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>
52 #include <asm/tlbflush.h>
53 #include <asm/div64.h>
57 * Array of node states.
59 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
60 [N_POSSIBLE] = NODE_MASK_ALL,
61 [N_ONLINE] = { { [0] = 1UL } },
63 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
65 [N_HIGH_MEMORY] = { { [0] = 1UL } },
67 [N_CPU] = { { [0] = 1UL } },
70 EXPORT_SYMBOL(node_states);
72 unsigned long totalram_pages __read_mostly;
73 unsigned long totalreserve_pages __read_mostly;
74 unsigned long highest_memmap_pfn __read_mostly;
75 int percpu_pagelist_fraction;
76 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
78 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
79 int pageblock_order __read_mostly;
82 static void __free_pages_ok(struct page *page, unsigned int order);
85 * results with 256, 32 in the lowmem_reserve sysctl:
86 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
87 * 1G machine -> (16M dma, 784M normal, 224M high)
88 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
89 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
90 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
92 * TBD: should special case ZONE_DMA32 machines here - in those we normally
93 * don't need any ZONE_NORMAL reservation
95 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
96 #ifdef CONFIG_ZONE_DMA
99 #ifdef CONFIG_ZONE_DMA32
102 #ifdef CONFIG_HIGHMEM
108 EXPORT_SYMBOL(totalram_pages);
110 static char * const zone_names[MAX_NR_ZONES] = {
111 #ifdef CONFIG_ZONE_DMA
114 #ifdef CONFIG_ZONE_DMA32
118 #ifdef CONFIG_HIGHMEM
124 int min_free_kbytes = 1024;
126 unsigned long __meminitdata nr_kernel_pages;
127 unsigned long __meminitdata nr_all_pages;
128 static unsigned long __meminitdata dma_reserve;
130 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
132 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
133 * ranges of memory (RAM) that may be registered with add_active_range().
134 * Ranges passed to add_active_range() will be merged if possible
135 * so the number of times add_active_range() can be called is
136 * related to the number of nodes and the number of holes
138 #ifdef CONFIG_MAX_ACTIVE_REGIONS
139 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
140 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
142 #if MAX_NUMNODES >= 32
143 /* If there can be many nodes, allow up to 50 holes per node */
144 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
146 /* By default, allow up to 256 distinct regions */
147 #define MAX_ACTIVE_REGIONS 256
151 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
152 static int __meminitdata nr_nodemap_entries;
153 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
154 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
155 static unsigned long __initdata required_kernelcore;
156 static unsigned long __initdata required_movablecore;
157 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
159 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
161 EXPORT_SYMBOL(movable_zone);
162 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
165 int nr_node_ids __read_mostly = MAX_NUMNODES;
166 int nr_online_nodes __read_mostly = 1;
167 EXPORT_SYMBOL(nr_node_ids);
168 EXPORT_SYMBOL(nr_online_nodes);
171 int page_group_by_mobility_disabled __read_mostly;
173 static void set_pageblock_migratetype(struct page *page, int migratetype)
176 if (unlikely(page_group_by_mobility_disabled))
177 migratetype = MIGRATE_UNMOVABLE;
179 set_pageblock_flags_group(page, (unsigned long)migratetype,
180 PB_migrate, PB_migrate_end);
183 bool oom_killer_disabled __read_mostly;
185 #ifdef CONFIG_DEBUG_VM
186 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
190 unsigned long pfn = page_to_pfn(page);
193 seq = zone_span_seqbegin(zone);
194 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
196 else if (pfn < zone->zone_start_pfn)
198 } while (zone_span_seqretry(zone, seq));
203 static int page_is_consistent(struct zone *zone, struct page *page)
205 if (!pfn_valid_within(page_to_pfn(page)))
207 if (zone != page_zone(page))
213 * Temporary debugging check for pages not lying within a given zone.
215 static int bad_range(struct zone *zone, struct page *page)
217 if (page_outside_zone_boundaries(zone, page))
219 if (!page_is_consistent(zone, page))
225 static inline int bad_range(struct zone *zone, struct page *page)
231 static void bad_page(struct page *page)
233 static unsigned long resume;
234 static unsigned long nr_shown;
235 static unsigned long nr_unshown;
238 * Allow a burst of 60 reports, then keep quiet for that minute;
239 * or allow a steady drip of one report per second.
241 if (nr_shown == 60) {
242 if (time_before(jiffies, resume)) {
248 "BUG: Bad page state: %lu messages suppressed\n",
255 resume = jiffies + 60 * HZ;
257 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
258 current->comm, page_to_pfn(page));
260 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
261 page, (void *)page->flags, page_count(page),
262 page_mapcount(page), page->mapping, page->index);
266 /* Leave bad fields for debug, except PageBuddy could make trouble */
267 __ClearPageBuddy(page);
268 add_taint(TAINT_BAD_PAGE);
272 * Higher-order pages are called "compound pages". They are structured thusly:
274 * The first PAGE_SIZE page is called the "head page".
276 * The remaining PAGE_SIZE pages are called "tail pages".
278 * All pages have PG_compound set. All pages have their ->private pointing at
279 * the head page (even the head page has this).
281 * The first tail page's ->lru.next holds the address of the compound page's
282 * put_page() function. Its ->lru.prev holds the order of allocation.
283 * This usage means that zero-order pages may not be compound.
286 static void free_compound_page(struct page *page)
288 __free_pages_ok(page, compound_order(page));
291 void prep_compound_page(struct page *page, unsigned long order)
294 int nr_pages = 1 << order;
296 set_compound_page_dtor(page, free_compound_page);
297 set_compound_order(page, order);
299 for (i = 1; i < nr_pages; i++) {
300 struct page *p = page + i;
303 p->first_page = page;
307 static int destroy_compound_page(struct page *page, unsigned long order)
310 int nr_pages = 1 << order;
313 if (unlikely(compound_order(page) != order) ||
314 unlikely(!PageHead(page))) {
319 __ClearPageHead(page);
321 for (i = 1; i < nr_pages; i++) {
322 struct page *p = page + i;
324 if (unlikely(!PageTail(p) || (p->first_page != page))) {
334 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
339 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
340 * and __GFP_HIGHMEM from hard or soft interrupt context.
342 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
343 for (i = 0; i < (1 << order); i++)
344 clear_highpage(page + i);
347 static inline void set_page_order(struct page *page, int order)
349 set_page_private(page, order);
350 __SetPageBuddy(page);
353 static inline void rmv_page_order(struct page *page)
355 __ClearPageBuddy(page);
356 set_page_private(page, 0);
360 * Locate the struct page for both the matching buddy in our
361 * pair (buddy1) and the combined O(n+1) page they form (page).
363 * 1) Any buddy B1 will have an order O twin B2 which satisfies
364 * the following equation:
366 * For example, if the starting buddy (buddy2) is #8 its order
368 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
370 * 2) Any buddy B will have an order O+1 parent P which
371 * satisfies the following equation:
374 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
376 static inline struct page *
377 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
379 unsigned long buddy_idx = page_idx ^ (1 << order);
381 return page + (buddy_idx - page_idx);
384 static inline unsigned long
385 __find_combined_index(unsigned long page_idx, unsigned int order)
387 return (page_idx & ~(1 << order));
391 * This function checks whether a page is free && is the buddy
392 * we can do coalesce a page and its buddy if
393 * (a) the buddy is not in a hole &&
394 * (b) the buddy is in the buddy system &&
395 * (c) a page and its buddy have the same order &&
396 * (d) a page and its buddy are in the same zone.
398 * For recording whether a page is in the buddy system, we use PG_buddy.
399 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
401 * For recording page's order, we use page_private(page).
403 static inline int page_is_buddy(struct page *page, struct page *buddy,
406 if (!pfn_valid_within(page_to_pfn(buddy)))
409 if (page_zone_id(page) != page_zone_id(buddy))
412 if (PageBuddy(buddy) && page_order(buddy) == order) {
413 VM_BUG_ON(page_count(buddy) != 0);
420 * Freeing function for a buddy system allocator.
422 * The concept of a buddy system is to maintain direct-mapped table
423 * (containing bit values) for memory blocks of various "orders".
424 * The bottom level table contains the map for the smallest allocatable
425 * units of memory (here, pages), and each level above it describes
426 * pairs of units from the levels below, hence, "buddies".
427 * At a high level, all that happens here is marking the table entry
428 * at the bottom level available, and propagating the changes upward
429 * as necessary, plus some accounting needed to play nicely with other
430 * parts of the VM system.
431 * At each level, we keep a list of pages, which are heads of continuous
432 * free pages of length of (1 << order) and marked with PG_buddy. Page's
433 * order is recorded in page_private(page) field.
434 * So when we are allocating or freeing one, we can derive the state of the
435 * other. That is, if we allocate a small block, and both were
436 * free, the remainder of the region must be split into blocks.
437 * If a block is freed, and its buddy is also free, then this
438 * triggers coalescing into a block of larger size.
443 static inline void __free_one_page(struct page *page,
444 struct zone *zone, unsigned int order,
447 unsigned long page_idx;
449 if (unlikely(PageCompound(page)))
450 if (unlikely(destroy_compound_page(page, order)))
453 VM_BUG_ON(migratetype == -1);
455 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
457 VM_BUG_ON(page_idx & ((1 << order) - 1));
458 VM_BUG_ON(bad_range(zone, page));
460 while (order < MAX_ORDER-1) {
461 unsigned long combined_idx;
464 buddy = __page_find_buddy(page, page_idx, order);
465 if (!page_is_buddy(page, buddy, order))
468 /* Our buddy is free, merge with it and move up one order. */
469 list_del(&buddy->lru);
470 zone->free_area[order].nr_free--;
471 rmv_page_order(buddy);
472 combined_idx = __find_combined_index(page_idx, order);
473 page = page + (combined_idx - page_idx);
474 page_idx = combined_idx;
477 set_page_order(page, order);
479 &zone->free_area[order].free_list[migratetype]);
480 zone->free_area[order].nr_free++;
483 #ifdef CONFIG_HAVE_MLOCKED_PAGE_BIT
485 * free_page_mlock() -- clean up attempts to free and mlocked() page.
486 * Page should not be on lru, so no need to fix that up.
487 * free_pages_check() will verify...
489 static inline void free_page_mlock(struct page *page)
491 __dec_zone_page_state(page, NR_MLOCK);
492 __count_vm_event(UNEVICTABLE_MLOCKFREED);
495 static void free_page_mlock(struct page *page) { }
498 static inline int free_pages_check(struct page *page)
500 if (unlikely(page_mapcount(page) |
501 (page->mapping != NULL) |
502 (atomic_read(&page->_count) != 0) |
503 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) {
507 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
508 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
513 * Frees a list of pages.
514 * Assumes all pages on list are in same zone, and of same order.
515 * count is the number of pages to free.
517 * If the zone was previously in an "all pages pinned" state then look to
518 * see if this freeing clears that state.
520 * And clear the zone's pages_scanned counter, to hold off the "all pages are
521 * pinned" detection logic.
523 static void free_pages_bulk(struct zone *zone, int count,
524 struct list_head *list, int order)
526 spin_lock(&zone->lock);
527 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
528 zone->pages_scanned = 0;
530 __mod_zone_page_state(zone, NR_FREE_PAGES, count << order);
534 VM_BUG_ON(list_empty(list));
535 page = list_entry(list->prev, struct page, lru);
536 /* have to delete it as __free_one_page list manipulates */
537 list_del(&page->lru);
538 __free_one_page(page, zone, order, page_private(page));
540 spin_unlock(&zone->lock);
543 static void free_one_page(struct zone *zone, struct page *page, int order,
546 spin_lock(&zone->lock);
547 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE);
548 zone->pages_scanned = 0;
550 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
551 __free_one_page(page, zone, order, migratetype);
552 spin_unlock(&zone->lock);
555 static void __free_pages_ok(struct page *page, unsigned int order)
560 int wasMlocked = __TestClearPageMlocked(page);
562 kmemcheck_free_shadow(page, order);
564 for (i = 0 ; i < (1 << order) ; ++i)
565 bad += free_pages_check(page + i);
569 if (!PageHighMem(page)) {
570 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
571 debug_check_no_obj_freed(page_address(page),
574 arch_free_page(page, order);
575 kernel_map_pages(page, 1 << order, 0);
577 local_irq_save(flags);
578 if (unlikely(wasMlocked))
579 free_page_mlock(page);
580 __count_vm_events(PGFREE, 1 << order);
581 free_one_page(page_zone(page), page, order,
582 get_pageblock_migratetype(page));
583 local_irq_restore(flags);
587 * permit the bootmem allocator to evade page validation on high-order frees
589 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
592 __ClearPageReserved(page);
593 set_page_count(page, 0);
594 set_page_refcounted(page);
600 for (loop = 0; loop < BITS_PER_LONG; loop++) {
601 struct page *p = &page[loop];
603 if (loop + 1 < BITS_PER_LONG)
605 __ClearPageReserved(p);
606 set_page_count(p, 0);
609 set_page_refcounted(page);
610 __free_pages(page, order);
616 * The order of subdivision here is critical for the IO subsystem.
617 * Please do not alter this order without good reasons and regression
618 * testing. Specifically, as large blocks of memory are subdivided,
619 * the order in which smaller blocks are delivered depends on the order
620 * they're subdivided in this function. This is the primary factor
621 * influencing the order in which pages are delivered to the IO
622 * subsystem according to empirical testing, and this is also justified
623 * by considering the behavior of a buddy system containing a single
624 * large block of memory acted on by a series of small allocations.
625 * This behavior is a critical factor in sglist merging's success.
629 static inline void expand(struct zone *zone, struct page *page,
630 int low, int high, struct free_area *area,
633 unsigned long size = 1 << high;
639 VM_BUG_ON(bad_range(zone, &page[size]));
640 list_add(&page[size].lru, &area->free_list[migratetype]);
642 set_page_order(&page[size], high);
647 * This page is about to be returned from the page allocator
649 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
651 if (unlikely(page_mapcount(page) |
652 (page->mapping != NULL) |
653 (atomic_read(&page->_count) != 0) |
654 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) {
659 set_page_private(page, 0);
660 set_page_refcounted(page);
662 arch_alloc_page(page, order);
663 kernel_map_pages(page, 1 << order, 1);
665 if (gfp_flags & __GFP_ZERO)
666 prep_zero_page(page, order, gfp_flags);
668 if (order && (gfp_flags & __GFP_COMP))
669 prep_compound_page(page, order);
675 * Go through the free lists for the given migratetype and remove
676 * the smallest available page from the freelists
679 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
682 unsigned int current_order;
683 struct free_area * area;
686 /* Find a page of the appropriate size in the preferred list */
687 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
688 area = &(zone->free_area[current_order]);
689 if (list_empty(&area->free_list[migratetype]))
692 page = list_entry(area->free_list[migratetype].next,
694 list_del(&page->lru);
695 rmv_page_order(page);
697 expand(zone, page, order, current_order, area, migratetype);
706 * This array describes the order lists are fallen back to when
707 * the free lists for the desirable migrate type are depleted
709 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
710 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
711 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
712 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
713 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
717 * Move the free pages in a range to the free lists of the requested type.
718 * Note that start_page and end_pages are not aligned on a pageblock
719 * boundary. If alignment is required, use move_freepages_block()
721 static int move_freepages(struct zone *zone,
722 struct page *start_page, struct page *end_page,
729 #ifndef CONFIG_HOLES_IN_ZONE
731 * page_zone is not safe to call in this context when
732 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
733 * anyway as we check zone boundaries in move_freepages_block().
734 * Remove at a later date when no bug reports exist related to
735 * grouping pages by mobility
737 BUG_ON(page_zone(start_page) != page_zone(end_page));
740 for (page = start_page; page <= end_page;) {
741 /* Make sure we are not inadvertently changing nodes */
742 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
744 if (!pfn_valid_within(page_to_pfn(page))) {
749 if (!PageBuddy(page)) {
754 order = page_order(page);
755 list_del(&page->lru);
757 &zone->free_area[order].free_list[migratetype]);
759 pages_moved += 1 << order;
765 static int move_freepages_block(struct zone *zone, struct page *page,
768 unsigned long start_pfn, end_pfn;
769 struct page *start_page, *end_page;
771 start_pfn = page_to_pfn(page);
772 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
773 start_page = pfn_to_page(start_pfn);
774 end_page = start_page + pageblock_nr_pages - 1;
775 end_pfn = start_pfn + pageblock_nr_pages - 1;
777 /* Do not cross zone boundaries */
778 if (start_pfn < zone->zone_start_pfn)
780 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
783 return move_freepages(zone, start_page, end_page, migratetype);
786 static void change_pageblock_range(struct page *pageblock_page,
787 int start_order, int migratetype)
789 int nr_pageblocks = 1 << (start_order - pageblock_order);
791 while (nr_pageblocks--) {
792 set_pageblock_migratetype(pageblock_page, migratetype);
793 pageblock_page += pageblock_nr_pages;
797 /* Remove an element from the buddy allocator from the fallback list */
798 static inline struct page *
799 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
801 struct free_area * area;
806 /* Find the largest possible block of pages in the other list */
807 for (current_order = MAX_ORDER-1; current_order >= order;
809 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
810 migratetype = fallbacks[start_migratetype][i];
812 /* MIGRATE_RESERVE handled later if necessary */
813 if (migratetype == MIGRATE_RESERVE)
816 area = &(zone->free_area[current_order]);
817 if (list_empty(&area->free_list[migratetype]))
820 page = list_entry(area->free_list[migratetype].next,
825 * If breaking a large block of pages, move all free
826 * pages to the preferred allocation list. If falling
827 * back for a reclaimable kernel allocation, be more
828 * agressive about taking ownership of free pages
830 if (unlikely(current_order >= (pageblock_order >> 1)) ||
831 start_migratetype == MIGRATE_RECLAIMABLE ||
832 page_group_by_mobility_disabled) {
834 pages = move_freepages_block(zone, page,
837 /* Claim the whole block if over half of it is free */
838 if (pages >= (1 << (pageblock_order-1)) ||
839 page_group_by_mobility_disabled)
840 set_pageblock_migratetype(page,
843 migratetype = start_migratetype;
846 /* Remove the page from the freelists */
847 list_del(&page->lru);
848 rmv_page_order(page);
850 /* Take ownership for orders >= pageblock_order */
851 if (current_order >= pageblock_order)
852 change_pageblock_range(page, current_order,
855 expand(zone, page, order, current_order, area, migratetype);
857 trace_mm_page_alloc_extfrag(page, order, current_order,
858 start_migratetype, migratetype);
868 * Do the hard work of removing an element from the buddy allocator.
869 * Call me with the zone->lock already held.
871 static struct page *__rmqueue(struct zone *zone, unsigned int order,
877 page = __rmqueue_smallest(zone, order, migratetype);
879 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
880 page = __rmqueue_fallback(zone, order, migratetype);
883 * Use MIGRATE_RESERVE rather than fail an allocation. goto
884 * is used because __rmqueue_smallest is an inline function
885 * and we want just one call site
888 migratetype = MIGRATE_RESERVE;
897 * Obtain a specified number of elements from the buddy allocator, all under
898 * a single hold of the lock, for efficiency. Add them to the supplied list.
899 * Returns the number of new pages which were placed at *list.
901 static int rmqueue_bulk(struct zone *zone, unsigned int order,
902 unsigned long count, struct list_head *list,
903 int migratetype, int cold)
907 spin_lock(&zone->lock);
908 for (i = 0; i < count; ++i) {
909 struct page *page = __rmqueue(zone, order, migratetype);
910 if (unlikely(page == NULL))
914 * Split buddy pages returned by expand() are received here
915 * in physical page order. The page is added to the callers and
916 * list and the list head then moves forward. From the callers
917 * perspective, the linked list is ordered by page number in
918 * some conditions. This is useful for IO devices that can
919 * merge IO requests if the physical pages are ordered
922 if (likely(cold == 0))
923 list_add(&page->lru, list);
925 list_add_tail(&page->lru, list);
926 set_page_private(page, migratetype);
929 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
930 spin_unlock(&zone->lock);
936 * Called from the vmstat counter updater to drain pagesets of this
937 * currently executing processor on remote nodes after they have
940 * Note that this function must be called with the thread pinned to
941 * a single processor.
943 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
948 local_irq_save(flags);
949 if (pcp->count >= pcp->batch)
950 to_drain = pcp->batch;
952 to_drain = pcp->count;
953 free_pages_bulk(zone, to_drain, &pcp->list, 0);
954 pcp->count -= to_drain;
955 local_irq_restore(flags);
960 * Drain pages of the indicated processor.
962 * The processor must either be the current processor and the
963 * thread pinned to the current processor or a processor that
966 static void drain_pages(unsigned int cpu)
971 for_each_populated_zone(zone) {
972 struct per_cpu_pageset *pset;
973 struct per_cpu_pages *pcp;
975 pset = zone_pcp(zone, cpu);
978 local_irq_save(flags);
979 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
981 local_irq_restore(flags);
986 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
988 void drain_local_pages(void *arg)
990 drain_pages(smp_processor_id());
994 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
996 void drain_all_pages(void)
998 on_each_cpu(drain_local_pages, NULL, 1);
1001 #ifdef CONFIG_HIBERNATION
1003 void mark_free_pages(struct zone *zone)
1005 unsigned long pfn, max_zone_pfn;
1006 unsigned long flags;
1008 struct list_head *curr;
1010 if (!zone->spanned_pages)
1013 spin_lock_irqsave(&zone->lock, flags);
1015 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1016 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1017 if (pfn_valid(pfn)) {
1018 struct page *page = pfn_to_page(pfn);
1020 if (!swsusp_page_is_forbidden(page))
1021 swsusp_unset_page_free(page);
1024 for_each_migratetype_order(order, t) {
1025 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1028 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1029 for (i = 0; i < (1UL << order); i++)
1030 swsusp_set_page_free(pfn_to_page(pfn + i));
1033 spin_unlock_irqrestore(&zone->lock, flags);
1035 #endif /* CONFIG_PM */
1038 * Free a 0-order page
1040 static void free_hot_cold_page(struct page *page, int cold)
1042 struct zone *zone = page_zone(page);
1043 struct per_cpu_pages *pcp;
1044 unsigned long flags;
1045 int wasMlocked = __TestClearPageMlocked(page);
1047 kmemcheck_free_shadow(page, 0);
1050 page->mapping = NULL;
1051 if (free_pages_check(page))
1054 if (!PageHighMem(page)) {
1055 debug_check_no_locks_freed(page_address(page), PAGE_SIZE);
1056 debug_check_no_obj_freed(page_address(page), PAGE_SIZE);
1058 arch_free_page(page, 0);
1059 kernel_map_pages(page, 1, 0);
1061 pcp = &zone_pcp(zone, get_cpu())->pcp;
1062 set_page_private(page, get_pageblock_migratetype(page));
1063 local_irq_save(flags);
1064 if (unlikely(wasMlocked))
1065 free_page_mlock(page);
1066 __count_vm_event(PGFREE);
1069 list_add_tail(&page->lru, &pcp->list);
1071 list_add(&page->lru, &pcp->list);
1073 if (pcp->count >= pcp->high) {
1074 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
1075 pcp->count -= pcp->batch;
1077 local_irq_restore(flags);
1081 void free_hot_page(struct page *page)
1083 trace_mm_page_free_direct(page, 0);
1084 free_hot_cold_page(page, 0);
1088 * split_page takes a non-compound higher-order page, and splits it into
1089 * n (1<<order) sub-pages: page[0..n]
1090 * Each sub-page must be freed individually.
1092 * Note: this is probably too low level an operation for use in drivers.
1093 * Please consult with lkml before using this in your driver.
1095 void split_page(struct page *page, unsigned int order)
1099 VM_BUG_ON(PageCompound(page));
1100 VM_BUG_ON(!page_count(page));
1102 #ifdef CONFIG_KMEMCHECK
1104 * Split shadow pages too, because free(page[0]) would
1105 * otherwise free the whole shadow.
1107 if (kmemcheck_page_is_tracked(page))
1108 split_page(virt_to_page(page[0].shadow), order);
1111 for (i = 1; i < (1 << order); i++)
1112 set_page_refcounted(page + i);
1116 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1117 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1121 struct page *buffered_rmqueue(struct zone *preferred_zone,
1122 struct zone *zone, int order, gfp_t gfp_flags,
1125 unsigned long flags;
1127 int cold = !!(gfp_flags & __GFP_COLD);
1132 if (likely(order == 0)) {
1133 struct per_cpu_pages *pcp;
1135 pcp = &zone_pcp(zone, cpu)->pcp;
1136 local_irq_save(flags);
1138 pcp->count = rmqueue_bulk(zone, 0,
1139 pcp->batch, &pcp->list,
1141 if (unlikely(!pcp->count))
1145 /* Find a page of the appropriate migrate type */
1147 list_for_each_entry_reverse(page, &pcp->list, lru)
1148 if (page_private(page) == migratetype)
1151 list_for_each_entry(page, &pcp->list, lru)
1152 if (page_private(page) == migratetype)
1156 /* Allocate more to the pcp list if necessary */
1157 if (unlikely(&page->lru == &pcp->list)) {
1158 int get_one_page = 0;
1160 pcp->count += rmqueue_bulk(zone, 0,
1161 pcp->batch, &pcp->list,
1163 list_for_each_entry(page, &pcp->list, lru) {
1164 if (get_pageblock_migratetype(page) !=
1174 list_del(&page->lru);
1177 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1179 * __GFP_NOFAIL is not to be used in new code.
1181 * All __GFP_NOFAIL callers should be fixed so that they
1182 * properly detect and handle allocation failures.
1184 * We most definitely don't want callers attempting to
1185 * allocate greater than order-1 page units with
1188 WARN_ON_ONCE(order > 1);
1190 spin_lock_irqsave(&zone->lock, flags);
1191 page = __rmqueue(zone, order, migratetype);
1192 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1193 spin_unlock(&zone->lock);
1198 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1199 zone_statistics(preferred_zone, zone);
1200 local_irq_restore(flags);
1203 VM_BUG_ON(bad_range(zone, page));
1204 if (prep_new_page(page, order, gfp_flags))
1209 local_irq_restore(flags);
1214 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1215 #define ALLOC_WMARK_MIN WMARK_MIN
1216 #define ALLOC_WMARK_LOW WMARK_LOW
1217 #define ALLOC_WMARK_HIGH WMARK_HIGH
1218 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1220 /* Mask to get the watermark bits */
1221 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1223 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1224 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1225 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1227 #ifdef CONFIG_FAIL_PAGE_ALLOC
1229 static struct fail_page_alloc_attr {
1230 struct fault_attr attr;
1232 u32 ignore_gfp_highmem;
1233 u32 ignore_gfp_wait;
1236 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1238 struct dentry *ignore_gfp_highmem_file;
1239 struct dentry *ignore_gfp_wait_file;
1240 struct dentry *min_order_file;
1242 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1244 } fail_page_alloc = {
1245 .attr = FAULT_ATTR_INITIALIZER,
1246 .ignore_gfp_wait = 1,
1247 .ignore_gfp_highmem = 1,
1251 static int __init setup_fail_page_alloc(char *str)
1253 return setup_fault_attr(&fail_page_alloc.attr, str);
1255 __setup("fail_page_alloc=", setup_fail_page_alloc);
1257 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1259 if (order < fail_page_alloc.min_order)
1261 if (gfp_mask & __GFP_NOFAIL)
1263 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1265 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1268 return should_fail(&fail_page_alloc.attr, 1 << order);
1271 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1273 static int __init fail_page_alloc_debugfs(void)
1275 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1279 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1283 dir = fail_page_alloc.attr.dentries.dir;
1285 fail_page_alloc.ignore_gfp_wait_file =
1286 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1287 &fail_page_alloc.ignore_gfp_wait);
1289 fail_page_alloc.ignore_gfp_highmem_file =
1290 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1291 &fail_page_alloc.ignore_gfp_highmem);
1292 fail_page_alloc.min_order_file =
1293 debugfs_create_u32("min-order", mode, dir,
1294 &fail_page_alloc.min_order);
1296 if (!fail_page_alloc.ignore_gfp_wait_file ||
1297 !fail_page_alloc.ignore_gfp_highmem_file ||
1298 !fail_page_alloc.min_order_file) {
1300 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1301 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1302 debugfs_remove(fail_page_alloc.min_order_file);
1303 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1309 late_initcall(fail_page_alloc_debugfs);
1311 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1313 #else /* CONFIG_FAIL_PAGE_ALLOC */
1315 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1320 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1323 * Return 1 if free pages are above 'mark'. This takes into account the order
1324 * of the allocation.
1326 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1327 int classzone_idx, int alloc_flags)
1329 /* free_pages my go negative - that's OK */
1331 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
1334 if (alloc_flags & ALLOC_HIGH)
1336 if (alloc_flags & ALLOC_HARDER)
1339 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1341 for (o = 0; o < order; o++) {
1342 /* At the next order, this order's pages become unavailable */
1343 free_pages -= z->free_area[o].nr_free << o;
1345 /* Require fewer higher order pages to be free */
1348 if (free_pages <= min)
1356 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1357 * skip over zones that are not allowed by the cpuset, or that have
1358 * been recently (in last second) found to be nearly full. See further
1359 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1360 * that have to skip over a lot of full or unallowed zones.
1362 * If the zonelist cache is present in the passed in zonelist, then
1363 * returns a pointer to the allowed node mask (either the current
1364 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1366 * If the zonelist cache is not available for this zonelist, does
1367 * nothing and returns NULL.
1369 * If the fullzones BITMAP in the zonelist cache is stale (more than
1370 * a second since last zap'd) then we zap it out (clear its bits.)
1372 * We hold off even calling zlc_setup, until after we've checked the
1373 * first zone in the zonelist, on the theory that most allocations will
1374 * be satisfied from that first zone, so best to examine that zone as
1375 * quickly as we can.
1377 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1379 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1380 nodemask_t *allowednodes; /* zonelist_cache approximation */
1382 zlc = zonelist->zlcache_ptr;
1386 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1387 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1388 zlc->last_full_zap = jiffies;
1391 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1392 &cpuset_current_mems_allowed :
1393 &node_states[N_HIGH_MEMORY];
1394 return allowednodes;
1398 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1399 * if it is worth looking at further for free memory:
1400 * 1) Check that the zone isn't thought to be full (doesn't have its
1401 * bit set in the zonelist_cache fullzones BITMAP).
1402 * 2) Check that the zones node (obtained from the zonelist_cache
1403 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1404 * Return true (non-zero) if zone is worth looking at further, or
1405 * else return false (zero) if it is not.
1407 * This check -ignores- the distinction between various watermarks,
1408 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1409 * found to be full for any variation of these watermarks, it will
1410 * be considered full for up to one second by all requests, unless
1411 * we are so low on memory on all allowed nodes that we are forced
1412 * into the second scan of the zonelist.
1414 * In the second scan we ignore this zonelist cache and exactly
1415 * apply the watermarks to all zones, even it is slower to do so.
1416 * We are low on memory in the second scan, and should leave no stone
1417 * unturned looking for a free page.
1419 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1420 nodemask_t *allowednodes)
1422 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1423 int i; /* index of *z in zonelist zones */
1424 int n; /* node that zone *z is on */
1426 zlc = zonelist->zlcache_ptr;
1430 i = z - zonelist->_zonerefs;
1433 /* This zone is worth trying if it is allowed but not full */
1434 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1438 * Given 'z' scanning a zonelist, set the corresponding bit in
1439 * zlc->fullzones, so that subsequent attempts to allocate a page
1440 * from that zone don't waste time re-examining it.
1442 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1444 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1445 int i; /* index of *z in zonelist zones */
1447 zlc = zonelist->zlcache_ptr;
1451 i = z - zonelist->_zonerefs;
1453 set_bit(i, zlc->fullzones);
1456 #else /* CONFIG_NUMA */
1458 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1463 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1464 nodemask_t *allowednodes)
1469 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1472 #endif /* CONFIG_NUMA */
1475 * get_page_from_freelist goes through the zonelist trying to allocate
1478 static struct page *
1479 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1480 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1481 struct zone *preferred_zone, int migratetype)
1484 struct page *page = NULL;
1487 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1488 int zlc_active = 0; /* set if using zonelist_cache */
1489 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1491 classzone_idx = zone_idx(preferred_zone);
1494 * Scan zonelist, looking for a zone with enough free.
1495 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1497 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1498 high_zoneidx, nodemask) {
1499 if (NUMA_BUILD && zlc_active &&
1500 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1502 if ((alloc_flags & ALLOC_CPUSET) &&
1503 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1506 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1507 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1511 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1512 if (zone_watermark_ok(zone, order, mark,
1513 classzone_idx, alloc_flags))
1516 if (zone_reclaim_mode == 0)
1517 goto this_zone_full;
1519 ret = zone_reclaim(zone, gfp_mask, order);
1521 case ZONE_RECLAIM_NOSCAN:
1524 case ZONE_RECLAIM_FULL:
1525 /* scanned but unreclaimable */
1526 goto this_zone_full;
1528 /* did we reclaim enough */
1529 if (!zone_watermark_ok(zone, order, mark,
1530 classzone_idx, alloc_flags))
1531 goto this_zone_full;
1536 page = buffered_rmqueue(preferred_zone, zone, order,
1537 gfp_mask, migratetype);
1542 zlc_mark_zone_full(zonelist, z);
1544 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1546 * we do zlc_setup after the first zone is tried but only
1547 * if there are multiple nodes make it worthwhile
1549 allowednodes = zlc_setup(zonelist, alloc_flags);
1555 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1556 /* Disable zlc cache for second zonelist scan */
1564 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1565 unsigned long pages_reclaimed)
1567 /* Do not loop if specifically requested */
1568 if (gfp_mask & __GFP_NORETRY)
1572 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1573 * means __GFP_NOFAIL, but that may not be true in other
1576 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1580 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1581 * specified, then we retry until we no longer reclaim any pages
1582 * (above), or we've reclaimed an order of pages at least as
1583 * large as the allocation's order. In both cases, if the
1584 * allocation still fails, we stop retrying.
1586 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1590 * Don't let big-order allocations loop unless the caller
1591 * explicitly requests that.
1593 if (gfp_mask & __GFP_NOFAIL)
1599 static inline struct page *
1600 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1601 struct zonelist *zonelist, enum zone_type high_zoneidx,
1602 nodemask_t *nodemask, struct zone *preferred_zone,
1607 /* Acquire the OOM killer lock for the zones in zonelist */
1608 if (!try_set_zone_oom(zonelist, gfp_mask)) {
1609 schedule_timeout_uninterruptible(1);
1614 * Go through the zonelist yet one more time, keep very high watermark
1615 * here, this is only to catch a parallel oom killing, we must fail if
1616 * we're still under heavy pressure.
1618 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1619 order, zonelist, high_zoneidx,
1620 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1621 preferred_zone, migratetype);
1625 /* The OOM killer will not help higher order allocs */
1626 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_NOFAIL))
1629 /* Exhausted what can be done so it's blamo time */
1630 out_of_memory(zonelist, gfp_mask, order);
1633 clear_zonelist_oom(zonelist, gfp_mask);
1637 /* The really slow allocator path where we enter direct reclaim */
1638 static inline struct page *
1639 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1640 struct zonelist *zonelist, enum zone_type high_zoneidx,
1641 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1642 int migratetype, unsigned long *did_some_progress)
1644 struct page *page = NULL;
1645 struct reclaim_state reclaim_state;
1646 struct task_struct *p = current;
1650 /* We now go into synchronous reclaim */
1651 cpuset_memory_pressure_bump();
1652 p->flags |= PF_MEMALLOC;
1653 lockdep_set_current_reclaim_state(gfp_mask);
1654 reclaim_state.reclaimed_slab = 0;
1655 p->reclaim_state = &reclaim_state;
1657 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1659 p->reclaim_state = NULL;
1660 lockdep_clear_current_reclaim_state();
1661 p->flags &= ~PF_MEMALLOC;
1668 if (likely(*did_some_progress))
1669 page = get_page_from_freelist(gfp_mask, nodemask, order,
1670 zonelist, high_zoneidx,
1671 alloc_flags, preferred_zone,
1677 * This is called in the allocator slow-path if the allocation request is of
1678 * sufficient urgency to ignore watermarks and take other desperate measures
1680 static inline struct page *
1681 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1682 struct zonelist *zonelist, enum zone_type high_zoneidx,
1683 nodemask_t *nodemask, struct zone *preferred_zone,
1689 page = get_page_from_freelist(gfp_mask, nodemask, order,
1690 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
1691 preferred_zone, migratetype);
1693 if (!page && gfp_mask & __GFP_NOFAIL)
1694 congestion_wait(BLK_RW_ASYNC, HZ/50);
1695 } while (!page && (gfp_mask & __GFP_NOFAIL));
1701 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
1702 enum zone_type high_zoneidx)
1707 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1708 wakeup_kswapd(zone, order);
1712 gfp_to_alloc_flags(gfp_t gfp_mask)
1714 struct task_struct *p = current;
1715 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
1716 const gfp_t wait = gfp_mask & __GFP_WAIT;
1718 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1719 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH);
1722 * The caller may dip into page reserves a bit more if the caller
1723 * cannot run direct reclaim, or if the caller has realtime scheduling
1724 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1725 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1727 alloc_flags |= (gfp_mask & __GFP_HIGH);
1730 alloc_flags |= ALLOC_HARDER;
1732 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1733 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1735 alloc_flags &= ~ALLOC_CPUSET;
1736 } else if (unlikely(rt_task(p)))
1737 alloc_flags |= ALLOC_HARDER;
1739 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
1740 if (!in_interrupt() &&
1741 ((p->flags & PF_MEMALLOC) ||
1742 unlikely(test_thread_flag(TIF_MEMDIE))))
1743 alloc_flags |= ALLOC_NO_WATERMARKS;
1749 static inline struct page *
1750 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
1751 struct zonelist *zonelist, enum zone_type high_zoneidx,
1752 nodemask_t *nodemask, struct zone *preferred_zone,
1755 const gfp_t wait = gfp_mask & __GFP_WAIT;
1756 struct page *page = NULL;
1758 unsigned long pages_reclaimed = 0;
1759 unsigned long did_some_progress;
1760 struct task_struct *p = current;
1763 * In the slowpath, we sanity check order to avoid ever trying to
1764 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1765 * be using allocators in order of preference for an area that is
1768 if (order >= MAX_ORDER) {
1769 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
1774 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1775 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1776 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1777 * using a larger set of nodes after it has established that the
1778 * allowed per node queues are empty and that nodes are
1781 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1784 wake_all_kswapd(order, zonelist, high_zoneidx);
1788 * OK, we're below the kswapd watermark and have kicked background
1789 * reclaim. Now things get more complex, so set up alloc_flags according
1790 * to how we want to proceed.
1792 alloc_flags = gfp_to_alloc_flags(gfp_mask);
1794 /* This is the last chance, in general, before the goto nopage. */
1795 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
1796 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
1797 preferred_zone, migratetype);
1802 /* Allocate without watermarks if the context allows */
1803 if (alloc_flags & ALLOC_NO_WATERMARKS) {
1804 page = __alloc_pages_high_priority(gfp_mask, order,
1805 zonelist, high_zoneidx, nodemask,
1806 preferred_zone, migratetype);
1811 /* Atomic allocations - we can't balance anything */
1815 /* Avoid recursion of direct reclaim */
1816 if (p->flags & PF_MEMALLOC)
1819 /* Avoid allocations with no watermarks from looping endlessly */
1820 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
1823 /* Try direct reclaim and then allocating */
1824 page = __alloc_pages_direct_reclaim(gfp_mask, order,
1825 zonelist, high_zoneidx,
1827 alloc_flags, preferred_zone,
1828 migratetype, &did_some_progress);
1833 * If we failed to make any progress reclaiming, then we are
1834 * running out of options and have to consider going OOM
1836 if (!did_some_progress) {
1837 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1838 if (oom_killer_disabled)
1840 page = __alloc_pages_may_oom(gfp_mask, order,
1841 zonelist, high_zoneidx,
1842 nodemask, preferred_zone,
1848 * The OOM killer does not trigger for high-order
1849 * ~__GFP_NOFAIL allocations so if no progress is being
1850 * made, there are no other options and retrying is
1853 if (order > PAGE_ALLOC_COSTLY_ORDER &&
1854 !(gfp_mask & __GFP_NOFAIL))
1861 /* Check if we should retry the allocation */
1862 pages_reclaimed += did_some_progress;
1863 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
1864 /* Wait for some write requests to complete then retry */
1865 congestion_wait(BLK_RW_ASYNC, HZ/50);
1870 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1871 printk(KERN_WARNING "%s: page allocation failure."
1872 " order:%d, mode:0x%x\n",
1873 p->comm, order, gfp_mask);
1879 if (kmemcheck_enabled)
1880 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
1886 * This is the 'heart' of the zoned buddy allocator.
1889 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
1890 struct zonelist *zonelist, nodemask_t *nodemask)
1892 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1893 struct zone *preferred_zone;
1895 int migratetype = allocflags_to_migratetype(gfp_mask);
1897 gfp_mask &= gfp_allowed_mask;
1899 lockdep_trace_alloc(gfp_mask);
1901 might_sleep_if(gfp_mask & __GFP_WAIT);
1903 if (should_fail_alloc_page(gfp_mask, order))
1907 * Check the zones suitable for the gfp_mask contain at least one
1908 * valid zone. It's possible to have an empty zonelist as a result
1909 * of GFP_THISNODE and a memoryless node
1911 if (unlikely(!zonelist->_zonerefs->zone))
1914 /* The preferred zone is used for statistics later */
1915 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone);
1916 if (!preferred_zone)
1919 /* First allocation attempt */
1920 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
1921 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
1922 preferred_zone, migratetype);
1923 if (unlikely(!page))
1924 page = __alloc_pages_slowpath(gfp_mask, order,
1925 zonelist, high_zoneidx, nodemask,
1926 preferred_zone, migratetype);
1928 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
1931 EXPORT_SYMBOL(__alloc_pages_nodemask);
1934 * Common helper functions.
1936 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1941 * __get_free_pages() returns a 32-bit address, which cannot represent
1944 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1946 page = alloc_pages(gfp_mask, order);
1949 return (unsigned long) page_address(page);
1951 EXPORT_SYMBOL(__get_free_pages);
1953 unsigned long get_zeroed_page(gfp_t gfp_mask)
1955 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
1957 EXPORT_SYMBOL(get_zeroed_page);
1959 void __pagevec_free(struct pagevec *pvec)
1961 int i = pagevec_count(pvec);
1964 trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
1965 free_hot_cold_page(pvec->pages[i], pvec->cold);
1969 void __free_pages(struct page *page, unsigned int order)
1971 if (put_page_testzero(page)) {
1972 trace_mm_page_free_direct(page, order);
1974 free_hot_page(page);
1976 __free_pages_ok(page, order);
1980 EXPORT_SYMBOL(__free_pages);
1982 void free_pages(unsigned long addr, unsigned int order)
1985 VM_BUG_ON(!virt_addr_valid((void *)addr));
1986 __free_pages(virt_to_page((void *)addr), order);
1990 EXPORT_SYMBOL(free_pages);
1993 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
1994 * @size: the number of bytes to allocate
1995 * @gfp_mask: GFP flags for the allocation
1997 * This function is similar to alloc_pages(), except that it allocates the
1998 * minimum number of pages to satisfy the request. alloc_pages() can only
1999 * allocate memory in power-of-two pages.
2001 * This function is also limited by MAX_ORDER.
2003 * Memory allocated by this function must be released by free_pages_exact().
2005 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2007 unsigned int order = get_order(size);
2010 addr = __get_free_pages(gfp_mask, order);
2012 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2013 unsigned long used = addr + PAGE_ALIGN(size);
2015 split_page(virt_to_page((void *)addr), order);
2016 while (used < alloc_end) {
2022 return (void *)addr;
2024 EXPORT_SYMBOL(alloc_pages_exact);
2027 * free_pages_exact - release memory allocated via alloc_pages_exact()
2028 * @virt: the value returned by alloc_pages_exact.
2029 * @size: size of allocation, same value as passed to alloc_pages_exact().
2031 * Release the memory allocated by a previous call to alloc_pages_exact.
2033 void free_pages_exact(void *virt, size_t size)
2035 unsigned long addr = (unsigned long)virt;
2036 unsigned long end = addr + PAGE_ALIGN(size);
2038 while (addr < end) {
2043 EXPORT_SYMBOL(free_pages_exact);
2045 static unsigned int nr_free_zone_pages(int offset)
2050 /* Just pick one node, since fallback list is circular */
2051 unsigned int sum = 0;
2053 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2055 for_each_zone_zonelist(zone, z, zonelist, offset) {
2056 unsigned long size = zone->present_pages;
2057 unsigned long high = high_wmark_pages(zone);
2066 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2068 unsigned int nr_free_buffer_pages(void)
2070 return nr_free_zone_pages(gfp_zone(GFP_USER));
2072 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2075 * Amount of free RAM allocatable within all zones
2077 unsigned int nr_free_pagecache_pages(void)
2079 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2082 static inline void show_node(struct zone *zone)
2085 printk("Node %d ", zone_to_nid(zone));
2088 void si_meminfo(struct sysinfo *val)
2090 val->totalram = totalram_pages;
2092 val->freeram = global_page_state(NR_FREE_PAGES);
2093 val->bufferram = nr_blockdev_pages();
2094 val->totalhigh = totalhigh_pages;
2095 val->freehigh = nr_free_highpages();
2096 val->mem_unit = PAGE_SIZE;
2099 EXPORT_SYMBOL(si_meminfo);
2102 void si_meminfo_node(struct sysinfo *val, int nid)
2104 pg_data_t *pgdat = NODE_DATA(nid);
2106 val->totalram = pgdat->node_present_pages;
2107 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2108 #ifdef CONFIG_HIGHMEM
2109 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2110 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2116 val->mem_unit = PAGE_SIZE;
2120 #define K(x) ((x) << (PAGE_SHIFT-10))
2123 * Show free area list (used inside shift_scroll-lock stuff)
2124 * We also calculate the percentage fragmentation. We do this by counting the
2125 * memory on each free list with the exception of the first item on the list.
2127 void show_free_areas(void)
2132 for_each_populated_zone(zone) {
2134 printk("%s per-cpu:\n", zone->name);
2136 for_each_online_cpu(cpu) {
2137 struct per_cpu_pageset *pageset;
2139 pageset = zone_pcp(zone, cpu);
2141 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2142 cpu, pageset->pcp.high,
2143 pageset->pcp.batch, pageset->pcp.count);
2147 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2148 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2150 " dirty:%lu writeback:%lu unstable:%lu buffer:%lu\n"
2151 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2152 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2153 global_page_state(NR_ACTIVE_ANON),
2154 global_page_state(NR_INACTIVE_ANON),
2155 global_page_state(NR_ISOLATED_ANON),
2156 global_page_state(NR_ACTIVE_FILE),
2157 global_page_state(NR_INACTIVE_FILE),
2158 global_page_state(NR_ISOLATED_FILE),
2159 global_page_state(NR_UNEVICTABLE),
2160 global_page_state(NR_FILE_DIRTY),
2161 global_page_state(NR_WRITEBACK),
2162 global_page_state(NR_UNSTABLE_NFS),
2163 nr_blockdev_pages(),
2164 global_page_state(NR_FREE_PAGES),
2165 global_page_state(NR_SLAB_RECLAIMABLE),
2166 global_page_state(NR_SLAB_UNRECLAIMABLE),
2167 global_page_state(NR_FILE_MAPPED),
2168 global_page_state(NR_SHMEM),
2169 global_page_state(NR_PAGETABLE),
2170 global_page_state(NR_BOUNCE));
2172 for_each_populated_zone(zone) {
2181 " active_anon:%lukB"
2182 " inactive_anon:%lukB"
2183 " active_file:%lukB"
2184 " inactive_file:%lukB"
2185 " unevictable:%lukB"
2186 " isolated(anon):%lukB"
2187 " isolated(file):%lukB"
2194 " slab_reclaimable:%lukB"
2195 " slab_unreclaimable:%lukB"
2196 " kernel_stack:%lukB"
2200 " writeback_tmp:%lukB"
2201 " pages_scanned:%lu"
2202 " all_unreclaimable? %s"
2205 K(zone_page_state(zone, NR_FREE_PAGES)),
2206 K(min_wmark_pages(zone)),
2207 K(low_wmark_pages(zone)),
2208 K(high_wmark_pages(zone)),
2209 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2210 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2211 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2212 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2213 K(zone_page_state(zone, NR_UNEVICTABLE)),
2214 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2215 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2216 K(zone->present_pages),
2217 K(zone_page_state(zone, NR_MLOCK)),
2218 K(zone_page_state(zone, NR_FILE_DIRTY)),
2219 K(zone_page_state(zone, NR_WRITEBACK)),
2220 K(zone_page_state(zone, NR_FILE_MAPPED)),
2221 K(zone_page_state(zone, NR_SHMEM)),
2222 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2223 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2224 zone_page_state(zone, NR_KERNEL_STACK) *
2226 K(zone_page_state(zone, NR_PAGETABLE)),
2227 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2228 K(zone_page_state(zone, NR_BOUNCE)),
2229 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2230 zone->pages_scanned,
2231 (zone_is_all_unreclaimable(zone) ? "yes" : "no")
2233 printk("lowmem_reserve[]:");
2234 for (i = 0; i < MAX_NR_ZONES; i++)
2235 printk(" %lu", zone->lowmem_reserve[i]);
2239 for_each_populated_zone(zone) {
2240 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2243 printk("%s: ", zone->name);
2245 spin_lock_irqsave(&zone->lock, flags);
2246 for (order = 0; order < MAX_ORDER; order++) {
2247 nr[order] = zone->free_area[order].nr_free;
2248 total += nr[order] << order;
2250 spin_unlock_irqrestore(&zone->lock, flags);
2251 for (order = 0; order < MAX_ORDER; order++)
2252 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2253 printk("= %lukB\n", K(total));
2256 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2258 show_swap_cache_info();
2261 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2263 zoneref->zone = zone;
2264 zoneref->zone_idx = zone_idx(zone);
2268 * Builds allocation fallback zone lists.
2270 * Add all populated zones of a node to the zonelist.
2272 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2273 int nr_zones, enum zone_type zone_type)
2277 BUG_ON(zone_type >= MAX_NR_ZONES);
2282 zone = pgdat->node_zones + zone_type;
2283 if (populated_zone(zone)) {
2284 zoneref_set_zone(zone,
2285 &zonelist->_zonerefs[nr_zones++]);
2286 check_highest_zone(zone_type);
2289 } while (zone_type);
2296 * 0 = automatic detection of better ordering.
2297 * 1 = order by ([node] distance, -zonetype)
2298 * 2 = order by (-zonetype, [node] distance)
2300 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2301 * the same zonelist. So only NUMA can configure this param.
2303 #define ZONELIST_ORDER_DEFAULT 0
2304 #define ZONELIST_ORDER_NODE 1
2305 #define ZONELIST_ORDER_ZONE 2
2307 /* zonelist order in the kernel.
2308 * set_zonelist_order() will set this to NODE or ZONE.
2310 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2311 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2315 /* The value user specified ....changed by config */
2316 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2317 /* string for sysctl */
2318 #define NUMA_ZONELIST_ORDER_LEN 16
2319 char numa_zonelist_order[16] = "default";
2322 * interface for configure zonelist ordering.
2323 * command line option "numa_zonelist_order"
2324 * = "[dD]efault - default, automatic configuration.
2325 * = "[nN]ode - order by node locality, then by zone within node
2326 * = "[zZ]one - order by zone, then by locality within zone
2329 static int __parse_numa_zonelist_order(char *s)
2331 if (*s == 'd' || *s == 'D') {
2332 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2333 } else if (*s == 'n' || *s == 'N') {
2334 user_zonelist_order = ZONELIST_ORDER_NODE;
2335 } else if (*s == 'z' || *s == 'Z') {
2336 user_zonelist_order = ZONELIST_ORDER_ZONE;
2339 "Ignoring invalid numa_zonelist_order value: "
2346 static __init int setup_numa_zonelist_order(char *s)
2349 return __parse_numa_zonelist_order(s);
2352 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2355 * sysctl handler for numa_zonelist_order
2357 int numa_zonelist_order_handler(ctl_table *table, int write,
2358 struct file *file, void __user *buffer, size_t *length,
2361 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2365 strncpy(saved_string, (char*)table->data,
2366 NUMA_ZONELIST_ORDER_LEN);
2367 ret = proc_dostring(table, write, file, buffer, length, ppos);
2371 int oldval = user_zonelist_order;
2372 if (__parse_numa_zonelist_order((char*)table->data)) {
2374 * bogus value. restore saved string
2376 strncpy((char*)table->data, saved_string,
2377 NUMA_ZONELIST_ORDER_LEN);
2378 user_zonelist_order = oldval;
2379 } else if (oldval != user_zonelist_order)
2380 build_all_zonelists();
2386 #define MAX_NODE_LOAD (nr_online_nodes)
2387 static int node_load[MAX_NUMNODES];
2390 * find_next_best_node - find the next node that should appear in a given node's fallback list
2391 * @node: node whose fallback list we're appending
2392 * @used_node_mask: nodemask_t of already used nodes
2394 * We use a number of factors to determine which is the next node that should
2395 * appear on a given node's fallback list. The node should not have appeared
2396 * already in @node's fallback list, and it should be the next closest node
2397 * according to the distance array (which contains arbitrary distance values
2398 * from each node to each node in the system), and should also prefer nodes
2399 * with no CPUs, since presumably they'll have very little allocation pressure
2400 * on them otherwise.
2401 * It returns -1 if no node is found.
2403 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2406 int min_val = INT_MAX;
2408 const struct cpumask *tmp = cpumask_of_node(0);
2410 /* Use the local node if we haven't already */
2411 if (!node_isset(node, *used_node_mask)) {
2412 node_set(node, *used_node_mask);
2416 for_each_node_state(n, N_HIGH_MEMORY) {
2418 /* Don't want a node to appear more than once */
2419 if (node_isset(n, *used_node_mask))
2422 /* Use the distance array to find the distance */
2423 val = node_distance(node, n);
2425 /* Penalize nodes under us ("prefer the next node") */
2428 /* Give preference to headless and unused nodes */
2429 tmp = cpumask_of_node(n);
2430 if (!cpumask_empty(tmp))
2431 val += PENALTY_FOR_NODE_WITH_CPUS;
2433 /* Slight preference for less loaded node */
2434 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2435 val += node_load[n];
2437 if (val < min_val) {
2444 node_set(best_node, *used_node_mask);
2451 * Build zonelists ordered by node and zones within node.
2452 * This results in maximum locality--normal zone overflows into local
2453 * DMA zone, if any--but risks exhausting DMA zone.
2455 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2458 struct zonelist *zonelist;
2460 zonelist = &pgdat->node_zonelists[0];
2461 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2463 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2465 zonelist->_zonerefs[j].zone = NULL;
2466 zonelist->_zonerefs[j].zone_idx = 0;
2470 * Build gfp_thisnode zonelists
2472 static void build_thisnode_zonelists(pg_data_t *pgdat)
2475 struct zonelist *zonelist;
2477 zonelist = &pgdat->node_zonelists[1];
2478 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2479 zonelist->_zonerefs[j].zone = NULL;
2480 zonelist->_zonerefs[j].zone_idx = 0;
2484 * Build zonelists ordered by zone and nodes within zones.
2485 * This results in conserving DMA zone[s] until all Normal memory is
2486 * exhausted, but results in overflowing to remote node while memory
2487 * may still exist in local DMA zone.
2489 static int node_order[MAX_NUMNODES];
2491 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2494 int zone_type; /* needs to be signed */
2496 struct zonelist *zonelist;
2498 zonelist = &pgdat->node_zonelists[0];
2500 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2501 for (j = 0; j < nr_nodes; j++) {
2502 node = node_order[j];
2503 z = &NODE_DATA(node)->node_zones[zone_type];
2504 if (populated_zone(z)) {
2506 &zonelist->_zonerefs[pos++]);
2507 check_highest_zone(zone_type);
2511 zonelist->_zonerefs[pos].zone = NULL;
2512 zonelist->_zonerefs[pos].zone_idx = 0;
2515 static int default_zonelist_order(void)
2518 unsigned long low_kmem_size,total_size;
2522 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem.
2523 * If they are really small and used heavily, the system can fall
2524 * into OOM very easily.
2525 * This function detect ZONE_DMA/DMA32 size and confgigures zone order.
2527 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2530 for_each_online_node(nid) {
2531 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2532 z = &NODE_DATA(nid)->node_zones[zone_type];
2533 if (populated_zone(z)) {
2534 if (zone_type < ZONE_NORMAL)
2535 low_kmem_size += z->present_pages;
2536 total_size += z->present_pages;
2540 if (!low_kmem_size || /* there are no DMA area. */
2541 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2542 return ZONELIST_ORDER_NODE;
2544 * look into each node's config.
2545 * If there is a node whose DMA/DMA32 memory is very big area on
2546 * local memory, NODE_ORDER may be suitable.
2548 average_size = total_size /
2549 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2550 for_each_online_node(nid) {
2553 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2554 z = &NODE_DATA(nid)->node_zones[zone_type];
2555 if (populated_zone(z)) {
2556 if (zone_type < ZONE_NORMAL)
2557 low_kmem_size += z->present_pages;
2558 total_size += z->present_pages;
2561 if (low_kmem_size &&
2562 total_size > average_size && /* ignore small node */
2563 low_kmem_size > total_size * 70/100)
2564 return ZONELIST_ORDER_NODE;
2566 return ZONELIST_ORDER_ZONE;
2569 static void set_zonelist_order(void)
2571 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2572 current_zonelist_order = default_zonelist_order();
2574 current_zonelist_order = user_zonelist_order;
2577 static void build_zonelists(pg_data_t *pgdat)
2581 nodemask_t used_mask;
2582 int local_node, prev_node;
2583 struct zonelist *zonelist;
2584 int order = current_zonelist_order;
2586 /* initialize zonelists */
2587 for (i = 0; i < MAX_ZONELISTS; i++) {
2588 zonelist = pgdat->node_zonelists + i;
2589 zonelist->_zonerefs[0].zone = NULL;
2590 zonelist->_zonerefs[0].zone_idx = 0;
2593 /* NUMA-aware ordering of nodes */
2594 local_node = pgdat->node_id;
2595 load = nr_online_nodes;
2596 prev_node = local_node;
2597 nodes_clear(used_mask);
2599 memset(node_order, 0, sizeof(node_order));
2602 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2603 int distance = node_distance(local_node, node);
2606 * If another node is sufficiently far away then it is better
2607 * to reclaim pages in a zone before going off node.
2609 if (distance > RECLAIM_DISTANCE)
2610 zone_reclaim_mode = 1;
2613 * We don't want to pressure a particular node.
2614 * So adding penalty to the first node in same
2615 * distance group to make it round-robin.
2617 if (distance != node_distance(local_node, prev_node))
2618 node_load[node] = load;
2622 if (order == ZONELIST_ORDER_NODE)
2623 build_zonelists_in_node_order(pgdat, node);
2625 node_order[j++] = node; /* remember order */
2628 if (order == ZONELIST_ORDER_ZONE) {
2629 /* calculate node order -- i.e., DMA last! */
2630 build_zonelists_in_zone_order(pgdat, j);
2633 build_thisnode_zonelists(pgdat);
2636 /* Construct the zonelist performance cache - see further mmzone.h */
2637 static void build_zonelist_cache(pg_data_t *pgdat)
2639 struct zonelist *zonelist;
2640 struct zonelist_cache *zlc;
2643 zonelist = &pgdat->node_zonelists[0];
2644 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2645 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2646 for (z = zonelist->_zonerefs; z->zone; z++)
2647 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2651 #else /* CONFIG_NUMA */
2653 static void set_zonelist_order(void)
2655 current_zonelist_order = ZONELIST_ORDER_ZONE;
2658 static void build_zonelists(pg_data_t *pgdat)
2660 int node, local_node;
2662 struct zonelist *zonelist;
2664 local_node = pgdat->node_id;
2666 zonelist = &pgdat->node_zonelists[0];
2667 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2670 * Now we build the zonelist so that it contains the zones
2671 * of all the other nodes.
2672 * We don't want to pressure a particular node, so when
2673 * building the zones for node N, we make sure that the
2674 * zones coming right after the local ones are those from
2675 * node N+1 (modulo N)
2677 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2678 if (!node_online(node))
2680 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2683 for (node = 0; node < local_node; node++) {
2684 if (!node_online(node))
2686 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2690 zonelist->_zonerefs[j].zone = NULL;
2691 zonelist->_zonerefs[j].zone_idx = 0;
2694 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2695 static void build_zonelist_cache(pg_data_t *pgdat)
2697 pgdat->node_zonelists[0].zlcache_ptr = NULL;
2700 #endif /* CONFIG_NUMA */
2702 /* return values int ....just for stop_machine() */
2703 static int __build_all_zonelists(void *dummy)
2708 memset(node_load, 0, sizeof(node_load));
2710 for_each_online_node(nid) {
2711 pg_data_t *pgdat = NODE_DATA(nid);
2713 build_zonelists(pgdat);
2714 build_zonelist_cache(pgdat);
2719 void build_all_zonelists(void)
2721 set_zonelist_order();
2723 if (system_state == SYSTEM_BOOTING) {
2724 __build_all_zonelists(NULL);
2725 mminit_verify_zonelist();
2726 cpuset_init_current_mems_allowed();
2728 /* we have to stop all cpus to guarantee there is no user
2730 stop_machine(__build_all_zonelists, NULL, NULL);
2731 /* cpuset refresh routine should be here */
2733 vm_total_pages = nr_free_pagecache_pages();
2735 * Disable grouping by mobility if the number of pages in the
2736 * system is too low to allow the mechanism to work. It would be
2737 * more accurate, but expensive to check per-zone. This check is
2738 * made on memory-hotadd so a system can start with mobility
2739 * disabled and enable it later
2741 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
2742 page_group_by_mobility_disabled = 1;
2744 page_group_by_mobility_disabled = 0;
2746 printk("Built %i zonelists in %s order, mobility grouping %s. "
2747 "Total pages: %ld\n",
2749 zonelist_order_name[current_zonelist_order],
2750 page_group_by_mobility_disabled ? "off" : "on",
2753 printk("Policy zone: %s\n", zone_names[policy_zone]);
2758 * Helper functions to size the waitqueue hash table.
2759 * Essentially these want to choose hash table sizes sufficiently
2760 * large so that collisions trying to wait on pages are rare.
2761 * But in fact, the number of active page waitqueues on typical
2762 * systems is ridiculously low, less than 200. So this is even
2763 * conservative, even though it seems large.
2765 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2766 * waitqueues, i.e. the size of the waitq table given the number of pages.
2768 #define PAGES_PER_WAITQUEUE 256
2770 #ifndef CONFIG_MEMORY_HOTPLUG
2771 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2773 unsigned long size = 1;
2775 pages /= PAGES_PER_WAITQUEUE;
2777 while (size < pages)
2781 * Once we have dozens or even hundreds of threads sleeping
2782 * on IO we've got bigger problems than wait queue collision.
2783 * Limit the size of the wait table to a reasonable size.
2785 size = min(size, 4096UL);
2787 return max(size, 4UL);
2791 * A zone's size might be changed by hot-add, so it is not possible to determine
2792 * a suitable size for its wait_table. So we use the maximum size now.
2794 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2796 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2797 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2798 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2800 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2801 * or more by the traditional way. (See above). It equals:
2803 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2804 * ia64(16K page size) : = ( 8G + 4M)byte.
2805 * powerpc (64K page size) : = (32G +16M)byte.
2807 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2814 * This is an integer logarithm so that shifts can be used later
2815 * to extract the more random high bits from the multiplicative
2816 * hash function before the remainder is taken.
2818 static inline unsigned long wait_table_bits(unsigned long size)
2823 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
2826 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
2827 * of blocks reserved is based on min_wmark_pages(zone). The memory within
2828 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
2829 * higher will lead to a bigger reserve which will get freed as contiguous
2830 * blocks as reclaim kicks in
2832 static void setup_zone_migrate_reserve(struct zone *zone)
2834 unsigned long start_pfn, pfn, end_pfn;
2836 unsigned long reserve, block_migratetype;
2838 /* Get the start pfn, end pfn and the number of blocks to reserve */
2839 start_pfn = zone->zone_start_pfn;
2840 end_pfn = start_pfn + zone->spanned_pages;
2841 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
2844 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
2845 if (!pfn_valid(pfn))
2847 page = pfn_to_page(pfn);
2849 /* Watch out for overlapping nodes */
2850 if (page_to_nid(page) != zone_to_nid(zone))
2853 /* Blocks with reserved pages will never free, skip them. */
2854 if (PageReserved(page))
2857 block_migratetype = get_pageblock_migratetype(page);
2859 /* If this block is reserved, account for it */
2860 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
2865 /* Suitable for reserving if this block is movable */
2866 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
2867 set_pageblock_migratetype(page, MIGRATE_RESERVE);
2868 move_freepages_block(zone, page, MIGRATE_RESERVE);
2874 * If the reserve is met and this is a previous reserved block,
2877 if (block_migratetype == MIGRATE_RESERVE) {
2878 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2879 move_freepages_block(zone, page, MIGRATE_MOVABLE);
2885 * Initially all pages are reserved - free ones are freed
2886 * up by free_all_bootmem() once the early boot process is
2887 * done. Non-atomic initialization, single-pass.
2889 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
2890 unsigned long start_pfn, enum memmap_context context)
2893 unsigned long end_pfn = start_pfn + size;
2897 if (highest_memmap_pfn < end_pfn - 1)
2898 highest_memmap_pfn = end_pfn - 1;
2900 z = &NODE_DATA(nid)->node_zones[zone];
2901 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
2903 * There can be holes in boot-time mem_map[]s
2904 * handed to this function. They do not
2905 * exist on hotplugged memory.
2907 if (context == MEMMAP_EARLY) {
2908 if (!early_pfn_valid(pfn))
2910 if (!early_pfn_in_nid(pfn, nid))
2913 page = pfn_to_page(pfn);
2914 set_page_links(page, zone, nid, pfn);
2915 mminit_verify_page_links(page, zone, nid, pfn);
2916 init_page_count(page);
2917 reset_page_mapcount(page);
2918 SetPageReserved(page);
2920 * Mark the block movable so that blocks are reserved for
2921 * movable at startup. This will force kernel allocations
2922 * to reserve their blocks rather than leaking throughout
2923 * the address space during boot when many long-lived
2924 * kernel allocations are made. Later some blocks near
2925 * the start are marked MIGRATE_RESERVE by
2926 * setup_zone_migrate_reserve()
2928 * bitmap is created for zone's valid pfn range. but memmap
2929 * can be created for invalid pages (for alignment)
2930 * check here not to call set_pageblock_migratetype() against
2933 if ((z->zone_start_pfn <= pfn)
2934 && (pfn < z->zone_start_pfn + z->spanned_pages)
2935 && !(pfn & (pageblock_nr_pages - 1)))
2936 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
2938 INIT_LIST_HEAD(&page->lru);
2939 #ifdef WANT_PAGE_VIRTUAL
2940 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
2941 if (!is_highmem_idx(zone))
2942 set_page_address(page, __va(pfn << PAGE_SHIFT));
2947 static void __meminit zone_init_free_lists(struct zone *zone)
2950 for_each_migratetype_order(order, t) {
2951 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
2952 zone->free_area[order].nr_free = 0;
2956 #ifndef __HAVE_ARCH_MEMMAP_INIT
2957 #define memmap_init(size, nid, zone, start_pfn) \
2958 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
2961 static int zone_batchsize(struct zone *zone)
2967 * The per-cpu-pages pools are set to around 1000th of the
2968 * size of the zone. But no more than 1/2 of a meg.
2970 * OK, so we don't know how big the cache is. So guess.
2972 batch = zone->present_pages / 1024;
2973 if (batch * PAGE_SIZE > 512 * 1024)
2974 batch = (512 * 1024) / PAGE_SIZE;
2975 batch /= 4; /* We effectively *= 4 below */
2980 * Clamp the batch to a 2^n - 1 value. Having a power
2981 * of 2 value was found to be more likely to have
2982 * suboptimal cache aliasing properties in some cases.
2984 * For example if 2 tasks are alternately allocating
2985 * batches of pages, one task can end up with a lot
2986 * of pages of one half of the possible page colors
2987 * and the other with pages of the other colors.
2989 batch = rounddown_pow_of_two(batch + batch/2) - 1;
2994 /* The deferral and batching of frees should be suppressed under NOMMU
2997 * The problem is that NOMMU needs to be able to allocate large chunks
2998 * of contiguous memory as there's no hardware page translation to
2999 * assemble apparent contiguous memory from discontiguous pages.
3001 * Queueing large contiguous runs of pages for batching, however,
3002 * causes the pages to actually be freed in smaller chunks. As there
3003 * can be a significant delay between the individual batches being
3004 * recycled, this leads to the once large chunks of space being
3005 * fragmented and becoming unavailable for high-order allocations.
3011 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3013 struct per_cpu_pages *pcp;
3015 memset(p, 0, sizeof(*p));
3019 pcp->high = 6 * batch;
3020 pcp->batch = max(1UL, 1 * batch);
3021 INIT_LIST_HEAD(&pcp->list);
3025 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3026 * to the value high for the pageset p.
3029 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3032 struct per_cpu_pages *pcp;
3036 pcp->batch = max(1UL, high/4);
3037 if ((high/4) > (PAGE_SHIFT * 8))
3038 pcp->batch = PAGE_SHIFT * 8;
3044 * Boot pageset table. One per cpu which is going to be used for all
3045 * zones and all nodes. The parameters will be set in such a way
3046 * that an item put on a list will immediately be handed over to
3047 * the buddy list. This is safe since pageset manipulation is done
3048 * with interrupts disabled.
3050 * Some NUMA counter updates may also be caught by the boot pagesets.
3052 * The boot_pagesets must be kept even after bootup is complete for
3053 * unused processors and/or zones. They do play a role for bootstrapping
3054 * hotplugged processors.
3056 * zoneinfo_show() and maybe other functions do
3057 * not check if the processor is online before following the pageset pointer.
3058 * Other parts of the kernel may not check if the zone is available.
3060 static struct per_cpu_pageset boot_pageset[NR_CPUS];
3063 * Dynamically allocate memory for the
3064 * per cpu pageset array in struct zone.
3066 static int __cpuinit process_zones(int cpu)
3068 struct zone *zone, *dzone;
3069 int node = cpu_to_node(cpu);
3071 node_set_state(node, N_CPU); /* this node has a cpu */
3073 for_each_populated_zone(zone) {
3074 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
3076 if (!zone_pcp(zone, cpu))
3079 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
3081 if (percpu_pagelist_fraction)
3082 setup_pagelist_highmark(zone_pcp(zone, cpu),
3083 (zone->present_pages / percpu_pagelist_fraction));
3088 for_each_zone(dzone) {
3089 if (!populated_zone(dzone))
3093 kfree(zone_pcp(dzone, cpu));
3094 zone_pcp(dzone, cpu) = &boot_pageset[cpu];
3099 static inline void free_zone_pagesets(int cpu)
3103 for_each_zone(zone) {
3104 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
3106 /* Free per_cpu_pageset if it is slab allocated */
3107 if (pset != &boot_pageset[cpu])
3109 zone_pcp(zone, cpu) = &boot_pageset[cpu];
3113 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
3114 unsigned long action,
3117 int cpu = (long)hcpu;
3118 int ret = NOTIFY_OK;
3121 case CPU_UP_PREPARE:
3122 case CPU_UP_PREPARE_FROZEN:
3123 if (process_zones(cpu))
3126 case CPU_UP_CANCELED:
3127 case CPU_UP_CANCELED_FROZEN:
3129 case CPU_DEAD_FROZEN:
3130 free_zone_pagesets(cpu);
3138 static struct notifier_block __cpuinitdata pageset_notifier =
3139 { &pageset_cpuup_callback, NULL, 0 };
3141 void __init setup_per_cpu_pageset(void)
3145 /* Initialize per_cpu_pageset for cpu 0.
3146 * A cpuup callback will do this for every cpu
3147 * as it comes online
3149 err = process_zones(smp_processor_id());
3151 register_cpu_notifier(&pageset_notifier);
3156 static noinline __init_refok
3157 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3160 struct pglist_data *pgdat = zone->zone_pgdat;
3164 * The per-page waitqueue mechanism uses hashed waitqueues
3167 zone->wait_table_hash_nr_entries =
3168 wait_table_hash_nr_entries(zone_size_pages);
3169 zone->wait_table_bits =
3170 wait_table_bits(zone->wait_table_hash_nr_entries);
3171 alloc_size = zone->wait_table_hash_nr_entries
3172 * sizeof(wait_queue_head_t);
3174 if (!slab_is_available()) {
3175 zone->wait_table = (wait_queue_head_t *)
3176 alloc_bootmem_node(pgdat, alloc_size);
3179 * This case means that a zone whose size was 0 gets new memory
3180 * via memory hot-add.
3181 * But it may be the case that a new node was hot-added. In
3182 * this case vmalloc() will not be able to use this new node's
3183 * memory - this wait_table must be initialized to use this new
3184 * node itself as well.
3185 * To use this new node's memory, further consideration will be
3188 zone->wait_table = vmalloc(alloc_size);
3190 if (!zone->wait_table)
3193 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3194 init_waitqueue_head(zone->wait_table + i);
3199 static int __zone_pcp_update(void *data)
3201 struct zone *zone = data;
3203 unsigned long batch = zone_batchsize(zone), flags;
3205 for (cpu = 0; cpu < NR_CPUS; cpu++) {
3206 struct per_cpu_pageset *pset;
3207 struct per_cpu_pages *pcp;
3209 pset = zone_pcp(zone, cpu);
3212 local_irq_save(flags);
3213 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
3214 setup_pageset(pset, batch);
3215 local_irq_restore(flags);
3220 void zone_pcp_update(struct zone *zone)
3222 stop_machine(__zone_pcp_update, zone, NULL);
3225 static __meminit void zone_pcp_init(struct zone *zone)
3228 unsigned long batch = zone_batchsize(zone);
3230 for (cpu = 0; cpu < NR_CPUS; cpu++) {
3232 /* Early boot. Slab allocator not functional yet */
3233 zone_pcp(zone, cpu) = &boot_pageset[cpu];
3234 setup_pageset(&boot_pageset[cpu],0);
3236 setup_pageset(zone_pcp(zone,cpu), batch);
3239 if (zone->present_pages)
3240 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
3241 zone->name, zone->present_pages, batch);
3244 __meminit int init_currently_empty_zone(struct zone *zone,
3245 unsigned long zone_start_pfn,
3247 enum memmap_context context)
3249 struct pglist_data *pgdat = zone->zone_pgdat;
3251 ret = zone_wait_table_init(zone, size);
3254 pgdat->nr_zones = zone_idx(zone) + 1;
3256 zone->zone_start_pfn = zone_start_pfn;
3258 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3259 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3261 (unsigned long)zone_idx(zone),
3262 zone_start_pfn, (zone_start_pfn + size));
3264 zone_init_free_lists(zone);
3269 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3271 * Basic iterator support. Return the first range of PFNs for a node
3272 * Note: nid == MAX_NUMNODES returns first region regardless of node
3274 static int __meminit first_active_region_index_in_nid(int nid)
3278 for (i = 0; i < nr_nodemap_entries; i++)
3279 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3286 * Basic iterator support. Return the next active range of PFNs for a node
3287 * Note: nid == MAX_NUMNODES returns next region regardless of node
3289 static int __meminit next_active_region_index_in_nid(int index, int nid)
3291 for (index = index + 1; index < nr_nodemap_entries; index++)
3292 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3298 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3300 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3301 * Architectures may implement their own version but if add_active_range()
3302 * was used and there are no special requirements, this is a convenient
3305 int __meminit __early_pfn_to_nid(unsigned long pfn)
3309 for (i = 0; i < nr_nodemap_entries; i++) {
3310 unsigned long start_pfn = early_node_map[i].start_pfn;
3311 unsigned long end_pfn = early_node_map[i].end_pfn;
3313 if (start_pfn <= pfn && pfn < end_pfn)
3314 return early_node_map[i].nid;
3316 /* This is a memory hole */
3319 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3321 int __meminit early_pfn_to_nid(unsigned long pfn)
3325 nid = __early_pfn_to_nid(pfn);
3328 /* just returns 0 */
3332 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3333 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3337 nid = __early_pfn_to_nid(pfn);
3338 if (nid >= 0 && nid != node)
3344 /* Basic iterator support to walk early_node_map[] */
3345 #define for_each_active_range_index_in_nid(i, nid) \
3346 for (i = first_active_region_index_in_nid(nid); i != -1; \
3347 i = next_active_region_index_in_nid(i, nid))
3350 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3351 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3352 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3354 * If an architecture guarantees that all ranges registered with
3355 * add_active_ranges() contain no holes and may be freed, this
3356 * this function may be used instead of calling free_bootmem() manually.
3358 void __init free_bootmem_with_active_regions(int nid,
3359 unsigned long max_low_pfn)
3363 for_each_active_range_index_in_nid(i, nid) {
3364 unsigned long size_pages = 0;
3365 unsigned long end_pfn = early_node_map[i].end_pfn;
3367 if (early_node_map[i].start_pfn >= max_low_pfn)
3370 if (end_pfn > max_low_pfn)
3371 end_pfn = max_low_pfn;
3373 size_pages = end_pfn - early_node_map[i].start_pfn;
3374 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3375 PFN_PHYS(early_node_map[i].start_pfn),
3376 size_pages << PAGE_SHIFT);
3380 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3385 for_each_active_range_index_in_nid(i, nid) {
3386 ret = work_fn(early_node_map[i].start_pfn,
3387 early_node_map[i].end_pfn, data);
3393 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3394 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3396 * If an architecture guarantees that all ranges registered with
3397 * add_active_ranges() contain no holes and may be freed, this
3398 * function may be used instead of calling memory_present() manually.
3400 void __init sparse_memory_present_with_active_regions(int nid)
3404 for_each_active_range_index_in_nid(i, nid)
3405 memory_present(early_node_map[i].nid,
3406 early_node_map[i].start_pfn,
3407 early_node_map[i].end_pfn);
3411 * get_pfn_range_for_nid - Return the start and end page frames for a node
3412 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3413 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3414 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3416 * It returns the start and end page frame of a node based on information
3417 * provided by an arch calling add_active_range(). If called for a node
3418 * with no available memory, a warning is printed and the start and end
3421 void __meminit get_pfn_range_for_nid(unsigned int nid,
3422 unsigned long *start_pfn, unsigned long *end_pfn)
3428 for_each_active_range_index_in_nid(i, nid) {
3429 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3430 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3433 if (*start_pfn == -1UL)
3438 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3439 * assumption is made that zones within a node are ordered in monotonic
3440 * increasing memory addresses so that the "highest" populated zone is used
3442 static void __init find_usable_zone_for_movable(void)
3445 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3446 if (zone_index == ZONE_MOVABLE)
3449 if (arch_zone_highest_possible_pfn[zone_index] >
3450 arch_zone_lowest_possible_pfn[zone_index])
3454 VM_BUG_ON(zone_index == -1);
3455 movable_zone = zone_index;
3459 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3460 * because it is sized independant of architecture. Unlike the other zones,
3461 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3462 * in each node depending on the size of each node and how evenly kernelcore
3463 * is distributed. This helper function adjusts the zone ranges
3464 * provided by the architecture for a given node by using the end of the
3465 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3466 * zones within a node are in order of monotonic increases memory addresses
3468 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3469 unsigned long zone_type,
3470 unsigned long node_start_pfn,
3471 unsigned long node_end_pfn,
3472 unsigned long *zone_start_pfn,
3473 unsigned long *zone_end_pfn)
3475 /* Only adjust if ZONE_MOVABLE is on this node */
3476 if (zone_movable_pfn[nid]) {
3477 /* Size ZONE_MOVABLE */
3478 if (zone_type == ZONE_MOVABLE) {
3479 *zone_start_pfn = zone_movable_pfn[nid];
3480 *zone_end_pfn = min(node_end_pfn,
3481 arch_zone_highest_possible_pfn[movable_zone]);
3483 /* Adjust for ZONE_MOVABLE starting within this range */
3484 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3485 *zone_end_pfn > zone_movable_pfn[nid]) {
3486 *zone_end_pfn = zone_movable_pfn[nid];
3488 /* Check if this whole range is within ZONE_MOVABLE */
3489 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3490 *zone_start_pfn = *zone_end_pfn;
3495 * Return the number of pages a zone spans in a node, including holes
3496 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3498 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3499 unsigned long zone_type,
3500 unsigned long *ignored)
3502 unsigned long node_start_pfn, node_end_pfn;
3503 unsigned long zone_start_pfn, zone_end_pfn;
3505 /* Get the start and end of the node and zone */
3506 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3507 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3508 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3509 adjust_zone_range_for_zone_movable(nid, zone_type,
3510 node_start_pfn, node_end_pfn,
3511 &zone_start_pfn, &zone_end_pfn);
3513 /* Check that this node has pages within the zone's required range */
3514 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3517 /* Move the zone boundaries inside the node if necessary */
3518 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3519 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3521 /* Return the spanned pages */
3522 return zone_end_pfn - zone_start_pfn;
3526 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3527 * then all holes in the requested range will be accounted for.
3529 static unsigned long __meminit __absent_pages_in_range(int nid,
3530 unsigned long range_start_pfn,
3531 unsigned long range_end_pfn)
3534 unsigned long prev_end_pfn = 0, hole_pages = 0;
3535 unsigned long start_pfn;
3537 /* Find the end_pfn of the first active range of pfns in the node */
3538 i = first_active_region_index_in_nid(nid);
3542 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3544 /* Account for ranges before physical memory on this node */
3545 if (early_node_map[i].start_pfn > range_start_pfn)
3546 hole_pages = prev_end_pfn - range_start_pfn;
3548 /* Find all holes for the zone within the node */
3549 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3551 /* No need to continue if prev_end_pfn is outside the zone */
3552 if (prev_end_pfn >= range_end_pfn)
3555 /* Make sure the end of the zone is not within the hole */
3556 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3557 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3559 /* Update the hole size cound and move on */
3560 if (start_pfn > range_start_pfn) {
3561 BUG_ON(prev_end_pfn > start_pfn);
3562 hole_pages += start_pfn - prev_end_pfn;
3564 prev_end_pfn = early_node_map[i].end_pfn;
3567 /* Account for ranges past physical memory on this node */
3568 if (range_end_pfn > prev_end_pfn)
3569 hole_pages += range_end_pfn -
3570 max(range_start_pfn, prev_end_pfn);
3576 * absent_pages_in_range - Return number of page frames in holes within a range
3577 * @start_pfn: The start PFN to start searching for holes
3578 * @end_pfn: The end PFN to stop searching for holes
3580 * It returns the number of pages frames in memory holes within a range.
3582 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3583 unsigned long end_pfn)
3585 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3588 /* Return the number of page frames in holes in a zone on a node */
3589 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3590 unsigned long zone_type,
3591 unsigned long *ignored)
3593 unsigned long node_start_pfn, node_end_pfn;
3594 unsigned long zone_start_pfn, zone_end_pfn;
3596 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3597 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3599 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3602 adjust_zone_range_for_zone_movable(nid, zone_type,
3603 node_start_pfn, node_end_pfn,
3604 &zone_start_pfn, &zone_end_pfn);
3605 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3609 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3610 unsigned long zone_type,
3611 unsigned long *zones_size)
3613 return zones_size[zone_type];
3616 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3617 unsigned long zone_type,
3618 unsigned long *zholes_size)
3623 return zholes_size[zone_type];
3628 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3629 unsigned long *zones_size, unsigned long *zholes_size)
3631 unsigned long realtotalpages, totalpages = 0;
3634 for (i = 0; i < MAX_NR_ZONES; i++)
3635 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3637 pgdat->node_spanned_pages = totalpages;
3639 realtotalpages = totalpages;
3640 for (i = 0; i < MAX_NR_ZONES; i++)
3642 zone_absent_pages_in_node(pgdat->node_id, i,
3644 pgdat->node_present_pages = realtotalpages;
3645 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3649 #ifndef CONFIG_SPARSEMEM
3651 * Calculate the size of the zone->blockflags rounded to an unsigned long
3652 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3653 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3654 * round what is now in bits to nearest long in bits, then return it in
3657 static unsigned long __init usemap_size(unsigned long zonesize)
3659 unsigned long usemapsize;
3661 usemapsize = roundup(zonesize, pageblock_nr_pages);
3662 usemapsize = usemapsize >> pageblock_order;
3663 usemapsize *= NR_PAGEBLOCK_BITS;
3664 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
3666 return usemapsize / 8;
3669 static void __init setup_usemap(struct pglist_data *pgdat,
3670 struct zone *zone, unsigned long zonesize)
3672 unsigned long usemapsize = usemap_size(zonesize);
3673 zone->pageblock_flags = NULL;
3675 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
3678 static void inline setup_usemap(struct pglist_data *pgdat,
3679 struct zone *zone, unsigned long zonesize) {}
3680 #endif /* CONFIG_SPARSEMEM */
3682 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3684 /* Return a sensible default order for the pageblock size. */
3685 static inline int pageblock_default_order(void)
3687 if (HPAGE_SHIFT > PAGE_SHIFT)
3688 return HUGETLB_PAGE_ORDER;
3693 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3694 static inline void __init set_pageblock_order(unsigned int order)
3696 /* Check that pageblock_nr_pages has not already been setup */
3697 if (pageblock_order)
3701 * Assume the largest contiguous order of interest is a huge page.
3702 * This value may be variable depending on boot parameters on IA64
3704 pageblock_order = order;
3706 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3709 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3710 * and pageblock_default_order() are unused as pageblock_order is set
3711 * at compile-time. See include/linux/pageblock-flags.h for the values of
3712 * pageblock_order based on the kernel config
3714 static inline int pageblock_default_order(unsigned int order)
3718 #define set_pageblock_order(x) do {} while (0)
3720 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3723 * Set up the zone data structures:
3724 * - mark all pages reserved
3725 * - mark all memory queues empty
3726 * - clear the memory bitmaps
3728 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
3729 unsigned long *zones_size, unsigned long *zholes_size)
3732 int nid = pgdat->node_id;
3733 unsigned long zone_start_pfn = pgdat->node_start_pfn;
3736 pgdat_resize_init(pgdat);
3737 pgdat->nr_zones = 0;
3738 init_waitqueue_head(&pgdat->kswapd_wait);
3739 pgdat->kswapd_max_order = 0;
3740 pgdat_page_cgroup_init(pgdat);
3742 for (j = 0; j < MAX_NR_ZONES; j++) {
3743 struct zone *zone = pgdat->node_zones + j;
3744 unsigned long size, realsize, memmap_pages;
3747 size = zone_spanned_pages_in_node(nid, j, zones_size);
3748 realsize = size - zone_absent_pages_in_node(nid, j,
3752 * Adjust realsize so that it accounts for how much memory
3753 * is used by this zone for memmap. This affects the watermark
3754 * and per-cpu initialisations
3757 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
3758 if (realsize >= memmap_pages) {
3759 realsize -= memmap_pages;
3762 " %s zone: %lu pages used for memmap\n",
3763 zone_names[j], memmap_pages);
3766 " %s zone: %lu pages exceeds realsize %lu\n",
3767 zone_names[j], memmap_pages, realsize);
3769 /* Account for reserved pages */
3770 if (j == 0 && realsize > dma_reserve) {
3771 realsize -= dma_reserve;
3772 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
3773 zone_names[0], dma_reserve);
3776 if (!is_highmem_idx(j))
3777 nr_kernel_pages += realsize;
3778 nr_all_pages += realsize;
3780 zone->spanned_pages = size;
3781 zone->present_pages = realsize;
3784 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
3786 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
3788 zone->name = zone_names[j];
3789 spin_lock_init(&zone->lock);
3790 spin_lock_init(&zone->lru_lock);
3791 zone_seqlock_init(zone);
3792 zone->zone_pgdat = pgdat;
3794 zone->prev_priority = DEF_PRIORITY;
3796 zone_pcp_init(zone);
3798 INIT_LIST_HEAD(&zone->lru[l].list);
3799 zone->lru[l].nr_saved_scan = 0;
3801 zone->reclaim_stat.recent_rotated[0] = 0;
3802 zone->reclaim_stat.recent_rotated[1] = 0;
3803 zone->reclaim_stat.recent_scanned[0] = 0;
3804 zone->reclaim_stat.recent_scanned[1] = 0;
3805 zap_zone_vm_stats(zone);
3810 set_pageblock_order(pageblock_default_order());
3811 setup_usemap(pgdat, zone, size);
3812 ret = init_currently_empty_zone(zone, zone_start_pfn,
3813 size, MEMMAP_EARLY);
3815 memmap_init(size, nid, j, zone_start_pfn);
3816 zone_start_pfn += size;
3820 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
3822 /* Skip empty nodes */
3823 if (!pgdat->node_spanned_pages)
3826 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3827 /* ia64 gets its own node_mem_map, before this, without bootmem */
3828 if (!pgdat->node_mem_map) {
3829 unsigned long size, start, end;
3833 * The zone's endpoints aren't required to be MAX_ORDER
3834 * aligned but the node_mem_map endpoints must be in order
3835 * for the buddy allocator to function correctly.
3837 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
3838 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
3839 end = ALIGN(end, MAX_ORDER_NR_PAGES);
3840 size = (end - start) * sizeof(struct page);
3841 map = alloc_remap(pgdat->node_id, size);
3843 map = alloc_bootmem_node(pgdat, size);
3844 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
3846 #ifndef CONFIG_NEED_MULTIPLE_NODES
3848 * With no DISCONTIG, the global mem_map is just set as node 0's
3850 if (pgdat == NODE_DATA(0)) {
3851 mem_map = NODE_DATA(0)->node_mem_map;
3852 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3853 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
3854 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
3855 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
3858 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
3861 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
3862 unsigned long node_start_pfn, unsigned long *zholes_size)
3864 pg_data_t *pgdat = NODE_DATA(nid);
3866 pgdat->node_id = nid;
3867 pgdat->node_start_pfn = node_start_pfn;
3868 calculate_node_totalpages(pgdat, zones_size, zholes_size);
3870 alloc_node_mem_map(pgdat);
3871 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3872 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
3873 nid, (unsigned long)pgdat,
3874 (unsigned long)pgdat->node_mem_map);
3877 free_area_init_core(pgdat, zones_size, zholes_size);
3880 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3882 #if MAX_NUMNODES > 1
3884 * Figure out the number of possible node ids.
3886 static void __init setup_nr_node_ids(void)
3889 unsigned int highest = 0;
3891 for_each_node_mask(node, node_possible_map)
3893 nr_node_ids = highest + 1;
3896 static inline void setup_nr_node_ids(void)
3902 * add_active_range - Register a range of PFNs backed by physical memory
3903 * @nid: The node ID the range resides on
3904 * @start_pfn: The start PFN of the available physical memory
3905 * @end_pfn: The end PFN of the available physical memory
3907 * These ranges are stored in an early_node_map[] and later used by
3908 * free_area_init_nodes() to calculate zone sizes and holes. If the
3909 * range spans a memory hole, it is up to the architecture to ensure
3910 * the memory is not freed by the bootmem allocator. If possible
3911 * the range being registered will be merged with existing ranges.
3913 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
3914 unsigned long end_pfn)
3918 mminit_dprintk(MMINIT_TRACE, "memory_register",
3919 "Entering add_active_range(%d, %#lx, %#lx) "
3920 "%d entries of %d used\n",
3921 nid, start_pfn, end_pfn,
3922 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
3924 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
3926 /* Merge with existing active regions if possible */
3927 for (i = 0; i < nr_nodemap_entries; i++) {
3928 if (early_node_map[i].nid != nid)
3931 /* Skip if an existing region covers this new one */
3932 if (start_pfn >= early_node_map[i].start_pfn &&
3933 end_pfn <= early_node_map[i].end_pfn)
3936 /* Merge forward if suitable */
3937 if (start_pfn <= early_node_map[i].end_pfn &&
3938 end_pfn > early_node_map[i].end_pfn) {
3939 early_node_map[i].end_pfn = end_pfn;
3943 /* Merge backward if suitable */
3944 if (start_pfn < early_node_map[i].end_pfn &&
3945 end_pfn >= early_node_map[i].start_pfn) {
3946 early_node_map[i].start_pfn = start_pfn;
3951 /* Check that early_node_map is large enough */
3952 if (i >= MAX_ACTIVE_REGIONS) {
3953 printk(KERN_CRIT "More than %d memory regions, truncating\n",
3954 MAX_ACTIVE_REGIONS);
3958 early_node_map[i].nid = nid;
3959 early_node_map[i].start_pfn = start_pfn;
3960 early_node_map[i].end_pfn = end_pfn;
3961 nr_nodemap_entries = i + 1;
3965 * remove_active_range - Shrink an existing registered range of PFNs
3966 * @nid: The node id the range is on that should be shrunk
3967 * @start_pfn: The new PFN of the range
3968 * @end_pfn: The new PFN of the range
3970 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
3971 * The map is kept near the end physical page range that has already been
3972 * registered. This function allows an arch to shrink an existing registered
3975 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
3976 unsigned long end_pfn)
3981 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
3982 nid, start_pfn, end_pfn);
3984 /* Find the old active region end and shrink */
3985 for_each_active_range_index_in_nid(i, nid) {
3986 if (early_node_map[i].start_pfn >= start_pfn &&
3987 early_node_map[i].end_pfn <= end_pfn) {
3989 early_node_map[i].start_pfn = 0;
3990 early_node_map[i].end_pfn = 0;
3994 if (early_node_map[i].start_pfn < start_pfn &&
3995 early_node_map[i].end_pfn > start_pfn) {
3996 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
3997 early_node_map[i].end_pfn = start_pfn;
3998 if (temp_end_pfn > end_pfn)
3999 add_active_range(nid, end_pfn, temp_end_pfn);
4002 if (early_node_map[i].start_pfn >= start_pfn &&
4003 early_node_map[i].end_pfn > end_pfn &&
4004 early_node_map[i].start_pfn < end_pfn) {
4005 early_node_map[i].start_pfn = end_pfn;
4013 /* remove the blank ones */
4014 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4015 if (early_node_map[i].nid != nid)
4017 if (early_node_map[i].end_pfn)
4019 /* we found it, get rid of it */
4020 for (j = i; j < nr_nodemap_entries - 1; j++)
4021 memcpy(&early_node_map[j], &early_node_map[j+1],
4022 sizeof(early_node_map[j]));
4023 j = nr_nodemap_entries - 1;
4024 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4025 nr_nodemap_entries--;
4030 * remove_all_active_ranges - Remove all currently registered regions
4032 * During discovery, it may be found that a table like SRAT is invalid
4033 * and an alternative discovery method must be used. This function removes
4034 * all currently registered regions.
4036 void __init remove_all_active_ranges(void)
4038 memset(early_node_map, 0, sizeof(early_node_map));
4039 nr_nodemap_entries = 0;
4042 /* Compare two active node_active_regions */
4043 static int __init cmp_node_active_region(const void *a, const void *b)
4045 struct node_active_region *arange = (struct node_active_region *)a;
4046 struct node_active_region *brange = (struct node_active_region *)b;
4048 /* Done this way to avoid overflows */
4049 if (arange->start_pfn > brange->start_pfn)
4051 if (arange->start_pfn < brange->start_pfn)
4057 /* sort the node_map by start_pfn */
4058 static void __init sort_node_map(void)
4060 sort(early_node_map, (size_t)nr_nodemap_entries,
4061 sizeof(struct node_active_region),
4062 cmp_node_active_region, NULL);
4065 /* Find the lowest pfn for a node */
4066 static unsigned long __init find_min_pfn_for_node(int nid)
4069 unsigned long min_pfn = ULONG_MAX;
4071 /* Assuming a sorted map, the first range found has the starting pfn */
4072 for_each_active_range_index_in_nid(i, nid)
4073 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4075 if (min_pfn == ULONG_MAX) {
4077 "Could not find start_pfn for node %d\n", nid);
4085 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4087 * It returns the minimum PFN based on information provided via
4088 * add_active_range().
4090 unsigned long __init find_min_pfn_with_active_regions(void)
4092 return find_min_pfn_for_node(MAX_NUMNODES);
4096 * early_calculate_totalpages()
4097 * Sum pages in active regions for movable zone.
4098 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4100 static unsigned long __init early_calculate_totalpages(void)
4103 unsigned long totalpages = 0;
4105 for (i = 0; i < nr_nodemap_entries; i++) {
4106 unsigned long pages = early_node_map[i].end_pfn -
4107 early_node_map[i].start_pfn;
4108 totalpages += pages;
4110 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4116 * Find the PFN the Movable zone begins in each node. Kernel memory
4117 * is spread evenly between nodes as long as the nodes have enough
4118 * memory. When they don't, some nodes will have more kernelcore than
4121 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4124 unsigned long usable_startpfn;
4125 unsigned long kernelcore_node, kernelcore_remaining;
4126 /* save the state before borrow the nodemask */
4127 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4128 unsigned long totalpages = early_calculate_totalpages();
4129 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4132 * If movablecore was specified, calculate what size of
4133 * kernelcore that corresponds so that memory usable for
4134 * any allocation type is evenly spread. If both kernelcore
4135 * and movablecore are specified, then the value of kernelcore
4136 * will be used for required_kernelcore if it's greater than
4137 * what movablecore would have allowed.
4139 if (required_movablecore) {
4140 unsigned long corepages;
4143 * Round-up so that ZONE_MOVABLE is at least as large as what
4144 * was requested by the user
4146 required_movablecore =
4147 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4148 corepages = totalpages - required_movablecore;
4150 required_kernelcore = max(required_kernelcore, corepages);
4153 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4154 if (!required_kernelcore)
4157 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4158 find_usable_zone_for_movable();
4159 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4162 /* Spread kernelcore memory as evenly as possible throughout nodes */
4163 kernelcore_node = required_kernelcore / usable_nodes;
4164 for_each_node_state(nid, N_HIGH_MEMORY) {
4166 * Recalculate kernelcore_node if the division per node
4167 * now exceeds what is necessary to satisfy the requested
4168 * amount of memory for the kernel
4170 if (required_kernelcore < kernelcore_node)
4171 kernelcore_node = required_kernelcore / usable_nodes;
4174 * As the map is walked, we track how much memory is usable
4175 * by the kernel using kernelcore_remaining. When it is
4176 * 0, the rest of the node is usable by ZONE_MOVABLE
4178 kernelcore_remaining = kernelcore_node;
4180 /* Go through each range of PFNs within this node */
4181 for_each_active_range_index_in_nid(i, nid) {
4182 unsigned long start_pfn, end_pfn;
4183 unsigned long size_pages;
4185 start_pfn = max(early_node_map[i].start_pfn,
4186 zone_movable_pfn[nid]);
4187 end_pfn = early_node_map[i].end_pfn;
4188 if (start_pfn >= end_pfn)
4191 /* Account for what is only usable for kernelcore */
4192 if (start_pfn < usable_startpfn) {
4193 unsigned long kernel_pages;
4194 kernel_pages = min(end_pfn, usable_startpfn)
4197 kernelcore_remaining -= min(kernel_pages,
4198 kernelcore_remaining);
4199 required_kernelcore -= min(kernel_pages,
4200 required_kernelcore);
4202 /* Continue if range is now fully accounted */
4203 if (end_pfn <= usable_startpfn) {
4206 * Push zone_movable_pfn to the end so
4207 * that if we have to rebalance
4208 * kernelcore across nodes, we will
4209 * not double account here
4211 zone_movable_pfn[nid] = end_pfn;
4214 start_pfn = usable_startpfn;
4218 * The usable PFN range for ZONE_MOVABLE is from
4219 * start_pfn->end_pfn. Calculate size_pages as the
4220 * number of pages used as kernelcore
4222 size_pages = end_pfn - start_pfn;
4223 if (size_pages > kernelcore_remaining)
4224 size_pages = kernelcore_remaining;
4225 zone_movable_pfn[nid] = start_pfn + size_pages;
4228 * Some kernelcore has been met, update counts and
4229 * break if the kernelcore for this node has been
4232 required_kernelcore -= min(required_kernelcore,
4234 kernelcore_remaining -= size_pages;
4235 if (!kernelcore_remaining)
4241 * If there is still required_kernelcore, we do another pass with one
4242 * less node in the count. This will push zone_movable_pfn[nid] further
4243 * along on the nodes that still have memory until kernelcore is
4247 if (usable_nodes && required_kernelcore > usable_nodes)
4250 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4251 for (nid = 0; nid < MAX_NUMNODES; nid++)
4252 zone_movable_pfn[nid] =
4253 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4256 /* restore the node_state */
4257 node_states[N_HIGH_MEMORY] = saved_node_state;
4260 /* Any regular memory on that node ? */
4261 static void check_for_regular_memory(pg_data_t *pgdat)
4263 #ifdef CONFIG_HIGHMEM
4264 enum zone_type zone_type;
4266 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4267 struct zone *zone = &pgdat->node_zones[zone_type];
4268 if (zone->present_pages)
4269 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4275 * free_area_init_nodes - Initialise all pg_data_t and zone data
4276 * @max_zone_pfn: an array of max PFNs for each zone
4278 * This will call free_area_init_node() for each active node in the system.
4279 * Using the page ranges provided by add_active_range(), the size of each
4280 * zone in each node and their holes is calculated. If the maximum PFN
4281 * between two adjacent zones match, it is assumed that the zone is empty.
4282 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4283 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4284 * starts where the previous one ended. For example, ZONE_DMA32 starts
4285 * at arch_max_dma_pfn.
4287 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4292 /* Sort early_node_map as initialisation assumes it is sorted */
4295 /* Record where the zone boundaries are */
4296 memset(arch_zone_lowest_possible_pfn, 0,
4297 sizeof(arch_zone_lowest_possible_pfn));
4298 memset(arch_zone_highest_possible_pfn, 0,
4299 sizeof(arch_zone_highest_possible_pfn));
4300 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4301 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4302 for (i = 1; i < MAX_NR_ZONES; i++) {
4303 if (i == ZONE_MOVABLE)
4305 arch_zone_lowest_possible_pfn[i] =
4306 arch_zone_highest_possible_pfn[i-1];
4307 arch_zone_highest_possible_pfn[i] =
4308 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4310 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4311 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4313 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4314 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4315 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4317 /* Print out the zone ranges */
4318 printk("Zone PFN ranges:\n");
4319 for (i = 0; i < MAX_NR_ZONES; i++) {
4320 if (i == ZONE_MOVABLE)
4322 printk(" %-8s %0#10lx -> %0#10lx\n",
4324 arch_zone_lowest_possible_pfn[i],
4325 arch_zone_highest_possible_pfn[i]);
4328 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4329 printk("Movable zone start PFN for each node\n");
4330 for (i = 0; i < MAX_NUMNODES; i++) {
4331 if (zone_movable_pfn[i])
4332 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4335 /* Print out the early_node_map[] */
4336 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4337 for (i = 0; i < nr_nodemap_entries; i++)
4338 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4339 early_node_map[i].start_pfn,
4340 early_node_map[i].end_pfn);
4342 /* Initialise every node */
4343 mminit_verify_pageflags_layout();
4344 setup_nr_node_ids();
4345 for_each_online_node(nid) {
4346 pg_data_t *pgdat = NODE_DATA(nid);
4347 free_area_init_node(nid, NULL,
4348 find_min_pfn_for_node(nid), NULL);
4350 /* Any memory on that node */
4351 if (pgdat->node_present_pages)
4352 node_set_state(nid, N_HIGH_MEMORY);
4353 check_for_regular_memory(pgdat);
4357 static int __init cmdline_parse_core(char *p, unsigned long *core)
4359 unsigned long long coremem;
4363 coremem = memparse(p, &p);
4364 *core = coremem >> PAGE_SHIFT;
4366 /* Paranoid check that UL is enough for the coremem value */
4367 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4373 * kernelcore=size sets the amount of memory for use for allocations that
4374 * cannot be reclaimed or migrated.
4376 static int __init cmdline_parse_kernelcore(char *p)
4378 return cmdline_parse_core(p, &required_kernelcore);
4382 * movablecore=size sets the amount of memory for use for allocations that
4383 * can be reclaimed or migrated.
4385 static int __init cmdline_parse_movablecore(char *p)
4387 return cmdline_parse_core(p, &required_movablecore);
4390 early_param("kernelcore", cmdline_parse_kernelcore);
4391 early_param("movablecore", cmdline_parse_movablecore);
4393 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4396 * set_dma_reserve - set the specified number of pages reserved in the first zone
4397 * @new_dma_reserve: The number of pages to mark reserved
4399 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4400 * In the DMA zone, a significant percentage may be consumed by kernel image
4401 * and other unfreeable allocations which can skew the watermarks badly. This
4402 * function may optionally be used to account for unfreeable pages in the
4403 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4404 * smaller per-cpu batchsize.
4406 void __init set_dma_reserve(unsigned long new_dma_reserve)
4408 dma_reserve = new_dma_reserve;
4411 #ifndef CONFIG_NEED_MULTIPLE_NODES
4412 struct pglist_data __refdata contig_page_data = { .bdata = &bootmem_node_data[0] };
4413 EXPORT_SYMBOL(contig_page_data);
4416 void __init free_area_init(unsigned long *zones_size)
4418 free_area_init_node(0, zones_size,
4419 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4422 static int page_alloc_cpu_notify(struct notifier_block *self,
4423 unsigned long action, void *hcpu)
4425 int cpu = (unsigned long)hcpu;
4427 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4431 * Spill the event counters of the dead processor
4432 * into the current processors event counters.
4433 * This artificially elevates the count of the current
4436 vm_events_fold_cpu(cpu);
4439 * Zero the differential counters of the dead processor
4440 * so that the vm statistics are consistent.
4442 * This is only okay since the processor is dead and cannot
4443 * race with what we are doing.
4445 refresh_cpu_vm_stats(cpu);
4450 void __init page_alloc_init(void)
4452 hotcpu_notifier(page_alloc_cpu_notify, 0);
4456 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4457 * or min_free_kbytes changes.
4459 static void calculate_totalreserve_pages(void)
4461 struct pglist_data *pgdat;
4462 unsigned long reserve_pages = 0;
4463 enum zone_type i, j;
4465 for_each_online_pgdat(pgdat) {
4466 for (i = 0; i < MAX_NR_ZONES; i++) {
4467 struct zone *zone = pgdat->node_zones + i;
4468 unsigned long max = 0;
4470 /* Find valid and maximum lowmem_reserve in the zone */
4471 for (j = i; j < MAX_NR_ZONES; j++) {
4472 if (zone->lowmem_reserve[j] > max)
4473 max = zone->lowmem_reserve[j];
4476 /* we treat the high watermark as reserved pages. */
4477 max += high_wmark_pages(zone);
4479 if (max > zone->present_pages)
4480 max = zone->present_pages;
4481 reserve_pages += max;
4484 totalreserve_pages = reserve_pages;
4488 * setup_per_zone_lowmem_reserve - called whenever
4489 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4490 * has a correct pages reserved value, so an adequate number of
4491 * pages are left in the zone after a successful __alloc_pages().
4493 static void setup_per_zone_lowmem_reserve(void)
4495 struct pglist_data *pgdat;
4496 enum zone_type j, idx;
4498 for_each_online_pgdat(pgdat) {
4499 for (j = 0; j < MAX_NR_ZONES; j++) {
4500 struct zone *zone = pgdat->node_zones + j;
4501 unsigned long present_pages = zone->present_pages;
4503 zone->lowmem_reserve[j] = 0;
4507 struct zone *lower_zone;
4511 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4512 sysctl_lowmem_reserve_ratio[idx] = 1;
4514 lower_zone = pgdat->node_zones + idx;
4515 lower_zone->lowmem_reserve[j] = present_pages /
4516 sysctl_lowmem_reserve_ratio[idx];
4517 present_pages += lower_zone->present_pages;
4522 /* update totalreserve_pages */
4523 calculate_totalreserve_pages();
4527 * setup_per_zone_wmarks - called when min_free_kbytes changes
4528 * or when memory is hot-{added|removed}
4530 * Ensures that the watermark[min,low,high] values for each zone are set
4531 * correctly with respect to min_free_kbytes.
4533 void setup_per_zone_wmarks(void)
4535 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4536 unsigned long lowmem_pages = 0;
4538 unsigned long flags;
4540 /* Calculate total number of !ZONE_HIGHMEM pages */
4541 for_each_zone(zone) {
4542 if (!is_highmem(zone))
4543 lowmem_pages += zone->present_pages;
4546 for_each_zone(zone) {
4549 spin_lock_irqsave(&zone->lock, flags);
4550 tmp = (u64)pages_min * zone->present_pages;
4551 do_div(tmp, lowmem_pages);
4552 if (is_highmem(zone)) {
4554 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4555 * need highmem pages, so cap pages_min to a small
4558 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4559 * deltas controls asynch page reclaim, and so should
4560 * not be capped for highmem.
4564 min_pages = zone->present_pages / 1024;
4565 if (min_pages < SWAP_CLUSTER_MAX)
4566 min_pages = SWAP_CLUSTER_MAX;
4567 if (min_pages > 128)
4569 zone->watermark[WMARK_MIN] = min_pages;
4572 * If it's a lowmem zone, reserve a number of pages
4573 * proportionate to the zone's size.
4575 zone->watermark[WMARK_MIN] = tmp;
4578 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
4579 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
4580 setup_zone_migrate_reserve(zone);
4581 spin_unlock_irqrestore(&zone->lock, flags);
4584 /* update totalreserve_pages */
4585 calculate_totalreserve_pages();
4589 * The inactive anon list should be small enough that the VM never has to
4590 * do too much work, but large enough that each inactive page has a chance
4591 * to be referenced again before it is swapped out.
4593 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4594 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4595 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4596 * the anonymous pages are kept on the inactive list.
4599 * memory ratio inactive anon
4600 * -------------------------------------
4609 void calculate_zone_inactive_ratio(struct zone *zone)
4611 unsigned int gb, ratio;
4613 /* Zone size in gigabytes */
4614 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4616 ratio = int_sqrt(10 * gb);
4620 zone->inactive_ratio = ratio;
4623 static void __init setup_per_zone_inactive_ratio(void)
4628 calculate_zone_inactive_ratio(zone);
4632 * Initialise min_free_kbytes.
4634 * For small machines we want it small (128k min). For large machines
4635 * we want it large (64MB max). But it is not linear, because network
4636 * bandwidth does not increase linearly with machine size. We use
4638 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4639 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4655 static int __init init_per_zone_wmark_min(void)
4657 unsigned long lowmem_kbytes;
4659 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4661 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4662 if (min_free_kbytes < 128)
4663 min_free_kbytes = 128;
4664 if (min_free_kbytes > 65536)
4665 min_free_kbytes = 65536;
4666 setup_per_zone_wmarks();
4667 setup_per_zone_lowmem_reserve();
4668 setup_per_zone_inactive_ratio();
4671 module_init(init_per_zone_wmark_min)
4674 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4675 * that we can call two helper functions whenever min_free_kbytes
4678 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
4679 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4681 proc_dointvec(table, write, file, buffer, length, ppos);
4683 setup_per_zone_wmarks();
4688 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
4689 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4694 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4699 zone->min_unmapped_pages = (zone->present_pages *
4700 sysctl_min_unmapped_ratio) / 100;
4704 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
4705 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4710 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4715 zone->min_slab_pages = (zone->present_pages *
4716 sysctl_min_slab_ratio) / 100;
4722 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4723 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4724 * whenever sysctl_lowmem_reserve_ratio changes.
4726 * The reserve ratio obviously has absolutely no relation with the
4727 * minimum watermarks. The lowmem reserve ratio can only make sense
4728 * if in function of the boot time zone sizes.
4730 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
4731 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4733 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4734 setup_per_zone_lowmem_reserve();
4739 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4740 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4741 * can have before it gets flushed back to buddy allocator.
4744 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
4745 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
4751 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
4752 if (!write || (ret == -EINVAL))
4754 for_each_populated_zone(zone) {
4755 for_each_online_cpu(cpu) {
4757 high = zone->present_pages / percpu_pagelist_fraction;
4758 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
4764 int hashdist = HASHDIST_DEFAULT;
4767 static int __init set_hashdist(char *str)
4771 hashdist = simple_strtoul(str, &str, 0);
4774 __setup("hashdist=", set_hashdist);
4778 * allocate a large system hash table from bootmem
4779 * - it is assumed that the hash table must contain an exact power-of-2
4780 * quantity of entries
4781 * - limit is the number of hash buckets, not the total allocation size
4783 void *__init alloc_large_system_hash(const char *tablename,
4784 unsigned long bucketsize,
4785 unsigned long numentries,
4788 unsigned int *_hash_shift,
4789 unsigned int *_hash_mask,
4790 unsigned long limit)
4792 unsigned long long max = limit;
4793 unsigned long log2qty, size;
4796 /* allow the kernel cmdline to have a say */
4798 /* round applicable memory size up to nearest megabyte */
4799 numentries = nr_kernel_pages;
4800 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
4801 numentries >>= 20 - PAGE_SHIFT;
4802 numentries <<= 20 - PAGE_SHIFT;
4804 /* limit to 1 bucket per 2^scale bytes of low memory */
4805 if (scale > PAGE_SHIFT)
4806 numentries >>= (scale - PAGE_SHIFT);
4808 numentries <<= (PAGE_SHIFT - scale);
4810 /* Make sure we've got at least a 0-order allocation.. */
4811 if (unlikely((numentries * bucketsize) < PAGE_SIZE))
4812 numentries = PAGE_SIZE / bucketsize;
4814 numentries = roundup_pow_of_two(numentries);
4816 /* limit allocation size to 1/16 total memory by default */
4818 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
4819 do_div(max, bucketsize);
4822 if (numentries > max)
4825 log2qty = ilog2(numentries);
4828 size = bucketsize << log2qty;
4829 if (flags & HASH_EARLY)
4830 table = alloc_bootmem_nopanic(size);
4832 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
4835 * If bucketsize is not a power-of-two, we may free
4836 * some pages at the end of hash table which
4837 * alloc_pages_exact() automatically does
4839 if (get_order(size) < MAX_ORDER) {
4840 table = alloc_pages_exact(size, GFP_ATOMIC);
4841 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
4844 } while (!table && size > PAGE_SIZE && --log2qty);
4847 panic("Failed to allocate %s hash table\n", tablename);
4849 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
4852 ilog2(size) - PAGE_SHIFT,
4856 *_hash_shift = log2qty;
4858 *_hash_mask = (1 << log2qty) - 1;
4863 /* Return a pointer to the bitmap storing bits affecting a block of pages */
4864 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
4867 #ifdef CONFIG_SPARSEMEM
4868 return __pfn_to_section(pfn)->pageblock_flags;
4870 return zone->pageblock_flags;
4871 #endif /* CONFIG_SPARSEMEM */
4874 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
4876 #ifdef CONFIG_SPARSEMEM
4877 pfn &= (PAGES_PER_SECTION-1);
4878 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4880 pfn = pfn - zone->zone_start_pfn;
4881 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
4882 #endif /* CONFIG_SPARSEMEM */
4886 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
4887 * @page: The page within the block of interest
4888 * @start_bitidx: The first bit of interest to retrieve
4889 * @end_bitidx: The last bit of interest
4890 * returns pageblock_bits flags
4892 unsigned long get_pageblock_flags_group(struct page *page,
4893 int start_bitidx, int end_bitidx)
4896 unsigned long *bitmap;
4897 unsigned long pfn, bitidx;
4898 unsigned long flags = 0;
4899 unsigned long value = 1;
4901 zone = page_zone(page);
4902 pfn = page_to_pfn(page);
4903 bitmap = get_pageblock_bitmap(zone, pfn);
4904 bitidx = pfn_to_bitidx(zone, pfn);
4906 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4907 if (test_bit(bitidx + start_bitidx, bitmap))
4914 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
4915 * @page: The page within the block of interest
4916 * @start_bitidx: The first bit of interest
4917 * @end_bitidx: The last bit of interest
4918 * @flags: The flags to set
4920 void set_pageblock_flags_group(struct page *page, unsigned long flags,
4921 int start_bitidx, int end_bitidx)
4924 unsigned long *bitmap;
4925 unsigned long pfn, bitidx;
4926 unsigned long value = 1;
4928 zone = page_zone(page);
4929 pfn = page_to_pfn(page);
4930 bitmap = get_pageblock_bitmap(zone, pfn);
4931 bitidx = pfn_to_bitidx(zone, pfn);
4932 VM_BUG_ON(pfn < zone->zone_start_pfn);
4933 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
4935 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
4937 __set_bit(bitidx + start_bitidx, bitmap);
4939 __clear_bit(bitidx + start_bitidx, bitmap);
4943 * This is designed as sub function...plz see page_isolation.c also.
4944 * set/clear page block's type to be ISOLATE.
4945 * page allocater never alloc memory from ISOLATE block.
4948 int set_migratetype_isolate(struct page *page)
4951 unsigned long flags;
4955 zone = page_zone(page);
4956 zone_idx = zone_idx(zone);
4957 spin_lock_irqsave(&zone->lock, flags);
4959 * In future, more migrate types will be able to be isolation target.
4961 if (get_pageblock_migratetype(page) != MIGRATE_MOVABLE &&
4962 zone_idx != ZONE_MOVABLE)
4964 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
4965 move_freepages_block(zone, page, MIGRATE_ISOLATE);
4968 spin_unlock_irqrestore(&zone->lock, flags);
4974 void unset_migratetype_isolate(struct page *page)
4977 unsigned long flags;
4978 zone = page_zone(page);
4979 spin_lock_irqsave(&zone->lock, flags);
4980 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
4982 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4983 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4985 spin_unlock_irqrestore(&zone->lock, flags);
4988 #ifdef CONFIG_MEMORY_HOTREMOVE
4990 * All pages in the range must be isolated before calling this.
4993 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
4999 unsigned long flags;
5000 /* find the first valid pfn */
5001 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5006 zone = page_zone(pfn_to_page(pfn));
5007 spin_lock_irqsave(&zone->lock, flags);
5009 while (pfn < end_pfn) {
5010 if (!pfn_valid(pfn)) {
5014 page = pfn_to_page(pfn);
5015 BUG_ON(page_count(page));
5016 BUG_ON(!PageBuddy(page));
5017 order = page_order(page);
5018 #ifdef CONFIG_DEBUG_VM
5019 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5020 pfn, 1 << order, end_pfn);
5022 list_del(&page->lru);
5023 rmv_page_order(page);
5024 zone->free_area[order].nr_free--;
5025 __mod_zone_page_state(zone, NR_FREE_PAGES,
5027 for (i = 0; i < (1 << order); i++)
5028 SetPageReserved((page+i));
5029 pfn += (1 << order);
5031 spin_unlock_irqrestore(&zone->lock, flags);