2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/compiler.h>
25 #include <linux/kernel.h>
26 #include <linux/kmemcheck.h>
27 #include <linux/module.h>
28 #include <linux/suspend.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/slab.h>
32 #include <linux/oom.h>
33 #include <linux/notifier.h>
34 #include <linux/topology.h>
35 #include <linux/sysctl.h>
36 #include <linux/cpu.h>
37 #include <linux/cpuset.h>
38 #include <linux/memory_hotplug.h>
39 #include <linux/nodemask.h>
40 #include <linux/vmalloc.h>
41 #include <linux/mempolicy.h>
42 #include <linux/stop_machine.h>
43 #include <linux/sort.h>
44 #include <linux/pfn.h>
45 #include <linux/backing-dev.h>
46 #include <linux/fault-inject.h>
47 #include <linux/page-isolation.h>
48 #include <linux/page_cgroup.h>
49 #include <linux/debugobjects.h>
50 #include <linux/kmemleak.h>
51 #include <trace/events/kmem.h>
53 #include <asm/tlbflush.h>
54 #include <asm/div64.h>
58 * Array of node states.
60 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
61 [N_POSSIBLE] = NODE_MASK_ALL,
62 [N_ONLINE] = { { [0] = 1UL } },
64 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
66 [N_HIGH_MEMORY] = { { [0] = 1UL } },
68 [N_CPU] = { { [0] = 1UL } },
71 EXPORT_SYMBOL(node_states);
73 unsigned long totalram_pages __read_mostly;
74 unsigned long totalreserve_pages __read_mostly;
75 unsigned long highest_memmap_pfn __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;
239 * Allow a burst of 60 reports, then keep quiet for that minute;
240 * or allow a steady drip of one report per second.
242 if (nr_shown == 60) {
243 if (time_before(jiffies, resume)) {
249 "BUG: Bad page state: %lu messages suppressed\n",
256 resume = jiffies + 60 * HZ;
258 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
259 current->comm, page_to_pfn(page));
261 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
262 page, (void *)page->flags, page_count(page),
263 page_mapcount(page), page->mapping, page->index);
267 /* Leave bad fields for debug, except PageBuddy could make trouble */
268 __ClearPageBuddy(page);
269 add_taint(TAINT_BAD_PAGE);
273 * Higher-order pages are called "compound pages". They are structured thusly:
275 * The first PAGE_SIZE page is called the "head page".
277 * The remaining PAGE_SIZE pages are called "tail pages".
279 * All pages have PG_compound set. All pages have their ->private pointing at
280 * the head page (even the head page has this).
282 * The first tail page's ->lru.next holds the address of the compound page's
283 * put_page() function. Its ->lru.prev holds the order of allocation.
284 * This usage means that zero-order pages may not be compound.
287 static void free_compound_page(struct page *page)
289 __free_pages_ok(page, compound_order(page));
292 void prep_compound_page(struct page *page, unsigned long order)
295 int nr_pages = 1 << order;
297 set_compound_page_dtor(page, free_compound_page);
298 set_compound_order(page, order);
300 for (i = 1; i < nr_pages; i++) {
301 struct page *p = page + i;
304 p->first_page = page;
308 static int destroy_compound_page(struct page *page, unsigned long order)
311 int nr_pages = 1 << order;
314 if (unlikely(compound_order(page) != order) ||
315 unlikely(!PageHead(page))) {
320 __ClearPageHead(page);
322 for (i = 1; i < nr_pages; i++) {
323 struct page *p = page + i;
325 if (unlikely(!PageTail(p) || (p->first_page != page))) {
335 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
340 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
341 * and __GFP_HIGHMEM from hard or soft interrupt context.
343 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
344 for (i = 0; i < (1 << order); i++)
345 clear_highpage(page + i);
348 static inline void set_page_order(struct page *page, int order)
350 set_page_private(page, order);
351 __SetPageBuddy(page);
354 static inline void rmv_page_order(struct page *page)
356 __ClearPageBuddy(page);
357 set_page_private(page, 0);
361 * Locate the struct page for both the matching buddy in our
362 * pair (buddy1) and the combined O(n+1) page they form (page).
364 * 1) Any buddy B1 will have an order O twin B2 which satisfies
365 * the following equation:
367 * For example, if the starting buddy (buddy2) is #8 its order
369 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
371 * 2) Any buddy B will have an order O+1 parent P which
372 * satisfies the following equation:
375 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
377 static inline struct page *
378 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
380 unsigned long buddy_idx = page_idx ^ (1 << order);
382 return page + (buddy_idx - page_idx);
385 static inline unsigned long
386 __find_combined_index(unsigned long page_idx, unsigned int order)
388 return (page_idx & ~(1 << order));
392 * This function checks whether a page is free && is the buddy
393 * we can do coalesce a page and its buddy if
394 * (a) the buddy is not in a hole &&
395 * (b) the buddy is in the buddy system &&
396 * (c) a page and its buddy have the same order &&
397 * (d) a page and its buddy are in the same zone.
399 * For recording whether a page is in the buddy system, we use PG_buddy.
400 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
402 * For recording page's order, we use page_private(page).
404 static inline int page_is_buddy(struct page *page, struct page *buddy,
407 if (!pfn_valid_within(page_to_pfn(buddy)))
410 if (page_zone_id(page) != page_zone_id(buddy))
413 if (PageBuddy(buddy) && page_order(buddy) == order) {
414 VM_BUG_ON(page_count(buddy) != 0);
421 * Freeing function for a buddy system allocator.
423 * The concept of a buddy system is to maintain direct-mapped table
424 * (containing bit values) for memory blocks of various "orders".
425 * The bottom level table contains the map for the smallest allocatable
426 * units of memory (here, pages), and each level above it describes
427 * pairs of units from the levels below, hence, "buddies".
428 * At a high level, all that happens here is marking the table entry
429 * at the bottom level available, and propagating the changes upward
430 * as necessary, plus some accounting needed to play nicely with other
431 * parts of the VM system.
432 * At each level, we keep a list of pages, which are heads of continuous
433 * free pages of length of (1 << order) and marked with PG_buddy. Page's
434 * order is recorded in page_private(page) field.
435 * So when we are allocating or freeing one, we can derive the state of the
436 * other. That is, if we allocate a small block, and both were
437 * free, the remainder of the region must be split into blocks.
438 * If a block is freed, and its buddy is also free, then this
439 * triggers coalescing into a block of larger size.
444 static inline void __free_one_page(struct page *page,
445 struct zone *zone, unsigned int order,
448 unsigned long page_idx;
450 if (unlikely(PageCompound(page)))
451 if (unlikely(destroy_compound_page(page, order)))
454 VM_BUG_ON(migratetype == -1);
456 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
458 VM_BUG_ON(page_idx & ((1 << order) - 1));
459 VM_BUG_ON(bad_range(zone, page));
461 while (order < MAX_ORDER-1) {
462 unsigned long combined_idx;
465 buddy = __page_find_buddy(page, page_idx, order);
466 if (!page_is_buddy(page, buddy, order))
469 /* Our buddy is free, merge with it and move up one order. */
470 list_del(&buddy->lru);
471 zone->free_area[order].nr_free--;
472 rmv_page_order(buddy);
473 combined_idx = __find_combined_index(page_idx, order);
474 page = page + (combined_idx - page_idx);
475 page_idx = combined_idx;
478 set_page_order(page, order);
480 &zone->free_area[order].free_list[migratetype]);
481 zone->free_area[order].nr_free++;
484 #ifdef CONFIG_HAVE_MLOCKED_PAGE_BIT
486 * free_page_mlock() -- clean up attempts to free and mlocked() page.
487 * Page should not be on lru, so no need to fix that up.
488 * free_pages_check() will verify...
490 static inline void free_page_mlock(struct page *page)
492 __dec_zone_page_state(page, NR_MLOCK);
493 __count_vm_event(UNEVICTABLE_MLOCKFREED);
496 static void free_page_mlock(struct page *page) { }
499 static inline int free_pages_check(struct page *page)
501 if (unlikely(page_mapcount(page) |
502 (page->mapping != NULL) |
503 (atomic_read(&page->_count) != 0) |
504 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) {
508 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
509 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
514 * Frees a list of pages.
515 * Assumes all pages on list are in same zone, and of same order.
516 * count is the number of pages to free.
518 * If the zone was previously in an "all pages pinned" state then look to
519 * see if this freeing clears that state.
521 * And clear the zone's pages_scanned counter, to hold off the "all pages are
522 * pinned" detection logic.
524 static void free_pages_bulk(struct zone *zone, int count,
525 struct list_head *list, int order)
527 spin_lock(&zone->lock);
528 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
529 zone->pages_scanned = 0;
531 __mod_zone_page_state(zone, NR_FREE_PAGES, count << order);
535 VM_BUG_ON(list_empty(list));
536 page = list_entry(list->prev, struct page, lru);
537 /* have to delete it as __free_one_page list manipulates */
538 list_del(&page->lru);
539 trace_mm_page_pcpu_drain(page, order, page_private(page));
540 __free_one_page(page, zone, order, page_private(page));
542 spin_unlock(&zone->lock);
545 static void free_one_page(struct zone *zone, struct page *page, int order,
548 spin_lock(&zone->lock);
549 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
550 zone->pages_scanned = 0;
552 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
553 __free_one_page(page, zone, order, migratetype);
554 spin_unlock(&zone->lock);
557 static void __free_pages_ok(struct page *page, unsigned int order)
562 int wasMlocked = __TestClearPageMlocked(page);
564 kmemcheck_free_shadow(page, order);
566 for (i = 0 ; i < (1 << order) ; ++i)
567 bad += free_pages_check(page + i);
571 if (!PageHighMem(page)) {
572 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
573 debug_check_no_obj_freed(page_address(page),
576 arch_free_page(page, order);
577 kernel_map_pages(page, 1 << order, 0);
579 local_irq_save(flags);
580 if (unlikely(wasMlocked))
581 free_page_mlock(page);
582 __count_vm_events(PGFREE, 1 << order);
583 free_one_page(page_zone(page), page, order,
584 get_pageblock_migratetype(page));
585 local_irq_restore(flags);
589 * permit the bootmem allocator to evade page validation on high-order frees
591 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
594 __ClearPageReserved(page);
595 set_page_count(page, 0);
596 set_page_refcounted(page);
602 for (loop = 0; loop < BITS_PER_LONG; loop++) {
603 struct page *p = &page[loop];
605 if (loop + 1 < BITS_PER_LONG)
607 __ClearPageReserved(p);
608 set_page_count(p, 0);
611 set_page_refcounted(page);
612 __free_pages(page, order);
618 * The order of subdivision here is critical for the IO subsystem.
619 * Please do not alter this order without good reasons and regression
620 * testing. Specifically, as large blocks of memory are subdivided,
621 * the order in which smaller blocks are delivered depends on the order
622 * they're subdivided in this function. This is the primary factor
623 * influencing the order in which pages are delivered to the IO
624 * subsystem according to empirical testing, and this is also justified
625 * by considering the behavior of a buddy system containing a single
626 * large block of memory acted on by a series of small allocations.
627 * This behavior is a critical factor in sglist merging's success.
631 static inline void expand(struct zone *zone, struct page *page,
632 int low, int high, struct free_area *area,
635 unsigned long size = 1 << high;
641 VM_BUG_ON(bad_range(zone, &page[size]));
642 list_add(&page[size].lru, &area->free_list[migratetype]);
644 set_page_order(&page[size], high);
649 * This page is about to be returned from the page allocator
651 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
653 if (unlikely(page_mapcount(page) |
654 (page->mapping != NULL) |
655 (atomic_read(&page->_count) != 0) |
656 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) {
661 set_page_private(page, 0);
662 set_page_refcounted(page);
664 arch_alloc_page(page, order);
665 kernel_map_pages(page, 1 << order, 1);
667 if (gfp_flags & __GFP_ZERO)
668 prep_zero_page(page, order, gfp_flags);
670 if (order && (gfp_flags & __GFP_COMP))
671 prep_compound_page(page, order);
677 * Go through the free lists for the given migratetype and remove
678 * the smallest available page from the freelists
681 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
684 unsigned int current_order;
685 struct free_area * area;
688 /* Find a page of the appropriate size in the preferred list */
689 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
690 area = &(zone->free_area[current_order]);
691 if (list_empty(&area->free_list[migratetype]))
694 page = list_entry(area->free_list[migratetype].next,
696 list_del(&page->lru);
697 rmv_page_order(page);
699 expand(zone, page, order, current_order, area, migratetype);
708 * This array describes the order lists are fallen back to when
709 * the free lists for the desirable migrate type are depleted
711 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
712 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
713 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
714 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
715 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
719 * Move the free pages in a range to the free lists of the requested type.
720 * Note that start_page and end_pages are not aligned on a pageblock
721 * boundary. If alignment is required, use move_freepages_block()
723 static int move_freepages(struct zone *zone,
724 struct page *start_page, struct page *end_page,
731 #ifndef CONFIG_HOLES_IN_ZONE
733 * page_zone is not safe to call in this context when
734 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
735 * anyway as we check zone boundaries in move_freepages_block().
736 * Remove at a later date when no bug reports exist related to
737 * grouping pages by mobility
739 BUG_ON(page_zone(start_page) != page_zone(end_page));
742 for (page = start_page; page <= end_page;) {
743 /* Make sure we are not inadvertently changing nodes */
744 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
746 if (!pfn_valid_within(page_to_pfn(page))) {
751 if (!PageBuddy(page)) {
756 order = page_order(page);
757 list_del(&page->lru);
759 &zone->free_area[order].free_list[migratetype]);
761 pages_moved += 1 << order;
767 static int move_freepages_block(struct zone *zone, struct page *page,
770 unsigned long start_pfn, end_pfn;
771 struct page *start_page, *end_page;
773 start_pfn = page_to_pfn(page);
774 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
775 start_page = pfn_to_page(start_pfn);
776 end_page = start_page + pageblock_nr_pages - 1;
777 end_pfn = start_pfn + pageblock_nr_pages - 1;
779 /* Do not cross zone boundaries */
780 if (start_pfn < zone->zone_start_pfn)
782 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
785 return move_freepages(zone, start_page, end_page, migratetype);
788 static void change_pageblock_range(struct page *pageblock_page,
789 int start_order, int migratetype)
791 int nr_pageblocks = 1 << (start_order - pageblock_order);
793 while (nr_pageblocks--) {
794 set_pageblock_migratetype(pageblock_page, migratetype);
795 pageblock_page += pageblock_nr_pages;
799 /* Remove an element from the buddy allocator from the fallback list */
800 static inline struct page *
801 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
803 struct free_area * area;
808 /* Find the largest possible block of pages in the other list */
809 for (current_order = MAX_ORDER-1; current_order >= order;
811 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
812 migratetype = fallbacks[start_migratetype][i];
814 /* MIGRATE_RESERVE handled later if necessary */
815 if (migratetype == MIGRATE_RESERVE)
818 area = &(zone->free_area[current_order]);
819 if (list_empty(&area->free_list[migratetype]))
822 page = list_entry(area->free_list[migratetype].next,
827 * If breaking a large block of pages, move all free
828 * pages to the preferred allocation list. If falling
829 * back for a reclaimable kernel allocation, be more
830 * agressive about taking ownership of free pages
832 if (unlikely(current_order >= (pageblock_order >> 1)) ||
833 start_migratetype == MIGRATE_RECLAIMABLE ||
834 page_group_by_mobility_disabled) {
836 pages = move_freepages_block(zone, page,
839 /* Claim the whole block if over half of it is free */
840 if (pages >= (1 << (pageblock_order-1)) ||
841 page_group_by_mobility_disabled)
842 set_pageblock_migratetype(page,
845 migratetype = start_migratetype;
848 /* Remove the page from the freelists */
849 list_del(&page->lru);
850 rmv_page_order(page);
852 /* Take ownership for orders >= pageblock_order */
853 if (current_order >= pageblock_order)
854 change_pageblock_range(page, current_order,
857 expand(zone, page, order, current_order, area, migratetype);
859 trace_mm_page_alloc_extfrag(page, order, current_order,
860 start_migratetype, migratetype);
870 * Do the hard work of removing an element from the buddy allocator.
871 * Call me with the zone->lock already held.
873 static struct page *__rmqueue(struct zone *zone, unsigned int order,
879 page = __rmqueue_smallest(zone, order, migratetype);
881 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
882 page = __rmqueue_fallback(zone, order, migratetype);
885 * Use MIGRATE_RESERVE rather than fail an allocation. goto
886 * is used because __rmqueue_smallest is an inline function
887 * and we want just one call site
890 migratetype = MIGRATE_RESERVE;
895 trace_mm_page_alloc_zone_locked(page, order, migratetype);
900 * Obtain a specified number of elements from the buddy allocator, all under
901 * a single hold of the lock, for efficiency. Add them to the supplied list.
902 * Returns the number of new pages which were placed at *list.
904 static int rmqueue_bulk(struct zone *zone, unsigned int order,
905 unsigned long count, struct list_head *list,
906 int migratetype, int cold)
910 spin_lock(&zone->lock);
911 for (i = 0; i < count; ++i) {
912 struct page *page = __rmqueue(zone, order, migratetype);
913 if (unlikely(page == NULL))
917 * Split buddy pages returned by expand() are received here
918 * in physical page order. The page is added to the callers and
919 * list and the list head then moves forward. From the callers
920 * perspective, the linked list is ordered by page number in
921 * some conditions. This is useful for IO devices that can
922 * merge IO requests if the physical pages are ordered
925 if (likely(cold == 0))
926 list_add(&page->lru, list);
928 list_add_tail(&page->lru, list);
929 set_page_private(page, migratetype);
932 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
933 spin_unlock(&zone->lock);
939 * Called from the vmstat counter updater to drain pagesets of this
940 * currently executing processor on remote nodes after they have
943 * Note that this function must be called with the thread pinned to
944 * a single processor.
946 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
951 local_irq_save(flags);
952 if (pcp->count >= pcp->batch)
953 to_drain = pcp->batch;
955 to_drain = pcp->count;
956 free_pages_bulk(zone, to_drain, &pcp->list, 0);
957 pcp->count -= to_drain;
958 local_irq_restore(flags);
963 * Drain pages of the indicated processor.
965 * The processor must either be the current processor and the
966 * thread pinned to the current processor or a processor that
969 static void drain_pages(unsigned int cpu)
974 for_each_populated_zone(zone) {
975 struct per_cpu_pageset *pset;
976 struct per_cpu_pages *pcp;
978 pset = zone_pcp(zone, cpu);
981 local_irq_save(flags);
982 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
984 local_irq_restore(flags);
989 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
991 void drain_local_pages(void *arg)
993 drain_pages(smp_processor_id());
997 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
999 void drain_all_pages(void)
1001 on_each_cpu(drain_local_pages, NULL, 1);
1004 #ifdef CONFIG_HIBERNATION
1006 void mark_free_pages(struct zone *zone)
1008 unsigned long pfn, max_zone_pfn;
1009 unsigned long flags;
1011 struct list_head *curr;
1013 if (!zone->spanned_pages)
1016 spin_lock_irqsave(&zone->lock, flags);
1018 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1019 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1020 if (pfn_valid(pfn)) {
1021 struct page *page = pfn_to_page(pfn);
1023 if (!swsusp_page_is_forbidden(page))
1024 swsusp_unset_page_free(page);
1027 for_each_migratetype_order(order, t) {
1028 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1031 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1032 for (i = 0; i < (1UL << order); i++)
1033 swsusp_set_page_free(pfn_to_page(pfn + i));
1036 spin_unlock_irqrestore(&zone->lock, flags);
1038 #endif /* CONFIG_PM */
1041 * Free a 0-order page
1043 static void free_hot_cold_page(struct page *page, int cold)
1045 struct zone *zone = page_zone(page);
1046 struct per_cpu_pages *pcp;
1047 unsigned long flags;
1048 int wasMlocked = __TestClearPageMlocked(page);
1050 kmemcheck_free_shadow(page, 0);
1053 page->mapping = NULL;
1054 if (free_pages_check(page))
1057 if (!PageHighMem(page)) {
1058 debug_check_no_locks_freed(page_address(page), PAGE_SIZE);
1059 debug_check_no_obj_freed(page_address(page), PAGE_SIZE);
1061 arch_free_page(page, 0);
1062 kernel_map_pages(page, 1, 0);
1064 pcp = &zone_pcp(zone, get_cpu())->pcp;
1065 set_page_private(page, get_pageblock_migratetype(page));
1066 local_irq_save(flags);
1067 if (unlikely(wasMlocked))
1068 free_page_mlock(page);
1069 __count_vm_event(PGFREE);
1072 list_add_tail(&page->lru, &pcp->list);
1074 list_add(&page->lru, &pcp->list);
1076 if (pcp->count >= pcp->high) {
1077 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
1078 pcp->count -= pcp->batch;
1080 local_irq_restore(flags);
1084 void free_hot_page(struct page *page)
1086 trace_mm_page_free_direct(page, 0);
1087 free_hot_cold_page(page, 0);
1091 * split_page takes a non-compound higher-order page, and splits it into
1092 * n (1<<order) sub-pages: page[0..n]
1093 * Each sub-page must be freed individually.
1095 * Note: this is probably too low level an operation for use in drivers.
1096 * Please consult with lkml before using this in your driver.
1098 void split_page(struct page *page, unsigned int order)
1102 VM_BUG_ON(PageCompound(page));
1103 VM_BUG_ON(!page_count(page));
1105 #ifdef CONFIG_KMEMCHECK
1107 * Split shadow pages too, because free(page[0]) would
1108 * otherwise free the whole shadow.
1110 if (kmemcheck_page_is_tracked(page))
1111 split_page(virt_to_page(page[0].shadow), order);
1114 for (i = 1; i < (1 << order); i++)
1115 set_page_refcounted(page + i);
1119 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1120 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1124 struct page *buffered_rmqueue(struct zone *preferred_zone,
1125 struct zone *zone, int order, gfp_t gfp_flags,
1128 unsigned long flags;
1130 int cold = !!(gfp_flags & __GFP_COLD);
1135 if (likely(order == 0)) {
1136 struct per_cpu_pages *pcp;
1138 pcp = &zone_pcp(zone, cpu)->pcp;
1139 local_irq_save(flags);
1141 pcp->count = rmqueue_bulk(zone, 0,
1142 pcp->batch, &pcp->list,
1144 if (unlikely(!pcp->count))
1148 /* Find a page of the appropriate migrate type */
1150 list_for_each_entry_reverse(page, &pcp->list, lru)
1151 if (page_private(page) == migratetype)
1154 list_for_each_entry(page, &pcp->list, lru)
1155 if (page_private(page) == migratetype)
1159 /* Allocate more to the pcp list if necessary */
1160 if (unlikely(&page->lru == &pcp->list)) {
1161 int get_one_page = 0;
1163 pcp->count += rmqueue_bulk(zone, 0,
1164 pcp->batch, &pcp->list,
1166 list_for_each_entry(page, &pcp->list, lru) {
1167 if (get_pageblock_migratetype(page) !=
1177 list_del(&page->lru);
1180 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1182 * __GFP_NOFAIL is not to be used in new code.
1184 * All __GFP_NOFAIL callers should be fixed so that they
1185 * properly detect and handle allocation failures.
1187 * We most definitely don't want callers attempting to
1188 * allocate greater than order-1 page units with
1191 WARN_ON_ONCE(order > 1);
1193 spin_lock_irqsave(&zone->lock, flags);
1194 page = __rmqueue(zone, order, migratetype);
1195 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1196 spin_unlock(&zone->lock);
1201 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1202 zone_statistics(preferred_zone, zone);
1203 local_irq_restore(flags);
1206 VM_BUG_ON(bad_range(zone, page));
1207 if (prep_new_page(page, order, gfp_flags))
1212 local_irq_restore(flags);
1217 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1218 #define ALLOC_WMARK_MIN WMARK_MIN
1219 #define ALLOC_WMARK_LOW WMARK_LOW
1220 #define ALLOC_WMARK_HIGH WMARK_HIGH
1221 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1223 /* Mask to get the watermark bits */
1224 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1226 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1227 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1228 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1230 #ifdef CONFIG_FAIL_PAGE_ALLOC
1232 static struct fail_page_alloc_attr {
1233 struct fault_attr attr;
1235 u32 ignore_gfp_highmem;
1236 u32 ignore_gfp_wait;
1239 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1241 struct dentry *ignore_gfp_highmem_file;
1242 struct dentry *ignore_gfp_wait_file;
1243 struct dentry *min_order_file;
1245 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1247 } fail_page_alloc = {
1248 .attr = FAULT_ATTR_INITIALIZER,
1249 .ignore_gfp_wait = 1,
1250 .ignore_gfp_highmem = 1,
1254 static int __init setup_fail_page_alloc(char *str)
1256 return setup_fault_attr(&fail_page_alloc.attr, str);
1258 __setup("fail_page_alloc=", setup_fail_page_alloc);
1260 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1262 if (order < fail_page_alloc.min_order)
1264 if (gfp_mask & __GFP_NOFAIL)
1266 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1268 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1271 return should_fail(&fail_page_alloc.attr, 1 << order);
1274 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1276 static int __init fail_page_alloc_debugfs(void)
1278 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1282 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1286 dir = fail_page_alloc.attr.dentries.dir;
1288 fail_page_alloc.ignore_gfp_wait_file =
1289 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1290 &fail_page_alloc.ignore_gfp_wait);
1292 fail_page_alloc.ignore_gfp_highmem_file =
1293 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1294 &fail_page_alloc.ignore_gfp_highmem);
1295 fail_page_alloc.min_order_file =
1296 debugfs_create_u32("min-order", mode, dir,
1297 &fail_page_alloc.min_order);
1299 if (!fail_page_alloc.ignore_gfp_wait_file ||
1300 !fail_page_alloc.ignore_gfp_highmem_file ||
1301 !fail_page_alloc.min_order_file) {
1303 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1304 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1305 debugfs_remove(fail_page_alloc.min_order_file);
1306 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1312 late_initcall(fail_page_alloc_debugfs);
1314 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1316 #else /* CONFIG_FAIL_PAGE_ALLOC */
1318 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1323 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1326 * Return 1 if free pages are above 'mark'. This takes into account the order
1327 * of the allocation.
1329 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1330 int classzone_idx, int alloc_flags)
1332 /* free_pages my go negative - that's OK */
1334 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
1337 if (alloc_flags & ALLOC_HIGH)
1339 if (alloc_flags & ALLOC_HARDER)
1342 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1344 for (o = 0; o < order; o++) {
1345 /* At the next order, this order's pages become unavailable */
1346 free_pages -= z->free_area[o].nr_free << o;
1348 /* Require fewer higher order pages to be free */
1351 if (free_pages <= min)
1359 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1360 * skip over zones that are not allowed by the cpuset, or that have
1361 * been recently (in last second) found to be nearly full. See further
1362 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1363 * that have to skip over a lot of full or unallowed zones.
1365 * If the zonelist cache is present in the passed in zonelist, then
1366 * returns a pointer to the allowed node mask (either the current
1367 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1369 * If the zonelist cache is not available for this zonelist, does
1370 * nothing and returns NULL.
1372 * If the fullzones BITMAP in the zonelist cache is stale (more than
1373 * a second since last zap'd) then we zap it out (clear its bits.)
1375 * We hold off even calling zlc_setup, until after we've checked the
1376 * first zone in the zonelist, on the theory that most allocations will
1377 * be satisfied from that first zone, so best to examine that zone as
1378 * quickly as we can.
1380 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1382 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1383 nodemask_t *allowednodes; /* zonelist_cache approximation */
1385 zlc = zonelist->zlcache_ptr;
1389 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1390 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1391 zlc->last_full_zap = jiffies;
1394 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1395 &cpuset_current_mems_allowed :
1396 &node_states[N_HIGH_MEMORY];
1397 return allowednodes;
1401 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1402 * if it is worth looking at further for free memory:
1403 * 1) Check that the zone isn't thought to be full (doesn't have its
1404 * bit set in the zonelist_cache fullzones BITMAP).
1405 * 2) Check that the zones node (obtained from the zonelist_cache
1406 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1407 * Return true (non-zero) if zone is worth looking at further, or
1408 * else return false (zero) if it is not.
1410 * This check -ignores- the distinction between various watermarks,
1411 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1412 * found to be full for any variation of these watermarks, it will
1413 * be considered full for up to one second by all requests, unless
1414 * we are so low on memory on all allowed nodes that we are forced
1415 * into the second scan of the zonelist.
1417 * In the second scan we ignore this zonelist cache and exactly
1418 * apply the watermarks to all zones, even it is slower to do so.
1419 * We are low on memory in the second scan, and should leave no stone
1420 * unturned looking for a free page.
1422 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1423 nodemask_t *allowednodes)
1425 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1426 int i; /* index of *z in zonelist zones */
1427 int n; /* node that zone *z is on */
1429 zlc = zonelist->zlcache_ptr;
1433 i = z - zonelist->_zonerefs;
1436 /* This zone is worth trying if it is allowed but not full */
1437 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1441 * Given 'z' scanning a zonelist, set the corresponding bit in
1442 * zlc->fullzones, so that subsequent attempts to allocate a page
1443 * from that zone don't waste time re-examining it.
1445 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1447 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1448 int i; /* index of *z in zonelist zones */
1450 zlc = zonelist->zlcache_ptr;
1454 i = z - zonelist->_zonerefs;
1456 set_bit(i, zlc->fullzones);
1459 #else /* CONFIG_NUMA */
1461 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1466 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1467 nodemask_t *allowednodes)
1472 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1475 #endif /* CONFIG_NUMA */
1478 * get_page_from_freelist goes through the zonelist trying to allocate
1481 static struct page *
1482 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1483 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1484 struct zone *preferred_zone, int migratetype)
1487 struct page *page = NULL;
1490 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1491 int zlc_active = 0; /* set if using zonelist_cache */
1492 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1494 classzone_idx = zone_idx(preferred_zone);
1497 * Scan zonelist, looking for a zone with enough free.
1498 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1500 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1501 high_zoneidx, nodemask) {
1502 if (NUMA_BUILD && zlc_active &&
1503 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1505 if ((alloc_flags & ALLOC_CPUSET) &&
1506 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1509 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1510 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1514 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1515 if (zone_watermark_ok(zone, order, mark,
1516 classzone_idx, alloc_flags))
1519 if (zone_reclaim_mode == 0)
1520 goto this_zone_full;
1522 ret = zone_reclaim(zone, gfp_mask, order);
1524 case ZONE_RECLAIM_NOSCAN:
1527 case ZONE_RECLAIM_FULL:
1528 /* scanned but unreclaimable */
1529 goto this_zone_full;
1531 /* did we reclaim enough */
1532 if (!zone_watermark_ok(zone, order, mark,
1533 classzone_idx, alloc_flags))
1534 goto this_zone_full;
1539 page = buffered_rmqueue(preferred_zone, zone, order,
1540 gfp_mask, migratetype);
1545 zlc_mark_zone_full(zonelist, z);
1547 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1549 * we do zlc_setup after the first zone is tried but only
1550 * if there are multiple nodes make it worthwhile
1552 allowednodes = zlc_setup(zonelist, alloc_flags);
1558 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1559 /* Disable zlc cache for second zonelist scan */
1567 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1568 unsigned long pages_reclaimed)
1570 /* Do not loop if specifically requested */
1571 if (gfp_mask & __GFP_NORETRY)
1575 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1576 * means __GFP_NOFAIL, but that may not be true in other
1579 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1583 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1584 * specified, then we retry until we no longer reclaim any pages
1585 * (above), or we've reclaimed an order of pages at least as
1586 * large as the allocation's order. In both cases, if the
1587 * allocation still fails, we stop retrying.
1589 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1593 * Don't let big-order allocations loop unless the caller
1594 * explicitly requests that.
1596 if (gfp_mask & __GFP_NOFAIL)
1602 static inline struct page *
1603 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1604 struct zonelist *zonelist, enum zone_type high_zoneidx,
1605 nodemask_t *nodemask, struct zone *preferred_zone,
1610 /* Acquire the OOM killer lock for the zones in zonelist */
1611 if (!try_set_zone_oom(zonelist, gfp_mask)) {
1612 schedule_timeout_uninterruptible(1);
1617 * Go through the zonelist yet one more time, keep very high watermark
1618 * here, this is only to catch a parallel oom killing, we must fail if
1619 * we're still under heavy pressure.
1621 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1622 order, zonelist, high_zoneidx,
1623 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1624 preferred_zone, migratetype);
1628 /* The OOM killer will not help higher order allocs */
1629 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_NOFAIL))
1632 /* Exhausted what can be done so it's blamo time */
1633 out_of_memory(zonelist, gfp_mask, order);
1636 clear_zonelist_oom(zonelist, gfp_mask);
1640 /* The really slow allocator path where we enter direct reclaim */
1641 static inline struct page *
1642 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1643 struct zonelist *zonelist, enum zone_type high_zoneidx,
1644 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1645 int migratetype, unsigned long *did_some_progress)
1647 struct page *page = NULL;
1648 struct reclaim_state reclaim_state;
1649 struct task_struct *p = current;
1653 /* We now go into synchronous reclaim */
1654 cpuset_memory_pressure_bump();
1655 p->flags |= PF_MEMALLOC;
1656 lockdep_set_current_reclaim_state(gfp_mask);
1657 reclaim_state.reclaimed_slab = 0;
1658 p->reclaim_state = &reclaim_state;
1660 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1662 p->reclaim_state = NULL;
1663 lockdep_clear_current_reclaim_state();
1664 p->flags &= ~PF_MEMALLOC;
1671 if (likely(*did_some_progress))
1672 page = get_page_from_freelist(gfp_mask, nodemask, order,
1673 zonelist, high_zoneidx,
1674 alloc_flags, preferred_zone,
1680 * This is called in the allocator slow-path if the allocation request is of
1681 * sufficient urgency to ignore watermarks and take other desperate measures
1683 static inline struct page *
1684 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1685 struct zonelist *zonelist, enum zone_type high_zoneidx,
1686 nodemask_t *nodemask, struct zone *preferred_zone,
1692 page = get_page_from_freelist(gfp_mask, nodemask, order,
1693 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
1694 preferred_zone, migratetype);
1696 if (!page && gfp_mask & __GFP_NOFAIL)
1697 congestion_wait(BLK_RW_ASYNC, HZ/50);
1698 } while (!page && (gfp_mask & __GFP_NOFAIL));
1704 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
1705 enum zone_type high_zoneidx)
1710 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1711 wakeup_kswapd(zone, order);
1715 gfp_to_alloc_flags(gfp_t gfp_mask)
1717 struct task_struct *p = current;
1718 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
1719 const gfp_t wait = gfp_mask & __GFP_WAIT;
1721 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1722 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH);
1725 * The caller may dip into page reserves a bit more if the caller
1726 * cannot run direct reclaim, or if the caller has realtime scheduling
1727 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1728 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1730 alloc_flags |= (gfp_mask & __GFP_HIGH);
1733 alloc_flags |= ALLOC_HARDER;
1735 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1736 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1738 alloc_flags &= ~ALLOC_CPUSET;
1739 } else if (unlikely(rt_task(p)))
1740 alloc_flags |= ALLOC_HARDER;
1742 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
1743 if (!in_interrupt() &&
1744 ((p->flags & PF_MEMALLOC) ||
1745 unlikely(test_thread_flag(TIF_MEMDIE))))
1746 alloc_flags |= ALLOC_NO_WATERMARKS;
1752 static inline struct page *
1753 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
1754 struct zonelist *zonelist, enum zone_type high_zoneidx,
1755 nodemask_t *nodemask, struct zone *preferred_zone,
1758 const gfp_t wait = gfp_mask & __GFP_WAIT;
1759 struct page *page = NULL;
1761 unsigned long pages_reclaimed = 0;
1762 unsigned long did_some_progress;
1763 struct task_struct *p = current;
1766 * In the slowpath, we sanity check order to avoid ever trying to
1767 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1768 * be using allocators in order of preference for an area that is
1771 if (order >= MAX_ORDER) {
1772 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
1777 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1778 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1779 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1780 * using a larger set of nodes after it has established that the
1781 * allowed per node queues are empty and that nodes are
1784 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1787 wake_all_kswapd(order, zonelist, high_zoneidx);
1791 * OK, we're below the kswapd watermark and have kicked background
1792 * reclaim. Now things get more complex, so set up alloc_flags according
1793 * to how we want to proceed.
1795 alloc_flags = gfp_to_alloc_flags(gfp_mask);
1797 /* This is the last chance, in general, before the goto nopage. */
1798 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
1799 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
1800 preferred_zone, migratetype);
1805 /* Allocate without watermarks if the context allows */
1806 if (alloc_flags & ALLOC_NO_WATERMARKS) {
1807 page = __alloc_pages_high_priority(gfp_mask, order,
1808 zonelist, high_zoneidx, nodemask,
1809 preferred_zone, migratetype);
1814 /* Atomic allocations - we can't balance anything */
1818 /* Avoid recursion of direct reclaim */
1819 if (p->flags & PF_MEMALLOC)
1822 /* Avoid allocations with no watermarks from looping endlessly */
1823 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
1826 /* Try direct reclaim and then allocating */
1827 page = __alloc_pages_direct_reclaim(gfp_mask, order,
1828 zonelist, high_zoneidx,
1830 alloc_flags, preferred_zone,
1831 migratetype, &did_some_progress);
1836 * If we failed to make any progress reclaiming, then we are
1837 * running out of options and have to consider going OOM
1839 if (!did_some_progress) {
1840 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1841 if (oom_killer_disabled)
1843 page = __alloc_pages_may_oom(gfp_mask, order,
1844 zonelist, high_zoneidx,
1845 nodemask, preferred_zone,
1851 * The OOM killer does not trigger for high-order
1852 * ~__GFP_NOFAIL allocations so if no progress is being
1853 * made, there are no other options and retrying is
1856 if (order > PAGE_ALLOC_COSTLY_ORDER &&
1857 !(gfp_mask & __GFP_NOFAIL))
1864 /* Check if we should retry the allocation */
1865 pages_reclaimed += did_some_progress;
1866 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
1867 /* Wait for some write requests to complete then retry */
1868 congestion_wait(BLK_RW_ASYNC, HZ/50);
1873 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1874 printk(KERN_WARNING "%s: page allocation failure."
1875 " order:%d, mode:0x%x\n",
1876 p->comm, order, gfp_mask);
1882 if (kmemcheck_enabled)
1883 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
1889 * This is the 'heart' of the zoned buddy allocator.
1892 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
1893 struct zonelist *zonelist, nodemask_t *nodemask)
1895 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1896 struct zone *preferred_zone;
1898 int migratetype = allocflags_to_migratetype(gfp_mask);
1900 gfp_mask &= gfp_allowed_mask;
1902 lockdep_trace_alloc(gfp_mask);
1904 might_sleep_if(gfp_mask & __GFP_WAIT);
1906 if (should_fail_alloc_page(gfp_mask, order))
1910 * Check the zones suitable for the gfp_mask contain at least one
1911 * valid zone. It's possible to have an empty zonelist as a result
1912 * of GFP_THISNODE and a memoryless node
1914 if (unlikely(!zonelist->_zonerefs->zone))
1917 /* The preferred zone is used for statistics later */
1918 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone);
1919 if (!preferred_zone)
1922 /* First allocation attempt */
1923 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
1924 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
1925 preferred_zone, migratetype);
1926 if (unlikely(!page))
1927 page = __alloc_pages_slowpath(gfp_mask, order,
1928 zonelist, high_zoneidx, nodemask,
1929 preferred_zone, migratetype);
1931 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
1934 EXPORT_SYMBOL(__alloc_pages_nodemask);
1937 * Common helper functions.
1939 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1944 * __get_free_pages() returns a 32-bit address, which cannot represent
1947 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1949 page = alloc_pages(gfp_mask, order);
1952 return (unsigned long) page_address(page);
1954 EXPORT_SYMBOL(__get_free_pages);
1956 unsigned long get_zeroed_page(gfp_t gfp_mask)
1958 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
1960 EXPORT_SYMBOL(get_zeroed_page);
1962 void __pagevec_free(struct pagevec *pvec)
1964 int i = pagevec_count(pvec);
1967 trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
1968 free_hot_cold_page(pvec->pages[i], pvec->cold);
1972 void __free_pages(struct page *page, unsigned int order)
1974 if (put_page_testzero(page)) {
1975 trace_mm_page_free_direct(page, order);
1977 free_hot_page(page);
1979 __free_pages_ok(page, order);
1983 EXPORT_SYMBOL(__free_pages);
1985 void free_pages(unsigned long addr, unsigned int order)
1988 VM_BUG_ON(!virt_addr_valid((void *)addr));
1989 __free_pages(virt_to_page((void *)addr), order);
1993 EXPORT_SYMBOL(free_pages);
1996 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
1997 * @size: the number of bytes to allocate
1998 * @gfp_mask: GFP flags for the allocation
2000 * This function is similar to alloc_pages(), except that it allocates the
2001 * minimum number of pages to satisfy the request. alloc_pages() can only
2002 * allocate memory in power-of-two pages.
2004 * This function is also limited by MAX_ORDER.
2006 * Memory allocated by this function must be released by free_pages_exact().
2008 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2010 unsigned int order = get_order(size);
2013 addr = __get_free_pages(gfp_mask, order);
2015 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2016 unsigned long used = addr + PAGE_ALIGN(size);
2018 split_page(virt_to_page((void *)addr), order);
2019 while (used < alloc_end) {
2025 return (void *)addr;
2027 EXPORT_SYMBOL(alloc_pages_exact);
2030 * free_pages_exact - release memory allocated via alloc_pages_exact()
2031 * @virt: the value returned by alloc_pages_exact.
2032 * @size: size of allocation, same value as passed to alloc_pages_exact().
2034 * Release the memory allocated by a previous call to alloc_pages_exact.
2036 void free_pages_exact(void *virt, size_t size)
2038 unsigned long addr = (unsigned long)virt;
2039 unsigned long end = addr + PAGE_ALIGN(size);
2041 while (addr < end) {
2046 EXPORT_SYMBOL(free_pages_exact);
2048 static unsigned int nr_free_zone_pages(int offset)
2053 /* Just pick one node, since fallback list is circular */
2054 unsigned int sum = 0;
2056 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2058 for_each_zone_zonelist(zone, z, zonelist, offset) {
2059 unsigned long size = zone->present_pages;
2060 unsigned long high = high_wmark_pages(zone);
2069 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2071 unsigned int nr_free_buffer_pages(void)
2073 return nr_free_zone_pages(gfp_zone(GFP_USER));
2075 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2078 * Amount of free RAM allocatable within all zones
2080 unsigned int nr_free_pagecache_pages(void)
2082 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2085 static inline void show_node(struct zone *zone)
2088 printk("Node %d ", zone_to_nid(zone));
2091 void si_meminfo(struct sysinfo *val)
2093 val->totalram = totalram_pages;
2095 val->freeram = global_page_state(NR_FREE_PAGES);
2096 val->bufferram = nr_blockdev_pages();
2097 val->totalhigh = totalhigh_pages;
2098 val->freehigh = nr_free_highpages();
2099 val->mem_unit = PAGE_SIZE;
2102 EXPORT_SYMBOL(si_meminfo);
2105 void si_meminfo_node(struct sysinfo *val, int nid)
2107 pg_data_t *pgdat = NODE_DATA(nid);
2109 val->totalram = pgdat->node_present_pages;
2110 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2111 #ifdef CONFIG_HIGHMEM
2112 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2113 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2119 val->mem_unit = PAGE_SIZE;
2123 #define K(x) ((x) << (PAGE_SHIFT-10))
2126 * Show free area list (used inside shift_scroll-lock stuff)
2127 * We also calculate the percentage fragmentation. We do this by counting the
2128 * memory on each free list with the exception of the first item on the list.
2130 void show_free_areas(void)
2135 for_each_populated_zone(zone) {
2137 printk("%s per-cpu:\n", zone->name);
2139 for_each_online_cpu(cpu) {
2140 struct per_cpu_pageset *pageset;
2142 pageset = zone_pcp(zone, cpu);
2144 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2145 cpu, pageset->pcp.high,
2146 pageset->pcp.batch, pageset->pcp.count);
2150 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2151 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2153 " dirty:%lu writeback:%lu unstable:%lu buffer:%lu\n"
2154 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2155 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2156 global_page_state(NR_ACTIVE_ANON),
2157 global_page_state(NR_INACTIVE_ANON),
2158 global_page_state(NR_ISOLATED_ANON),
2159 global_page_state(NR_ACTIVE_FILE),
2160 global_page_state(NR_INACTIVE_FILE),
2161 global_page_state(NR_ISOLATED_FILE),
2162 global_page_state(NR_UNEVICTABLE),
2163 global_page_state(NR_FILE_DIRTY),
2164 global_page_state(NR_WRITEBACK),
2165 global_page_state(NR_UNSTABLE_NFS),
2166 nr_blockdev_pages(),
2167 global_page_state(NR_FREE_PAGES),
2168 global_page_state(NR_SLAB_RECLAIMABLE),
2169 global_page_state(NR_SLAB_UNRECLAIMABLE),
2170 global_page_state(NR_FILE_MAPPED),
2171 global_page_state(NR_SHMEM),
2172 global_page_state(NR_PAGETABLE),
2173 global_page_state(NR_BOUNCE));
2175 for_each_populated_zone(zone) {
2184 " active_anon:%lukB"
2185 " inactive_anon:%lukB"
2186 " active_file:%lukB"
2187 " inactive_file:%lukB"
2188 " unevictable:%lukB"
2189 " isolated(anon):%lukB"
2190 " isolated(file):%lukB"
2197 " slab_reclaimable:%lukB"
2198 " slab_unreclaimable:%lukB"
2199 " kernel_stack:%lukB"
2203 " writeback_tmp:%lukB"
2204 " pages_scanned:%lu"
2205 " all_unreclaimable? %s"
2208 K(zone_page_state(zone, NR_FREE_PAGES)),
2209 K(min_wmark_pages(zone)),
2210 K(low_wmark_pages(zone)),
2211 K(high_wmark_pages(zone)),
2212 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2213 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2214 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2215 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2216 K(zone_page_state(zone, NR_UNEVICTABLE)),
2217 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2218 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2219 K(zone->present_pages),
2220 K(zone_page_state(zone, NR_MLOCK)),
2221 K(zone_page_state(zone, NR_FILE_DIRTY)),
2222 K(zone_page_state(zone, NR_WRITEBACK)),
2223 K(zone_page_state(zone, NR_FILE_MAPPED)),
2224 K(zone_page_state(zone, NR_SHMEM)),
2225 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2226 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2227 zone_page_state(zone, NR_KERNEL_STACK) *
2229 K(zone_page_state(zone, NR_PAGETABLE)),
2230 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2231 K(zone_page_state(zone, NR_BOUNCE)),
2232 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2233 zone->pages_scanned,
2234 (zone_is_all_unreclaimable(zone) ? "yes" : "no")
2236 printk("lowmem_reserve[]:");
2237 for (i = 0; i < MAX_NR_ZONES; i++)
2238 printk(" %lu", zone->lowmem_reserve[i]);
2242 for_each_populated_zone(zone) {
2243 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2246 printk("%s: ", zone->name);
2248 spin_lock_irqsave(&zone->lock, flags);
2249 for (order = 0; order < MAX_ORDER; order++) {
2250 nr[order] = zone->free_area[order].nr_free;
2251 total += nr[order] << order;
2253 spin_unlock_irqrestore(&zone->lock, flags);
2254 for (order = 0; order < MAX_ORDER; order++)
2255 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2256 printk("= %lukB\n", K(total));
2259 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2261 show_swap_cache_info();
2264 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2266 zoneref->zone = zone;
2267 zoneref->zone_idx = zone_idx(zone);
2271 * Builds allocation fallback zone lists.
2273 * Add all populated zones of a node to the zonelist.
2275 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2276 int nr_zones, enum zone_type zone_type)
2280 BUG_ON(zone_type >= MAX_NR_ZONES);
2285 zone = pgdat->node_zones + zone_type;
2286 if (populated_zone(zone)) {
2287 zoneref_set_zone(zone,
2288 &zonelist->_zonerefs[nr_zones++]);
2289 check_highest_zone(zone_type);
2292 } while (zone_type);
2299 * 0 = automatic detection of better ordering.
2300 * 1 = order by ([node] distance, -zonetype)
2301 * 2 = order by (-zonetype, [node] distance)
2303 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2304 * the same zonelist. So only NUMA can configure this param.
2306 #define ZONELIST_ORDER_DEFAULT 0
2307 #define ZONELIST_ORDER_NODE 1
2308 #define ZONELIST_ORDER_ZONE 2
2310 /* zonelist order in the kernel.
2311 * set_zonelist_order() will set this to NODE or ZONE.
2313 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2314 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2318 /* The value user specified ....changed by config */
2319 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2320 /* string for sysctl */
2321 #define NUMA_ZONELIST_ORDER_LEN 16
2322 char numa_zonelist_order[16] = "default";
2325 * interface for configure zonelist ordering.
2326 * command line option "numa_zonelist_order"
2327 * = "[dD]efault - default, automatic configuration.
2328 * = "[nN]ode - order by node locality, then by zone within node
2329 * = "[zZ]one - order by zone, then by locality within zone
2332 static int __parse_numa_zonelist_order(char *s)
2334 if (*s == 'd' || *s == 'D') {
2335 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2336 } else if (*s == 'n' || *s == 'N') {
2337 user_zonelist_order = ZONELIST_ORDER_NODE;
2338 } else if (*s == 'z' || *s == 'Z') {
2339 user_zonelist_order = ZONELIST_ORDER_ZONE;
2342 "Ignoring invalid numa_zonelist_order value: "
2349 static __init int setup_numa_zonelist_order(char *s)
2352 return __parse_numa_zonelist_order(s);
2355 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2358 * sysctl handler for numa_zonelist_order
2360 int numa_zonelist_order_handler(ctl_table *table, int write,
2361 struct file *file, void __user *buffer, size_t *length,
2364 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2368 strncpy(saved_string, (char*)table->data,
2369 NUMA_ZONELIST_ORDER_LEN);
2370 ret = proc_dostring(table, write, file, buffer, length, ppos);
2374 int oldval = user_zonelist_order;
2375 if (__parse_numa_zonelist_order((char*)table->data)) {
2377 * bogus value. restore saved string
2379 strncpy((char*)table->data, saved_string,
2380 NUMA_ZONELIST_ORDER_LEN);
2381 user_zonelist_order = oldval;
2382 } else if (oldval != user_zonelist_order)
2383 build_all_zonelists();
2389 #define MAX_NODE_LOAD (nr_online_nodes)
2390 static int node_load[MAX_NUMNODES];
2393 * find_next_best_node - find the next node that should appear in a given node's fallback list
2394 * @node: node whose fallback list we're appending
2395 * @used_node_mask: nodemask_t of already used nodes
2397 * We use a number of factors to determine which is the next node that should
2398 * appear on a given node's fallback list. The node should not have appeared
2399 * already in @node's fallback list, and it should be the next closest node
2400 * according to the distance array (which contains arbitrary distance values
2401 * from each node to each node in the system), and should also prefer nodes
2402 * with no CPUs, since presumably they'll have very little allocation pressure
2403 * on them otherwise.
2404 * It returns -1 if no node is found.
2406 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2409 int min_val = INT_MAX;
2411 const struct cpumask *tmp = cpumask_of_node(0);
2413 /* Use the local node if we haven't already */
2414 if (!node_isset(node, *used_node_mask)) {
2415 node_set(node, *used_node_mask);
2419 for_each_node_state(n, N_HIGH_MEMORY) {
2421 /* Don't want a node to appear more than once */
2422 if (node_isset(n, *used_node_mask))
2425 /* Use the distance array to find the distance */
2426 val = node_distance(node, n);
2428 /* Penalize nodes under us ("prefer the next node") */
2431 /* Give preference to headless and unused nodes */
2432 tmp = cpumask_of_node(n);
2433 if (!cpumask_empty(tmp))
2434 val += PENALTY_FOR_NODE_WITH_CPUS;
2436 /* Slight preference for less loaded node */
2437 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2438 val += node_load[n];
2440 if (val < min_val) {
2447 node_set(best_node, *used_node_mask);
2454 * Build zonelists ordered by node and zones within node.
2455 * This results in maximum locality--normal zone overflows into local
2456 * DMA zone, if any--but risks exhausting DMA zone.
2458 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2461 struct zonelist *zonelist;
2463 zonelist = &pgdat->node_zonelists[0];
2464 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2466 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2468 zonelist->_zonerefs[j].zone = NULL;
2469 zonelist->_zonerefs[j].zone_idx = 0;
2473 * Build gfp_thisnode zonelists
2475 static void build_thisnode_zonelists(pg_data_t *pgdat)
2478 struct zonelist *zonelist;
2480 zonelist = &pgdat->node_zonelists[1];
2481 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2482 zonelist->_zonerefs[j].zone = NULL;
2483 zonelist->_zonerefs[j].zone_idx = 0;
2487 * Build zonelists ordered by zone and nodes within zones.
2488 * This results in conserving DMA zone[s] until all Normal memory is
2489 * exhausted, but results in overflowing to remote node while memory
2490 * may still exist in local DMA zone.
2492 static int node_order[MAX_NUMNODES];
2494 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2497 int zone_type; /* needs to be signed */
2499 struct zonelist *zonelist;
2501 zonelist = &pgdat->node_zonelists[0];
2503 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2504 for (j = 0; j < nr_nodes; j++) {
2505 node = node_order[j];
2506 z = &NODE_DATA(node)->node_zones[zone_type];
2507 if (populated_zone(z)) {
2509 &zonelist->_zonerefs[pos++]);
2510 check_highest_zone(zone_type);
2514 zonelist->_zonerefs[pos].zone = NULL;
2515 zonelist->_zonerefs[pos].zone_idx = 0;
2518 static int default_zonelist_order(void)
2521 unsigned long low_kmem_size,total_size;
2525 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2526 * If they are really small and used heavily, the system can fall
2527 * into OOM very easily.
2528 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2530 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2533 for_each_online_node(nid) {
2534 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2535 z = &NODE_DATA(nid)->node_zones[zone_type];
2536 if (populated_zone(z)) {
2537 if (zone_type < ZONE_NORMAL)
2538 low_kmem_size += z->present_pages;
2539 total_size += z->present_pages;
2543 if (!low_kmem_size || /* there are no DMA area. */
2544 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2545 return ZONELIST_ORDER_NODE;
2547 * look into each node's config.
2548 * If there is a node whose DMA/DMA32 memory is very big area on
2549 * local memory, NODE_ORDER may be suitable.
2551 average_size = total_size /
2552 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2553 for_each_online_node(nid) {
2556 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2557 z = &NODE_DATA(nid)->node_zones[zone_type];
2558 if (populated_zone(z)) {
2559 if (zone_type < ZONE_NORMAL)
2560 low_kmem_size += z->present_pages;
2561 total_size += z->present_pages;
2564 if (low_kmem_size &&
2565 total_size > average_size && /* ignore small node */
2566 low_kmem_size > total_size * 70/100)
2567 return ZONELIST_ORDER_NODE;
2569 return ZONELIST_ORDER_ZONE;
2572 static void set_zonelist_order(void)
2574 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2575 current_zonelist_order = default_zonelist_order();
2577 current_zonelist_order = user_zonelist_order;
2580 static void build_zonelists(pg_data_t *pgdat)
2584 nodemask_t used_mask;
2585 int local_node, prev_node;
2586 struct zonelist *zonelist;
2587 int order = current_zonelist_order;
2589 /* initialize zonelists */
2590 for (i = 0; i < MAX_ZONELISTS; i++) {
2591 zonelist = pgdat->node_zonelists + i;
2592 zonelist->_zonerefs[0].zone = NULL;
2593 zonelist->_zonerefs[0].zone_idx = 0;
2596 /* NUMA-aware ordering of nodes */
2597 local_node = pgdat->node_id;
2598 load = nr_online_nodes;
2599 prev_node = local_node;
2600 nodes_clear(used_mask);
2602 memset(node_order, 0, sizeof(node_order));
2605 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2606 int distance = node_distance(local_node, node);
2609 * If another node is sufficiently far away then it is better
2610 * to reclaim pages in a zone before going off node.
2612 if (distance > RECLAIM_DISTANCE)
2613 zone_reclaim_mode = 1;
2616 * We don't want to pressure a particular node.
2617 * So adding penalty to the first node in same
2618 * distance group to make it round-robin.
2620 if (distance != node_distance(local_node, prev_node))
2621 node_load[node] = load;
2625 if (order == ZONELIST_ORDER_NODE)
2626 build_zonelists_in_node_order(pgdat, node);
2628 node_order[j++] = node; /* remember order */
2631 if (order == ZONELIST_ORDER_ZONE) {
2632 /* calculate node order -- i.e., DMA last! */
2633 build_zonelists_in_zone_order(pgdat, j);
2636 build_thisnode_zonelists(pgdat);
2639 /* Construct the zonelist performance cache - see further mmzone.h */
2640 static void build_zonelist_cache(pg_data_t *pgdat)
2642 struct zonelist *zonelist;
2643 struct zonelist_cache *zlc;
2646 zonelist = &pgdat->node_zonelists[0];
2647 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2648 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2649 for (z = zonelist->_zonerefs; z->zone; z++)
2650 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2654 #else /* CONFIG_NUMA */
2656 static void set_zonelist_order(void)
2658 current_zonelist_order = ZONELIST_ORDER_ZONE;
2661 static void build_zonelists(pg_data_t *pgdat)
2663 int node, local_node;
2665 struct zonelist *zonelist;
2667 local_node = pgdat->node_id;
2669 zonelist = &pgdat->node_zonelists[0];
2670 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2673 * Now we build the zonelist so that it contains the zones
2674 * of all the other nodes.
2675 * We don't want to pressure a particular node, so when
2676 * building the zones for node N, we make sure that the
2677 * zones coming right after the local ones are those from
2678 * node N+1 (modulo N)
2680 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2681 if (!node_online(node))
2683 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2686 for (node = 0; node < local_node; node++) {
2687 if (!node_online(node))
2689 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2693 zonelist->_zonerefs[j].zone = NULL;
2694 zonelist->_zonerefs[j].zone_idx = 0;
2697 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2698 static void build_zonelist_cache(pg_data_t *pgdat)
2700 pgdat->node_zonelists[0].zlcache_ptr = NULL;
2703 #endif /* CONFIG_NUMA */
2705 /* return values int ....just for stop_machine() */
2706 static int __build_all_zonelists(void *dummy)
2711 memset(node_load, 0, sizeof(node_load));
2713 for_each_online_node(nid) {
2714 pg_data_t *pgdat = NODE_DATA(nid);
2716 build_zonelists(pgdat);
2717 build_zonelist_cache(pgdat);
2722 void build_all_zonelists(void)
2724 set_zonelist_order();
2726 if (system_state == SYSTEM_BOOTING) {
2727 __build_all_zonelists(NULL);
2728 mminit_verify_zonelist();
2729 cpuset_init_current_mems_allowed();
2731 /* we have to stop all cpus to guarantee there is no user
2733 stop_machine(__build_all_zonelists, NULL, NULL);
2734 /* cpuset refresh routine should be here */
2736 vm_total_pages = nr_free_pagecache_pages();
2738 * Disable grouping by mobility if the number of pages in the
2739 * system is too low to allow the mechanism to work. It would be
2740 * more accurate, but expensive to check per-zone. This check is
2741 * made on memory-hotadd so a system can start with mobility
2742 * disabled and enable it later
2744 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
2745 page_group_by_mobility_disabled = 1;
2747 page_group_by_mobility_disabled = 0;
2749 printk("Built %i zonelists in %s order, mobility grouping %s. "
2750 "Total pages: %ld\n",
2752 zonelist_order_name[current_zonelist_order],
2753 page_group_by_mobility_disabled ? "off" : "on",
2756 printk("Policy zone: %s\n", zone_names[policy_zone]);
2761 * Helper functions to size the waitqueue hash table.
2762 * Essentially these want to choose hash table sizes sufficiently
2763 * large so that collisions trying to wait on pages are rare.
2764 * But in fact, the number of active page waitqueues on typical
2765 * systems is ridiculously low, less than 200. So this is even
2766 * conservative, even though it seems large.
2768 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2769 * waitqueues, i.e. the size of the waitq table given the number of pages.
2771 #define PAGES_PER_WAITQUEUE 256
2773 #ifndef CONFIG_MEMORY_HOTPLUG
2774 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2776 unsigned long size = 1;
2778 pages /= PAGES_PER_WAITQUEUE;
2780 while (size < pages)
2784 * Once we have dozens or even hundreds of threads sleeping
2785 * on IO we've got bigger problems than wait queue collision.
2786 * Limit the size of the wait table to a reasonable size.
2788 size = min(size, 4096UL);
2790 return max(size, 4UL);
2794 * A zone's size might be changed by hot-add, so it is not possible to determine
2795 * a suitable size for its wait_table. So we use the maximum size now.
2797 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2799 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2800 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2801 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2803 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2804 * or more by the traditional way. (See above). It equals:
2806 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2807 * ia64(16K page size) : = ( 8G + 4M)byte.
2808 * powerpc (64K page size) : = (32G +16M)byte.
2810 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2817 * This is an integer logarithm so that shifts can be used later
2818 * to extract the more random high bits from the multiplicative
2819 * hash function before the remainder is taken.
2821 static inline unsigned long wait_table_bits(unsigned long size)
2826 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2829 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2830 * of blocks reserved is based on min_wmark_pages(zone). The memory within
2831 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
2832 * higher will lead to a bigger reserve which will get freed as contiguous
2833 * blocks as reclaim kicks in
2835 static void setup_zone_migrate_reserve(struct zone *zone)
2837 unsigned long start_pfn, pfn, end_pfn;
2839 unsigned long block_migratetype;
2842 /* Get the start pfn, end pfn and the number of blocks to reserve */
2843 start_pfn = zone->zone_start_pfn;
2844 end_pfn = start_pfn + zone->spanned_pages;
2845 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
2849 * Reserve blocks are generally in place to help high-order atomic
2850 * allocations that are short-lived. A min_free_kbytes value that
2851 * would result in more than 2 reserve blocks for atomic allocations
2852 * is assumed to be in place to help anti-fragmentation for the
2853 * future allocation of hugepages at runtime.
2855 reserve = min(2, reserve);
2857 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
2858 if (!pfn_valid(pfn))
2860 page = pfn_to_page(pfn);
2862 /* Watch out for overlapping nodes */
2863 if (page_to_nid(page) != zone_to_nid(zone))
2866 /* Blocks with reserved pages will never free, skip them. */
2867 if (PageReserved(page))
2870 block_migratetype = get_pageblock_migratetype(page);
2872 /* If this block is reserved, account for it */
2873 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
2878 /* Suitable for reserving if this block is movable */
2879 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
2880 set_pageblock_migratetype(page, MIGRATE_RESERVE);
2881 move_freepages_block(zone, page, MIGRATE_RESERVE);
2887 * If the reserve is met and this is a previous reserved block,
2890 if (block_migratetype == MIGRATE_RESERVE) {
2891 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2892 move_freepages_block(zone, page, MIGRATE_MOVABLE);
2898 * Initially all pages are reserved - free ones are freed
2899 * up by free_all_bootmem() once the early boot process is
2900 * done. Non-atomic initialization, single-pass.
2902 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
2903 unsigned long start_pfn, enum memmap_context context)
2906 unsigned long end_pfn = start_pfn + size;
2910 if (highest_memmap_pfn < end_pfn - 1)
2911 highest_memmap_pfn = end_pfn - 1;
2913 z = &NODE_DATA(nid)->node_zones[zone];
2914 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
2916 * There can be holes in boot-time mem_map[]s
2917 * handed to this function. They do not
2918 * exist on hotplugged memory.
2920 if (context == MEMMAP_EARLY) {
2921 if (!early_pfn_valid(pfn))
2923 if (!early_pfn_in_nid(pfn, nid))
2926 page = pfn_to_page(pfn);
2927 set_page_links(page, zone, nid, pfn);
2928 mminit_verify_page_links(page, zone, nid, pfn);
2929 init_page_count(page);
2930 reset_page_mapcount(page);
2931 SetPageReserved(page);
2933 * Mark the block movable so that blocks are reserved for
2934 * movable at startup. This will force kernel allocations
2935 * to reserve their blocks rather than leaking throughout
2936 * the address space during boot when many long-lived
2937 * kernel allocations are made. Later some blocks near
2938 * the start are marked MIGRATE_RESERVE by
2939 * setup_zone_migrate_reserve()
2941 * bitmap is created for zone's valid pfn range. but memmap
2942 * can be created for invalid pages (for alignment)
2943 * check here not to call set_pageblock_migratetype() against
2946 if ((z->zone_start_pfn <= pfn)
2947 && (pfn < z->zone_start_pfn + z->spanned_pages)
2948 && !(pfn & (pageblock_nr_pages - 1)))
2949 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2951 INIT_LIST_HEAD(&page->lru);
2952 #ifdef WANT_PAGE_VIRTUAL
2953 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2954 if (!is_highmem_idx(zone))
2955 set_page_address(page, __va(pfn << PAGE_SHIFT));
2960 static void __meminit zone_init_free_lists(struct zone *zone)
2963 for_each_migratetype_order(order, t) {
2964 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
2965 zone->free_area[order].nr_free = 0;
2969 #ifndef __HAVE_ARCH_MEMMAP_INIT
2970 #define memmap_init(size, nid, zone, start_pfn) \
2971 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
2974 static int zone_batchsize(struct zone *zone)
2980 * The per-cpu-pages pools are set to around 1000th of the
2981 * size of the zone. But no more than 1/2 of a meg.
2983 * OK, so we don't know how big the cache is. So guess.
2985 batch = zone->present_pages / 1024;
2986 if (batch * PAGE_SIZE > 512 * 1024)
2987 batch = (512 * 1024) / PAGE_SIZE;
2988 batch /= 4; /* We effectively *= 4 below */
2993 * Clamp the batch to a 2^n - 1 value. Having a power
2994 * of 2 value was found to be more likely to have
2995 * suboptimal cache aliasing properties in some cases.
2997 * For example if 2 tasks are alternately allocating
2998 * batches of pages, one task can end up with a lot
2999 * of pages of one half of the possible page colors
3000 * and the other with pages of the other colors.
3002 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3007 /* The deferral and batching of frees should be suppressed under NOMMU
3010 * The problem is that NOMMU needs to be able to allocate large chunks
3011 * of contiguous memory as there's no hardware page translation to
3012 * assemble apparent contiguous memory from discontiguous pages.
3014 * Queueing large contiguous runs of pages for batching, however,
3015 * causes the pages to actually be freed in smaller chunks. As there
3016 * can be a significant delay between the individual batches being
3017 * recycled, this leads to the once large chunks of space being
3018 * fragmented and becoming unavailable for high-order allocations.
3024 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3026 struct per_cpu_pages *pcp;
3028 memset(p, 0, sizeof(*p));
3032 pcp->high = 6 * batch;
3033 pcp->batch = max(1UL, 1 * batch);
3034 INIT_LIST_HEAD(&pcp->list);
3038 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3039 * to the value high for the pageset p.
3042 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3045 struct per_cpu_pages *pcp;
3049 pcp->batch = max(1UL, high/4);
3050 if ((high/4) > (PAGE_SHIFT * 8))
3051 pcp->batch = PAGE_SHIFT * 8;
3057 * Boot pageset table. One per cpu which is going to be used for all
3058 * zones and all nodes. The parameters will be set in such a way
3059 * that an item put on a list will immediately be handed over to
3060 * the buddy list. This is safe since pageset manipulation is done
3061 * with interrupts disabled.
3063 * Some NUMA counter updates may also be caught by the boot pagesets.
3065 * The boot_pagesets must be kept even after bootup is complete for
3066 * unused processors and/or zones. They do play a role for bootstrapping
3067 * hotplugged processors.
3069 * zoneinfo_show() and maybe other functions do
3070 * not check if the processor is online before following the pageset pointer.
3071 * Other parts of the kernel may not check if the zone is available.
3073 static struct per_cpu_pageset boot_pageset[NR_CPUS];
3076 * Dynamically allocate memory for the
3077 * per cpu pageset array in struct zone.
3079 static int __cpuinit process_zones(int cpu)
3081 struct zone *zone, *dzone;
3082 int node = cpu_to_node(cpu);
3084 node_set_state(node, N_CPU); /* this node has a cpu */
3086 for_each_populated_zone(zone) {
3087 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
3089 if (!zone_pcp(zone, cpu))
3092 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
3094 if (percpu_pagelist_fraction)
3095 setup_pagelist_highmark(zone_pcp(zone, cpu),
3096 (zone->present_pages / percpu_pagelist_fraction));
3101 for_each_zone(dzone) {
3102 if (!populated_zone(dzone))
3106 kfree(zone_pcp(dzone, cpu));
3107 zone_pcp(dzone, cpu) = &boot_pageset[cpu];
3112 static inline void free_zone_pagesets(int cpu)
3116 for_each_zone(zone) {
3117 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
3119 /* Free per_cpu_pageset if it is slab allocated */
3120 if (pset != &boot_pageset[cpu])
3122 zone_pcp(zone, cpu) = &boot_pageset[cpu];
3126 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
3127 unsigned long action,
3130 int cpu = (long)hcpu;
3131 int ret = NOTIFY_OK;
3134 case CPU_UP_PREPARE:
3135 case CPU_UP_PREPARE_FROZEN:
3136 if (process_zones(cpu))
3139 case CPU_UP_CANCELED:
3140 case CPU_UP_CANCELED_FROZEN:
3142 case CPU_DEAD_FROZEN:
3143 free_zone_pagesets(cpu);
3151 static struct notifier_block __cpuinitdata pageset_notifier =
3152 { &pageset_cpuup_callback, NULL, 0 };
3154 void __init setup_per_cpu_pageset(void)
3158 /* Initialize per_cpu_pageset for cpu 0.
3159 * A cpuup callback will do this for every cpu
3160 * as it comes online
3162 err = process_zones(smp_processor_id());
3164 register_cpu_notifier(&pageset_notifier);
3169 static noinline __init_refok
3170 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3173 struct pglist_data *pgdat = zone->zone_pgdat;
3177 * The per-page waitqueue mechanism uses hashed waitqueues
3180 zone->wait_table_hash_nr_entries =
3181 wait_table_hash_nr_entries(zone_size_pages);
3182 zone->wait_table_bits =
3183 wait_table_bits(zone->wait_table_hash_nr_entries);
3184 alloc_size = zone->wait_table_hash_nr_entries
3185 * sizeof(wait_queue_head_t);
3187 if (!slab_is_available()) {
3188 zone->wait_table = (wait_queue_head_t *)
3189 alloc_bootmem_node(pgdat, alloc_size);
3192 * This case means that a zone whose size was 0 gets new memory
3193 * via memory hot-add.
3194 * But it may be the case that a new node was hot-added. In
3195 * this case vmalloc() will not be able to use this new node's
3196 * memory - this wait_table must be initialized to use this new
3197 * node itself as well.
3198 * To use this new node's memory, further consideration will be
3201 zone->wait_table = vmalloc(alloc_size);
3203 if (!zone->wait_table)
3206 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3207 init_waitqueue_head(zone->wait_table + i);
3212 static int __zone_pcp_update(void *data)
3214 struct zone *zone = data;
3216 unsigned long batch = zone_batchsize(zone), flags;
3218 for (cpu = 0; cpu < NR_CPUS; cpu++) {
3219 struct per_cpu_pageset *pset;
3220 struct per_cpu_pages *pcp;
3222 pset = zone_pcp(zone, cpu);
3225 local_irq_save(flags);
3226 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
3227 setup_pageset(pset, batch);
3228 local_irq_restore(flags);
3233 void zone_pcp_update(struct zone *zone)
3235 stop_machine(__zone_pcp_update, zone, NULL);
3238 static __meminit void zone_pcp_init(struct zone *zone)
3241 unsigned long batch = zone_batchsize(zone);
3243 for (cpu = 0; cpu < NR_CPUS; cpu++) {
3245 /* Early boot. Slab allocator not functional yet */
3246 zone_pcp(zone, cpu) = &boot_pageset[cpu];
3247 setup_pageset(&boot_pageset[cpu],0);
3249 setup_pageset(zone_pcp(zone,cpu), batch);
3252 if (zone->present_pages)
3253 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
3254 zone->name, zone->present_pages, batch);
3257 __meminit int init_currently_empty_zone(struct zone *zone,
3258 unsigned long zone_start_pfn,
3260 enum memmap_context context)
3262 struct pglist_data *pgdat = zone->zone_pgdat;
3264 ret = zone_wait_table_init(zone, size);
3267 pgdat->nr_zones = zone_idx(zone) + 1;
3269 zone->zone_start_pfn = zone_start_pfn;
3271 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3272 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3274 (unsigned long)zone_idx(zone),
3275 zone_start_pfn, (zone_start_pfn + size));
3277 zone_init_free_lists(zone);
3282 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3284 * Basic iterator support. Return the first range of PFNs for a node
3285 * Note: nid == MAX_NUMNODES returns first region regardless of node
3287 static int __meminit first_active_region_index_in_nid(int nid)
3291 for (i = 0; i < nr_nodemap_entries; i++)
3292 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3299 * Basic iterator support. Return the next active range of PFNs for a node
3300 * Note: nid == MAX_NUMNODES returns next region regardless of node
3302 static int __meminit next_active_region_index_in_nid(int index, int nid)
3304 for (index = index + 1; index < nr_nodemap_entries; index++)
3305 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3311 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3313 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3314 * Architectures may implement their own version but if add_active_range()
3315 * was used and there are no special requirements, this is a convenient
3318 int __meminit __early_pfn_to_nid(unsigned long pfn)
3322 for (i = 0; i < nr_nodemap_entries; i++) {
3323 unsigned long start_pfn = early_node_map[i].start_pfn;
3324 unsigned long end_pfn = early_node_map[i].end_pfn;
3326 if (start_pfn <= pfn && pfn < end_pfn)
3327 return early_node_map[i].nid;
3329 /* This is a memory hole */
3332 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3334 int __meminit early_pfn_to_nid(unsigned long pfn)
3338 nid = __early_pfn_to_nid(pfn);
3341 /* just returns 0 */
3345 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3346 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3350 nid = __early_pfn_to_nid(pfn);
3351 if (nid >= 0 && nid != node)
3357 /* Basic iterator support to walk early_node_map[] */
3358 #define for_each_active_range_index_in_nid(i, nid) \
3359 for (i = first_active_region_index_in_nid(nid); i != -1; \
3360 i = next_active_region_index_in_nid(i, nid))
3363 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3364 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3365 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3367 * If an architecture guarantees that all ranges registered with
3368 * add_active_ranges() contain no holes and may be freed, this
3369 * this function may be used instead of calling free_bootmem() manually.
3371 void __init free_bootmem_with_active_regions(int nid,
3372 unsigned long max_low_pfn)
3376 for_each_active_range_index_in_nid(i, nid) {
3377 unsigned long size_pages = 0;
3378 unsigned long end_pfn = early_node_map[i].end_pfn;
3380 if (early_node_map[i].start_pfn >= max_low_pfn)
3383 if (end_pfn > max_low_pfn)
3384 end_pfn = max_low_pfn;
3386 size_pages = end_pfn - early_node_map[i].start_pfn;
3387 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3388 PFN_PHYS(early_node_map[i].start_pfn),
3389 size_pages << PAGE_SHIFT);
3393 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3398 for_each_active_range_index_in_nid(i, nid) {
3399 ret = work_fn(early_node_map[i].start_pfn,
3400 early_node_map[i].end_pfn, data);
3406 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3407 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3409 * If an architecture guarantees that all ranges registered with
3410 * add_active_ranges() contain no holes and may be freed, this
3411 * function may be used instead of calling memory_present() manually.
3413 void __init sparse_memory_present_with_active_regions(int nid)
3417 for_each_active_range_index_in_nid(i, nid)
3418 memory_present(early_node_map[i].nid,
3419 early_node_map[i].start_pfn,
3420 early_node_map[i].end_pfn);
3424 * get_pfn_range_for_nid - Return the start and end page frames for a node
3425 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3426 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3427 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3429 * It returns the start and end page frame of a node based on information
3430 * provided by an arch calling add_active_range(). If called for a node
3431 * with no available memory, a warning is printed and the start and end
3434 void __meminit get_pfn_range_for_nid(unsigned int nid,
3435 unsigned long *start_pfn, unsigned long *end_pfn)
3441 for_each_active_range_index_in_nid(i, nid) {
3442 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3443 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3446 if (*start_pfn == -1UL)
3451 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3452 * assumption is made that zones within a node are ordered in monotonic
3453 * increasing memory addresses so that the "highest" populated zone is used
3455 static void __init find_usable_zone_for_movable(void)
3458 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3459 if (zone_index == ZONE_MOVABLE)
3462 if (arch_zone_highest_possible_pfn[zone_index] >
3463 arch_zone_lowest_possible_pfn[zone_index])
3467 VM_BUG_ON(zone_index == -1);
3468 movable_zone = zone_index;
3472 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3473 * because it is sized independant of architecture. Unlike the other zones,
3474 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3475 * in each node depending on the size of each node and how evenly kernelcore
3476 * is distributed. This helper function adjusts the zone ranges
3477 * provided by the architecture for a given node by using the end of the
3478 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3479 * zones within a node are in order of monotonic increases memory addresses
3481 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3482 unsigned long zone_type,
3483 unsigned long node_start_pfn,
3484 unsigned long node_end_pfn,
3485 unsigned long *zone_start_pfn,
3486 unsigned long *zone_end_pfn)
3488 /* Only adjust if ZONE_MOVABLE is on this node */
3489 if (zone_movable_pfn[nid]) {
3490 /* Size ZONE_MOVABLE */
3491 if (zone_type == ZONE_MOVABLE) {
3492 *zone_start_pfn = zone_movable_pfn[nid];
3493 *zone_end_pfn = min(node_end_pfn,
3494 arch_zone_highest_possible_pfn[movable_zone]);
3496 /* Adjust for ZONE_MOVABLE starting within this range */
3497 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3498 *zone_end_pfn > zone_movable_pfn[nid]) {
3499 *zone_end_pfn = zone_movable_pfn[nid];
3501 /* Check if this whole range is within ZONE_MOVABLE */
3502 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3503 *zone_start_pfn = *zone_end_pfn;
3508 * Return the number of pages a zone spans in a node, including holes
3509 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3511 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3512 unsigned long zone_type,
3513 unsigned long *ignored)
3515 unsigned long node_start_pfn, node_end_pfn;
3516 unsigned long zone_start_pfn, zone_end_pfn;
3518 /* Get the start and end of the node and zone */
3519 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3520 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3521 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3522 adjust_zone_range_for_zone_movable(nid, zone_type,
3523 node_start_pfn, node_end_pfn,
3524 &zone_start_pfn, &zone_end_pfn);
3526 /* Check that this node has pages within the zone's required range */
3527 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3530 /* Move the zone boundaries inside the node if necessary */
3531 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3532 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3534 /* Return the spanned pages */
3535 return zone_end_pfn - zone_start_pfn;
3539 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3540 * then all holes in the requested range will be accounted for.
3542 static unsigned long __meminit __absent_pages_in_range(int nid,
3543 unsigned long range_start_pfn,
3544 unsigned long range_end_pfn)
3547 unsigned long prev_end_pfn = 0, hole_pages = 0;
3548 unsigned long start_pfn;
3550 /* Find the end_pfn of the first active range of pfns in the node */
3551 i = first_active_region_index_in_nid(nid);
3555 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3557 /* Account for ranges before physical memory on this node */
3558 if (early_node_map[i].start_pfn > range_start_pfn)
3559 hole_pages = prev_end_pfn - range_start_pfn;
3561 /* Find all holes for the zone within the node */
3562 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3564 /* No need to continue if prev_end_pfn is outside the zone */
3565 if (prev_end_pfn >= range_end_pfn)
3568 /* Make sure the end of the zone is not within the hole */
3569 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3570 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3572 /* Update the hole size cound and move on */
3573 if (start_pfn > range_start_pfn) {
3574 BUG_ON(prev_end_pfn > start_pfn);
3575 hole_pages += start_pfn - prev_end_pfn;
3577 prev_end_pfn = early_node_map[i].end_pfn;
3580 /* Account for ranges past physical memory on this node */
3581 if (range_end_pfn > prev_end_pfn)
3582 hole_pages += range_end_pfn -
3583 max(range_start_pfn, prev_end_pfn);
3589 * absent_pages_in_range - Return number of page frames in holes within a range
3590 * @start_pfn: The start PFN to start searching for holes
3591 * @end_pfn: The end PFN to stop searching for holes
3593 * It returns the number of pages frames in memory holes within a range.
3595 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3596 unsigned long end_pfn)
3598 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3601 /* Return the number of page frames in holes in a zone on a node */
3602 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3603 unsigned long zone_type,
3604 unsigned long *ignored)
3606 unsigned long node_start_pfn, node_end_pfn;
3607 unsigned long zone_start_pfn, zone_end_pfn;
3609 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3610 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3612 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3615 adjust_zone_range_for_zone_movable(nid, zone_type,
3616 node_start_pfn, node_end_pfn,
3617 &zone_start_pfn, &zone_end_pfn);
3618 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3622 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3623 unsigned long zone_type,
3624 unsigned long *zones_size)
3626 return zones_size[zone_type];
3629 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3630 unsigned long zone_type,
3631 unsigned long *zholes_size)
3636 return zholes_size[zone_type];
3641 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3642 unsigned long *zones_size, unsigned long *zholes_size)
3644 unsigned long realtotalpages, totalpages = 0;
3647 for (i = 0; i < MAX_NR_ZONES; i++)
3648 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3650 pgdat->node_spanned_pages = totalpages;
3652 realtotalpages = totalpages;
3653 for (i = 0; i < MAX_NR_ZONES; i++)
3655 zone_absent_pages_in_node(pgdat->node_id, i,
3657 pgdat->node_present_pages = realtotalpages;
3658 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3662 #ifndef CONFIG_SPARSEMEM
3664 * Calculate the size of the zone->blockflags rounded to an unsigned long
3665 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3666 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3667 * round what is now in bits to nearest long in bits, then return it in
3670 static unsigned long __init usemap_size(unsigned long zonesize)
3672 unsigned long usemapsize;
3674 usemapsize = roundup(zonesize, pageblock_nr_pages);
3675 usemapsize = usemapsize >> pageblock_order;
3676 usemapsize *= NR_PAGEBLOCK_BITS;
3677 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
3679 return usemapsize / 8;
3682 static void __init setup_usemap(struct pglist_data *pgdat,
3683 struct zone *zone, unsigned long zonesize)
3685 unsigned long usemapsize = usemap_size(zonesize);
3686 zone->pageblock_flags = NULL;
3688 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
3691 static void inline setup_usemap(struct pglist_data *pgdat,
3692 struct zone *zone, unsigned long zonesize) {}
3693 #endif /* CONFIG_SPARSEMEM */
3695 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3697 /* Return a sensible default order for the pageblock size. */
3698 static inline int pageblock_default_order(void)
3700 if (HPAGE_SHIFT > PAGE_SHIFT)
3701 return HUGETLB_PAGE_ORDER;
3706 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3707 static inline void __init set_pageblock_order(unsigned int order)
3709 /* Check that pageblock_nr_pages has not already been setup */
3710 if (pageblock_order)
3714 * Assume the largest contiguous order of interest is a huge page.
3715 * This value may be variable depending on boot parameters on IA64
3717 pageblock_order = order;
3719 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3722 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3723 * and pageblock_default_order() are unused as pageblock_order is set
3724 * at compile-time. See include/linux/pageblock-flags.h for the values of
3725 * pageblock_order based on the kernel config
3727 static inline int pageblock_default_order(unsigned int order)
3731 #define set_pageblock_order(x) do {} while (0)
3733 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3736 * Set up the zone data structures:
3737 * - mark all pages reserved
3738 * - mark all memory queues empty
3739 * - clear the memory bitmaps
3741 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
3742 unsigned long *zones_size, unsigned long *zholes_size)
3745 int nid = pgdat->node_id;
3746 unsigned long zone_start_pfn = pgdat->node_start_pfn;
3749 pgdat_resize_init(pgdat);
3750 pgdat->nr_zones = 0;
3751 init_waitqueue_head(&pgdat->kswapd_wait);
3752 pgdat->kswapd_max_order = 0;
3753 pgdat_page_cgroup_init(pgdat);
3755 for (j = 0; j < MAX_NR_ZONES; j++) {
3756 struct zone *zone = pgdat->node_zones + j;
3757 unsigned long size, realsize, memmap_pages;
3760 size = zone_spanned_pages_in_node(nid, j, zones_size);
3761 realsize = size - zone_absent_pages_in_node(nid, j,
3765 * Adjust realsize so that it accounts for how much memory
3766 * is used by this zone for memmap. This affects the watermark
3767 * and per-cpu initialisations
3770 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
3771 if (realsize >= memmap_pages) {
3772 realsize -= memmap_pages;
3775 " %s zone: %lu pages used for memmap\n",
3776 zone_names[j], memmap_pages);
3779 " %s zone: %lu pages exceeds realsize %lu\n",
3780 zone_names[j], memmap_pages, realsize);
3782 /* Account for reserved pages */
3783 if (j == 0 && realsize > dma_reserve) {
3784 realsize -= dma_reserve;
3785 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
3786 zone_names[0], dma_reserve);
3789 if (!is_highmem_idx(j))
3790 nr_kernel_pages += realsize;
3791 nr_all_pages += realsize;
3793 zone->spanned_pages = size;
3794 zone->present_pages = realsize;
3797 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
3799 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
3801 zone->name = zone_names[j];
3802 spin_lock_init(&zone->lock);
3803 spin_lock_init(&zone->lru_lock);
3804 zone_seqlock_init(zone);
3805 zone->zone_pgdat = pgdat;
3807 zone->prev_priority = DEF_PRIORITY;
3809 zone_pcp_init(zone);
3811 INIT_LIST_HEAD(&zone->lru[l].list);
3812 zone->lru[l].nr_saved_scan = 0;
3814 zone->reclaim_stat.recent_rotated[0] = 0;
3815 zone->reclaim_stat.recent_rotated[1] = 0;
3816 zone->reclaim_stat.recent_scanned[0] = 0;
3817 zone->reclaim_stat.recent_scanned[1] = 0;
3818 zap_zone_vm_stats(zone);
3823 set_pageblock_order(pageblock_default_order());
3824 setup_usemap(pgdat, zone, size);
3825 ret = init_currently_empty_zone(zone, zone_start_pfn,
3826 size, MEMMAP_EARLY);
3828 memmap_init(size, nid, j, zone_start_pfn);
3829 zone_start_pfn += size;
3833 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
3835 /* Skip empty nodes */
3836 if (!pgdat->node_spanned_pages)
3839 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3840 /* ia64 gets its own node_mem_map, before this, without bootmem */
3841 if (!pgdat->node_mem_map) {
3842 unsigned long size, start, end;
3846 * The zone's endpoints aren't required to be MAX_ORDER
3847 * aligned but the node_mem_map endpoints must be in order
3848 * for the buddy allocator to function correctly.
3850 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
3851 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
3852 end = ALIGN(end, MAX_ORDER_NR_PAGES);
3853 size = (end - start) * sizeof(struct page);
3854 map = alloc_remap(pgdat->node_id, size);
3856 map = alloc_bootmem_node(pgdat, size);
3857 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
3859 #ifndef CONFIG_NEED_MULTIPLE_NODES
3861 * With no DISCONTIG, the global mem_map is just set as node 0's
3863 if (pgdat == NODE_DATA(0)) {
3864 mem_map = NODE_DATA(0)->node_mem_map;
3865 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3866 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
3867 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
3868 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3871 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3874 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
3875 unsigned long node_start_pfn, unsigned long *zholes_size)
3877 pg_data_t *pgdat = NODE_DATA(nid);
3879 pgdat->node_id = nid;
3880 pgdat->node_start_pfn = node_start_pfn;
3881 calculate_node_totalpages(pgdat, zones_size, zholes_size);
3883 alloc_node_mem_map(pgdat);
3884 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3885 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3886 nid, (unsigned long)pgdat,
3887 (unsigned long)pgdat->node_mem_map);
3890 free_area_init_core(pgdat, zones_size, zholes_size);
3893 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3895 #if MAX_NUMNODES > 1
3897 * Figure out the number of possible node ids.
3899 static void __init setup_nr_node_ids(void)
3902 unsigned int highest = 0;
3904 for_each_node_mask(node, node_possible_map)
3906 nr_node_ids = highest + 1;
3909 static inline void setup_nr_node_ids(void)
3915 * add_active_range - Register a range of PFNs backed by physical memory
3916 * @nid: The node ID the range resides on
3917 * @start_pfn: The start PFN of the available physical memory
3918 * @end_pfn: The end PFN of the available physical memory
3920 * These ranges are stored in an early_node_map[] and later used by
3921 * free_area_init_nodes() to calculate zone sizes and holes. If the
3922 * range spans a memory hole, it is up to the architecture to ensure
3923 * the memory is not freed by the bootmem allocator. If possible
3924 * the range being registered will be merged with existing ranges.
3926 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
3927 unsigned long end_pfn)
3931 mminit_dprintk(MMINIT_TRACE, "memory_register",
3932 "Entering add_active_range(%d, %#lx, %#lx) "
3933 "%d entries of %d used\n",
3934 nid, start_pfn, end_pfn,
3935 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
3937 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
3939 /* Merge with existing active regions if possible */
3940 for (i = 0; i < nr_nodemap_entries; i++) {
3941 if (early_node_map[i].nid != nid)
3944 /* Skip if an existing region covers this new one */
3945 if (start_pfn >= early_node_map[i].start_pfn &&
3946 end_pfn <= early_node_map[i].end_pfn)
3949 /* Merge forward if suitable */
3950 if (start_pfn <= early_node_map[i].end_pfn &&
3951 end_pfn > early_node_map[i].end_pfn) {
3952 early_node_map[i].end_pfn = end_pfn;
3956 /* Merge backward if suitable */
3957 if (start_pfn < early_node_map[i].end_pfn &&
3958 end_pfn >= early_node_map[i].start_pfn) {
3959 early_node_map[i].start_pfn = start_pfn;
3964 /* Check that early_node_map is large enough */
3965 if (i >= MAX_ACTIVE_REGIONS) {
3966 printk(KERN_CRIT "More than %d memory regions, truncating\n",
3967 MAX_ACTIVE_REGIONS);
3971 early_node_map[i].nid = nid;
3972 early_node_map[i].start_pfn = start_pfn;
3973 early_node_map[i].end_pfn = end_pfn;
3974 nr_nodemap_entries = i + 1;
3978 * remove_active_range - Shrink an existing registered range of PFNs
3979 * @nid: The node id the range is on that should be shrunk
3980 * @start_pfn: The new PFN of the range
3981 * @end_pfn: The new PFN of the range
3983 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
3984 * The map is kept near the end physical page range that has already been
3985 * registered. This function allows an arch to shrink an existing registered
3988 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
3989 unsigned long end_pfn)
3994 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
3995 nid, start_pfn, end_pfn);
3997 /* Find the old active region end and shrink */
3998 for_each_active_range_index_in_nid(i, nid) {
3999 if (early_node_map[i].start_pfn >= start_pfn &&
4000 early_node_map[i].end_pfn <= end_pfn) {
4002 early_node_map[i].start_pfn = 0;
4003 early_node_map[i].end_pfn = 0;
4007 if (early_node_map[i].start_pfn < start_pfn &&
4008 early_node_map[i].end_pfn > start_pfn) {
4009 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4010 early_node_map[i].end_pfn = start_pfn;
4011 if (temp_end_pfn > end_pfn)
4012 add_active_range(nid, end_pfn, temp_end_pfn);
4015 if (early_node_map[i].start_pfn >= start_pfn &&
4016 early_node_map[i].end_pfn > end_pfn &&
4017 early_node_map[i].start_pfn < end_pfn) {
4018 early_node_map[i].start_pfn = end_pfn;
4026 /* remove the blank ones */
4027 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4028 if (early_node_map[i].nid != nid)
4030 if (early_node_map[i].end_pfn)
4032 /* we found it, get rid of it */
4033 for (j = i; j < nr_nodemap_entries - 1; j++)
4034 memcpy(&early_node_map[j], &early_node_map[j+1],
4035 sizeof(early_node_map[j]));
4036 j = nr_nodemap_entries - 1;
4037 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4038 nr_nodemap_entries--;
4043 * remove_all_active_ranges - Remove all currently registered regions
4045 * During discovery, it may be found that a table like SRAT is invalid
4046 * and an alternative discovery method must be used. This function removes
4047 * all currently registered regions.
4049 void __init remove_all_active_ranges(void)
4051 memset(early_node_map, 0, sizeof(early_node_map));
4052 nr_nodemap_entries = 0;
4055 /* Compare two active node_active_regions */
4056 static int __init cmp_node_active_region(const void *a, const void *b)
4058 struct node_active_region *arange = (struct node_active_region *)a;
4059 struct node_active_region *brange = (struct node_active_region *)b;
4061 /* Done this way to avoid overflows */
4062 if (arange->start_pfn > brange->start_pfn)
4064 if (arange->start_pfn < brange->start_pfn)
4070 /* sort the node_map by start_pfn */
4071 static void __init sort_node_map(void)
4073 sort(early_node_map, (size_t)nr_nodemap_entries,
4074 sizeof(struct node_active_region),
4075 cmp_node_active_region, NULL);
4078 /* Find the lowest pfn for a node */
4079 static unsigned long __init find_min_pfn_for_node(int nid)
4082 unsigned long min_pfn = ULONG_MAX;
4084 /* Assuming a sorted map, the first range found has the starting pfn */
4085 for_each_active_range_index_in_nid(i, nid)
4086 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4088 if (min_pfn == ULONG_MAX) {
4090 "Could not find start_pfn for node %d\n", nid);
4098 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4100 * It returns the minimum PFN based on information provided via
4101 * add_active_range().
4103 unsigned long __init find_min_pfn_with_active_regions(void)
4105 return find_min_pfn_for_node(MAX_NUMNODES);
4109 * early_calculate_totalpages()
4110 * Sum pages in active regions for movable zone.
4111 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4113 static unsigned long __init early_calculate_totalpages(void)
4116 unsigned long totalpages = 0;
4118 for (i = 0; i < nr_nodemap_entries; i++) {
4119 unsigned long pages = early_node_map[i].end_pfn -
4120 early_node_map[i].start_pfn;
4121 totalpages += pages;
4123 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4129 * Find the PFN the Movable zone begins in each node. Kernel memory
4130 * is spread evenly between nodes as long as the nodes have enough
4131 * memory. When they don't, some nodes will have more kernelcore than
4134 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4137 unsigned long usable_startpfn;
4138 unsigned long kernelcore_node, kernelcore_remaining;
4139 /* save the state before borrow the nodemask */
4140 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4141 unsigned long totalpages = early_calculate_totalpages();
4142 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4145 * If movablecore was specified, calculate what size of
4146 * kernelcore that corresponds so that memory usable for
4147 * any allocation type is evenly spread. If both kernelcore
4148 * and movablecore are specified, then the value of kernelcore
4149 * will be used for required_kernelcore if it's greater than
4150 * what movablecore would have allowed.
4152 if (required_movablecore) {
4153 unsigned long corepages;
4156 * Round-up so that ZONE_MOVABLE is at least as large as what
4157 * was requested by the user
4159 required_movablecore =
4160 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4161 corepages = totalpages - required_movablecore;
4163 required_kernelcore = max(required_kernelcore, corepages);
4166 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4167 if (!required_kernelcore)
4170 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4171 find_usable_zone_for_movable();
4172 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4175 /* Spread kernelcore memory as evenly as possible throughout nodes */
4176 kernelcore_node = required_kernelcore / usable_nodes;
4177 for_each_node_state(nid, N_HIGH_MEMORY) {
4179 * Recalculate kernelcore_node if the division per node
4180 * now exceeds what is necessary to satisfy the requested
4181 * amount of memory for the kernel
4183 if (required_kernelcore < kernelcore_node)
4184 kernelcore_node = required_kernelcore / usable_nodes;
4187 * As the map is walked, we track how much memory is usable
4188 * by the kernel using kernelcore_remaining. When it is
4189 * 0, the rest of the node is usable by ZONE_MOVABLE
4191 kernelcore_remaining = kernelcore_node;
4193 /* Go through each range of PFNs within this node */
4194 for_each_active_range_index_in_nid(i, nid) {
4195 unsigned long start_pfn, end_pfn;
4196 unsigned long size_pages;
4198 start_pfn = max(early_node_map[i].start_pfn,
4199 zone_movable_pfn[nid]);
4200 end_pfn = early_node_map[i].end_pfn;
4201 if (start_pfn >= end_pfn)
4204 /* Account for what is only usable for kernelcore */
4205 if (start_pfn < usable_startpfn) {
4206 unsigned long kernel_pages;
4207 kernel_pages = min(end_pfn, usable_startpfn)
4210 kernelcore_remaining -= min(kernel_pages,
4211 kernelcore_remaining);
4212 required_kernelcore -= min(kernel_pages,
4213 required_kernelcore);
4215 /* Continue if range is now fully accounted */
4216 if (end_pfn <= usable_startpfn) {
4219 * Push zone_movable_pfn to the end so
4220 * that if we have to rebalance
4221 * kernelcore across nodes, we will
4222 * not double account here
4224 zone_movable_pfn[nid] = end_pfn;
4227 start_pfn = usable_startpfn;
4231 * The usable PFN range for ZONE_MOVABLE is from
4232 * start_pfn->end_pfn. Calculate size_pages as the
4233 * number of pages used as kernelcore
4235 size_pages = end_pfn - start_pfn;
4236 if (size_pages > kernelcore_remaining)
4237 size_pages = kernelcore_remaining;
4238 zone_movable_pfn[nid] = start_pfn + size_pages;
4241 * Some kernelcore has been met, update counts and
4242 * break if the kernelcore for this node has been
4245 required_kernelcore -= min(required_kernelcore,
4247 kernelcore_remaining -= size_pages;
4248 if (!kernelcore_remaining)
4254 * If there is still required_kernelcore, we do another pass with one
4255 * less node in the count. This will push zone_movable_pfn[nid] further
4256 * along on the nodes that still have memory until kernelcore is
4260 if (usable_nodes && required_kernelcore > usable_nodes)
4263 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4264 for (nid = 0; nid < MAX_NUMNODES; nid++)
4265 zone_movable_pfn[nid] =
4266 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4269 /* restore the node_state */
4270 node_states[N_HIGH_MEMORY] = saved_node_state;
4273 /* Any regular memory on that node ? */
4274 static void check_for_regular_memory(pg_data_t *pgdat)
4276 #ifdef CONFIG_HIGHMEM
4277 enum zone_type zone_type;
4279 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4280 struct zone *zone = &pgdat->node_zones[zone_type];
4281 if (zone->present_pages)
4282 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4288 * free_area_init_nodes - Initialise all pg_data_t and zone data
4289 * @max_zone_pfn: an array of max PFNs for each zone
4291 * This will call free_area_init_node() for each active node in the system.
4292 * Using the page ranges provided by add_active_range(), the size of each
4293 * zone in each node and their holes is calculated. If the maximum PFN
4294 * between two adjacent zones match, it is assumed that the zone is empty.
4295 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4296 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4297 * starts where the previous one ended. For example, ZONE_DMA32 starts
4298 * at arch_max_dma_pfn.
4300 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4305 /* Sort early_node_map as initialisation assumes it is sorted */
4308 /* Record where the zone boundaries are */
4309 memset(arch_zone_lowest_possible_pfn, 0,
4310 sizeof(arch_zone_lowest_possible_pfn));
4311 memset(arch_zone_highest_possible_pfn, 0,
4312 sizeof(arch_zone_highest_possible_pfn));
4313 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4314 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4315 for (i = 1; i < MAX_NR_ZONES; i++) {
4316 if (i == ZONE_MOVABLE)
4318 arch_zone_lowest_possible_pfn[i] =
4319 arch_zone_highest_possible_pfn[i-1];
4320 arch_zone_highest_possible_pfn[i] =
4321 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4323 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4324 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4326 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4327 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4328 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4330 /* Print out the zone ranges */
4331 printk("Zone PFN ranges:\n");
4332 for (i = 0; i < MAX_NR_ZONES; i++) {
4333 if (i == ZONE_MOVABLE)
4335 printk(" %-8s %0#10lx -> %0#10lx\n",
4337 arch_zone_lowest_possible_pfn[i],
4338 arch_zone_highest_possible_pfn[i]);
4341 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4342 printk("Movable zone start PFN for each node\n");
4343 for (i = 0; i < MAX_NUMNODES; i++) {
4344 if (zone_movable_pfn[i])
4345 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4348 /* Print out the early_node_map[] */
4349 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4350 for (i = 0; i < nr_nodemap_entries; i++)
4351 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4352 early_node_map[i].start_pfn,
4353 early_node_map[i].end_pfn);
4355 /* Initialise every node */
4356 mminit_verify_pageflags_layout();
4357 setup_nr_node_ids();
4358 for_each_online_node(nid) {
4359 pg_data_t *pgdat = NODE_DATA(nid);
4360 free_area_init_node(nid, NULL,
4361 find_min_pfn_for_node(nid), NULL);
4363 /* Any memory on that node */
4364 if (pgdat->node_present_pages)
4365 node_set_state(nid, N_HIGH_MEMORY);
4366 check_for_regular_memory(pgdat);
4370 static int __init cmdline_parse_core(char *p, unsigned long *core)
4372 unsigned long long coremem;
4376 coremem = memparse(p, &p);
4377 *core = coremem >> PAGE_SHIFT;
4379 /* Paranoid check that UL is enough for the coremem value */
4380 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4386 * kernelcore=size sets the amount of memory for use for allocations that
4387 * cannot be reclaimed or migrated.
4389 static int __init cmdline_parse_kernelcore(char *p)
4391 return cmdline_parse_core(p, &required_kernelcore);
4395 * movablecore=size sets the amount of memory for use for allocations that
4396 * can be reclaimed or migrated.
4398 static int __init cmdline_parse_movablecore(char *p)
4400 return cmdline_parse_core(p, &required_movablecore);
4403 early_param("kernelcore", cmdline_parse_kernelcore);
4404 early_param("movablecore", cmdline_parse_movablecore);
4406 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4409 * set_dma_reserve - set the specified number of pages reserved in the first zone
4410 * @new_dma_reserve: The number of pages to mark reserved
4412 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4413 * In the DMA zone, a significant percentage may be consumed by kernel image
4414 * and other unfreeable allocations which can skew the watermarks badly. This
4415 * function may optionally be used to account for unfreeable pages in the
4416 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4417 * smaller per-cpu batchsize.
4419 void __init set_dma_reserve(unsigned long new_dma_reserve)
4421 dma_reserve = new_dma_reserve;
4424 #ifndef CONFIG_NEED_MULTIPLE_NODES
4425 struct pglist_data __refdata contig_page_data = { .bdata = &bootmem_node_data[0] };
4426 EXPORT_SYMBOL(contig_page_data);
4429 void __init free_area_init(unsigned long *zones_size)
4431 free_area_init_node(0, zones_size,
4432 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4435 static int page_alloc_cpu_notify(struct notifier_block *self,
4436 unsigned long action, void *hcpu)
4438 int cpu = (unsigned long)hcpu;
4440 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4444 * Spill the event counters of the dead processor
4445 * into the current processors event counters.
4446 * This artificially elevates the count of the current
4449 vm_events_fold_cpu(cpu);
4452 * Zero the differential counters of the dead processor
4453 * so that the vm statistics are consistent.
4455 * This is only okay since the processor is dead and cannot
4456 * race with what we are doing.
4458 refresh_cpu_vm_stats(cpu);
4463 void __init page_alloc_init(void)
4465 hotcpu_notifier(page_alloc_cpu_notify, 0);
4469 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4470 * or min_free_kbytes changes.
4472 static void calculate_totalreserve_pages(void)
4474 struct pglist_data *pgdat;
4475 unsigned long reserve_pages = 0;
4476 enum zone_type i, j;
4478 for_each_online_pgdat(pgdat) {
4479 for (i = 0; i < MAX_NR_ZONES; i++) {
4480 struct zone *zone = pgdat->node_zones + i;
4481 unsigned long max = 0;
4483 /* Find valid and maximum lowmem_reserve in the zone */
4484 for (j = i; j < MAX_NR_ZONES; j++) {
4485 if (zone->lowmem_reserve[j] > max)
4486 max = zone->lowmem_reserve[j];
4489 /* we treat the high watermark as reserved pages. */
4490 max += high_wmark_pages(zone);
4492 if (max > zone->present_pages)
4493 max = zone->present_pages;
4494 reserve_pages += max;
4497 totalreserve_pages = reserve_pages;
4501 * setup_per_zone_lowmem_reserve - called whenever
4502 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4503 * has a correct pages reserved value, so an adequate number of
4504 * pages are left in the zone after a successful __alloc_pages().
4506 static void setup_per_zone_lowmem_reserve(void)
4508 struct pglist_data *pgdat;
4509 enum zone_type j, idx;
4511 for_each_online_pgdat(pgdat) {
4512 for (j = 0; j < MAX_NR_ZONES; j++) {
4513 struct zone *zone = pgdat->node_zones + j;
4514 unsigned long present_pages = zone->present_pages;
4516 zone->lowmem_reserve[j] = 0;
4520 struct zone *lower_zone;
4524 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4525 sysctl_lowmem_reserve_ratio[idx] = 1;
4527 lower_zone = pgdat->node_zones + idx;
4528 lower_zone->lowmem_reserve[j] = present_pages /
4529 sysctl_lowmem_reserve_ratio[idx];
4530 present_pages += lower_zone->present_pages;
4535 /* update totalreserve_pages */
4536 calculate_totalreserve_pages();
4540 * setup_per_zone_wmarks - called when min_free_kbytes changes
4541 * or when memory is hot-{added|removed}
4543 * Ensures that the watermark[min,low,high] values for each zone are set
4544 * correctly with respect to min_free_kbytes.
4546 void setup_per_zone_wmarks(void)
4548 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4549 unsigned long lowmem_pages = 0;
4551 unsigned long flags;
4553 /* Calculate total number of !ZONE_HIGHMEM pages */
4554 for_each_zone(zone) {
4555 if (!is_highmem(zone))
4556 lowmem_pages += zone->present_pages;
4559 for_each_zone(zone) {
4562 spin_lock_irqsave(&zone->lock, flags);
4563 tmp = (u64)pages_min * zone->present_pages;
4564 do_div(tmp, lowmem_pages);
4565 if (is_highmem(zone)) {
4567 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4568 * need highmem pages, so cap pages_min to a small
4571 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4572 * deltas controls asynch page reclaim, and so should
4573 * not be capped for highmem.
4577 min_pages = zone->present_pages / 1024;
4578 if (min_pages < SWAP_CLUSTER_MAX)
4579 min_pages = SWAP_CLUSTER_MAX;
4580 if (min_pages > 128)
4582 zone->watermark[WMARK_MIN] = min_pages;
4585 * If it's a lowmem zone, reserve a number of pages
4586 * proportionate to the zone's size.
4588 zone->watermark[WMARK_MIN] = tmp;
4591 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
4592 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
4593 setup_zone_migrate_reserve(zone);
4594 spin_unlock_irqrestore(&zone->lock, flags);
4597 /* update totalreserve_pages */
4598 calculate_totalreserve_pages();
4602 * The inactive anon list should be small enough that the VM never has to
4603 * do too much work, but large enough that each inactive page has a chance
4604 * to be referenced again before it is swapped out.
4606 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4607 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4608 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4609 * the anonymous pages are kept on the inactive list.
4612 * memory ratio inactive anon
4613 * -------------------------------------
4622 void calculate_zone_inactive_ratio(struct zone *zone)
4624 unsigned int gb, ratio;
4626 /* Zone size in gigabytes */
4627 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4629 ratio = int_sqrt(10 * gb);
4633 zone->inactive_ratio = ratio;
4636 static void __init setup_per_zone_inactive_ratio(void)
4641 calculate_zone_inactive_ratio(zone);
4645 * Initialise min_free_kbytes.
4647 * For small machines we want it small (128k min). For large machines
4648 * we want it large (64MB max). But it is not linear, because network
4649 * bandwidth does not increase linearly with machine size. We use
4651 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4652 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4668 static int __init init_per_zone_wmark_min(void)
4670 unsigned long lowmem_kbytes;
4672 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4674 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4675 if (min_free_kbytes < 128)
4676 min_free_kbytes = 128;
4677 if (min_free_kbytes > 65536)
4678 min_free_kbytes = 65536;
4679 setup_per_zone_wmarks();
4680 setup_per_zone_lowmem_reserve();
4681 setup_per_zone_inactive_ratio();
4684 module_init(init_per_zone_wmark_min)
4687 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4688 * that we can call two helper functions whenever min_free_kbytes
4691 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
4692 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4694 proc_dointvec(table, write, file, buffer, length, ppos);
4696 setup_per_zone_wmarks();
4701 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
4702 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4707 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4712 zone->min_unmapped_pages = (zone->present_pages *
4713 sysctl_min_unmapped_ratio) / 100;
4717 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
4718 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4723 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4728 zone->min_slab_pages = (zone->present_pages *
4729 sysctl_min_slab_ratio) / 100;
4735 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4736 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4737 * whenever sysctl_lowmem_reserve_ratio changes.
4739 * The reserve ratio obviously has absolutely no relation with the
4740 * minimum watermarks. The lowmem reserve ratio can only make sense
4741 * if in function of the boot time zone sizes.
4743 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
4744 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4746 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4747 setup_per_zone_lowmem_reserve();
4752 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4753 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4754 * can have before it gets flushed back to buddy allocator.
4757 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
4758 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4764 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4765 if (!write || (ret == -EINVAL))
4767 for_each_populated_zone(zone) {
4768 for_each_online_cpu(cpu) {
4770 high = zone->present_pages / percpu_pagelist_fraction;
4771 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
4777 int hashdist = HASHDIST_DEFAULT;
4780 static int __init set_hashdist(char *str)
4784 hashdist = simple_strtoul(str, &str, 0);
4787 __setup("hashdist=", set_hashdist);
4791 * allocate a large system hash table from bootmem
4792 * - it is assumed that the hash table must contain an exact power-of-2
4793 * quantity of entries
4794 * - limit is the number of hash buckets, not the total allocation size
4796 void *__init alloc_large_system_hash(const char *tablename,
4797 unsigned long bucketsize,
4798 unsigned long numentries,
4801 unsigned int *_hash_shift,
4802 unsigned int *_hash_mask,
4803 unsigned long limit)
4805 unsigned long long max = limit;
4806 unsigned long log2qty, size;
4809 /* allow the kernel cmdline to have a say */
4811 /* round applicable memory size up to nearest megabyte */
4812 numentries = nr_kernel_pages;
4813 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
4814 numentries >>= 20 - PAGE_SHIFT;
4815 numentries <<= 20 - PAGE_SHIFT;
4817 /* limit to 1 bucket per 2^scale bytes of low memory */
4818 if (scale > PAGE_SHIFT)
4819 numentries >>= (scale - PAGE_SHIFT);
4821 numentries <<= (PAGE_SHIFT - scale);
4823 /* Make sure we've got at least a 0-order allocation.. */
4824 if (unlikely(flags & HASH_SMALL)) {
4825 /* Makes no sense without HASH_EARLY */
4826 WARN_ON(!(flags & HASH_EARLY));
4827 if (!(numentries >> *_hash_shift)) {
4828 numentries = 1UL << *_hash_shift;
4829 BUG_ON(!numentries);
4831 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
4832 numentries = PAGE_SIZE / bucketsize;
4834 numentries = roundup_pow_of_two(numentries);
4836 /* limit allocation size to 1/16 total memory by default */
4838 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
4839 do_div(max, bucketsize);
4842 if (numentries > max)
4845 log2qty = ilog2(numentries);
4848 size = bucketsize << log2qty;
4849 if (flags & HASH_EARLY)
4850 table = alloc_bootmem_nopanic(size);
4852 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
4855 * If bucketsize is not a power-of-two, we may free
4856 * some pages at the end of hash table which
4857 * alloc_pages_exact() automatically does
4859 if (get_order(size) < MAX_ORDER) {
4860 table = alloc_pages_exact(size, GFP_ATOMIC);
4861 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
4864 } while (!table && size > PAGE_SIZE && --log2qty);
4867 panic("Failed to allocate %s hash table\n", tablename);
4869 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
4872 ilog2(size) - PAGE_SHIFT,
4876 *_hash_shift = log2qty;
4878 *_hash_mask = (1 << log2qty) - 1;
4883 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4884 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
4887 #ifdef CONFIG_SPARSEMEM
4888 return __pfn_to_section(pfn)->pageblock_flags;
4890 return zone->pageblock_flags;
4891 #endif /* CONFIG_SPARSEMEM */
4894 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
4896 #ifdef CONFIG_SPARSEMEM
4897 pfn &= (PAGES_PER_SECTION-1);
4898 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4900 pfn = pfn - zone->zone_start_pfn;
4901 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4902 #endif /* CONFIG_SPARSEMEM */
4906 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4907 * @page: The page within the block of interest
4908 * @start_bitidx: The first bit of interest to retrieve
4909 * @end_bitidx: The last bit of interest
4910 * returns pageblock_bits flags
4912 unsigned long get_pageblock_flags_group(struct page *page,
4913 int start_bitidx, int end_bitidx)
4916 unsigned long *bitmap;
4917 unsigned long pfn, bitidx;
4918 unsigned long flags = 0;
4919 unsigned long value = 1;
4921 zone = page_zone(page);
4922 pfn = page_to_pfn(page);
4923 bitmap = get_pageblock_bitmap(zone, pfn);
4924 bitidx = pfn_to_bitidx(zone, pfn);
4926 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4927 if (test_bit(bitidx + start_bitidx, bitmap))
4934 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4935 * @page: The page within the block of interest
4936 * @start_bitidx: The first bit of interest
4937 * @end_bitidx: The last bit of interest
4938 * @flags: The flags to set
4940 void set_pageblock_flags_group(struct page *page, unsigned long flags,
4941 int start_bitidx, int end_bitidx)
4944 unsigned long *bitmap;
4945 unsigned long pfn, bitidx;
4946 unsigned long value = 1;
4948 zone = page_zone(page);
4949 pfn = page_to_pfn(page);
4950 bitmap = get_pageblock_bitmap(zone, pfn);
4951 bitidx = pfn_to_bitidx(zone, pfn);
4952 VM_BUG_ON(pfn < zone->zone_start_pfn);
4953 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
4955 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4957 __set_bit(bitidx + start_bitidx, bitmap);
4959 __clear_bit(bitidx + start_bitidx, bitmap);
4963 * This is designed as sub function...plz see page_isolation.c also.
4964 * set/clear page block's type to be ISOLATE.
4965 * page allocater never alloc memory from ISOLATE block.
4968 int set_migratetype_isolate(struct page *page)
4971 unsigned long flags;
4975 zone = page_zone(page);
4976 zone_idx = zone_idx(zone);
4977 spin_lock_irqsave(&zone->lock, flags);
4979 * In future, more migrate types will be able to be isolation target.
4981 if (get_pageblock_migratetype(page) != MIGRATE_MOVABLE &&
4982 zone_idx != ZONE_MOVABLE)
4984 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
4985 move_freepages_block(zone, page, MIGRATE_ISOLATE);
4988 spin_unlock_irqrestore(&zone->lock, flags);
4994 void unset_migratetype_isolate(struct page *page)
4997 unsigned long flags;
4998 zone = page_zone(page);
4999 spin_lock_irqsave(&zone->lock, flags);
5000 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5002 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5003 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5005 spin_unlock_irqrestore(&zone->lock, flags);
5008 #ifdef CONFIG_MEMORY_HOTREMOVE
5010 * All pages in the range must be isolated before calling this.
5013 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5019 unsigned long flags;
5020 /* find the first valid pfn */
5021 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5026 zone = page_zone(pfn_to_page(pfn));
5027 spin_lock_irqsave(&zone->lock, flags);
5029 while (pfn < end_pfn) {
5030 if (!pfn_valid(pfn)) {
5034 page = pfn_to_page(pfn);
5035 BUG_ON(page_count(page));
5036 BUG_ON(!PageBuddy(page));
5037 order = page_order(page);
5038 #ifdef CONFIG_DEBUG_VM
5039 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5040 pfn, 1 << order, end_pfn);
5042 list_del(&page->lru);
5043 rmv_page_order(page);
5044 zone->free_area[order].nr_free--;
5045 __mod_zone_page_state(zone, NR_FREE_PAGES,
5047 for (i = 0; i < (1 << order); i++)
5048 SetPageReserved((page+i));
5049 pfn += (1 << order);
5051 spin_unlock_irqrestore(&zone->lock, flags);