2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/compiler.h>
25 #include <linux/kernel.h>
26 #include <linux/kmemcheck.h>
27 #include <linux/module.h>
28 #include <linux/suspend.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/slab.h>
32 #include <linux/oom.h>
33 #include <linux/notifier.h>
34 #include <linux/topology.h>
35 #include <linux/sysctl.h>
36 #include <linux/cpu.h>
37 #include <linux/cpuset.h>
38 #include <linux/memory_hotplug.h>
39 #include <linux/nodemask.h>
40 #include <linux/vmalloc.h>
41 #include <linux/mempolicy.h>
42 #include <linux/stop_machine.h>
43 #include <linux/sort.h>
44 #include <linux/pfn.h>
45 #include <linux/backing-dev.h>
46 #include <linux/fault-inject.h>
47 #include <linux/page-isolation.h>
48 #include <linux/page_cgroup.h>
49 #include <linux/debugobjects.h>
50 #include <linux/kmemleak.h>
51 #include <linux/memory.h>
52 #include <trace/events/kmem.h>
54 #include <asm/tlbflush.h>
55 #include <asm/div64.h>
59 * Array of node states.
61 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
62 [N_POSSIBLE] = NODE_MASK_ALL,
63 [N_ONLINE] = { { [0] = 1UL } },
65 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
67 [N_HIGH_MEMORY] = { { [0] = 1UL } },
69 [N_CPU] = { { [0] = 1UL } },
72 EXPORT_SYMBOL(node_states);
74 unsigned long totalram_pages __read_mostly;
75 unsigned long totalreserve_pages __read_mostly;
76 int percpu_pagelist_fraction;
77 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
79 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
80 int pageblock_order __read_mostly;
83 static void __free_pages_ok(struct page *page, unsigned int order);
86 * results with 256, 32 in the lowmem_reserve sysctl:
87 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
88 * 1G machine -> (16M dma, 784M normal, 224M high)
89 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
90 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
91 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
93 * TBD: should special case ZONE_DMA32 machines here - in those we normally
94 * don't need any ZONE_NORMAL reservation
96 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
97 #ifdef CONFIG_ZONE_DMA
100 #ifdef CONFIG_ZONE_DMA32
103 #ifdef CONFIG_HIGHMEM
109 EXPORT_SYMBOL(totalram_pages);
111 static char * const zone_names[MAX_NR_ZONES] = {
112 #ifdef CONFIG_ZONE_DMA
115 #ifdef CONFIG_ZONE_DMA32
119 #ifdef CONFIG_HIGHMEM
125 int min_free_kbytes = 1024;
127 static unsigned long __meminitdata nr_kernel_pages;
128 static unsigned long __meminitdata nr_all_pages;
129 static unsigned long __meminitdata dma_reserve;
131 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
133 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
134 * ranges of memory (RAM) that may be registered with add_active_range().
135 * Ranges passed to add_active_range() will be merged if possible
136 * so the number of times add_active_range() can be called is
137 * related to the number of nodes and the number of holes
139 #ifdef CONFIG_MAX_ACTIVE_REGIONS
140 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
141 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
143 #if MAX_NUMNODES >= 32
144 /* If there can be many nodes, allow up to 50 holes per node */
145 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
147 /* By default, allow up to 256 distinct regions */
148 #define MAX_ACTIVE_REGIONS 256
152 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
153 static int __meminitdata nr_nodemap_entries;
154 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
155 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
156 static unsigned long __initdata required_kernelcore;
157 static unsigned long __initdata required_movablecore;
158 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
160 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
162 EXPORT_SYMBOL(movable_zone);
163 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
166 int nr_node_ids __read_mostly = MAX_NUMNODES;
167 int nr_online_nodes __read_mostly = 1;
168 EXPORT_SYMBOL(nr_node_ids);
169 EXPORT_SYMBOL(nr_online_nodes);
172 int page_group_by_mobility_disabled __read_mostly;
174 static void set_pageblock_migratetype(struct page *page, int migratetype)
177 if (unlikely(page_group_by_mobility_disabled))
178 migratetype = MIGRATE_UNMOVABLE;
180 set_pageblock_flags_group(page, (unsigned long)migratetype,
181 PB_migrate, PB_migrate_end);
184 bool oom_killer_disabled __read_mostly;
186 #ifdef CONFIG_DEBUG_VM
187 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
191 unsigned long pfn = page_to_pfn(page);
194 seq = zone_span_seqbegin(zone);
195 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
197 else if (pfn < zone->zone_start_pfn)
199 } while (zone_span_seqretry(zone, seq));
204 static int page_is_consistent(struct zone *zone, struct page *page)
206 if (!pfn_valid_within(page_to_pfn(page)))
208 if (zone != page_zone(page))
214 * Temporary debugging check for pages not lying within a given zone.
216 static int bad_range(struct zone *zone, struct page *page)
218 if (page_outside_zone_boundaries(zone, page))
220 if (!page_is_consistent(zone, page))
226 static inline int bad_range(struct zone *zone, struct page *page)
232 static void bad_page(struct page *page)
234 static unsigned long resume;
235 static unsigned long nr_shown;
236 static unsigned long nr_unshown;
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++;
491 * free_page_mlock() -- clean up attempts to free and mlocked() page.
492 * Page should not be on lru, so no need to fix that up.
493 * free_pages_check() will verify...
495 static inline void free_page_mlock(struct page *page)
497 __dec_zone_page_state(page, NR_MLOCK);
498 __count_vm_event(UNEVICTABLE_MLOCKFREED);
501 static inline int free_pages_check(struct page *page)
503 if (unlikely(page_mapcount(page) |
504 (page->mapping != NULL) |
505 (atomic_read(&page->_count) != 0) |
506 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) {
510 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
511 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
516 * Frees a number of pages from the PCP lists
517 * Assumes all pages on list are in same zone, and of same order.
518 * count is the number of pages to free.
520 * If the zone was previously in an "all pages pinned" state then look to
521 * see if this freeing clears that state.
523 * And clear the zone's pages_scanned counter, to hold off the "all pages are
524 * pinned" detection logic.
526 static void free_pcppages_bulk(struct zone *zone, int count,
527 struct per_cpu_pages *pcp)
532 spin_lock(&zone->lock);
533 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
534 zone->pages_scanned = 0;
536 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
539 struct list_head *list;
542 * Remove pages from lists in a round-robin fashion. A
543 * batch_free count is maintained that is incremented when an
544 * empty list is encountered. This is so more pages are freed
545 * off fuller lists instead of spinning excessively around empty
550 if (++migratetype == MIGRATE_PCPTYPES)
552 list = &pcp->lists[migratetype];
553 } while (list_empty(list));
556 page = list_entry(list->prev, struct page, lru);
557 /* must delete as __free_one_page list manipulates */
558 list_del(&page->lru);
559 __free_one_page(page, zone, 0, migratetype);
560 trace_mm_page_pcpu_drain(page, 0, migratetype);
561 } while (--count && --batch_free && !list_empty(list));
563 spin_unlock(&zone->lock);
566 static void free_one_page(struct zone *zone, struct page *page, int order,
569 spin_lock(&zone->lock);
570 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
571 zone->pages_scanned = 0;
573 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
574 __free_one_page(page, zone, order, migratetype);
575 spin_unlock(&zone->lock);
578 static void __free_pages_ok(struct page *page, unsigned int order)
583 int wasMlocked = __TestClearPageMlocked(page);
585 kmemcheck_free_shadow(page, order);
587 for (i = 0 ; i < (1 << order) ; ++i)
588 bad += free_pages_check(page + i);
592 if (!PageHighMem(page)) {
593 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
594 debug_check_no_obj_freed(page_address(page),
597 arch_free_page(page, order);
598 kernel_map_pages(page, 1 << order, 0);
600 local_irq_save(flags);
601 if (unlikely(wasMlocked))
602 free_page_mlock(page);
603 __count_vm_events(PGFREE, 1 << order);
604 free_one_page(page_zone(page), page, order,
605 get_pageblock_migratetype(page));
606 local_irq_restore(flags);
610 * permit the bootmem allocator to evade page validation on high-order frees
612 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
615 __ClearPageReserved(page);
616 set_page_count(page, 0);
617 set_page_refcounted(page);
623 for (loop = 0; loop < BITS_PER_LONG; loop++) {
624 struct page *p = &page[loop];
626 if (loop + 1 < BITS_PER_LONG)
628 __ClearPageReserved(p);
629 set_page_count(p, 0);
632 set_page_refcounted(page);
633 __free_pages(page, order);
639 * The order of subdivision here is critical for the IO subsystem.
640 * Please do not alter this order without good reasons and regression
641 * testing. Specifically, as large blocks of memory are subdivided,
642 * the order in which smaller blocks are delivered depends on the order
643 * they're subdivided in this function. This is the primary factor
644 * influencing the order in which pages are delivered to the IO
645 * subsystem according to empirical testing, and this is also justified
646 * by considering the behavior of a buddy system containing a single
647 * large block of memory acted on by a series of small allocations.
648 * This behavior is a critical factor in sglist merging's success.
652 static inline void expand(struct zone *zone, struct page *page,
653 int low, int high, struct free_area *area,
656 unsigned long size = 1 << high;
662 VM_BUG_ON(bad_range(zone, &page[size]));
663 list_add(&page[size].lru, &area->free_list[migratetype]);
665 set_page_order(&page[size], high);
670 * This page is about to be returned from the page allocator
672 static inline int check_new_page(struct page *page)
674 if (unlikely(page_mapcount(page) |
675 (page->mapping != NULL) |
676 (atomic_read(&page->_count) != 0) |
677 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) {
684 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
688 for (i = 0; i < (1 << order); i++) {
689 struct page *p = page + i;
690 if (unlikely(check_new_page(p)))
694 set_page_private(page, 0);
695 set_page_refcounted(page);
697 arch_alloc_page(page, order);
698 kernel_map_pages(page, 1 << order, 1);
700 if (gfp_flags & __GFP_ZERO)
701 prep_zero_page(page, order, gfp_flags);
703 if (order && (gfp_flags & __GFP_COMP))
704 prep_compound_page(page, order);
710 * Go through the free lists for the given migratetype and remove
711 * the smallest available page from the freelists
714 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
717 unsigned int current_order;
718 struct free_area * area;
721 /* Find a page of the appropriate size in the preferred list */
722 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
723 area = &(zone->free_area[current_order]);
724 if (list_empty(&area->free_list[migratetype]))
727 page = list_entry(area->free_list[migratetype].next,
729 list_del(&page->lru);
730 rmv_page_order(page);
732 expand(zone, page, order, current_order, area, migratetype);
741 * This array describes the order lists are fallen back to when
742 * the free lists for the desirable migrate type are depleted
744 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
745 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
746 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
747 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
748 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
752 * Move the free pages in a range to the free lists of the requested type.
753 * Note that start_page and end_pages are not aligned on a pageblock
754 * boundary. If alignment is required, use move_freepages_block()
756 static int move_freepages(struct zone *zone,
757 struct page *start_page, struct page *end_page,
764 #ifndef CONFIG_HOLES_IN_ZONE
766 * page_zone is not safe to call in this context when
767 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
768 * anyway as we check zone boundaries in move_freepages_block().
769 * Remove at a later date when no bug reports exist related to
770 * grouping pages by mobility
772 BUG_ON(page_zone(start_page) != page_zone(end_page));
775 for (page = start_page; page <= end_page;) {
776 /* Make sure we are not inadvertently changing nodes */
777 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
779 if (!pfn_valid_within(page_to_pfn(page))) {
784 if (!PageBuddy(page)) {
789 order = page_order(page);
790 list_del(&page->lru);
792 &zone->free_area[order].free_list[migratetype]);
794 pages_moved += 1 << order;
800 static int move_freepages_block(struct zone *zone, struct page *page,
803 unsigned long start_pfn, end_pfn;
804 struct page *start_page, *end_page;
806 start_pfn = page_to_pfn(page);
807 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
808 start_page = pfn_to_page(start_pfn);
809 end_page = start_page + pageblock_nr_pages - 1;
810 end_pfn = start_pfn + pageblock_nr_pages - 1;
812 /* Do not cross zone boundaries */
813 if (start_pfn < zone->zone_start_pfn)
815 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
818 return move_freepages(zone, start_page, end_page, migratetype);
821 static void change_pageblock_range(struct page *pageblock_page,
822 int start_order, int migratetype)
824 int nr_pageblocks = 1 << (start_order - pageblock_order);
826 while (nr_pageblocks--) {
827 set_pageblock_migratetype(pageblock_page, migratetype);
828 pageblock_page += pageblock_nr_pages;
832 /* Remove an element from the buddy allocator from the fallback list */
833 static inline struct page *
834 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
836 struct free_area * area;
841 /* Find the largest possible block of pages in the other list */
842 for (current_order = MAX_ORDER-1; current_order >= order;
844 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
845 migratetype = fallbacks[start_migratetype][i];
847 /* MIGRATE_RESERVE handled later if necessary */
848 if (migratetype == MIGRATE_RESERVE)
851 area = &(zone->free_area[current_order]);
852 if (list_empty(&area->free_list[migratetype]))
855 page = list_entry(area->free_list[migratetype].next,
860 * If breaking a large block of pages, move all free
861 * pages to the preferred allocation list. If falling
862 * back for a reclaimable kernel allocation, be more
863 * agressive about taking ownership of free pages
865 if (unlikely(current_order >= (pageblock_order >> 1)) ||
866 start_migratetype == MIGRATE_RECLAIMABLE ||
867 page_group_by_mobility_disabled) {
869 pages = move_freepages_block(zone, page,
872 /* Claim the whole block if over half of it is free */
873 if (pages >= (1 << (pageblock_order-1)) ||
874 page_group_by_mobility_disabled)
875 set_pageblock_migratetype(page,
878 migratetype = start_migratetype;
881 /* Remove the page from the freelists */
882 list_del(&page->lru);
883 rmv_page_order(page);
885 /* Take ownership for orders >= pageblock_order */
886 if (current_order >= pageblock_order)
887 change_pageblock_range(page, current_order,
890 expand(zone, page, order, current_order, area, migratetype);
892 trace_mm_page_alloc_extfrag(page, order, current_order,
893 start_migratetype, migratetype);
903 * Do the hard work of removing an element from the buddy allocator.
904 * Call me with the zone->lock already held.
906 static struct page *__rmqueue(struct zone *zone, unsigned int order,
912 page = __rmqueue_smallest(zone, order, migratetype);
914 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
915 page = __rmqueue_fallback(zone, order, migratetype);
918 * Use MIGRATE_RESERVE rather than fail an allocation. goto
919 * is used because __rmqueue_smallest is an inline function
920 * and we want just one call site
923 migratetype = MIGRATE_RESERVE;
928 trace_mm_page_alloc_zone_locked(page, order, migratetype);
933 * Obtain a specified number of elements from the buddy allocator, all under
934 * a single hold of the lock, for efficiency. Add them to the supplied list.
935 * Returns the number of new pages which were placed at *list.
937 static int rmqueue_bulk(struct zone *zone, unsigned int order,
938 unsigned long count, struct list_head *list,
939 int migratetype, int cold)
943 spin_lock(&zone->lock);
944 for (i = 0; i < count; ++i) {
945 struct page *page = __rmqueue(zone, order, migratetype);
946 if (unlikely(page == NULL))
950 * Split buddy pages returned by expand() are received here
951 * in physical page order. The page is added to the callers and
952 * list and the list head then moves forward. From the callers
953 * perspective, the linked list is ordered by page number in
954 * some conditions. This is useful for IO devices that can
955 * merge IO requests if the physical pages are ordered
958 if (likely(cold == 0))
959 list_add(&page->lru, list);
961 list_add_tail(&page->lru, list);
962 set_page_private(page, migratetype);
965 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
966 spin_unlock(&zone->lock);
972 * Called from the vmstat counter updater to drain pagesets of this
973 * currently executing processor on remote nodes after they have
976 * Note that this function must be called with the thread pinned to
977 * a single processor.
979 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
984 local_irq_save(flags);
985 if (pcp->count >= pcp->batch)
986 to_drain = pcp->batch;
988 to_drain = pcp->count;
989 free_pcppages_bulk(zone, to_drain, pcp);
990 pcp->count -= to_drain;
991 local_irq_restore(flags);
996 * Drain pages of the indicated processor.
998 * The processor must either be the current processor and the
999 * thread pinned to the current processor or a processor that
1002 static void drain_pages(unsigned int cpu)
1004 unsigned long flags;
1007 for_each_populated_zone(zone) {
1008 struct per_cpu_pageset *pset;
1009 struct per_cpu_pages *pcp;
1011 local_irq_save(flags);
1012 pset = per_cpu_ptr(zone->pageset, cpu);
1015 free_pcppages_bulk(zone, pcp->count, pcp);
1017 local_irq_restore(flags);
1022 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1024 void drain_local_pages(void *arg)
1026 drain_pages(smp_processor_id());
1030 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1032 void drain_all_pages(void)
1034 on_each_cpu(drain_local_pages, NULL, 1);
1037 #ifdef CONFIG_HIBERNATION
1039 void mark_free_pages(struct zone *zone)
1041 unsigned long pfn, max_zone_pfn;
1042 unsigned long flags;
1044 struct list_head *curr;
1046 if (!zone->spanned_pages)
1049 spin_lock_irqsave(&zone->lock, flags);
1051 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1052 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1053 if (pfn_valid(pfn)) {
1054 struct page *page = pfn_to_page(pfn);
1056 if (!swsusp_page_is_forbidden(page))
1057 swsusp_unset_page_free(page);
1060 for_each_migratetype_order(order, t) {
1061 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1064 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1065 for (i = 0; i < (1UL << order); i++)
1066 swsusp_set_page_free(pfn_to_page(pfn + i));
1069 spin_unlock_irqrestore(&zone->lock, flags);
1071 #endif /* CONFIG_PM */
1074 * Free a 0-order page
1076 static void free_hot_cold_page(struct page *page, int cold)
1078 struct zone *zone = page_zone(page);
1079 struct per_cpu_pages *pcp;
1080 unsigned long flags;
1082 int wasMlocked = __TestClearPageMlocked(page);
1084 kmemcheck_free_shadow(page, 0);
1087 page->mapping = NULL;
1088 if (free_pages_check(page))
1091 if (!PageHighMem(page)) {
1092 debug_check_no_locks_freed(page_address(page), PAGE_SIZE);
1093 debug_check_no_obj_freed(page_address(page), PAGE_SIZE);
1095 arch_free_page(page, 0);
1096 kernel_map_pages(page, 1, 0);
1098 migratetype = get_pageblock_migratetype(page);
1099 set_page_private(page, migratetype);
1100 local_irq_save(flags);
1101 if (unlikely(wasMlocked))
1102 free_page_mlock(page);
1103 __count_vm_event(PGFREE);
1106 * We only track unmovable, reclaimable and movable on pcp lists.
1107 * Free ISOLATE pages back to the allocator because they are being
1108 * offlined but treat RESERVE as movable pages so we can get those
1109 * areas back if necessary. Otherwise, we may have to free
1110 * excessively into the page allocator
1112 if (migratetype >= MIGRATE_PCPTYPES) {
1113 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1114 free_one_page(zone, page, 0, migratetype);
1117 migratetype = MIGRATE_MOVABLE;
1120 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1122 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1124 list_add(&page->lru, &pcp->lists[migratetype]);
1126 if (pcp->count >= pcp->high) {
1127 free_pcppages_bulk(zone, pcp->batch, pcp);
1128 pcp->count -= pcp->batch;
1132 local_irq_restore(flags);
1135 void free_hot_page(struct page *page)
1137 trace_mm_page_free_direct(page, 0);
1138 free_hot_cold_page(page, 0);
1142 * split_page takes a non-compound higher-order page, and splits it into
1143 * n (1<<order) sub-pages: page[0..n]
1144 * Each sub-page must be freed individually.
1146 * Note: this is probably too low level an operation for use in drivers.
1147 * Please consult with lkml before using this in your driver.
1149 void split_page(struct page *page, unsigned int order)
1153 VM_BUG_ON(PageCompound(page));
1154 VM_BUG_ON(!page_count(page));
1156 #ifdef CONFIG_KMEMCHECK
1158 * Split shadow pages too, because free(page[0]) would
1159 * otherwise free the whole shadow.
1161 if (kmemcheck_page_is_tracked(page))
1162 split_page(virt_to_page(page[0].shadow), order);
1165 for (i = 1; i < (1 << order); i++)
1166 set_page_refcounted(page + i);
1170 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1171 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1175 struct page *buffered_rmqueue(struct zone *preferred_zone,
1176 struct zone *zone, int order, gfp_t gfp_flags,
1179 unsigned long flags;
1181 int cold = !!(gfp_flags & __GFP_COLD);
1184 if (likely(order == 0)) {
1185 struct per_cpu_pages *pcp;
1186 struct list_head *list;
1188 local_irq_save(flags);
1189 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1190 list = &pcp->lists[migratetype];
1191 if (list_empty(list)) {
1192 pcp->count += rmqueue_bulk(zone, 0,
1195 if (unlikely(list_empty(list)))
1200 page = list_entry(list->prev, struct page, lru);
1202 page = list_entry(list->next, struct page, lru);
1204 list_del(&page->lru);
1207 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1209 * __GFP_NOFAIL is not to be used in new code.
1211 * All __GFP_NOFAIL callers should be fixed so that they
1212 * properly detect and handle allocation failures.
1214 * We most definitely don't want callers attempting to
1215 * allocate greater than order-1 page units with
1218 WARN_ON_ONCE(order > 1);
1220 spin_lock_irqsave(&zone->lock, flags);
1221 page = __rmqueue(zone, order, migratetype);
1222 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1223 spin_unlock(&zone->lock);
1228 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1229 zone_statistics(preferred_zone, zone);
1230 local_irq_restore(flags);
1232 VM_BUG_ON(bad_range(zone, page));
1233 if (prep_new_page(page, order, gfp_flags))
1238 local_irq_restore(flags);
1242 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1243 #define ALLOC_WMARK_MIN WMARK_MIN
1244 #define ALLOC_WMARK_LOW WMARK_LOW
1245 #define ALLOC_WMARK_HIGH WMARK_HIGH
1246 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1248 /* Mask to get the watermark bits */
1249 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1251 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1252 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1253 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1255 #ifdef CONFIG_FAIL_PAGE_ALLOC
1257 static struct fail_page_alloc_attr {
1258 struct fault_attr attr;
1260 u32 ignore_gfp_highmem;
1261 u32 ignore_gfp_wait;
1264 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1266 struct dentry *ignore_gfp_highmem_file;
1267 struct dentry *ignore_gfp_wait_file;
1268 struct dentry *min_order_file;
1270 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1272 } fail_page_alloc = {
1273 .attr = FAULT_ATTR_INITIALIZER,
1274 .ignore_gfp_wait = 1,
1275 .ignore_gfp_highmem = 1,
1279 static int __init setup_fail_page_alloc(char *str)
1281 return setup_fault_attr(&fail_page_alloc.attr, str);
1283 __setup("fail_page_alloc=", setup_fail_page_alloc);
1285 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1287 if (order < fail_page_alloc.min_order)
1289 if (gfp_mask & __GFP_NOFAIL)
1291 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1293 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1296 return should_fail(&fail_page_alloc.attr, 1 << order);
1299 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1301 static int __init fail_page_alloc_debugfs(void)
1303 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1307 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1311 dir = fail_page_alloc.attr.dentries.dir;
1313 fail_page_alloc.ignore_gfp_wait_file =
1314 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1315 &fail_page_alloc.ignore_gfp_wait);
1317 fail_page_alloc.ignore_gfp_highmem_file =
1318 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1319 &fail_page_alloc.ignore_gfp_highmem);
1320 fail_page_alloc.min_order_file =
1321 debugfs_create_u32("min-order", mode, dir,
1322 &fail_page_alloc.min_order);
1324 if (!fail_page_alloc.ignore_gfp_wait_file ||
1325 !fail_page_alloc.ignore_gfp_highmem_file ||
1326 !fail_page_alloc.min_order_file) {
1328 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1329 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1330 debugfs_remove(fail_page_alloc.min_order_file);
1331 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1337 late_initcall(fail_page_alloc_debugfs);
1339 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1341 #else /* CONFIG_FAIL_PAGE_ALLOC */
1343 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1348 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1351 * Return 1 if free pages are above 'mark'. This takes into account the order
1352 * of the allocation.
1354 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1355 int classzone_idx, int alloc_flags)
1357 /* free_pages my go negative - that's OK */
1359 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
1362 if (alloc_flags & ALLOC_HIGH)
1364 if (alloc_flags & ALLOC_HARDER)
1367 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1369 for (o = 0; o < order; o++) {
1370 /* At the next order, this order's pages become unavailable */
1371 free_pages -= z->free_area[o].nr_free << o;
1373 /* Require fewer higher order pages to be free */
1376 if (free_pages <= min)
1384 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1385 * skip over zones that are not allowed by the cpuset, or that have
1386 * been recently (in last second) found to be nearly full. See further
1387 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1388 * that have to skip over a lot of full or unallowed zones.
1390 * If the zonelist cache is present in the passed in zonelist, then
1391 * returns a pointer to the allowed node mask (either the current
1392 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1394 * If the zonelist cache is not available for this zonelist, does
1395 * nothing and returns NULL.
1397 * If the fullzones BITMAP in the zonelist cache is stale (more than
1398 * a second since last zap'd) then we zap it out (clear its bits.)
1400 * We hold off even calling zlc_setup, until after we've checked the
1401 * first zone in the zonelist, on the theory that most allocations will
1402 * be satisfied from that first zone, so best to examine that zone as
1403 * quickly as we can.
1405 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1407 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1408 nodemask_t *allowednodes; /* zonelist_cache approximation */
1410 zlc = zonelist->zlcache_ptr;
1414 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1415 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1416 zlc->last_full_zap = jiffies;
1419 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1420 &cpuset_current_mems_allowed :
1421 &node_states[N_HIGH_MEMORY];
1422 return allowednodes;
1426 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1427 * if it is worth looking at further for free memory:
1428 * 1) Check that the zone isn't thought to be full (doesn't have its
1429 * bit set in the zonelist_cache fullzones BITMAP).
1430 * 2) Check that the zones node (obtained from the zonelist_cache
1431 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1432 * Return true (non-zero) if zone is worth looking at further, or
1433 * else return false (zero) if it is not.
1435 * This check -ignores- the distinction between various watermarks,
1436 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1437 * found to be full for any variation of these watermarks, it will
1438 * be considered full for up to one second by all requests, unless
1439 * we are so low on memory on all allowed nodes that we are forced
1440 * into the second scan of the zonelist.
1442 * In the second scan we ignore this zonelist cache and exactly
1443 * apply the watermarks to all zones, even it is slower to do so.
1444 * We are low on memory in the second scan, and should leave no stone
1445 * unturned looking for a free page.
1447 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1448 nodemask_t *allowednodes)
1450 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1451 int i; /* index of *z in zonelist zones */
1452 int n; /* node that zone *z is on */
1454 zlc = zonelist->zlcache_ptr;
1458 i = z - zonelist->_zonerefs;
1461 /* This zone is worth trying if it is allowed but not full */
1462 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1466 * Given 'z' scanning a zonelist, set the corresponding bit in
1467 * zlc->fullzones, so that subsequent attempts to allocate a page
1468 * from that zone don't waste time re-examining it.
1470 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1472 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1473 int i; /* index of *z in zonelist zones */
1475 zlc = zonelist->zlcache_ptr;
1479 i = z - zonelist->_zonerefs;
1481 set_bit(i, zlc->fullzones);
1484 #else /* CONFIG_NUMA */
1486 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1491 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1492 nodemask_t *allowednodes)
1497 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1500 #endif /* CONFIG_NUMA */
1503 * get_page_from_freelist goes through the zonelist trying to allocate
1506 static struct page *
1507 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1508 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1509 struct zone *preferred_zone, int migratetype)
1512 struct page *page = NULL;
1515 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1516 int zlc_active = 0; /* set if using zonelist_cache */
1517 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1519 classzone_idx = zone_idx(preferred_zone);
1522 * Scan zonelist, looking for a zone with enough free.
1523 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1525 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1526 high_zoneidx, nodemask) {
1527 if (NUMA_BUILD && zlc_active &&
1528 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1530 if ((alloc_flags & ALLOC_CPUSET) &&
1531 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1534 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1535 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1539 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1540 if (zone_watermark_ok(zone, order, mark,
1541 classzone_idx, alloc_flags))
1544 if (zone_reclaim_mode == 0)
1545 goto this_zone_full;
1547 ret = zone_reclaim(zone, gfp_mask, order);
1549 case ZONE_RECLAIM_NOSCAN:
1552 case ZONE_RECLAIM_FULL:
1553 /* scanned but unreclaimable */
1554 goto this_zone_full;
1556 /* did we reclaim enough */
1557 if (!zone_watermark_ok(zone, order, mark,
1558 classzone_idx, alloc_flags))
1559 goto this_zone_full;
1564 page = buffered_rmqueue(preferred_zone, zone, order,
1565 gfp_mask, migratetype);
1570 zlc_mark_zone_full(zonelist, z);
1572 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1574 * we do zlc_setup after the first zone is tried but only
1575 * if there are multiple nodes make it worthwhile
1577 allowednodes = zlc_setup(zonelist, alloc_flags);
1583 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1584 /* Disable zlc cache for second zonelist scan */
1592 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1593 unsigned long pages_reclaimed)
1595 /* Do not loop if specifically requested */
1596 if (gfp_mask & __GFP_NORETRY)
1600 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1601 * means __GFP_NOFAIL, but that may not be true in other
1604 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1608 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1609 * specified, then we retry until we no longer reclaim any pages
1610 * (above), or we've reclaimed an order of pages at least as
1611 * large as the allocation's order. In both cases, if the
1612 * allocation still fails, we stop retrying.
1614 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1618 * Don't let big-order allocations loop unless the caller
1619 * explicitly requests that.
1621 if (gfp_mask & __GFP_NOFAIL)
1627 static inline struct page *
1628 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1629 struct zonelist *zonelist, enum zone_type high_zoneidx,
1630 nodemask_t *nodemask, struct zone *preferred_zone,
1635 /* Acquire the OOM killer lock for the zones in zonelist */
1636 if (!try_set_zone_oom(zonelist, gfp_mask)) {
1637 schedule_timeout_uninterruptible(1);
1642 * Go through the zonelist yet one more time, keep very high watermark
1643 * here, this is only to catch a parallel oom killing, we must fail if
1644 * we're still under heavy pressure.
1646 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1647 order, zonelist, high_zoneidx,
1648 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1649 preferred_zone, migratetype);
1653 if (!(gfp_mask & __GFP_NOFAIL)) {
1654 /* The OOM killer will not help higher order allocs */
1655 if (order > PAGE_ALLOC_COSTLY_ORDER)
1658 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1659 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1660 * The caller should handle page allocation failure by itself if
1661 * it specifies __GFP_THISNODE.
1662 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1664 if (gfp_mask & __GFP_THISNODE)
1667 /* Exhausted what can be done so it's blamo time */
1668 out_of_memory(zonelist, gfp_mask, order, nodemask);
1671 clear_zonelist_oom(zonelist, gfp_mask);
1675 /* The really slow allocator path where we enter direct reclaim */
1676 static inline struct page *
1677 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1678 struct zonelist *zonelist, enum zone_type high_zoneidx,
1679 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1680 int migratetype, unsigned long *did_some_progress)
1682 struct page *page = NULL;
1683 struct reclaim_state reclaim_state;
1684 struct task_struct *p = current;
1688 /* We now go into synchronous reclaim */
1689 cpuset_memory_pressure_bump();
1690 p->flags |= PF_MEMALLOC;
1691 lockdep_set_current_reclaim_state(gfp_mask);
1692 reclaim_state.reclaimed_slab = 0;
1693 p->reclaim_state = &reclaim_state;
1695 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1697 p->reclaim_state = NULL;
1698 lockdep_clear_current_reclaim_state();
1699 p->flags &= ~PF_MEMALLOC;
1706 if (likely(*did_some_progress))
1707 page = get_page_from_freelist(gfp_mask, nodemask, order,
1708 zonelist, high_zoneidx,
1709 alloc_flags, preferred_zone,
1715 * This is called in the allocator slow-path if the allocation request is of
1716 * sufficient urgency to ignore watermarks and take other desperate measures
1718 static inline struct page *
1719 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1720 struct zonelist *zonelist, enum zone_type high_zoneidx,
1721 nodemask_t *nodemask, struct zone *preferred_zone,
1727 page = get_page_from_freelist(gfp_mask, nodemask, order,
1728 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
1729 preferred_zone, migratetype);
1731 if (!page && gfp_mask & __GFP_NOFAIL)
1732 congestion_wait(BLK_RW_ASYNC, HZ/50);
1733 } while (!page && (gfp_mask & __GFP_NOFAIL));
1739 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
1740 enum zone_type high_zoneidx)
1745 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1746 wakeup_kswapd(zone, order);
1750 gfp_to_alloc_flags(gfp_t gfp_mask)
1752 struct task_struct *p = current;
1753 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
1754 const gfp_t wait = gfp_mask & __GFP_WAIT;
1756 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1757 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH);
1760 * The caller may dip into page reserves a bit more if the caller
1761 * cannot run direct reclaim, or if the caller has realtime scheduling
1762 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1763 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1765 alloc_flags |= (gfp_mask & __GFP_HIGH);
1768 alloc_flags |= ALLOC_HARDER;
1770 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1771 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1773 alloc_flags &= ~ALLOC_CPUSET;
1774 } else if (unlikely(rt_task(p)) && !in_interrupt())
1775 alloc_flags |= ALLOC_HARDER;
1777 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
1778 if (!in_interrupt() &&
1779 ((p->flags & PF_MEMALLOC) ||
1780 unlikely(test_thread_flag(TIF_MEMDIE))))
1781 alloc_flags |= ALLOC_NO_WATERMARKS;
1787 static inline struct page *
1788 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
1789 struct zonelist *zonelist, enum zone_type high_zoneidx,
1790 nodemask_t *nodemask, struct zone *preferred_zone,
1793 const gfp_t wait = gfp_mask & __GFP_WAIT;
1794 struct page *page = NULL;
1796 unsigned long pages_reclaimed = 0;
1797 unsigned long did_some_progress;
1798 struct task_struct *p = current;
1801 * In the slowpath, we sanity check order to avoid ever trying to
1802 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1803 * be using allocators in order of preference for an area that is
1806 if (order >= MAX_ORDER) {
1807 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
1812 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1813 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1814 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1815 * using a larger set of nodes after it has established that the
1816 * allowed per node queues are empty and that nodes are
1819 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1823 wake_all_kswapd(order, zonelist, high_zoneidx);
1826 * OK, we're below the kswapd watermark and have kicked background
1827 * reclaim. Now things get more complex, so set up alloc_flags according
1828 * to how we want to proceed.
1830 alloc_flags = gfp_to_alloc_flags(gfp_mask);
1832 /* This is the last chance, in general, before the goto nopage. */
1833 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
1834 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
1835 preferred_zone, migratetype);
1840 /* Allocate without watermarks if the context allows */
1841 if (alloc_flags & ALLOC_NO_WATERMARKS) {
1842 page = __alloc_pages_high_priority(gfp_mask, order,
1843 zonelist, high_zoneidx, nodemask,
1844 preferred_zone, migratetype);
1849 /* Atomic allocations - we can't balance anything */
1853 /* Avoid recursion of direct reclaim */
1854 if (p->flags & PF_MEMALLOC)
1857 /* Avoid allocations with no watermarks from looping endlessly */
1858 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
1861 /* Try direct reclaim and then allocating */
1862 page = __alloc_pages_direct_reclaim(gfp_mask, order,
1863 zonelist, high_zoneidx,
1865 alloc_flags, preferred_zone,
1866 migratetype, &did_some_progress);
1871 * If we failed to make any progress reclaiming, then we are
1872 * running out of options and have to consider going OOM
1874 if (!did_some_progress) {
1875 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1876 if (oom_killer_disabled)
1878 page = __alloc_pages_may_oom(gfp_mask, order,
1879 zonelist, high_zoneidx,
1880 nodemask, preferred_zone,
1886 * The OOM killer does not trigger for high-order
1887 * ~__GFP_NOFAIL allocations so if no progress is being
1888 * made, there are no other options and retrying is
1891 if (order > PAGE_ALLOC_COSTLY_ORDER &&
1892 !(gfp_mask & __GFP_NOFAIL))
1899 /* Check if we should retry the allocation */
1900 pages_reclaimed += did_some_progress;
1901 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
1902 /* Wait for some write requests to complete then retry */
1903 congestion_wait(BLK_RW_ASYNC, HZ/50);
1908 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1909 printk(KERN_WARNING "%s: page allocation failure."
1910 " order:%d, mode:0x%x\n",
1911 p->comm, order, gfp_mask);
1917 if (kmemcheck_enabled)
1918 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
1924 * This is the 'heart' of the zoned buddy allocator.
1927 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
1928 struct zonelist *zonelist, nodemask_t *nodemask)
1930 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1931 struct zone *preferred_zone;
1933 int migratetype = allocflags_to_migratetype(gfp_mask);
1935 gfp_mask &= gfp_allowed_mask;
1937 lockdep_trace_alloc(gfp_mask);
1939 might_sleep_if(gfp_mask & __GFP_WAIT);
1941 if (should_fail_alloc_page(gfp_mask, order))
1945 * Check the zones suitable for the gfp_mask contain at least one
1946 * valid zone. It's possible to have an empty zonelist as a result
1947 * of GFP_THISNODE and a memoryless node
1949 if (unlikely(!zonelist->_zonerefs->zone))
1952 /* The preferred zone is used for statistics later */
1953 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone);
1954 if (!preferred_zone)
1957 /* First allocation attempt */
1958 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
1959 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
1960 preferred_zone, migratetype);
1961 if (unlikely(!page))
1962 page = __alloc_pages_slowpath(gfp_mask, order,
1963 zonelist, high_zoneidx, nodemask,
1964 preferred_zone, migratetype);
1966 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
1969 EXPORT_SYMBOL(__alloc_pages_nodemask);
1972 * Common helper functions.
1974 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1979 * __get_free_pages() returns a 32-bit address, which cannot represent
1982 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1984 page = alloc_pages(gfp_mask, order);
1987 return (unsigned long) page_address(page);
1989 EXPORT_SYMBOL(__get_free_pages);
1991 unsigned long get_zeroed_page(gfp_t gfp_mask)
1993 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
1995 EXPORT_SYMBOL(get_zeroed_page);
1997 void __pagevec_free(struct pagevec *pvec)
1999 int i = pagevec_count(pvec);
2002 trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
2003 free_hot_cold_page(pvec->pages[i], pvec->cold);
2007 void __free_pages(struct page *page, unsigned int order)
2009 if (put_page_testzero(page)) {
2010 trace_mm_page_free_direct(page, order);
2012 free_hot_page(page);
2014 __free_pages_ok(page, order);
2018 EXPORT_SYMBOL(__free_pages);
2020 void free_pages(unsigned long addr, unsigned int order)
2023 VM_BUG_ON(!virt_addr_valid((void *)addr));
2024 __free_pages(virt_to_page((void *)addr), order);
2028 EXPORT_SYMBOL(free_pages);
2031 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2032 * @size: the number of bytes to allocate
2033 * @gfp_mask: GFP flags for the allocation
2035 * This function is similar to alloc_pages(), except that it allocates the
2036 * minimum number of pages to satisfy the request. alloc_pages() can only
2037 * allocate memory in power-of-two pages.
2039 * This function is also limited by MAX_ORDER.
2041 * Memory allocated by this function must be released by free_pages_exact().
2043 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2045 unsigned int order = get_order(size);
2048 addr = __get_free_pages(gfp_mask, order);
2050 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2051 unsigned long used = addr + PAGE_ALIGN(size);
2053 split_page(virt_to_page((void *)addr), order);
2054 while (used < alloc_end) {
2060 return (void *)addr;
2062 EXPORT_SYMBOL(alloc_pages_exact);
2065 * free_pages_exact - release memory allocated via alloc_pages_exact()
2066 * @virt: the value returned by alloc_pages_exact.
2067 * @size: size of allocation, same value as passed to alloc_pages_exact().
2069 * Release the memory allocated by a previous call to alloc_pages_exact.
2071 void free_pages_exact(void *virt, size_t size)
2073 unsigned long addr = (unsigned long)virt;
2074 unsigned long end = addr + PAGE_ALIGN(size);
2076 while (addr < end) {
2081 EXPORT_SYMBOL(free_pages_exact);
2083 static unsigned int nr_free_zone_pages(int offset)
2088 /* Just pick one node, since fallback list is circular */
2089 unsigned int sum = 0;
2091 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2093 for_each_zone_zonelist(zone, z, zonelist, offset) {
2094 unsigned long size = zone->present_pages;
2095 unsigned long high = high_wmark_pages(zone);
2104 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2106 unsigned int nr_free_buffer_pages(void)
2108 return nr_free_zone_pages(gfp_zone(GFP_USER));
2110 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2113 * Amount of free RAM allocatable within all zones
2115 unsigned int nr_free_pagecache_pages(void)
2117 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2120 static inline void show_node(struct zone *zone)
2123 printk("Node %d ", zone_to_nid(zone));
2126 void si_meminfo(struct sysinfo *val)
2128 val->totalram = totalram_pages;
2130 val->freeram = global_page_state(NR_FREE_PAGES);
2131 val->bufferram = nr_blockdev_pages();
2132 val->totalhigh = totalhigh_pages;
2133 val->freehigh = nr_free_highpages();
2134 val->mem_unit = PAGE_SIZE;
2137 EXPORT_SYMBOL(si_meminfo);
2140 void si_meminfo_node(struct sysinfo *val, int nid)
2142 pg_data_t *pgdat = NODE_DATA(nid);
2144 val->totalram = pgdat->node_present_pages;
2145 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2146 #ifdef CONFIG_HIGHMEM
2147 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2148 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2154 val->mem_unit = PAGE_SIZE;
2158 #define K(x) ((x) << (PAGE_SHIFT-10))
2161 * Show free area list (used inside shift_scroll-lock stuff)
2162 * We also calculate the percentage fragmentation. We do this by counting the
2163 * memory on each free list with the exception of the first item on the list.
2165 void show_free_areas(void)
2170 for_each_populated_zone(zone) {
2172 printk("%s per-cpu:\n", zone->name);
2174 for_each_online_cpu(cpu) {
2175 struct per_cpu_pageset *pageset;
2177 pageset = per_cpu_ptr(zone->pageset, cpu);
2179 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2180 cpu, pageset->pcp.high,
2181 pageset->pcp.batch, pageset->pcp.count);
2185 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2186 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2188 " dirty:%lu writeback:%lu unstable:%lu\n"
2189 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2190 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2191 global_page_state(NR_ACTIVE_ANON),
2192 global_page_state(NR_INACTIVE_ANON),
2193 global_page_state(NR_ISOLATED_ANON),
2194 global_page_state(NR_ACTIVE_FILE),
2195 global_page_state(NR_INACTIVE_FILE),
2196 global_page_state(NR_ISOLATED_FILE),
2197 global_page_state(NR_UNEVICTABLE),
2198 global_page_state(NR_FILE_DIRTY),
2199 global_page_state(NR_WRITEBACK),
2200 global_page_state(NR_UNSTABLE_NFS),
2201 global_page_state(NR_FREE_PAGES),
2202 global_page_state(NR_SLAB_RECLAIMABLE),
2203 global_page_state(NR_SLAB_UNRECLAIMABLE),
2204 global_page_state(NR_FILE_MAPPED),
2205 global_page_state(NR_SHMEM),
2206 global_page_state(NR_PAGETABLE),
2207 global_page_state(NR_BOUNCE));
2209 for_each_populated_zone(zone) {
2218 " active_anon:%lukB"
2219 " inactive_anon:%lukB"
2220 " active_file:%lukB"
2221 " inactive_file:%lukB"
2222 " unevictable:%lukB"
2223 " isolated(anon):%lukB"
2224 " isolated(file):%lukB"
2231 " slab_reclaimable:%lukB"
2232 " slab_unreclaimable:%lukB"
2233 " kernel_stack:%lukB"
2237 " writeback_tmp:%lukB"
2238 " pages_scanned:%lu"
2239 " all_unreclaimable? %s"
2242 K(zone_page_state(zone, NR_FREE_PAGES)),
2243 K(min_wmark_pages(zone)),
2244 K(low_wmark_pages(zone)),
2245 K(high_wmark_pages(zone)),
2246 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2247 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2248 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2249 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2250 K(zone_page_state(zone, NR_UNEVICTABLE)),
2251 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2252 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2253 K(zone->present_pages),
2254 K(zone_page_state(zone, NR_MLOCK)),
2255 K(zone_page_state(zone, NR_FILE_DIRTY)),
2256 K(zone_page_state(zone, NR_WRITEBACK)),
2257 K(zone_page_state(zone, NR_FILE_MAPPED)),
2258 K(zone_page_state(zone, NR_SHMEM)),
2259 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2260 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2261 zone_page_state(zone, NR_KERNEL_STACK) *
2263 K(zone_page_state(zone, NR_PAGETABLE)),
2264 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2265 K(zone_page_state(zone, NR_BOUNCE)),
2266 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2267 zone->pages_scanned,
2268 (zone_is_all_unreclaimable(zone) ? "yes" : "no")
2270 printk("lowmem_reserve[]:");
2271 for (i = 0; i < MAX_NR_ZONES; i++)
2272 printk(" %lu", zone->lowmem_reserve[i]);
2276 for_each_populated_zone(zone) {
2277 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2280 printk("%s: ", zone->name);
2282 spin_lock_irqsave(&zone->lock, flags);
2283 for (order = 0; order < MAX_ORDER; order++) {
2284 nr[order] = zone->free_area[order].nr_free;
2285 total += nr[order] << order;
2287 spin_unlock_irqrestore(&zone->lock, flags);
2288 for (order = 0; order < MAX_ORDER; order++)
2289 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2290 printk("= %lukB\n", K(total));
2293 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2295 show_swap_cache_info();
2298 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2300 zoneref->zone = zone;
2301 zoneref->zone_idx = zone_idx(zone);
2305 * Builds allocation fallback zone lists.
2307 * Add all populated zones of a node to the zonelist.
2309 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2310 int nr_zones, enum zone_type zone_type)
2314 BUG_ON(zone_type >= MAX_NR_ZONES);
2319 zone = pgdat->node_zones + zone_type;
2320 if (populated_zone(zone)) {
2321 zoneref_set_zone(zone,
2322 &zonelist->_zonerefs[nr_zones++]);
2323 check_highest_zone(zone_type);
2326 } while (zone_type);
2333 * 0 = automatic detection of better ordering.
2334 * 1 = order by ([node] distance, -zonetype)
2335 * 2 = order by (-zonetype, [node] distance)
2337 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2338 * the same zonelist. So only NUMA can configure this param.
2340 #define ZONELIST_ORDER_DEFAULT 0
2341 #define ZONELIST_ORDER_NODE 1
2342 #define ZONELIST_ORDER_ZONE 2
2344 /* zonelist order in the kernel.
2345 * set_zonelist_order() will set this to NODE or ZONE.
2347 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2348 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2352 /* The value user specified ....changed by config */
2353 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2354 /* string for sysctl */
2355 #define NUMA_ZONELIST_ORDER_LEN 16
2356 char numa_zonelist_order[16] = "default";
2359 * interface for configure zonelist ordering.
2360 * command line option "numa_zonelist_order"
2361 * = "[dD]efault - default, automatic configuration.
2362 * = "[nN]ode - order by node locality, then by zone within node
2363 * = "[zZ]one - order by zone, then by locality within zone
2366 static int __parse_numa_zonelist_order(char *s)
2368 if (*s == 'd' || *s == 'D') {
2369 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2370 } else if (*s == 'n' || *s == 'N') {
2371 user_zonelist_order = ZONELIST_ORDER_NODE;
2372 } else if (*s == 'z' || *s == 'Z') {
2373 user_zonelist_order = ZONELIST_ORDER_ZONE;
2376 "Ignoring invalid numa_zonelist_order value: "
2383 static __init int setup_numa_zonelist_order(char *s)
2386 return __parse_numa_zonelist_order(s);
2389 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2392 * sysctl handler for numa_zonelist_order
2394 int numa_zonelist_order_handler(ctl_table *table, int write,
2395 void __user *buffer, size_t *length,
2398 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2400 static DEFINE_MUTEX(zl_order_mutex);
2402 mutex_lock(&zl_order_mutex);
2404 strcpy(saved_string, (char*)table->data);
2405 ret = proc_dostring(table, write, buffer, length, ppos);
2409 int oldval = user_zonelist_order;
2410 if (__parse_numa_zonelist_order((char*)table->data)) {
2412 * bogus value. restore saved string
2414 strncpy((char*)table->data, saved_string,
2415 NUMA_ZONELIST_ORDER_LEN);
2416 user_zonelist_order = oldval;
2417 } else if (oldval != user_zonelist_order)
2418 build_all_zonelists();
2421 mutex_unlock(&zl_order_mutex);
2426 #define MAX_NODE_LOAD (nr_online_nodes)
2427 static int node_load[MAX_NUMNODES];
2430 * find_next_best_node - find the next node that should appear in a given node's fallback list
2431 * @node: node whose fallback list we're appending
2432 * @used_node_mask: nodemask_t of already used nodes
2434 * We use a number of factors to determine which is the next node that should
2435 * appear on a given node's fallback list. The node should not have appeared
2436 * already in @node's fallback list, and it should be the next closest node
2437 * according to the distance array (which contains arbitrary distance values
2438 * from each node to each node in the system), and should also prefer nodes
2439 * with no CPUs, since presumably they'll have very little allocation pressure
2440 * on them otherwise.
2441 * It returns -1 if no node is found.
2443 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2446 int min_val = INT_MAX;
2448 const struct cpumask *tmp = cpumask_of_node(0);
2450 /* Use the local node if we haven't already */
2451 if (!node_isset(node, *used_node_mask)) {
2452 node_set(node, *used_node_mask);
2456 for_each_node_state(n, N_HIGH_MEMORY) {
2458 /* Don't want a node to appear more than once */
2459 if (node_isset(n, *used_node_mask))
2462 /* Use the distance array to find the distance */
2463 val = node_distance(node, n);
2465 /* Penalize nodes under us ("prefer the next node") */
2468 /* Give preference to headless and unused nodes */
2469 tmp = cpumask_of_node(n);
2470 if (!cpumask_empty(tmp))
2471 val += PENALTY_FOR_NODE_WITH_CPUS;
2473 /* Slight preference for less loaded node */
2474 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2475 val += node_load[n];
2477 if (val < min_val) {
2484 node_set(best_node, *used_node_mask);
2491 * Build zonelists ordered by node and zones within node.
2492 * This results in maximum locality--normal zone overflows into local
2493 * DMA zone, if any--but risks exhausting DMA zone.
2495 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2498 struct zonelist *zonelist;
2500 zonelist = &pgdat->node_zonelists[0];
2501 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2503 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2505 zonelist->_zonerefs[j].zone = NULL;
2506 zonelist->_zonerefs[j].zone_idx = 0;
2510 * Build gfp_thisnode zonelists
2512 static void build_thisnode_zonelists(pg_data_t *pgdat)
2515 struct zonelist *zonelist;
2517 zonelist = &pgdat->node_zonelists[1];
2518 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2519 zonelist->_zonerefs[j].zone = NULL;
2520 zonelist->_zonerefs[j].zone_idx = 0;
2524 * Build zonelists ordered by zone and nodes within zones.
2525 * This results in conserving DMA zone[s] until all Normal memory is
2526 * exhausted, but results in overflowing to remote node while memory
2527 * may still exist in local DMA zone.
2529 static int node_order[MAX_NUMNODES];
2531 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2534 int zone_type; /* needs to be signed */
2536 struct zonelist *zonelist;
2538 zonelist = &pgdat->node_zonelists[0];
2540 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2541 for (j = 0; j < nr_nodes; j++) {
2542 node = node_order[j];
2543 z = &NODE_DATA(node)->node_zones[zone_type];
2544 if (populated_zone(z)) {
2546 &zonelist->_zonerefs[pos++]);
2547 check_highest_zone(zone_type);
2551 zonelist->_zonerefs[pos].zone = NULL;
2552 zonelist->_zonerefs[pos].zone_idx = 0;
2555 static int default_zonelist_order(void)
2558 unsigned long low_kmem_size,total_size;
2562 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2563 * If they are really small and used heavily, the system can fall
2564 * into OOM very easily.
2565 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2567 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2570 for_each_online_node(nid) {
2571 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2572 z = &NODE_DATA(nid)->node_zones[zone_type];
2573 if (populated_zone(z)) {
2574 if (zone_type < ZONE_NORMAL)
2575 low_kmem_size += z->present_pages;
2576 total_size += z->present_pages;
2580 if (!low_kmem_size || /* there are no DMA area. */
2581 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2582 return ZONELIST_ORDER_NODE;
2584 * look into each node's config.
2585 * If there is a node whose DMA/DMA32 memory is very big area on
2586 * local memory, NODE_ORDER may be suitable.
2588 average_size = total_size /
2589 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2590 for_each_online_node(nid) {
2593 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2594 z = &NODE_DATA(nid)->node_zones[zone_type];
2595 if (populated_zone(z)) {
2596 if (zone_type < ZONE_NORMAL)
2597 low_kmem_size += z->present_pages;
2598 total_size += z->present_pages;
2601 if (low_kmem_size &&
2602 total_size > average_size && /* ignore small node */
2603 low_kmem_size > total_size * 70/100)
2604 return ZONELIST_ORDER_NODE;
2606 return ZONELIST_ORDER_ZONE;
2609 static void set_zonelist_order(void)
2611 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2612 current_zonelist_order = default_zonelist_order();
2614 current_zonelist_order = user_zonelist_order;
2617 static void build_zonelists(pg_data_t *pgdat)
2621 nodemask_t used_mask;
2622 int local_node, prev_node;
2623 struct zonelist *zonelist;
2624 int order = current_zonelist_order;
2626 /* initialize zonelists */
2627 for (i = 0; i < MAX_ZONELISTS; i++) {
2628 zonelist = pgdat->node_zonelists + i;
2629 zonelist->_zonerefs[0].zone = NULL;
2630 zonelist->_zonerefs[0].zone_idx = 0;
2633 /* NUMA-aware ordering of nodes */
2634 local_node = pgdat->node_id;
2635 load = nr_online_nodes;
2636 prev_node = local_node;
2637 nodes_clear(used_mask);
2639 memset(node_order, 0, sizeof(node_order));
2642 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2643 int distance = node_distance(local_node, node);
2646 * If another node is sufficiently far away then it is better
2647 * to reclaim pages in a zone before going off node.
2649 if (distance > RECLAIM_DISTANCE)
2650 zone_reclaim_mode = 1;
2653 * We don't want to pressure a particular node.
2654 * So adding penalty to the first node in same
2655 * distance group to make it round-robin.
2657 if (distance != node_distance(local_node, prev_node))
2658 node_load[node] = load;
2662 if (order == ZONELIST_ORDER_NODE)
2663 build_zonelists_in_node_order(pgdat, node);
2665 node_order[j++] = node; /* remember order */
2668 if (order == ZONELIST_ORDER_ZONE) {
2669 /* calculate node order -- i.e., DMA last! */
2670 build_zonelists_in_zone_order(pgdat, j);
2673 build_thisnode_zonelists(pgdat);
2676 /* Construct the zonelist performance cache - see further mmzone.h */
2677 static void build_zonelist_cache(pg_data_t *pgdat)
2679 struct zonelist *zonelist;
2680 struct zonelist_cache *zlc;
2683 zonelist = &pgdat->node_zonelists[0];
2684 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2685 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2686 for (z = zonelist->_zonerefs; z->zone; z++)
2687 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2691 #else /* CONFIG_NUMA */
2693 static void set_zonelist_order(void)
2695 current_zonelist_order = ZONELIST_ORDER_ZONE;
2698 static void build_zonelists(pg_data_t *pgdat)
2700 int node, local_node;
2702 struct zonelist *zonelist;
2704 local_node = pgdat->node_id;
2706 zonelist = &pgdat->node_zonelists[0];
2707 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2710 * Now we build the zonelist so that it contains the zones
2711 * of all the other nodes.
2712 * We don't want to pressure a particular node, so when
2713 * building the zones for node N, we make sure that the
2714 * zones coming right after the local ones are those from
2715 * node N+1 (modulo N)
2717 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2718 if (!node_online(node))
2720 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2723 for (node = 0; node < local_node; node++) {
2724 if (!node_online(node))
2726 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2730 zonelist->_zonerefs[j].zone = NULL;
2731 zonelist->_zonerefs[j].zone_idx = 0;
2734 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2735 static void build_zonelist_cache(pg_data_t *pgdat)
2737 pgdat->node_zonelists[0].zlcache_ptr = NULL;
2740 #endif /* CONFIG_NUMA */
2743 * Boot pageset table. One per cpu which is going to be used for all
2744 * zones and all nodes. The parameters will be set in such a way
2745 * that an item put on a list will immediately be handed over to
2746 * the buddy list. This is safe since pageset manipulation is done
2747 * with interrupts disabled.
2749 * The boot_pagesets must be kept even after bootup is complete for
2750 * unused processors and/or zones. They do play a role for bootstrapping
2751 * hotplugged processors.
2753 * zoneinfo_show() and maybe other functions do
2754 * not check if the processor is online before following the pageset pointer.
2755 * Other parts of the kernel may not check if the zone is available.
2757 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
2758 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
2760 /* return values int ....just for stop_machine() */
2761 static int __build_all_zonelists(void *dummy)
2767 memset(node_load, 0, sizeof(node_load));
2769 for_each_online_node(nid) {
2770 pg_data_t *pgdat = NODE_DATA(nid);
2772 build_zonelists(pgdat);
2773 build_zonelist_cache(pgdat);
2777 * Initialize the boot_pagesets that are going to be used
2778 * for bootstrapping processors. The real pagesets for
2779 * each zone will be allocated later when the per cpu
2780 * allocator is available.
2782 * boot_pagesets are used also for bootstrapping offline
2783 * cpus if the system is already booted because the pagesets
2784 * are needed to initialize allocators on a specific cpu too.
2785 * F.e. the percpu allocator needs the page allocator which
2786 * needs the percpu allocator in order to allocate its pagesets
2787 * (a chicken-egg dilemma).
2789 for_each_possible_cpu(cpu)
2790 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
2795 void build_all_zonelists(void)
2797 set_zonelist_order();
2799 if (system_state == SYSTEM_BOOTING) {
2800 __build_all_zonelists(NULL);
2801 mminit_verify_zonelist();
2802 cpuset_init_current_mems_allowed();
2804 /* we have to stop all cpus to guarantee there is no user
2806 stop_machine(__build_all_zonelists, NULL, NULL);
2807 /* cpuset refresh routine should be here */
2809 vm_total_pages = nr_free_pagecache_pages();
2811 * Disable grouping by mobility if the number of pages in the
2812 * system is too low to allow the mechanism to work. It would be
2813 * more accurate, but expensive to check per-zone. This check is
2814 * made on memory-hotadd so a system can start with mobility
2815 * disabled and enable it later
2817 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
2818 page_group_by_mobility_disabled = 1;
2820 page_group_by_mobility_disabled = 0;
2822 printk("Built %i zonelists in %s order, mobility grouping %s. "
2823 "Total pages: %ld\n",
2825 zonelist_order_name[current_zonelist_order],
2826 page_group_by_mobility_disabled ? "off" : "on",
2829 printk("Policy zone: %s\n", zone_names[policy_zone]);
2834 * Helper functions to size the waitqueue hash table.
2835 * Essentially these want to choose hash table sizes sufficiently
2836 * large so that collisions trying to wait on pages are rare.
2837 * But in fact, the number of active page waitqueues on typical
2838 * systems is ridiculously low, less than 200. So this is even
2839 * conservative, even though it seems large.
2841 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2842 * waitqueues, i.e. the size of the waitq table given the number of pages.
2844 #define PAGES_PER_WAITQUEUE 256
2846 #ifndef CONFIG_MEMORY_HOTPLUG
2847 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2849 unsigned long size = 1;
2851 pages /= PAGES_PER_WAITQUEUE;
2853 while (size < pages)
2857 * Once we have dozens or even hundreds of threads sleeping
2858 * on IO we've got bigger problems than wait queue collision.
2859 * Limit the size of the wait table to a reasonable size.
2861 size = min(size, 4096UL);
2863 return max(size, 4UL);
2867 * A zone's size might be changed by hot-add, so it is not possible to determine
2868 * a suitable size for its wait_table. So we use the maximum size now.
2870 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2872 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2873 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2874 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2876 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2877 * or more by the traditional way. (See above). It equals:
2879 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2880 * ia64(16K page size) : = ( 8G + 4M)byte.
2881 * powerpc (64K page size) : = (32G +16M)byte.
2883 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2890 * This is an integer logarithm so that shifts can be used later
2891 * to extract the more random high bits from the multiplicative
2892 * hash function before the remainder is taken.
2894 static inline unsigned long wait_table_bits(unsigned long size)
2899 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2902 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2903 * of blocks reserved is based on min_wmark_pages(zone). The memory within
2904 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
2905 * higher will lead to a bigger reserve which will get freed as contiguous
2906 * blocks as reclaim kicks in
2908 static void setup_zone_migrate_reserve(struct zone *zone)
2910 unsigned long start_pfn, pfn, end_pfn;
2912 unsigned long block_migratetype;
2915 /* Get the start pfn, end pfn and the number of blocks to reserve */
2916 start_pfn = zone->zone_start_pfn;
2917 end_pfn = start_pfn + zone->spanned_pages;
2918 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
2922 * Reserve blocks are generally in place to help high-order atomic
2923 * allocations that are short-lived. A min_free_kbytes value that
2924 * would result in more than 2 reserve blocks for atomic allocations
2925 * is assumed to be in place to help anti-fragmentation for the
2926 * future allocation of hugepages at runtime.
2928 reserve = min(2, reserve);
2930 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
2931 if (!pfn_valid(pfn))
2933 page = pfn_to_page(pfn);
2935 /* Watch out for overlapping nodes */
2936 if (page_to_nid(page) != zone_to_nid(zone))
2939 /* Blocks with reserved pages will never free, skip them. */
2940 if (PageReserved(page))
2943 block_migratetype = get_pageblock_migratetype(page);
2945 /* If this block is reserved, account for it */
2946 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
2951 /* Suitable for reserving if this block is movable */
2952 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
2953 set_pageblock_migratetype(page, MIGRATE_RESERVE);
2954 move_freepages_block(zone, page, MIGRATE_RESERVE);
2960 * If the reserve is met and this is a previous reserved block,
2963 if (block_migratetype == MIGRATE_RESERVE) {
2964 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2965 move_freepages_block(zone, page, MIGRATE_MOVABLE);
2971 * Initially all pages are reserved - free ones are freed
2972 * up by free_all_bootmem() once the early boot process is
2973 * done. Non-atomic initialization, single-pass.
2975 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
2976 unsigned long start_pfn, enum memmap_context context)
2979 unsigned long end_pfn = start_pfn + size;
2983 if (highest_memmap_pfn < end_pfn - 1)
2984 highest_memmap_pfn = end_pfn - 1;
2986 z = &NODE_DATA(nid)->node_zones[zone];
2987 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
2989 * There can be holes in boot-time mem_map[]s
2990 * handed to this function. They do not
2991 * exist on hotplugged memory.
2993 if (context == MEMMAP_EARLY) {
2994 if (!early_pfn_valid(pfn))
2996 if (!early_pfn_in_nid(pfn, nid))
2999 page = pfn_to_page(pfn);
3000 set_page_links(page, zone, nid, pfn);
3001 mminit_verify_page_links(page, zone, nid, pfn);
3002 init_page_count(page);
3003 reset_page_mapcount(page);
3004 SetPageReserved(page);
3006 * Mark the block movable so that blocks are reserved for
3007 * movable at startup. This will force kernel allocations
3008 * to reserve their blocks rather than leaking throughout
3009 * the address space during boot when many long-lived
3010 * kernel allocations are made. Later some blocks near
3011 * the start are marked MIGRATE_RESERVE by
3012 * setup_zone_migrate_reserve()
3014 * bitmap is created for zone's valid pfn range. but memmap
3015 * can be created for invalid pages (for alignment)
3016 * check here not to call set_pageblock_migratetype() against
3019 if ((z->zone_start_pfn <= pfn)
3020 && (pfn < z->zone_start_pfn + z->spanned_pages)
3021 && !(pfn & (pageblock_nr_pages - 1)))
3022 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3024 INIT_LIST_HEAD(&page->lru);
3025 #ifdef WANT_PAGE_VIRTUAL
3026 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3027 if (!is_highmem_idx(zone))
3028 set_page_address(page, __va(pfn << PAGE_SHIFT));
3033 static void __meminit zone_init_free_lists(struct zone *zone)
3036 for_each_migratetype_order(order, t) {
3037 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3038 zone->free_area[order].nr_free = 0;
3042 #ifndef __HAVE_ARCH_MEMMAP_INIT
3043 #define memmap_init(size, nid, zone, start_pfn) \
3044 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3047 static int zone_batchsize(struct zone *zone)
3053 * The per-cpu-pages pools are set to around 1000th of the
3054 * size of the zone. But no more than 1/2 of a meg.
3056 * OK, so we don't know how big the cache is. So guess.
3058 batch = zone->present_pages / 1024;
3059 if (batch * PAGE_SIZE > 512 * 1024)
3060 batch = (512 * 1024) / PAGE_SIZE;
3061 batch /= 4; /* We effectively *= 4 below */
3066 * Clamp the batch to a 2^n - 1 value. Having a power
3067 * of 2 value was found to be more likely to have
3068 * suboptimal cache aliasing properties in some cases.
3070 * For example if 2 tasks are alternately allocating
3071 * batches of pages, one task can end up with a lot
3072 * of pages of one half of the possible page colors
3073 * and the other with pages of the other colors.
3075 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3080 /* The deferral and batching of frees should be suppressed under NOMMU
3083 * The problem is that NOMMU needs to be able to allocate large chunks
3084 * of contiguous memory as there's no hardware page translation to
3085 * assemble apparent contiguous memory from discontiguous pages.
3087 * Queueing large contiguous runs of pages for batching, however,
3088 * causes the pages to actually be freed in smaller chunks. As there
3089 * can be a significant delay between the individual batches being
3090 * recycled, this leads to the once large chunks of space being
3091 * fragmented and becoming unavailable for high-order allocations.
3097 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3099 struct per_cpu_pages *pcp;
3102 memset(p, 0, sizeof(*p));
3106 pcp->high = 6 * batch;
3107 pcp->batch = max(1UL, 1 * batch);
3108 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3109 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3113 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3114 * to the value high for the pageset p.
3117 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3120 struct per_cpu_pages *pcp;
3124 pcp->batch = max(1UL, high/4);
3125 if ((high/4) > (PAGE_SHIFT * 8))
3126 pcp->batch = PAGE_SHIFT * 8;
3130 * Allocate per cpu pagesets and initialize them.
3131 * Before this call only boot pagesets were available.
3132 * Boot pagesets will no longer be used by this processorr
3133 * after setup_per_cpu_pageset().
3135 void __init setup_per_cpu_pageset(void)
3140 for_each_populated_zone(zone) {
3141 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3143 for_each_possible_cpu(cpu) {
3144 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3146 setup_pageset(pcp, zone_batchsize(zone));
3148 if (percpu_pagelist_fraction)
3149 setup_pagelist_highmark(pcp,
3150 (zone->present_pages /
3151 percpu_pagelist_fraction));
3156 static noinline __init_refok
3157 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3160 struct pglist_data *pgdat = zone->zone_pgdat;
3164 * The per-page waitqueue mechanism uses hashed waitqueues
3167 zone->wait_table_hash_nr_entries =
3168 wait_table_hash_nr_entries(zone_size_pages);
3169 zone->wait_table_bits =
3170 wait_table_bits(zone->wait_table_hash_nr_entries);
3171 alloc_size = zone->wait_table_hash_nr_entries
3172 * sizeof(wait_queue_head_t);
3174 if (!slab_is_available()) {
3175 zone->wait_table = (wait_queue_head_t *)
3176 alloc_bootmem_node(pgdat, alloc_size);
3179 * This case means that a zone whose size was 0 gets new memory
3180 * via memory hot-add.
3181 * But it may be the case that a new node was hot-added. In
3182 * this case vmalloc() will not be able to use this new node's
3183 * memory - this wait_table must be initialized to use this new
3184 * node itself as well.
3185 * To use this new node's memory, further consideration will be
3188 zone->wait_table = vmalloc(alloc_size);
3190 if (!zone->wait_table)
3193 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3194 init_waitqueue_head(zone->wait_table + i);
3199 static int __zone_pcp_update(void *data)
3201 struct zone *zone = data;
3203 unsigned long batch = zone_batchsize(zone), flags;
3205 for (cpu = 0; cpu < NR_CPUS; cpu++) {
3206 struct per_cpu_pageset *pset;
3207 struct per_cpu_pages *pcp;
3209 pset = per_cpu_ptr(zone->pageset, cpu);
3212 local_irq_save(flags);
3213 free_pcppages_bulk(zone, pcp->count, pcp);
3214 setup_pageset(pset, batch);
3215 local_irq_restore(flags);
3220 void zone_pcp_update(struct zone *zone)
3222 stop_machine(__zone_pcp_update, zone, NULL);
3225 static __meminit void zone_pcp_init(struct zone *zone)
3228 * per cpu subsystem is not up at this point. The following code
3229 * relies on the ability of the linker to provide the
3230 * offset of a (static) per cpu variable into the per cpu area.
3232 zone->pageset = &boot_pageset;
3234 if (zone->present_pages)
3235 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3236 zone->name, zone->present_pages,
3237 zone_batchsize(zone));
3240 __meminit int init_currently_empty_zone(struct zone *zone,
3241 unsigned long zone_start_pfn,
3243 enum memmap_context context)
3245 struct pglist_data *pgdat = zone->zone_pgdat;
3247 ret = zone_wait_table_init(zone, size);
3250 pgdat->nr_zones = zone_idx(zone) + 1;
3252 zone->zone_start_pfn = zone_start_pfn;
3254 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3255 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3257 (unsigned long)zone_idx(zone),
3258 zone_start_pfn, (zone_start_pfn + size));
3260 zone_init_free_lists(zone);
3265 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3267 * Basic iterator support. Return the first range of PFNs for a node
3268 * Note: nid == MAX_NUMNODES returns first region regardless of node
3270 static int __meminit first_active_region_index_in_nid(int nid)
3274 for (i = 0; i < nr_nodemap_entries; i++)
3275 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3282 * Basic iterator support. Return the next active range of PFNs for a node
3283 * Note: nid == MAX_NUMNODES returns next region regardless of node
3285 static int __meminit next_active_region_index_in_nid(int index, int nid)
3287 for (index = index + 1; index < nr_nodemap_entries; index++)
3288 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3294 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3296 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3297 * Architectures may implement their own version but if add_active_range()
3298 * was used and there are no special requirements, this is a convenient
3301 int __meminit __early_pfn_to_nid(unsigned long pfn)
3305 for (i = 0; i < nr_nodemap_entries; i++) {
3306 unsigned long start_pfn = early_node_map[i].start_pfn;
3307 unsigned long end_pfn = early_node_map[i].end_pfn;
3309 if (start_pfn <= pfn && pfn < end_pfn)
3310 return early_node_map[i].nid;
3312 /* This is a memory hole */
3315 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3317 int __meminit early_pfn_to_nid(unsigned long pfn)
3321 nid = __early_pfn_to_nid(pfn);
3324 /* just returns 0 */
3328 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3329 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3333 nid = __early_pfn_to_nid(pfn);
3334 if (nid >= 0 && nid != node)
3340 /* Basic iterator support to walk early_node_map[] */
3341 #define for_each_active_range_index_in_nid(i, nid) \
3342 for (i = first_active_region_index_in_nid(nid); i != -1; \
3343 i = next_active_region_index_in_nid(i, nid))
3346 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3347 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3348 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3350 * If an architecture guarantees that all ranges registered with
3351 * add_active_ranges() contain no holes and may be freed, this
3352 * this function may be used instead of calling free_bootmem() manually.
3354 void __init free_bootmem_with_active_regions(int nid,
3355 unsigned long max_low_pfn)
3359 for_each_active_range_index_in_nid(i, nid) {
3360 unsigned long size_pages = 0;
3361 unsigned long end_pfn = early_node_map[i].end_pfn;
3363 if (early_node_map[i].start_pfn >= max_low_pfn)
3366 if (end_pfn > max_low_pfn)
3367 end_pfn = max_low_pfn;
3369 size_pages = end_pfn - early_node_map[i].start_pfn;
3370 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3371 PFN_PHYS(early_node_map[i].start_pfn),
3372 size_pages << PAGE_SHIFT);
3376 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3381 for_each_active_range_index_in_nid(i, nid) {
3382 ret = work_fn(early_node_map[i].start_pfn,
3383 early_node_map[i].end_pfn, data);
3389 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3390 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3392 * If an architecture guarantees that all ranges registered with
3393 * add_active_ranges() contain no holes and may be freed, this
3394 * function may be used instead of calling memory_present() manually.
3396 void __init sparse_memory_present_with_active_regions(int nid)
3400 for_each_active_range_index_in_nid(i, nid)
3401 memory_present(early_node_map[i].nid,
3402 early_node_map[i].start_pfn,
3403 early_node_map[i].end_pfn);
3407 * get_pfn_range_for_nid - Return the start and end page frames for a node
3408 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3409 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3410 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3412 * It returns the start and end page frame of a node based on information
3413 * provided by an arch calling add_active_range(). If called for a node
3414 * with no available memory, a warning is printed and the start and end
3417 void __meminit get_pfn_range_for_nid(unsigned int nid,
3418 unsigned long *start_pfn, unsigned long *end_pfn)
3424 for_each_active_range_index_in_nid(i, nid) {
3425 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3426 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3429 if (*start_pfn == -1UL)
3434 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3435 * assumption is made that zones within a node are ordered in monotonic
3436 * increasing memory addresses so that the "highest" populated zone is used
3438 static void __init find_usable_zone_for_movable(void)
3441 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3442 if (zone_index == ZONE_MOVABLE)
3445 if (arch_zone_highest_possible_pfn[zone_index] >
3446 arch_zone_lowest_possible_pfn[zone_index])
3450 VM_BUG_ON(zone_index == -1);
3451 movable_zone = zone_index;
3455 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3456 * because it is sized independant of architecture. Unlike the other zones,
3457 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3458 * in each node depending on the size of each node and how evenly kernelcore
3459 * is distributed. This helper function adjusts the zone ranges
3460 * provided by the architecture for a given node by using the end of the
3461 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3462 * zones within a node are in order of monotonic increases memory addresses
3464 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3465 unsigned long zone_type,
3466 unsigned long node_start_pfn,
3467 unsigned long node_end_pfn,
3468 unsigned long *zone_start_pfn,
3469 unsigned long *zone_end_pfn)
3471 /* Only adjust if ZONE_MOVABLE is on this node */
3472 if (zone_movable_pfn[nid]) {
3473 /* Size ZONE_MOVABLE */
3474 if (zone_type == ZONE_MOVABLE) {
3475 *zone_start_pfn = zone_movable_pfn[nid];
3476 *zone_end_pfn = min(node_end_pfn,
3477 arch_zone_highest_possible_pfn[movable_zone]);
3479 /* Adjust for ZONE_MOVABLE starting within this range */
3480 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3481 *zone_end_pfn > zone_movable_pfn[nid]) {
3482 *zone_end_pfn = zone_movable_pfn[nid];
3484 /* Check if this whole range is within ZONE_MOVABLE */
3485 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3486 *zone_start_pfn = *zone_end_pfn;
3491 * Return the number of pages a zone spans in a node, including holes
3492 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3494 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3495 unsigned long zone_type,
3496 unsigned long *ignored)
3498 unsigned long node_start_pfn, node_end_pfn;
3499 unsigned long zone_start_pfn, zone_end_pfn;
3501 /* Get the start and end of the node and zone */
3502 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3503 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3504 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3505 adjust_zone_range_for_zone_movable(nid, zone_type,
3506 node_start_pfn, node_end_pfn,
3507 &zone_start_pfn, &zone_end_pfn);
3509 /* Check that this node has pages within the zone's required range */
3510 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3513 /* Move the zone boundaries inside the node if necessary */
3514 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3515 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3517 /* Return the spanned pages */
3518 return zone_end_pfn - zone_start_pfn;
3522 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3523 * then all holes in the requested range will be accounted for.
3525 unsigned long __meminit __absent_pages_in_range(int nid,
3526 unsigned long range_start_pfn,
3527 unsigned long range_end_pfn)
3530 unsigned long prev_end_pfn = 0, hole_pages = 0;
3531 unsigned long start_pfn;
3533 /* Find the end_pfn of the first active range of pfns in the node */
3534 i = first_active_region_index_in_nid(nid);
3538 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3540 /* Account for ranges before physical memory on this node */
3541 if (early_node_map[i].start_pfn > range_start_pfn)
3542 hole_pages = prev_end_pfn - range_start_pfn;
3544 /* Find all holes for the zone within the node */
3545 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3547 /* No need to continue if prev_end_pfn is outside the zone */
3548 if (prev_end_pfn >= range_end_pfn)
3551 /* Make sure the end of the zone is not within the hole */
3552 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3553 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3555 /* Update the hole size cound and move on */
3556 if (start_pfn > range_start_pfn) {
3557 BUG_ON(prev_end_pfn > start_pfn);
3558 hole_pages += start_pfn - prev_end_pfn;
3560 prev_end_pfn = early_node_map[i].end_pfn;
3563 /* Account for ranges past physical memory on this node */
3564 if (range_end_pfn > prev_end_pfn)
3565 hole_pages += range_end_pfn -
3566 max(range_start_pfn, prev_end_pfn);
3572 * absent_pages_in_range - Return number of page frames in holes within a range
3573 * @start_pfn: The start PFN to start searching for holes
3574 * @end_pfn: The end PFN to stop searching for holes
3576 * It returns the number of pages frames in memory holes within a range.
3578 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3579 unsigned long end_pfn)
3581 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3584 /* Return the number of page frames in holes in a zone on a node */
3585 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3586 unsigned long zone_type,
3587 unsigned long *ignored)
3589 unsigned long node_start_pfn, node_end_pfn;
3590 unsigned long zone_start_pfn, zone_end_pfn;
3592 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3593 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3595 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3598 adjust_zone_range_for_zone_movable(nid, zone_type,
3599 node_start_pfn, node_end_pfn,
3600 &zone_start_pfn, &zone_end_pfn);
3601 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3605 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3606 unsigned long zone_type,
3607 unsigned long *zones_size)
3609 return zones_size[zone_type];
3612 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3613 unsigned long zone_type,
3614 unsigned long *zholes_size)
3619 return zholes_size[zone_type];
3624 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3625 unsigned long *zones_size, unsigned long *zholes_size)
3627 unsigned long realtotalpages, totalpages = 0;
3630 for (i = 0; i < MAX_NR_ZONES; i++)
3631 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3633 pgdat->node_spanned_pages = totalpages;
3635 realtotalpages = totalpages;
3636 for (i = 0; i < MAX_NR_ZONES; i++)
3638 zone_absent_pages_in_node(pgdat->node_id, i,
3640 pgdat->node_present_pages = realtotalpages;
3641 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3645 #ifndef CONFIG_SPARSEMEM
3647 * Calculate the size of the zone->blockflags rounded to an unsigned long
3648 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3649 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3650 * round what is now in bits to nearest long in bits, then return it in
3653 static unsigned long __init usemap_size(unsigned long zonesize)
3655 unsigned long usemapsize;
3657 usemapsize = roundup(zonesize, pageblock_nr_pages);
3658 usemapsize = usemapsize >> pageblock_order;
3659 usemapsize *= NR_PAGEBLOCK_BITS;
3660 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
3662 return usemapsize / 8;
3665 static void __init setup_usemap(struct pglist_data *pgdat,
3666 struct zone *zone, unsigned long zonesize)
3668 unsigned long usemapsize = usemap_size(zonesize);
3669 zone->pageblock_flags = NULL;
3671 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
3674 static void inline setup_usemap(struct pglist_data *pgdat,
3675 struct zone *zone, unsigned long zonesize) {}
3676 #endif /* CONFIG_SPARSEMEM */
3678 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3680 /* Return a sensible default order for the pageblock size. */
3681 static inline int pageblock_default_order(void)
3683 if (HPAGE_SHIFT > PAGE_SHIFT)
3684 return HUGETLB_PAGE_ORDER;
3689 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3690 static inline void __init set_pageblock_order(unsigned int order)
3692 /* Check that pageblock_nr_pages has not already been setup */
3693 if (pageblock_order)
3697 * Assume the largest contiguous order of interest is a huge page.
3698 * This value may be variable depending on boot parameters on IA64
3700 pageblock_order = order;
3702 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3705 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3706 * and pageblock_default_order() are unused as pageblock_order is set
3707 * at compile-time. See include/linux/pageblock-flags.h for the values of
3708 * pageblock_order based on the kernel config
3710 static inline int pageblock_default_order(unsigned int order)
3714 #define set_pageblock_order(x) do {} while (0)
3716 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3719 * Set up the zone data structures:
3720 * - mark all pages reserved
3721 * - mark all memory queues empty
3722 * - clear the memory bitmaps
3724 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
3725 unsigned long *zones_size, unsigned long *zholes_size)
3728 int nid = pgdat->node_id;
3729 unsigned long zone_start_pfn = pgdat->node_start_pfn;
3732 pgdat_resize_init(pgdat);
3733 pgdat->nr_zones = 0;
3734 init_waitqueue_head(&pgdat->kswapd_wait);
3735 pgdat->kswapd_max_order = 0;
3736 pgdat_page_cgroup_init(pgdat);
3738 for (j = 0; j < MAX_NR_ZONES; j++) {
3739 struct zone *zone = pgdat->node_zones + j;
3740 unsigned long size, realsize, memmap_pages;
3743 size = zone_spanned_pages_in_node(nid, j, zones_size);
3744 realsize = size - zone_absent_pages_in_node(nid, j,
3748 * Adjust realsize so that it accounts for how much memory
3749 * is used by this zone for memmap. This affects the watermark
3750 * and per-cpu initialisations
3753 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
3754 if (realsize >= memmap_pages) {
3755 realsize -= memmap_pages;
3758 " %s zone: %lu pages used for memmap\n",
3759 zone_names[j], memmap_pages);
3762 " %s zone: %lu pages exceeds realsize %lu\n",
3763 zone_names[j], memmap_pages, realsize);
3765 /* Account for reserved pages */
3766 if (j == 0 && realsize > dma_reserve) {
3767 realsize -= dma_reserve;
3768 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
3769 zone_names[0], dma_reserve);
3772 if (!is_highmem_idx(j))
3773 nr_kernel_pages += realsize;
3774 nr_all_pages += realsize;
3776 zone->spanned_pages = size;
3777 zone->present_pages = realsize;
3780 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
3782 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
3784 zone->name = zone_names[j];
3785 spin_lock_init(&zone->lock);
3786 spin_lock_init(&zone->lru_lock);
3787 zone_seqlock_init(zone);
3788 zone->zone_pgdat = pgdat;
3790 zone->prev_priority = DEF_PRIORITY;
3792 zone_pcp_init(zone);
3794 INIT_LIST_HEAD(&zone->lru[l].list);
3795 zone->reclaim_stat.nr_saved_scan[l] = 0;
3797 zone->reclaim_stat.recent_rotated[0] = 0;
3798 zone->reclaim_stat.recent_rotated[1] = 0;
3799 zone->reclaim_stat.recent_scanned[0] = 0;
3800 zone->reclaim_stat.recent_scanned[1] = 0;
3801 zap_zone_vm_stats(zone);
3806 set_pageblock_order(pageblock_default_order());
3807 setup_usemap(pgdat, zone, size);
3808 ret = init_currently_empty_zone(zone, zone_start_pfn,
3809 size, MEMMAP_EARLY);
3811 memmap_init(size, nid, j, zone_start_pfn);
3812 zone_start_pfn += size;
3816 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
3818 /* Skip empty nodes */
3819 if (!pgdat->node_spanned_pages)
3822 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3823 /* ia64 gets its own node_mem_map, before this, without bootmem */
3824 if (!pgdat->node_mem_map) {
3825 unsigned long size, start, end;
3829 * The zone's endpoints aren't required to be MAX_ORDER
3830 * aligned but the node_mem_map endpoints must be in order
3831 * for the buddy allocator to function correctly.
3833 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
3834 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
3835 end = ALIGN(end, MAX_ORDER_NR_PAGES);
3836 size = (end - start) * sizeof(struct page);
3837 map = alloc_remap(pgdat->node_id, size);
3839 map = alloc_bootmem_node(pgdat, size);
3840 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
3842 #ifndef CONFIG_NEED_MULTIPLE_NODES
3844 * With no DISCONTIG, the global mem_map is just set as node 0's
3846 if (pgdat == NODE_DATA(0)) {
3847 mem_map = NODE_DATA(0)->node_mem_map;
3848 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3849 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
3850 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
3851 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3854 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3857 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
3858 unsigned long node_start_pfn, unsigned long *zholes_size)
3860 pg_data_t *pgdat = NODE_DATA(nid);
3862 pgdat->node_id = nid;
3863 pgdat->node_start_pfn = node_start_pfn;
3864 calculate_node_totalpages(pgdat, zones_size, zholes_size);
3866 alloc_node_mem_map(pgdat);
3867 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3868 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3869 nid, (unsigned long)pgdat,
3870 (unsigned long)pgdat->node_mem_map);
3873 free_area_init_core(pgdat, zones_size, zholes_size);
3876 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3878 #if MAX_NUMNODES > 1
3880 * Figure out the number of possible node ids.
3882 static void __init setup_nr_node_ids(void)
3885 unsigned int highest = 0;
3887 for_each_node_mask(node, node_possible_map)
3889 nr_node_ids = highest + 1;
3892 static inline void setup_nr_node_ids(void)
3898 * add_active_range - Register a range of PFNs backed by physical memory
3899 * @nid: The node ID the range resides on
3900 * @start_pfn: The start PFN of the available physical memory
3901 * @end_pfn: The end PFN of the available physical memory
3903 * These ranges are stored in an early_node_map[] and later used by
3904 * free_area_init_nodes() to calculate zone sizes and holes. If the
3905 * range spans a memory hole, it is up to the architecture to ensure
3906 * the memory is not freed by the bootmem allocator. If possible
3907 * the range being registered will be merged with existing ranges.
3909 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
3910 unsigned long end_pfn)
3914 mminit_dprintk(MMINIT_TRACE, "memory_register",
3915 "Entering add_active_range(%d, %#lx, %#lx) "
3916 "%d entries of %d used\n",
3917 nid, start_pfn, end_pfn,
3918 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
3920 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
3922 /* Merge with existing active regions if possible */
3923 for (i = 0; i < nr_nodemap_entries; i++) {
3924 if (early_node_map[i].nid != nid)
3927 /* Skip if an existing region covers this new one */
3928 if (start_pfn >= early_node_map[i].start_pfn &&
3929 end_pfn <= early_node_map[i].end_pfn)
3932 /* Merge forward if suitable */
3933 if (start_pfn <= early_node_map[i].end_pfn &&
3934 end_pfn > early_node_map[i].end_pfn) {
3935 early_node_map[i].end_pfn = end_pfn;
3939 /* Merge backward if suitable */
3940 if (start_pfn < early_node_map[i].end_pfn &&
3941 end_pfn >= early_node_map[i].start_pfn) {
3942 early_node_map[i].start_pfn = start_pfn;
3947 /* Check that early_node_map is large enough */
3948 if (i >= MAX_ACTIVE_REGIONS) {
3949 printk(KERN_CRIT "More than %d memory regions, truncating\n",
3950 MAX_ACTIVE_REGIONS);
3954 early_node_map[i].nid = nid;
3955 early_node_map[i].start_pfn = start_pfn;
3956 early_node_map[i].end_pfn = end_pfn;
3957 nr_nodemap_entries = i + 1;
3961 * remove_active_range - Shrink an existing registered range of PFNs
3962 * @nid: The node id the range is on that should be shrunk
3963 * @start_pfn: The new PFN of the range
3964 * @end_pfn: The new PFN of the range
3966 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
3967 * The map is kept near the end physical page range that has already been
3968 * registered. This function allows an arch to shrink an existing registered
3971 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
3972 unsigned long end_pfn)
3977 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
3978 nid, start_pfn, end_pfn);
3980 /* Find the old active region end and shrink */
3981 for_each_active_range_index_in_nid(i, nid) {
3982 if (early_node_map[i].start_pfn >= start_pfn &&
3983 early_node_map[i].end_pfn <= end_pfn) {
3985 early_node_map[i].start_pfn = 0;
3986 early_node_map[i].end_pfn = 0;
3990 if (early_node_map[i].start_pfn < start_pfn &&
3991 early_node_map[i].end_pfn > start_pfn) {
3992 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
3993 early_node_map[i].end_pfn = start_pfn;
3994 if (temp_end_pfn > end_pfn)
3995 add_active_range(nid, end_pfn, temp_end_pfn);
3998 if (early_node_map[i].start_pfn >= start_pfn &&
3999 early_node_map[i].end_pfn > end_pfn &&
4000 early_node_map[i].start_pfn < end_pfn) {
4001 early_node_map[i].start_pfn = end_pfn;
4009 /* remove the blank ones */
4010 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4011 if (early_node_map[i].nid != nid)
4013 if (early_node_map[i].end_pfn)
4015 /* we found it, get rid of it */
4016 for (j = i; j < nr_nodemap_entries - 1; j++)
4017 memcpy(&early_node_map[j], &early_node_map[j+1],
4018 sizeof(early_node_map[j]));
4019 j = nr_nodemap_entries - 1;
4020 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4021 nr_nodemap_entries--;
4026 * remove_all_active_ranges - Remove all currently registered regions
4028 * During discovery, it may be found that a table like SRAT is invalid
4029 * and an alternative discovery method must be used. This function removes
4030 * all currently registered regions.
4032 void __init remove_all_active_ranges(void)
4034 memset(early_node_map, 0, sizeof(early_node_map));
4035 nr_nodemap_entries = 0;
4038 /* Compare two active node_active_regions */
4039 static int __init cmp_node_active_region(const void *a, const void *b)
4041 struct node_active_region *arange = (struct node_active_region *)a;
4042 struct node_active_region *brange = (struct node_active_region *)b;
4044 /* Done this way to avoid overflows */
4045 if (arange->start_pfn > brange->start_pfn)
4047 if (arange->start_pfn < brange->start_pfn)
4053 /* sort the node_map by start_pfn */
4054 void __init sort_node_map(void)
4056 sort(early_node_map, (size_t)nr_nodemap_entries,
4057 sizeof(struct node_active_region),
4058 cmp_node_active_region, NULL);
4061 /* Find the lowest pfn for a node */
4062 static unsigned long __init find_min_pfn_for_node(int nid)
4065 unsigned long min_pfn = ULONG_MAX;
4067 /* Assuming a sorted map, the first range found has the starting pfn */
4068 for_each_active_range_index_in_nid(i, nid)
4069 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4071 if (min_pfn == ULONG_MAX) {
4073 "Could not find start_pfn for node %d\n", nid);
4081 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4083 * It returns the minimum PFN based on information provided via
4084 * add_active_range().
4086 unsigned long __init find_min_pfn_with_active_regions(void)
4088 return find_min_pfn_for_node(MAX_NUMNODES);
4092 * early_calculate_totalpages()
4093 * Sum pages in active regions for movable zone.
4094 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4096 static unsigned long __init early_calculate_totalpages(void)
4099 unsigned long totalpages = 0;
4101 for (i = 0; i < nr_nodemap_entries; i++) {
4102 unsigned long pages = early_node_map[i].end_pfn -
4103 early_node_map[i].start_pfn;
4104 totalpages += pages;
4106 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4112 * Find the PFN the Movable zone begins in each node. Kernel memory
4113 * is spread evenly between nodes as long as the nodes have enough
4114 * memory. When they don't, some nodes will have more kernelcore than
4117 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4120 unsigned long usable_startpfn;
4121 unsigned long kernelcore_node, kernelcore_remaining;
4122 /* save the state before borrow the nodemask */
4123 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4124 unsigned long totalpages = early_calculate_totalpages();
4125 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4128 * If movablecore was specified, calculate what size of
4129 * kernelcore that corresponds so that memory usable for
4130 * any allocation type is evenly spread. If both kernelcore
4131 * and movablecore are specified, then the value of kernelcore
4132 * will be used for required_kernelcore if it's greater than
4133 * what movablecore would have allowed.
4135 if (required_movablecore) {
4136 unsigned long corepages;
4139 * Round-up so that ZONE_MOVABLE is at least as large as what
4140 * was requested by the user
4142 required_movablecore =
4143 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4144 corepages = totalpages - required_movablecore;
4146 required_kernelcore = max(required_kernelcore, corepages);
4149 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4150 if (!required_kernelcore)
4153 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4154 find_usable_zone_for_movable();
4155 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4158 /* Spread kernelcore memory as evenly as possible throughout nodes */
4159 kernelcore_node = required_kernelcore / usable_nodes;
4160 for_each_node_state(nid, N_HIGH_MEMORY) {
4162 * Recalculate kernelcore_node if the division per node
4163 * now exceeds what is necessary to satisfy the requested
4164 * amount of memory for the kernel
4166 if (required_kernelcore < kernelcore_node)
4167 kernelcore_node = required_kernelcore / usable_nodes;
4170 * As the map is walked, we track how much memory is usable
4171 * by the kernel using kernelcore_remaining. When it is
4172 * 0, the rest of the node is usable by ZONE_MOVABLE
4174 kernelcore_remaining = kernelcore_node;
4176 /* Go through each range of PFNs within this node */
4177 for_each_active_range_index_in_nid(i, nid) {
4178 unsigned long start_pfn, end_pfn;
4179 unsigned long size_pages;
4181 start_pfn = max(early_node_map[i].start_pfn,
4182 zone_movable_pfn[nid]);
4183 end_pfn = early_node_map[i].end_pfn;
4184 if (start_pfn >= end_pfn)
4187 /* Account for what is only usable for kernelcore */
4188 if (start_pfn < usable_startpfn) {
4189 unsigned long kernel_pages;
4190 kernel_pages = min(end_pfn, usable_startpfn)
4193 kernelcore_remaining -= min(kernel_pages,
4194 kernelcore_remaining);
4195 required_kernelcore -= min(kernel_pages,
4196 required_kernelcore);
4198 /* Continue if range is now fully accounted */
4199 if (end_pfn <= usable_startpfn) {
4202 * Push zone_movable_pfn to the end so
4203 * that if we have to rebalance
4204 * kernelcore across nodes, we will
4205 * not double account here
4207 zone_movable_pfn[nid] = end_pfn;
4210 start_pfn = usable_startpfn;
4214 * The usable PFN range for ZONE_MOVABLE is from
4215 * start_pfn->end_pfn. Calculate size_pages as the
4216 * number of pages used as kernelcore
4218 size_pages = end_pfn - start_pfn;
4219 if (size_pages > kernelcore_remaining)
4220 size_pages = kernelcore_remaining;
4221 zone_movable_pfn[nid] = start_pfn + size_pages;
4224 * Some kernelcore has been met, update counts and
4225 * break if the kernelcore for this node has been
4228 required_kernelcore -= min(required_kernelcore,
4230 kernelcore_remaining -= size_pages;
4231 if (!kernelcore_remaining)
4237 * If there is still required_kernelcore, we do another pass with one
4238 * less node in the count. This will push zone_movable_pfn[nid] further
4239 * along on the nodes that still have memory until kernelcore is
4243 if (usable_nodes && required_kernelcore > usable_nodes)
4246 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4247 for (nid = 0; nid < MAX_NUMNODES; nid++)
4248 zone_movable_pfn[nid] =
4249 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4252 /* restore the node_state */
4253 node_states[N_HIGH_MEMORY] = saved_node_state;
4256 /* Any regular memory on that node ? */
4257 static void check_for_regular_memory(pg_data_t *pgdat)
4259 #ifdef CONFIG_HIGHMEM
4260 enum zone_type zone_type;
4262 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4263 struct zone *zone = &pgdat->node_zones[zone_type];
4264 if (zone->present_pages)
4265 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4271 * free_area_init_nodes - Initialise all pg_data_t and zone data
4272 * @max_zone_pfn: an array of max PFNs for each zone
4274 * This will call free_area_init_node() for each active node in the system.
4275 * Using the page ranges provided by add_active_range(), the size of each
4276 * zone in each node and their holes is calculated. If the maximum PFN
4277 * between two adjacent zones match, it is assumed that the zone is empty.
4278 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4279 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4280 * starts where the previous one ended. For example, ZONE_DMA32 starts
4281 * at arch_max_dma_pfn.
4283 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4288 /* Sort early_node_map as initialisation assumes it is sorted */
4291 /* Record where the zone boundaries are */
4292 memset(arch_zone_lowest_possible_pfn, 0,
4293 sizeof(arch_zone_lowest_possible_pfn));
4294 memset(arch_zone_highest_possible_pfn, 0,
4295 sizeof(arch_zone_highest_possible_pfn));
4296 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4297 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4298 for (i = 1; i < MAX_NR_ZONES; i++) {
4299 if (i == ZONE_MOVABLE)
4301 arch_zone_lowest_possible_pfn[i] =
4302 arch_zone_highest_possible_pfn[i-1];
4303 arch_zone_highest_possible_pfn[i] =
4304 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4306 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4307 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4309 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4310 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4311 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4313 /* Print out the zone ranges */
4314 printk("Zone PFN ranges:\n");
4315 for (i = 0; i < MAX_NR_ZONES; i++) {
4316 if (i == ZONE_MOVABLE)
4318 printk(" %-8s %0#10lx -> %0#10lx\n",
4320 arch_zone_lowest_possible_pfn[i],
4321 arch_zone_highest_possible_pfn[i]);
4324 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4325 printk("Movable zone start PFN for each node\n");
4326 for (i = 0; i < MAX_NUMNODES; i++) {
4327 if (zone_movable_pfn[i])
4328 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4331 /* Print out the early_node_map[] */
4332 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4333 for (i = 0; i < nr_nodemap_entries; i++)
4334 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4335 early_node_map[i].start_pfn,
4336 early_node_map[i].end_pfn);
4338 /* Initialise every node */
4339 mminit_verify_pageflags_layout();
4340 setup_nr_node_ids();
4341 for_each_online_node(nid) {
4342 pg_data_t *pgdat = NODE_DATA(nid);
4343 free_area_init_node(nid, NULL,
4344 find_min_pfn_for_node(nid), NULL);
4346 /* Any memory on that node */
4347 if (pgdat->node_present_pages)
4348 node_set_state(nid, N_HIGH_MEMORY);
4349 check_for_regular_memory(pgdat);
4353 static int __init cmdline_parse_core(char *p, unsigned long *core)
4355 unsigned long long coremem;
4359 coremem = memparse(p, &p);
4360 *core = coremem >> PAGE_SHIFT;
4362 /* Paranoid check that UL is enough for the coremem value */
4363 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4369 * kernelcore=size sets the amount of memory for use for allocations that
4370 * cannot be reclaimed or migrated.
4372 static int __init cmdline_parse_kernelcore(char *p)
4374 return cmdline_parse_core(p, &required_kernelcore);
4378 * movablecore=size sets the amount of memory for use for allocations that
4379 * can be reclaimed or migrated.
4381 static int __init cmdline_parse_movablecore(char *p)
4383 return cmdline_parse_core(p, &required_movablecore);
4386 early_param("kernelcore", cmdline_parse_kernelcore);
4387 early_param("movablecore", cmdline_parse_movablecore);
4389 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4392 * set_dma_reserve - set the specified number of pages reserved in the first zone
4393 * @new_dma_reserve: The number of pages to mark reserved
4395 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4396 * In the DMA zone, a significant percentage may be consumed by kernel image
4397 * and other unfreeable allocations which can skew the watermarks badly. This
4398 * function may optionally be used to account for unfreeable pages in the
4399 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4400 * smaller per-cpu batchsize.
4402 void __init set_dma_reserve(unsigned long new_dma_reserve)
4404 dma_reserve = new_dma_reserve;
4407 #ifndef CONFIG_NEED_MULTIPLE_NODES
4408 struct pglist_data __refdata contig_page_data = { .bdata = &bootmem_node_data[0] };
4409 EXPORT_SYMBOL(contig_page_data);
4412 void __init free_area_init(unsigned long *zones_size)
4414 free_area_init_node(0, zones_size,
4415 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4418 static int page_alloc_cpu_notify(struct notifier_block *self,
4419 unsigned long action, void *hcpu)
4421 int cpu = (unsigned long)hcpu;
4423 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4427 * Spill the event counters of the dead processor
4428 * into the current processors event counters.
4429 * This artificially elevates the count of the current
4432 vm_events_fold_cpu(cpu);
4435 * Zero the differential counters of the dead processor
4436 * so that the vm statistics are consistent.
4438 * This is only okay since the processor is dead and cannot
4439 * race with what we are doing.
4441 refresh_cpu_vm_stats(cpu);
4446 void __init page_alloc_init(void)
4448 hotcpu_notifier(page_alloc_cpu_notify, 0);
4452 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4453 * or min_free_kbytes changes.
4455 static void calculate_totalreserve_pages(void)
4457 struct pglist_data *pgdat;
4458 unsigned long reserve_pages = 0;
4459 enum zone_type i, j;
4461 for_each_online_pgdat(pgdat) {
4462 for (i = 0; i < MAX_NR_ZONES; i++) {
4463 struct zone *zone = pgdat->node_zones + i;
4464 unsigned long max = 0;
4466 /* Find valid and maximum lowmem_reserve in the zone */
4467 for (j = i; j < MAX_NR_ZONES; j++) {
4468 if (zone->lowmem_reserve[j] > max)
4469 max = zone->lowmem_reserve[j];
4472 /* we treat the high watermark as reserved pages. */
4473 max += high_wmark_pages(zone);
4475 if (max > zone->present_pages)
4476 max = zone->present_pages;
4477 reserve_pages += max;
4480 totalreserve_pages = reserve_pages;
4484 * setup_per_zone_lowmem_reserve - called whenever
4485 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4486 * has a correct pages reserved value, so an adequate number of
4487 * pages are left in the zone after a successful __alloc_pages().
4489 static void setup_per_zone_lowmem_reserve(void)
4491 struct pglist_data *pgdat;
4492 enum zone_type j, idx;
4494 for_each_online_pgdat(pgdat) {
4495 for (j = 0; j < MAX_NR_ZONES; j++) {
4496 struct zone *zone = pgdat->node_zones + j;
4497 unsigned long present_pages = zone->present_pages;
4499 zone->lowmem_reserve[j] = 0;
4503 struct zone *lower_zone;
4507 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4508 sysctl_lowmem_reserve_ratio[idx] = 1;
4510 lower_zone = pgdat->node_zones + idx;
4511 lower_zone->lowmem_reserve[j] = present_pages /
4512 sysctl_lowmem_reserve_ratio[idx];
4513 present_pages += lower_zone->present_pages;
4518 /* update totalreserve_pages */
4519 calculate_totalreserve_pages();
4523 * setup_per_zone_wmarks - called when min_free_kbytes changes
4524 * or when memory is hot-{added|removed}
4526 * Ensures that the watermark[min,low,high] values for each zone are set
4527 * correctly with respect to min_free_kbytes.
4529 void setup_per_zone_wmarks(void)
4531 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4532 unsigned long lowmem_pages = 0;
4534 unsigned long flags;
4536 /* Calculate total number of !ZONE_HIGHMEM pages */
4537 for_each_zone(zone) {
4538 if (!is_highmem(zone))
4539 lowmem_pages += zone->present_pages;
4542 for_each_zone(zone) {
4545 spin_lock_irqsave(&zone->lock, flags);
4546 tmp = (u64)pages_min * zone->present_pages;
4547 do_div(tmp, lowmem_pages);
4548 if (is_highmem(zone)) {
4550 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4551 * need highmem pages, so cap pages_min to a small
4554 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4555 * deltas controls asynch page reclaim, and so should
4556 * not be capped for highmem.
4560 min_pages = zone->present_pages / 1024;
4561 if (min_pages < SWAP_CLUSTER_MAX)
4562 min_pages = SWAP_CLUSTER_MAX;
4563 if (min_pages > 128)
4565 zone->watermark[WMARK_MIN] = min_pages;
4568 * If it's a lowmem zone, reserve a number of pages
4569 * proportionate to the zone's size.
4571 zone->watermark[WMARK_MIN] = tmp;
4574 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
4575 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
4576 setup_zone_migrate_reserve(zone);
4577 spin_unlock_irqrestore(&zone->lock, flags);
4580 /* update totalreserve_pages */
4581 calculate_totalreserve_pages();
4585 * The inactive anon list should be small enough that the VM never has to
4586 * do too much work, but large enough that each inactive page has a chance
4587 * to be referenced again before it is swapped out.
4589 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4590 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4591 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4592 * the anonymous pages are kept on the inactive list.
4595 * memory ratio inactive anon
4596 * -------------------------------------
4605 void calculate_zone_inactive_ratio(struct zone *zone)
4607 unsigned int gb, ratio;
4609 /* Zone size in gigabytes */
4610 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4612 ratio = int_sqrt(10 * gb);
4616 zone->inactive_ratio = ratio;
4619 static void __init setup_per_zone_inactive_ratio(void)
4624 calculate_zone_inactive_ratio(zone);
4628 * Initialise min_free_kbytes.
4630 * For small machines we want it small (128k min). For large machines
4631 * we want it large (64MB max). But it is not linear, because network
4632 * bandwidth does not increase linearly with machine size. We use
4634 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4635 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4651 static int __init init_per_zone_wmark_min(void)
4653 unsigned long lowmem_kbytes;
4655 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4657 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4658 if (min_free_kbytes < 128)
4659 min_free_kbytes = 128;
4660 if (min_free_kbytes > 65536)
4661 min_free_kbytes = 65536;
4662 setup_per_zone_wmarks();
4663 setup_per_zone_lowmem_reserve();
4664 setup_per_zone_inactive_ratio();
4667 module_init(init_per_zone_wmark_min)
4670 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4671 * that we can call two helper functions whenever min_free_kbytes
4674 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
4675 void __user *buffer, size_t *length, loff_t *ppos)
4677 proc_dointvec(table, write, buffer, length, ppos);
4679 setup_per_zone_wmarks();
4684 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
4685 void __user *buffer, size_t *length, loff_t *ppos)
4690 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
4695 zone->min_unmapped_pages = (zone->present_pages *
4696 sysctl_min_unmapped_ratio) / 100;
4700 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
4701 void __user *buffer, size_t *length, loff_t *ppos)
4706 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
4711 zone->min_slab_pages = (zone->present_pages *
4712 sysctl_min_slab_ratio) / 100;
4718 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4719 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4720 * whenever sysctl_lowmem_reserve_ratio changes.
4722 * The reserve ratio obviously has absolutely no relation with the
4723 * minimum watermarks. The lowmem reserve ratio can only make sense
4724 * if in function of the boot time zone sizes.
4726 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
4727 void __user *buffer, size_t *length, loff_t *ppos)
4729 proc_dointvec_minmax(table, write, buffer, length, ppos);
4730 setup_per_zone_lowmem_reserve();
4735 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4736 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4737 * can have before it gets flushed back to buddy allocator.
4740 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
4741 void __user *buffer, size_t *length, loff_t *ppos)
4747 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
4748 if (!write || (ret == -EINVAL))
4750 for_each_populated_zone(zone) {
4751 for_each_possible_cpu(cpu) {
4753 high = zone->present_pages / percpu_pagelist_fraction;
4754 setup_pagelist_highmark(
4755 per_cpu_ptr(zone->pageset, cpu), high);
4761 int hashdist = HASHDIST_DEFAULT;
4764 static int __init set_hashdist(char *str)
4768 hashdist = simple_strtoul(str, &str, 0);
4771 __setup("hashdist=", set_hashdist);
4775 * allocate a large system hash table from bootmem
4776 * - it is assumed that the hash table must contain an exact power-of-2
4777 * quantity of entries
4778 * - limit is the number of hash buckets, not the total allocation size
4780 void *__init alloc_large_system_hash(const char *tablename,
4781 unsigned long bucketsize,
4782 unsigned long numentries,
4785 unsigned int *_hash_shift,
4786 unsigned int *_hash_mask,
4787 unsigned long limit)
4789 unsigned long long max = limit;
4790 unsigned long log2qty, size;
4793 /* allow the kernel cmdline to have a say */
4795 /* round applicable memory size up to nearest megabyte */
4796 numentries = nr_kernel_pages;
4797 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
4798 numentries >>= 20 - PAGE_SHIFT;
4799 numentries <<= 20 - PAGE_SHIFT;
4801 /* limit to 1 bucket per 2^scale bytes of low memory */
4802 if (scale > PAGE_SHIFT)
4803 numentries >>= (scale - PAGE_SHIFT);
4805 numentries <<= (PAGE_SHIFT - scale);
4807 /* Make sure we've got at least a 0-order allocation.. */
4808 if (unlikely(flags & HASH_SMALL)) {
4809 /* Makes no sense without HASH_EARLY */
4810 WARN_ON(!(flags & HASH_EARLY));
4811 if (!(numentries >> *_hash_shift)) {
4812 numentries = 1UL << *_hash_shift;
4813 BUG_ON(!numentries);
4815 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
4816 numentries = PAGE_SIZE / bucketsize;
4818 numentries = roundup_pow_of_two(numentries);
4820 /* limit allocation size to 1/16 total memory by default */
4822 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
4823 do_div(max, bucketsize);
4826 if (numentries > max)
4829 log2qty = ilog2(numentries);
4832 size = bucketsize << log2qty;
4833 if (flags & HASH_EARLY)
4834 table = alloc_bootmem_nopanic(size);
4836 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
4839 * If bucketsize is not a power-of-two, we may free
4840 * some pages at the end of hash table which
4841 * alloc_pages_exact() automatically does
4843 if (get_order(size) < MAX_ORDER) {
4844 table = alloc_pages_exact(size, GFP_ATOMIC);
4845 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
4848 } while (!table && size > PAGE_SIZE && --log2qty);
4851 panic("Failed to allocate %s hash table\n", tablename);
4853 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
4856 ilog2(size) - PAGE_SHIFT,
4860 *_hash_shift = log2qty;
4862 *_hash_mask = (1 << log2qty) - 1;
4867 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4868 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
4871 #ifdef CONFIG_SPARSEMEM
4872 return __pfn_to_section(pfn)->pageblock_flags;
4874 return zone->pageblock_flags;
4875 #endif /* CONFIG_SPARSEMEM */
4878 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
4880 #ifdef CONFIG_SPARSEMEM
4881 pfn &= (PAGES_PER_SECTION-1);
4882 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4884 pfn = pfn - zone->zone_start_pfn;
4885 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4886 #endif /* CONFIG_SPARSEMEM */
4890 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4891 * @page: The page within the block of interest
4892 * @start_bitidx: The first bit of interest to retrieve
4893 * @end_bitidx: The last bit of interest
4894 * returns pageblock_bits flags
4896 unsigned long get_pageblock_flags_group(struct page *page,
4897 int start_bitidx, int end_bitidx)
4900 unsigned long *bitmap;
4901 unsigned long pfn, bitidx;
4902 unsigned long flags = 0;
4903 unsigned long value = 1;
4905 zone = page_zone(page);
4906 pfn = page_to_pfn(page);
4907 bitmap = get_pageblock_bitmap(zone, pfn);
4908 bitidx = pfn_to_bitidx(zone, pfn);
4910 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4911 if (test_bit(bitidx + start_bitidx, bitmap))
4918 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4919 * @page: The page within the block of interest
4920 * @start_bitidx: The first bit of interest
4921 * @end_bitidx: The last bit of interest
4922 * @flags: The flags to set
4924 void set_pageblock_flags_group(struct page *page, unsigned long flags,
4925 int start_bitidx, int end_bitidx)
4928 unsigned long *bitmap;
4929 unsigned long pfn, bitidx;
4930 unsigned long value = 1;
4932 zone = page_zone(page);
4933 pfn = page_to_pfn(page);
4934 bitmap = get_pageblock_bitmap(zone, pfn);
4935 bitidx = pfn_to_bitidx(zone, pfn);
4936 VM_BUG_ON(pfn < zone->zone_start_pfn);
4937 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
4939 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4941 __set_bit(bitidx + start_bitidx, bitmap);
4943 __clear_bit(bitidx + start_bitidx, bitmap);
4947 * This is designed as sub function...plz see page_isolation.c also.
4948 * set/clear page block's type to be ISOLATE.
4949 * page allocater never alloc memory from ISOLATE block.
4952 int set_migratetype_isolate(struct page *page)
4955 struct page *curr_page;
4956 unsigned long flags, pfn, iter;
4957 unsigned long immobile = 0;
4958 struct memory_isolate_notify arg;
4963 zone = page_zone(page);
4964 zone_idx = zone_idx(zone);
4966 spin_lock_irqsave(&zone->lock, flags);
4967 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE ||
4968 zone_idx == ZONE_MOVABLE) {
4973 pfn = page_to_pfn(page);
4974 arg.start_pfn = pfn;
4975 arg.nr_pages = pageblock_nr_pages;
4976 arg.pages_found = 0;
4979 * It may be possible to isolate a pageblock even if the
4980 * migratetype is not MIGRATE_MOVABLE. The memory isolation
4981 * notifier chain is used by balloon drivers to return the
4982 * number of pages in a range that are held by the balloon
4983 * driver to shrink memory. If all the pages are accounted for
4984 * by balloons, are free, or on the LRU, isolation can continue.
4985 * Later, for example, when memory hotplug notifier runs, these
4986 * pages reported as "can be isolated" should be isolated(freed)
4987 * by the balloon driver through the memory notifier chain.
4989 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
4990 notifier_ret = notifier_to_errno(notifier_ret);
4991 if (notifier_ret || !arg.pages_found)
4994 for (iter = pfn; iter < (pfn + pageblock_nr_pages); iter++) {
4995 if (!pfn_valid_within(pfn))
4998 curr_page = pfn_to_page(iter);
4999 if (!page_count(curr_page) || PageLRU(curr_page))
5005 if (arg.pages_found == immobile)
5010 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5011 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5014 spin_unlock_irqrestore(&zone->lock, flags);
5020 void unset_migratetype_isolate(struct page *page)
5023 unsigned long flags;
5024 zone = page_zone(page);
5025 spin_lock_irqsave(&zone->lock, flags);
5026 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5028 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5029 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5031 spin_unlock_irqrestore(&zone->lock, flags);
5034 #ifdef CONFIG_MEMORY_HOTREMOVE
5036 * All pages in the range must be isolated before calling this.
5039 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5045 unsigned long flags;
5046 /* find the first valid pfn */
5047 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5052 zone = page_zone(pfn_to_page(pfn));
5053 spin_lock_irqsave(&zone->lock, flags);
5055 while (pfn < end_pfn) {
5056 if (!pfn_valid(pfn)) {
5060 page = pfn_to_page(pfn);
5061 BUG_ON(page_count(page));
5062 BUG_ON(!PageBuddy(page));
5063 order = page_order(page);
5064 #ifdef CONFIG_DEBUG_VM
5065 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5066 pfn, 1 << order, end_pfn);
5068 list_del(&page->lru);
5069 rmv_page_order(page);
5070 zone->free_area[order].nr_free--;
5071 __mod_zone_page_state(zone, NR_FREE_PAGES,
5073 for (i = 0; i < (1 << order); i++)
5074 SetPageReserved((page+i));
5075 pfn += (1 << order);
5077 spin_unlock_irqrestore(&zone->lock, flags);
5081 #ifdef CONFIG_MEMORY_FAILURE
5082 bool is_free_buddy_page(struct page *page)
5084 struct zone *zone = page_zone(page);
5085 unsigned long pfn = page_to_pfn(page);
5086 unsigned long flags;
5089 spin_lock_irqsave(&zone->lock, flags);
5090 for (order = 0; order < MAX_ORDER; order++) {
5091 struct page *page_head = page - (pfn & ((1 << order) - 1));
5093 if (PageBuddy(page_head) && page_order(page_head) >= order)
5096 spin_unlock_irqrestore(&zone->lock, flags);
5098 return order < MAX_ORDER;