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);
215 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
219 BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
221 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
222 * and __GFP_HIGHMEM from hard or soft interrupt context.
224 BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
225 for (i = 0; i < (1 << order); i++)
226 clear_highpage(page + i);
230 * function for dealing with page's order in buddy system.
231 * zone->lock is already acquired when we use these.
232 * So, we don't need atomic page->flags operations here.
234 static inline unsigned long page_order(struct page *page) {
235 return page_private(page);
238 static inline void set_page_order(struct page *page, int order) {
239 set_page_private(page, order);
240 __SetPagePrivate(page);
243 static inline void rmv_page_order(struct page *page)
245 __ClearPagePrivate(page);
246 set_page_private(page, 0);
250 * Locate the struct page for both the matching buddy in our
251 * pair (buddy1) and the combined O(n+1) page they form (page).
253 * 1) Any buddy B1 will have an order O twin B2 which satisfies
254 * the following equation:
256 * For example, if the starting buddy (buddy2) is #8 its order
258 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
260 * 2) Any buddy B will have an order O+1 parent P which
261 * satisfies the following equation:
264 * Assumption: *_mem_map is contigious at least up to MAX_ORDER
266 static inline struct page *
267 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
269 unsigned long buddy_idx = page_idx ^ (1 << order);
271 return page + (buddy_idx - page_idx);
274 static inline unsigned long
275 __find_combined_index(unsigned long page_idx, unsigned int order)
277 return (page_idx & ~(1 << order));
281 * This function checks whether a page is free && is the buddy
282 * we can do coalesce a page and its buddy if
283 * (a) the buddy is not in a hole &&
284 * (b) the buddy is free &&
285 * (c) the buddy is on the buddy system &&
286 * (d) a page and its buddy have the same order.
287 * for recording page's order, we use page_private(page) and PG_private.
290 static inline int page_is_buddy(struct page *page, int order)
292 #ifdef CONFIG_HOLES_IN_ZONE
293 if (!pfn_valid(page_to_pfn(page)))
297 if (PagePrivate(page) &&
298 (page_order(page) == order) &&
299 page_count(page) == 0)
305 * Freeing function for a buddy system allocator.
307 * The concept of a buddy system is to maintain direct-mapped table
308 * (containing bit values) for memory blocks of various "orders".
309 * The bottom level table contains the map for the smallest allocatable
310 * units of memory (here, pages), and each level above it describes
311 * pairs of units from the levels below, hence, "buddies".
312 * At a high level, all that happens here is marking the table entry
313 * at the bottom level available, and propagating the changes upward
314 * as necessary, plus some accounting needed to play nicely with other
315 * parts of the VM system.
316 * At each level, we keep a list of pages, which are heads of continuous
317 * free pages of length of (1 << order) and marked with PG_Private.Page's
318 * order is recorded in page_private(page) field.
319 * So when we are allocating or freeing one, we can derive the state of the
320 * other. That is, if we allocate a small block, and both were
321 * free, the remainder of the region must be split into blocks.
322 * If a block is freed, and its buddy is also free, then this
323 * triggers coalescing into a block of larger size.
328 static inline void __free_one_page(struct page *page,
329 struct zone *zone, unsigned int order)
331 unsigned long page_idx;
332 int order_size = 1 << order;
334 if (unlikely(PageCompound(page)))
335 destroy_compound_page(page, order);
337 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
339 BUG_ON(page_idx & (order_size - 1));
340 BUG_ON(bad_range(zone, page));
342 zone->free_pages += order_size;
343 while (order < MAX_ORDER-1) {
344 unsigned long combined_idx;
345 struct free_area *area;
348 buddy = __page_find_buddy(page, page_idx, order);
349 if (!page_is_buddy(buddy, order))
350 break; /* Move the buddy up one level. */
352 list_del(&buddy->lru);
353 area = zone->free_area + order;
355 rmv_page_order(buddy);
356 combined_idx = __find_combined_index(page_idx, order);
357 page = page + (combined_idx - page_idx);
358 page_idx = combined_idx;
361 set_page_order(page, order);
362 list_add(&page->lru, &zone->free_area[order].free_list);
363 zone->free_area[order].nr_free++;
366 static inline int free_pages_check(struct page *page)
368 if (unlikely(page_mapcount(page) |
369 (page->mapping != NULL) |
370 (page_count(page) != 0) |
380 1 << PG_reserved ))))
383 __ClearPageDirty(page);
385 * For now, we report if PG_reserved was found set, but do not
386 * clear it, and do not free the page. But we shall soon need
387 * to do more, for when the ZERO_PAGE count wraps negative.
389 return PageReserved(page);
393 * Frees a list of pages.
394 * Assumes all pages on list are in same zone, and of same order.
395 * count is the number of pages to free.
397 * If the zone was previously in an "all pages pinned" state then look to
398 * see if this freeing clears that state.
400 * And clear the zone's pages_scanned counter, to hold off the "all pages are
401 * pinned" detection logic.
403 static void free_pages_bulk(struct zone *zone, int count,
404 struct list_head *list, int order)
406 spin_lock(&zone->lock);
407 zone->all_unreclaimable = 0;
408 zone->pages_scanned = 0;
412 BUG_ON(list_empty(list));
413 page = list_entry(list->prev, struct page, lru);
414 /* have to delete it as __free_one_page list manipulates */
415 list_del(&page->lru);
416 __free_one_page(page, zone, order);
418 spin_unlock(&zone->lock);
421 static void free_one_page(struct zone *zone, struct page *page, int order)
424 list_add(&page->lru, &list);
425 free_pages_bulk(zone, 1, &list, order);
428 static void __free_pages_ok(struct page *page, unsigned int order)
434 arch_free_page(page, order);
435 if (!PageHighMem(page))
436 mutex_debug_check_no_locks_freed(page_address(page),
439 for (i = 0 ; i < (1 << order) ; ++i)
440 reserved += free_pages_check(page + i);
444 kernel_map_pages(page, 1 << order, 0);
445 local_irq_save(flags);
446 __mod_page_state(pgfree, 1 << order);
447 free_one_page(page_zone(page), page, order);
448 local_irq_restore(flags);
452 * permit the bootmem allocator to evade page validation on high-order frees
454 void fastcall __init __free_pages_bootmem(struct page *page, unsigned int order)
457 __ClearPageReserved(page);
458 set_page_count(page, 0);
459 set_page_refcounted(page);
465 for (loop = 0; loop < BITS_PER_LONG; loop++) {
466 struct page *p = &page[loop];
468 if (loop + 1 < BITS_PER_LONG)
470 __ClearPageReserved(p);
471 set_page_count(p, 0);
474 set_page_refcounted(page);
475 __free_pages(page, order);
481 * The order of subdivision here is critical for the IO subsystem.
482 * Please do not alter this order without good reasons and regression
483 * testing. Specifically, as large blocks of memory are subdivided,
484 * the order in which smaller blocks are delivered depends on the order
485 * they're subdivided in this function. This is the primary factor
486 * influencing the order in which pages are delivered to the IO
487 * subsystem according to empirical testing, and this is also justified
488 * by considering the behavior of a buddy system containing a single
489 * large block of memory acted on by a series of small allocations.
490 * This behavior is a critical factor in sglist merging's success.
494 static inline void expand(struct zone *zone, struct page *page,
495 int low, int high, struct free_area *area)
497 unsigned long size = 1 << high;
503 BUG_ON(bad_range(zone, &page[size]));
504 list_add(&page[size].lru, &area->free_list);
506 set_page_order(&page[size], high);
511 * This page is about to be returned from the page allocator
513 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
515 if (unlikely(page_mapcount(page) |
516 (page->mapping != NULL) |
517 (page_count(page) != 0) |
528 1 << PG_reserved ))))
532 * For now, we report if PG_reserved was found set, but do not
533 * clear it, and do not allocate the page: as a safety net.
535 if (PageReserved(page))
538 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
539 1 << PG_referenced | 1 << PG_arch_1 |
540 1 << PG_checked | 1 << PG_mappedtodisk);
541 set_page_private(page, 0);
542 set_page_refcounted(page);
543 kernel_map_pages(page, 1 << order, 1);
545 if (gfp_flags & __GFP_ZERO)
546 prep_zero_page(page, order, gfp_flags);
548 if (order && (gfp_flags & __GFP_COMP))
549 prep_compound_page(page, order);
555 * Do the hard work of removing an element from the buddy allocator.
556 * Call me with the zone->lock already held.
558 static struct page *__rmqueue(struct zone *zone, unsigned int order)
560 struct free_area * area;
561 unsigned int current_order;
564 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
565 area = zone->free_area + current_order;
566 if (list_empty(&area->free_list))
569 page = list_entry(area->free_list.next, struct page, lru);
570 list_del(&page->lru);
571 rmv_page_order(page);
573 zone->free_pages -= 1UL << order;
574 expand(zone, page, order, current_order, area);
582 * Obtain a specified number of elements from the buddy allocator, all under
583 * a single hold of the lock, for efficiency. Add them to the supplied list.
584 * Returns the number of new pages which were placed at *list.
586 static int rmqueue_bulk(struct zone *zone, unsigned int order,
587 unsigned long count, struct list_head *list)
591 spin_lock(&zone->lock);
592 for (i = 0; i < count; ++i) {
593 struct page *page = __rmqueue(zone, order);
594 if (unlikely(page == NULL))
596 list_add_tail(&page->lru, list);
598 spin_unlock(&zone->lock);
604 * Called from the slab reaper to drain pagesets on a particular node that
605 * belong to the currently executing processor.
607 void drain_node_pages(int nodeid)
612 local_irq_save(flags);
613 for (z = 0; z < MAX_NR_ZONES; z++) {
614 struct zone *zone = NODE_DATA(nodeid)->node_zones + z;
615 struct per_cpu_pageset *pset;
617 pset = zone_pcp(zone, smp_processor_id());
618 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
619 struct per_cpu_pages *pcp;
622 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
626 local_irq_restore(flags);
630 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
631 static void __drain_pages(unsigned int cpu)
637 for_each_zone(zone) {
638 struct per_cpu_pageset *pset;
640 pset = zone_pcp(zone, cpu);
641 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
642 struct per_cpu_pages *pcp;
645 local_irq_save(flags);
646 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
648 local_irq_restore(flags);
652 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
656 void mark_free_pages(struct zone *zone)
658 unsigned long zone_pfn, flags;
660 struct list_head *curr;
662 if (!zone->spanned_pages)
665 spin_lock_irqsave(&zone->lock, flags);
666 for (zone_pfn = 0; zone_pfn < zone->spanned_pages; ++zone_pfn)
667 ClearPageNosaveFree(pfn_to_page(zone_pfn + zone->zone_start_pfn));
669 for (order = MAX_ORDER - 1; order >= 0; --order)
670 list_for_each(curr, &zone->free_area[order].free_list) {
671 unsigned long start_pfn, i;
673 start_pfn = page_to_pfn(list_entry(curr, struct page, lru));
675 for (i=0; i < (1<<order); i++)
676 SetPageNosaveFree(pfn_to_page(start_pfn+i));
678 spin_unlock_irqrestore(&zone->lock, flags);
682 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
684 void drain_local_pages(void)
688 local_irq_save(flags);
689 __drain_pages(smp_processor_id());
690 local_irq_restore(flags);
692 #endif /* CONFIG_PM */
694 static void zone_statistics(struct zonelist *zonelist, struct zone *z, int cpu)
697 pg_data_t *pg = z->zone_pgdat;
698 pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
699 struct per_cpu_pageset *p;
701 p = zone_pcp(z, cpu);
706 zone_pcp(zonelist->zones[0], cpu)->numa_foreign++;
708 if (pg == NODE_DATA(numa_node_id()))
716 * Free a 0-order page
718 static void fastcall free_hot_cold_page(struct page *page, int cold)
720 struct zone *zone = page_zone(page);
721 struct per_cpu_pages *pcp;
724 arch_free_page(page, 0);
727 page->mapping = NULL;
728 if (free_pages_check(page))
731 kernel_map_pages(page, 1, 0);
733 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
734 local_irq_save(flags);
735 __inc_page_state(pgfree);
736 list_add(&page->lru, &pcp->list);
738 if (pcp->count >= pcp->high) {
739 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
740 pcp->count -= pcp->batch;
742 local_irq_restore(flags);
746 void fastcall free_hot_page(struct page *page)
748 free_hot_cold_page(page, 0);
751 void fastcall free_cold_page(struct page *page)
753 free_hot_cold_page(page, 1);
757 * split_page takes a non-compound higher-order page, and splits it into
758 * n (1<<order) sub-pages: page[0..n]
759 * Each sub-page must be freed individually.
761 * Note: this is probably too low level an operation for use in drivers.
762 * Please consult with lkml before using this in your driver.
764 void split_page(struct page *page, unsigned int order)
768 BUG_ON(PageCompound(page));
769 BUG_ON(!page_count(page));
770 for (i = 1; i < (1 << order); i++)
771 set_page_refcounted(page + i);
775 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
776 * we cheat by calling it from here, in the order > 0 path. Saves a branch
779 static struct page *buffered_rmqueue(struct zonelist *zonelist,
780 struct zone *zone, int order, gfp_t gfp_flags)
784 int cold = !!(gfp_flags & __GFP_COLD);
789 if (likely(order == 0)) {
790 struct per_cpu_pages *pcp;
792 pcp = &zone_pcp(zone, cpu)->pcp[cold];
793 local_irq_save(flags);
795 pcp->count += rmqueue_bulk(zone, 0,
796 pcp->batch, &pcp->list);
797 if (unlikely(!pcp->count))
800 page = list_entry(pcp->list.next, struct page, lru);
801 list_del(&page->lru);
804 spin_lock_irqsave(&zone->lock, flags);
805 page = __rmqueue(zone, order);
806 spin_unlock(&zone->lock);
811 __mod_page_state_zone(zone, pgalloc, 1 << order);
812 zone_statistics(zonelist, zone, cpu);
813 local_irq_restore(flags);
816 BUG_ON(bad_range(zone, page));
817 if (prep_new_page(page, order, gfp_flags))
822 local_irq_restore(flags);
827 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
828 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
829 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
830 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
831 #define ALLOC_HARDER 0x10 /* try to alloc harder */
832 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
833 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
836 * Return 1 if free pages are above 'mark'. This takes into account the order
839 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
840 int classzone_idx, int alloc_flags)
842 /* free_pages my go negative - that's OK */
843 long min = mark, free_pages = z->free_pages - (1 << order) + 1;
846 if (alloc_flags & ALLOC_HIGH)
848 if (alloc_flags & ALLOC_HARDER)
851 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
853 for (o = 0; o < order; o++) {
854 /* At the next order, this order's pages become unavailable */
855 free_pages -= z->free_area[o].nr_free << o;
857 /* Require fewer higher order pages to be free */
860 if (free_pages <= min)
867 * get_page_from_freeliest goes through the zonelist trying to allocate
871 get_page_from_freelist(gfp_t gfp_mask, unsigned int order,
872 struct zonelist *zonelist, int alloc_flags)
874 struct zone **z = zonelist->zones;
875 struct page *page = NULL;
876 int classzone_idx = zone_idx(*z);
879 * Go through the zonelist once, looking for a zone with enough free.
880 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
883 if ((alloc_flags & ALLOC_CPUSET) &&
884 !cpuset_zone_allowed(*z, gfp_mask))
887 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
889 if (alloc_flags & ALLOC_WMARK_MIN)
890 mark = (*z)->pages_min;
891 else if (alloc_flags & ALLOC_WMARK_LOW)
892 mark = (*z)->pages_low;
894 mark = (*z)->pages_high;
895 if (!zone_watermark_ok(*z, order, mark,
896 classzone_idx, alloc_flags))
897 if (!zone_reclaim_mode ||
898 !zone_reclaim(*z, gfp_mask, order))
902 page = buffered_rmqueue(zonelist, *z, order, gfp_mask);
906 } while (*(++z) != NULL);
911 * This is the 'heart' of the zoned buddy allocator.
913 struct page * fastcall
914 __alloc_pages(gfp_t gfp_mask, unsigned int order,
915 struct zonelist *zonelist)
917 const gfp_t wait = gfp_mask & __GFP_WAIT;
920 struct reclaim_state reclaim_state;
921 struct task_struct *p = current;
924 int did_some_progress;
926 might_sleep_if(wait);
929 z = zonelist->zones; /* the list of zones suitable for gfp_mask */
931 if (unlikely(*z == NULL)) {
932 /* Should this ever happen?? */
936 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
937 zonelist, ALLOC_WMARK_LOW|ALLOC_CPUSET);
942 wakeup_kswapd(*z, order);
946 * OK, we're below the kswapd watermark and have kicked background
947 * reclaim. Now things get more complex, so set up alloc_flags according
948 * to how we want to proceed.
950 * The caller may dip into page reserves a bit more if the caller
951 * cannot run direct reclaim, or if the caller has realtime scheduling
952 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
953 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
955 alloc_flags = ALLOC_WMARK_MIN;
956 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
957 alloc_flags |= ALLOC_HARDER;
958 if (gfp_mask & __GFP_HIGH)
959 alloc_flags |= ALLOC_HIGH;
960 alloc_flags |= ALLOC_CPUSET;
963 * Go through the zonelist again. Let __GFP_HIGH and allocations
964 * coming from realtime tasks go deeper into reserves.
966 * This is the last chance, in general, before the goto nopage.
967 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
968 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
970 page = get_page_from_freelist(gfp_mask, order, zonelist, alloc_flags);
974 /* This allocation should allow future memory freeing. */
976 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
977 && !in_interrupt()) {
978 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
980 /* go through the zonelist yet again, ignoring mins */
981 page = get_page_from_freelist(gfp_mask, order,
982 zonelist, ALLOC_NO_WATERMARKS);
985 if (gfp_mask & __GFP_NOFAIL) {
986 blk_congestion_wait(WRITE, HZ/50);
993 /* Atomic allocations - we can't balance anything */
1000 /* We now go into synchronous reclaim */
1001 cpuset_memory_pressure_bump();
1002 p->flags |= PF_MEMALLOC;
1003 reclaim_state.reclaimed_slab = 0;
1004 p->reclaim_state = &reclaim_state;
1006 did_some_progress = try_to_free_pages(zonelist->zones, gfp_mask);
1008 p->reclaim_state = NULL;
1009 p->flags &= ~PF_MEMALLOC;
1013 if (likely(did_some_progress)) {
1014 page = get_page_from_freelist(gfp_mask, order,
1015 zonelist, alloc_flags);
1018 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1020 * Go through the zonelist yet one more time, keep
1021 * very high watermark here, this is only to catch
1022 * a parallel oom killing, we must fail if we're still
1023 * under heavy pressure.
1025 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1026 zonelist, ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1030 out_of_memory(zonelist, gfp_mask, order);
1035 * Don't let big-order allocations loop unless the caller explicitly
1036 * requests that. Wait for some write requests to complete then retry.
1038 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
1039 * <= 3, but that may not be true in other implementations.
1042 if (!(gfp_mask & __GFP_NORETRY)) {
1043 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
1045 if (gfp_mask & __GFP_NOFAIL)
1049 blk_congestion_wait(WRITE, HZ/50);
1054 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1055 printk(KERN_WARNING "%s: page allocation failure."
1056 " order:%d, mode:0x%x\n",
1057 p->comm, order, gfp_mask);
1065 EXPORT_SYMBOL(__alloc_pages);
1068 * Common helper functions.
1070 fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1073 page = alloc_pages(gfp_mask, order);
1076 return (unsigned long) page_address(page);
1079 EXPORT_SYMBOL(__get_free_pages);
1081 fastcall unsigned long get_zeroed_page(gfp_t gfp_mask)
1086 * get_zeroed_page() returns a 32-bit address, which cannot represent
1089 BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1091 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1093 return (unsigned long) page_address(page);
1097 EXPORT_SYMBOL(get_zeroed_page);
1099 void __pagevec_free(struct pagevec *pvec)
1101 int i = pagevec_count(pvec);
1104 free_hot_cold_page(pvec->pages[i], pvec->cold);
1107 fastcall void __free_pages(struct page *page, unsigned int order)
1109 if (put_page_testzero(page)) {
1111 free_hot_page(page);
1113 __free_pages_ok(page, order);
1117 EXPORT_SYMBOL(__free_pages);
1119 fastcall void free_pages(unsigned long addr, unsigned int order)
1122 BUG_ON(!virt_addr_valid((void *)addr));
1123 __free_pages(virt_to_page((void *)addr), order);
1127 EXPORT_SYMBOL(free_pages);
1130 * Total amount of free (allocatable) RAM:
1132 unsigned int nr_free_pages(void)
1134 unsigned int sum = 0;
1138 sum += zone->free_pages;
1143 EXPORT_SYMBOL(nr_free_pages);
1146 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
1148 unsigned int i, sum = 0;
1150 for (i = 0; i < MAX_NR_ZONES; i++)
1151 sum += pgdat->node_zones[i].free_pages;
1157 static unsigned int nr_free_zone_pages(int offset)
1159 /* Just pick one node, since fallback list is circular */
1160 pg_data_t *pgdat = NODE_DATA(numa_node_id());
1161 unsigned int sum = 0;
1163 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1164 struct zone **zonep = zonelist->zones;
1167 for (zone = *zonep++; zone; zone = *zonep++) {
1168 unsigned long size = zone->present_pages;
1169 unsigned long high = zone->pages_high;
1178 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1180 unsigned int nr_free_buffer_pages(void)
1182 return nr_free_zone_pages(gfp_zone(GFP_USER));
1186 * Amount of free RAM allocatable within all zones
1188 unsigned int nr_free_pagecache_pages(void)
1190 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER));
1193 #ifdef CONFIG_HIGHMEM
1194 unsigned int nr_free_highpages (void)
1197 unsigned int pages = 0;
1199 for_each_pgdat(pgdat)
1200 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1207 static void show_node(struct zone *zone)
1209 printk("Node %d ", zone->zone_pgdat->node_id);
1212 #define show_node(zone) do { } while (0)
1216 * Accumulate the page_state information across all CPUs.
1217 * The result is unavoidably approximate - it can change
1218 * during and after execution of this function.
1220 static DEFINE_PER_CPU(struct page_state, page_states) = {0};
1222 atomic_t nr_pagecache = ATOMIC_INIT(0);
1223 EXPORT_SYMBOL(nr_pagecache);
1225 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
1228 static void __get_page_state(struct page_state *ret, int nr, cpumask_t *cpumask)
1232 memset(ret, 0, nr * sizeof(unsigned long));
1233 cpus_and(*cpumask, *cpumask, cpu_online_map);
1235 for_each_cpu_mask(cpu, *cpumask) {
1241 in = (unsigned long *)&per_cpu(page_states, cpu);
1243 next_cpu = next_cpu(cpu, *cpumask);
1244 if (likely(next_cpu < NR_CPUS))
1245 prefetch(&per_cpu(page_states, next_cpu));
1247 out = (unsigned long *)ret;
1248 for (off = 0; off < nr; off++)
1253 void get_page_state_node(struct page_state *ret, int node)
1256 cpumask_t mask = node_to_cpumask(node);
1258 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
1259 nr /= sizeof(unsigned long);
1261 __get_page_state(ret, nr+1, &mask);
1264 void get_page_state(struct page_state *ret)
1267 cpumask_t mask = CPU_MASK_ALL;
1269 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
1270 nr /= sizeof(unsigned long);
1272 __get_page_state(ret, nr + 1, &mask);
1275 void get_full_page_state(struct page_state *ret)
1277 cpumask_t mask = CPU_MASK_ALL;
1279 __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long), &mask);
1282 unsigned long read_page_state_offset(unsigned long offset)
1284 unsigned long ret = 0;
1287 for_each_online_cpu(cpu) {
1290 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
1291 ret += *((unsigned long *)in);
1296 void __mod_page_state_offset(unsigned long offset, unsigned long delta)
1300 ptr = &__get_cpu_var(page_states);
1301 *(unsigned long *)(ptr + offset) += delta;
1303 EXPORT_SYMBOL(__mod_page_state_offset);
1305 void mod_page_state_offset(unsigned long offset, unsigned long delta)
1307 unsigned long flags;
1310 local_irq_save(flags);
1311 ptr = &__get_cpu_var(page_states);
1312 *(unsigned long *)(ptr + offset) += delta;
1313 local_irq_restore(flags);
1315 EXPORT_SYMBOL(mod_page_state_offset);
1317 void __get_zone_counts(unsigned long *active, unsigned long *inactive,
1318 unsigned long *free, struct pglist_data *pgdat)
1320 struct zone *zones = pgdat->node_zones;
1326 for (i = 0; i < MAX_NR_ZONES; i++) {
1327 *active += zones[i].nr_active;
1328 *inactive += zones[i].nr_inactive;
1329 *free += zones[i].free_pages;
1333 void get_zone_counts(unsigned long *active,
1334 unsigned long *inactive, unsigned long *free)
1336 struct pglist_data *pgdat;
1341 for_each_pgdat(pgdat) {
1342 unsigned long l, m, n;
1343 __get_zone_counts(&l, &m, &n, pgdat);
1350 void si_meminfo(struct sysinfo *val)
1352 val->totalram = totalram_pages;
1354 val->freeram = nr_free_pages();
1355 val->bufferram = nr_blockdev_pages();
1356 #ifdef CONFIG_HIGHMEM
1357 val->totalhigh = totalhigh_pages;
1358 val->freehigh = nr_free_highpages();
1363 val->mem_unit = PAGE_SIZE;
1366 EXPORT_SYMBOL(si_meminfo);
1369 void si_meminfo_node(struct sysinfo *val, int nid)
1371 pg_data_t *pgdat = NODE_DATA(nid);
1373 val->totalram = pgdat->node_present_pages;
1374 val->freeram = nr_free_pages_pgdat(pgdat);
1375 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1376 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1377 val->mem_unit = PAGE_SIZE;
1381 #define K(x) ((x) << (PAGE_SHIFT-10))
1384 * Show free area list (used inside shift_scroll-lock stuff)
1385 * We also calculate the percentage fragmentation. We do this by counting the
1386 * memory on each free list with the exception of the first item on the list.
1388 void show_free_areas(void)
1390 struct page_state ps;
1391 int cpu, temperature;
1392 unsigned long active;
1393 unsigned long inactive;
1397 for_each_zone(zone) {
1399 printk("%s per-cpu:", zone->name);
1401 if (!populated_zone(zone)) {
1407 for_each_online_cpu(cpu) {
1408 struct per_cpu_pageset *pageset;
1410 pageset = zone_pcp(zone, cpu);
1412 for (temperature = 0; temperature < 2; temperature++)
1413 printk("cpu %d %s: high %d, batch %d used:%d\n",
1415 temperature ? "cold" : "hot",
1416 pageset->pcp[temperature].high,
1417 pageset->pcp[temperature].batch,
1418 pageset->pcp[temperature].count);
1422 get_page_state(&ps);
1423 get_zone_counts(&active, &inactive, &free);
1425 printk("Free pages: %11ukB (%ukB HighMem)\n",
1427 K(nr_free_highpages()));
1429 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1430 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1439 ps.nr_page_table_pages);
1441 for_each_zone(zone) {
1453 " pages_scanned:%lu"
1454 " all_unreclaimable? %s"
1457 K(zone->free_pages),
1460 K(zone->pages_high),
1462 K(zone->nr_inactive),
1463 K(zone->present_pages),
1464 zone->pages_scanned,
1465 (zone->all_unreclaimable ? "yes" : "no")
1467 printk("lowmem_reserve[]:");
1468 for (i = 0; i < MAX_NR_ZONES; i++)
1469 printk(" %lu", zone->lowmem_reserve[i]);
1473 for_each_zone(zone) {
1474 unsigned long nr, flags, order, total = 0;
1477 printk("%s: ", zone->name);
1478 if (!populated_zone(zone)) {
1483 spin_lock_irqsave(&zone->lock, flags);
1484 for (order = 0; order < MAX_ORDER; order++) {
1485 nr = zone->free_area[order].nr_free;
1486 total += nr << order;
1487 printk("%lu*%lukB ", nr, K(1UL) << order);
1489 spin_unlock_irqrestore(&zone->lock, flags);
1490 printk("= %lukB\n", K(total));
1493 show_swap_cache_info();
1497 * Builds allocation fallback zone lists.
1499 * Add all populated zones of a node to the zonelist.
1501 static int __init build_zonelists_node(pg_data_t *pgdat,
1502 struct zonelist *zonelist, int nr_zones, int zone_type)
1506 BUG_ON(zone_type > ZONE_HIGHMEM);
1509 zone = pgdat->node_zones + zone_type;
1510 if (populated_zone(zone)) {
1511 #ifndef CONFIG_HIGHMEM
1512 BUG_ON(zone_type > ZONE_NORMAL);
1514 zonelist->zones[nr_zones++] = zone;
1515 check_highest_zone(zone_type);
1519 } while (zone_type >= 0);
1523 static inline int highest_zone(int zone_bits)
1525 int res = ZONE_NORMAL;
1526 if (zone_bits & (__force int)__GFP_HIGHMEM)
1528 if (zone_bits & (__force int)__GFP_DMA32)
1530 if (zone_bits & (__force int)__GFP_DMA)
1536 #define MAX_NODE_LOAD (num_online_nodes())
1537 static int __initdata node_load[MAX_NUMNODES];
1539 * find_next_best_node - find the next node that should appear in a given node's fallback list
1540 * @node: node whose fallback list we're appending
1541 * @used_node_mask: nodemask_t of already used nodes
1543 * We use a number of factors to determine which is the next node that should
1544 * appear on a given node's fallback list. The node should not have appeared
1545 * already in @node's fallback list, and it should be the next closest node
1546 * according to the distance array (which contains arbitrary distance values
1547 * from each node to each node in the system), and should also prefer nodes
1548 * with no CPUs, since presumably they'll have very little allocation pressure
1549 * on them otherwise.
1550 * It returns -1 if no node is found.
1552 static int __init find_next_best_node(int node, nodemask_t *used_node_mask)
1555 int min_val = INT_MAX;
1558 /* Use the local node if we haven't already */
1559 if (!node_isset(node, *used_node_mask)) {
1560 node_set(node, *used_node_mask);
1564 for_each_online_node(n) {
1567 /* Don't want a node to appear more than once */
1568 if (node_isset(n, *used_node_mask))
1571 /* Use the distance array to find the distance */
1572 val = node_distance(node, n);
1574 /* Penalize nodes under us ("prefer the next node") */
1577 /* Give preference to headless and unused nodes */
1578 tmp = node_to_cpumask(n);
1579 if (!cpus_empty(tmp))
1580 val += PENALTY_FOR_NODE_WITH_CPUS;
1582 /* Slight preference for less loaded node */
1583 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1584 val += node_load[n];
1586 if (val < min_val) {
1593 node_set(best_node, *used_node_mask);
1598 static void __init build_zonelists(pg_data_t *pgdat)
1600 int i, j, k, node, local_node;
1601 int prev_node, load;
1602 struct zonelist *zonelist;
1603 nodemask_t used_mask;
1605 /* initialize zonelists */
1606 for (i = 0; i < GFP_ZONETYPES; i++) {
1607 zonelist = pgdat->node_zonelists + i;
1608 zonelist->zones[0] = NULL;
1611 /* NUMA-aware ordering of nodes */
1612 local_node = pgdat->node_id;
1613 load = num_online_nodes();
1614 prev_node = local_node;
1615 nodes_clear(used_mask);
1616 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1617 int distance = node_distance(local_node, node);
1620 * If another node is sufficiently far away then it is better
1621 * to reclaim pages in a zone before going off node.
1623 if (distance > RECLAIM_DISTANCE)
1624 zone_reclaim_mode = 1;
1627 * We don't want to pressure a particular node.
1628 * So adding penalty to the first node in same
1629 * distance group to make it round-robin.
1632 if (distance != node_distance(local_node, prev_node))
1633 node_load[node] += load;
1636 for (i = 0; i < GFP_ZONETYPES; i++) {
1637 zonelist = pgdat->node_zonelists + i;
1638 for (j = 0; zonelist->zones[j] != NULL; j++);
1640 k = highest_zone(i);
1642 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1643 zonelist->zones[j] = NULL;
1648 #else /* CONFIG_NUMA */
1650 static void __init build_zonelists(pg_data_t *pgdat)
1652 int i, j, k, node, local_node;
1654 local_node = pgdat->node_id;
1655 for (i = 0; i < GFP_ZONETYPES; i++) {
1656 struct zonelist *zonelist;
1658 zonelist = pgdat->node_zonelists + i;
1661 k = highest_zone(i);
1662 j = build_zonelists_node(pgdat, zonelist, j, k);
1664 * Now we build the zonelist so that it contains the zones
1665 * of all the other nodes.
1666 * We don't want to pressure a particular node, so when
1667 * building the zones for node N, we make sure that the
1668 * zones coming right after the local ones are those from
1669 * node N+1 (modulo N)
1671 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1672 if (!node_online(node))
1674 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1676 for (node = 0; node < local_node; node++) {
1677 if (!node_online(node))
1679 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1682 zonelist->zones[j] = NULL;
1686 #endif /* CONFIG_NUMA */
1688 void __init build_all_zonelists(void)
1692 for_each_online_node(i)
1693 build_zonelists(NODE_DATA(i));
1694 printk("Built %i zonelists\n", num_online_nodes());
1695 cpuset_init_current_mems_allowed();
1699 * Helper functions to size the waitqueue hash table.
1700 * Essentially these want to choose hash table sizes sufficiently
1701 * large so that collisions trying to wait on pages are rare.
1702 * But in fact, the number of active page waitqueues on typical
1703 * systems is ridiculously low, less than 200. So this is even
1704 * conservative, even though it seems large.
1706 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1707 * waitqueues, i.e. the size of the waitq table given the number of pages.
1709 #define PAGES_PER_WAITQUEUE 256
1711 static inline unsigned long wait_table_size(unsigned long pages)
1713 unsigned long size = 1;
1715 pages /= PAGES_PER_WAITQUEUE;
1717 while (size < pages)
1721 * Once we have dozens or even hundreds of threads sleeping
1722 * on IO we've got bigger problems than wait queue collision.
1723 * Limit the size of the wait table to a reasonable size.
1725 size = min(size, 4096UL);
1727 return max(size, 4UL);
1731 * This is an integer logarithm so that shifts can be used later
1732 * to extract the more random high bits from the multiplicative
1733 * hash function before the remainder is taken.
1735 static inline unsigned long wait_table_bits(unsigned long size)
1740 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1742 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1743 unsigned long *zones_size, unsigned long *zholes_size)
1745 unsigned long realtotalpages, totalpages = 0;
1748 for (i = 0; i < MAX_NR_ZONES; i++)
1749 totalpages += zones_size[i];
1750 pgdat->node_spanned_pages = totalpages;
1752 realtotalpages = totalpages;
1754 for (i = 0; i < MAX_NR_ZONES; i++)
1755 realtotalpages -= zholes_size[i];
1756 pgdat->node_present_pages = realtotalpages;
1757 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1762 * Initially all pages are reserved - free ones are freed
1763 * up by free_all_bootmem() once the early boot process is
1764 * done. Non-atomic initialization, single-pass.
1766 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1767 unsigned long start_pfn)
1770 unsigned long end_pfn = start_pfn + size;
1773 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1774 if (!early_pfn_valid(pfn))
1776 page = pfn_to_page(pfn);
1777 set_page_links(page, zone, nid, pfn);
1778 init_page_count(page);
1779 reset_page_mapcount(page);
1780 SetPageReserved(page);
1781 INIT_LIST_HEAD(&page->lru);
1782 #ifdef WANT_PAGE_VIRTUAL
1783 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1784 if (!is_highmem_idx(zone))
1785 set_page_address(page, __va(pfn << PAGE_SHIFT));
1790 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1794 for (order = 0; order < MAX_ORDER ; order++) {
1795 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1796 zone->free_area[order].nr_free = 0;
1800 #define ZONETABLE_INDEX(x, zone_nr) ((x << ZONES_SHIFT) | zone_nr)
1801 void zonetable_add(struct zone *zone, int nid, int zid, unsigned long pfn,
1804 unsigned long snum = pfn_to_section_nr(pfn);
1805 unsigned long end = pfn_to_section_nr(pfn + size);
1808 zone_table[ZONETABLE_INDEX(nid, zid)] = zone;
1810 for (; snum <= end; snum++)
1811 zone_table[ZONETABLE_INDEX(snum, zid)] = zone;
1814 #ifndef __HAVE_ARCH_MEMMAP_INIT
1815 #define memmap_init(size, nid, zone, start_pfn) \
1816 memmap_init_zone((size), (nid), (zone), (start_pfn))
1819 static int __cpuinit zone_batchsize(struct zone *zone)
1824 * The per-cpu-pages pools are set to around 1000th of the
1825 * size of the zone. But no more than 1/2 of a meg.
1827 * OK, so we don't know how big the cache is. So guess.
1829 batch = zone->present_pages / 1024;
1830 if (batch * PAGE_SIZE > 512 * 1024)
1831 batch = (512 * 1024) / PAGE_SIZE;
1832 batch /= 4; /* We effectively *= 4 below */
1837 * Clamp the batch to a 2^n - 1 value. Having a power
1838 * of 2 value was found to be more likely to have
1839 * suboptimal cache aliasing properties in some cases.
1841 * For example if 2 tasks are alternately allocating
1842 * batches of pages, one task can end up with a lot
1843 * of pages of one half of the possible page colors
1844 * and the other with pages of the other colors.
1846 batch = (1 << (fls(batch + batch/2)-1)) - 1;
1851 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
1853 struct per_cpu_pages *pcp;
1855 memset(p, 0, sizeof(*p));
1857 pcp = &p->pcp[0]; /* hot */
1859 pcp->high = 6 * batch;
1860 pcp->batch = max(1UL, 1 * batch);
1861 INIT_LIST_HEAD(&pcp->list);
1863 pcp = &p->pcp[1]; /* cold*/
1865 pcp->high = 2 * batch;
1866 pcp->batch = max(1UL, batch/2);
1867 INIT_LIST_HEAD(&pcp->list);
1871 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
1872 * to the value high for the pageset p.
1875 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
1878 struct per_cpu_pages *pcp;
1880 pcp = &p->pcp[0]; /* hot list */
1882 pcp->batch = max(1UL, high/4);
1883 if ((high/4) > (PAGE_SHIFT * 8))
1884 pcp->batch = PAGE_SHIFT * 8;
1890 * Boot pageset table. One per cpu which is going to be used for all
1891 * zones and all nodes. The parameters will be set in such a way
1892 * that an item put on a list will immediately be handed over to
1893 * the buddy list. This is safe since pageset manipulation is done
1894 * with interrupts disabled.
1896 * Some NUMA counter updates may also be caught by the boot pagesets.
1898 * The boot_pagesets must be kept even after bootup is complete for
1899 * unused processors and/or zones. They do play a role for bootstrapping
1900 * hotplugged processors.
1902 * zoneinfo_show() and maybe other functions do
1903 * not check if the processor is online before following the pageset pointer.
1904 * Other parts of the kernel may not check if the zone is available.
1906 static struct per_cpu_pageset boot_pageset[NR_CPUS];
1909 * Dynamically allocate memory for the
1910 * per cpu pageset array in struct zone.
1912 static int __cpuinit process_zones(int cpu)
1914 struct zone *zone, *dzone;
1916 for_each_zone(zone) {
1918 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
1919 GFP_KERNEL, cpu_to_node(cpu));
1920 if (!zone_pcp(zone, cpu))
1923 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
1925 if (percpu_pagelist_fraction)
1926 setup_pagelist_highmark(zone_pcp(zone, cpu),
1927 (zone->present_pages / percpu_pagelist_fraction));
1932 for_each_zone(dzone) {
1935 kfree(zone_pcp(dzone, cpu));
1936 zone_pcp(dzone, cpu) = NULL;
1941 static inline void free_zone_pagesets(int cpu)
1945 for_each_zone(zone) {
1946 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
1948 zone_pcp(zone, cpu) = NULL;
1953 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
1954 unsigned long action,
1957 int cpu = (long)hcpu;
1958 int ret = NOTIFY_OK;
1961 case CPU_UP_PREPARE:
1962 if (process_zones(cpu))
1965 case CPU_UP_CANCELED:
1967 free_zone_pagesets(cpu);
1975 static struct notifier_block pageset_notifier =
1976 { &pageset_cpuup_callback, NULL, 0 };
1978 void __init setup_per_cpu_pageset(void)
1982 /* Initialize per_cpu_pageset for cpu 0.
1983 * A cpuup callback will do this for every cpu
1984 * as it comes online
1986 err = process_zones(smp_processor_id());
1988 register_cpu_notifier(&pageset_notifier);
1994 void zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
1997 struct pglist_data *pgdat = zone->zone_pgdat;
2000 * The per-page waitqueue mechanism uses hashed waitqueues
2003 zone->wait_table_size = wait_table_size(zone_size_pages);
2004 zone->wait_table_bits = wait_table_bits(zone->wait_table_size);
2005 zone->wait_table = (wait_queue_head_t *)
2006 alloc_bootmem_node(pgdat, zone->wait_table_size
2007 * sizeof(wait_queue_head_t));
2009 for(i = 0; i < zone->wait_table_size; ++i)
2010 init_waitqueue_head(zone->wait_table + i);
2013 static __meminit void zone_pcp_init(struct zone *zone)
2016 unsigned long batch = zone_batchsize(zone);
2018 for (cpu = 0; cpu < NR_CPUS; cpu++) {
2020 /* Early boot. Slab allocator not functional yet */
2021 zone_pcp(zone, cpu) = &boot_pageset[cpu];
2022 setup_pageset(&boot_pageset[cpu],0);
2024 setup_pageset(zone_pcp(zone,cpu), batch);
2027 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
2028 zone->name, zone->present_pages, batch);
2031 static __meminit void init_currently_empty_zone(struct zone *zone,
2032 unsigned long zone_start_pfn, unsigned long size)
2034 struct pglist_data *pgdat = zone->zone_pgdat;
2036 zone_wait_table_init(zone, size);
2037 pgdat->nr_zones = zone_idx(zone) + 1;
2039 zone->zone_mem_map = pfn_to_page(zone_start_pfn);
2040 zone->zone_start_pfn = zone_start_pfn;
2042 memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn);
2044 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
2048 * Set up the zone data structures:
2049 * - mark all pages reserved
2050 * - mark all memory queues empty
2051 * - clear the memory bitmaps
2053 static void __init free_area_init_core(struct pglist_data *pgdat,
2054 unsigned long *zones_size, unsigned long *zholes_size)
2057 int nid = pgdat->node_id;
2058 unsigned long zone_start_pfn = pgdat->node_start_pfn;
2060 pgdat_resize_init(pgdat);
2061 pgdat->nr_zones = 0;
2062 init_waitqueue_head(&pgdat->kswapd_wait);
2063 pgdat->kswapd_max_order = 0;
2065 for (j = 0; j < MAX_NR_ZONES; j++) {
2066 struct zone *zone = pgdat->node_zones + j;
2067 unsigned long size, realsize;
2069 realsize = size = zones_size[j];
2071 realsize -= zholes_size[j];
2073 if (j < ZONE_HIGHMEM)
2074 nr_kernel_pages += realsize;
2075 nr_all_pages += realsize;
2077 zone->spanned_pages = size;
2078 zone->present_pages = realsize;
2079 zone->name = zone_names[j];
2080 spin_lock_init(&zone->lock);
2081 spin_lock_init(&zone->lru_lock);
2082 zone_seqlock_init(zone);
2083 zone->zone_pgdat = pgdat;
2084 zone->free_pages = 0;
2086 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
2088 zone_pcp_init(zone);
2089 INIT_LIST_HEAD(&zone->active_list);
2090 INIT_LIST_HEAD(&zone->inactive_list);
2091 zone->nr_scan_active = 0;
2092 zone->nr_scan_inactive = 0;
2093 zone->nr_active = 0;
2094 zone->nr_inactive = 0;
2095 atomic_set(&zone->reclaim_in_progress, 0);
2099 zonetable_add(zone, nid, j, zone_start_pfn, size);
2100 init_currently_empty_zone(zone, zone_start_pfn, size);
2101 zone_start_pfn += size;
2105 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
2107 /* Skip empty nodes */
2108 if (!pgdat->node_spanned_pages)
2111 #ifdef CONFIG_FLAT_NODE_MEM_MAP
2112 /* ia64 gets its own node_mem_map, before this, without bootmem */
2113 if (!pgdat->node_mem_map) {
2117 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
2118 map = alloc_remap(pgdat->node_id, size);
2120 map = alloc_bootmem_node(pgdat, size);
2121 pgdat->node_mem_map = map;
2123 #ifdef CONFIG_FLATMEM
2125 * With no DISCONTIG, the global mem_map is just set as node 0's
2127 if (pgdat == NODE_DATA(0))
2128 mem_map = NODE_DATA(0)->node_mem_map;
2130 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
2133 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
2134 unsigned long *zones_size, unsigned long node_start_pfn,
2135 unsigned long *zholes_size)
2137 pgdat->node_id = nid;
2138 pgdat->node_start_pfn = node_start_pfn;
2139 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
2141 alloc_node_mem_map(pgdat);
2143 free_area_init_core(pgdat, zones_size, zholes_size);
2146 #ifndef CONFIG_NEED_MULTIPLE_NODES
2147 static bootmem_data_t contig_bootmem_data;
2148 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
2150 EXPORT_SYMBOL(contig_page_data);
2153 void __init free_area_init(unsigned long *zones_size)
2155 free_area_init_node(0, NODE_DATA(0), zones_size,
2156 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
2159 #ifdef CONFIG_PROC_FS
2161 #include <linux/seq_file.h>
2163 static void *frag_start(struct seq_file *m, loff_t *pos)
2168 for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
2174 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
2176 pg_data_t *pgdat = (pg_data_t *)arg;
2179 return pgdat->pgdat_next;
2182 static void frag_stop(struct seq_file *m, void *arg)
2187 * This walks the free areas for each zone.
2189 static int frag_show(struct seq_file *m, void *arg)
2191 pg_data_t *pgdat = (pg_data_t *)arg;
2193 struct zone *node_zones = pgdat->node_zones;
2194 unsigned long flags;
2197 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2198 if (!populated_zone(zone))
2201 spin_lock_irqsave(&zone->lock, flags);
2202 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
2203 for (order = 0; order < MAX_ORDER; ++order)
2204 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
2205 spin_unlock_irqrestore(&zone->lock, flags);
2211 struct seq_operations fragmentation_op = {
2212 .start = frag_start,
2219 * Output information about zones in @pgdat.
2221 static int zoneinfo_show(struct seq_file *m, void *arg)
2223 pg_data_t *pgdat = arg;
2225 struct zone *node_zones = pgdat->node_zones;
2226 unsigned long flags;
2228 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; zone++) {
2231 if (!populated_zone(zone))
2234 spin_lock_irqsave(&zone->lock, flags);
2235 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
2243 "\n scanned %lu (a: %lu i: %lu)"
2252 zone->pages_scanned,
2253 zone->nr_scan_active, zone->nr_scan_inactive,
2254 zone->spanned_pages,
2255 zone->present_pages);
2257 "\n protection: (%lu",
2258 zone->lowmem_reserve[0]);
2259 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
2260 seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
2264 for_each_online_cpu(i) {
2265 struct per_cpu_pageset *pageset;
2268 pageset = zone_pcp(zone, i);
2269 for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2270 if (pageset->pcp[j].count)
2273 if (j == ARRAY_SIZE(pageset->pcp))
2275 for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2277 "\n cpu: %i pcp: %i"
2282 pageset->pcp[j].count,
2283 pageset->pcp[j].high,
2284 pageset->pcp[j].batch);
2290 "\n numa_foreign: %lu"
2291 "\n interleave_hit: %lu"
2292 "\n local_node: %lu"
2293 "\n other_node: %lu",
2296 pageset->numa_foreign,
2297 pageset->interleave_hit,
2298 pageset->local_node,
2299 pageset->other_node);
2303 "\n all_unreclaimable: %u"
2304 "\n prev_priority: %i"
2305 "\n temp_priority: %i"
2306 "\n start_pfn: %lu",
2307 zone->all_unreclaimable,
2308 zone->prev_priority,
2309 zone->temp_priority,
2310 zone->zone_start_pfn);
2311 spin_unlock_irqrestore(&zone->lock, flags);
2317 struct seq_operations zoneinfo_op = {
2318 .start = frag_start, /* iterate over all zones. The same as in
2322 .show = zoneinfo_show,
2325 static char *vmstat_text[] = {
2329 "nr_page_table_pages",
2360 "pgscan_kswapd_high",
2361 "pgscan_kswapd_normal",
2362 "pgscan_kswapd_dma32",
2363 "pgscan_kswapd_dma",
2365 "pgscan_direct_high",
2366 "pgscan_direct_normal",
2367 "pgscan_direct_dma32",
2368 "pgscan_direct_dma",
2373 "kswapd_inodesteal",
2381 static void *vmstat_start(struct seq_file *m, loff_t *pos)
2383 struct page_state *ps;
2385 if (*pos >= ARRAY_SIZE(vmstat_text))
2388 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
2391 return ERR_PTR(-ENOMEM);
2392 get_full_page_state(ps);
2393 ps->pgpgin /= 2; /* sectors -> kbytes */
2395 return (unsigned long *)ps + *pos;
2398 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
2401 if (*pos >= ARRAY_SIZE(vmstat_text))
2403 return (unsigned long *)m->private + *pos;
2406 static int vmstat_show(struct seq_file *m, void *arg)
2408 unsigned long *l = arg;
2409 unsigned long off = l - (unsigned long *)m->private;
2411 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
2415 static void vmstat_stop(struct seq_file *m, void *arg)
2421 struct seq_operations vmstat_op = {
2422 .start = vmstat_start,
2423 .next = vmstat_next,
2424 .stop = vmstat_stop,
2425 .show = vmstat_show,
2428 #endif /* CONFIG_PROC_FS */
2430 #ifdef CONFIG_HOTPLUG_CPU
2431 static int page_alloc_cpu_notify(struct notifier_block *self,
2432 unsigned long action, void *hcpu)
2434 int cpu = (unsigned long)hcpu;
2436 unsigned long *src, *dest;
2438 if (action == CPU_DEAD) {
2441 /* Drain local pagecache count. */
2442 count = &per_cpu(nr_pagecache_local, cpu);
2443 atomic_add(*count, &nr_pagecache);
2445 local_irq_disable();
2448 /* Add dead cpu's page_states to our own. */
2449 dest = (unsigned long *)&__get_cpu_var(page_states);
2450 src = (unsigned long *)&per_cpu(page_states, cpu);
2452 for (i = 0; i < sizeof(struct page_state)/sizeof(unsigned long);
2462 #endif /* CONFIG_HOTPLUG_CPU */
2464 void __init page_alloc_init(void)
2466 hotcpu_notifier(page_alloc_cpu_notify, 0);
2470 * setup_per_zone_lowmem_reserve - called whenever
2471 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
2472 * has a correct pages reserved value, so an adequate number of
2473 * pages are left in the zone after a successful __alloc_pages().
2475 static void setup_per_zone_lowmem_reserve(void)
2477 struct pglist_data *pgdat;
2480 for_each_pgdat(pgdat) {
2481 for (j = 0; j < MAX_NR_ZONES; j++) {
2482 struct zone *zone = pgdat->node_zones + j;
2483 unsigned long present_pages = zone->present_pages;
2485 zone->lowmem_reserve[j] = 0;
2487 for (idx = j-1; idx >= 0; idx--) {
2488 struct zone *lower_zone;
2490 if (sysctl_lowmem_reserve_ratio[idx] < 1)
2491 sysctl_lowmem_reserve_ratio[idx] = 1;
2493 lower_zone = pgdat->node_zones + idx;
2494 lower_zone->lowmem_reserve[j] = present_pages /
2495 sysctl_lowmem_reserve_ratio[idx];
2496 present_pages += lower_zone->present_pages;
2503 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
2504 * that the pages_{min,low,high} values for each zone are set correctly
2505 * with respect to min_free_kbytes.
2507 void setup_per_zone_pages_min(void)
2509 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
2510 unsigned long lowmem_pages = 0;
2512 unsigned long flags;
2514 /* Calculate total number of !ZONE_HIGHMEM pages */
2515 for_each_zone(zone) {
2516 if (!is_highmem(zone))
2517 lowmem_pages += zone->present_pages;
2520 for_each_zone(zone) {
2522 spin_lock_irqsave(&zone->lru_lock, flags);
2523 tmp = (pages_min * zone->present_pages) / lowmem_pages;
2524 if (is_highmem(zone)) {
2526 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
2527 * need highmem pages, so cap pages_min to a small
2530 * The (pages_high-pages_low) and (pages_low-pages_min)
2531 * deltas controls asynch page reclaim, and so should
2532 * not be capped for highmem.
2536 min_pages = zone->present_pages / 1024;
2537 if (min_pages < SWAP_CLUSTER_MAX)
2538 min_pages = SWAP_CLUSTER_MAX;
2539 if (min_pages > 128)
2541 zone->pages_min = min_pages;
2544 * If it's a lowmem zone, reserve a number of pages
2545 * proportionate to the zone's size.
2547 zone->pages_min = tmp;
2550 zone->pages_low = zone->pages_min + tmp / 4;
2551 zone->pages_high = zone->pages_min + tmp / 2;
2552 spin_unlock_irqrestore(&zone->lru_lock, flags);
2557 * Initialise min_free_kbytes.
2559 * For small machines we want it small (128k min). For large machines
2560 * we want it large (64MB max). But it is not linear, because network
2561 * bandwidth does not increase linearly with machine size. We use
2563 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2564 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2580 static int __init init_per_zone_pages_min(void)
2582 unsigned long lowmem_kbytes;
2584 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2586 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2587 if (min_free_kbytes < 128)
2588 min_free_kbytes = 128;
2589 if (min_free_kbytes > 65536)
2590 min_free_kbytes = 65536;
2591 setup_per_zone_pages_min();
2592 setup_per_zone_lowmem_reserve();
2595 module_init(init_per_zone_pages_min)
2598 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2599 * that we can call two helper functions whenever min_free_kbytes
2602 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2603 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2605 proc_dointvec(table, write, file, buffer, length, ppos);
2606 setup_per_zone_pages_min();
2611 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2612 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2613 * whenever sysctl_lowmem_reserve_ratio changes.
2615 * The reserve ratio obviously has absolutely no relation with the
2616 * pages_min watermarks. The lowmem reserve ratio can only make sense
2617 * if in function of the boot time zone sizes.
2619 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
2620 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2622 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2623 setup_per_zone_lowmem_reserve();
2628 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
2629 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
2630 * can have before it gets flushed back to buddy allocator.
2633 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
2634 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2640 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2641 if (!write || (ret == -EINVAL))
2643 for_each_zone(zone) {
2644 for_each_online_cpu(cpu) {
2646 high = zone->present_pages / percpu_pagelist_fraction;
2647 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
2653 __initdata int hashdist = HASHDIST_DEFAULT;
2656 static int __init set_hashdist(char *str)
2660 hashdist = simple_strtoul(str, &str, 0);
2663 __setup("hashdist=", set_hashdist);
2667 * allocate a large system hash table from bootmem
2668 * - it is assumed that the hash table must contain an exact power-of-2
2669 * quantity of entries
2670 * - limit is the number of hash buckets, not the total allocation size
2672 void *__init alloc_large_system_hash(const char *tablename,
2673 unsigned long bucketsize,
2674 unsigned long numentries,
2677 unsigned int *_hash_shift,
2678 unsigned int *_hash_mask,
2679 unsigned long limit)
2681 unsigned long long max = limit;
2682 unsigned long log2qty, size;
2685 /* allow the kernel cmdline to have a say */
2687 /* round applicable memory size up to nearest megabyte */
2688 numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages;
2689 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2690 numentries >>= 20 - PAGE_SHIFT;
2691 numentries <<= 20 - PAGE_SHIFT;
2693 /* limit to 1 bucket per 2^scale bytes of low memory */
2694 if (scale > PAGE_SHIFT)
2695 numentries >>= (scale - PAGE_SHIFT);
2697 numentries <<= (PAGE_SHIFT - scale);
2699 /* rounded up to nearest power of 2 in size */
2700 numentries = 1UL << (long_log2(numentries) + 1);
2702 /* limit allocation size to 1/16 total memory by default */
2704 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2705 do_div(max, bucketsize);
2708 if (numentries > max)
2711 log2qty = long_log2(numentries);
2714 size = bucketsize << log2qty;
2715 if (flags & HASH_EARLY)
2716 table = alloc_bootmem(size);
2718 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
2720 unsigned long order;
2721 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
2723 table = (void*) __get_free_pages(GFP_ATOMIC, order);
2725 } while (!table && size > PAGE_SIZE && --log2qty);
2728 panic("Failed to allocate %s hash table\n", tablename);
2730 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2733 long_log2(size) - PAGE_SHIFT,
2737 *_hash_shift = log2qty;
2739 *_hash_mask = (1 << log2qty) - 1;