2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/compiler.h>
25 #include <linux/kernel.h>
26 #include <linux/kmemcheck.h>
27 #include <linux/module.h>
28 #include <linux/suspend.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/slab.h>
32 #include <linux/oom.h>
33 #include <linux/notifier.h>
34 #include <linux/topology.h>
35 #include <linux/sysctl.h>
36 #include <linux/cpu.h>
37 #include <linux/cpuset.h>
38 #include <linux/memory_hotplug.h>
39 #include <linux/nodemask.h>
40 #include <linux/vmalloc.h>
41 #include <linux/mempolicy.h>
42 #include <linux/stop_machine.h>
43 #include <linux/sort.h>
44 #include <linux/pfn.h>
45 #include <linux/backing-dev.h>
46 #include <linux/fault-inject.h>
47 #include <linux/page-isolation.h>
48 #include <linux/page_cgroup.h>
49 #include <linux/debugobjects.h>
50 #include <linux/kmemleak.h>
51 #include <linux/memory.h>
52 #include <trace/events/kmem.h>
53 #include <linux/ftrace_event.h>
55 #include <asm/tlbflush.h>
56 #include <asm/div64.h>
60 * Array of node states.
62 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
63 [N_POSSIBLE] = NODE_MASK_ALL,
64 [N_ONLINE] = { { [0] = 1UL } },
66 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
68 [N_HIGH_MEMORY] = { { [0] = 1UL } },
70 [N_CPU] = { { [0] = 1UL } },
73 EXPORT_SYMBOL(node_states);
75 unsigned long totalram_pages __read_mostly;
76 unsigned long totalreserve_pages __read_mostly;
77 int percpu_pagelist_fraction;
78 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
80 #ifdef CONFIG_PM_SLEEP
82 * The following functions are used by the suspend/hibernate code to temporarily
83 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
84 * while devices are suspended. To avoid races with the suspend/hibernate code,
85 * they should always be called with pm_mutex held (gfp_allowed_mask also should
86 * only be modified with pm_mutex held, unless the suspend/hibernate code is
87 * guaranteed not to run in parallel with that modification).
89 void set_gfp_allowed_mask(gfp_t mask)
91 WARN_ON(!mutex_is_locked(&pm_mutex));
92 gfp_allowed_mask = mask;
95 gfp_t clear_gfp_allowed_mask(gfp_t mask)
97 gfp_t ret = gfp_allowed_mask;
99 WARN_ON(!mutex_is_locked(&pm_mutex));
100 gfp_allowed_mask &= ~mask;
103 #endif /* CONFIG_PM_SLEEP */
105 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
106 int pageblock_order __read_mostly;
109 static void __free_pages_ok(struct page *page, unsigned int order);
112 * results with 256, 32 in the lowmem_reserve sysctl:
113 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
114 * 1G machine -> (16M dma, 784M normal, 224M high)
115 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
116 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
117 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
119 * TBD: should special case ZONE_DMA32 machines here - in those we normally
120 * don't need any ZONE_NORMAL reservation
122 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
123 #ifdef CONFIG_ZONE_DMA
126 #ifdef CONFIG_ZONE_DMA32
129 #ifdef CONFIG_HIGHMEM
135 EXPORT_SYMBOL(totalram_pages);
137 static char * const zone_names[MAX_NR_ZONES] = {
138 #ifdef CONFIG_ZONE_DMA
141 #ifdef CONFIG_ZONE_DMA32
145 #ifdef CONFIG_HIGHMEM
151 int min_free_kbytes = 1024;
153 static unsigned long __meminitdata nr_kernel_pages;
154 static unsigned long __meminitdata nr_all_pages;
155 static unsigned long __meminitdata dma_reserve;
157 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
159 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
160 * ranges of memory (RAM) that may be registered with add_active_range().
161 * Ranges passed to add_active_range() will be merged if possible
162 * so the number of times add_active_range() can be called is
163 * related to the number of nodes and the number of holes
165 #ifdef CONFIG_MAX_ACTIVE_REGIONS
166 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
167 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
169 #if MAX_NUMNODES >= 32
170 /* If there can be many nodes, allow up to 50 holes per node */
171 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
173 /* By default, allow up to 256 distinct regions */
174 #define MAX_ACTIVE_REGIONS 256
178 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
179 static int __meminitdata nr_nodemap_entries;
180 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
181 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
182 static unsigned long __initdata required_kernelcore;
183 static unsigned long __initdata required_movablecore;
184 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
186 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
188 EXPORT_SYMBOL(movable_zone);
189 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
192 int nr_node_ids __read_mostly = MAX_NUMNODES;
193 int nr_online_nodes __read_mostly = 1;
194 EXPORT_SYMBOL(nr_node_ids);
195 EXPORT_SYMBOL(nr_online_nodes);
198 int page_group_by_mobility_disabled __read_mostly;
200 static void set_pageblock_migratetype(struct page *page, int migratetype)
203 if (unlikely(page_group_by_mobility_disabled))
204 migratetype = MIGRATE_UNMOVABLE;
206 set_pageblock_flags_group(page, (unsigned long)migratetype,
207 PB_migrate, PB_migrate_end);
210 bool oom_killer_disabled __read_mostly;
212 #ifdef CONFIG_DEBUG_VM
213 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
217 unsigned long pfn = page_to_pfn(page);
220 seq = zone_span_seqbegin(zone);
221 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
223 else if (pfn < zone->zone_start_pfn)
225 } while (zone_span_seqretry(zone, seq));
230 static int page_is_consistent(struct zone *zone, struct page *page)
232 if (!pfn_valid_within(page_to_pfn(page)))
234 if (zone != page_zone(page))
240 * Temporary debugging check for pages not lying within a given zone.
242 static int bad_range(struct zone *zone, struct page *page)
244 if (page_outside_zone_boundaries(zone, page))
246 if (!page_is_consistent(zone, page))
252 static inline int bad_range(struct zone *zone, struct page *page)
258 static void bad_page(struct page *page)
260 static unsigned long resume;
261 static unsigned long nr_shown;
262 static unsigned long nr_unshown;
264 /* Don't complain about poisoned pages */
265 if (PageHWPoison(page)) {
266 __ClearPageBuddy(page);
271 * Allow a burst of 60 reports, then keep quiet for that minute;
272 * or allow a steady drip of one report per second.
274 if (nr_shown == 60) {
275 if (time_before(jiffies, resume)) {
281 "BUG: Bad page state: %lu messages suppressed\n",
288 resume = jiffies + 60 * HZ;
290 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
291 current->comm, page_to_pfn(page));
296 /* Leave bad fields for debug, except PageBuddy could make trouble */
297 __ClearPageBuddy(page);
298 add_taint(TAINT_BAD_PAGE);
302 * Higher-order pages are called "compound pages". They are structured thusly:
304 * The first PAGE_SIZE page is called the "head page".
306 * The remaining PAGE_SIZE pages are called "tail pages".
308 * All pages have PG_compound set. All pages have their ->private pointing at
309 * the head page (even the head page has this).
311 * The first tail page's ->lru.next holds the address of the compound page's
312 * put_page() function. Its ->lru.prev holds the order of allocation.
313 * This usage means that zero-order pages may not be compound.
316 static void free_compound_page(struct page *page)
318 __free_pages_ok(page, compound_order(page));
321 void prep_compound_page(struct page *page, unsigned long order)
324 int nr_pages = 1 << order;
326 set_compound_page_dtor(page, free_compound_page);
327 set_compound_order(page, order);
329 for (i = 1; i < nr_pages; i++) {
330 struct page *p = page + i;
333 p->first_page = page;
337 static int destroy_compound_page(struct page *page, unsigned long order)
340 int nr_pages = 1 << order;
343 if (unlikely(compound_order(page) != order) ||
344 unlikely(!PageHead(page))) {
349 __ClearPageHead(page);
351 for (i = 1; i < nr_pages; i++) {
352 struct page *p = page + i;
354 if (unlikely(!PageTail(p) || (p->first_page != page))) {
364 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
369 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
370 * and __GFP_HIGHMEM from hard or soft interrupt context.
372 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
373 for (i = 0; i < (1 << order); i++)
374 clear_highpage(page + i);
377 static inline void set_page_order(struct page *page, int order)
379 set_page_private(page, order);
380 __SetPageBuddy(page);
383 static inline void rmv_page_order(struct page *page)
385 __ClearPageBuddy(page);
386 set_page_private(page, 0);
390 * Locate the struct page for both the matching buddy in our
391 * pair (buddy1) and the combined O(n+1) page they form (page).
393 * 1) Any buddy B1 will have an order O twin B2 which satisfies
394 * the following equation:
396 * For example, if the starting buddy (buddy2) is #8 its order
398 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
400 * 2) Any buddy B will have an order O+1 parent P which
401 * satisfies the following equation:
404 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
406 static inline struct page *
407 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
409 unsigned long buddy_idx = page_idx ^ (1 << order);
411 return page + (buddy_idx - page_idx);
414 static inline unsigned long
415 __find_combined_index(unsigned long page_idx, unsigned int order)
417 return (page_idx & ~(1 << order));
421 * This function checks whether a page is free && is the buddy
422 * we can do coalesce a page and its buddy if
423 * (a) the buddy is not in a hole &&
424 * (b) the buddy is in the buddy system &&
425 * (c) a page and its buddy have the same order &&
426 * (d) a page and its buddy are in the same zone.
428 * For recording whether a page is in the buddy system, we use PG_buddy.
429 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
431 * For recording page's order, we use page_private(page).
433 static inline int page_is_buddy(struct page *page, struct page *buddy,
436 if (!pfn_valid_within(page_to_pfn(buddy)))
439 if (page_zone_id(page) != page_zone_id(buddy))
442 if (PageBuddy(buddy) && page_order(buddy) == order) {
443 VM_BUG_ON(page_count(buddy) != 0);
450 * Freeing function for a buddy system allocator.
452 * The concept of a buddy system is to maintain direct-mapped table
453 * (containing bit values) for memory blocks of various "orders".
454 * The bottom level table contains the map for the smallest allocatable
455 * units of memory (here, pages), and each level above it describes
456 * pairs of units from the levels below, hence, "buddies".
457 * At a high level, all that happens here is marking the table entry
458 * at the bottom level available, and propagating the changes upward
459 * as necessary, plus some accounting needed to play nicely with other
460 * parts of the VM system.
461 * At each level, we keep a list of pages, which are heads of continuous
462 * free pages of length of (1 << order) and marked with PG_buddy. Page's
463 * order is recorded in page_private(page) field.
464 * So when we are allocating or freeing one, we can derive the state of the
465 * other. That is, if we allocate a small block, and both were
466 * free, the remainder of the region must be split into blocks.
467 * If a block is freed, and its buddy is also free, then this
468 * triggers coalescing into a block of larger size.
473 static inline void __free_one_page(struct page *page,
474 struct zone *zone, unsigned int order,
477 unsigned long page_idx;
478 unsigned long combined_idx;
481 if (unlikely(PageCompound(page)))
482 if (unlikely(destroy_compound_page(page, order)))
485 VM_BUG_ON(migratetype == -1);
487 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
489 VM_BUG_ON(page_idx & ((1 << order) - 1));
490 VM_BUG_ON(bad_range(zone, page));
492 while (order < MAX_ORDER-1) {
493 buddy = __page_find_buddy(page, page_idx, order);
494 if (!page_is_buddy(page, buddy, order))
497 /* Our buddy is free, merge with it and move up one order. */
498 list_del(&buddy->lru);
499 zone->free_area[order].nr_free--;
500 rmv_page_order(buddy);
501 combined_idx = __find_combined_index(page_idx, order);
502 page = page + (combined_idx - page_idx);
503 page_idx = combined_idx;
506 set_page_order(page, order);
509 * If this is not the largest possible page, check if the buddy
510 * of the next-highest order is free. If it is, it's possible
511 * that pages are being freed that will coalesce soon. In case,
512 * that is happening, add the free page to the tail of the list
513 * so it's less likely to be used soon and more likely to be merged
514 * as a higher order page
516 if ((order < MAX_ORDER-1) && pfn_valid_within(page_to_pfn(buddy))) {
517 struct page *higher_page, *higher_buddy;
518 combined_idx = __find_combined_index(page_idx, order);
519 higher_page = page + combined_idx - page_idx;
520 higher_buddy = __page_find_buddy(higher_page, combined_idx, order + 1);
521 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
522 list_add_tail(&page->lru,
523 &zone->free_area[order].free_list[migratetype]);
528 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
530 zone->free_area[order].nr_free++;
534 * free_page_mlock() -- clean up attempts to free and mlocked() page.
535 * Page should not be on lru, so no need to fix that up.
536 * free_pages_check() will verify...
538 static inline void free_page_mlock(struct page *page)
540 __dec_zone_page_state(page, NR_MLOCK);
541 __count_vm_event(UNEVICTABLE_MLOCKFREED);
544 static inline int free_pages_check(struct page *page)
546 if (unlikely(page_mapcount(page) |
547 (page->mapping != NULL) |
548 (atomic_read(&page->_count) != 0) |
549 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) {
553 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
554 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
559 * Frees a number of pages from the PCP lists
560 * Assumes all pages on list are in same zone, and of same order.
561 * count is the number of pages to free.
563 * If the zone was previously in an "all pages pinned" state then look to
564 * see if this freeing clears that state.
566 * And clear the zone's pages_scanned counter, to hold off the "all pages are
567 * pinned" detection logic.
569 static void free_pcppages_bulk(struct zone *zone, int count,
570 struct per_cpu_pages *pcp)
575 spin_lock(&zone->lock);
576 zone->all_unreclaimable = 0;
577 zone->pages_scanned = 0;
579 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
582 struct list_head *list;
585 * Remove pages from lists in a round-robin fashion. A
586 * batch_free count is maintained that is incremented when an
587 * empty list is encountered. This is so more pages are freed
588 * off fuller lists instead of spinning excessively around empty
593 if (++migratetype == MIGRATE_PCPTYPES)
595 list = &pcp->lists[migratetype];
596 } while (list_empty(list));
599 page = list_entry(list->prev, struct page, lru);
600 /* must delete as __free_one_page list manipulates */
601 list_del(&page->lru);
602 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
603 __free_one_page(page, zone, 0, page_private(page));
604 trace_mm_page_pcpu_drain(page, 0, page_private(page));
605 } while (--count && --batch_free && !list_empty(list));
607 spin_unlock(&zone->lock);
610 static void free_one_page(struct zone *zone, struct page *page, int order,
613 spin_lock(&zone->lock);
614 zone->all_unreclaimable = 0;
615 zone->pages_scanned = 0;
617 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
618 __free_one_page(page, zone, order, migratetype);
619 spin_unlock(&zone->lock);
622 static void __free_pages_ok(struct page *page, unsigned int order)
627 int wasMlocked = __TestClearPageMlocked(page);
629 trace_mm_page_free_direct(page, order);
630 kmemcheck_free_shadow(page, order);
632 for (i = 0 ; i < (1 << order) ; ++i)
633 bad += free_pages_check(page + i);
637 if (!PageHighMem(page)) {
638 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
639 debug_check_no_obj_freed(page_address(page),
642 arch_free_page(page, order);
643 kernel_map_pages(page, 1 << order, 0);
645 local_irq_save(flags);
646 if (unlikely(wasMlocked))
647 free_page_mlock(page);
648 __count_vm_events(PGFREE, 1 << order);
649 free_one_page(page_zone(page), page, order,
650 get_pageblock_migratetype(page));
651 local_irq_restore(flags);
655 * permit the bootmem allocator to evade page validation on high-order frees
657 void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
660 __ClearPageReserved(page);
661 set_page_count(page, 0);
662 set_page_refcounted(page);
668 for (loop = 0; loop < BITS_PER_LONG; loop++) {
669 struct page *p = &page[loop];
671 if (loop + 1 < BITS_PER_LONG)
673 __ClearPageReserved(p);
674 set_page_count(p, 0);
677 set_page_refcounted(page);
678 __free_pages(page, order);
684 * The order of subdivision here is critical for the IO subsystem.
685 * Please do not alter this order without good reasons and regression
686 * testing. Specifically, as large blocks of memory are subdivided,
687 * the order in which smaller blocks are delivered depends on the order
688 * they're subdivided in this function. This is the primary factor
689 * influencing the order in which pages are delivered to the IO
690 * subsystem according to empirical testing, and this is also justified
691 * by considering the behavior of a buddy system containing a single
692 * large block of memory acted on by a series of small allocations.
693 * This behavior is a critical factor in sglist merging's success.
697 static inline void expand(struct zone *zone, struct page *page,
698 int low, int high, struct free_area *area,
701 unsigned long size = 1 << high;
707 VM_BUG_ON(bad_range(zone, &page[size]));
708 list_add(&page[size].lru, &area->free_list[migratetype]);
710 set_page_order(&page[size], high);
715 * This page is about to be returned from the page allocator
717 static inline int check_new_page(struct page *page)
719 if (unlikely(page_mapcount(page) |
720 (page->mapping != NULL) |
721 (atomic_read(&page->_count) != 0) |
722 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) {
729 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
733 for (i = 0; i < (1 << order); i++) {
734 struct page *p = page + i;
735 if (unlikely(check_new_page(p)))
739 set_page_private(page, 0);
740 set_page_refcounted(page);
742 arch_alloc_page(page, order);
743 kernel_map_pages(page, 1 << order, 1);
745 if (gfp_flags & __GFP_ZERO)
746 prep_zero_page(page, order, gfp_flags);
748 if (order && (gfp_flags & __GFP_COMP))
749 prep_compound_page(page, order);
755 * Go through the free lists for the given migratetype and remove
756 * the smallest available page from the freelists
759 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
762 unsigned int current_order;
763 struct free_area * area;
766 /* Find a page of the appropriate size in the preferred list */
767 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
768 area = &(zone->free_area[current_order]);
769 if (list_empty(&area->free_list[migratetype]))
772 page = list_entry(area->free_list[migratetype].next,
774 list_del(&page->lru);
775 rmv_page_order(page);
777 expand(zone, page, order, current_order, area, migratetype);
786 * This array describes the order lists are fallen back to when
787 * the free lists for the desirable migrate type are depleted
789 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
790 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
791 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
792 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
793 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
797 * Move the free pages in a range to the free lists of the requested type.
798 * Note that start_page and end_pages are not aligned on a pageblock
799 * boundary. If alignment is required, use move_freepages_block()
801 static int move_freepages(struct zone *zone,
802 struct page *start_page, struct page *end_page,
809 #ifndef CONFIG_HOLES_IN_ZONE
811 * page_zone is not safe to call in this context when
812 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
813 * anyway as we check zone boundaries in move_freepages_block().
814 * Remove at a later date when no bug reports exist related to
815 * grouping pages by mobility
817 BUG_ON(page_zone(start_page) != page_zone(end_page));
820 for (page = start_page; page <= end_page;) {
821 /* Make sure we are not inadvertently changing nodes */
822 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
824 if (!pfn_valid_within(page_to_pfn(page))) {
829 if (!PageBuddy(page)) {
834 order = page_order(page);
835 list_del(&page->lru);
837 &zone->free_area[order].free_list[migratetype]);
839 pages_moved += 1 << order;
845 static int move_freepages_block(struct zone *zone, struct page *page,
848 unsigned long start_pfn, end_pfn;
849 struct page *start_page, *end_page;
851 start_pfn = page_to_pfn(page);
852 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
853 start_page = pfn_to_page(start_pfn);
854 end_page = start_page + pageblock_nr_pages - 1;
855 end_pfn = start_pfn + pageblock_nr_pages - 1;
857 /* Do not cross zone boundaries */
858 if (start_pfn < zone->zone_start_pfn)
860 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
863 return move_freepages(zone, start_page, end_page, migratetype);
866 static void change_pageblock_range(struct page *pageblock_page,
867 int start_order, int migratetype)
869 int nr_pageblocks = 1 << (start_order - pageblock_order);
871 while (nr_pageblocks--) {
872 set_pageblock_migratetype(pageblock_page, migratetype);
873 pageblock_page += pageblock_nr_pages;
877 /* Remove an element from the buddy allocator from the fallback list */
878 static inline struct page *
879 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
881 struct free_area * area;
886 /* Find the largest possible block of pages in the other list */
887 for (current_order = MAX_ORDER-1; current_order >= order;
889 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
890 migratetype = fallbacks[start_migratetype][i];
892 /* MIGRATE_RESERVE handled later if necessary */
893 if (migratetype == MIGRATE_RESERVE)
896 area = &(zone->free_area[current_order]);
897 if (list_empty(&area->free_list[migratetype]))
900 page = list_entry(area->free_list[migratetype].next,
905 * If breaking a large block of pages, move all free
906 * pages to the preferred allocation list. If falling
907 * back for a reclaimable kernel allocation, be more
908 * agressive about taking ownership of free pages
910 if (unlikely(current_order >= (pageblock_order >> 1)) ||
911 start_migratetype == MIGRATE_RECLAIMABLE ||
912 page_group_by_mobility_disabled) {
914 pages = move_freepages_block(zone, page,
917 /* Claim the whole block if over half of it is free */
918 if (pages >= (1 << (pageblock_order-1)) ||
919 page_group_by_mobility_disabled)
920 set_pageblock_migratetype(page,
923 migratetype = start_migratetype;
926 /* Remove the page from the freelists */
927 list_del(&page->lru);
928 rmv_page_order(page);
930 /* Take ownership for orders >= pageblock_order */
931 if (current_order >= pageblock_order)
932 change_pageblock_range(page, current_order,
935 expand(zone, page, order, current_order, area, migratetype);
937 trace_mm_page_alloc_extfrag(page, order, current_order,
938 start_migratetype, migratetype);
948 * Do the hard work of removing an element from the buddy allocator.
949 * Call me with the zone->lock already held.
951 static struct page *__rmqueue(struct zone *zone, unsigned int order,
957 page = __rmqueue_smallest(zone, order, migratetype);
959 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
960 page = __rmqueue_fallback(zone, order, migratetype);
963 * Use MIGRATE_RESERVE rather than fail an allocation. goto
964 * is used because __rmqueue_smallest is an inline function
965 * and we want just one call site
968 migratetype = MIGRATE_RESERVE;
973 trace_mm_page_alloc_zone_locked(page, order, migratetype);
978 * Obtain a specified number of elements from the buddy allocator, all under
979 * a single hold of the lock, for efficiency. Add them to the supplied list.
980 * Returns the number of new pages which were placed at *list.
982 static int rmqueue_bulk(struct zone *zone, unsigned int order,
983 unsigned long count, struct list_head *list,
984 int migratetype, int cold)
988 spin_lock(&zone->lock);
989 for (i = 0; i < count; ++i) {
990 struct page *page = __rmqueue(zone, order, migratetype);
991 if (unlikely(page == NULL))
995 * Split buddy pages returned by expand() are received here
996 * in physical page order. The page is added to the callers and
997 * list and the list head then moves forward. From the callers
998 * perspective, the linked list is ordered by page number in
999 * some conditions. This is useful for IO devices that can
1000 * merge IO requests if the physical pages are ordered
1003 if (likely(cold == 0))
1004 list_add(&page->lru, list);
1006 list_add_tail(&page->lru, list);
1007 set_page_private(page, migratetype);
1010 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1011 spin_unlock(&zone->lock);
1017 * Called from the vmstat counter updater to drain pagesets of this
1018 * currently executing processor on remote nodes after they have
1021 * Note that this function must be called with the thread pinned to
1022 * a single processor.
1024 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1026 unsigned long flags;
1029 local_irq_save(flags);
1030 if (pcp->count >= pcp->batch)
1031 to_drain = pcp->batch;
1033 to_drain = pcp->count;
1034 free_pcppages_bulk(zone, to_drain, pcp);
1035 pcp->count -= to_drain;
1036 local_irq_restore(flags);
1041 * Drain pages of the indicated processor.
1043 * The processor must either be the current processor and the
1044 * thread pinned to the current processor or a processor that
1047 static void drain_pages(unsigned int cpu)
1049 unsigned long flags;
1052 for_each_populated_zone(zone) {
1053 struct per_cpu_pageset *pset;
1054 struct per_cpu_pages *pcp;
1056 local_irq_save(flags);
1057 pset = per_cpu_ptr(zone->pageset, cpu);
1060 free_pcppages_bulk(zone, pcp->count, pcp);
1062 local_irq_restore(flags);
1067 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1069 void drain_local_pages(void *arg)
1071 drain_pages(smp_processor_id());
1075 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1077 void drain_all_pages(void)
1079 on_each_cpu(drain_local_pages, NULL, 1);
1082 #ifdef CONFIG_HIBERNATION
1084 void mark_free_pages(struct zone *zone)
1086 unsigned long pfn, max_zone_pfn;
1087 unsigned long flags;
1089 struct list_head *curr;
1091 if (!zone->spanned_pages)
1094 spin_lock_irqsave(&zone->lock, flags);
1096 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1097 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1098 if (pfn_valid(pfn)) {
1099 struct page *page = pfn_to_page(pfn);
1101 if (!swsusp_page_is_forbidden(page))
1102 swsusp_unset_page_free(page);
1105 for_each_migratetype_order(order, t) {
1106 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1109 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1110 for (i = 0; i < (1UL << order); i++)
1111 swsusp_set_page_free(pfn_to_page(pfn + i));
1114 spin_unlock_irqrestore(&zone->lock, flags);
1116 #endif /* CONFIG_PM */
1119 * Free a 0-order page
1120 * cold == 1 ? free a cold page : free a hot page
1122 void free_hot_cold_page(struct page *page, int cold)
1124 struct zone *zone = page_zone(page);
1125 struct per_cpu_pages *pcp;
1126 unsigned long flags;
1128 int wasMlocked = __TestClearPageMlocked(page);
1130 trace_mm_page_free_direct(page, 0);
1131 kmemcheck_free_shadow(page, 0);
1134 page->mapping = NULL;
1135 if (free_pages_check(page))
1138 if (!PageHighMem(page)) {
1139 debug_check_no_locks_freed(page_address(page), PAGE_SIZE);
1140 debug_check_no_obj_freed(page_address(page), PAGE_SIZE);
1142 arch_free_page(page, 0);
1143 kernel_map_pages(page, 1, 0);
1145 migratetype = get_pageblock_migratetype(page);
1146 set_page_private(page, migratetype);
1147 local_irq_save(flags);
1148 if (unlikely(wasMlocked))
1149 free_page_mlock(page);
1150 __count_vm_event(PGFREE);
1153 * We only track unmovable, reclaimable and movable on pcp lists.
1154 * Free ISOLATE pages back to the allocator because they are being
1155 * offlined but treat RESERVE as movable pages so we can get those
1156 * areas back if necessary. Otherwise, we may have to free
1157 * excessively into the page allocator
1159 if (migratetype >= MIGRATE_PCPTYPES) {
1160 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1161 free_one_page(zone, page, 0, migratetype);
1164 migratetype = MIGRATE_MOVABLE;
1167 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1169 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1171 list_add(&page->lru, &pcp->lists[migratetype]);
1173 if (pcp->count >= pcp->high) {
1174 free_pcppages_bulk(zone, pcp->batch, pcp);
1175 pcp->count -= pcp->batch;
1179 local_irq_restore(flags);
1183 * split_page takes a non-compound higher-order page, and splits it into
1184 * n (1<<order) sub-pages: page[0..n]
1185 * Each sub-page must be freed individually.
1187 * Note: this is probably too low level an operation for use in drivers.
1188 * Please consult with lkml before using this in your driver.
1190 void split_page(struct page *page, unsigned int order)
1194 VM_BUG_ON(PageCompound(page));
1195 VM_BUG_ON(!page_count(page));
1197 #ifdef CONFIG_KMEMCHECK
1199 * Split shadow pages too, because free(page[0]) would
1200 * otherwise free the whole shadow.
1202 if (kmemcheck_page_is_tracked(page))
1203 split_page(virt_to_page(page[0].shadow), order);
1206 for (i = 1; i < (1 << order); i++)
1207 set_page_refcounted(page + i);
1211 * Similar to split_page except the page is already free. As this is only
1212 * being used for migration, the migratetype of the block also changes.
1213 * As this is called with interrupts disabled, the caller is responsible
1214 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1217 * Note: this is probably too low level an operation for use in drivers.
1218 * Please consult with lkml before using this in your driver.
1220 int split_free_page(struct page *page)
1223 unsigned long watermark;
1226 BUG_ON(!PageBuddy(page));
1228 zone = page_zone(page);
1229 order = page_order(page);
1231 /* Obey watermarks as if the page was being allocated */
1232 watermark = low_wmark_pages(zone) + (1 << order);
1233 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1236 /* Remove page from free list */
1237 list_del(&page->lru);
1238 zone->free_area[order].nr_free--;
1239 rmv_page_order(page);
1240 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1242 /* Split into individual pages */
1243 set_page_refcounted(page);
1244 split_page(page, order);
1246 if (order >= pageblock_order - 1) {
1247 struct page *endpage = page + (1 << order) - 1;
1248 for (; page < endpage; page += pageblock_nr_pages)
1249 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1256 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1257 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1261 struct page *buffered_rmqueue(struct zone *preferred_zone,
1262 struct zone *zone, int order, gfp_t gfp_flags,
1265 unsigned long flags;
1267 int cold = !!(gfp_flags & __GFP_COLD);
1270 if (likely(order == 0)) {
1271 struct per_cpu_pages *pcp;
1272 struct list_head *list;
1274 local_irq_save(flags);
1275 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1276 list = &pcp->lists[migratetype];
1277 if (list_empty(list)) {
1278 pcp->count += rmqueue_bulk(zone, 0,
1281 if (unlikely(list_empty(list)))
1286 page = list_entry(list->prev, struct page, lru);
1288 page = list_entry(list->next, struct page, lru);
1290 list_del(&page->lru);
1293 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1295 * __GFP_NOFAIL is not to be used in new code.
1297 * All __GFP_NOFAIL callers should be fixed so that they
1298 * properly detect and handle allocation failures.
1300 * We most definitely don't want callers attempting to
1301 * allocate greater than order-1 page units with
1304 WARN_ON_ONCE(order > 1);
1306 spin_lock_irqsave(&zone->lock, flags);
1307 page = __rmqueue(zone, order, migratetype);
1308 spin_unlock(&zone->lock);
1311 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1314 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1315 zone_statistics(preferred_zone, zone);
1316 local_irq_restore(flags);
1318 VM_BUG_ON(bad_range(zone, page));
1319 if (prep_new_page(page, order, gfp_flags))
1324 local_irq_restore(flags);
1328 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1329 #define ALLOC_WMARK_MIN WMARK_MIN
1330 #define ALLOC_WMARK_LOW WMARK_LOW
1331 #define ALLOC_WMARK_HIGH WMARK_HIGH
1332 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1334 /* Mask to get the watermark bits */
1335 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1337 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1338 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1339 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1341 #ifdef CONFIG_FAIL_PAGE_ALLOC
1343 static struct fail_page_alloc_attr {
1344 struct fault_attr attr;
1346 u32 ignore_gfp_highmem;
1347 u32 ignore_gfp_wait;
1350 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1352 struct dentry *ignore_gfp_highmem_file;
1353 struct dentry *ignore_gfp_wait_file;
1354 struct dentry *min_order_file;
1356 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1358 } fail_page_alloc = {
1359 .attr = FAULT_ATTR_INITIALIZER,
1360 .ignore_gfp_wait = 1,
1361 .ignore_gfp_highmem = 1,
1365 static int __init setup_fail_page_alloc(char *str)
1367 return setup_fault_attr(&fail_page_alloc.attr, str);
1369 __setup("fail_page_alloc=", setup_fail_page_alloc);
1371 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1373 if (order < fail_page_alloc.min_order)
1375 if (gfp_mask & __GFP_NOFAIL)
1377 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1379 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1382 return should_fail(&fail_page_alloc.attr, 1 << order);
1385 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1387 static int __init fail_page_alloc_debugfs(void)
1389 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1393 err = init_fault_attr_dentries(&fail_page_alloc.attr,
1397 dir = fail_page_alloc.attr.dentries.dir;
1399 fail_page_alloc.ignore_gfp_wait_file =
1400 debugfs_create_bool("ignore-gfp-wait", mode, dir,
1401 &fail_page_alloc.ignore_gfp_wait);
1403 fail_page_alloc.ignore_gfp_highmem_file =
1404 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1405 &fail_page_alloc.ignore_gfp_highmem);
1406 fail_page_alloc.min_order_file =
1407 debugfs_create_u32("min-order", mode, dir,
1408 &fail_page_alloc.min_order);
1410 if (!fail_page_alloc.ignore_gfp_wait_file ||
1411 !fail_page_alloc.ignore_gfp_highmem_file ||
1412 !fail_page_alloc.min_order_file) {
1414 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file);
1415 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file);
1416 debugfs_remove(fail_page_alloc.min_order_file);
1417 cleanup_fault_attr_dentries(&fail_page_alloc.attr);
1423 late_initcall(fail_page_alloc_debugfs);
1425 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1427 #else /* CONFIG_FAIL_PAGE_ALLOC */
1429 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1434 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1437 * Return 1 if free pages are above 'mark'. This takes into account the order
1438 * of the allocation.
1440 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1441 int classzone_idx, int alloc_flags)
1443 /* free_pages my go negative - that's OK */
1445 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1;
1448 if (alloc_flags & ALLOC_HIGH)
1450 if (alloc_flags & ALLOC_HARDER)
1453 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1455 for (o = 0; o < order; o++) {
1456 /* At the next order, this order's pages become unavailable */
1457 free_pages -= z->free_area[o].nr_free << o;
1459 /* Require fewer higher order pages to be free */
1462 if (free_pages <= min)
1470 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1471 * skip over zones that are not allowed by the cpuset, or that have
1472 * been recently (in last second) found to be nearly full. See further
1473 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1474 * that have to skip over a lot of full or unallowed zones.
1476 * If the zonelist cache is present in the passed in zonelist, then
1477 * returns a pointer to the allowed node mask (either the current
1478 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1480 * If the zonelist cache is not available for this zonelist, does
1481 * nothing and returns NULL.
1483 * If the fullzones BITMAP in the zonelist cache is stale (more than
1484 * a second since last zap'd) then we zap it out (clear its bits.)
1486 * We hold off even calling zlc_setup, until after we've checked the
1487 * first zone in the zonelist, on the theory that most allocations will
1488 * be satisfied from that first zone, so best to examine that zone as
1489 * quickly as we can.
1491 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1493 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1494 nodemask_t *allowednodes; /* zonelist_cache approximation */
1496 zlc = zonelist->zlcache_ptr;
1500 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1501 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1502 zlc->last_full_zap = jiffies;
1505 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1506 &cpuset_current_mems_allowed :
1507 &node_states[N_HIGH_MEMORY];
1508 return allowednodes;
1512 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1513 * if it is worth looking at further for free memory:
1514 * 1) Check that the zone isn't thought to be full (doesn't have its
1515 * bit set in the zonelist_cache fullzones BITMAP).
1516 * 2) Check that the zones node (obtained from the zonelist_cache
1517 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1518 * Return true (non-zero) if zone is worth looking at further, or
1519 * else return false (zero) if it is not.
1521 * This check -ignores- the distinction between various watermarks,
1522 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1523 * found to be full for any variation of these watermarks, it will
1524 * be considered full for up to one second by all requests, unless
1525 * we are so low on memory on all allowed nodes that we are forced
1526 * into the second scan of the zonelist.
1528 * In the second scan we ignore this zonelist cache and exactly
1529 * apply the watermarks to all zones, even it is slower to do so.
1530 * We are low on memory in the second scan, and should leave no stone
1531 * unturned looking for a free page.
1533 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1534 nodemask_t *allowednodes)
1536 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1537 int i; /* index of *z in zonelist zones */
1538 int n; /* node that zone *z is on */
1540 zlc = zonelist->zlcache_ptr;
1544 i = z - zonelist->_zonerefs;
1547 /* This zone is worth trying if it is allowed but not full */
1548 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1552 * Given 'z' scanning a zonelist, set the corresponding bit in
1553 * zlc->fullzones, so that subsequent attempts to allocate a page
1554 * from that zone don't waste time re-examining it.
1556 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1558 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1559 int i; /* index of *z in zonelist zones */
1561 zlc = zonelist->zlcache_ptr;
1565 i = z - zonelist->_zonerefs;
1567 set_bit(i, zlc->fullzones);
1570 #else /* CONFIG_NUMA */
1572 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1577 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1578 nodemask_t *allowednodes)
1583 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1586 #endif /* CONFIG_NUMA */
1589 * get_page_from_freelist goes through the zonelist trying to allocate
1592 static struct page *
1593 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1594 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1595 struct zone *preferred_zone, int migratetype)
1598 struct page *page = NULL;
1601 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1602 int zlc_active = 0; /* set if using zonelist_cache */
1603 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1605 classzone_idx = zone_idx(preferred_zone);
1608 * Scan zonelist, looking for a zone with enough free.
1609 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1611 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1612 high_zoneidx, nodemask) {
1613 if (NUMA_BUILD && zlc_active &&
1614 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1616 if ((alloc_flags & ALLOC_CPUSET) &&
1617 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1620 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1621 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1625 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1626 if (zone_watermark_ok(zone, order, mark,
1627 classzone_idx, alloc_flags))
1630 if (zone_reclaim_mode == 0)
1631 goto this_zone_full;
1633 ret = zone_reclaim(zone, gfp_mask, order);
1635 case ZONE_RECLAIM_NOSCAN:
1638 case ZONE_RECLAIM_FULL:
1639 /* scanned but unreclaimable */
1640 goto this_zone_full;
1642 /* did we reclaim enough */
1643 if (!zone_watermark_ok(zone, order, mark,
1644 classzone_idx, alloc_flags))
1645 goto this_zone_full;
1650 page = buffered_rmqueue(preferred_zone, zone, order,
1651 gfp_mask, migratetype);
1656 zlc_mark_zone_full(zonelist, z);
1658 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1660 * we do zlc_setup after the first zone is tried but only
1661 * if there are multiple nodes make it worthwhile
1663 allowednodes = zlc_setup(zonelist, alloc_flags);
1669 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1670 /* Disable zlc cache for second zonelist scan */
1678 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1679 unsigned long pages_reclaimed)
1681 /* Do not loop if specifically requested */
1682 if (gfp_mask & __GFP_NORETRY)
1686 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1687 * means __GFP_NOFAIL, but that may not be true in other
1690 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1694 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1695 * specified, then we retry until we no longer reclaim any pages
1696 * (above), or we've reclaimed an order of pages at least as
1697 * large as the allocation's order. In both cases, if the
1698 * allocation still fails, we stop retrying.
1700 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1704 * Don't let big-order allocations loop unless the caller
1705 * explicitly requests that.
1707 if (gfp_mask & __GFP_NOFAIL)
1713 static inline struct page *
1714 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1715 struct zonelist *zonelist, enum zone_type high_zoneidx,
1716 nodemask_t *nodemask, struct zone *preferred_zone,
1721 /* Acquire the OOM killer lock for the zones in zonelist */
1722 if (!try_set_zone_oom(zonelist, gfp_mask)) {
1723 schedule_timeout_uninterruptible(1);
1728 * Go through the zonelist yet one more time, keep very high watermark
1729 * here, this is only to catch a parallel oom killing, we must fail if
1730 * we're still under heavy pressure.
1732 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1733 order, zonelist, high_zoneidx,
1734 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1735 preferred_zone, migratetype);
1739 if (!(gfp_mask & __GFP_NOFAIL)) {
1740 /* The OOM killer will not help higher order allocs */
1741 if (order > PAGE_ALLOC_COSTLY_ORDER)
1744 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1745 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1746 * The caller should handle page allocation failure by itself if
1747 * it specifies __GFP_THISNODE.
1748 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1750 if (gfp_mask & __GFP_THISNODE)
1753 /* Exhausted what can be done so it's blamo time */
1754 out_of_memory(zonelist, gfp_mask, order, nodemask);
1757 clear_zonelist_oom(zonelist, gfp_mask);
1761 /* The really slow allocator path where we enter direct reclaim */
1762 static inline struct page *
1763 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1764 struct zonelist *zonelist, enum zone_type high_zoneidx,
1765 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1766 int migratetype, unsigned long *did_some_progress)
1768 struct page *page = NULL;
1769 struct reclaim_state reclaim_state;
1770 struct task_struct *p = current;
1774 /* We now go into synchronous reclaim */
1775 cpuset_memory_pressure_bump();
1776 p->flags |= PF_MEMALLOC;
1777 lockdep_set_current_reclaim_state(gfp_mask);
1778 reclaim_state.reclaimed_slab = 0;
1779 p->reclaim_state = &reclaim_state;
1781 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1783 p->reclaim_state = NULL;
1784 lockdep_clear_current_reclaim_state();
1785 p->flags &= ~PF_MEMALLOC;
1792 if (likely(*did_some_progress))
1793 page = get_page_from_freelist(gfp_mask, nodemask, order,
1794 zonelist, high_zoneidx,
1795 alloc_flags, preferred_zone,
1801 * This is called in the allocator slow-path if the allocation request is of
1802 * sufficient urgency to ignore watermarks and take other desperate measures
1804 static inline struct page *
1805 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1806 struct zonelist *zonelist, enum zone_type high_zoneidx,
1807 nodemask_t *nodemask, struct zone *preferred_zone,
1813 page = get_page_from_freelist(gfp_mask, nodemask, order,
1814 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
1815 preferred_zone, migratetype);
1817 if (!page && gfp_mask & __GFP_NOFAIL)
1818 congestion_wait(BLK_RW_ASYNC, HZ/50);
1819 } while (!page && (gfp_mask & __GFP_NOFAIL));
1825 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
1826 enum zone_type high_zoneidx)
1831 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
1832 wakeup_kswapd(zone, order);
1836 gfp_to_alloc_flags(gfp_t gfp_mask)
1838 struct task_struct *p = current;
1839 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
1840 const gfp_t wait = gfp_mask & __GFP_WAIT;
1842 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
1843 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH);
1846 * The caller may dip into page reserves a bit more if the caller
1847 * cannot run direct reclaim, or if the caller has realtime scheduling
1848 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1849 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1851 alloc_flags |= (gfp_mask & __GFP_HIGH);
1854 alloc_flags |= ALLOC_HARDER;
1856 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1857 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1859 alloc_flags &= ~ALLOC_CPUSET;
1860 } else if (unlikely(rt_task(p)) && !in_interrupt())
1861 alloc_flags |= ALLOC_HARDER;
1863 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
1864 if (!in_interrupt() &&
1865 ((p->flags & PF_MEMALLOC) ||
1866 unlikely(test_thread_flag(TIF_MEMDIE))))
1867 alloc_flags |= ALLOC_NO_WATERMARKS;
1873 static inline struct page *
1874 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
1875 struct zonelist *zonelist, enum zone_type high_zoneidx,
1876 nodemask_t *nodemask, struct zone *preferred_zone,
1879 const gfp_t wait = gfp_mask & __GFP_WAIT;
1880 struct page *page = NULL;
1882 unsigned long pages_reclaimed = 0;
1883 unsigned long did_some_progress;
1884 struct task_struct *p = current;
1887 * In the slowpath, we sanity check order to avoid ever trying to
1888 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
1889 * be using allocators in order of preference for an area that is
1892 if (order >= MAX_ORDER) {
1893 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
1898 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
1899 * __GFP_NOWARN set) should not cause reclaim since the subsystem
1900 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
1901 * using a larger set of nodes after it has established that the
1902 * allowed per node queues are empty and that nodes are
1905 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
1909 wake_all_kswapd(order, zonelist, high_zoneidx);
1912 * OK, we're below the kswapd watermark and have kicked background
1913 * reclaim. Now things get more complex, so set up alloc_flags according
1914 * to how we want to proceed.
1916 alloc_flags = gfp_to_alloc_flags(gfp_mask);
1918 /* This is the last chance, in general, before the goto nopage. */
1919 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
1920 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
1921 preferred_zone, migratetype);
1926 /* Allocate without watermarks if the context allows */
1927 if (alloc_flags & ALLOC_NO_WATERMARKS) {
1928 page = __alloc_pages_high_priority(gfp_mask, order,
1929 zonelist, high_zoneidx, nodemask,
1930 preferred_zone, migratetype);
1935 /* Atomic allocations - we can't balance anything */
1939 /* Avoid recursion of direct reclaim */
1940 if (p->flags & PF_MEMALLOC)
1943 /* Avoid allocations with no watermarks from looping endlessly */
1944 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
1947 /* Try direct reclaim and then allocating */
1948 page = __alloc_pages_direct_reclaim(gfp_mask, order,
1949 zonelist, high_zoneidx,
1951 alloc_flags, preferred_zone,
1952 migratetype, &did_some_progress);
1957 * If we failed to make any progress reclaiming, then we are
1958 * running out of options and have to consider going OOM
1960 if (!did_some_progress) {
1961 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1962 if (oom_killer_disabled)
1964 page = __alloc_pages_may_oom(gfp_mask, order,
1965 zonelist, high_zoneidx,
1966 nodemask, preferred_zone,
1972 * The OOM killer does not trigger for high-order
1973 * ~__GFP_NOFAIL allocations so if no progress is being
1974 * made, there are no other options and retrying is
1977 if (order > PAGE_ALLOC_COSTLY_ORDER &&
1978 !(gfp_mask & __GFP_NOFAIL))
1985 /* Check if we should retry the allocation */
1986 pages_reclaimed += did_some_progress;
1987 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
1988 /* Wait for some write requests to complete then retry */
1989 congestion_wait(BLK_RW_ASYNC, HZ/50);
1994 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1995 printk(KERN_WARNING "%s: page allocation failure."
1996 " order:%d, mode:0x%x\n",
1997 p->comm, order, gfp_mask);
2003 if (kmemcheck_enabled)
2004 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2010 * This is the 'heart' of the zoned buddy allocator.
2013 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2014 struct zonelist *zonelist, nodemask_t *nodemask)
2016 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2017 struct zone *preferred_zone;
2019 int migratetype = allocflags_to_migratetype(gfp_mask);
2021 gfp_mask &= gfp_allowed_mask;
2023 lockdep_trace_alloc(gfp_mask);
2025 might_sleep_if(gfp_mask & __GFP_WAIT);
2027 if (should_fail_alloc_page(gfp_mask, order))
2031 * Check the zones suitable for the gfp_mask contain at least one
2032 * valid zone. It's possible to have an empty zonelist as a result
2033 * of GFP_THISNODE and a memoryless node
2035 if (unlikely(!zonelist->_zonerefs->zone))
2039 /* The preferred zone is used for statistics later */
2040 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone);
2041 if (!preferred_zone) {
2046 /* First allocation attempt */
2047 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2048 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2049 preferred_zone, migratetype);
2050 if (unlikely(!page))
2051 page = __alloc_pages_slowpath(gfp_mask, order,
2052 zonelist, high_zoneidx, nodemask,
2053 preferred_zone, migratetype);
2056 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2059 EXPORT_SYMBOL(__alloc_pages_nodemask);
2062 * Common helper functions.
2064 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2069 * __get_free_pages() returns a 32-bit address, which cannot represent
2072 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2074 page = alloc_pages(gfp_mask, order);
2077 return (unsigned long) page_address(page);
2079 EXPORT_SYMBOL(__get_free_pages);
2081 unsigned long get_zeroed_page(gfp_t gfp_mask)
2083 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2085 EXPORT_SYMBOL(get_zeroed_page);
2087 void __pagevec_free(struct pagevec *pvec)
2089 int i = pagevec_count(pvec);
2092 trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
2093 free_hot_cold_page(pvec->pages[i], pvec->cold);
2097 void __free_pages(struct page *page, unsigned int order)
2099 if (put_page_testzero(page)) {
2101 free_hot_cold_page(page, 0);
2103 __free_pages_ok(page, order);
2107 EXPORT_SYMBOL(__free_pages);
2109 void free_pages(unsigned long addr, unsigned int order)
2112 VM_BUG_ON(!virt_addr_valid((void *)addr));
2113 __free_pages(virt_to_page((void *)addr), order);
2117 EXPORT_SYMBOL(free_pages);
2120 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2121 * @size: the number of bytes to allocate
2122 * @gfp_mask: GFP flags for the allocation
2124 * This function is similar to alloc_pages(), except that it allocates the
2125 * minimum number of pages to satisfy the request. alloc_pages() can only
2126 * allocate memory in power-of-two pages.
2128 * This function is also limited by MAX_ORDER.
2130 * Memory allocated by this function must be released by free_pages_exact().
2132 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2134 unsigned int order = get_order(size);
2137 addr = __get_free_pages(gfp_mask, order);
2139 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2140 unsigned long used = addr + PAGE_ALIGN(size);
2142 split_page(virt_to_page((void *)addr), order);
2143 while (used < alloc_end) {
2149 return (void *)addr;
2151 EXPORT_SYMBOL(alloc_pages_exact);
2154 * free_pages_exact - release memory allocated via alloc_pages_exact()
2155 * @virt: the value returned by alloc_pages_exact.
2156 * @size: size of allocation, same value as passed to alloc_pages_exact().
2158 * Release the memory allocated by a previous call to alloc_pages_exact.
2160 void free_pages_exact(void *virt, size_t size)
2162 unsigned long addr = (unsigned long)virt;
2163 unsigned long end = addr + PAGE_ALIGN(size);
2165 while (addr < end) {
2170 EXPORT_SYMBOL(free_pages_exact);
2172 static unsigned int nr_free_zone_pages(int offset)
2177 /* Just pick one node, since fallback list is circular */
2178 unsigned int sum = 0;
2180 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2182 for_each_zone_zonelist(zone, z, zonelist, offset) {
2183 unsigned long size = zone->present_pages;
2184 unsigned long high = high_wmark_pages(zone);
2193 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2195 unsigned int nr_free_buffer_pages(void)
2197 return nr_free_zone_pages(gfp_zone(GFP_USER));
2199 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2202 * Amount of free RAM allocatable within all zones
2204 unsigned int nr_free_pagecache_pages(void)
2206 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2209 static inline void show_node(struct zone *zone)
2212 printk("Node %d ", zone_to_nid(zone));
2215 void si_meminfo(struct sysinfo *val)
2217 val->totalram = totalram_pages;
2219 val->freeram = global_page_state(NR_FREE_PAGES);
2220 val->bufferram = nr_blockdev_pages();
2221 val->totalhigh = totalhigh_pages;
2222 val->freehigh = nr_free_highpages();
2223 val->mem_unit = PAGE_SIZE;
2226 EXPORT_SYMBOL(si_meminfo);
2229 void si_meminfo_node(struct sysinfo *val, int nid)
2231 pg_data_t *pgdat = NODE_DATA(nid);
2233 val->totalram = pgdat->node_present_pages;
2234 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2235 #ifdef CONFIG_HIGHMEM
2236 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2237 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2243 val->mem_unit = PAGE_SIZE;
2247 #define K(x) ((x) << (PAGE_SHIFT-10))
2250 * Show free area list (used inside shift_scroll-lock stuff)
2251 * We also calculate the percentage fragmentation. We do this by counting the
2252 * memory on each free list with the exception of the first item on the list.
2254 void show_free_areas(void)
2259 for_each_populated_zone(zone) {
2261 printk("%s per-cpu:\n", zone->name);
2263 for_each_online_cpu(cpu) {
2264 struct per_cpu_pageset *pageset;
2266 pageset = per_cpu_ptr(zone->pageset, cpu);
2268 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2269 cpu, pageset->pcp.high,
2270 pageset->pcp.batch, pageset->pcp.count);
2274 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2275 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2277 " dirty:%lu writeback:%lu unstable:%lu\n"
2278 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2279 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2280 global_page_state(NR_ACTIVE_ANON),
2281 global_page_state(NR_INACTIVE_ANON),
2282 global_page_state(NR_ISOLATED_ANON),
2283 global_page_state(NR_ACTIVE_FILE),
2284 global_page_state(NR_INACTIVE_FILE),
2285 global_page_state(NR_ISOLATED_FILE),
2286 global_page_state(NR_UNEVICTABLE),
2287 global_page_state(NR_FILE_DIRTY),
2288 global_page_state(NR_WRITEBACK),
2289 global_page_state(NR_UNSTABLE_NFS),
2290 global_page_state(NR_FREE_PAGES),
2291 global_page_state(NR_SLAB_RECLAIMABLE),
2292 global_page_state(NR_SLAB_UNRECLAIMABLE),
2293 global_page_state(NR_FILE_MAPPED),
2294 global_page_state(NR_SHMEM),
2295 global_page_state(NR_PAGETABLE),
2296 global_page_state(NR_BOUNCE));
2298 for_each_populated_zone(zone) {
2307 " active_anon:%lukB"
2308 " inactive_anon:%lukB"
2309 " active_file:%lukB"
2310 " inactive_file:%lukB"
2311 " unevictable:%lukB"
2312 " isolated(anon):%lukB"
2313 " isolated(file):%lukB"
2320 " slab_reclaimable:%lukB"
2321 " slab_unreclaimable:%lukB"
2322 " kernel_stack:%lukB"
2326 " writeback_tmp:%lukB"
2327 " pages_scanned:%lu"
2328 " all_unreclaimable? %s"
2331 K(zone_page_state(zone, NR_FREE_PAGES)),
2332 K(min_wmark_pages(zone)),
2333 K(low_wmark_pages(zone)),
2334 K(high_wmark_pages(zone)),
2335 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2336 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2337 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2338 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2339 K(zone_page_state(zone, NR_UNEVICTABLE)),
2340 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2341 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2342 K(zone->present_pages),
2343 K(zone_page_state(zone, NR_MLOCK)),
2344 K(zone_page_state(zone, NR_FILE_DIRTY)),
2345 K(zone_page_state(zone, NR_WRITEBACK)),
2346 K(zone_page_state(zone, NR_FILE_MAPPED)),
2347 K(zone_page_state(zone, NR_SHMEM)),
2348 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2349 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2350 zone_page_state(zone, NR_KERNEL_STACK) *
2352 K(zone_page_state(zone, NR_PAGETABLE)),
2353 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2354 K(zone_page_state(zone, NR_BOUNCE)),
2355 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2356 zone->pages_scanned,
2357 (zone->all_unreclaimable ? "yes" : "no")
2359 printk("lowmem_reserve[]:");
2360 for (i = 0; i < MAX_NR_ZONES; i++)
2361 printk(" %lu", zone->lowmem_reserve[i]);
2365 for_each_populated_zone(zone) {
2366 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2369 printk("%s: ", zone->name);
2371 spin_lock_irqsave(&zone->lock, flags);
2372 for (order = 0; order < MAX_ORDER; order++) {
2373 nr[order] = zone->free_area[order].nr_free;
2374 total += nr[order] << order;
2376 spin_unlock_irqrestore(&zone->lock, flags);
2377 for (order = 0; order < MAX_ORDER; order++)
2378 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2379 printk("= %lukB\n", K(total));
2382 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2384 show_swap_cache_info();
2387 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2389 zoneref->zone = zone;
2390 zoneref->zone_idx = zone_idx(zone);
2394 * Builds allocation fallback zone lists.
2396 * Add all populated zones of a node to the zonelist.
2398 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2399 int nr_zones, enum zone_type zone_type)
2403 BUG_ON(zone_type >= MAX_NR_ZONES);
2408 zone = pgdat->node_zones + zone_type;
2409 if (populated_zone(zone)) {
2410 zoneref_set_zone(zone,
2411 &zonelist->_zonerefs[nr_zones++]);
2412 check_highest_zone(zone_type);
2415 } while (zone_type);
2422 * 0 = automatic detection of better ordering.
2423 * 1 = order by ([node] distance, -zonetype)
2424 * 2 = order by (-zonetype, [node] distance)
2426 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2427 * the same zonelist. So only NUMA can configure this param.
2429 #define ZONELIST_ORDER_DEFAULT 0
2430 #define ZONELIST_ORDER_NODE 1
2431 #define ZONELIST_ORDER_ZONE 2
2433 /* zonelist order in the kernel.
2434 * set_zonelist_order() will set this to NODE or ZONE.
2436 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2437 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2441 /* The value user specified ....changed by config */
2442 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2443 /* string for sysctl */
2444 #define NUMA_ZONELIST_ORDER_LEN 16
2445 char numa_zonelist_order[16] = "default";
2448 * interface for configure zonelist ordering.
2449 * command line option "numa_zonelist_order"
2450 * = "[dD]efault - default, automatic configuration.
2451 * = "[nN]ode - order by node locality, then by zone within node
2452 * = "[zZ]one - order by zone, then by locality within zone
2455 static int __parse_numa_zonelist_order(char *s)
2457 if (*s == 'd' || *s == 'D') {
2458 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2459 } else if (*s == 'n' || *s == 'N') {
2460 user_zonelist_order = ZONELIST_ORDER_NODE;
2461 } else if (*s == 'z' || *s == 'Z') {
2462 user_zonelist_order = ZONELIST_ORDER_ZONE;
2465 "Ignoring invalid numa_zonelist_order value: "
2472 static __init int setup_numa_zonelist_order(char *s)
2475 return __parse_numa_zonelist_order(s);
2478 early_param("numa_zonelist_order", setup_numa_zonelist_order);
2481 * sysctl handler for numa_zonelist_order
2483 int numa_zonelist_order_handler(ctl_table *table, int write,
2484 void __user *buffer, size_t *length,
2487 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2489 static DEFINE_MUTEX(zl_order_mutex);
2491 mutex_lock(&zl_order_mutex);
2493 strcpy(saved_string, (char*)table->data);
2494 ret = proc_dostring(table, write, buffer, length, ppos);
2498 int oldval = user_zonelist_order;
2499 if (__parse_numa_zonelist_order((char*)table->data)) {
2501 * bogus value. restore saved string
2503 strncpy((char*)table->data, saved_string,
2504 NUMA_ZONELIST_ORDER_LEN);
2505 user_zonelist_order = oldval;
2506 } else if (oldval != user_zonelist_order)
2507 build_all_zonelists();
2510 mutex_unlock(&zl_order_mutex);
2515 #define MAX_NODE_LOAD (nr_online_nodes)
2516 static int node_load[MAX_NUMNODES];
2519 * find_next_best_node - find the next node that should appear in a given node's fallback list
2520 * @node: node whose fallback list we're appending
2521 * @used_node_mask: nodemask_t of already used nodes
2523 * We use a number of factors to determine which is the next node that should
2524 * appear on a given node's fallback list. The node should not have appeared
2525 * already in @node's fallback list, and it should be the next closest node
2526 * according to the distance array (which contains arbitrary distance values
2527 * from each node to each node in the system), and should also prefer nodes
2528 * with no CPUs, since presumably they'll have very little allocation pressure
2529 * on them otherwise.
2530 * It returns -1 if no node is found.
2532 static int find_next_best_node(int node, nodemask_t *used_node_mask)
2535 int min_val = INT_MAX;
2537 const struct cpumask *tmp = cpumask_of_node(0);
2539 /* Use the local node if we haven't already */
2540 if (!node_isset(node, *used_node_mask)) {
2541 node_set(node, *used_node_mask);
2545 for_each_node_state(n, N_HIGH_MEMORY) {
2547 /* Don't want a node to appear more than once */
2548 if (node_isset(n, *used_node_mask))
2551 /* Use the distance array to find the distance */
2552 val = node_distance(node, n);
2554 /* Penalize nodes under us ("prefer the next node") */
2557 /* Give preference to headless and unused nodes */
2558 tmp = cpumask_of_node(n);
2559 if (!cpumask_empty(tmp))
2560 val += PENALTY_FOR_NODE_WITH_CPUS;
2562 /* Slight preference for less loaded node */
2563 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2564 val += node_load[n];
2566 if (val < min_val) {
2573 node_set(best_node, *used_node_mask);
2580 * Build zonelists ordered by node and zones within node.
2581 * This results in maximum locality--normal zone overflows into local
2582 * DMA zone, if any--but risks exhausting DMA zone.
2584 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2587 struct zonelist *zonelist;
2589 zonelist = &pgdat->node_zonelists[0];
2590 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2592 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2594 zonelist->_zonerefs[j].zone = NULL;
2595 zonelist->_zonerefs[j].zone_idx = 0;
2599 * Build gfp_thisnode zonelists
2601 static void build_thisnode_zonelists(pg_data_t *pgdat)
2604 struct zonelist *zonelist;
2606 zonelist = &pgdat->node_zonelists[1];
2607 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2608 zonelist->_zonerefs[j].zone = NULL;
2609 zonelist->_zonerefs[j].zone_idx = 0;
2613 * Build zonelists ordered by zone and nodes within zones.
2614 * This results in conserving DMA zone[s] until all Normal memory is
2615 * exhausted, but results in overflowing to remote node while memory
2616 * may still exist in local DMA zone.
2618 static int node_order[MAX_NUMNODES];
2620 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2623 int zone_type; /* needs to be signed */
2625 struct zonelist *zonelist;
2627 zonelist = &pgdat->node_zonelists[0];
2629 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2630 for (j = 0; j < nr_nodes; j++) {
2631 node = node_order[j];
2632 z = &NODE_DATA(node)->node_zones[zone_type];
2633 if (populated_zone(z)) {
2635 &zonelist->_zonerefs[pos++]);
2636 check_highest_zone(zone_type);
2640 zonelist->_zonerefs[pos].zone = NULL;
2641 zonelist->_zonerefs[pos].zone_idx = 0;
2644 static int default_zonelist_order(void)
2647 unsigned long low_kmem_size,total_size;
2651 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2652 * If they are really small and used heavily, the system can fall
2653 * into OOM very easily.
2654 * This function detect ZONE_DMA/DMA32 size and configures zone order.
2656 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2659 for_each_online_node(nid) {
2660 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2661 z = &NODE_DATA(nid)->node_zones[zone_type];
2662 if (populated_zone(z)) {
2663 if (zone_type < ZONE_NORMAL)
2664 low_kmem_size += z->present_pages;
2665 total_size += z->present_pages;
2666 } else if (zone_type == ZONE_NORMAL) {
2668 * If any node has only lowmem, then node order
2669 * is preferred to allow kernel allocations
2670 * locally; otherwise, they can easily infringe
2671 * on other nodes when there is an abundance of
2672 * lowmem available to allocate from.
2674 return ZONELIST_ORDER_NODE;
2678 if (!low_kmem_size || /* there are no DMA area. */
2679 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2680 return ZONELIST_ORDER_NODE;
2682 * look into each node's config.
2683 * If there is a node whose DMA/DMA32 memory is very big area on
2684 * local memory, NODE_ORDER may be suitable.
2686 average_size = total_size /
2687 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2688 for_each_online_node(nid) {
2691 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2692 z = &NODE_DATA(nid)->node_zones[zone_type];
2693 if (populated_zone(z)) {
2694 if (zone_type < ZONE_NORMAL)
2695 low_kmem_size += z->present_pages;
2696 total_size += z->present_pages;
2699 if (low_kmem_size &&
2700 total_size > average_size && /* ignore small node */
2701 low_kmem_size > total_size * 70/100)
2702 return ZONELIST_ORDER_NODE;
2704 return ZONELIST_ORDER_ZONE;
2707 static void set_zonelist_order(void)
2709 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
2710 current_zonelist_order = default_zonelist_order();
2712 current_zonelist_order = user_zonelist_order;
2715 static void build_zonelists(pg_data_t *pgdat)
2719 nodemask_t used_mask;
2720 int local_node, prev_node;
2721 struct zonelist *zonelist;
2722 int order = current_zonelist_order;
2724 /* initialize zonelists */
2725 for (i = 0; i < MAX_ZONELISTS; i++) {
2726 zonelist = pgdat->node_zonelists + i;
2727 zonelist->_zonerefs[0].zone = NULL;
2728 zonelist->_zonerefs[0].zone_idx = 0;
2731 /* NUMA-aware ordering of nodes */
2732 local_node = pgdat->node_id;
2733 load = nr_online_nodes;
2734 prev_node = local_node;
2735 nodes_clear(used_mask);
2737 memset(node_order, 0, sizeof(node_order));
2740 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
2741 int distance = node_distance(local_node, node);
2744 * If another node is sufficiently far away then it is better
2745 * to reclaim pages in a zone before going off node.
2747 if (distance > RECLAIM_DISTANCE)
2748 zone_reclaim_mode = 1;
2751 * We don't want to pressure a particular node.
2752 * So adding penalty to the first node in same
2753 * distance group to make it round-robin.
2755 if (distance != node_distance(local_node, prev_node))
2756 node_load[node] = load;
2760 if (order == ZONELIST_ORDER_NODE)
2761 build_zonelists_in_node_order(pgdat, node);
2763 node_order[j++] = node; /* remember order */
2766 if (order == ZONELIST_ORDER_ZONE) {
2767 /* calculate node order -- i.e., DMA last! */
2768 build_zonelists_in_zone_order(pgdat, j);
2771 build_thisnode_zonelists(pgdat);
2774 /* Construct the zonelist performance cache - see further mmzone.h */
2775 static void build_zonelist_cache(pg_data_t *pgdat)
2777 struct zonelist *zonelist;
2778 struct zonelist_cache *zlc;
2781 zonelist = &pgdat->node_zonelists[0];
2782 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
2783 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2784 for (z = zonelist->_zonerefs; z->zone; z++)
2785 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
2789 #else /* CONFIG_NUMA */
2791 static void set_zonelist_order(void)
2793 current_zonelist_order = ZONELIST_ORDER_ZONE;
2796 static void build_zonelists(pg_data_t *pgdat)
2798 int node, local_node;
2800 struct zonelist *zonelist;
2802 local_node = pgdat->node_id;
2804 zonelist = &pgdat->node_zonelists[0];
2805 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2808 * Now we build the zonelist so that it contains the zones
2809 * of all the other nodes.
2810 * We don't want to pressure a particular node, so when
2811 * building the zones for node N, we make sure that the
2812 * zones coming right after the local ones are those from
2813 * node N+1 (modulo N)
2815 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
2816 if (!node_online(node))
2818 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2821 for (node = 0; node < local_node; node++) {
2822 if (!node_online(node))
2824 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2828 zonelist->_zonerefs[j].zone = NULL;
2829 zonelist->_zonerefs[j].zone_idx = 0;
2832 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
2833 static void build_zonelist_cache(pg_data_t *pgdat)
2835 pgdat->node_zonelists[0].zlcache_ptr = NULL;
2838 #endif /* CONFIG_NUMA */
2841 * Boot pageset table. One per cpu which is going to be used for all
2842 * zones and all nodes. The parameters will be set in such a way
2843 * that an item put on a list will immediately be handed over to
2844 * the buddy list. This is safe since pageset manipulation is done
2845 * with interrupts disabled.
2847 * The boot_pagesets must be kept even after bootup is complete for
2848 * unused processors and/or zones. They do play a role for bootstrapping
2849 * hotplugged processors.
2851 * zoneinfo_show() and maybe other functions do
2852 * not check if the processor is online before following the pageset pointer.
2853 * Other parts of the kernel may not check if the zone is available.
2855 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
2856 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
2858 /* return values int ....just for stop_machine() */
2859 static int __build_all_zonelists(void *dummy)
2865 memset(node_load, 0, sizeof(node_load));
2867 for_each_online_node(nid) {
2868 pg_data_t *pgdat = NODE_DATA(nid);
2870 build_zonelists(pgdat);
2871 build_zonelist_cache(pgdat);
2875 * Initialize the boot_pagesets that are going to be used
2876 * for bootstrapping processors. The real pagesets for
2877 * each zone will be allocated later when the per cpu
2878 * allocator is available.
2880 * boot_pagesets are used also for bootstrapping offline
2881 * cpus if the system is already booted because the pagesets
2882 * are needed to initialize allocators on a specific cpu too.
2883 * F.e. the percpu allocator needs the page allocator which
2884 * needs the percpu allocator in order to allocate its pagesets
2885 * (a chicken-egg dilemma).
2887 for_each_possible_cpu(cpu)
2888 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
2893 void build_all_zonelists(void)
2895 set_zonelist_order();
2897 if (system_state == SYSTEM_BOOTING) {
2898 __build_all_zonelists(NULL);
2899 mminit_verify_zonelist();
2900 cpuset_init_current_mems_allowed();
2902 /* we have to stop all cpus to guarantee there is no user
2904 stop_machine(__build_all_zonelists, NULL, NULL);
2905 /* cpuset refresh routine should be here */
2907 vm_total_pages = nr_free_pagecache_pages();
2909 * Disable grouping by mobility if the number of pages in the
2910 * system is too low to allow the mechanism to work. It would be
2911 * more accurate, but expensive to check per-zone. This check is
2912 * made on memory-hotadd so a system can start with mobility
2913 * disabled and enable it later
2915 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
2916 page_group_by_mobility_disabled = 1;
2918 page_group_by_mobility_disabled = 0;
2920 printk("Built %i zonelists in %s order, mobility grouping %s. "
2921 "Total pages: %ld\n",
2923 zonelist_order_name[current_zonelist_order],
2924 page_group_by_mobility_disabled ? "off" : "on",
2927 printk("Policy zone: %s\n", zone_names[policy_zone]);
2932 * Helper functions to size the waitqueue hash table.
2933 * Essentially these want to choose hash table sizes sufficiently
2934 * large so that collisions trying to wait on pages are rare.
2935 * But in fact, the number of active page waitqueues on typical
2936 * systems is ridiculously low, less than 200. So this is even
2937 * conservative, even though it seems large.
2939 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
2940 * waitqueues, i.e. the size of the waitq table given the number of pages.
2942 #define PAGES_PER_WAITQUEUE 256
2944 #ifndef CONFIG_MEMORY_HOTPLUG
2945 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2947 unsigned long size = 1;
2949 pages /= PAGES_PER_WAITQUEUE;
2951 while (size < pages)
2955 * Once we have dozens or even hundreds of threads sleeping
2956 * on IO we've got bigger problems than wait queue collision.
2957 * Limit the size of the wait table to a reasonable size.
2959 size = min(size, 4096UL);
2961 return max(size, 4UL);
2965 * A zone's size might be changed by hot-add, so it is not possible to determine
2966 * a suitable size for its wait_table. So we use the maximum size now.
2968 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
2970 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
2971 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
2972 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
2974 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
2975 * or more by the traditional way. (See above). It equals:
2977 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
2978 * ia64(16K page size) : = ( 8G + 4M)byte.
2979 * powerpc (64K page size) : = (32G +16M)byte.
2981 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
2988 * This is an integer logarithm so that shifts can be used later
2989 * to extract the more random high bits from the multiplicative
2990 * hash function before the remainder is taken.
2992 static inline unsigned long wait_table_bits(unsigned long size)
2997 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3000 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3001 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3002 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3003 * higher will lead to a bigger reserve which will get freed as contiguous
3004 * blocks as reclaim kicks in
3006 static void setup_zone_migrate_reserve(struct zone *zone)
3008 unsigned long start_pfn, pfn, end_pfn;
3010 unsigned long block_migratetype;
3013 /* Get the start pfn, end pfn and the number of blocks to reserve */
3014 start_pfn = zone->zone_start_pfn;
3015 end_pfn = start_pfn + zone->spanned_pages;
3016 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3020 * Reserve blocks are generally in place to help high-order atomic
3021 * allocations that are short-lived. A min_free_kbytes value that
3022 * would result in more than 2 reserve blocks for atomic allocations
3023 * is assumed to be in place to help anti-fragmentation for the
3024 * future allocation of hugepages at runtime.
3026 reserve = min(2, reserve);
3028 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3029 if (!pfn_valid(pfn))
3031 page = pfn_to_page(pfn);
3033 /* Watch out for overlapping nodes */
3034 if (page_to_nid(page) != zone_to_nid(zone))
3037 /* Blocks with reserved pages will never free, skip them. */
3038 if (PageReserved(page))
3041 block_migratetype = get_pageblock_migratetype(page);
3043 /* If this block is reserved, account for it */
3044 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
3049 /* Suitable for reserving if this block is movable */
3050 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
3051 set_pageblock_migratetype(page, MIGRATE_RESERVE);
3052 move_freepages_block(zone, page, MIGRATE_RESERVE);
3058 * If the reserve is met and this is a previous reserved block,
3061 if (block_migratetype == MIGRATE_RESERVE) {
3062 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3063 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3069 * Initially all pages are reserved - free ones are freed
3070 * up by free_all_bootmem() once the early boot process is
3071 * done. Non-atomic initialization, single-pass.
3073 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3074 unsigned long start_pfn, enum memmap_context context)
3077 unsigned long end_pfn = start_pfn + size;
3081 if (highest_memmap_pfn < end_pfn - 1)
3082 highest_memmap_pfn = end_pfn - 1;
3084 z = &NODE_DATA(nid)->node_zones[zone];
3085 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3087 * There can be holes in boot-time mem_map[]s
3088 * handed to this function. They do not
3089 * exist on hotplugged memory.
3091 if (context == MEMMAP_EARLY) {
3092 if (!early_pfn_valid(pfn))
3094 if (!early_pfn_in_nid(pfn, nid))
3097 page = pfn_to_page(pfn);
3098 set_page_links(page, zone, nid, pfn);
3099 mminit_verify_page_links(page, zone, nid, pfn);
3100 init_page_count(page);
3101 reset_page_mapcount(page);
3102 SetPageReserved(page);
3104 * Mark the block movable so that blocks are reserved for
3105 * movable at startup. This will force kernel allocations
3106 * to reserve their blocks rather than leaking throughout
3107 * the address space during boot when many long-lived
3108 * kernel allocations are made. Later some blocks near
3109 * the start are marked MIGRATE_RESERVE by
3110 * setup_zone_migrate_reserve()
3112 * bitmap is created for zone's valid pfn range. but memmap
3113 * can be created for invalid pages (for alignment)
3114 * check here not to call set_pageblock_migratetype() against
3117 if ((z->zone_start_pfn <= pfn)
3118 && (pfn < z->zone_start_pfn + z->spanned_pages)
3119 && !(pfn & (pageblock_nr_pages - 1)))
3120 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3122 INIT_LIST_HEAD(&page->lru);
3123 #ifdef WANT_PAGE_VIRTUAL
3124 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3125 if (!is_highmem_idx(zone))
3126 set_page_address(page, __va(pfn << PAGE_SHIFT));
3131 static void __meminit zone_init_free_lists(struct zone *zone)
3134 for_each_migratetype_order(order, t) {
3135 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3136 zone->free_area[order].nr_free = 0;
3140 #ifndef __HAVE_ARCH_MEMMAP_INIT
3141 #define memmap_init(size, nid, zone, start_pfn) \
3142 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3145 static int zone_batchsize(struct zone *zone)
3151 * The per-cpu-pages pools are set to around 1000th of the
3152 * size of the zone. But no more than 1/2 of a meg.
3154 * OK, so we don't know how big the cache is. So guess.
3156 batch = zone->present_pages / 1024;
3157 if (batch * PAGE_SIZE > 512 * 1024)
3158 batch = (512 * 1024) / PAGE_SIZE;
3159 batch /= 4; /* We effectively *= 4 below */
3164 * Clamp the batch to a 2^n - 1 value. Having a power
3165 * of 2 value was found to be more likely to have
3166 * suboptimal cache aliasing properties in some cases.
3168 * For example if 2 tasks are alternately allocating
3169 * batches of pages, one task can end up with a lot
3170 * of pages of one half of the possible page colors
3171 * and the other with pages of the other colors.
3173 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3178 /* The deferral and batching of frees should be suppressed under NOMMU
3181 * The problem is that NOMMU needs to be able to allocate large chunks
3182 * of contiguous memory as there's no hardware page translation to
3183 * assemble apparent contiguous memory from discontiguous pages.
3185 * Queueing large contiguous runs of pages for batching, however,
3186 * causes the pages to actually be freed in smaller chunks. As there
3187 * can be a significant delay between the individual batches being
3188 * recycled, this leads to the once large chunks of space being
3189 * fragmented and becoming unavailable for high-order allocations.
3195 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3197 struct per_cpu_pages *pcp;
3200 memset(p, 0, sizeof(*p));
3204 pcp->high = 6 * batch;
3205 pcp->batch = max(1UL, 1 * batch);
3206 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3207 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3211 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3212 * to the value high for the pageset p.
3215 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3218 struct per_cpu_pages *pcp;
3222 pcp->batch = max(1UL, high/4);
3223 if ((high/4) > (PAGE_SHIFT * 8))
3224 pcp->batch = PAGE_SHIFT * 8;
3228 * Allocate per cpu pagesets and initialize them.
3229 * Before this call only boot pagesets were available.
3230 * Boot pagesets will no longer be used by this processorr
3231 * after setup_per_cpu_pageset().
3233 void __init setup_per_cpu_pageset(void)
3238 for_each_populated_zone(zone) {
3239 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3241 for_each_possible_cpu(cpu) {
3242 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3244 setup_pageset(pcp, zone_batchsize(zone));
3246 if (percpu_pagelist_fraction)
3247 setup_pagelist_highmark(pcp,
3248 (zone->present_pages /
3249 percpu_pagelist_fraction));
3254 static noinline __init_refok
3255 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3258 struct pglist_data *pgdat = zone->zone_pgdat;
3262 * The per-page waitqueue mechanism uses hashed waitqueues
3265 zone->wait_table_hash_nr_entries =
3266 wait_table_hash_nr_entries(zone_size_pages);
3267 zone->wait_table_bits =
3268 wait_table_bits(zone->wait_table_hash_nr_entries);
3269 alloc_size = zone->wait_table_hash_nr_entries
3270 * sizeof(wait_queue_head_t);
3272 if (!slab_is_available()) {
3273 zone->wait_table = (wait_queue_head_t *)
3274 alloc_bootmem_node(pgdat, alloc_size);
3277 * This case means that a zone whose size was 0 gets new memory
3278 * via memory hot-add.
3279 * But it may be the case that a new node was hot-added. In
3280 * this case vmalloc() will not be able to use this new node's
3281 * memory - this wait_table must be initialized to use this new
3282 * node itself as well.
3283 * To use this new node's memory, further consideration will be
3286 zone->wait_table = vmalloc(alloc_size);
3288 if (!zone->wait_table)
3291 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3292 init_waitqueue_head(zone->wait_table + i);
3297 static int __zone_pcp_update(void *data)
3299 struct zone *zone = data;
3301 unsigned long batch = zone_batchsize(zone), flags;
3303 for_each_possible_cpu(cpu) {
3304 struct per_cpu_pageset *pset;
3305 struct per_cpu_pages *pcp;
3307 pset = per_cpu_ptr(zone->pageset, cpu);
3310 local_irq_save(flags);
3311 free_pcppages_bulk(zone, pcp->count, pcp);
3312 setup_pageset(pset, batch);
3313 local_irq_restore(flags);
3318 void zone_pcp_update(struct zone *zone)
3320 stop_machine(__zone_pcp_update, zone, NULL);
3323 static __meminit void zone_pcp_init(struct zone *zone)
3326 * per cpu subsystem is not up at this point. The following code
3327 * relies on the ability of the linker to provide the
3328 * offset of a (static) per cpu variable into the per cpu area.
3330 zone->pageset = &boot_pageset;
3332 if (zone->present_pages)
3333 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3334 zone->name, zone->present_pages,
3335 zone_batchsize(zone));
3338 __meminit int init_currently_empty_zone(struct zone *zone,
3339 unsigned long zone_start_pfn,
3341 enum memmap_context context)
3343 struct pglist_data *pgdat = zone->zone_pgdat;
3345 ret = zone_wait_table_init(zone, size);
3348 pgdat->nr_zones = zone_idx(zone) + 1;
3350 zone->zone_start_pfn = zone_start_pfn;
3352 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3353 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3355 (unsigned long)zone_idx(zone),
3356 zone_start_pfn, (zone_start_pfn + size));
3358 zone_init_free_lists(zone);
3363 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3365 * Basic iterator support. Return the first range of PFNs for a node
3366 * Note: nid == MAX_NUMNODES returns first region regardless of node
3368 static int __meminit first_active_region_index_in_nid(int nid)
3372 for (i = 0; i < nr_nodemap_entries; i++)
3373 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3380 * Basic iterator support. Return the next active range of PFNs for a node
3381 * Note: nid == MAX_NUMNODES returns next region regardless of node
3383 static int __meminit next_active_region_index_in_nid(int index, int nid)
3385 for (index = index + 1; index < nr_nodemap_entries; index++)
3386 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3392 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3394 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3395 * Architectures may implement their own version but if add_active_range()
3396 * was used and there are no special requirements, this is a convenient
3399 int __meminit __early_pfn_to_nid(unsigned long pfn)
3403 for (i = 0; i < nr_nodemap_entries; i++) {
3404 unsigned long start_pfn = early_node_map[i].start_pfn;
3405 unsigned long end_pfn = early_node_map[i].end_pfn;
3407 if (start_pfn <= pfn && pfn < end_pfn)
3408 return early_node_map[i].nid;
3410 /* This is a memory hole */
3413 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3415 int __meminit early_pfn_to_nid(unsigned long pfn)
3419 nid = __early_pfn_to_nid(pfn);
3422 /* just returns 0 */
3426 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3427 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3431 nid = __early_pfn_to_nid(pfn);
3432 if (nid >= 0 && nid != node)
3438 /* Basic iterator support to walk early_node_map[] */
3439 #define for_each_active_range_index_in_nid(i, nid) \
3440 for (i = first_active_region_index_in_nid(nid); i != -1; \
3441 i = next_active_region_index_in_nid(i, nid))
3444 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3445 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3446 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3448 * If an architecture guarantees that all ranges registered with
3449 * add_active_ranges() contain no holes and may be freed, this
3450 * this function may be used instead of calling free_bootmem() manually.
3452 void __init free_bootmem_with_active_regions(int nid,
3453 unsigned long max_low_pfn)
3457 for_each_active_range_index_in_nid(i, nid) {
3458 unsigned long size_pages = 0;
3459 unsigned long end_pfn = early_node_map[i].end_pfn;
3461 if (early_node_map[i].start_pfn >= max_low_pfn)
3464 if (end_pfn > max_low_pfn)
3465 end_pfn = max_low_pfn;
3467 size_pages = end_pfn - early_node_map[i].start_pfn;
3468 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3469 PFN_PHYS(early_node_map[i].start_pfn),
3470 size_pages << PAGE_SHIFT);
3474 int __init add_from_early_node_map(struct range *range, int az,
3475 int nr_range, int nid)
3480 /* need to go over early_node_map to find out good range for node */
3481 for_each_active_range_index_in_nid(i, nid) {
3482 start = early_node_map[i].start_pfn;
3483 end = early_node_map[i].end_pfn;
3484 nr_range = add_range(range, az, nr_range, start, end);
3489 #ifdef CONFIG_NO_BOOTMEM
3490 void * __init __alloc_memory_core_early(int nid, u64 size, u64 align,
3491 u64 goal, u64 limit)
3496 /* need to go over early_node_map to find out good range for node */
3497 for_each_active_range_index_in_nid(i, nid) {
3499 u64 ei_start, ei_last;
3501 ei_last = early_node_map[i].end_pfn;
3502 ei_last <<= PAGE_SHIFT;
3503 ei_start = early_node_map[i].start_pfn;
3504 ei_start <<= PAGE_SHIFT;
3505 addr = find_early_area(ei_start, ei_last,
3506 goal, limit, size, align);
3512 printk(KERN_DEBUG "alloc (nid=%d %llx - %llx) (%llx - %llx) %llx %llx => %llx\n",
3514 ei_start, ei_last, goal, limit, size,
3518 ptr = phys_to_virt(addr);
3519 memset(ptr, 0, size);
3520 reserve_early_without_check(addr, addr + size, "BOOTMEM");
3529 void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3534 for_each_active_range_index_in_nid(i, nid) {
3535 ret = work_fn(early_node_map[i].start_pfn,
3536 early_node_map[i].end_pfn, data);
3542 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3543 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3545 * If an architecture guarantees that all ranges registered with
3546 * add_active_ranges() contain no holes and may be freed, this
3547 * function may be used instead of calling memory_present() manually.
3549 void __init sparse_memory_present_with_active_regions(int nid)
3553 for_each_active_range_index_in_nid(i, nid)
3554 memory_present(early_node_map[i].nid,
3555 early_node_map[i].start_pfn,
3556 early_node_map[i].end_pfn);
3560 * get_pfn_range_for_nid - Return the start and end page frames for a node
3561 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3562 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3563 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3565 * It returns the start and end page frame of a node based on information
3566 * provided by an arch calling add_active_range(). If called for a node
3567 * with no available memory, a warning is printed and the start and end
3570 void __meminit get_pfn_range_for_nid(unsigned int nid,
3571 unsigned long *start_pfn, unsigned long *end_pfn)
3577 for_each_active_range_index_in_nid(i, nid) {
3578 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3579 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3582 if (*start_pfn == -1UL)
3587 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3588 * assumption is made that zones within a node are ordered in monotonic
3589 * increasing memory addresses so that the "highest" populated zone is used
3591 static void __init find_usable_zone_for_movable(void)
3594 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3595 if (zone_index == ZONE_MOVABLE)
3598 if (arch_zone_highest_possible_pfn[zone_index] >
3599 arch_zone_lowest_possible_pfn[zone_index])
3603 VM_BUG_ON(zone_index == -1);
3604 movable_zone = zone_index;
3608 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
3609 * because it is sized independant of architecture. Unlike the other zones,
3610 * the starting point for ZONE_MOVABLE is not fixed. It may be different
3611 * in each node depending on the size of each node and how evenly kernelcore
3612 * is distributed. This helper function adjusts the zone ranges
3613 * provided by the architecture for a given node by using the end of the
3614 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
3615 * zones within a node are in order of monotonic increases memory addresses
3617 static void __meminit adjust_zone_range_for_zone_movable(int nid,
3618 unsigned long zone_type,
3619 unsigned long node_start_pfn,
3620 unsigned long node_end_pfn,
3621 unsigned long *zone_start_pfn,
3622 unsigned long *zone_end_pfn)
3624 /* Only adjust if ZONE_MOVABLE is on this node */
3625 if (zone_movable_pfn[nid]) {
3626 /* Size ZONE_MOVABLE */
3627 if (zone_type == ZONE_MOVABLE) {
3628 *zone_start_pfn = zone_movable_pfn[nid];
3629 *zone_end_pfn = min(node_end_pfn,
3630 arch_zone_highest_possible_pfn[movable_zone]);
3632 /* Adjust for ZONE_MOVABLE starting within this range */
3633 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
3634 *zone_end_pfn > zone_movable_pfn[nid]) {
3635 *zone_end_pfn = zone_movable_pfn[nid];
3637 /* Check if this whole range is within ZONE_MOVABLE */
3638 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
3639 *zone_start_pfn = *zone_end_pfn;
3644 * Return the number of pages a zone spans in a node, including holes
3645 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
3647 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
3648 unsigned long zone_type,
3649 unsigned long *ignored)
3651 unsigned long node_start_pfn, node_end_pfn;
3652 unsigned long zone_start_pfn, zone_end_pfn;
3654 /* Get the start and end of the node and zone */
3655 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3656 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
3657 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
3658 adjust_zone_range_for_zone_movable(nid, zone_type,
3659 node_start_pfn, node_end_pfn,
3660 &zone_start_pfn, &zone_end_pfn);
3662 /* Check that this node has pages within the zone's required range */
3663 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
3666 /* Move the zone boundaries inside the node if necessary */
3667 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
3668 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
3670 /* Return the spanned pages */
3671 return zone_end_pfn - zone_start_pfn;
3675 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
3676 * then all holes in the requested range will be accounted for.
3678 unsigned long __meminit __absent_pages_in_range(int nid,
3679 unsigned long range_start_pfn,
3680 unsigned long range_end_pfn)
3683 unsigned long prev_end_pfn = 0, hole_pages = 0;
3684 unsigned long start_pfn;
3686 /* Find the end_pfn of the first active range of pfns in the node */
3687 i = first_active_region_index_in_nid(nid);
3691 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3693 /* Account for ranges before physical memory on this node */
3694 if (early_node_map[i].start_pfn > range_start_pfn)
3695 hole_pages = prev_end_pfn - range_start_pfn;
3697 /* Find all holes for the zone within the node */
3698 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
3700 /* No need to continue if prev_end_pfn is outside the zone */
3701 if (prev_end_pfn >= range_end_pfn)
3704 /* Make sure the end of the zone is not within the hole */
3705 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
3706 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
3708 /* Update the hole size cound and move on */
3709 if (start_pfn > range_start_pfn) {
3710 BUG_ON(prev_end_pfn > start_pfn);
3711 hole_pages += start_pfn - prev_end_pfn;
3713 prev_end_pfn = early_node_map[i].end_pfn;
3716 /* Account for ranges past physical memory on this node */
3717 if (range_end_pfn > prev_end_pfn)
3718 hole_pages += range_end_pfn -
3719 max(range_start_pfn, prev_end_pfn);
3725 * absent_pages_in_range - Return number of page frames in holes within a range
3726 * @start_pfn: The start PFN to start searching for holes
3727 * @end_pfn: The end PFN to stop searching for holes
3729 * It returns the number of pages frames in memory holes within a range.
3731 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
3732 unsigned long end_pfn)
3734 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
3737 /* Return the number of page frames in holes in a zone on a node */
3738 static unsigned long __meminit zone_absent_pages_in_node(int nid,
3739 unsigned long zone_type,
3740 unsigned long *ignored)
3742 unsigned long node_start_pfn, node_end_pfn;
3743 unsigned long zone_start_pfn, zone_end_pfn;
3745 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
3746 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
3748 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
3751 adjust_zone_range_for_zone_movable(nid, zone_type,
3752 node_start_pfn, node_end_pfn,
3753 &zone_start_pfn, &zone_end_pfn);
3754 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
3758 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
3759 unsigned long zone_type,
3760 unsigned long *zones_size)
3762 return zones_size[zone_type];
3765 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
3766 unsigned long zone_type,
3767 unsigned long *zholes_size)
3772 return zholes_size[zone_type];
3777 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
3778 unsigned long *zones_size, unsigned long *zholes_size)
3780 unsigned long realtotalpages, totalpages = 0;
3783 for (i = 0; i < MAX_NR_ZONES; i++)
3784 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
3786 pgdat->node_spanned_pages = totalpages;
3788 realtotalpages = totalpages;
3789 for (i = 0; i < MAX_NR_ZONES; i++)
3791 zone_absent_pages_in_node(pgdat->node_id, i,
3793 pgdat->node_present_pages = realtotalpages;
3794 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
3798 #ifndef CONFIG_SPARSEMEM
3800 * Calculate the size of the zone->blockflags rounded to an unsigned long
3801 * Start by making sure zonesize is a multiple of pageblock_order by rounding
3802 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
3803 * round what is now in bits to nearest long in bits, then return it in
3806 static unsigned long __init usemap_size(unsigned long zonesize)
3808 unsigned long usemapsize;
3810 usemapsize = roundup(zonesize, pageblock_nr_pages);
3811 usemapsize = usemapsize >> pageblock_order;
3812 usemapsize *= NR_PAGEBLOCK_BITS;
3813 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
3815 return usemapsize / 8;
3818 static void __init setup_usemap(struct pglist_data *pgdat,
3819 struct zone *zone, unsigned long zonesize)
3821 unsigned long usemapsize = usemap_size(zonesize);
3822 zone->pageblock_flags = NULL;
3824 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize);
3827 static void inline setup_usemap(struct pglist_data *pgdat,
3828 struct zone *zone, unsigned long zonesize) {}
3829 #endif /* CONFIG_SPARSEMEM */
3831 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
3833 /* Return a sensible default order for the pageblock size. */
3834 static inline int pageblock_default_order(void)
3836 if (HPAGE_SHIFT > PAGE_SHIFT)
3837 return HUGETLB_PAGE_ORDER;
3842 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
3843 static inline void __init set_pageblock_order(unsigned int order)
3845 /* Check that pageblock_nr_pages has not already been setup */
3846 if (pageblock_order)
3850 * Assume the largest contiguous order of interest is a huge page.
3851 * This value may be variable depending on boot parameters on IA64
3853 pageblock_order = order;
3855 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3858 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
3859 * and pageblock_default_order() are unused as pageblock_order is set
3860 * at compile-time. See include/linux/pageblock-flags.h for the values of
3861 * pageblock_order based on the kernel config
3863 static inline int pageblock_default_order(unsigned int order)
3867 #define set_pageblock_order(x) do {} while (0)
3869 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
3872 * Set up the zone data structures:
3873 * - mark all pages reserved
3874 * - mark all memory queues empty
3875 * - clear the memory bitmaps
3877 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
3878 unsigned long *zones_size, unsigned long *zholes_size)
3881 int nid = pgdat->node_id;
3882 unsigned long zone_start_pfn = pgdat->node_start_pfn;
3885 pgdat_resize_init(pgdat);
3886 pgdat->nr_zones = 0;
3887 init_waitqueue_head(&pgdat->kswapd_wait);
3888 pgdat->kswapd_max_order = 0;
3889 pgdat_page_cgroup_init(pgdat);
3891 for (j = 0; j < MAX_NR_ZONES; j++) {
3892 struct zone *zone = pgdat->node_zones + j;
3893 unsigned long size, realsize, memmap_pages;
3896 size = zone_spanned_pages_in_node(nid, j, zones_size);
3897 realsize = size - zone_absent_pages_in_node(nid, j,
3901 * Adjust realsize so that it accounts for how much memory
3902 * is used by this zone for memmap. This affects the watermark
3903 * and per-cpu initialisations
3906 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
3907 if (realsize >= memmap_pages) {
3908 realsize -= memmap_pages;
3911 " %s zone: %lu pages used for memmap\n",
3912 zone_names[j], memmap_pages);
3915 " %s zone: %lu pages exceeds realsize %lu\n",
3916 zone_names[j], memmap_pages, realsize);
3918 /* Account for reserved pages */
3919 if (j == 0 && realsize > dma_reserve) {
3920 realsize -= dma_reserve;
3921 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
3922 zone_names[0], dma_reserve);
3925 if (!is_highmem_idx(j))
3926 nr_kernel_pages += realsize;
3927 nr_all_pages += realsize;
3929 zone->spanned_pages = size;
3930 zone->present_pages = realsize;
3933 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
3935 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
3937 zone->name = zone_names[j];
3938 spin_lock_init(&zone->lock);
3939 spin_lock_init(&zone->lru_lock);
3940 zone_seqlock_init(zone);
3941 zone->zone_pgdat = pgdat;
3943 zone->prev_priority = DEF_PRIORITY;
3945 zone_pcp_init(zone);
3947 INIT_LIST_HEAD(&zone->lru[l].list);
3948 zone->reclaim_stat.nr_saved_scan[l] = 0;
3950 zone->reclaim_stat.recent_rotated[0] = 0;
3951 zone->reclaim_stat.recent_rotated[1] = 0;
3952 zone->reclaim_stat.recent_scanned[0] = 0;
3953 zone->reclaim_stat.recent_scanned[1] = 0;
3954 zap_zone_vm_stats(zone);
3959 set_pageblock_order(pageblock_default_order());
3960 setup_usemap(pgdat, zone, size);
3961 ret = init_currently_empty_zone(zone, zone_start_pfn,
3962 size, MEMMAP_EARLY);
3964 memmap_init(size, nid, j, zone_start_pfn);
3965 zone_start_pfn += size;
3969 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
3971 /* Skip empty nodes */
3972 if (!pgdat->node_spanned_pages)
3975 #ifdef CONFIG_FLAT_NODE_MEM_MAP
3976 /* ia64 gets its own node_mem_map, before this, without bootmem */
3977 if (!pgdat->node_mem_map) {
3978 unsigned long size, start, end;
3982 * The zone's endpoints aren't required to be MAX_ORDER
3983 * aligned but the node_mem_map endpoints must be in order
3984 * for the buddy allocator to function correctly.
3986 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
3987 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
3988 end = ALIGN(end, MAX_ORDER_NR_PAGES);
3989 size = (end - start) * sizeof(struct page);
3990 map = alloc_remap(pgdat->node_id, size);
3992 map = alloc_bootmem_node(pgdat, size);
3993 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
3995 #ifndef CONFIG_NEED_MULTIPLE_NODES
3997 * With no DISCONTIG, the global mem_map is just set as node 0's
3999 if (pgdat == NODE_DATA(0)) {
4000 mem_map = NODE_DATA(0)->node_mem_map;
4001 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4002 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4003 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4004 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4007 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4010 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4011 unsigned long node_start_pfn, unsigned long *zholes_size)
4013 pg_data_t *pgdat = NODE_DATA(nid);
4015 pgdat->node_id = nid;
4016 pgdat->node_start_pfn = node_start_pfn;
4017 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4019 alloc_node_mem_map(pgdat);
4020 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4021 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4022 nid, (unsigned long)pgdat,
4023 (unsigned long)pgdat->node_mem_map);
4026 free_area_init_core(pgdat, zones_size, zholes_size);
4029 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4031 #if MAX_NUMNODES > 1
4033 * Figure out the number of possible node ids.
4035 static void __init setup_nr_node_ids(void)
4038 unsigned int highest = 0;
4040 for_each_node_mask(node, node_possible_map)
4042 nr_node_ids = highest + 1;
4045 static inline void setup_nr_node_ids(void)
4051 * add_active_range - Register a range of PFNs backed by physical memory
4052 * @nid: The node ID the range resides on
4053 * @start_pfn: The start PFN of the available physical memory
4054 * @end_pfn: The end PFN of the available physical memory
4056 * These ranges are stored in an early_node_map[] and later used by
4057 * free_area_init_nodes() to calculate zone sizes and holes. If the
4058 * range spans a memory hole, it is up to the architecture to ensure
4059 * the memory is not freed by the bootmem allocator. If possible
4060 * the range being registered will be merged with existing ranges.
4062 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4063 unsigned long end_pfn)
4067 mminit_dprintk(MMINIT_TRACE, "memory_register",
4068 "Entering add_active_range(%d, %#lx, %#lx) "
4069 "%d entries of %d used\n",
4070 nid, start_pfn, end_pfn,
4071 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4073 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4075 /* Merge with existing active regions if possible */
4076 for (i = 0; i < nr_nodemap_entries; i++) {
4077 if (early_node_map[i].nid != nid)
4080 /* Skip if an existing region covers this new one */
4081 if (start_pfn >= early_node_map[i].start_pfn &&
4082 end_pfn <= early_node_map[i].end_pfn)
4085 /* Merge forward if suitable */
4086 if (start_pfn <= early_node_map[i].end_pfn &&
4087 end_pfn > early_node_map[i].end_pfn) {
4088 early_node_map[i].end_pfn = end_pfn;
4092 /* Merge backward if suitable */
4093 if (start_pfn < early_node_map[i].start_pfn &&
4094 end_pfn >= early_node_map[i].start_pfn) {
4095 early_node_map[i].start_pfn = start_pfn;
4100 /* Check that early_node_map is large enough */
4101 if (i >= MAX_ACTIVE_REGIONS) {
4102 printk(KERN_CRIT "More than %d memory regions, truncating\n",
4103 MAX_ACTIVE_REGIONS);
4107 early_node_map[i].nid = nid;
4108 early_node_map[i].start_pfn = start_pfn;
4109 early_node_map[i].end_pfn = end_pfn;
4110 nr_nodemap_entries = i + 1;
4114 * remove_active_range - Shrink an existing registered range of PFNs
4115 * @nid: The node id the range is on that should be shrunk
4116 * @start_pfn: The new PFN of the range
4117 * @end_pfn: The new PFN of the range
4119 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4120 * The map is kept near the end physical page range that has already been
4121 * registered. This function allows an arch to shrink an existing registered
4124 void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4125 unsigned long end_pfn)
4130 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4131 nid, start_pfn, end_pfn);
4133 /* Find the old active region end and shrink */
4134 for_each_active_range_index_in_nid(i, nid) {
4135 if (early_node_map[i].start_pfn >= start_pfn &&
4136 early_node_map[i].end_pfn <= end_pfn) {
4138 early_node_map[i].start_pfn = 0;
4139 early_node_map[i].end_pfn = 0;
4143 if (early_node_map[i].start_pfn < start_pfn &&
4144 early_node_map[i].end_pfn > start_pfn) {
4145 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4146 early_node_map[i].end_pfn = start_pfn;
4147 if (temp_end_pfn > end_pfn)
4148 add_active_range(nid, end_pfn, temp_end_pfn);
4151 if (early_node_map[i].start_pfn >= start_pfn &&
4152 early_node_map[i].end_pfn > end_pfn &&
4153 early_node_map[i].start_pfn < end_pfn) {
4154 early_node_map[i].start_pfn = end_pfn;
4162 /* remove the blank ones */
4163 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4164 if (early_node_map[i].nid != nid)
4166 if (early_node_map[i].end_pfn)
4168 /* we found it, get rid of it */
4169 for (j = i; j < nr_nodemap_entries - 1; j++)
4170 memcpy(&early_node_map[j], &early_node_map[j+1],
4171 sizeof(early_node_map[j]));
4172 j = nr_nodemap_entries - 1;
4173 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4174 nr_nodemap_entries--;
4179 * remove_all_active_ranges - Remove all currently registered regions
4181 * During discovery, it may be found that a table like SRAT is invalid
4182 * and an alternative discovery method must be used. This function removes
4183 * all currently registered regions.
4185 void __init remove_all_active_ranges(void)
4187 memset(early_node_map, 0, sizeof(early_node_map));
4188 nr_nodemap_entries = 0;
4191 /* Compare two active node_active_regions */
4192 static int __init cmp_node_active_region(const void *a, const void *b)
4194 struct node_active_region *arange = (struct node_active_region *)a;
4195 struct node_active_region *brange = (struct node_active_region *)b;
4197 /* Done this way to avoid overflows */
4198 if (arange->start_pfn > brange->start_pfn)
4200 if (arange->start_pfn < brange->start_pfn)
4206 /* sort the node_map by start_pfn */
4207 void __init sort_node_map(void)
4209 sort(early_node_map, (size_t)nr_nodemap_entries,
4210 sizeof(struct node_active_region),
4211 cmp_node_active_region, NULL);
4214 /* Find the lowest pfn for a node */
4215 static unsigned long __init find_min_pfn_for_node(int nid)
4218 unsigned long min_pfn = ULONG_MAX;
4220 /* Assuming a sorted map, the first range found has the starting pfn */
4221 for_each_active_range_index_in_nid(i, nid)
4222 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4224 if (min_pfn == ULONG_MAX) {
4226 "Could not find start_pfn for node %d\n", nid);
4234 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4236 * It returns the minimum PFN based on information provided via
4237 * add_active_range().
4239 unsigned long __init find_min_pfn_with_active_regions(void)
4241 return find_min_pfn_for_node(MAX_NUMNODES);
4245 * early_calculate_totalpages()
4246 * Sum pages in active regions for movable zone.
4247 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4249 static unsigned long __init early_calculate_totalpages(void)
4252 unsigned long totalpages = 0;
4254 for (i = 0; i < nr_nodemap_entries; i++) {
4255 unsigned long pages = early_node_map[i].end_pfn -
4256 early_node_map[i].start_pfn;
4257 totalpages += pages;
4259 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4265 * Find the PFN the Movable zone begins in each node. Kernel memory
4266 * is spread evenly between nodes as long as the nodes have enough
4267 * memory. When they don't, some nodes will have more kernelcore than
4270 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4273 unsigned long usable_startpfn;
4274 unsigned long kernelcore_node, kernelcore_remaining;
4275 /* save the state before borrow the nodemask */
4276 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4277 unsigned long totalpages = early_calculate_totalpages();
4278 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4281 * If movablecore was specified, calculate what size of
4282 * kernelcore that corresponds so that memory usable for
4283 * any allocation type is evenly spread. If both kernelcore
4284 * and movablecore are specified, then the value of kernelcore
4285 * will be used for required_kernelcore if it's greater than
4286 * what movablecore would have allowed.
4288 if (required_movablecore) {
4289 unsigned long corepages;
4292 * Round-up so that ZONE_MOVABLE is at least as large as what
4293 * was requested by the user
4295 required_movablecore =
4296 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4297 corepages = totalpages - required_movablecore;
4299 required_kernelcore = max(required_kernelcore, corepages);
4302 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4303 if (!required_kernelcore)
4306 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4307 find_usable_zone_for_movable();
4308 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4311 /* Spread kernelcore memory as evenly as possible throughout nodes */
4312 kernelcore_node = required_kernelcore / usable_nodes;
4313 for_each_node_state(nid, N_HIGH_MEMORY) {
4315 * Recalculate kernelcore_node if the division per node
4316 * now exceeds what is necessary to satisfy the requested
4317 * amount of memory for the kernel
4319 if (required_kernelcore < kernelcore_node)
4320 kernelcore_node = required_kernelcore / usable_nodes;
4323 * As the map is walked, we track how much memory is usable
4324 * by the kernel using kernelcore_remaining. When it is
4325 * 0, the rest of the node is usable by ZONE_MOVABLE
4327 kernelcore_remaining = kernelcore_node;
4329 /* Go through each range of PFNs within this node */
4330 for_each_active_range_index_in_nid(i, nid) {
4331 unsigned long start_pfn, end_pfn;
4332 unsigned long size_pages;
4334 start_pfn = max(early_node_map[i].start_pfn,
4335 zone_movable_pfn[nid]);
4336 end_pfn = early_node_map[i].end_pfn;
4337 if (start_pfn >= end_pfn)
4340 /* Account for what is only usable for kernelcore */
4341 if (start_pfn < usable_startpfn) {
4342 unsigned long kernel_pages;
4343 kernel_pages = min(end_pfn, usable_startpfn)
4346 kernelcore_remaining -= min(kernel_pages,
4347 kernelcore_remaining);
4348 required_kernelcore -= min(kernel_pages,
4349 required_kernelcore);
4351 /* Continue if range is now fully accounted */
4352 if (end_pfn <= usable_startpfn) {
4355 * Push zone_movable_pfn to the end so
4356 * that if we have to rebalance
4357 * kernelcore across nodes, we will
4358 * not double account here
4360 zone_movable_pfn[nid] = end_pfn;
4363 start_pfn = usable_startpfn;
4367 * The usable PFN range for ZONE_MOVABLE is from
4368 * start_pfn->end_pfn. Calculate size_pages as the
4369 * number of pages used as kernelcore
4371 size_pages = end_pfn - start_pfn;
4372 if (size_pages > kernelcore_remaining)
4373 size_pages = kernelcore_remaining;
4374 zone_movable_pfn[nid] = start_pfn + size_pages;
4377 * Some kernelcore has been met, update counts and
4378 * break if the kernelcore for this node has been
4381 required_kernelcore -= min(required_kernelcore,
4383 kernelcore_remaining -= size_pages;
4384 if (!kernelcore_remaining)
4390 * If there is still required_kernelcore, we do another pass with one
4391 * less node in the count. This will push zone_movable_pfn[nid] further
4392 * along on the nodes that still have memory until kernelcore is
4396 if (usable_nodes && required_kernelcore > usable_nodes)
4399 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4400 for (nid = 0; nid < MAX_NUMNODES; nid++)
4401 zone_movable_pfn[nid] =
4402 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4405 /* restore the node_state */
4406 node_states[N_HIGH_MEMORY] = saved_node_state;
4409 /* Any regular memory on that node ? */
4410 static void check_for_regular_memory(pg_data_t *pgdat)
4412 #ifdef CONFIG_HIGHMEM
4413 enum zone_type zone_type;
4415 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4416 struct zone *zone = &pgdat->node_zones[zone_type];
4417 if (zone->present_pages)
4418 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4424 * free_area_init_nodes - Initialise all pg_data_t and zone data
4425 * @max_zone_pfn: an array of max PFNs for each zone
4427 * This will call free_area_init_node() for each active node in the system.
4428 * Using the page ranges provided by add_active_range(), the size of each
4429 * zone in each node and their holes is calculated. If the maximum PFN
4430 * between two adjacent zones match, it is assumed that the zone is empty.
4431 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4432 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4433 * starts where the previous one ended. For example, ZONE_DMA32 starts
4434 * at arch_max_dma_pfn.
4436 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4441 /* Sort early_node_map as initialisation assumes it is sorted */
4444 /* Record where the zone boundaries are */
4445 memset(arch_zone_lowest_possible_pfn, 0,
4446 sizeof(arch_zone_lowest_possible_pfn));
4447 memset(arch_zone_highest_possible_pfn, 0,
4448 sizeof(arch_zone_highest_possible_pfn));
4449 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4450 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4451 for (i = 1; i < MAX_NR_ZONES; i++) {
4452 if (i == ZONE_MOVABLE)
4454 arch_zone_lowest_possible_pfn[i] =
4455 arch_zone_highest_possible_pfn[i-1];
4456 arch_zone_highest_possible_pfn[i] =
4457 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4459 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4460 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4462 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4463 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4464 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4466 /* Print out the zone ranges */
4467 printk("Zone PFN ranges:\n");
4468 for (i = 0; i < MAX_NR_ZONES; i++) {
4469 if (i == ZONE_MOVABLE)
4471 printk(" %-8s ", zone_names[i]);
4472 if (arch_zone_lowest_possible_pfn[i] ==
4473 arch_zone_highest_possible_pfn[i])
4476 printk("%0#10lx -> %0#10lx\n",
4477 arch_zone_lowest_possible_pfn[i],
4478 arch_zone_highest_possible_pfn[i]);
4481 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4482 printk("Movable zone start PFN for each node\n");
4483 for (i = 0; i < MAX_NUMNODES; i++) {
4484 if (zone_movable_pfn[i])
4485 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4488 /* Print out the early_node_map[] */
4489 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4490 for (i = 0; i < nr_nodemap_entries; i++)
4491 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4492 early_node_map[i].start_pfn,
4493 early_node_map[i].end_pfn);
4495 /* Initialise every node */
4496 mminit_verify_pageflags_layout();
4497 setup_nr_node_ids();
4498 for_each_online_node(nid) {
4499 pg_data_t *pgdat = NODE_DATA(nid);
4500 free_area_init_node(nid, NULL,
4501 find_min_pfn_for_node(nid), NULL);
4503 /* Any memory on that node */
4504 if (pgdat->node_present_pages)
4505 node_set_state(nid, N_HIGH_MEMORY);
4506 check_for_regular_memory(pgdat);
4510 static int __init cmdline_parse_core(char *p, unsigned long *core)
4512 unsigned long long coremem;
4516 coremem = memparse(p, &p);
4517 *core = coremem >> PAGE_SHIFT;
4519 /* Paranoid check that UL is enough for the coremem value */
4520 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4526 * kernelcore=size sets the amount of memory for use for allocations that
4527 * cannot be reclaimed or migrated.
4529 static int __init cmdline_parse_kernelcore(char *p)
4531 return cmdline_parse_core(p, &required_kernelcore);
4535 * movablecore=size sets the amount of memory for use for allocations that
4536 * can be reclaimed or migrated.
4538 static int __init cmdline_parse_movablecore(char *p)
4540 return cmdline_parse_core(p, &required_movablecore);
4543 early_param("kernelcore", cmdline_parse_kernelcore);
4544 early_param("movablecore", cmdline_parse_movablecore);
4546 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4549 * set_dma_reserve - set the specified number of pages reserved in the first zone
4550 * @new_dma_reserve: The number of pages to mark reserved
4552 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4553 * In the DMA zone, a significant percentage may be consumed by kernel image
4554 * and other unfreeable allocations which can skew the watermarks badly. This
4555 * function may optionally be used to account for unfreeable pages in the
4556 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4557 * smaller per-cpu batchsize.
4559 void __init set_dma_reserve(unsigned long new_dma_reserve)
4561 dma_reserve = new_dma_reserve;
4564 #ifndef CONFIG_NEED_MULTIPLE_NODES
4565 struct pglist_data __refdata contig_page_data = {
4566 #ifndef CONFIG_NO_BOOTMEM
4567 .bdata = &bootmem_node_data[0]
4570 EXPORT_SYMBOL(contig_page_data);
4573 void __init free_area_init(unsigned long *zones_size)
4575 free_area_init_node(0, zones_size,
4576 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
4579 static int page_alloc_cpu_notify(struct notifier_block *self,
4580 unsigned long action, void *hcpu)
4582 int cpu = (unsigned long)hcpu;
4584 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
4588 * Spill the event counters of the dead processor
4589 * into the current processors event counters.
4590 * This artificially elevates the count of the current
4593 vm_events_fold_cpu(cpu);
4596 * Zero the differential counters of the dead processor
4597 * so that the vm statistics are consistent.
4599 * This is only okay since the processor is dead and cannot
4600 * race with what we are doing.
4602 refresh_cpu_vm_stats(cpu);
4607 void __init page_alloc_init(void)
4609 hotcpu_notifier(page_alloc_cpu_notify, 0);
4613 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4614 * or min_free_kbytes changes.
4616 static void calculate_totalreserve_pages(void)
4618 struct pglist_data *pgdat;
4619 unsigned long reserve_pages = 0;
4620 enum zone_type i, j;
4622 for_each_online_pgdat(pgdat) {
4623 for (i = 0; i < MAX_NR_ZONES; i++) {
4624 struct zone *zone = pgdat->node_zones + i;
4625 unsigned long max = 0;
4627 /* Find valid and maximum lowmem_reserve in the zone */
4628 for (j = i; j < MAX_NR_ZONES; j++) {
4629 if (zone->lowmem_reserve[j] > max)
4630 max = zone->lowmem_reserve[j];
4633 /* we treat the high watermark as reserved pages. */
4634 max += high_wmark_pages(zone);
4636 if (max > zone->present_pages)
4637 max = zone->present_pages;
4638 reserve_pages += max;
4641 totalreserve_pages = reserve_pages;
4645 * setup_per_zone_lowmem_reserve - called whenever
4646 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4647 * has a correct pages reserved value, so an adequate number of
4648 * pages are left in the zone after a successful __alloc_pages().
4650 static void setup_per_zone_lowmem_reserve(void)
4652 struct pglist_data *pgdat;
4653 enum zone_type j, idx;
4655 for_each_online_pgdat(pgdat) {
4656 for (j = 0; j < MAX_NR_ZONES; j++) {
4657 struct zone *zone = pgdat->node_zones + j;
4658 unsigned long present_pages = zone->present_pages;
4660 zone->lowmem_reserve[j] = 0;
4664 struct zone *lower_zone;
4668 if (sysctl_lowmem_reserve_ratio[idx] < 1)
4669 sysctl_lowmem_reserve_ratio[idx] = 1;
4671 lower_zone = pgdat->node_zones + idx;
4672 lower_zone->lowmem_reserve[j] = present_pages /
4673 sysctl_lowmem_reserve_ratio[idx];
4674 present_pages += lower_zone->present_pages;
4679 /* update totalreserve_pages */
4680 calculate_totalreserve_pages();
4684 * setup_per_zone_wmarks - called when min_free_kbytes changes
4685 * or when memory is hot-{added|removed}
4687 * Ensures that the watermark[min,low,high] values for each zone are set
4688 * correctly with respect to min_free_kbytes.
4690 void setup_per_zone_wmarks(void)
4692 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
4693 unsigned long lowmem_pages = 0;
4695 unsigned long flags;
4697 /* Calculate total number of !ZONE_HIGHMEM pages */
4698 for_each_zone(zone) {
4699 if (!is_highmem(zone))
4700 lowmem_pages += zone->present_pages;
4703 for_each_zone(zone) {
4706 spin_lock_irqsave(&zone->lock, flags);
4707 tmp = (u64)pages_min * zone->present_pages;
4708 do_div(tmp, lowmem_pages);
4709 if (is_highmem(zone)) {
4711 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4712 * need highmem pages, so cap pages_min to a small
4715 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4716 * deltas controls asynch page reclaim, and so should
4717 * not be capped for highmem.
4721 min_pages = zone->present_pages / 1024;
4722 if (min_pages < SWAP_CLUSTER_MAX)
4723 min_pages = SWAP_CLUSTER_MAX;
4724 if (min_pages > 128)
4726 zone->watermark[WMARK_MIN] = min_pages;
4729 * If it's a lowmem zone, reserve a number of pages
4730 * proportionate to the zone's size.
4732 zone->watermark[WMARK_MIN] = tmp;
4735 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
4736 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
4737 setup_zone_migrate_reserve(zone);
4738 spin_unlock_irqrestore(&zone->lock, flags);
4741 /* update totalreserve_pages */
4742 calculate_totalreserve_pages();
4746 * The inactive anon list should be small enough that the VM never has to
4747 * do too much work, but large enough that each inactive page has a chance
4748 * to be referenced again before it is swapped out.
4750 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
4751 * INACTIVE_ANON pages on this zone's LRU, maintained by the
4752 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
4753 * the anonymous pages are kept on the inactive list.
4756 * memory ratio inactive anon
4757 * -------------------------------------
4766 void calculate_zone_inactive_ratio(struct zone *zone)
4768 unsigned int gb, ratio;
4770 /* Zone size in gigabytes */
4771 gb = zone->present_pages >> (30 - PAGE_SHIFT);
4773 ratio = int_sqrt(10 * gb);
4777 zone->inactive_ratio = ratio;
4780 static void __init setup_per_zone_inactive_ratio(void)
4785 calculate_zone_inactive_ratio(zone);
4789 * Initialise min_free_kbytes.
4791 * For small machines we want it small (128k min). For large machines
4792 * we want it large (64MB max). But it is not linear, because network
4793 * bandwidth does not increase linearly with machine size. We use
4795 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
4796 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
4812 static int __init init_per_zone_wmark_min(void)
4814 unsigned long lowmem_kbytes;
4816 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
4818 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
4819 if (min_free_kbytes < 128)
4820 min_free_kbytes = 128;
4821 if (min_free_kbytes > 65536)
4822 min_free_kbytes = 65536;
4823 setup_per_zone_wmarks();
4824 setup_per_zone_lowmem_reserve();
4825 setup_per_zone_inactive_ratio();
4828 module_init(init_per_zone_wmark_min)
4831 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
4832 * that we can call two helper functions whenever min_free_kbytes
4835 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
4836 void __user *buffer, size_t *length, loff_t *ppos)
4838 proc_dointvec(table, write, buffer, length, ppos);
4840 setup_per_zone_wmarks();
4845 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
4846 void __user *buffer, size_t *length, loff_t *ppos)
4851 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
4856 zone->min_unmapped_pages = (zone->present_pages *
4857 sysctl_min_unmapped_ratio) / 100;
4861 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
4862 void __user *buffer, size_t *length, loff_t *ppos)
4867 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
4872 zone->min_slab_pages = (zone->present_pages *
4873 sysctl_min_slab_ratio) / 100;
4879 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
4880 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
4881 * whenever sysctl_lowmem_reserve_ratio changes.
4883 * The reserve ratio obviously has absolutely no relation with the
4884 * minimum watermarks. The lowmem reserve ratio can only make sense
4885 * if in function of the boot time zone sizes.
4887 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
4888 void __user *buffer, size_t *length, loff_t *ppos)
4890 proc_dointvec_minmax(table, write, buffer, length, ppos);
4891 setup_per_zone_lowmem_reserve();
4896 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
4897 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
4898 * can have before it gets flushed back to buddy allocator.
4901 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
4902 void __user *buffer, size_t *length, loff_t *ppos)
4908 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
4909 if (!write || (ret == -EINVAL))
4911 for_each_populated_zone(zone) {
4912 for_each_possible_cpu(cpu) {
4914 high = zone->present_pages / percpu_pagelist_fraction;
4915 setup_pagelist_highmark(
4916 per_cpu_ptr(zone->pageset, cpu), high);
4922 int hashdist = HASHDIST_DEFAULT;
4925 static int __init set_hashdist(char *str)
4929 hashdist = simple_strtoul(str, &str, 0);
4932 __setup("hashdist=", set_hashdist);
4936 * allocate a large system hash table from bootmem
4937 * - it is assumed that the hash table must contain an exact power-of-2
4938 * quantity of entries
4939 * - limit is the number of hash buckets, not the total allocation size
4941 void *__init alloc_large_system_hash(const char *tablename,
4942 unsigned long bucketsize,
4943 unsigned long numentries,
4946 unsigned int *_hash_shift,
4947 unsigned int *_hash_mask,
4948 unsigned long limit)
4950 unsigned long long max = limit;
4951 unsigned long log2qty, size;
4954 /* allow the kernel cmdline to have a say */
4956 /* round applicable memory size up to nearest megabyte */
4957 numentries = nr_kernel_pages;
4958 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
4959 numentries >>= 20 - PAGE_SHIFT;
4960 numentries <<= 20 - PAGE_SHIFT;
4962 /* limit to 1 bucket per 2^scale bytes of low memory */
4963 if (scale > PAGE_SHIFT)
4964 numentries >>= (scale - PAGE_SHIFT);
4966 numentries <<= (PAGE_SHIFT - scale);
4968 /* Make sure we've got at least a 0-order allocation.. */
4969 if (unlikely(flags & HASH_SMALL)) {
4970 /* Makes no sense without HASH_EARLY */
4971 WARN_ON(!(flags & HASH_EARLY));
4972 if (!(numentries >> *_hash_shift)) {
4973 numentries = 1UL << *_hash_shift;
4974 BUG_ON(!numentries);
4976 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
4977 numentries = PAGE_SIZE / bucketsize;
4979 numentries = roundup_pow_of_two(numentries);
4981 /* limit allocation size to 1/16 total memory by default */
4983 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
4984 do_div(max, bucketsize);
4987 if (numentries > max)
4990 log2qty = ilog2(numentries);
4993 size = bucketsize << log2qty;
4994 if (flags & HASH_EARLY)
4995 table = alloc_bootmem_nopanic(size);
4997 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5000 * If bucketsize is not a power-of-two, we may free
5001 * some pages at the end of hash table which
5002 * alloc_pages_exact() automatically does
5004 if (get_order(size) < MAX_ORDER) {
5005 table = alloc_pages_exact(size, GFP_ATOMIC);
5006 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5009 } while (!table && size > PAGE_SIZE && --log2qty);
5012 panic("Failed to allocate %s hash table\n", tablename);
5014 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n",
5017 ilog2(size) - PAGE_SHIFT,
5021 *_hash_shift = log2qty;
5023 *_hash_mask = (1 << log2qty) - 1;
5028 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5029 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5032 #ifdef CONFIG_SPARSEMEM
5033 return __pfn_to_section(pfn)->pageblock_flags;
5035 return zone->pageblock_flags;
5036 #endif /* CONFIG_SPARSEMEM */
5039 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5041 #ifdef CONFIG_SPARSEMEM
5042 pfn &= (PAGES_PER_SECTION-1);
5043 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5045 pfn = pfn - zone->zone_start_pfn;
5046 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5047 #endif /* CONFIG_SPARSEMEM */
5051 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5052 * @page: The page within the block of interest
5053 * @start_bitidx: The first bit of interest to retrieve
5054 * @end_bitidx: The last bit of interest
5055 * returns pageblock_bits flags
5057 unsigned long get_pageblock_flags_group(struct page *page,
5058 int start_bitidx, int end_bitidx)
5061 unsigned long *bitmap;
5062 unsigned long pfn, bitidx;
5063 unsigned long flags = 0;
5064 unsigned long value = 1;
5066 zone = page_zone(page);
5067 pfn = page_to_pfn(page);
5068 bitmap = get_pageblock_bitmap(zone, pfn);
5069 bitidx = pfn_to_bitidx(zone, pfn);
5071 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5072 if (test_bit(bitidx + start_bitidx, bitmap))
5079 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5080 * @page: The page within the block of interest
5081 * @start_bitidx: The first bit of interest
5082 * @end_bitidx: The last bit of interest
5083 * @flags: The flags to set
5085 void set_pageblock_flags_group(struct page *page, unsigned long flags,
5086 int start_bitidx, int end_bitidx)
5089 unsigned long *bitmap;
5090 unsigned long pfn, bitidx;
5091 unsigned long value = 1;
5093 zone = page_zone(page);
5094 pfn = page_to_pfn(page);
5095 bitmap = get_pageblock_bitmap(zone, pfn);
5096 bitidx = pfn_to_bitidx(zone, pfn);
5097 VM_BUG_ON(pfn < zone->zone_start_pfn);
5098 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5100 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5102 __set_bit(bitidx + start_bitidx, bitmap);
5104 __clear_bit(bitidx + start_bitidx, bitmap);
5108 * This is designed as sub function...plz see page_isolation.c also.
5109 * set/clear page block's type to be ISOLATE.
5110 * page allocater never alloc memory from ISOLATE block.
5113 int set_migratetype_isolate(struct page *page)
5116 struct page *curr_page;
5117 unsigned long flags, pfn, iter;
5118 unsigned long immobile = 0;
5119 struct memory_isolate_notify arg;
5124 zone = page_zone(page);
5125 zone_idx = zone_idx(zone);
5127 spin_lock_irqsave(&zone->lock, flags);
5128 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE ||
5129 zone_idx == ZONE_MOVABLE) {
5134 pfn = page_to_pfn(page);
5135 arg.start_pfn = pfn;
5136 arg.nr_pages = pageblock_nr_pages;
5137 arg.pages_found = 0;
5140 * It may be possible to isolate a pageblock even if the
5141 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5142 * notifier chain is used by balloon drivers to return the
5143 * number of pages in a range that are held by the balloon
5144 * driver to shrink memory. If all the pages are accounted for
5145 * by balloons, are free, or on the LRU, isolation can continue.
5146 * Later, for example, when memory hotplug notifier runs, these
5147 * pages reported as "can be isolated" should be isolated(freed)
5148 * by the balloon driver through the memory notifier chain.
5150 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5151 notifier_ret = notifier_to_errno(notifier_ret);
5152 if (notifier_ret || !arg.pages_found)
5155 for (iter = pfn; iter < (pfn + pageblock_nr_pages); iter++) {
5156 if (!pfn_valid_within(pfn))
5159 curr_page = pfn_to_page(iter);
5160 if (!page_count(curr_page) || PageLRU(curr_page))
5166 if (arg.pages_found == immobile)
5171 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5172 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5175 spin_unlock_irqrestore(&zone->lock, flags);
5181 void unset_migratetype_isolate(struct page *page)
5184 unsigned long flags;
5185 zone = page_zone(page);
5186 spin_lock_irqsave(&zone->lock, flags);
5187 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5189 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5190 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5192 spin_unlock_irqrestore(&zone->lock, flags);
5195 #ifdef CONFIG_MEMORY_HOTREMOVE
5197 * All pages in the range must be isolated before calling this.
5200 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5206 unsigned long flags;
5207 /* find the first valid pfn */
5208 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5213 zone = page_zone(pfn_to_page(pfn));
5214 spin_lock_irqsave(&zone->lock, flags);
5216 while (pfn < end_pfn) {
5217 if (!pfn_valid(pfn)) {
5221 page = pfn_to_page(pfn);
5222 BUG_ON(page_count(page));
5223 BUG_ON(!PageBuddy(page));
5224 order = page_order(page);
5225 #ifdef CONFIG_DEBUG_VM
5226 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5227 pfn, 1 << order, end_pfn);
5229 list_del(&page->lru);
5230 rmv_page_order(page);
5231 zone->free_area[order].nr_free--;
5232 __mod_zone_page_state(zone, NR_FREE_PAGES,
5234 for (i = 0; i < (1 << order); i++)
5235 SetPageReserved((page+i));
5236 pfn += (1 << order);
5238 spin_unlock_irqrestore(&zone->lock, flags);
5242 #ifdef CONFIG_MEMORY_FAILURE
5243 bool is_free_buddy_page(struct page *page)
5245 struct zone *zone = page_zone(page);
5246 unsigned long pfn = page_to_pfn(page);
5247 unsigned long flags;
5250 spin_lock_irqsave(&zone->lock, flags);
5251 for (order = 0; order < MAX_ORDER; order++) {
5252 struct page *page_head = page - (pfn & ((1 << order) - 1));
5254 if (PageBuddy(page_head) && page_order(page_head) >= order)
5257 spin_unlock_irqrestore(&zone->lock, flags);
5259 return order < MAX_ORDER;
5263 static struct trace_print_flags pageflag_names[] = {
5264 {1UL << PG_locked, "locked" },
5265 {1UL << PG_error, "error" },
5266 {1UL << PG_referenced, "referenced" },
5267 {1UL << PG_uptodate, "uptodate" },
5268 {1UL << PG_dirty, "dirty" },
5269 {1UL << PG_lru, "lru" },
5270 {1UL << PG_active, "active" },
5271 {1UL << PG_slab, "slab" },
5272 {1UL << PG_owner_priv_1, "owner_priv_1" },
5273 {1UL << PG_arch_1, "arch_1" },
5274 {1UL << PG_reserved, "reserved" },
5275 {1UL << PG_private, "private" },
5276 {1UL << PG_private_2, "private_2" },
5277 {1UL << PG_writeback, "writeback" },
5278 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5279 {1UL << PG_head, "head" },
5280 {1UL << PG_tail, "tail" },
5282 {1UL << PG_compound, "compound" },
5284 {1UL << PG_swapcache, "swapcache" },
5285 {1UL << PG_mappedtodisk, "mappedtodisk" },
5286 {1UL << PG_reclaim, "reclaim" },
5287 {1UL << PG_buddy, "buddy" },
5288 {1UL << PG_swapbacked, "swapbacked" },
5289 {1UL << PG_unevictable, "unevictable" },
5291 {1UL << PG_mlocked, "mlocked" },
5293 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5294 {1UL << PG_uncached, "uncached" },
5296 #ifdef CONFIG_MEMORY_FAILURE
5297 {1UL << PG_hwpoison, "hwpoison" },
5302 static void dump_page_flags(unsigned long flags)
5304 const char *delim = "";
5308 printk(KERN_ALERT "page flags: %#lx(", flags);
5310 /* remove zone id */
5311 flags &= (1UL << NR_PAGEFLAGS) - 1;
5313 for (i = 0; pageflag_names[i].name && flags; i++) {
5315 mask = pageflag_names[i].mask;
5316 if ((flags & mask) != mask)
5320 printk("%s%s", delim, pageflag_names[i].name);
5324 /* check for left over flags */
5326 printk("%s%#lx", delim, flags);
5331 void dump_page(struct page *page)
5334 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5335 page, page_count(page), page_mapcount(page),
5336 page->mapping, page->index);
5337 dump_page_flags(page->flags);