set in such a way
3122 * that an item put on a list will immediately be handed over to
3123 * the buddy list. This is safe since pageset manipulation is done
3124 * with interrupts disabled.
3126 * Some NUMA counter updates may also be caught by the boot pagesets.
3128 * The boot_pagesets must be kept even after bootup is complete for
3129 * unused processors and/or zones. They do play a role for bootstrapping
3130 * hotplugged processors.
3132 * zoneinfo_show() and maybe other functions do
3133 * not check if the processor is online before following the pageset pointer.
3134 * Other parts of the kernel may not check if the zone is available.
3136 static struct per_cpu_pageset boot_pageset[NR_CPUS];
3139 * Dynamically allocate memory for the
3140 * per cpu pageset array in struct zone.
3142 static int __cpuinit process_zones(int cpu)
3144 struct zone *zone, *dzone;
3145 int node = cpu_to_node(cpu);
3147 node_set_state(node, N_CPU); /* this node has a cpu */
3149 for_each_populated_zone(zone) {
3150 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
3152 if (!zone_pcp(zone, cpu))
3155 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
3157 if (percpu_pagelist_fraction)
3158 setup_pagelist_highmark(zone_pcp(zone, cpu),
3159 (zone->present_pages / percpu_pagelist_fraction));
3164 for_each_zone(dzone) {
3165 if (!populated_zone(dzone))
3169 kfree(zone_pcp(dzone, cpu));
3170 zone_pcp(dzone, cpu) = &boot_pageset[cpu];
3175 static inline void free_zone_pagesets(int cpu)
3179 for_each_zone(zone) {
3180 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
3182 /* Free per_cpu_pageset if it is slab allocated */
3183 if (pset != &boot_pageset[cpu])
3185 zone_pcp(zone, cpu) = &boot_pageset[cpu];
3189 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
3190 unsigned long action,
3193 int cpu = (long)hcpu;
3194 int ret = NOTIFY_OK;
3197 case CPU_UP_PREPARE:
3198 case CPU_UP_PREPARE_FROZEN:
3199 if (process_zones(cpu))
3202 case CPU_UP_CANCELED:
3203 case CPU_UP_CANCELED_FROZEN:
3205 case CPU_DEAD_FROZEN:
3206 free_zone_pagesets(cpu);
3214 static struct notifier_block __cpuinitdata pageset_notifier =
3215 { &pageset_cpuup_callback, NULL, 0 };
3217 void __init setup_per_cpu_pageset(void)
3221 /* Initialize per_cpu_pageset for cpu 0.
3222 * A cpuup callback will do this for every cpu
3223 * as it comes online
3225 err = process_zones(smp_processor_id());
3227 register_cpu_notifier(&pageset_notifier);
3232 static noinline __init_refok
3233 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3236 struct pglist_data *pgdat = zone->zone_pgdat;
3240 * The per-page waitqueue mechanism uses hashed waitqueues
3243 zone->wait_table_hash_nr_entries =
3244 wait_table_hash_nr_entries(zone_size_pages);
3245 zone->wait_table_bits =
3246 wait_table_bits(zone->wait_table_hash_nr_entries);
3247 alloc_size = zone->wait_table_hash_nr_entries
3248 * sizeof(wait_queue_head_t);
3250 if (!slab_is_available()) {
3251 zone->wait_table = (wait_queue_head_t *)
3252 alloc_bootmem_node(pgdat, alloc_size);
3255 * This case means that a zone whose size was 0 gets new memory
3256 * via memory hot-add.
3257 * But it may be the case that a new node was hot-added. In
3258 * this case vmalloc() will not be able to use this new node's
3259 * memory - this wait_table must be initialized to use this new
3260 * node itself as well.
3261 * To use this new node's memory, further consideration will be
3264 zone->wait_table = vmalloc(alloc_size);
3266 if (!zone->wait_table)
3269 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3270 init_waitqueue_head(zone->wait_table + i);
3275 static int __zone_pcp_update(void *data)
3277 struct zone *zone = data;
3279 unsigned long batch = zone_batchsize(zone), flags;
3281 for (cpu = 0; cpu < NR_CPUS; cpu++) {
3282 struct per_cpu_pageset *pset;
3283 struct per_cpu_pages *pcp;
3285 pset = zone_pcp(zone, cpu);
3288 local_irq_save(flags);
3289 free_pcppages_bulk(zone, pcp->count, pcp);
3290 setup_pageset(pset, batch);
3291 local_irq_restore(flags);
3296 void zone_pcp_update(struct zone *zone)
3298 stop_machine(__zone_pcp_update, zone, NULL);
3301 static __meminit void zone_pcp_init(struct zone *zone)
3304 unsigned long batch = zone_batchsize(zone);
3306 for (cpu = 0; cpu < NR_CPUS; cpu++) {
3308 /* Early boot. Slab allocator not functional yet */
3309 zone_pcp(zone, cpu) = &boot_pageset[cpu];
3310 setup_pageset(&boot_pageset[cpu],0);
3312 setup_pageset(zone_pcp(zone,cpu), batch);
3315 if (zone->present_pages)
3316 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
3317 zone->name, zone->present_pages, batch);
3320 __meminit int init_currently_empty_zone(struct zone *zone,
3321 unsigned long zone_start_pfn,
3323 enum memmap_context context)
3325 struct pglist_data *pgdat = zone->zone_pgdat;
3327 ret = zone_wait_table_init(zone, size);
3330 pgdat->nr_zones = zone_idx(zone) + 1;
3332 zone->zone_start_pfn = zone_start_pfn;
3334 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3335 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3337 (unsigned long)zone_idx(zone),
3338 zone_start_pfn, (zone_start_pfn + size));
3340 zone_init_free_lists(zone);
3345 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3347 * Basic iterator support. Return the first range of PFNs for a node
3348 * Note: nid == MAX_NUMNODES returns first region regardless of node
3350 static int __meminit first_active_region_index_in_nid(int nid)
3354 for (i = 0; i < nr_nodemap_entries; i++)
3355 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3362 * Basic iterator support. Return the next active range of PFNs for a node
3363 * Note: nid == MAX_NUMNODES returns next region regardless of node
3365 static int __meminit next_active_region_index_in_nid(int index, int nid)
3367 for (index = index + 1; index < nr_nodemap_entries; index++)
3368 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3374 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3376 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3377 * Architectures may implement their own version but if add_active_range()
3378 * was used and there are no special requirements, this is a convenient
3381 int __meminit __early_pfn_to_nid(unsigned long pfn)
3385 for (i = 0; i < nr_nodemap_entries; i++) {
3386 unsigned long start_pfn = early_node_map[i].start_pfn;
3387 unsigned long end_pfn = early_node_map[i].end_pfn;
3389 if (start_pfn <= pfn && pfn < end_pfn)
3390 return early_node_map[i].nid;
3392 /* This is a memory hole */
3395 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3397 int __meminit early_pfn_to_nid(unsigned long pfn)
3401 nid = __early_pfn_to_nid(pfn);
3404 /* just returns 0 */
3408 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3409 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3413 nid = __early_pfn_to_nid(pfn);
3414 if (nid >= 0 && nid != node)
3420 /* Basic iterator support to walk early_node_map[] */
3421 #define for_each_active_range_index_in_nid(i, nid) \
3422 for (i = first_active_region_index_in_nid(nid); i != -1; \
3423 i = next_active_region_index_in_nid(i, nid))
3426 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3427 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3428 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3430 * If an architecture guarantees that all ranges registered with
3431 * add_active_ranges() contain no holes and may be freed, this
3432 * this function may be used instead of calling free_bootmem() manually.
3434 void __init free_bootmem_with_active_regions(int nid,
3435 unsigned long max_low_pfn)
3439 for_each_active_range_index_in_nid(i, nid) {
3440 unsigned long size_pages = 0;
3441 unsigned long end_pfn = early_node_map[i].end_pfn;
3443 if (early_node_map[i].start_pfn >= max_low_pfn)
3446 if (end_pfn > max_low_pfn)
3447 end_pfn = max_low_pfn;
3449 size_pages = end_pfn - early_node_map[i].start_pfn;
3450 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3451 PFN_PHYS(early_node_map[i].start_pfn),
3452 size_pages << PAGE_SHIFT);
3456 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3461 for_each_active_range_index_in_nid(i, nid) {
3462 ret = work_fn(early_node_map[i].start_pfn,
3463 early_node_map[i].end_pfn, data);
3469 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3470 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3472 * If an architecture guarantees that all ranges registered with
3473 * add_active_ranges() contain no holes and may be freed, this
3474 * function may be used instead of calling memory_present() manually.
3476 void __init sparse_memory_present_with_active_regions(int nid)
3480 for_each_active_range_index_in_nid(i, nid)
3481 memory_present(early_node_map[i].nid,
3482 early_node_map[i].start_pfn,
3483 early_node_map[i].end_pfn);
3487 * get_pfn_range_for_nid - Return the start and end page frames for a node
3488 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3489 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3490 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3492 * It returns the start and end page frame of a node based on information
3493 * provided by an arch calling add_active_range(). If called for a node
3494 * with no available memory, a warning is printed and the start and end
3497 void __meminit get_pfn_range_for_nid(unsigned int nid,
3498 unsigned long *start_pfn, unsigned long *end_pfn)
3504 for_each_active_range_index_in_nid(i, nid) {
3505 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3506 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3509 if (*start_pfn == -1UL)
3514 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3515 * assumption is made that zones within a node are ordered in monotonic
3516 * increasing memory addresses so that the "highest" populated zone is used
3518 static void __init find_usable_zone_for_movable(void)
3521 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3522 if (zone_index == ZONE_MOVABLE)
3525 if (arch_zone_highest_possible_pfn[zone_index] >
3526 arch_zone_lowest_possible_pfn[zone_index])
3530 VM_BUG_ON(zone_index == -1);
3531 movable_zone = zone_index;
3535 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3536 * because it is sized independant of architecture. Unlike the other zones,
3537 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3538 * in each node depending on the size of each node and how evenly kernelcore
3539 * is distributed. This helper function adjusts the zone ranges
3540 * provided by the architecture for a given node by using the end of the
3541 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3542 * zones within a node are in order of monotonic increases memory addresses
3544 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3545 unsigned long zone_type,
3546 unsigned long node_start_pfn,
3547 unsigned long node_end_pfn,
3548 unsigned long *zone_start_pfn,
3549 unsigned long *zone_end_pfn)
3551 /* Only adjust if ZONE_MOVABLE is on this node */
3552 if (zone_movable_pfn[nid]) {
3553 /* Size ZONE_MOVABLE */
3554 if (zone_type == ZONE_MOVABLE) {
3555 *zone_start_pfn = zone_movable_pfn[nid];
3556 *zone_end_pfn = min(node_end_pfn,
3557 arch_zone_highest_possible_pfn[movable_zone]);
3559 /* Adjust for ZONE_MOVABLE starting within this range */
3560 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3561 *zone_end_pfn > zone_movable_pfn[nid]) {
3562 *zone_end_pfn = zone_movable_pfn[nid];
3564 /* Check if this whole range is within ZONE_MOVABLE */
3565 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3566 *zone_start_pfn = *zone_end_pfn;
3571 * Return the number of pages a zone spans in a node, including holes
3572 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3574 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3575 unsigned long zone_type,
3576 unsigned long *ignored)
3578 unsigned long node_start_pfn, node_end_pfn;
3579 unsigned long zone_start_pfn, zone_end_pfn;
3581 /* Get the start and end of the node and zone */
3582 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3583 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3584 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3585 adjust_zone_range_for_zone_movable(nid, zone_type,
3586 node_start_pfn, node_end_pfn,
3587 &zone_start_pfn, &zone_end_pfn);
3589 /* Check that this node has pages within the zone's required range */
3590 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3593 /* Move the zone boundaries inside the node if necessary */
3594 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3595 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3597 /* Return the spanned pages */
3598 return zone_end_pfn - zone_start_pfn;
3602 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3603 * then all holes in the requested range will be accounted for.
3605 static unsigned long __meminit __absent_pages_in_range(int nid,
3606 unsigned long range_start_pfn,
3607 unsigned long range_end_pfn)
3610 unsigned long prev_end_pfn = 0, hole_pages = 0;
3611 unsigned long start_pfn;
3613 /* Find the end_pfn of the first active range of pfns in the node */
3614 i = first_active_region_index_in_nid(nid);
3618 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3620 /* Account for ranges before physical memory on this node */
3621 if (early_node_map[i].start_pfn > range_start_pfn)
3622 hole_pages = prev_end_pfn - range_start_pfn;
3624 /* Find all holes for the zone within the node */
3625 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3627 /* No need to continue if prev_end_pfn is outside the zone */
3628 if (prev_end_pfn >= range_end_pfn)
3631 /* Make sure the end of the zone is not within the hole */
3632 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3633 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3635 /* Update the hole size cound and move on */
3636 if (start_pfn > range_start_pfn) {
3637 BUG_ON(prev_end_pfn > start_pfn);
3638 hole_pages += start_pfn - prev_end_pfn;
3640 prev_end_pfn = early_node_map[i].end_pfn;
3643 /* Account for ranges past physical memory on this node */
3644 if (range_end_pfn > prev_end_pfn)
3645 hole_pages += range_end_pfn -
3646 max(range_start_pfn, prev_end_pfn);
3652 * absent_pages_in_range - Return number of page frames in holes within a range
3653 * @start_pfn: The start PFN to start searching for holes
3654 * @end_pfn: The end PFN to stop searching for holes
3656 * It returns the number of pages frames in memory holes within a range.
3658 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3659 unsigned long end_pfn)
3661 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3664 /* Return the number of page frames in holes in a zone on a node */
3665 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3666 unsigned long zone_type,
3667 unsigned long *ignored)
3669 unsigned long node_start_pfn, node_end_pfn;
3670 unsigned long zone_start_pfn, zone_end_pfn;
3672 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3673 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3675 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3678 adjust_zone_range_for_zone_movable(nid, zone_type,
3679 node_start_pfn, node_end_pfn,
3680 &zone_start_pfn, &zone_end_pfn);
3681 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3685 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3686 unsigned long zone_type,
3687 unsigned long *zones_size)
3689 return zones_size[zone_type];
3692 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3693 unsigned long zone_type,
3694 unsigned long *zholes_size)
3699 return zholes_size[zone_type];
3704 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3705 unsigned long *zones_size, unsigned long *zholes_size)
3707 unsigned long realtotalpages, totalpages = 0;
3710 for (i = 0; i < MAX_NR_ZONES; i++)
3711 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3713 pgdat->node_spanned_pages = totalpages;
3715 realtotalpages = totalpages;
3716 for (i = 0; i < MAX_NR_ZONES; i++)
3718 zone_absent_pages_in_node(pgdat->node_id, i,
3720 pgdat->node_present_pages = realtotalpages;
3721 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3725 #ifndef CONFIG_SPARSEMEM
3727 * Calculate the size of the zone->blockflags rounded to an unsigned long
3728 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3729 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3730 * round what is now in bits to nearest long in bits, then return it in
3733 static unsigned long __init usemap_size(unsigned long zonesize)
3735 unsigned long usemapsize;
3737 usemapsize = roundup(zonesize, pageblock_nr_pages);
3738 usemapsize = usemapsize >> pageblock_order;
3739 usemapsize *= NR_PAGEBLOCK_BITS;
3740 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
3742 return usemapsize / 8;
3745 static void __init setup_usemap(struct pglist_data *pgdat,
3746 struct zone *zone, unsigned long zonesize)
3748 unsigned long usemapsize = usemap_size(zonesize);
3749 zone->pageblock_flags = NULL;
3751 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
3754 static void inline setup_usemap(struct pglist_data *pgdat,
3755 struct zone *zone, unsigned long zonesize) {}
3756 #endif /* CONFIG_SPARSEMEM */
3758 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3760 /* Return a sensible default order for the pageblock size. */
3761 static inline int pageblock_default_order(void)
3763 if (HPAGE_SHIFT > PAGE_SHIFT)
3764 return HUGETLB_PAGE_ORDER;
3769 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3770 static inline void __init set_pageblock_order(unsigned int order)
3772 /* Check that pageblock_nr_pages has not already been setup */
3773 if (pageblock_order)
3777 * Assume the largest contiguous order of interest is a huge page.
3778 * This value may be variable depending on boot parameters on IA64
3780 pageblock_order = order;
3782 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3785 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3786 * and pageblock_default_order() are unused as pageblock_order is set
3787 * at compile-time. See include/linux/pageblock-flags.h for the values of
3788 * pageblock_order based on the kernel config
3790 static inline int pageblock_default_order(unsigned int order)
3794 #define set_pageblock_order(x) do {} while (0)
3796 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3799 * Set up the zone data structures:
3800 * - mark all pages reserved
3801 * - mark all memory queues empty
3802 * - clear the memory bitmaps
3804 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
3805 unsigned long *zones_size, unsigned long *zholes_size)
3808 int nid = pgdat->node_id;
3809 unsigned long zone_start_pfn = pgdat->node_start_pfn;
3812 pgdat_resize_init(pgdat);
3813 pgdat->nr_zones = 0;
3814 init_waitqueue_head(&pgdat->kswapd_wait);
3815 pgdat->kswapd_max_order = 0;
3816 pgdat_page_cgroup_init(pgdat);
3818 for (j = 0; j < MAX_NR_ZONES; j++) {
3819 struct zone *zone = pgdat->node_zones + j;
3820 unsigned long size, realsize, memmap_pages;
3823 size = zone_spanned_pages_in_node(nid, j, zones_size);
3824 realsize = size - zone_absent_pages_in_node(nid, j,
3828 * Adjust realsize so that it accounts for how much memory
3829 * is used by this zone for memmap. This affects the watermark
3830 * and per-cpu initialisations
3833 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
3834 if (realsize >= memmap_pages) {
3835 realsize -= memmap_pages;
3838 " %s zone: %lu pages used for memmap\n",
3839 zone_names[j], memmap_pages);
3842 " %s zone: %lu pages exceeds realsize %lu\n",
3843 zone_names[j], memmap_pages, realsize);
3845 /* Account for reserved pages */
3846 if (j == 0 && realsize > dma_reserve) {
3847 realsize -= dma_reserve;
3848 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
3849 zone_names[0], dma_reserve);
3852 if (!is_highmem_idx(j))
3853 nr_kernel_pages += realsize;
3854 nr_all_pages += realsize;
3856 zone->spanned_pages = size;
3857 zone->present_pages = realsize;
3860 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
3862 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
3864 zone->name = zone_names[j];
3865 spin_lock_init(&zone->lock);
3866 spin_lock_init(&zone->lru_lock);
3867 zone_seqlock_init(zone);
3868 zone->zone_pgdat = pgdat;
3870 zone->prev_priority = DEF_PRIORITY;
3872 zone_pcp_init(zone);
3874 INIT_LIST_HEAD(&zone->lru[l].list);
3875 zone->reclaim_stat.nr_saved_scan[l] = 0;
3877 zone->reclaim_stat.recent_rotated[0] = 0;
3878 zone->reclaim_stat.recent_rotated[1] = 0;
3879 zone->reclaim_stat.recent_scanned[0] = 0;
3880 zone->reclaim_stat.recent_scanned[1] = 0;
3881 zap_zone_vm_stats(zone);
3886 set_pageblock_order(pageblock_default_order());
3887 setup_usemap(pgdat, zone, size);
3888 ret = init_currently_empty_zone(zone, zone_start_pfn,
3889 size, MEMMAP_EARLY);
3891 memmap_init(size, nid, j, zone_start_pfn);
3892 zone_start_pfn += size;
3896 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
3898 /* Skip empty nodes */
3899 if (!pgdat->node_spanned_pages)
3902 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3903 /* ia64 gets its own node_mem_map, before this, without bootmem */
3904 if (!pgdat->node_mem_map) {
3905 unsigned long size, start, end;
3909 * The zone's endpoints aren't required to be MAX_ORDER
3910 * aligned but the node_mem_map endpoints must be in order
3911 * for the buddy allocator to function correctly.
3913 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
3914 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
3915 end = ALIGN(end, MAX_ORDER_NR_PAGES);
3916 size = (end - start) * sizeof(struct page);
3917 map = alloc_remap(pgdat->node_id, size);
3919 map = alloc_bootmem_node(pgdat, size);
3920 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
3922 #ifndef CONFIG_NEED_MULTIPLE_NODES
3924 * With no DISCONTIG, the global mem_map is just set as node 0's
3926 if (pgdat == NODE_DATA(0)) {
3927 mem_map = NODE_DATA(0)->node_mem_map;
3928 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3929 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
3930 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
3931 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3934 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3937 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
3938 unsigned long node_start_pfn, unsigned long *zholes_size)
3940 pg_data_t *pgdat = NODE_DATA(nid);
3942 pgdat->node_id = nid;
3943 pgdat->node_start_pfn = node_start_pfn;
3944 calculate_node_totalpages(pgdat, zones_size, zholes_size);
3946 alloc_node_mem_map(pgdat);
3947 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3948 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3949 nid, (unsigned long)pgdat,
3950 (unsigned long)pgdat->node_mem_map);
3953 free_area_init_core(pgdat, zones_size, zholes_size);
3956 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3958 #if MAX_NUMNODES > 1
3960 * Figure out the number of possible node ids.
3962 static void __init setup_nr_node_ids(void)
3965 unsigned int highest = 0;
3967 for_each_node_mask(node, node_possible_map)
3969 nr_node_ids = highest + 1;
3972 static inline void setup_nr_node_ids(void)
3978 * add_active_range - Register a range of PFNs backed by physical memory
3979 * @nid: The node ID the range resides on
3980 * @start_pfn: The start PFN of the available physical memory
3981 * @end_pfn: The end PFN of the available physical memory
3983 * These ranges are stored in an early_node_map[] and later used by
3984 * free_area_init_nodes() to calculate zone sizes and holes. If the
3985 * range spans a memory hole, it is up to the architecture to ensure
3986 * the memory is not freed by the bootmem allocator. If possible
3987 * the range being registered will be merged with existing ranges.
3989 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
3990 unsigned long end_pfn)
3994 mminit_dprintk(MMINIT_TRACE, "memory_register",
3995 "Entering add_active_range(%d, %#lx, %#lx) "
3996 "%d entries of %d used\n",
3997 nid, start_pfn, end_pfn,
3998 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4000 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4002 /* Merge with existing active regions if possible */
4003 for (i = 0; i < nr_nodemap_entries; i++) {
4004 if (early_node_map[i].nid != nid)
4007 /* Skip if an existing region covers this new one */
4008 if (start_pfn >= early_node_map[i].start_pfn &&
4009 end_pfn <= early_node_map[i].end_pfn)
4012 /* Merge forward if suitable */
4013 if (start_pfn <= early_node_map[i].end_pfn &&
4014 end_pfn > early_node_map[i].end_pfn) {
4015 early_node_map[i].end_pfn = end_pfn;
4019 /* Merge backward if suitable */
4020 if (start_pfn < early_node_map[i].end_pfn &&
4021 end_pfn >= early_node_map[i].start_pfn) {
4022 early_node_map[i].start_pfn = start_pfn;
4027 /* Check that early_node_map is large enough */
4028 if (i >= MAX_ACTIVE_REGIONS) {
4029 printk(KERN_CRIT "More than %d memory regions, truncating\n",
4030 MAX_ACTIVE_REGIONS);
4034 early_node_map[i].nid = nid;
4035 early_node_map[i].start_pfn = start_pfn;
4036 early_node_map[i].end_pfn = end_pfn;
4037 nr_nodemap_entries = i + 1;
4041 * remove_active_range - Shrink an existing registered range of PFNs
4042 * @nid: The node id the range is on that should be shrunk
4043 * @start_pfn: The new PFN of the range
4044 * @end_pfn: The new PFN of the range
4046 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4047 * The map is kept near the end physical page range that has already been
4048 * registered. This function allows an arch to shrink an existing registered
4051 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4052 unsigned long end_pfn)
4057 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4058 nid, start_pfn, end_pfn);
4060 /* Find the old active region end and shrink */
4061 for_each_active_range_index_in_nid(i, nid) {
4062 if (early_node_map[i].start_pfn >= start_pfn &&
4063 early_node_map[i].end_pfn <= end_pfn) {
4065 early_node_map[i].start_pfn = 0;
4066 early_node_map[i].end_pfn = 0;
4070 if (early_node_map[i].start_pfn < start_pfn &&
4071 early_node_map[i].end_pfn > start_pfn) {
4072 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4073 early_node_map[i].end_pfn = start_pfn;
4074 if (temp_end_pfn > end_pfn)
4075 add_active_range(nid, end_pfn, temp_end_pfn);
4078 if (early_node_map[i].start_pfn >= start_pfn &&
4079 early_node_map[i].end_pfn > end_pfn &&
4080 early_node_map[i].start_pfn < end_pfn) {
4081 early_node_map[i].start_pfn = end_pfn;
4089 /* remove the blank ones */
4090 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4091 if (early_node_map[i].nid != nid)
4093 if (early_node_map[i].end_pfn)
4095 /* we found it, get rid of it */
4096 for (j = i; j < nr_nodemap_entries - 1; j++)
4097 memcpy(&early_node_map[j], &early_node_map[j+1],
4098 sizeof(early_node_map[j]));
4099 j = nr_nodemap_entries - 1;
4100 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4101 nr_nodemap_entries--;
4106 * remove_all_active_ranges - Remove all currently registered regions
4108 * During discovery, it may be found that a table like SRAT is invalid
4109 * and an alternative discovery method must be used. This function removes
4110 * all currently registered regions.
4112 void __init remove_all_active_ranges(void)
4114 memset(early_node_map, 0, sizeof(early_node_map));
4115 nr_nodemap_entries = 0;
4118 /* Compare two active node_active_regions */
4119 static int __init cmp_node_active_region(const void *a, const void *b)
4121 struct node_active_region *arange = (struct node_active_region *)a;
4122 struct node_active_region *brange = (struct node_active_region *)b;
4124 /* Done this way to avoid overflows */
4125 if (arange->start_pfn > brange->start_pfn)
4127 if (arange->start_pfn < brange->start_pfn)
4133 /* sort the node_map by start_pfn */
4134 static void __init sort_node_map(void)
4136 sort(early_node_map, (size_t)nr_nodemap_entries,
4137 sizeof(struct node_active_region),
4138 cmp_node_active_region, NULL);
4141 /* Find the lowest pfn for a node */
4142 static unsigned long __init find_min_pfn_for_node(int nid)
4145 unsigned long min_pfn = ULONG_MAX;
4147 /* Assuming a sorted map, the first range found has the starting pfn */
4148 for_each_active_range_index_in_nid(i, nid)
4149 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4151 if (min_pfn == ULONG_MAX) {
4153 "Could not find start_pfn for node %d\n", nid);
4161 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4163 * It returns the minimum PFN based on information provided via
4164 * add_active_range().
4166 unsigned long __init find_min_pfn_with_active_regions(void)
4168 return find_min_pfn_for_node(MAX_NUMNODES);
4172 * early_calculate_totalpages()
4173 * Sum pages in active regions for movable zone.
4174 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4176 static unsigned long __init early_calculate_totalpages(void)
4179 unsigned long totalpages = 0;
4181 for (i = 0; i < nr_nodemap_entries; i++) {
4182 unsigned long pages = early_node_map[i].end_pfn -
4183 early_node_map[i].start_pfn;
4184 totalpages += pages;
4186 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4192 * Find the PFN the Movable zone begins in each node. Kernel memory
4193 * is spread evenly between nodes as long as the nodes have enough
4194 * memory. When they don't, some nodes will have more kernelcore than
4197 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4200 unsigned long usable_startpfn;
4201 unsigned long kernelcore_node, kernelcore_remaining;
4202 /* save the state before borrow the nodemask */
4203 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4204 unsigned long totalpages = early_calculate_totalpages();
4205 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4208 * If movablecore was specified, calculate what size of
4209 * kernelcore that corresponds so that memory usable for
4210 * any allocation type is evenly spread. If both kernelcore
4211 * and movablecore are specified, then the value of kernelcore
4212 * will be used for required_kernelcore if it's greater than
4213 * what movablecore would have allowed.
4215 if (required_movablecore) {
4216 unsigned long corepages;
4219 * Round-up so that ZONE_MOVABLE is at least as large as what
4220 * was requested by the user
4222 required_movablecore =
4223 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4224 corepages = totalpages - required_movablecore;
4226 required_kernelcore = max(required_kernelcore, corepages);
4229 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4230 if (!required_kernelcore)
4233 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4234 find_usable_zone_for_movable();
4235 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4238 /* Spread kernelcore memory as evenly as possible throughout nodes */
4239 kernelcore_node = required_kernelcore / usable_nodes;
4240 for_each_node_state(nid, N_HIGH_MEMORY) {
4242 * Recalculate kernelcore_node if the division per node
4243 * now exceeds what is necessary to satisfy the requested
4244 * amount of memory for the kernel
4246 if (required_kernelcore < kernelcore_node)
4247 kernelcore_node = required_kernelcore / usable_nodes;
4250 * As the map is walked, we track how much memory is usable
4251 * by the kernel using kernelcore_remaining. When it is
4252 * 0, the rest of the node is usable by ZONE_MOVABLE
4254 kernelcore_remaining = kernelcore_node;
4256 /* Go through each range of PFNs within this node */
4257 for_each_active_range_index_in_nid(i, nid) {
4258 unsigned long start_pfn, end_pfn;
4259 unsigned long size_pages;
4261 start_pfn = max(early_node_map[i].start_pfn,
4262 zone_movable_pfn[nid]);
4263 end_pfn = early_node_map[i].end_pfn;
4264 if (start_pfn >= end_pfn)
4267 /* Account for what is only usable for kernelcore */
4268 if (start_pfn < usable_startpfn) {
4269 unsigned long kernel_pages;
4270 kernel_pages = min(end_pfn, usable_startpfn)
4273 kernelcore_remaining -= min(kernel_pages,
4274 kernelcore_remaining);
4275 required_kernelcore -= min(kernel_pages,
4276 required_kernelcore);
4278 /* Continue if range is now fully accounted */
4279 if (end_pfn <= usable_startpfn) {
4282 * Push zone_movable_pfn to the end so
4283 * that if we have to rebalance
4284 * kernelcore across nodes, we will
4285 * not double account here
4287 zone_movable_pfn[nid] = end_pfn;
4290 start_pfn = usable_startpfn;
4294 * The usable PFN range for ZONE_MOVABLE is from
4295 * start_pfn->end_pfn. Calculate size_pages as the
4296 * number of pages used as kernelcore
4298 size_pages = end_pfn - start_pfn;
4299 if (size_pages > kernelcore_remaining)
4300 size_pages = kernelcore_remaining;
4301 zone_movable_pfn[nid] = start_pfn + size_pages;
4304 * Some kernelcore has been met, update counts and
4305 * break if the kernelcore for this node has been
4308 required_kernelcore -= min(required_kernelcore,
4310 kernelcore_remaining -= size_pages;
4311 if (!kernelcore_remaining)
4317 * If there is still required_kernelcore, we do another pass with one
4318 * less node in the count. This will push zone_movable_pfn[nid] further
4319 * along on the nodes that still have memory until kernelcore is
4323 if (usable_nodes && required_kernelcore > usable_nodes)
4326 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4327 for (nid = 0; nid < MAX_NUMNODES; nid++)
4328 zone_movable_pfn[nid] =
4329 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4332 /* restore the node_state */
4333 node_states[N_HIGH_MEMORY] = saved_node_state;
4336 /* Any regular memory on that node ? */
4337 static void check_for_regular_memory(pg_data_t *pgdat)
4339 #ifdef CONFIG_HIGHMEM
4340 enum zone_type zone_type;
4342 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4343 struct zone *zone = &pgdat->node_zones[zone_type];
4344 if (zone->present_pages)
4345 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4351 * free_area_init_nodes - Initialise all pg_data_t and zone data
4352 * @max_zone_pfn: an array of max PFNs for each zone
4354 * This will call free_area_init_node() for each active node in the system.
4355 * Using the page ranges provided by add_active_range(), the size of each
4356 * zone in each node and their holes is calculated. If the maximum PFN
4357 * between two adjacent zones match, it is assumed that the zone is empty.
4358 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4359 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4360 * starts where the previous one ended. For example, ZONE_DMA32 starts
4361 * at arch_max_dma_pfn.
4363 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4368 /* Sort early_node_map as initialisation assumes it is sorted */
4371 /* Record where the zone boundaries are */
4372 memset(arch_zone_lowest_possible_pfn, 0,
4373 sizeof(arch_zone_lowest_possible_pfn));
4374 memset(arch_zone_highest_possible_pfn, 0,
4375 sizeof(arch_zone_highest_possible_pfn));
4376 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4377 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4378 for (i = 1; i < MAX_NR_ZONES; i++) {
4379 if (i == ZONE_MOVABLE)
4381 arch_zone_lowest_possible_pfn[i] =
4382 arch_zone_highest_possible_pfn[i-1];
4383 arch_zone_highest_possible_pfn[i] =
4384 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4386 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4387 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4389 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4390 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4391 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4393 /* Print out the zone ranges */
4394 printk("Zone PFN ranges:\n");
4395 for (i = 0; i < MAX_NR_ZONES; i++) {
4396 if (i == ZONE_MOVABLE)
4398 printk(" %-8s %0#10lx -> %0#10lx\n",
4400 arch_zone_lowest_possible_pfn[i],
4401 arch_zone_highest_possible_pfn[i]);
4404 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4405 printk("Movable zone start PFN for each node\n");
4406 for (i = 0; i < MAX_NUMNODES; i++) {
4407 if (zone_movable_pfn[i])
4408 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4411 /* Print out the early_node_map[] */
4412 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4413 for (i = 0; i < nr_nodemap_entries; i++)
4414 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4415 early_node_map[i].start_pfn,
4416 early_node_map[i].end_pfn);
4418 /* Initialise every node */
4419 mminit_verify_pageflags_layout();
4420 setup_nr_node_ids();
4421 for_each_online_node(nid) {
4422 pg_data_t *pgdat = NODE_DATA(nid);
4423 free_area_init_node(nid, NULL,
4424 find_min_pfn_for_node(nid), NULL);
4426 /* Any memory on that node */
4427 if (pgdat->node_present_pages)
4428 node_set_state(nid, N_HIGH_MEMORY);
4429 check_for_regular_memory(pgdat);
4433 static int __init cmdline_parse_core(char *p, unsigned long *core)
4435 unsigned long long coremem;
4439 coremem = memparse(p, &p);
4440 *core = coremem >> PAGE_SHIFT;
4442 /* Paranoid check that UL is enough for the coremem value */
4443 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4449 * kernelcore=size sets the amount of memory for use for allocations that
4450 * cannot be reclaimed or migrated.
4452 static int __init cmdline_parse_kernelcore(char *p)
4454 return cmdline_parse_core(p, &required_kernelcore);
4458 * movablecore=size sets the amount of memory for use for allocations that
4459 * can be reclaimed or migrated.
4461 static int __init cmdline_parse_movablecore(char *p)
4463 return cmdline_parse_core(p, &required_movablecore);
4466 early_param("kernelcore", cmdline_parse_kernelcore);
4467 early_param("movablecore", cmdline_parse_movablecore);
4469 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4472 * set_dma_reserve - set the specified number of pages reserved in the first zone
4473 * @new_dma_reserve: The number of pages to mark reserved
4475 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4476 * In the DMA zone, a significant percentage may be consumed by kernel image
4477 * and other unfreeable allocations which can skew the watermarks badly. This
4478 * function may optionally be used to account for unfreeable pages in the
4479 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4480 * smaller per-cpu batchsize.
4482 void __init set_dma_reserve(unsigned long new_dma_reserve)
4484 dma_reserve = new_dma_reserve;
4487 #ifndef CONFIG_NEED_MULTIPLE_NODES
4488 struct pglist_data __refdata contig_page_data = { .bdata = &bootmem_node_data[0] };
4489 EXPORT_SYMBOL(contig_page_data);
4492 void __init free_area_init(unsigned long *zones_size)
4494 free_area_init_node(0, zones_size,
4495 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4498 static int page_alloc_cpu_notify(struct notifier_block *self,
4499 unsigned long action, void *hcpu)
4501 int cpu = (unsigned long)hcpu;
4503 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4507 * Spill the event counters of the dead processor
4508 * into the current processors event counters.
4509 * This artificially elevates the count of the current
4512 vm_events_fold_cpu(cpu);
4515 * Zero the differential counters of the dead processor
4516 * so that the vm statistics are consistent.
4518 * This is only okay since the processor is dead and cannot
4519 * race with what we are doing.
4521 refresh_cpu_vm_stats(cpu);
4526 void __init page_alloc_init(void)
4528 hotcpu_notifier(page_alloc_cpu_notify, 0);
4532 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4533 * or min_free_kbytes changes.
4535 static void calculate_totalreserve_pages(void)
4537 struct pglist_data *pgdat;
4538 unsigned long reserve_pages = 0;
4539 enum zone_type i, j;
4541 for_each_online_pgdat(pgdat) {
4542 for (i = 0; i < MAX_NR_ZONES; i++) {
4543 struct zone *zone = pgdat->node_zones + i;
4544 unsigned long max = 0;
4546 /* Find valid and maximum lowmem_reserve in the zone */
4547 for (j = i; j < MAX_NR_ZONES; j++) {
4548 if (zone->lowmem_reserve[j] > max)
4549 max = zone->lowmem_reserve[j];
4552 /* we treat the high watermark as reserved pages. */
4553 max += high_wmark_pages(zone);
4555 if (max > zone->present_pages)
4556 max = zone->present_pages;
4557 reserve_pages += max;
4560 totalreserve_pages = reserve_pages;
4564 * setup_per_zone_lowmem_reserve - called whenever
4565 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4566 * has a correct pages reserved value, so an adequate number of
4567 * pages are left in the zone after a successful __alloc_pages().
4569 static void setup_per_zone_lowmem_reserve(void)
4571 struct pglist_data *pgdat;
4572 enum zone_type j, idx;
4574 for_each_online_pgdat(pgdat) {
4575 for (j = 0; j < MAX_NR_ZONES; j++) {
4576 struct zone *zone = pgdat->node_zones + j;
4577 unsigned long present_pages = zone->present_pages;
4579 zone->lowmem_reserve[j] = 0;
4583 struct zone *lower_zone;
4587 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4588 sysctl_lowmem_reserve_ratio[idx] = 1;
4590 lower_zone = pgdat->node_zones + idx;
4591 lower_zone->lowmem_reserve[j] = present_pages /
4592 sysctl_lowmem_reserve_ratio[idx];
4593 present_pages += lower_zone->present_pages;
4598 /* update totalreserve_pages */
4599 calculate_totalreserve_pages();
4603 * setup_per_zone_wmarks - called when min_free_kbytes changes
4604 * or when memory is hot-{added|removed}
4606 * Ensures that the watermark[min,low,high] values for each zone are set
4607 * correctly with respect to min_free_kbytes.
4609 void setup_per_zone_wmarks(void)
4611 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4612 unsigned long lowmem_pages = 0;
4614 unsigned long flags;
4616 /* Calculate total number of !ZONE_HIGHMEM pages */
4617 for_each_zone(zone) {
4618 if (!is_highmem(zone))
4619 lowmem_pages += zone->present_pages;
4622 for_each_zone(zone) {
4625 spin_lock_irqsave(&zone->lock, flags);
4626 tmp = (u64)pages_min * zone->present_pages;
4627 do_div(tmp, lowmem_pages);
4628 if (is_highmem(zone)) {
4630 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4631 * need highmem pages, so cap pages_min to a small
4634 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4635 * deltas controls asynch page reclaim, and so should
4636 * not be capped for highmem.
4640 min_pages = zone->present_pages / 1024;
4641 if (min_pages < SWAP_CLUSTER_MAX)
4642 min_pages = SWAP_CLUSTER_MAX;
4643 if (min_pages > 128)
4645 zone->watermark[WMARK_MIN] = min_pages;
4648 * If it's a lowmem zone, reserve a number of pages
4649 * proportionate to the zone's size.
4651 zone->watermark[WMARK_MIN] = tmp;
4654 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
4655 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
4656 setup_zone_migrate_reserve(zone);
4657 spin_unlock_irqrestore(&zone->lock, flags);
4660 /* update totalreserve_pages */
4661 calculate_totalreserve_pages();
4665 * The inactive anon list should be small enough that the VM never has to
4666 * do too much work, but large enough that each inactive page has a chance
4667 * to be referenced again before it is swapped out.
4669 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4670 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4671 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4672 * the anonymous pages are kept on the inactive list.
4675 * memory ratio inactive anon
4676 * -------------------------------------
4685 void calculate_zone_inactive_ratio(struct zone *zone)
4687 unsigned int gb, ratio;
4689 /* Zone size in gigabytes */
4690 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4692 ratio = int_sqrt(10 * gb);
4696 zone->inactive_ratio = ratio;
4699 static void __init setup_per_zone_inactive_ratio(void)
4704 calculate_zone_inactive_ratio(zone);
4708 * Initialise min_free_kbytes.
4710 * For small machines we want it small (128k min). For large machines
4711 * we want it large (64MB max). But it is not linear, because network
4712 * bandwidth does not increase linearly with machine size. We use
4714 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4715 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4731 static int __init init_per_zone_wmark_min(void)
4733 unsigned long lowmem_kbytes;
4735 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4737 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4738 if (min_free_kbytes < 128)
4739 min_free_kbytes = 128;
4740 if (min_free_kbytes > 65536)
4741 min_free_kbytes = 65536;
4742 setup_per_zone_wmarks();
4743 setup_per_zone_lowmem_reserve();
4744 setup_per_zone_inactive_ratio();
4747 module_init(init_per_zone_wmark_min)
4750 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4751 * that we can call two helper functions whenever min_free_kbytes
4754 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
4755 void __user *buffer, size_t *length, loff_t *ppos)
4757 proc_dointvec(table, write, buffer, length, ppos);
4759 setup_per_zone_wmarks();
4764 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
4765 void __user *buffer, size_t *length, loff_t *ppos)
4770 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
4775 zone->min_unmapped_pages = (zone->present_pages *
4776 sysctl_min_unmapped_ratio) / 100;
4780 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
4781 void __user *buffer, size_t *length, loff_t *ppos)
4786 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
4791 zone->min_slab_pages = (zone->present_pages *
4792 sysctl_min_slab_ratio) / 100;
4798 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4799 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4800 * whenever sysctl_lowmem_reserve_ratio changes.
4802 * The reserve ratio obviously has absolutely no relation with the
4803 * minimum watermarks. The lowmem reserve ratio can only make sense
4804 * if in function of the boot time zone sizes.
4806 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
4807 void __user *buffer, size_t *length, loff_t *ppos)
4809 proc_dointvec_minmax(table, write, buffer, length, ppos);
4810 setup_per_zone_lowmem_reserve();
4815 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4816 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4817 * can have before it gets flushed back to buddy allocator.
4820 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
4821 void __user *buffer, size_t *length, loff_t *ppos)
4827 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
4828 if (!write || (ret == -EINVAL))
4830 for_each_populated_zone(zone) {
4831 for_each_online_cpu(cpu) {
4833 high = zone->present_pages / percpu_pagelist_fraction;
4834 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
4840 int hashdist = HASHDIST_DEFAULT;
4843 static int __init set_hashdist(char *str)
4847 hashdist = simple_strtoul(str, &str, 0);
4850 __setup("hashdist=", set_hashdist);
4854 * allocate a large system hash table from bootmem
4855 * - it is assumed that the hash table must contain an exact power-of-2
4856 * quantity of entries
4857 * - limit is the number of hash buckets, not the total allocation size
4859 void *__init alloc_large_system_hash(const char *tablename,
4860 unsigned long bucketsize,
4861 unsigned long numentries,
4864 unsigned int *_hash_shift,
4865 unsigned int *_hash_mask,
4866 unsigned long limit)
4868 unsigned long long max = limit;
4869 unsigned long log2qty, size;
4872 /* allow the kernel cmdline to have a say */
4874 /* round applicable memory size up to nearest megabyte */
4875 numentries = nr_kernel_pages;
4876 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
4877 numentries >>= 20 - PAGE_SHIFT;
4878 numentries <<= 20 - PAGE_SHIFT;
4880 /* limit to 1 bucket per 2^scale bytes of low memory */
4881 if (scale > PAGE_SHIFT)
4882 numentries >>= (scale - PAGE_SHIFT);
4884 numentries <<= (PAGE_SHIFT - scale);
4886 /* Make sure we've got at least a 0-order allocation.. */
4887 if (unlikely(flags & HASH_SMALL)) {
4888 /* Makes no sense without HASH_EARLY */
4889 WARN_ON(!(flags & HASH_EARLY));
4890 if (!(numentries >> *_hash_shift)) {
4891 numentries = 1UL << *_hash_shift;
4892 BUG_ON(!numentries);
4894 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
4895 numentries = PAGE_SIZE / bucketsize;
4897 numentries = roundup_pow_of_two(numentries);
4899 /* limit allocation size to 1/16 total memory by default */
4901 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
4902 do_div(max, bucketsize);
4905 if (numentries > max)
4908 log2qty = ilog2(numentries);
4911 size = bucketsize << log2qty;
4912 if (flags & HASH_EARLY)
4913 table = alloc_bootmem_nopanic(size);
4915 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
4918 * If bucketsize is not a power-of-two, we may free
4919 * some pages at the end of hash table which
4920 * alloc_pages_exact() automatically does
4922 if (get_order(size) < MAX_ORDER) {
4923 table = alloc_pages_exact(size, GFP_ATOMIC);
4924 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
4927 } while (!table && size > PAGE_SIZE && --log2qty);
4930 panic("Failed to allocate %s hash table\n", tablename);
4932 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
4935 ilog2(size) - PAGE_SHIFT,
4939 *_hash_shift = log2qty;
4941 *_hash_mask = (1 << log2qty) - 1;
4946 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4947 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
4950 #ifdef CONFIG_SPARSEMEM
4951 return __pfn_to_section(pfn)->pageblock_flags;
4953 return zone->pageblock_flags;
4954 #endif /* CONFIG_SPARSEMEM */
4957 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
4959 #ifdef CONFIG_SPARSEMEM
4960 pfn &= (PAGES_PER_SECTION-1);
4961 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4963 pfn = pfn - zone->zone_start_pfn;
4964 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4965 #endif /* CONFIG_SPARSEMEM */
4969 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4970 * @page: The page within the block of interest
4971 * @start_bitidx: The first bit of interest to retrieve
4972 * @end_bitidx: The last bit of interest
4973 * returns pageblock_bits flags
4975 unsigned long get_pageblock_flags_group(struct page *page,
4976 int start_bitidx, int end_bitidx)
4979 unsigned long *bitmap;
4980 unsigned long pfn, bitidx;
4981 unsigned long flags = 0;
4982 unsigned long value = 1;
4984 zone = page_zone(page);
4985 pfn = page_to_pfn(page);
4986 bitmap = get_pageblock_bitmap(zone, pfn);
4987 bitidx = pfn_to_bitidx(zone, pfn);
4989 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4990 if (test_bit(bitidx + start_bitidx, bitmap))
4997 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4998 * @page: The page within the block of interest
4999 * @start_bitidx: The first bit of interest
5000 * @end_bitidx: The last bit of interest
5001 * @flags: The flags to set
5003 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5004 int start_bitidx, int end_bitidx)
5007 unsigned long *bitmap;
5008 unsigned long pfn, bitidx;
5009 unsigned long value = 1;
5011 zone = page_zone(page);
5012 pfn = page_to_pfn(page);
5013 bitmap = get_pageblock_bitmap(zone, pfn);
5014 bitidx = pfn_to_bitidx(zone, pfn);
5015 VM_BUG_ON(pfn < zone->zone_start_pfn);
5016 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5018 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5020 __set_bit(bitidx + start_bitidx, bitmap);
5022 __clear_bit(bitidx + start_bitidx, bitmap);
5026 * This is designed as sub function...plz see page_isolation.c also.
5027 * set/clear page block's type to be ISOLATE.
5028 * page allocater never alloc memory from ISOLATE block.
5031 int set_migratetype_isolate(struct page *page)
5034 unsigned long flags;
5038 zone = page_zone(page);
5039 zone_idx = zone_idx(zone);
5040 spin_lock_irqsave(&zone->lock, flags);
5042 * In future, more migrate types will be able to be isolation target.
5044 if (get_pageblock_migratetype(page) != MIGRATE_MOVABLE &&
5045 zone_idx != ZONE_MOVABLE)
5047 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5048 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5051 spin_unlock_irqrestore(&zone->lock, flags);
5057 void unset_migratetype_isolate(struct page *page)
5060 unsigned long flags;
5061 zone = page_zone(page);
5062 spin_lock_irqsave(&zone->lock, flags);
5063 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5065 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5066 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5068 spin_unlock_irqrestore(&zone->lock, flags);
5071 #ifdef CONFIG_MEMORY_HOTREMOVE
5073 * All pages in the range must be isolated before calling this.
5076 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5082 unsigned long flags;
5083 /* find the first valid pfn */
5084 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5089 zone = page_zone(pfn_to_page(pfn));
5090 spin_lock_irqsave(&zone->lock, flags);
5092 while (pfn < end_pfn) {
5093 if (!pfn_valid(pfn)) {
5097 page = pfn_to_page(pfn);
5098 BUG_ON(page_count(page));
5099 BUG_ON(!PageBuddy(page));
5100 order = page_order(page);
5101 #ifdef CONFIG_DEBUG_VM
5102 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5103 pfn, 1 << order, end_pfn);
5105 list_del(&page->lru);
5106 rmv_page_order(page);
5107 zone->free_area[order].nr_free--;
5108 __mod_zone_page_state(zone, NR_FREE_PAGES,
5110 for (i = 0; i < (1 << order); i++)
5111 SetPageReserved((page+i));
5112 pfn += (1 << order);
5114 spin_unlock_irqrestore(&zone->lock, flags);