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 fastcall free_hot_cold_page(struct page *page, int cold);
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.
176 static void prep_compound_page(struct page *page, unsigned long order)
179 int nr_pages = 1 << order;
181 page[1].lru.next = NULL; /* set dtor */
182 page[1].lru.prev = (void *)order;
183 for (i = 0; i < nr_pages; i++) {
184 struct page *p = page + i;
187 set_page_private(p, (unsigned long)page);
191 static void destroy_compound_page(struct page *page, unsigned long order)
194 int nr_pages = 1 << order;
196 if (unlikely((unsigned long)page[1].lru.prev != order))
199 for (i = 0; i < nr_pages; i++) {
200 struct page *p = page + i;
202 if (unlikely(!PageCompound(p) |
203 (page_private(p) != (unsigned long)page)))
205 ClearPageCompound(p);
210 * function for dealing with page's order in buddy system.
211 * zone->lock is already acquired when we use these.
212 * So, we don't need atomic page->flags operations here.
214 static inline unsigned long page_order(struct page *page) {
215 return page_private(page);
218 static inline void set_page_order(struct page *page, int order) {
219 set_page_private(page, order);
220 __SetPagePrivate(page);
223 static inline void rmv_page_order(struct page *page)
225 __ClearPagePrivate(page);
226 set_page_private(page, 0);
230 * Locate the struct page for both the matching buddy in our
231 * pair (buddy1) and the combined O(n+1) page they form (page).
233 * 1) Any buddy B1 will have an order O twin B2 which satisfies
234 * the following equation:
236 * For example, if the starting buddy (buddy2) is #8 its order
238 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
240 * 2) Any buddy B will have an order O+1 parent P which
241 * satisfies the following equation:
244 * Assumption: *_mem_map is contigious at least up to MAX_ORDER
246 static inline struct page *
247 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
249 unsigned long buddy_idx = page_idx ^ (1 << order);
251 return page + (buddy_idx - page_idx);
254 static inline unsigned long
255 __find_combined_index(unsigned long page_idx, unsigned int order)
257 return (page_idx & ~(1 << order));
261 * This function checks whether a page is free && is the buddy
262 * we can do coalesce a page and its buddy if
263 * (a) the buddy is not in a hole &&
264 * (b) the buddy is free &&
265 * (c) the buddy is on the buddy system &&
266 * (d) a page and its buddy have the same order.
267 * for recording page's order, we use page_private(page) and PG_private.
270 static inline int page_is_buddy(struct page *page, int order)
272 #ifdef CONFIG_HOLES_IN_ZONE
273 if (!pfn_valid(page_to_pfn(page)))
277 if (PagePrivate(page) &&
278 (page_order(page) == order) &&
279 page_count(page) == 0)
285 * Freeing function for a buddy system allocator.
287 * The concept of a buddy system is to maintain direct-mapped table
288 * (containing bit values) for memory blocks of various "orders".
289 * The bottom level table contains the map for the smallest allocatable
290 * units of memory (here, pages), and each level above it describes
291 * pairs of units from the levels below, hence, "buddies".
292 * At a high level, all that happens here is marking the table entry
293 * at the bottom level available, and propagating the changes upward
294 * as necessary, plus some accounting needed to play nicely with other
295 * parts of the VM system.
296 * At each level, we keep a list of pages, which are heads of continuous
297 * free pages of length of (1 << order) and marked with PG_Private.Page's
298 * order is recorded in page_private(page) field.
299 * So when we are allocating or freeing one, we can derive the state of the
300 * other. That is, if we allocate a small block, and both were
301 * free, the remainder of the region must be split into blocks.
302 * If a block is freed, and its buddy is also free, then this
303 * triggers coalescing into a block of larger size.
308 static inline void __free_one_page(struct page *page,
309 struct zone *zone, unsigned int order)
311 unsigned long page_idx;
312 int order_size = 1 << order;
314 if (unlikely(PageCompound(page)))
315 destroy_compound_page(page, order);
317 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
319 BUG_ON(page_idx & (order_size - 1));
320 BUG_ON(bad_range(zone, page));
322 zone->free_pages += order_size;
323 while (order < MAX_ORDER-1) {
324 unsigned long combined_idx;
325 struct free_area *area;
328 buddy = __page_find_buddy(page, page_idx, order);
329 if (!page_is_buddy(buddy, order))
330 break; /* Move the buddy up one level. */
332 list_del(&buddy->lru);
333 area = zone->free_area + order;
335 rmv_page_order(buddy);
336 combined_idx = __find_combined_index(page_idx, order);
337 page = page + (combined_idx - page_idx);
338 page_idx = combined_idx;
341 set_page_order(page, order);
342 list_add(&page->lru, &zone->free_area[order].free_list);
343 zone->free_area[order].nr_free++;
346 static inline int free_pages_check(struct page *page)
348 if (unlikely(page_mapcount(page) |
349 (page->mapping != NULL) |
350 (page_count(page) != 0) |
360 1 << PG_reserved ))))
363 __ClearPageDirty(page);
365 * For now, we report if PG_reserved was found set, but do not
366 * clear it, and do not free the page. But we shall soon need
367 * to do more, for when the ZERO_PAGE count wraps negative.
369 return PageReserved(page);
373 * Frees a list of pages.
374 * Assumes all pages on list are in same zone, and of same order.
375 * count is the number of pages to free.
377 * If the zone was previously in an "all pages pinned" state then look to
378 * see if this freeing clears that state.
380 * And clear the zone's pages_scanned counter, to hold off the "all pages are
381 * pinned" detection logic.
383 static void free_pages_bulk(struct zone *zone, int count,
384 struct list_head *list, int order)
386 spin_lock(&zone->lock);
387 zone->all_unreclaimable = 0;
388 zone->pages_scanned = 0;
392 BUG_ON(list_empty(list));
393 page = list_entry(list->prev, struct page, lru);
394 /* have to delete it as __free_one_page list manipulates */
395 list_del(&page->lru);
396 __free_one_page(page, zone, order);
398 spin_unlock(&zone->lock);
401 static void free_one_page(struct zone *zone, struct page *page, int order)
404 list_add(&page->lru, &list);
405 free_pages_bulk(zone, 1, &list, order);
408 static void __free_pages_ok(struct page *page, unsigned int order)
414 arch_free_page(page, order);
415 if (!PageHighMem(page))
416 mutex_debug_check_no_locks_freed(page_address(page),
420 for (i = 1 ; i < (1 << order) ; ++i)
421 __put_page(page + i);
424 for (i = 0 ; i < (1 << order) ; ++i)
425 reserved += free_pages_check(page + i);
429 kernel_map_pages(page, 1 << order, 0);
430 local_irq_save(flags);
431 __mod_page_state(pgfree, 1 << order);
432 free_one_page(page_zone(page), page, order);
433 local_irq_restore(flags);
437 * permit the bootmem allocator to evade page validation on high-order frees
439 void fastcall __init __free_pages_bootmem(struct page *page, unsigned int order)
442 __ClearPageReserved(page);
443 set_page_count(page, 0);
445 free_hot_cold_page(page, 0);
450 for (loop = 0; loop < BITS_PER_LONG; loop++) {
451 struct page *p = &page[loop];
453 if (loop + 16 < BITS_PER_LONG)
455 __ClearPageReserved(p);
456 set_page_count(p, 0);
459 arch_free_page(page, order);
461 mod_page_state(pgfree, 1 << order);
463 list_add(&page->lru, &list);
464 kernel_map_pages(page, 1 << order, 0);
465 free_pages_bulk(page_zone(page), 1, &list, order);
471 * The order of subdivision here is critical for the IO subsystem.
472 * Please do not alter this order without good reasons and regression
473 * testing. Specifically, as large blocks of memory are subdivided,
474 * the order in which smaller blocks are delivered depends on the order
475 * they're subdivided in this function. This is the primary factor
476 * influencing the order in which pages are delivered to the IO
477 * subsystem according to empirical testing, and this is also justified
478 * by considering the behavior of a buddy system containing a single
479 * large block of memory acted on by a series of small allocations.
480 * This behavior is a critical factor in sglist merging's success.
484 static inline void expand(struct zone *zone, struct page *page,
485 int low, int high, struct free_area *area)
487 unsigned long size = 1 << high;
493 BUG_ON(bad_range(zone, &page[size]));
494 list_add(&page[size].lru, &area->free_list);
496 set_page_order(&page[size], high);
501 * This page is about to be returned from the page allocator
503 static int prep_new_page(struct page *page, int order)
505 if (unlikely(page_mapcount(page) |
506 (page->mapping != NULL) |
507 (page_count(page) != 0) |
518 1 << PG_reserved ))))
522 * For now, we report if PG_reserved was found set, but do not
523 * clear it, and do not allocate the page: as a safety net.
525 if (PageReserved(page))
528 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
529 1 << PG_referenced | 1 << PG_arch_1 |
530 1 << PG_checked | 1 << PG_mappedtodisk);
531 set_page_private(page, 0);
532 set_page_refs(page, order);
533 kernel_map_pages(page, 1 << order, 1);
538 * Do the hard work of removing an element from the buddy allocator.
539 * Call me with the zone->lock already held.
541 static struct page *__rmqueue(struct zone *zone, unsigned int order)
543 struct free_area * area;
544 unsigned int current_order;
547 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
548 area = zone->free_area + current_order;
549 if (list_empty(&area->free_list))
552 page = list_entry(area->free_list.next, struct page, lru);
553 list_del(&page->lru);
554 rmv_page_order(page);
556 zone->free_pages -= 1UL << order;
557 expand(zone, page, order, current_order, area);
565 * Obtain a specified number of elements from the buddy allocator, all under
566 * a single hold of the lock, for efficiency. Add them to the supplied list.
567 * Returns the number of new pages which were placed at *list.
569 static int rmqueue_bulk(struct zone *zone, unsigned int order,
570 unsigned long count, struct list_head *list)
574 spin_lock(&zone->lock);
575 for (i = 0; i < count; ++i) {
576 struct page *page = __rmqueue(zone, order);
577 if (unlikely(page == NULL))
579 list_add_tail(&page->lru, list);
581 spin_unlock(&zone->lock);
586 /* Called from the slab reaper to drain remote pagesets */
587 void drain_remote_pages(void)
593 local_irq_save(flags);
594 for_each_zone(zone) {
595 struct per_cpu_pageset *pset;
597 /* Do not drain local pagesets */
598 if (zone->zone_pgdat->node_id == numa_node_id())
601 pset = zone_pcp(zone, smp_processor_id());
602 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
603 struct per_cpu_pages *pcp;
606 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
610 local_irq_restore(flags);
614 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
615 static void __drain_pages(unsigned int cpu)
621 for_each_zone(zone) {
622 struct per_cpu_pageset *pset;
624 pset = zone_pcp(zone, cpu);
625 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
626 struct per_cpu_pages *pcp;
629 local_irq_save(flags);
630 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
632 local_irq_restore(flags);
636 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
640 void mark_free_pages(struct zone *zone)
642 unsigned long zone_pfn, flags;
644 struct list_head *curr;
646 if (!zone->spanned_pages)
649 spin_lock_irqsave(&zone->lock, flags);
650 for (zone_pfn = 0; zone_pfn < zone->spanned_pages; ++zone_pfn)
651 ClearPageNosaveFree(pfn_to_page(zone_pfn + zone->zone_start_pfn));
653 for (order = MAX_ORDER - 1; order >= 0; --order)
654 list_for_each(curr, &zone->free_area[order].free_list) {
655 unsigned long start_pfn, i;
657 start_pfn = page_to_pfn(list_entry(curr, struct page, lru));
659 for (i=0; i < (1<<order); i++)
660 SetPageNosaveFree(pfn_to_page(start_pfn+i));
662 spin_unlock_irqrestore(&zone->lock, flags);
666 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
668 void drain_local_pages(void)
672 local_irq_save(flags);
673 __drain_pages(smp_processor_id());
674 local_irq_restore(flags);
676 #endif /* CONFIG_PM */
678 static void zone_statistics(struct zonelist *zonelist, struct zone *z, int cpu)
681 pg_data_t *pg = z->zone_pgdat;
682 pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
683 struct per_cpu_pageset *p;
685 p = zone_pcp(z, cpu);
690 zone_pcp(zonelist->zones[0], cpu)->numa_foreign++;
692 if (pg == NODE_DATA(numa_node_id()))
700 * Free a 0-order page
702 static void fastcall free_hot_cold_page(struct page *page, int cold)
704 struct zone *zone = page_zone(page);
705 struct per_cpu_pages *pcp;
708 arch_free_page(page, 0);
711 page->mapping = NULL;
712 if (free_pages_check(page))
715 kernel_map_pages(page, 1, 0);
717 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
718 local_irq_save(flags);
719 __inc_page_state(pgfree);
720 list_add(&page->lru, &pcp->list);
722 if (pcp->count >= pcp->high) {
723 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
724 pcp->count -= pcp->batch;
726 local_irq_restore(flags);
730 void fastcall free_hot_page(struct page *page)
732 free_hot_cold_page(page, 0);
735 void fastcall free_cold_page(struct page *page)
737 free_hot_cold_page(page, 1);
740 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
744 BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
745 for(i = 0; i < (1 << order); i++)
746 clear_highpage(page + i);
750 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
751 * we cheat by calling it from here, in the order > 0 path. Saves a branch
754 static struct page *buffered_rmqueue(struct zonelist *zonelist,
755 struct zone *zone, int order, gfp_t gfp_flags)
759 int cold = !!(gfp_flags & __GFP_COLD);
764 if (likely(order == 0)) {
765 struct per_cpu_pages *pcp;
767 pcp = &zone_pcp(zone, cpu)->pcp[cold];
768 local_irq_save(flags);
770 pcp->count += rmqueue_bulk(zone, 0,
771 pcp->batch, &pcp->list);
772 if (unlikely(!pcp->count))
775 page = list_entry(pcp->list.next, struct page, lru);
776 list_del(&page->lru);
779 spin_lock_irqsave(&zone->lock, flags);
780 page = __rmqueue(zone, order);
781 spin_unlock(&zone->lock);
786 __mod_page_state_zone(zone, pgalloc, 1 << order);
787 zone_statistics(zonelist, zone, cpu);
788 local_irq_restore(flags);
791 BUG_ON(bad_range(zone, page));
792 if (prep_new_page(page, order))
795 if (gfp_flags & __GFP_ZERO)
796 prep_zero_page(page, order, gfp_flags);
798 if (order && (gfp_flags & __GFP_COMP))
799 prep_compound_page(page, order);
803 local_irq_restore(flags);
808 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
809 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
810 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
811 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
812 #define ALLOC_HARDER 0x10 /* try to alloc harder */
813 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
814 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
817 * Return 1 if free pages are above 'mark'. This takes into account the order
820 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
821 int classzone_idx, int alloc_flags)
823 /* free_pages my go negative - that's OK */
824 long min = mark, free_pages = z->free_pages - (1 << order) + 1;
827 if (alloc_flags & ALLOC_HIGH)
829 if (alloc_flags & ALLOC_HARDER)
832 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
834 for (o = 0; o < order; o++) {
835 /* At the next order, this order's pages become unavailable */
836 free_pages -= z->free_area[o].nr_free << o;
838 /* Require fewer higher order pages to be free */
841 if (free_pages <= min)
848 * get_page_from_freeliest goes through the zonelist trying to allocate
852 get_page_from_freelist(gfp_t gfp_mask, unsigned int order,
853 struct zonelist *zonelist, int alloc_flags)
855 struct zone **z = zonelist->zones;
856 struct page *page = NULL;
857 int classzone_idx = zone_idx(*z);
860 * Go through the zonelist once, looking for a zone with enough free.
861 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
864 if ((alloc_flags & ALLOC_CPUSET) &&
865 !cpuset_zone_allowed(*z, gfp_mask))
868 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
870 if (alloc_flags & ALLOC_WMARK_MIN)
871 mark = (*z)->pages_min;
872 else if (alloc_flags & ALLOC_WMARK_LOW)
873 mark = (*z)->pages_low;
875 mark = (*z)->pages_high;
876 if (!zone_watermark_ok(*z, order, mark,
877 classzone_idx, alloc_flags))
878 if (!zone_reclaim_mode ||
879 !zone_reclaim(*z, gfp_mask, order))
883 page = buffered_rmqueue(zonelist, *z, order, gfp_mask);
887 } while (*(++z) != NULL);
892 * This is the 'heart' of the zoned buddy allocator.
894 struct page * fastcall
895 __alloc_pages(gfp_t gfp_mask, unsigned int order,
896 struct zonelist *zonelist)
898 const gfp_t wait = gfp_mask & __GFP_WAIT;
901 struct reclaim_state reclaim_state;
902 struct task_struct *p = current;
905 int did_some_progress;
907 might_sleep_if(wait);
910 z = zonelist->zones; /* the list of zones suitable for gfp_mask */
912 if (unlikely(*z == NULL)) {
913 /* Should this ever happen?? */
917 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
918 zonelist, ALLOC_WMARK_LOW|ALLOC_CPUSET);
923 wakeup_kswapd(*z, order);
927 * OK, we're below the kswapd watermark and have kicked background
928 * reclaim. Now things get more complex, so set up alloc_flags according
929 * to how we want to proceed.
931 * The caller may dip into page reserves a bit more if the caller
932 * cannot run direct reclaim, or if the caller has realtime scheduling
933 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
934 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
936 alloc_flags = ALLOC_WMARK_MIN;
937 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
938 alloc_flags |= ALLOC_HARDER;
939 if (gfp_mask & __GFP_HIGH)
940 alloc_flags |= ALLOC_HIGH;
941 alloc_flags |= ALLOC_CPUSET;
944 * Go through the zonelist again. Let __GFP_HIGH and allocations
945 * coming from realtime tasks go deeper into reserves.
947 * This is the last chance, in general, before the goto nopage.
948 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
949 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
951 page = get_page_from_freelist(gfp_mask, order, zonelist, alloc_flags);
955 /* This allocation should allow future memory freeing. */
957 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
958 && !in_interrupt()) {
959 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
961 /* go through the zonelist yet again, ignoring mins */
962 page = get_page_from_freelist(gfp_mask, order,
963 zonelist, ALLOC_NO_WATERMARKS);
966 if (gfp_mask & __GFP_NOFAIL) {
967 blk_congestion_wait(WRITE, HZ/50);
974 /* Atomic allocations - we can't balance anything */
981 /* We now go into synchronous reclaim */
982 cpuset_memory_pressure_bump();
983 p->flags |= PF_MEMALLOC;
984 reclaim_state.reclaimed_slab = 0;
985 p->reclaim_state = &reclaim_state;
987 did_some_progress = try_to_free_pages(zonelist->zones, gfp_mask);
989 p->reclaim_state = NULL;
990 p->flags &= ~PF_MEMALLOC;
994 if (likely(did_some_progress)) {
995 page = get_page_from_freelist(gfp_mask, order,
996 zonelist, alloc_flags);
999 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1001 * Go through the zonelist yet one more time, keep
1002 * very high watermark here, this is only to catch
1003 * a parallel oom killing, we must fail if we're still
1004 * under heavy pressure.
1006 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1007 zonelist, ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1011 out_of_memory(gfp_mask, order);
1016 * Don't let big-order allocations loop unless the caller explicitly
1017 * requests that. Wait for some write requests to complete then retry.
1019 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
1020 * <= 3, but that may not be true in other implementations.
1023 if (!(gfp_mask & __GFP_NORETRY)) {
1024 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
1026 if (gfp_mask & __GFP_NOFAIL)
1030 blk_congestion_wait(WRITE, HZ/50);
1035 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1036 printk(KERN_WARNING "%s: page allocation failure."
1037 " order:%d, mode:0x%x\n",
1038 p->comm, order, gfp_mask);
1046 EXPORT_SYMBOL(__alloc_pages);
1049 * Common helper functions.
1051 fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1054 page = alloc_pages(gfp_mask, order);
1057 return (unsigned long) page_address(page);
1060 EXPORT_SYMBOL(__get_free_pages);
1062 fastcall unsigned long get_zeroed_page(gfp_t gfp_mask)
1067 * get_zeroed_page() returns a 32-bit address, which cannot represent
1070 BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1072 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1074 return (unsigned long) page_address(page);
1078 EXPORT_SYMBOL(get_zeroed_page);
1080 void __pagevec_free(struct pagevec *pvec)
1082 int i = pagevec_count(pvec);
1085 free_hot_cold_page(pvec->pages[i], pvec->cold);
1088 fastcall void __free_pages(struct page *page, unsigned int order)
1090 if (put_page_testzero(page)) {
1092 free_hot_page(page);
1094 __free_pages_ok(page, order);
1098 EXPORT_SYMBOL(__free_pages);
1100 fastcall void free_pages(unsigned long addr, unsigned int order)
1103 BUG_ON(!virt_addr_valid((void *)addr));
1104 __free_pages(virt_to_page((void *)addr), order);
1108 EXPORT_SYMBOL(free_pages);
1111 * Total amount of free (allocatable) RAM:
1113 unsigned int nr_free_pages(void)
1115 unsigned int sum = 0;
1119 sum += zone->free_pages;
1124 EXPORT_SYMBOL(nr_free_pages);
1127 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
1129 unsigned int i, sum = 0;
1131 for (i = 0; i < MAX_NR_ZONES; i++)
1132 sum += pgdat->node_zones[i].free_pages;
1138 static unsigned int nr_free_zone_pages(int offset)
1140 /* Just pick one node, since fallback list is circular */
1141 pg_data_t *pgdat = NODE_DATA(numa_node_id());
1142 unsigned int sum = 0;
1144 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1145 struct zone **zonep = zonelist->zones;
1148 for (zone = *zonep++; zone; zone = *zonep++) {
1149 unsigned long size = zone->present_pages;
1150 unsigned long high = zone->pages_high;
1159 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1161 unsigned int nr_free_buffer_pages(void)
1163 return nr_free_zone_pages(gfp_zone(GFP_USER));
1167 * Amount of free RAM allocatable within all zones
1169 unsigned int nr_free_pagecache_pages(void)
1171 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER));
1174 #ifdef CONFIG_HIGHMEM
1175 unsigned int nr_free_highpages (void)
1178 unsigned int pages = 0;
1180 for_each_pgdat(pgdat)
1181 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1188 static void show_node(struct zone *zone)
1190 printk("Node %d ", zone->zone_pgdat->node_id);
1193 #define show_node(zone) do { } while (0)
1197 * Accumulate the page_state information across all CPUs.
1198 * The result is unavoidably approximate - it can change
1199 * during and after execution of this function.
1201 static DEFINE_PER_CPU(struct page_state, page_states) = {0};
1203 atomic_t nr_pagecache = ATOMIC_INIT(0);
1204 EXPORT_SYMBOL(nr_pagecache);
1206 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
1209 static void __get_page_state(struct page_state *ret, int nr, cpumask_t *cpumask)
1213 memset(ret, 0, nr * sizeof(unsigned long));
1214 cpus_and(*cpumask, *cpumask, cpu_online_map);
1216 cpu = first_cpu(*cpumask);
1217 while (cpu < NR_CPUS) {
1218 unsigned long *in, *out, off;
1220 if (!cpu_isset(cpu, *cpumask))
1223 in = (unsigned long *)&per_cpu(page_states, cpu);
1225 cpu = next_cpu(cpu, *cpumask);
1227 if (likely(cpu < NR_CPUS))
1228 prefetch(&per_cpu(page_states, cpu));
1230 out = (unsigned long *)ret;
1231 for (off = 0; off < nr; off++)
1236 void get_page_state_node(struct page_state *ret, int node)
1239 cpumask_t mask = node_to_cpumask(node);
1241 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
1242 nr /= sizeof(unsigned long);
1244 __get_page_state(ret, nr+1, &mask);
1247 void get_page_state(struct page_state *ret)
1250 cpumask_t mask = CPU_MASK_ALL;
1252 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
1253 nr /= sizeof(unsigned long);
1255 __get_page_state(ret, nr + 1, &mask);
1258 void get_full_page_state(struct page_state *ret)
1260 cpumask_t mask = CPU_MASK_ALL;
1262 __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long), &mask);
1265 unsigned long read_page_state_offset(unsigned long offset)
1267 unsigned long ret = 0;
1270 for_each_online_cpu(cpu) {
1273 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
1274 ret += *((unsigned long *)in);
1279 void __mod_page_state_offset(unsigned long offset, unsigned long delta)
1283 ptr = &__get_cpu_var(page_states);
1284 *(unsigned long *)(ptr + offset) += delta;
1286 EXPORT_SYMBOL(__mod_page_state_offset);
1288 void mod_page_state_offset(unsigned long offset, unsigned long delta)
1290 unsigned long flags;
1293 local_irq_save(flags);
1294 ptr = &__get_cpu_var(page_states);
1295 *(unsigned long *)(ptr + offset) += delta;
1296 local_irq_restore(flags);
1298 EXPORT_SYMBOL(mod_page_state_offset);
1300 void __get_zone_counts(unsigned long *active, unsigned long *inactive,
1301 unsigned long *free, struct pglist_data *pgdat)
1303 struct zone *zones = pgdat->node_zones;
1309 for (i = 0; i < MAX_NR_ZONES; i++) {
1310 *active += zones[i].nr_active;
1311 *inactive += zones[i].nr_inactive;
1312 *free += zones[i].free_pages;
1316 void get_zone_counts(unsigned long *active,
1317 unsigned long *inactive, unsigned long *free)
1319 struct pglist_data *pgdat;
1324 for_each_pgdat(pgdat) {
1325 unsigned long l, m, n;
1326 __get_zone_counts(&l, &m, &n, pgdat);
1333 void si_meminfo(struct sysinfo *val)
1335 val->totalram = totalram_pages;
1337 val->freeram = nr_free_pages();
1338 val->bufferram = nr_blockdev_pages();
1339 #ifdef CONFIG_HIGHMEM
1340 val->totalhigh = totalhigh_pages;
1341 val->freehigh = nr_free_highpages();
1346 val->mem_unit = PAGE_SIZE;
1349 EXPORT_SYMBOL(si_meminfo);
1352 void si_meminfo_node(struct sysinfo *val, int nid)
1354 pg_data_t *pgdat = NODE_DATA(nid);
1356 val->totalram = pgdat->node_present_pages;
1357 val->freeram = nr_free_pages_pgdat(pgdat);
1358 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1359 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1360 val->mem_unit = PAGE_SIZE;
1364 #define K(x) ((x) << (PAGE_SHIFT-10))
1367 * Show free area list (used inside shift_scroll-lock stuff)
1368 * We also calculate the percentage fragmentation. We do this by counting the
1369 * memory on each free list with the exception of the first item on the list.
1371 void show_free_areas(void)
1373 struct page_state ps;
1374 int cpu, temperature;
1375 unsigned long active;
1376 unsigned long inactive;
1380 for_each_zone(zone) {
1382 printk("%s per-cpu:", zone->name);
1384 if (!populated_zone(zone)) {
1390 for_each_online_cpu(cpu) {
1391 struct per_cpu_pageset *pageset;
1393 pageset = zone_pcp(zone, cpu);
1395 for (temperature = 0; temperature < 2; temperature++)
1396 printk("cpu %d %s: high %d, batch %d used:%d\n",
1398 temperature ? "cold" : "hot",
1399 pageset->pcp[temperature].high,
1400 pageset->pcp[temperature].batch,
1401 pageset->pcp[temperature].count);
1405 get_page_state(&ps);
1406 get_zone_counts(&active, &inactive, &free);
1408 printk("Free pages: %11ukB (%ukB HighMem)\n",
1410 K(nr_free_highpages()));
1412 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1413 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1422 ps.nr_page_table_pages);
1424 for_each_zone(zone) {
1436 " pages_scanned:%lu"
1437 " all_unreclaimable? %s"
1440 K(zone->free_pages),
1443 K(zone->pages_high),
1445 K(zone->nr_inactive),
1446 K(zone->present_pages),
1447 zone->pages_scanned,
1448 (zone->all_unreclaimable ? "yes" : "no")
1450 printk("lowmem_reserve[]:");
1451 for (i = 0; i < MAX_NR_ZONES; i++)
1452 printk(" %lu", zone->lowmem_reserve[i]);
1456 for_each_zone(zone) {
1457 unsigned long nr, flags, order, total = 0;
1460 printk("%s: ", zone->name);
1461 if (!populated_zone(zone)) {
1466 spin_lock_irqsave(&zone->lock, flags);
1467 for (order = 0; order < MAX_ORDER; order++) {
1468 nr = zone->free_area[order].nr_free;
1469 total += nr << order;
1470 printk("%lu*%lukB ", nr, K(1UL) << order);
1472 spin_unlock_irqrestore(&zone->lock, flags);
1473 printk("= %lukB\n", K(total));
1476 show_swap_cache_info();
1480 * Builds allocation fallback zone lists.
1482 * Add all populated zones of a node to the zonelist.
1484 static int __init build_zonelists_node(pg_data_t *pgdat,
1485 struct zonelist *zonelist, int nr_zones, int zone_type)
1489 BUG_ON(zone_type > ZONE_HIGHMEM);
1492 zone = pgdat->node_zones + zone_type;
1493 if (populated_zone(zone)) {
1494 #ifndef CONFIG_HIGHMEM
1495 BUG_ON(zone_type > ZONE_NORMAL);
1497 zonelist->zones[nr_zones++] = zone;
1498 check_highest_zone(zone_type);
1502 } while (zone_type >= 0);
1506 static inline int highest_zone(int zone_bits)
1508 int res = ZONE_NORMAL;
1509 if (zone_bits & (__force int)__GFP_HIGHMEM)
1511 if (zone_bits & (__force int)__GFP_DMA32)
1513 if (zone_bits & (__force int)__GFP_DMA)
1519 #define MAX_NODE_LOAD (num_online_nodes())
1520 static int __initdata node_load[MAX_NUMNODES];
1522 * find_next_best_node - find the next node that should appear in a given node's fallback list
1523 * @node: node whose fallback list we're appending
1524 * @used_node_mask: nodemask_t of already used nodes
1526 * We use a number of factors to determine which is the next node that should
1527 * appear on a given node's fallback list. The node should not have appeared
1528 * already in @node's fallback list, and it should be the next closest node
1529 * according to the distance array (which contains arbitrary distance values
1530 * from each node to each node in the system), and should also prefer nodes
1531 * with no CPUs, since presumably they'll have very little allocation pressure
1532 * on them otherwise.
1533 * It returns -1 if no node is found.
1535 static int __init find_next_best_node(int node, nodemask_t *used_node_mask)
1538 int min_val = INT_MAX;
1541 for_each_online_node(i) {
1544 /* Start from local node */
1545 n = (node+i) % num_online_nodes();
1547 /* Don't want a node to appear more than once */
1548 if (node_isset(n, *used_node_mask))
1551 /* Use the local node if we haven't already */
1552 if (!node_isset(node, *used_node_mask)) {
1557 /* Use the distance array to find the distance */
1558 val = node_distance(node, n);
1560 /* Give preference to headless and unused nodes */
1561 tmp = node_to_cpumask(n);
1562 if (!cpus_empty(tmp))
1563 val += PENALTY_FOR_NODE_WITH_CPUS;
1565 /* Slight preference for less loaded node */
1566 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1567 val += node_load[n];
1569 if (val < min_val) {
1576 node_set(best_node, *used_node_mask);
1581 static void __init build_zonelists(pg_data_t *pgdat)
1583 int i, j, k, node, local_node;
1584 int prev_node, load;
1585 struct zonelist *zonelist;
1586 nodemask_t used_mask;
1588 /* initialize zonelists */
1589 for (i = 0; i < GFP_ZONETYPES; i++) {
1590 zonelist = pgdat->node_zonelists + i;
1591 zonelist->zones[0] = NULL;
1594 /* NUMA-aware ordering of nodes */
1595 local_node = pgdat->node_id;
1596 load = num_online_nodes();
1597 prev_node = local_node;
1598 nodes_clear(used_mask);
1599 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1600 int distance = node_distance(local_node, node);
1603 * If another node is sufficiently far away then it is better
1604 * to reclaim pages in a zone before going off node.
1606 if (distance > RECLAIM_DISTANCE)
1607 zone_reclaim_mode = 1;
1610 * We don't want to pressure a particular node.
1611 * So adding penalty to the first node in same
1612 * distance group to make it round-robin.
1615 if (distance != node_distance(local_node, prev_node))
1616 node_load[node] += load;
1619 for (i = 0; i < GFP_ZONETYPES; i++) {
1620 zonelist = pgdat->node_zonelists + i;
1621 for (j = 0; zonelist->zones[j] != NULL; j++);
1623 k = highest_zone(i);
1625 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1626 zonelist->zones[j] = NULL;
1631 #else /* CONFIG_NUMA */
1633 static void __init build_zonelists(pg_data_t *pgdat)
1635 int i, j, k, node, local_node;
1637 local_node = pgdat->node_id;
1638 for (i = 0; i < GFP_ZONETYPES; i++) {
1639 struct zonelist *zonelist;
1641 zonelist = pgdat->node_zonelists + i;
1644 k = highest_zone(i);
1645 j = build_zonelists_node(pgdat, zonelist, j, k);
1647 * Now we build the zonelist so that it contains the zones
1648 * of all the other nodes.
1649 * We don't want to pressure a particular node, so when
1650 * building the zones for node N, we make sure that the
1651 * zones coming right after the local ones are those from
1652 * node N+1 (modulo N)
1654 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1655 if (!node_online(node))
1657 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1659 for (node = 0; node < local_node; node++) {
1660 if (!node_online(node))
1662 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1665 zonelist->zones[j] = NULL;
1669 #endif /* CONFIG_NUMA */
1671 void __init build_all_zonelists(void)
1675 for_each_online_node(i)
1676 build_zonelists(NODE_DATA(i));
1677 printk("Built %i zonelists\n", num_online_nodes());
1678 cpuset_init_current_mems_allowed();
1682 * Helper functions to size the waitqueue hash table.
1683 * Essentially these want to choose hash table sizes sufficiently
1684 * large so that collisions trying to wait on pages are rare.
1685 * But in fact, the number of active page waitqueues on typical
1686 * systems is ridiculously low, less than 200. So this is even
1687 * conservative, even though it seems large.
1689 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1690 * waitqueues, i.e. the size of the waitq table given the number of pages.
1692 #define PAGES_PER_WAITQUEUE 256
1694 static inline unsigned long wait_table_size(unsigned long pages)
1696 unsigned long size = 1;
1698 pages /= PAGES_PER_WAITQUEUE;
1700 while (size < pages)
1704 * Once we have dozens or even hundreds of threads sleeping
1705 * on IO we've got bigger problems than wait queue collision.
1706 * Limit the size of the wait table to a reasonable size.
1708 size = min(size, 4096UL);
1710 return max(size, 4UL);
1714 * This is an integer logarithm so that shifts can be used later
1715 * to extract the more random high bits from the multiplicative
1716 * hash function before the remainder is taken.
1718 static inline unsigned long wait_table_bits(unsigned long size)
1723 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1725 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1726 unsigned long *zones_size, unsigned long *zholes_size)
1728 unsigned long realtotalpages, totalpages = 0;
1731 for (i = 0; i < MAX_NR_ZONES; i++)
1732 totalpages += zones_size[i];
1733 pgdat->node_spanned_pages = totalpages;
1735 realtotalpages = totalpages;
1737 for (i = 0; i < MAX_NR_ZONES; i++)
1738 realtotalpages -= zholes_size[i];
1739 pgdat->node_present_pages = realtotalpages;
1740 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1745 * Initially all pages are reserved - free ones are freed
1746 * up by free_all_bootmem() once the early boot process is
1747 * done. Non-atomic initialization, single-pass.
1749 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1750 unsigned long start_pfn)
1753 unsigned long end_pfn = start_pfn + size;
1756 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1757 if (!early_pfn_valid(pfn))
1759 page = pfn_to_page(pfn);
1760 set_page_links(page, zone, nid, pfn);
1761 set_page_count(page, 1);
1762 reset_page_mapcount(page);
1763 SetPageReserved(page);
1764 INIT_LIST_HEAD(&page->lru);
1765 #ifdef WANT_PAGE_VIRTUAL
1766 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1767 if (!is_highmem_idx(zone))
1768 set_page_address(page, __va(pfn << PAGE_SHIFT));
1773 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1777 for (order = 0; order < MAX_ORDER ; order++) {
1778 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1779 zone->free_area[order].nr_free = 0;
1783 #define ZONETABLE_INDEX(x, zone_nr) ((x << ZONES_SHIFT) | zone_nr)
1784 void zonetable_add(struct zone *zone, int nid, int zid, unsigned long pfn,
1787 unsigned long snum = pfn_to_section_nr(pfn);
1788 unsigned long end = pfn_to_section_nr(pfn + size);
1791 zone_table[ZONETABLE_INDEX(nid, zid)] = zone;
1793 for (; snum <= end; snum++)
1794 zone_table[ZONETABLE_INDEX(snum, zid)] = zone;
1797 #ifndef __HAVE_ARCH_MEMMAP_INIT
1798 #define memmap_init(size, nid, zone, start_pfn) \
1799 memmap_init_zone((size), (nid), (zone), (start_pfn))
1802 static int __cpuinit zone_batchsize(struct zone *zone)
1807 * The per-cpu-pages pools are set to around 1000th of the
1808 * size of the zone. But no more than 1/2 of a meg.
1810 * OK, so we don't know how big the cache is. So guess.
1812 batch = zone->present_pages / 1024;
1813 if (batch * PAGE_SIZE > 512 * 1024)
1814 batch = (512 * 1024) / PAGE_SIZE;
1815 batch /= 4; /* We effectively *= 4 below */
1820 * Clamp the batch to a 2^n - 1 value. Having a power
1821 * of 2 value was found to be more likely to have
1822 * suboptimal cache aliasing properties in some cases.
1824 * For example if 2 tasks are alternately allocating
1825 * batches of pages, one task can end up with a lot
1826 * of pages of one half of the possible page colors
1827 * and the other with pages of the other colors.
1829 batch = (1 << (fls(batch + batch/2)-1)) - 1;
1834 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
1836 struct per_cpu_pages *pcp;
1838 memset(p, 0, sizeof(*p));
1840 pcp = &p->pcp[0]; /* hot */
1842 pcp->high = 6 * batch;
1843 pcp->batch = max(1UL, 1 * batch);
1844 INIT_LIST_HEAD(&pcp->list);
1846 pcp = &p->pcp[1]; /* cold*/
1848 pcp->high = 2 * batch;
1849 pcp->batch = max(1UL, batch/2);
1850 INIT_LIST_HEAD(&pcp->list);
1854 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
1855 * to the value high for the pageset p.
1858 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
1861 struct per_cpu_pages *pcp;
1863 pcp = &p->pcp[0]; /* hot list */
1865 pcp->batch = max(1UL, high/4);
1866 if ((high/4) > (PAGE_SHIFT * 8))
1867 pcp->batch = PAGE_SHIFT * 8;
1873 * Boot pageset table. One per cpu which is going to be used for all
1874 * zones and all nodes. The parameters will be set in such a way
1875 * that an item put on a list will immediately be handed over to
1876 * the buddy list. This is safe since pageset manipulation is done
1877 * with interrupts disabled.
1879 * Some NUMA counter updates may also be caught by the boot pagesets.
1881 * The boot_pagesets must be kept even after bootup is complete for
1882 * unused processors and/or zones. They do play a role for bootstrapping
1883 * hotplugged processors.
1885 * zoneinfo_show() and maybe other functions do
1886 * not check if the processor is online before following the pageset pointer.
1887 * Other parts of the kernel may not check if the zone is available.
1889 static struct per_cpu_pageset boot_pageset[NR_CPUS];
1892 * Dynamically allocate memory for the
1893 * per cpu pageset array in struct zone.
1895 static int __cpuinit process_zones(int cpu)
1897 struct zone *zone, *dzone;
1899 for_each_zone(zone) {
1901 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
1902 GFP_KERNEL, cpu_to_node(cpu));
1903 if (!zone_pcp(zone, cpu))
1906 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
1908 if (percpu_pagelist_fraction)
1909 setup_pagelist_highmark(zone_pcp(zone, cpu),
1910 (zone->present_pages / percpu_pagelist_fraction));
1915 for_each_zone(dzone) {
1918 kfree(zone_pcp(dzone, cpu));
1919 zone_pcp(dzone, cpu) = NULL;
1924 static inline void free_zone_pagesets(int cpu)
1928 for_each_zone(zone) {
1929 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
1931 zone_pcp(zone, cpu) = NULL;
1936 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
1937 unsigned long action,
1940 int cpu = (long)hcpu;
1941 int ret = NOTIFY_OK;
1944 case CPU_UP_PREPARE:
1945 if (process_zones(cpu))
1948 case CPU_UP_CANCELED:
1950 free_zone_pagesets(cpu);
1958 static struct notifier_block pageset_notifier =
1959 { &pageset_cpuup_callback, NULL, 0 };
1961 void __init setup_per_cpu_pageset(void)
1965 /* Initialize per_cpu_pageset for cpu 0.
1966 * A cpuup callback will do this for every cpu
1967 * as it comes online
1969 err = process_zones(smp_processor_id());
1971 register_cpu_notifier(&pageset_notifier);
1977 void zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
1980 struct pglist_data *pgdat = zone->zone_pgdat;
1983 * The per-page waitqueue mechanism uses hashed waitqueues
1986 zone->wait_table_size = wait_table_size(zone_size_pages);
1987 zone->wait_table_bits = wait_table_bits(zone->wait_table_size);
1988 zone->wait_table = (wait_queue_head_t *)
1989 alloc_bootmem_node(pgdat, zone->wait_table_size
1990 * sizeof(wait_queue_head_t));
1992 for(i = 0; i < zone->wait_table_size; ++i)
1993 init_waitqueue_head(zone->wait_table + i);
1996 static __meminit void zone_pcp_init(struct zone *zone)
1999 unsigned long batch = zone_batchsize(zone);
2001 for (cpu = 0; cpu < NR_CPUS; cpu++) {
2003 /* Early boot. Slab allocator not functional yet */
2004 zone_pcp(zone, cpu) = &boot_pageset[cpu];
2005 setup_pageset(&boot_pageset[cpu],0);
2007 setup_pageset(zone_pcp(zone,cpu), batch);
2010 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
2011 zone->name, zone->present_pages, batch);
2014 static __meminit void init_currently_empty_zone(struct zone *zone,
2015 unsigned long zone_start_pfn, unsigned long size)
2017 struct pglist_data *pgdat = zone->zone_pgdat;
2019 zone_wait_table_init(zone, size);
2020 pgdat->nr_zones = zone_idx(zone) + 1;
2022 zone->zone_mem_map = pfn_to_page(zone_start_pfn);
2023 zone->zone_start_pfn = zone_start_pfn;
2025 memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn);
2027 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
2031 * Set up the zone data structures:
2032 * - mark all pages reserved
2033 * - mark all memory queues empty
2034 * - clear the memory bitmaps
2036 static void __init free_area_init_core(struct pglist_data *pgdat,
2037 unsigned long *zones_size, unsigned long *zholes_size)
2040 int nid = pgdat->node_id;
2041 unsigned long zone_start_pfn = pgdat->node_start_pfn;
2043 pgdat_resize_init(pgdat);
2044 pgdat->nr_zones = 0;
2045 init_waitqueue_head(&pgdat->kswapd_wait);
2046 pgdat->kswapd_max_order = 0;
2048 for (j = 0; j < MAX_NR_ZONES; j++) {
2049 struct zone *zone = pgdat->node_zones + j;
2050 unsigned long size, realsize;
2052 realsize = size = zones_size[j];
2054 realsize -= zholes_size[j];
2056 if (j < ZONE_HIGHMEM)
2057 nr_kernel_pages += realsize;
2058 nr_all_pages += realsize;
2060 zone->spanned_pages = size;
2061 zone->present_pages = realsize;
2062 zone->name = zone_names[j];
2063 spin_lock_init(&zone->lock);
2064 spin_lock_init(&zone->lru_lock);
2065 zone_seqlock_init(zone);
2066 zone->zone_pgdat = pgdat;
2067 zone->free_pages = 0;
2069 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
2071 zone_pcp_init(zone);
2072 INIT_LIST_HEAD(&zone->active_list);
2073 INIT_LIST_HEAD(&zone->inactive_list);
2074 zone->nr_scan_active = 0;
2075 zone->nr_scan_inactive = 0;
2076 zone->nr_active = 0;
2077 zone->nr_inactive = 0;
2078 atomic_set(&zone->reclaim_in_progress, 0);
2082 zonetable_add(zone, nid, j, zone_start_pfn, size);
2083 init_currently_empty_zone(zone, zone_start_pfn, size);
2084 zone_start_pfn += size;
2088 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
2090 /* Skip empty nodes */
2091 if (!pgdat->node_spanned_pages)
2094 #ifdef CONFIG_FLAT_NODE_MEM_MAP
2095 /* ia64 gets its own node_mem_map, before this, without bootmem */
2096 if (!pgdat->node_mem_map) {
2100 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
2101 map = alloc_remap(pgdat->node_id, size);
2103 map = alloc_bootmem_node(pgdat, size);
2104 pgdat->node_mem_map = map;
2106 #ifdef CONFIG_FLATMEM
2108 * With no DISCONTIG, the global mem_map is just set as node 0's
2110 if (pgdat == NODE_DATA(0))
2111 mem_map = NODE_DATA(0)->node_mem_map;
2113 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
2116 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
2117 unsigned long *zones_size, unsigned long node_start_pfn,
2118 unsigned long *zholes_size)
2120 pgdat->node_id = nid;
2121 pgdat->node_start_pfn = node_start_pfn;
2122 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
2124 alloc_node_mem_map(pgdat);
2126 free_area_init_core(pgdat, zones_size, zholes_size);
2129 #ifndef CONFIG_NEED_MULTIPLE_NODES
2130 static bootmem_data_t contig_bootmem_data;
2131 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
2133 EXPORT_SYMBOL(contig_page_data);
2136 void __init free_area_init(unsigned long *zones_size)
2138 free_area_init_node(0, NODE_DATA(0), zones_size,
2139 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
2142 #ifdef CONFIG_PROC_FS
2144 #include <linux/seq_file.h>
2146 static void *frag_start(struct seq_file *m, loff_t *pos)
2151 for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
2157 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
2159 pg_data_t *pgdat = (pg_data_t *)arg;
2162 return pgdat->pgdat_next;
2165 static void frag_stop(struct seq_file *m, void *arg)
2170 * This walks the free areas for each zone.
2172 static int frag_show(struct seq_file *m, void *arg)
2174 pg_data_t *pgdat = (pg_data_t *)arg;
2176 struct zone *node_zones = pgdat->node_zones;
2177 unsigned long flags;
2180 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2181 if (!populated_zone(zone))
2184 spin_lock_irqsave(&zone->lock, flags);
2185 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
2186 for (order = 0; order < MAX_ORDER; ++order)
2187 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
2188 spin_unlock_irqrestore(&zone->lock, flags);
2194 struct seq_operations fragmentation_op = {
2195 .start = frag_start,
2202 * Output information about zones in @pgdat.
2204 static int zoneinfo_show(struct seq_file *m, void *arg)
2206 pg_data_t *pgdat = arg;
2208 struct zone *node_zones = pgdat->node_zones;
2209 unsigned long flags;
2211 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; zone++) {
2214 if (!populated_zone(zone))
2217 spin_lock_irqsave(&zone->lock, flags);
2218 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
2226 "\n scanned %lu (a: %lu i: %lu)"
2235 zone->pages_scanned,
2236 zone->nr_scan_active, zone->nr_scan_inactive,
2237 zone->spanned_pages,
2238 zone->present_pages);
2240 "\n protection: (%lu",
2241 zone->lowmem_reserve[0]);
2242 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
2243 seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
2247 for_each_online_cpu(i) {
2248 struct per_cpu_pageset *pageset;
2251 pageset = zone_pcp(zone, i);
2252 for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2253 if (pageset->pcp[j].count)
2256 if (j == ARRAY_SIZE(pageset->pcp))
2258 for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2260 "\n cpu: %i pcp: %i"
2265 pageset->pcp[j].count,
2266 pageset->pcp[j].high,
2267 pageset->pcp[j].batch);
2273 "\n numa_foreign: %lu"
2274 "\n interleave_hit: %lu"
2275 "\n local_node: %lu"
2276 "\n other_node: %lu",
2279 pageset->numa_foreign,
2280 pageset->interleave_hit,
2281 pageset->local_node,
2282 pageset->other_node);
2286 "\n all_unreclaimable: %u"
2287 "\n prev_priority: %i"
2288 "\n temp_priority: %i"
2289 "\n start_pfn: %lu",
2290 zone->all_unreclaimable,
2291 zone->prev_priority,
2292 zone->temp_priority,
2293 zone->zone_start_pfn);
2294 spin_unlock_irqrestore(&zone->lock, flags);
2300 struct seq_operations zoneinfo_op = {
2301 .start = frag_start, /* iterate over all zones. The same as in
2305 .show = zoneinfo_show,
2308 static char *vmstat_text[] = {
2312 "nr_page_table_pages",
2343 "pgscan_kswapd_high",
2344 "pgscan_kswapd_normal",
2345 "pgscan_kswapd_dma32",
2346 "pgscan_kswapd_dma",
2348 "pgscan_direct_high",
2349 "pgscan_direct_normal",
2350 "pgscan_direct_dma32",
2351 "pgscan_direct_dma",
2356 "kswapd_inodesteal",
2364 static void *vmstat_start(struct seq_file *m, loff_t *pos)
2366 struct page_state *ps;
2368 if (*pos >= ARRAY_SIZE(vmstat_text))
2371 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
2374 return ERR_PTR(-ENOMEM);
2375 get_full_page_state(ps);
2376 ps->pgpgin /= 2; /* sectors -> kbytes */
2378 return (unsigned long *)ps + *pos;
2381 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
2384 if (*pos >= ARRAY_SIZE(vmstat_text))
2386 return (unsigned long *)m->private + *pos;
2389 static int vmstat_show(struct seq_file *m, void *arg)
2391 unsigned long *l = arg;
2392 unsigned long off = l - (unsigned long *)m->private;
2394 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
2398 static void vmstat_stop(struct seq_file *m, void *arg)
2404 struct seq_operations vmstat_op = {
2405 .start = vmstat_start,
2406 .next = vmstat_next,
2407 .stop = vmstat_stop,
2408 .show = vmstat_show,
2411 #endif /* CONFIG_PROC_FS */
2413 #ifdef CONFIG_HOTPLUG_CPU
2414 static int page_alloc_cpu_notify(struct notifier_block *self,
2415 unsigned long action, void *hcpu)
2417 int cpu = (unsigned long)hcpu;
2419 unsigned long *src, *dest;
2421 if (action == CPU_DEAD) {
2424 /* Drain local pagecache count. */
2425 count = &per_cpu(nr_pagecache_local, cpu);
2426 atomic_add(*count, &nr_pagecache);
2428 local_irq_disable();
2431 /* Add dead cpu's page_states to our own. */
2432 dest = (unsigned long *)&__get_cpu_var(page_states);
2433 src = (unsigned long *)&per_cpu(page_states, cpu);
2435 for (i = 0; i < sizeof(struct page_state)/sizeof(unsigned long);
2445 #endif /* CONFIG_HOTPLUG_CPU */
2447 void __init page_alloc_init(void)
2449 hotcpu_notifier(page_alloc_cpu_notify, 0);
2453 * setup_per_zone_lowmem_reserve - called whenever
2454 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
2455 * has a correct pages reserved value, so an adequate number of
2456 * pages are left in the zone after a successful __alloc_pages().
2458 static void setup_per_zone_lowmem_reserve(void)
2460 struct pglist_data *pgdat;
2463 for_each_pgdat(pgdat) {
2464 for (j = 0; j < MAX_NR_ZONES; j++) {
2465 struct zone *zone = pgdat->node_zones + j;
2466 unsigned long present_pages = zone->present_pages;
2468 zone->lowmem_reserve[j] = 0;
2470 for (idx = j-1; idx >= 0; idx--) {
2471 struct zone *lower_zone;
2473 if (sysctl_lowmem_reserve_ratio[idx] < 1)
2474 sysctl_lowmem_reserve_ratio[idx] = 1;
2476 lower_zone = pgdat->node_zones + idx;
2477 lower_zone->lowmem_reserve[j] = present_pages /
2478 sysctl_lowmem_reserve_ratio[idx];
2479 present_pages += lower_zone->present_pages;
2486 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
2487 * that the pages_{min,low,high} values for each zone are set correctly
2488 * with respect to min_free_kbytes.
2490 void setup_per_zone_pages_min(void)
2492 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
2493 unsigned long lowmem_pages = 0;
2495 unsigned long flags;
2497 /* Calculate total number of !ZONE_HIGHMEM pages */
2498 for_each_zone(zone) {
2499 if (!is_highmem(zone))
2500 lowmem_pages += zone->present_pages;
2503 for_each_zone(zone) {
2505 spin_lock_irqsave(&zone->lru_lock, flags);
2506 tmp = (pages_min * zone->present_pages) / lowmem_pages;
2507 if (is_highmem(zone)) {
2509 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
2510 * need highmem pages, so cap pages_min to a small
2513 * The (pages_high-pages_low) and (pages_low-pages_min)
2514 * deltas controls asynch page reclaim, and so should
2515 * not be capped for highmem.
2519 min_pages = zone->present_pages / 1024;
2520 if (min_pages < SWAP_CLUSTER_MAX)
2521 min_pages = SWAP_CLUSTER_MAX;
2522 if (min_pages > 128)
2524 zone->pages_min = min_pages;
2527 * If it's a lowmem zone, reserve a number of pages
2528 * proportionate to the zone's size.
2530 zone->pages_min = tmp;
2533 zone->pages_low = zone->pages_min + tmp / 4;
2534 zone->pages_high = zone->pages_min + tmp / 2;
2535 spin_unlock_irqrestore(&zone->lru_lock, flags);
2540 * Initialise min_free_kbytes.
2542 * For small machines we want it small (128k min). For large machines
2543 * we want it large (64MB max). But it is not linear, because network
2544 * bandwidth does not increase linearly with machine size. We use
2546 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2547 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2563 static int __init init_per_zone_pages_min(void)
2565 unsigned long lowmem_kbytes;
2567 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2569 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2570 if (min_free_kbytes < 128)
2571 min_free_kbytes = 128;
2572 if (min_free_kbytes > 65536)
2573 min_free_kbytes = 65536;
2574 setup_per_zone_pages_min();
2575 setup_per_zone_lowmem_reserve();
2578 module_init(init_per_zone_pages_min)
2581 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2582 * that we can call two helper functions whenever min_free_kbytes
2585 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2586 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2588 proc_dointvec(table, write, file, buffer, length, ppos);
2589 setup_per_zone_pages_min();
2594 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2595 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2596 * whenever sysctl_lowmem_reserve_ratio changes.
2598 * The reserve ratio obviously has absolutely no relation with the
2599 * pages_min watermarks. The lowmem reserve ratio can only make sense
2600 * if in function of the boot time zone sizes.
2602 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
2603 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2605 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2606 setup_per_zone_lowmem_reserve();
2611 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
2612 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
2613 * can have before it gets flushed back to buddy allocator.
2616 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
2617 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2623 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2624 if (!write || (ret == -EINVAL))
2626 for_each_zone(zone) {
2627 for_each_online_cpu(cpu) {
2629 high = zone->present_pages / percpu_pagelist_fraction;
2630 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
2636 __initdata int hashdist = HASHDIST_DEFAULT;
2639 static int __init set_hashdist(char *str)
2643 hashdist = simple_strtoul(str, &str, 0);
2646 __setup("hashdist=", set_hashdist);
2650 * allocate a large system hash table from bootmem
2651 * - it is assumed that the hash table must contain an exact power-of-2
2652 * quantity of entries
2653 * - limit is the number of hash buckets, not the total allocation size
2655 void *__init alloc_large_system_hash(const char *tablename,
2656 unsigned long bucketsize,
2657 unsigned long numentries,
2660 unsigned int *_hash_shift,
2661 unsigned int *_hash_mask,
2662 unsigned long limit)
2664 unsigned long long max = limit;
2665 unsigned long log2qty, size;
2668 /* allow the kernel cmdline to have a say */
2670 /* round applicable memory size up to nearest megabyte */
2671 numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages;
2672 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2673 numentries >>= 20 - PAGE_SHIFT;
2674 numentries <<= 20 - PAGE_SHIFT;
2676 /* limit to 1 bucket per 2^scale bytes of low memory */
2677 if (scale > PAGE_SHIFT)
2678 numentries >>= (scale - PAGE_SHIFT);
2680 numentries <<= (PAGE_SHIFT - scale);
2682 /* rounded up to nearest power of 2 in size */
2683 numentries = 1UL << (long_log2(numentries) + 1);
2685 /* limit allocation size to 1/16 total memory by default */
2687 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2688 do_div(max, bucketsize);
2691 if (numentries > max)
2694 log2qty = long_log2(numentries);
2697 size = bucketsize << log2qty;
2698 if (flags & HASH_EARLY)
2699 table = alloc_bootmem(size);
2701 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
2703 unsigned long order;
2704 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
2706 table = (void*) __get_free_pages(GFP_ATOMIC, order);
2708 } while (!table && size > PAGE_SIZE && --log2qty);
2711 panic("Failed to allocate %s hash table\n", tablename);
2713 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2716 long_log2(size) - PAGE_SHIFT,
2720 *_hash_shift = log2qty;
2722 *_hash_mask = (1 << log2qty) - 1;