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/config.h>
18 #include <linux/stddef.h>
20 #include <linux/swap.h>
21 #include <linux/interrupt.h>
22 #include <linux/pagemap.h>
23 #include <linux/bootmem.h>
24 #include <linux/compiler.h>
25 #include <linux/kernel.h>
26 #include <linux/module.h>
27 #include <linux/suspend.h>
28 #include <linux/pagevec.h>
29 #include <linux/blkdev.h>
30 #include <linux/slab.h>
31 #include <linux/notifier.h>
32 #include <linux/topology.h>
33 #include <linux/sysctl.h>
34 #include <linux/cpu.h>
35 #include <linux/cpuset.h>
36 #include <linux/memory_hotplug.h>
37 #include <linux/nodemask.h>
38 #include <linux/vmalloc.h>
39 #include <linux/mempolicy.h>
41 #include <asm/tlbflush.h>
45 * MCD - HACK: Find somewhere to initialize this EARLY, or make this
48 nodemask_t node_online_map __read_mostly = { { [0] = 1UL } };
49 EXPORT_SYMBOL(node_online_map);
50 nodemask_t node_possible_map __read_mostly = NODE_MASK_ALL;
51 EXPORT_SYMBOL(node_possible_map);
52 struct pglist_data *pgdat_list __read_mostly;
53 unsigned long totalram_pages __read_mostly;
54 unsigned long totalhigh_pages __read_mostly;
56 int percpu_pagelist_fraction;
58 static void __free_pages_ok(struct page *page, unsigned int order);
61 * results with 256, 32 in the lowmem_reserve sysctl:
62 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
63 * 1G machine -> (16M dma, 784M normal, 224M high)
64 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
65 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
66 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
68 * TBD: should special case ZONE_DMA32 machines here - in those we normally
69 * don't need any ZONE_NORMAL reservation
71 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 256, 256, 32 };
73 EXPORT_SYMBOL(totalram_pages);
76 * Used by page_zone() to look up the address of the struct zone whose
77 * id is encoded in the upper bits of page->flags
79 struct zone *zone_table[1 << ZONETABLE_SHIFT] __read_mostly;
80 EXPORT_SYMBOL(zone_table);
82 static char *zone_names[MAX_NR_ZONES] = { "DMA", "DMA32", "Normal", "HighMem" };
83 int min_free_kbytes = 1024;
85 unsigned long __initdata nr_kernel_pages;
86 unsigned long __initdata nr_all_pages;
88 #ifdef CONFIG_DEBUG_VM
89 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
93 unsigned long pfn = page_to_pfn(page);
96 seq = zone_span_seqbegin(zone);
97 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
99 else if (pfn < zone->zone_start_pfn)
101 } while (zone_span_seqretry(zone, seq));
106 static int page_is_consistent(struct zone *zone, struct page *page)
108 #ifdef CONFIG_HOLES_IN_ZONE
109 if (!pfn_valid(page_to_pfn(page)))
112 if (zone != page_zone(page))
118 * Temporary debugging check for pages not lying within a given zone.
120 static int bad_range(struct zone *zone, struct page *page)
122 if (page_outside_zone_boundaries(zone, page))
124 if (!page_is_consistent(zone, page))
131 static inline int bad_range(struct zone *zone, struct page *page)
137 static void bad_page(struct page *page)
139 printk(KERN_EMERG "Bad page state in process '%s'\n"
140 KERN_EMERG "page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d\n"
141 KERN_EMERG "Trying to fix it up, but a reboot is needed\n"
142 KERN_EMERG "Backtrace:\n",
143 current->comm, page, (int)(2*sizeof(unsigned long)),
144 (unsigned long)page->flags, page->mapping,
145 page_mapcount(page), page_count(page));
147 page->flags &= ~(1 << PG_lru |
156 set_page_count(page, 0);
157 reset_page_mapcount(page);
158 page->mapping = NULL;
159 add_taint(TAINT_BAD_PAGE);
163 * Higher-order pages are called "compound pages". They are structured thusly:
165 * The first PAGE_SIZE page is called the "head page".
167 * The remaining PAGE_SIZE pages are called "tail pages".
169 * All pages have PG_compound set. All pages have their ->private pointing at
170 * the head page (even the head page has this).
172 * The first tail page's ->lru.next holds the address of the compound page's
173 * put_page() function. Its ->lru.prev holds the order of allocation.
174 * This usage means that zero-order pages may not be compound.
177 static void free_compound_page(struct page *page)
179 __free_pages_ok(page, (unsigned long)page[1].lru.prev);
182 static void prep_compound_page(struct page *page, unsigned long order)
185 int nr_pages = 1 << order;
187 page[1].lru.next = (void *)free_compound_page; /* set dtor */
188 page[1].lru.prev = (void *)order;
189 for (i = 0; i < nr_pages; i++) {
190 struct page *p = page + i;
192 __SetPageCompound(p);
193 set_page_private(p, (unsigned long)page);
197 static void destroy_compound_page(struct page *page, unsigned long order)
200 int nr_pages = 1 << order;
202 if (unlikely((unsigned long)page[1].lru.prev != order))
205 for (i = 0; i < nr_pages; i++) {
206 struct page *p = page + i;
208 if (unlikely(!PageCompound(p) |
209 (page_private(p) != (unsigned long)page)))
211 __ClearPageCompound(p);
216 * function for dealing with page's order in buddy system.
217 * zone->lock is already acquired when we use these.
218 * So, we don't need atomic page->flags operations here.
220 static inline unsigned long page_order(struct page *page) {
221 return page_private(page);
224 static inline void set_page_order(struct page *page, int order) {
225 set_page_private(page, order);
226 __SetPagePrivate(page);
229 static inline void rmv_page_order(struct page *page)
231 __ClearPagePrivate(page);
232 set_page_private(page, 0);
236 * Locate the struct page for both the matching buddy in our
237 * pair (buddy1) and the combined O(n+1) page they form (page).
239 * 1) Any buddy B1 will have an order O twin B2 which satisfies
240 * the following equation:
242 * For example, if the starting buddy (buddy2) is #8 its order
244 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
246 * 2) Any buddy B will have an order O+1 parent P which
247 * satisfies the following equation:
250 * Assumption: *_mem_map is contigious at least up to MAX_ORDER
252 static inline struct page *
253 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
255 unsigned long buddy_idx = page_idx ^ (1 << order);
257 return page + (buddy_idx - page_idx);
260 static inline unsigned long
261 __find_combined_index(unsigned long page_idx, unsigned int order)
263 return (page_idx & ~(1 << order));
267 * This function checks whether a page is free && is the buddy
268 * we can do coalesce a page and its buddy if
269 * (a) the buddy is not in a hole &&
270 * (b) the buddy is free &&
271 * (c) the buddy is on the buddy system &&
272 * (d) a page and its buddy have the same order.
273 * for recording page's order, we use page_private(page) and PG_private.
276 static inline int page_is_buddy(struct page *page, int order)
278 #ifdef CONFIG_HOLES_IN_ZONE
279 if (!pfn_valid(page_to_pfn(page)))
283 if (PagePrivate(page) &&
284 (page_order(page) == order) &&
285 page_count(page) == 0)
291 * Freeing function for a buddy system allocator.
293 * The concept of a buddy system is to maintain direct-mapped table
294 * (containing bit values) for memory blocks of various "orders".
295 * The bottom level table contains the map for the smallest allocatable
296 * units of memory (here, pages), and each level above it describes
297 * pairs of units from the levels below, hence, "buddies".
298 * At a high level, all that happens here is marking the table entry
299 * at the bottom level available, and propagating the changes upward
300 * as necessary, plus some accounting needed to play nicely with other
301 * parts of the VM system.
302 * At each level, we keep a list of pages, which are heads of continuous
303 * free pages of length of (1 << order) and marked with PG_Private.Page's
304 * order is recorded in page_private(page) field.
305 * So when we are allocating or freeing one, we can derive the state of the
306 * other. That is, if we allocate a small block, and both were
307 * free, the remainder of the region must be split into blocks.
308 * If a block is freed, and its buddy is also free, then this
309 * triggers coalescing into a block of larger size.
314 static inline void __free_one_page(struct page *page,
315 struct zone *zone, unsigned int order)
317 unsigned long page_idx;
318 int order_size = 1 << order;
320 if (unlikely(PageCompound(page)))
321 destroy_compound_page(page, order);
323 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
325 BUG_ON(page_idx & (order_size - 1));
326 BUG_ON(bad_range(zone, page));
328 zone->free_pages += order_size;
329 while (order < MAX_ORDER-1) {
330 unsigned long combined_idx;
331 struct free_area *area;
334 buddy = __page_find_buddy(page, page_idx, order);
335 if (!page_is_buddy(buddy, order))
336 break; /* Move the buddy up one level. */
338 list_del(&buddy->lru);
339 area = zone->free_area + order;
341 rmv_page_order(buddy);
342 combined_idx = __find_combined_index(page_idx, order);
343 page = page + (combined_idx - page_idx);
344 page_idx = combined_idx;
347 set_page_order(page, order);
348 list_add(&page->lru, &zone->free_area[order].free_list);
349 zone->free_area[order].nr_free++;
352 static inline int free_pages_check(struct page *page)
354 if (unlikely(page_mapcount(page) |
355 (page->mapping != NULL) |
356 (page_count(page) != 0) |
366 1 << PG_reserved ))))
369 __ClearPageDirty(page);
371 * For now, we report if PG_reserved was found set, but do not
372 * clear it, and do not free the page. But we shall soon need
373 * to do more, for when the ZERO_PAGE count wraps negative.
375 return PageReserved(page);
379 * Frees a list of pages.
380 * Assumes all pages on list are in same zone, and of same order.
381 * count is the number of pages to free.
383 * If the zone was previously in an "all pages pinned" state then look to
384 * see if this freeing clears that state.
386 * And clear the zone's pages_scanned counter, to hold off the "all pages are
387 * pinned" detection logic.
389 static void free_pages_bulk(struct zone *zone, int count,
390 struct list_head *list, int order)
392 spin_lock(&zone->lock);
393 zone->all_unreclaimable = 0;
394 zone->pages_scanned = 0;
398 BUG_ON(list_empty(list));
399 page = list_entry(list->prev, struct page, lru);
400 /* have to delete it as __free_one_page list manipulates */
401 list_del(&page->lru);
402 __free_one_page(page, zone, order);
404 spin_unlock(&zone->lock);
407 static void free_one_page(struct zone *zone, struct page *page, int order)
410 list_add(&page->lru, &list);
411 free_pages_bulk(zone, 1, &list, order);
414 static void __free_pages_ok(struct page *page, unsigned int order)
420 arch_free_page(page, order);
421 if (!PageHighMem(page))
422 mutex_debug_check_no_locks_freed(page_address(page),
425 for (i = 0 ; i < (1 << order) ; ++i)
426 reserved += free_pages_check(page + i);
430 kernel_map_pages(page, 1 << order, 0);
431 local_irq_save(flags);
432 __mod_page_state(pgfree, 1 << order);
433 free_one_page(page_zone(page), page, order);
434 local_irq_restore(flags);
438 * permit the bootmem allocator to evade page validation on high-order frees
440 void fastcall __init __free_pages_bootmem(struct page *page, unsigned int order)
443 __ClearPageReserved(page);
444 set_page_count(page, 0);
445 set_page_refs(page, 0);
451 for (loop = 0; loop < BITS_PER_LONG; loop++) {
452 struct page *p = &page[loop];
454 if (loop + 1 < BITS_PER_LONG)
456 __ClearPageReserved(p);
457 set_page_count(p, 0);
460 set_page_refs(page, order);
461 __free_pages(page, order);
467 * The order of subdivision here is critical for the IO subsystem.
468 * Please do not alter this order without good reasons and regression
469 * testing. Specifically, as large blocks of memory are subdivided,
470 * the order in which smaller blocks are delivered depends on the order
471 * they're subdivided in this function. This is the primary factor
472 * influencing the order in which pages are delivered to the IO
473 * subsystem according to empirical testing, and this is also justified
474 * by considering the behavior of a buddy system containing a single
475 * large block of memory acted on by a series of small allocations.
476 * This behavior is a critical factor in sglist merging's success.
480 static inline void expand(struct zone *zone, struct page *page,
481 int low, int high, struct free_area *area)
483 unsigned long size = 1 << high;
489 BUG_ON(bad_range(zone, &page[size]));
490 list_add(&page[size].lru, &area->free_list);
492 set_page_order(&page[size], high);
497 * This page is about to be returned from the page allocator
499 static int prep_new_page(struct page *page, int order)
501 if (unlikely(page_mapcount(page) |
502 (page->mapping != NULL) |
503 (page_count(page) != 0) |
514 1 << PG_reserved ))))
518 * For now, we report if PG_reserved was found set, but do not
519 * clear it, and do not allocate the page: as a safety net.
521 if (PageReserved(page))
524 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
525 1 << PG_referenced | 1 << PG_arch_1 |
526 1 << PG_checked | 1 << PG_mappedtodisk);
527 set_page_private(page, 0);
528 set_page_refs(page, order);
529 kernel_map_pages(page, 1 << order, 1);
534 * Do the hard work of removing an element from the buddy allocator.
535 * Call me with the zone->lock already held.
537 static struct page *__rmqueue(struct zone *zone, unsigned int order)
539 struct free_area * area;
540 unsigned int current_order;
543 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
544 area = zone->free_area + current_order;
545 if (list_empty(&area->free_list))
548 page = list_entry(area->free_list.next, struct page, lru);
549 list_del(&page->lru);
550 rmv_page_order(page);
552 zone->free_pages -= 1UL << order;
553 expand(zone, page, order, current_order, area);
561 * Obtain a specified number of elements from the buddy allocator, all under
562 * a single hold of the lock, for efficiency. Add them to the supplied list.
563 * Returns the number of new pages which were placed at *list.
565 static int rmqueue_bulk(struct zone *zone, unsigned int order,
566 unsigned long count, struct list_head *list)
570 spin_lock(&zone->lock);
571 for (i = 0; i < count; ++i) {
572 struct page *page = __rmqueue(zone, order);
573 if (unlikely(page == NULL))
575 list_add_tail(&page->lru, list);
577 spin_unlock(&zone->lock);
583 * Called from the slab reaper to drain pagesets on a particular node that
584 * belong to the currently executing processor.
586 void drain_node_pages(int nodeid)
591 local_irq_save(flags);
592 for (z = 0; z < MAX_NR_ZONES; z++) {
593 struct zone *zone = NODE_DATA(nodeid)->node_zones + z;
594 struct per_cpu_pageset *pset;
596 pset = zone_pcp(zone, smp_processor_id());
597 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
598 struct per_cpu_pages *pcp;
601 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
605 local_irq_restore(flags);
609 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
610 static void __drain_pages(unsigned int cpu)
616 for_each_zone(zone) {
617 struct per_cpu_pageset *pset;
619 pset = zone_pcp(zone, cpu);
620 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
621 struct per_cpu_pages *pcp;
624 local_irq_save(flags);
625 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
627 local_irq_restore(flags);
631 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
635 void mark_free_pages(struct zone *zone)
637 unsigned long zone_pfn, flags;
639 struct list_head *curr;
641 if (!zone->spanned_pages)
644 spin_lock_irqsave(&zone->lock, flags);
645 for (zone_pfn = 0; zone_pfn < zone->spanned_pages; ++zone_pfn)
646 ClearPageNosaveFree(pfn_to_page(zone_pfn + zone->zone_start_pfn));
648 for (order = MAX_ORDER - 1; order >= 0; --order)
649 list_for_each(curr, &zone->free_area[order].free_list) {
650 unsigned long start_pfn, i;
652 start_pfn = page_to_pfn(list_entry(curr, struct page, lru));
654 for (i=0; i < (1<<order); i++)
655 SetPageNosaveFree(pfn_to_page(start_pfn+i));
657 spin_unlock_irqrestore(&zone->lock, flags);
661 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
663 void drain_local_pages(void)
667 local_irq_save(flags);
668 __drain_pages(smp_processor_id());
669 local_irq_restore(flags);
671 #endif /* CONFIG_PM */
673 static void zone_statistics(struct zonelist *zonelist, struct zone *z, int cpu)
676 pg_data_t *pg = z->zone_pgdat;
677 pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
678 struct per_cpu_pageset *p;
680 p = zone_pcp(z, cpu);
685 zone_pcp(zonelist->zones[0], cpu)->numa_foreign++;
687 if (pg == NODE_DATA(numa_node_id()))
695 * Free a 0-order page
697 static void fastcall free_hot_cold_page(struct page *page, int cold)
699 struct zone *zone = page_zone(page);
700 struct per_cpu_pages *pcp;
703 arch_free_page(page, 0);
706 page->mapping = NULL;
707 if (free_pages_check(page))
710 kernel_map_pages(page, 1, 0);
712 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
713 local_irq_save(flags);
714 __inc_page_state(pgfree);
715 list_add(&page->lru, &pcp->list);
717 if (pcp->count >= pcp->high) {
718 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
719 pcp->count -= pcp->batch;
721 local_irq_restore(flags);
725 void fastcall free_hot_page(struct page *page)
727 free_hot_cold_page(page, 0);
730 void fastcall free_cold_page(struct page *page)
732 free_hot_cold_page(page, 1);
735 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
739 BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
740 for(i = 0; i < (1 << order); i++)
741 clear_highpage(page + i);
745 * split_page takes a non-compound higher-order page, and splits it into
746 * n (1<<order) sub-pages: page[0..n]
747 * Each sub-page must be freed individually.
749 * Note: this is probably too low level an operation for use in drivers.
750 * Please consult with lkml before using this in your driver.
752 void split_page(struct page *page, unsigned int order)
756 BUG_ON(PageCompound(page));
757 BUG_ON(!page_count(page));
758 for (i = 1; i < (1 << order); i++) {
759 BUG_ON(page_count(page + i));
760 set_page_count(page + i, 1);
765 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
766 * we cheat by calling it from here, in the order > 0 path. Saves a branch
769 static struct page *buffered_rmqueue(struct zonelist *zonelist,
770 struct zone *zone, int order, gfp_t gfp_flags)
774 int cold = !!(gfp_flags & __GFP_COLD);
779 if (likely(order == 0)) {
780 struct per_cpu_pages *pcp;
782 pcp = &zone_pcp(zone, cpu)->pcp[cold];
783 local_irq_save(flags);
785 pcp->count += rmqueue_bulk(zone, 0,
786 pcp->batch, &pcp->list);
787 if (unlikely(!pcp->count))
790 page = list_entry(pcp->list.next, struct page, lru);
791 list_del(&page->lru);
794 spin_lock_irqsave(&zone->lock, flags);
795 page = __rmqueue(zone, order);
796 spin_unlock(&zone->lock);
801 __mod_page_state_zone(zone, pgalloc, 1 << order);
802 zone_statistics(zonelist, zone, cpu);
803 local_irq_restore(flags);
806 BUG_ON(bad_range(zone, page));
807 if (prep_new_page(page, order))
810 if (gfp_flags & __GFP_ZERO)
811 prep_zero_page(page, order, gfp_flags);
813 if (order && (gfp_flags & __GFP_COMP))
814 prep_compound_page(page, order);
818 local_irq_restore(flags);
823 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
824 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
825 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
826 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
827 #define ALLOC_HARDER 0x10 /* try to alloc harder */
828 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
829 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
832 * Return 1 if free pages are above 'mark'. This takes into account the order
835 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
836 int classzone_idx, int alloc_flags)
838 /* free_pages my go negative - that's OK */
839 long min = mark, free_pages = z->free_pages - (1 << order) + 1;
842 if (alloc_flags & ALLOC_HIGH)
844 if (alloc_flags & ALLOC_HARDER)
847 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
849 for (o = 0; o < order; o++) {
850 /* At the next order, this order's pages become unavailable */
851 free_pages -= z->free_area[o].nr_free << o;
853 /* Require fewer higher order pages to be free */
856 if (free_pages <= min)
863 * get_page_from_freeliest goes through the zonelist trying to allocate
867 get_page_from_freelist(gfp_t gfp_mask, unsigned int order,
868 struct zonelist *zonelist, int alloc_flags)
870 struct zone **z = zonelist->zones;
871 struct page *page = NULL;
872 int classzone_idx = zone_idx(*z);
875 * Go through the zonelist once, looking for a zone with enough free.
876 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
879 if ((alloc_flags & ALLOC_CPUSET) &&
880 !cpuset_zone_allowed(*z, gfp_mask))
883 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
885 if (alloc_flags & ALLOC_WMARK_MIN)
886 mark = (*z)->pages_min;
887 else if (alloc_flags & ALLOC_WMARK_LOW)
888 mark = (*z)->pages_low;
890 mark = (*z)->pages_high;
891 if (!zone_watermark_ok(*z, order, mark,
892 classzone_idx, alloc_flags))
893 if (!zone_reclaim_mode ||
894 !zone_reclaim(*z, gfp_mask, order))
898 page = buffered_rmqueue(zonelist, *z, order, gfp_mask);
902 } while (*(++z) != NULL);
907 * This is the 'heart' of the zoned buddy allocator.
909 struct page * fastcall
910 __alloc_pages(gfp_t gfp_mask, unsigned int order,
911 struct zonelist *zonelist)
913 const gfp_t wait = gfp_mask & __GFP_WAIT;
916 struct reclaim_state reclaim_state;
917 struct task_struct *p = current;
920 int did_some_progress;
922 might_sleep_if(wait);
925 z = zonelist->zones; /* the list of zones suitable for gfp_mask */
927 if (unlikely(*z == NULL)) {
928 /* Should this ever happen?? */
932 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
933 zonelist, ALLOC_WMARK_LOW|ALLOC_CPUSET);
938 wakeup_kswapd(*z, order);
942 * OK, we're below the kswapd watermark and have kicked background
943 * reclaim. Now things get more complex, so set up alloc_flags according
944 * to how we want to proceed.
946 * The caller may dip into page reserves a bit more if the caller
947 * cannot run direct reclaim, or if the caller has realtime scheduling
948 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
949 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
951 alloc_flags = ALLOC_WMARK_MIN;
952 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
953 alloc_flags |= ALLOC_HARDER;
954 if (gfp_mask & __GFP_HIGH)
955 alloc_flags |= ALLOC_HIGH;
956 alloc_flags |= ALLOC_CPUSET;
959 * Go through the zonelist again. Let __GFP_HIGH and allocations
960 * coming from realtime tasks go deeper into reserves.
962 * This is the last chance, in general, before the goto nopage.
963 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
964 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
966 page = get_page_from_freelist(gfp_mask, order, zonelist, alloc_flags);
970 /* This allocation should allow future memory freeing. */
972 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
973 && !in_interrupt()) {
974 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
976 /* go through the zonelist yet again, ignoring mins */
977 page = get_page_from_freelist(gfp_mask, order,
978 zonelist, ALLOC_NO_WATERMARKS);
981 if (gfp_mask & __GFP_NOFAIL) {
982 blk_congestion_wait(WRITE, HZ/50);
989 /* Atomic allocations - we can't balance anything */
996 /* We now go into synchronous reclaim */
997 cpuset_memory_pressure_bump();
998 p->flags |= PF_MEMALLOC;
999 reclaim_state.reclaimed_slab = 0;
1000 p->reclaim_state = &reclaim_state;
1002 did_some_progress = try_to_free_pages(zonelist->zones, gfp_mask);
1004 p->reclaim_state = NULL;
1005 p->flags &= ~PF_MEMALLOC;
1009 if (likely(did_some_progress)) {
1010 page = get_page_from_freelist(gfp_mask, order,
1011 zonelist, alloc_flags);
1014 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1016 * Go through the zonelist yet one more time, keep
1017 * very high watermark here, this is only to catch
1018 * a parallel oom killing, we must fail if we're still
1019 * under heavy pressure.
1021 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1022 zonelist, ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1026 out_of_memory(zonelist, gfp_mask, order);
1031 * Don't let big-order allocations loop unless the caller explicitly
1032 * requests that. Wait for some write requests to complete then retry.
1034 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
1035 * <= 3, but that may not be true in other implementations.
1038 if (!(gfp_mask & __GFP_NORETRY)) {
1039 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
1041 if (gfp_mask & __GFP_NOFAIL)
1045 blk_congestion_wait(WRITE, HZ/50);
1050 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1051 printk(KERN_WARNING "%s: page allocation failure."
1052 " order:%d, mode:0x%x\n",
1053 p->comm, order, gfp_mask);
1061 EXPORT_SYMBOL(__alloc_pages);
1064 * Common helper functions.
1066 fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1069 page = alloc_pages(gfp_mask, order);
1072 return (unsigned long) page_address(page);
1075 EXPORT_SYMBOL(__get_free_pages);
1077 fastcall unsigned long get_zeroed_page(gfp_t gfp_mask)
1082 * get_zeroed_page() returns a 32-bit address, which cannot represent
1085 BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1087 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1089 return (unsigned long) page_address(page);
1093 EXPORT_SYMBOL(get_zeroed_page);
1095 void __pagevec_free(struct pagevec *pvec)
1097 int i = pagevec_count(pvec);
1100 free_hot_cold_page(pvec->pages[i], pvec->cold);
1103 fastcall void __free_pages(struct page *page, unsigned int order)
1105 if (put_page_testzero(page)) {
1107 free_hot_page(page);
1109 __free_pages_ok(page, order);
1113 EXPORT_SYMBOL(__free_pages);
1115 fastcall void free_pages(unsigned long addr, unsigned int order)
1118 BUG_ON(!virt_addr_valid((void *)addr));
1119 __free_pages(virt_to_page((void *)addr), order);
1123 EXPORT_SYMBOL(free_pages);
1126 * Total amount of free (allocatable) RAM:
1128 unsigned int nr_free_pages(void)
1130 unsigned int sum = 0;
1134 sum += zone->free_pages;
1139 EXPORT_SYMBOL(nr_free_pages);
1142 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
1144 unsigned int i, sum = 0;
1146 for (i = 0; i < MAX_NR_ZONES; i++)
1147 sum += pgdat->node_zones[i].free_pages;
1153 static unsigned int nr_free_zone_pages(int offset)
1155 /* Just pick one node, since fallback list is circular */
1156 pg_data_t *pgdat = NODE_DATA(numa_node_id());
1157 unsigned int sum = 0;
1159 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1160 struct zone **zonep = zonelist->zones;
1163 for (zone = *zonep++; zone; zone = *zonep++) {
1164 unsigned long size = zone->present_pages;
1165 unsigned long high = zone->pages_high;
1174 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1176 unsigned int nr_free_buffer_pages(void)
1178 return nr_free_zone_pages(gfp_zone(GFP_USER));
1182 * Amount of free RAM allocatable within all zones
1184 unsigned int nr_free_pagecache_pages(void)
1186 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER));
1189 #ifdef CONFIG_HIGHMEM
1190 unsigned int nr_free_highpages (void)
1193 unsigned int pages = 0;
1195 for_each_pgdat(pgdat)
1196 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1203 static void show_node(struct zone *zone)
1205 printk("Node %d ", zone->zone_pgdat->node_id);
1208 #define show_node(zone) do { } while (0)
1212 * Accumulate the page_state information across all CPUs.
1213 * The result is unavoidably approximate - it can change
1214 * during and after execution of this function.
1216 static DEFINE_PER_CPU(struct page_state, page_states) = {0};
1218 atomic_t nr_pagecache = ATOMIC_INIT(0);
1219 EXPORT_SYMBOL(nr_pagecache);
1221 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
1224 static void __get_page_state(struct page_state *ret, int nr, cpumask_t *cpumask)
1228 memset(ret, 0, nr * sizeof(unsigned long));
1229 cpus_and(*cpumask, *cpumask, cpu_online_map);
1231 for_each_cpu_mask(cpu, *cpumask) {
1237 in = (unsigned long *)&per_cpu(page_states, cpu);
1239 next_cpu = next_cpu(cpu, *cpumask);
1240 if (likely(next_cpu < NR_CPUS))
1241 prefetch(&per_cpu(page_states, next_cpu));
1243 out = (unsigned long *)ret;
1244 for (off = 0; off < nr; off++)
1249 void get_page_state_node(struct page_state *ret, int node)
1252 cpumask_t mask = node_to_cpumask(node);
1254 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
1255 nr /= sizeof(unsigned long);
1257 __get_page_state(ret, nr+1, &mask);
1260 void get_page_state(struct page_state *ret)
1263 cpumask_t mask = CPU_MASK_ALL;
1265 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
1266 nr /= sizeof(unsigned long);
1268 __get_page_state(ret, nr + 1, &mask);
1271 void get_full_page_state(struct page_state *ret)
1273 cpumask_t mask = CPU_MASK_ALL;
1275 __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long), &mask);
1278 unsigned long read_page_state_offset(unsigned long offset)
1280 unsigned long ret = 0;
1283 for_each_online_cpu(cpu) {
1286 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
1287 ret += *((unsigned long *)in);
1292 void __mod_page_state_offset(unsigned long offset, unsigned long delta)
1296 ptr = &__get_cpu_var(page_states);
1297 *(unsigned long *)(ptr + offset) += delta;
1299 EXPORT_SYMBOL(__mod_page_state_offset);
1301 void mod_page_state_offset(unsigned long offset, unsigned long delta)
1303 unsigned long flags;
1306 local_irq_save(flags);
1307 ptr = &__get_cpu_var(page_states);
1308 *(unsigned long *)(ptr + offset) += delta;
1309 local_irq_restore(flags);
1311 EXPORT_SYMBOL(mod_page_state_offset);
1313 void __get_zone_counts(unsigned long *active, unsigned long *inactive,
1314 unsigned long *free, struct pglist_data *pgdat)
1316 struct zone *zones = pgdat->node_zones;
1322 for (i = 0; i < MAX_NR_ZONES; i++) {
1323 *active += zones[i].nr_active;
1324 *inactive += zones[i].nr_inactive;
1325 *free += zones[i].free_pages;
1329 void get_zone_counts(unsigned long *active,
1330 unsigned long *inactive, unsigned long *free)
1332 struct pglist_data *pgdat;
1337 for_each_pgdat(pgdat) {
1338 unsigned long l, m, n;
1339 __get_zone_counts(&l, &m, &n, pgdat);
1346 void si_meminfo(struct sysinfo *val)
1348 val->totalram = totalram_pages;
1350 val->freeram = nr_free_pages();
1351 val->bufferram = nr_blockdev_pages();
1352 #ifdef CONFIG_HIGHMEM
1353 val->totalhigh = totalhigh_pages;
1354 val->freehigh = nr_free_highpages();
1359 val->mem_unit = PAGE_SIZE;
1362 EXPORT_SYMBOL(si_meminfo);
1365 void si_meminfo_node(struct sysinfo *val, int nid)
1367 pg_data_t *pgdat = NODE_DATA(nid);
1369 val->totalram = pgdat->node_present_pages;
1370 val->freeram = nr_free_pages_pgdat(pgdat);
1371 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1372 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1373 val->mem_unit = PAGE_SIZE;
1377 #define K(x) ((x) << (PAGE_SHIFT-10))
1380 * Show free area list (used inside shift_scroll-lock stuff)
1381 * We also calculate the percentage fragmentation. We do this by counting the
1382 * memory on each free list with the exception of the first item on the list.
1384 void show_free_areas(void)
1386 struct page_state ps;
1387 int cpu, temperature;
1388 unsigned long active;
1389 unsigned long inactive;
1393 for_each_zone(zone) {
1395 printk("%s per-cpu:", zone->name);
1397 if (!populated_zone(zone)) {
1403 for_each_online_cpu(cpu) {
1404 struct per_cpu_pageset *pageset;
1406 pageset = zone_pcp(zone, cpu);
1408 for (temperature = 0; temperature < 2; temperature++)
1409 printk("cpu %d %s: high %d, batch %d used:%d\n",
1411 temperature ? "cold" : "hot",
1412 pageset->pcp[temperature].high,
1413 pageset->pcp[temperature].batch,
1414 pageset->pcp[temperature].count);
1418 get_page_state(&ps);
1419 get_zone_counts(&active, &inactive, &free);
1421 printk("Free pages: %11ukB (%ukB HighMem)\n",
1423 K(nr_free_highpages()));
1425 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1426 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1435 ps.nr_page_table_pages);
1437 for_each_zone(zone) {
1449 " pages_scanned:%lu"
1450 " all_unreclaimable? %s"
1453 K(zone->free_pages),
1456 K(zone->pages_high),
1458 K(zone->nr_inactive),
1459 K(zone->present_pages),
1460 zone->pages_scanned,
1461 (zone->all_unreclaimable ? "yes" : "no")
1463 printk("lowmem_reserve[]:");
1464 for (i = 0; i < MAX_NR_ZONES; i++)
1465 printk(" %lu", zone->lowmem_reserve[i]);
1469 for_each_zone(zone) {
1470 unsigned long nr, flags, order, total = 0;
1473 printk("%s: ", zone->name);
1474 if (!populated_zone(zone)) {
1479 spin_lock_irqsave(&zone->lock, flags);
1480 for (order = 0; order < MAX_ORDER; order++) {
1481 nr = zone->free_area[order].nr_free;
1482 total += nr << order;
1483 printk("%lu*%lukB ", nr, K(1UL) << order);
1485 spin_unlock_irqrestore(&zone->lock, flags);
1486 printk("= %lukB\n", K(total));
1489 show_swap_cache_info();
1493 * Builds allocation fallback zone lists.
1495 * Add all populated zones of a node to the zonelist.
1497 static int __init build_zonelists_node(pg_data_t *pgdat,
1498 struct zonelist *zonelist, int nr_zones, int zone_type)
1502 BUG_ON(zone_type > ZONE_HIGHMEM);
1505 zone = pgdat->node_zones + zone_type;
1506 if (populated_zone(zone)) {
1507 #ifndef CONFIG_HIGHMEM
1508 BUG_ON(zone_type > ZONE_NORMAL);
1510 zonelist->zones[nr_zones++] = zone;
1511 check_highest_zone(zone_type);
1515 } while (zone_type >= 0);
1519 static inline int highest_zone(int zone_bits)
1521 int res = ZONE_NORMAL;
1522 if (zone_bits & (__force int)__GFP_HIGHMEM)
1524 if (zone_bits & (__force int)__GFP_DMA32)
1526 if (zone_bits & (__force int)__GFP_DMA)
1532 #define MAX_NODE_LOAD (num_online_nodes())
1533 static int __initdata node_load[MAX_NUMNODES];
1535 * find_next_best_node - find the next node that should appear in a given node's fallback list
1536 * @node: node whose fallback list we're appending
1537 * @used_node_mask: nodemask_t of already used nodes
1539 * We use a number of factors to determine which is the next node that should
1540 * appear on a given node's fallback list. The node should not have appeared
1541 * already in @node's fallback list, and it should be the next closest node
1542 * according to the distance array (which contains arbitrary distance values
1543 * from each node to each node in the system), and should also prefer nodes
1544 * with no CPUs, since presumably they'll have very little allocation pressure
1545 * on them otherwise.
1546 * It returns -1 if no node is found.
1548 static int __init find_next_best_node(int node, nodemask_t *used_node_mask)
1551 int min_val = INT_MAX;
1554 /* Use the local node if we haven't already */
1555 if (!node_isset(node, *used_node_mask)) {
1556 node_set(node, *used_node_mask);
1560 for_each_online_node(n) {
1563 /* Don't want a node to appear more than once */
1564 if (node_isset(n, *used_node_mask))
1567 /* Use the distance array to find the distance */
1568 val = node_distance(node, n);
1570 /* Penalize nodes under us ("prefer the next node") */
1573 /* Give preference to headless and unused nodes */
1574 tmp = node_to_cpumask(n);
1575 if (!cpus_empty(tmp))
1576 val += PENALTY_FOR_NODE_WITH_CPUS;
1578 /* Slight preference for less loaded node */
1579 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1580 val += node_load[n];
1582 if (val < min_val) {
1589 node_set(best_node, *used_node_mask);
1594 static void __init build_zonelists(pg_data_t *pgdat)
1596 int i, j, k, node, local_node;
1597 int prev_node, load;
1598 struct zonelist *zonelist;
1599 nodemask_t used_mask;
1601 /* initialize zonelists */
1602 for (i = 0; i < GFP_ZONETYPES; i++) {
1603 zonelist = pgdat->node_zonelists + i;
1604 zonelist->zones[0] = NULL;
1607 /* NUMA-aware ordering of nodes */
1608 local_node = pgdat->node_id;
1609 load = num_online_nodes();
1610 prev_node = local_node;
1611 nodes_clear(used_mask);
1612 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1613 int distance = node_distance(local_node, node);
1616 * If another node is sufficiently far away then it is better
1617 * to reclaim pages in a zone before going off node.
1619 if (distance > RECLAIM_DISTANCE)
1620 zone_reclaim_mode = 1;
1623 * We don't want to pressure a particular node.
1624 * So adding penalty to the first node in same
1625 * distance group to make it round-robin.
1628 if (distance != node_distance(local_node, prev_node))
1629 node_load[node] += load;
1632 for (i = 0; i < GFP_ZONETYPES; i++) {
1633 zonelist = pgdat->node_zonelists + i;
1634 for (j = 0; zonelist->zones[j] != NULL; j++);
1636 k = highest_zone(i);
1638 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1639 zonelist->zones[j] = NULL;
1644 #else /* CONFIG_NUMA */
1646 static void __init build_zonelists(pg_data_t *pgdat)
1648 int i, j, k, node, local_node;
1650 local_node = pgdat->node_id;
1651 for (i = 0; i < GFP_ZONETYPES; i++) {
1652 struct zonelist *zonelist;
1654 zonelist = pgdat->node_zonelists + i;
1657 k = highest_zone(i);
1658 j = build_zonelists_node(pgdat, zonelist, j, k);
1660 * Now we build the zonelist so that it contains the zones
1661 * of all the other nodes.
1662 * We don't want to pressure a particular node, so when
1663 * building the zones for node N, we make sure that the
1664 * zones coming right after the local ones are those from
1665 * node N+1 (modulo N)
1667 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1668 if (!node_online(node))
1670 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1672 for (node = 0; node < local_node; node++) {
1673 if (!node_online(node))
1675 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1678 zonelist->zones[j] = NULL;
1682 #endif /* CONFIG_NUMA */
1684 void __init build_all_zonelists(void)
1688 for_each_online_node(i)
1689 build_zonelists(NODE_DATA(i));
1690 printk("Built %i zonelists\n", num_online_nodes());
1691 cpuset_init_current_mems_allowed();
1695 * Helper functions to size the waitqueue hash table.
1696 * Essentially these want to choose hash table sizes sufficiently
1697 * large so that collisions trying to wait on pages are rare.
1698 * But in fact, the number of active page waitqueues on typical
1699 * systems is ridiculously low, less than 200. So this is even
1700 * conservative, even though it seems large.
1702 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1703 * waitqueues, i.e. the size of the waitq table given the number of pages.
1705 #define PAGES_PER_WAITQUEUE 256
1707 static inline unsigned long wait_table_size(unsigned long pages)
1709 unsigned long size = 1;
1711 pages /= PAGES_PER_WAITQUEUE;
1713 while (size < pages)
1717 * Once we have dozens or even hundreds of threads sleeping
1718 * on IO we've got bigger problems than wait queue collision.
1719 * Limit the size of the wait table to a reasonable size.
1721 size = min(size, 4096UL);
1723 return max(size, 4UL);
1727 * This is an integer logarithm so that shifts can be used later
1728 * to extract the more random high bits from the multiplicative
1729 * hash function before the remainder is taken.
1731 static inline unsigned long wait_table_bits(unsigned long size)
1736 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1738 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1739 unsigned long *zones_size, unsigned long *zholes_size)
1741 unsigned long realtotalpages, totalpages = 0;
1744 for (i = 0; i < MAX_NR_ZONES; i++)
1745 totalpages += zones_size[i];
1746 pgdat->node_spanned_pages = totalpages;
1748 realtotalpages = totalpages;
1750 for (i = 0; i < MAX_NR_ZONES; i++)
1751 realtotalpages -= zholes_size[i];
1752 pgdat->node_present_pages = realtotalpages;
1753 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1758 * Initially all pages are reserved - free ones are freed
1759 * up by free_all_bootmem() once the early boot process is
1760 * done. Non-atomic initialization, single-pass.
1762 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1763 unsigned long start_pfn)
1766 unsigned long end_pfn = start_pfn + size;
1769 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1770 if (!early_pfn_valid(pfn))
1772 page = pfn_to_page(pfn);
1773 set_page_links(page, zone, nid, pfn);
1774 set_page_count(page, 1);
1775 reset_page_mapcount(page);
1776 SetPageReserved(page);
1777 INIT_LIST_HEAD(&page->lru);
1778 #ifdef WANT_PAGE_VIRTUAL
1779 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1780 if (!is_highmem_idx(zone))
1781 set_page_address(page, __va(pfn << PAGE_SHIFT));
1786 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1790 for (order = 0; order < MAX_ORDER ; order++) {
1791 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1792 zone->free_area[order].nr_free = 0;
1796 #define ZONETABLE_INDEX(x, zone_nr) ((x << ZONES_SHIFT) | zone_nr)
1797 void zonetable_add(struct zone *zone, int nid, int zid, unsigned long pfn,
1800 unsigned long snum = pfn_to_section_nr(pfn);
1801 unsigned long end = pfn_to_section_nr(pfn + size);
1804 zone_table[ZONETABLE_INDEX(nid, zid)] = zone;
1806 for (; snum <= end; snum++)
1807 zone_table[ZONETABLE_INDEX(snum, zid)] = zone;
1810 #ifndef __HAVE_ARCH_MEMMAP_INIT
1811 #define memmap_init(size, nid, zone, start_pfn) \
1812 memmap_init_zone((size), (nid), (zone), (start_pfn))
1815 static int __cpuinit zone_batchsize(struct zone *zone)
1820 * The per-cpu-pages pools are set to around 1000th of the
1821 * size of the zone. But no more than 1/2 of a meg.
1823 * OK, so we don't know how big the cache is. So guess.
1825 batch = zone->present_pages / 1024;
1826 if (batch * PAGE_SIZE > 512 * 1024)
1827 batch = (512 * 1024) / PAGE_SIZE;
1828 batch /= 4; /* We effectively *= 4 below */
1833 * Clamp the batch to a 2^n - 1 value. Having a power
1834 * of 2 value was found to be more likely to have
1835 * suboptimal cache aliasing properties in some cases.
1837 * For example if 2 tasks are alternately allocating
1838 * batches of pages, one task can end up with a lot
1839 * of pages of one half of the possible page colors
1840 * and the other with pages of the other colors.
1842 batch = (1 << (fls(batch + batch/2)-1)) - 1;
1847 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
1849 struct per_cpu_pages *pcp;
1851 memset(p, 0, sizeof(*p));
1853 pcp = &p->pcp[0]; /* hot */
1855 pcp->high = 6 * batch;
1856 pcp->batch = max(1UL, 1 * batch);
1857 INIT_LIST_HEAD(&pcp->list);
1859 pcp = &p->pcp[1]; /* cold*/
1861 pcp->high = 2 * batch;
1862 pcp->batch = max(1UL, batch/2);
1863 INIT_LIST_HEAD(&pcp->list);
1867 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
1868 * to the value high for the pageset p.
1871 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
1874 struct per_cpu_pages *pcp;
1876 pcp = &p->pcp[0]; /* hot list */
1878 pcp->batch = max(1UL, high/4);
1879 if ((high/4) > (PAGE_SHIFT * 8))
1880 pcp->batch = PAGE_SHIFT * 8;
1886 * Boot pageset table. One per cpu which is going to be used for all
1887 * zones and all nodes. The parameters will be set in such a way
1888 * that an item put on a list will immediately be handed over to
1889 * the buddy list. This is safe since pageset manipulation is done
1890 * with interrupts disabled.
1892 * Some NUMA counter updates may also be caught by the boot pagesets.
1894 * The boot_pagesets must be kept even after bootup is complete for
1895 * unused processors and/or zones. They do play a role for bootstrapping
1896 * hotplugged processors.
1898 * zoneinfo_show() and maybe other functions do
1899 * not check if the processor is online before following the pageset pointer.
1900 * Other parts of the kernel may not check if the zone is available.
1902 static struct per_cpu_pageset boot_pageset[NR_CPUS];
1905 * Dynamically allocate memory for the
1906 * per cpu pageset array in struct zone.
1908 static int __cpuinit process_zones(int cpu)
1910 struct zone *zone, *dzone;
1912 for_each_zone(zone) {
1914 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
1915 GFP_KERNEL, cpu_to_node(cpu));
1916 if (!zone_pcp(zone, cpu))
1919 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
1921 if (percpu_pagelist_fraction)
1922 setup_pagelist_highmark(zone_pcp(zone, cpu),
1923 (zone->present_pages / percpu_pagelist_fraction));
1928 for_each_zone(dzone) {
1931 kfree(zone_pcp(dzone, cpu));
1932 zone_pcp(dzone, cpu) = NULL;
1937 static inline void free_zone_pagesets(int cpu)
1941 for_each_zone(zone) {
1942 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
1944 zone_pcp(zone, cpu) = NULL;
1949 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
1950 unsigned long action,
1953 int cpu = (long)hcpu;
1954 int ret = NOTIFY_OK;
1957 case CPU_UP_PREPARE:
1958 if (process_zones(cpu))
1961 case CPU_UP_CANCELED:
1963 free_zone_pagesets(cpu);
1971 static struct notifier_block pageset_notifier =
1972 { &pageset_cpuup_callback, NULL, 0 };
1974 void __init setup_per_cpu_pageset(void)
1978 /* Initialize per_cpu_pageset for cpu 0.
1979 * A cpuup callback will do this for every cpu
1980 * as it comes online
1982 err = process_zones(smp_processor_id());
1984 register_cpu_notifier(&pageset_notifier);
1990 void zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
1993 struct pglist_data *pgdat = zone->zone_pgdat;
1996 * The per-page waitqueue mechanism uses hashed waitqueues
1999 zone->wait_table_size = wait_table_size(zone_size_pages);
2000 zone->wait_table_bits = wait_table_bits(zone->wait_table_size);
2001 zone->wait_table = (wait_queue_head_t *)
2002 alloc_bootmem_node(pgdat, zone->wait_table_size
2003 * sizeof(wait_queue_head_t));
2005 for(i = 0; i < zone->wait_table_size; ++i)
2006 init_waitqueue_head(zone->wait_table + i);
2009 static __meminit void zone_pcp_init(struct zone *zone)
2012 unsigned long batch = zone_batchsize(zone);
2014 for (cpu = 0; cpu < NR_CPUS; cpu++) {
2016 /* Early boot. Slab allocator not functional yet */
2017 zone_pcp(zone, cpu) = &boot_pageset[cpu];
2018 setup_pageset(&boot_pageset[cpu],0);
2020 setup_pageset(zone_pcp(zone,cpu), batch);
2023 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
2024 zone->name, zone->present_pages, batch);
2027 static __meminit void init_currently_empty_zone(struct zone *zone,
2028 unsigned long zone_start_pfn, unsigned long size)
2030 struct pglist_data *pgdat = zone->zone_pgdat;
2032 zone_wait_table_init(zone, size);
2033 pgdat->nr_zones = zone_idx(zone) + 1;
2035 zone->zone_mem_map = pfn_to_page(zone_start_pfn);
2036 zone->zone_start_pfn = zone_start_pfn;
2038 memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn);
2040 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
2044 * Set up the zone data structures:
2045 * - mark all pages reserved
2046 * - mark all memory queues empty
2047 * - clear the memory bitmaps
2049 static void __init free_area_init_core(struct pglist_data *pgdat,
2050 unsigned long *zones_size, unsigned long *zholes_size)
2053 int nid = pgdat->node_id;
2054 unsigned long zone_start_pfn = pgdat->node_start_pfn;
2056 pgdat_resize_init(pgdat);
2057 pgdat->nr_zones = 0;
2058 init_waitqueue_head(&pgdat->kswapd_wait);
2059 pgdat->kswapd_max_order = 0;
2061 for (j = 0; j < MAX_NR_ZONES; j++) {
2062 struct zone *zone = pgdat->node_zones + j;
2063 unsigned long size, realsize;
2065 realsize = size = zones_size[j];
2067 realsize -= zholes_size[j];
2069 if (j < ZONE_HIGHMEM)
2070 nr_kernel_pages += realsize;
2071 nr_all_pages += realsize;
2073 zone->spanned_pages = size;
2074 zone->present_pages = realsize;
2075 zone->name = zone_names[j];
2076 spin_lock_init(&zone->lock);
2077 spin_lock_init(&zone->lru_lock);
2078 zone_seqlock_init(zone);
2079 zone->zone_pgdat = pgdat;
2080 zone->free_pages = 0;
2082 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
2084 zone_pcp_init(zone);
2085 INIT_LIST_HEAD(&zone->active_list);
2086 INIT_LIST_HEAD(&zone->inactive_list);
2087 zone->nr_scan_active = 0;
2088 zone->nr_scan_inactive = 0;
2089 zone->nr_active = 0;
2090 zone->nr_inactive = 0;
2091 atomic_set(&zone->reclaim_in_progress, 0);
2095 zonetable_add(zone, nid, j, zone_start_pfn, size);
2096 init_currently_empty_zone(zone, zone_start_pfn, size);
2097 zone_start_pfn += size;
2101 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
2103 /* Skip empty nodes */
2104 if (!pgdat->node_spanned_pages)
2107 #ifdef CONFIG_FLAT_NODE_MEM_MAP
2108 /* ia64 gets its own node_mem_map, before this, without bootmem */
2109 if (!pgdat->node_mem_map) {
2113 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
2114 map = alloc_remap(pgdat->node_id, size);
2116 map = alloc_bootmem_node(pgdat, size);
2117 pgdat->node_mem_map = map;
2119 #ifdef CONFIG_FLATMEM
2121 * With no DISCONTIG, the global mem_map is just set as node 0's
2123 if (pgdat == NODE_DATA(0))
2124 mem_map = NODE_DATA(0)->node_mem_map;
2126 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
2129 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
2130 unsigned long *zones_size, unsigned long node_start_pfn,
2131 unsigned long *zholes_size)
2133 pgdat->node_id = nid;
2134 pgdat->node_start_pfn = node_start_pfn;
2135 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
2137 alloc_node_mem_map(pgdat);
2139 free_area_init_core(pgdat, zones_size, zholes_size);
2142 #ifndef CONFIG_NEED_MULTIPLE_NODES
2143 static bootmem_data_t contig_bootmem_data;
2144 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
2146 EXPORT_SYMBOL(contig_page_data);
2149 void __init free_area_init(unsigned long *zones_size)
2151 free_area_init_node(0, NODE_DATA(0), zones_size,
2152 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
2155 #ifdef CONFIG_PROC_FS
2157 #include <linux/seq_file.h>
2159 static void *frag_start(struct seq_file *m, loff_t *pos)
2164 for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
2170 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
2172 pg_data_t *pgdat = (pg_data_t *)arg;
2175 return pgdat->pgdat_next;
2178 static void frag_stop(struct seq_file *m, void *arg)
2183 * This walks the free areas for each zone.
2185 static int frag_show(struct seq_file *m, void *arg)
2187 pg_data_t *pgdat = (pg_data_t *)arg;
2189 struct zone *node_zones = pgdat->node_zones;
2190 unsigned long flags;
2193 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2194 if (!populated_zone(zone))
2197 spin_lock_irqsave(&zone->lock, flags);
2198 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
2199 for (order = 0; order < MAX_ORDER; ++order)
2200 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
2201 spin_unlock_irqrestore(&zone->lock, flags);
2207 struct seq_operations fragmentation_op = {
2208 .start = frag_start,
2215 * Output information about zones in @pgdat.
2217 static int zoneinfo_show(struct seq_file *m, void *arg)
2219 pg_data_t *pgdat = arg;
2221 struct zone *node_zones = pgdat->node_zones;
2222 unsigned long flags;
2224 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; zone++) {
2227 if (!populated_zone(zone))
2230 spin_lock_irqsave(&zone->lock, flags);
2231 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
2239 "\n scanned %lu (a: %lu i: %lu)"
2248 zone->pages_scanned,
2249 zone->nr_scan_active, zone->nr_scan_inactive,
2250 zone->spanned_pages,
2251 zone->present_pages);
2253 "\n protection: (%lu",
2254 zone->lowmem_reserve[0]);
2255 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
2256 seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
2260 for_each_online_cpu(i) {
2261 struct per_cpu_pageset *pageset;
2264 pageset = zone_pcp(zone, i);
2265 for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2266 if (pageset->pcp[j].count)
2269 if (j == ARRAY_SIZE(pageset->pcp))
2271 for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2273 "\n cpu: %i pcp: %i"
2278 pageset->pcp[j].count,
2279 pageset->pcp[j].high,
2280 pageset->pcp[j].batch);
2286 "\n numa_foreign: %lu"
2287 "\n interleave_hit: %lu"
2288 "\n local_node: %lu"
2289 "\n other_node: %lu",
2292 pageset->numa_foreign,
2293 pageset->interleave_hit,
2294 pageset->local_node,
2295 pageset->other_node);
2299 "\n all_unreclaimable: %u"
2300 "\n prev_priority: %i"
2301 "\n temp_priority: %i"
2302 "\n start_pfn: %lu",
2303 zone->all_unreclaimable,
2304 zone->prev_priority,
2305 zone->temp_priority,
2306 zone->zone_start_pfn);
2307 spin_unlock_irqrestore(&zone->lock, flags);
2313 struct seq_operations zoneinfo_op = {
2314 .start = frag_start, /* iterate over all zones. The same as in
2318 .show = zoneinfo_show,
2321 static char *vmstat_text[] = {
2325 "nr_page_table_pages",
2356 "pgscan_kswapd_high",
2357 "pgscan_kswapd_normal",
2358 "pgscan_kswapd_dma32",
2359 "pgscan_kswapd_dma",
2361 "pgscan_direct_high",
2362 "pgscan_direct_normal",
2363 "pgscan_direct_dma32",
2364 "pgscan_direct_dma",
2369 "kswapd_inodesteal",
2377 static void *vmstat_start(struct seq_file *m, loff_t *pos)
2379 struct page_state *ps;
2381 if (*pos >= ARRAY_SIZE(vmstat_text))
2384 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
2387 return ERR_PTR(-ENOMEM);
2388 get_full_page_state(ps);
2389 ps->pgpgin /= 2; /* sectors -> kbytes */
2391 return (unsigned long *)ps + *pos;
2394 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
2397 if (*pos >= ARRAY_SIZE(vmstat_text))
2399 return (unsigned long *)m->private + *pos;
2402 static int vmstat_show(struct seq_file *m, void *arg)
2404 unsigned long *l = arg;
2405 unsigned long off = l - (unsigned long *)m->private;
2407 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
2411 static void vmstat_stop(struct seq_file *m, void *arg)
2417 struct seq_operations vmstat_op = {
2418 .start = vmstat_start,
2419 .next = vmstat_next,
2420 .stop = vmstat_stop,
2421 .show = vmstat_show,
2424 #endif /* CONFIG_PROC_FS */
2426 #ifdef CONFIG_HOTPLUG_CPU
2427 static int page_alloc_cpu_notify(struct notifier_block *self,
2428 unsigned long action, void *hcpu)
2430 int cpu = (unsigned long)hcpu;
2432 unsigned long *src, *dest;
2434 if (action == CPU_DEAD) {
2437 /* Drain local pagecache count. */
2438 count = &per_cpu(nr_pagecache_local, cpu);
2439 atomic_add(*count, &nr_pagecache);
2441 local_irq_disable();
2444 /* Add dead cpu's page_states to our own. */
2445 dest = (unsigned long *)&__get_cpu_var(page_states);
2446 src = (unsigned long *)&per_cpu(page_states, cpu);
2448 for (i = 0; i < sizeof(struct page_state)/sizeof(unsigned long);
2458 #endif /* CONFIG_HOTPLUG_CPU */
2460 void __init page_alloc_init(void)
2462 hotcpu_notifier(page_alloc_cpu_notify, 0);
2466 * setup_per_zone_lowmem_reserve - called whenever
2467 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
2468 * has a correct pages reserved value, so an adequate number of
2469 * pages are left in the zone after a successful __alloc_pages().
2471 static void setup_per_zone_lowmem_reserve(void)
2473 struct pglist_data *pgdat;
2476 for_each_pgdat(pgdat) {
2477 for (j = 0; j < MAX_NR_ZONES; j++) {
2478 struct zone *zone = pgdat->node_zones + j;
2479 unsigned long present_pages = zone->present_pages;
2481 zone->lowmem_reserve[j] = 0;
2483 for (idx = j-1; idx >= 0; idx--) {
2484 struct zone *lower_zone;
2486 if (sysctl_lowmem_reserve_ratio[idx] < 1)
2487 sysctl_lowmem_reserve_ratio[idx] = 1;
2489 lower_zone = pgdat->node_zones + idx;
2490 lower_zone->lowmem_reserve[j] = present_pages /
2491 sysctl_lowmem_reserve_ratio[idx];
2492 present_pages += lower_zone->present_pages;
2499 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
2500 * that the pages_{min,low,high} values for each zone are set correctly
2501 * with respect to min_free_kbytes.
2503 void setup_per_zone_pages_min(void)
2505 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
2506 unsigned long lowmem_pages = 0;
2508 unsigned long flags;
2510 /* Calculate total number of !ZONE_HIGHMEM pages */
2511 for_each_zone(zone) {
2512 if (!is_highmem(zone))
2513 lowmem_pages += zone->present_pages;
2516 for_each_zone(zone) {
2518 spin_lock_irqsave(&zone->lru_lock, flags);
2519 tmp = (pages_min * zone->present_pages) / lowmem_pages;
2520 if (is_highmem(zone)) {
2522 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
2523 * need highmem pages, so cap pages_min to a small
2526 * The (pages_high-pages_low) and (pages_low-pages_min)
2527 * deltas controls asynch page reclaim, and so should
2528 * not be capped for highmem.
2532 min_pages = zone->present_pages / 1024;
2533 if (min_pages < SWAP_CLUSTER_MAX)
2534 min_pages = SWAP_CLUSTER_MAX;
2535 if (min_pages > 128)
2537 zone->pages_min = min_pages;
2540 * If it's a lowmem zone, reserve a number of pages
2541 * proportionate to the zone's size.
2543 zone->pages_min = tmp;
2546 zone->pages_low = zone->pages_min + tmp / 4;
2547 zone->pages_high = zone->pages_min + tmp / 2;
2548 spin_unlock_irqrestore(&zone->lru_lock, flags);
2553 * Initialise min_free_kbytes.
2555 * For small machines we want it small (128k min). For large machines
2556 * we want it large (64MB max). But it is not linear, because network
2557 * bandwidth does not increase linearly with machine size. We use
2559 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2560 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2576 static int __init init_per_zone_pages_min(void)
2578 unsigned long lowmem_kbytes;
2580 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2582 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2583 if (min_free_kbytes < 128)
2584 min_free_kbytes = 128;
2585 if (min_free_kbytes > 65536)
2586 min_free_kbytes = 65536;
2587 setup_per_zone_pages_min();
2588 setup_per_zone_lowmem_reserve();
2591 module_init(init_per_zone_pages_min)
2594 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2595 * that we can call two helper functions whenever min_free_kbytes
2598 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2599 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2601 proc_dointvec(table, write, file, buffer, length, ppos);
2602 setup_per_zone_pages_min();
2607 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2608 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2609 * whenever sysctl_lowmem_reserve_ratio changes.
2611 * The reserve ratio obviously has absolutely no relation with the
2612 * pages_min watermarks. The lowmem reserve ratio can only make sense
2613 * if in function of the boot time zone sizes.
2615 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
2616 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2618 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2619 setup_per_zone_lowmem_reserve();
2624 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
2625 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
2626 * can have before it gets flushed back to buddy allocator.
2629 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
2630 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2636 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2637 if (!write || (ret == -EINVAL))
2639 for_each_zone(zone) {
2640 for_each_online_cpu(cpu) {
2642 high = zone->present_pages / percpu_pagelist_fraction;
2643 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
2649 __initdata int hashdist = HASHDIST_DEFAULT;
2652 static int __init set_hashdist(char *str)
2656 hashdist = simple_strtoul(str, &str, 0);
2659 __setup("hashdist=", set_hashdist);
2663 * allocate a large system hash table from bootmem
2664 * - it is assumed that the hash table must contain an exact power-of-2
2665 * quantity of entries
2666 * - limit is the number of hash buckets, not the total allocation size
2668 void *__init alloc_large_system_hash(const char *tablename,
2669 unsigned long bucketsize,
2670 unsigned long numentries,
2673 unsigned int *_hash_shift,
2674 unsigned int *_hash_mask,
2675 unsigned long limit)
2677 unsigned long long max = limit;
2678 unsigned long log2qty, size;
2681 /* allow the kernel cmdline to have a say */
2683 /* round applicable memory size up to nearest megabyte */
2684 numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages;
2685 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2686 numentries >>= 20 - PAGE_SHIFT;
2687 numentries <<= 20 - PAGE_SHIFT;
2689 /* limit to 1 bucket per 2^scale bytes of low memory */
2690 if (scale > PAGE_SHIFT)
2691 numentries >>= (scale - PAGE_SHIFT);
2693 numentries <<= (PAGE_SHIFT - scale);
2695 /* rounded up to nearest power of 2 in size */
2696 numentries = 1UL << (long_log2(numentries) + 1);
2698 /* limit allocation size to 1/16 total memory by default */
2700 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2701 do_div(max, bucketsize);
2704 if (numentries > max)
2707 log2qty = long_log2(numentries);
2710 size = bucketsize << log2qty;
2711 if (flags & HASH_EARLY)
2712 table = alloc_bootmem(size);
2714 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
2716 unsigned long order;
2717 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
2719 table = (void*) __get_free_pages(GFP_ATOMIC, order);
2721 } while (!table && size > PAGE_SIZE && --log2qty);
2724 panic("Failed to allocate %s hash table\n", tablename);
2726 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2729 long_log2(size) - PAGE_SHIFT,
2733 *_hash_shift = log2qty;
2735 *_hash_mask = (1 << log2qty) - 1;