2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/bootmem.h>
23 #include <linux/compiler.h>
24 #include <linux/kernel.h>
25 #include <linux/module.h>
26 #include <linux/suspend.h>
27 #include <linux/pagevec.h>
28 #include <linux/blkdev.h>
29 #include <linux/slab.h>
30 #include <linux/notifier.h>
31 #include <linux/topology.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/memory_hotplug.h>
36 #include <linux/nodemask.h>
37 #include <linux/vmalloc.h>
38 #include <linux/mempolicy.h>
39 #include <linux/stop_machine.h>
41 #include <asm/tlbflush.h>
42 #include <asm/div64.h>
46 * MCD - HACK: Find somewhere to initialize this EARLY, or make this
49 nodemask_t node_online_map __read_mostly = { { [0] = 1UL } };
50 EXPORT_SYMBOL(node_online_map);
51 nodemask_t node_possible_map __read_mostly = NODE_MASK_ALL;
52 EXPORT_SYMBOL(node_possible_map);
53 unsigned long totalram_pages __read_mostly;
54 unsigned long totalreserve_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] = {
73 #ifdef CONFIG_ZONE_DMA32
81 EXPORT_SYMBOL(totalram_pages);
84 * Used by page_zone() to look up the address of the struct zone whose
85 * id is encoded in the upper bits of page->flags
87 struct zone *zone_table[1 << ZONETABLE_SHIFT] __read_mostly;
88 EXPORT_SYMBOL(zone_table);
90 static char *zone_names[MAX_NR_ZONES] = {
92 #ifdef CONFIG_ZONE_DMA32
101 int min_free_kbytes = 1024;
103 unsigned long __meminitdata nr_kernel_pages;
104 unsigned long __meminitdata nr_all_pages;
106 #ifdef CONFIG_DEBUG_VM
107 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
111 unsigned long pfn = page_to_pfn(page);
114 seq = zone_span_seqbegin(zone);
115 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
117 else if (pfn < zone->zone_start_pfn)
119 } while (zone_span_seqretry(zone, seq));
124 static int page_is_consistent(struct zone *zone, struct page *page)
126 #ifdef CONFIG_HOLES_IN_ZONE
127 if (!pfn_valid(page_to_pfn(page)))
130 if (zone != page_zone(page))
136 * Temporary debugging check for pages not lying within a given zone.
138 static int bad_range(struct zone *zone, struct page *page)
140 if (page_outside_zone_boundaries(zone, page))
142 if (!page_is_consistent(zone, page))
148 static inline int bad_range(struct zone *zone, struct page *page)
154 static void bad_page(struct page *page)
156 printk(KERN_EMERG "Bad page state in process '%s'\n"
157 KERN_EMERG "page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d\n"
158 KERN_EMERG "Trying to fix it up, but a reboot is needed\n"
159 KERN_EMERG "Backtrace:\n",
160 current->comm, page, (int)(2*sizeof(unsigned long)),
161 (unsigned long)page->flags, page->mapping,
162 page_mapcount(page), page_count(page));
164 page->flags &= ~(1 << PG_lru |
174 set_page_count(page, 0);
175 reset_page_mapcount(page);
176 page->mapping = NULL;
177 add_taint(TAINT_BAD_PAGE);
181 * Higher-order pages are called "compound pages". They are structured thusly:
183 * The first PAGE_SIZE page is called the "head page".
185 * The remaining PAGE_SIZE pages are called "tail pages".
187 * All pages have PG_compound set. All pages have their ->private pointing at
188 * the head page (even the head page has this).
190 * The first tail page's ->lru.next holds the address of the compound page's
191 * put_page() function. Its ->lru.prev holds the order of allocation.
192 * This usage means that zero-order pages may not be compound.
195 static void free_compound_page(struct page *page)
197 __free_pages_ok(page, (unsigned long)page[1].lru.prev);
200 static void prep_compound_page(struct page *page, unsigned long order)
203 int nr_pages = 1 << order;
205 page[1].lru.next = (void *)free_compound_page; /* set dtor */
206 page[1].lru.prev = (void *)order;
207 for (i = 0; i < nr_pages; i++) {
208 struct page *p = page + i;
210 __SetPageCompound(p);
211 set_page_private(p, (unsigned long)page);
215 static void destroy_compound_page(struct page *page, unsigned long order)
218 int nr_pages = 1 << order;
220 if (unlikely((unsigned long)page[1].lru.prev != order))
223 for (i = 0; i < nr_pages; i++) {
224 struct page *p = page + i;
226 if (unlikely(!PageCompound(p) |
227 (page_private(p) != (unsigned long)page)))
229 __ClearPageCompound(p);
233 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
237 VM_BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
239 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
240 * and __GFP_HIGHMEM from hard or soft interrupt context.
242 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
243 for (i = 0; i < (1 << order); i++)
244 clear_highpage(page + i);
248 * function for dealing with page's order in buddy system.
249 * zone->lock is already acquired when we use these.
250 * So, we don't need atomic page->flags operations here.
252 static inline unsigned long page_order(struct page *page)
254 return page_private(page);
257 static inline void set_page_order(struct page *page, int order)
259 set_page_private(page, order);
260 __SetPageBuddy(page);
263 static inline void rmv_page_order(struct page *page)
265 __ClearPageBuddy(page);
266 set_page_private(page, 0);
270 * Locate the struct page for both the matching buddy in our
271 * pair (buddy1) and the combined O(n+1) page they form (page).
273 * 1) Any buddy B1 will have an order O twin B2 which satisfies
274 * the following equation:
276 * For example, if the starting buddy (buddy2) is #8 its order
278 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
280 * 2) Any buddy B will have an order O+1 parent P which
281 * satisfies the following equation:
284 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
286 static inline struct page *
287 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
289 unsigned long buddy_idx = page_idx ^ (1 << order);
291 return page + (buddy_idx - page_idx);
294 static inline unsigned long
295 __find_combined_index(unsigned long page_idx, unsigned int order)
297 return (page_idx & ~(1 << order));
301 * This function checks whether a page is free && is the buddy
302 * we can do coalesce a page and its buddy if
303 * (a) the buddy is not in a hole &&
304 * (b) the buddy is in the buddy system &&
305 * (c) a page and its buddy have the same order &&
306 * (d) a page and its buddy are in the same zone.
308 * For recording whether a page is in the buddy system, we use PG_buddy.
309 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
311 * For recording page's order, we use page_private(page).
313 static inline int page_is_buddy(struct page *page, struct page *buddy,
316 #ifdef CONFIG_HOLES_IN_ZONE
317 if (!pfn_valid(page_to_pfn(buddy)))
321 if (page_zone_id(page) != page_zone_id(buddy))
324 if (PageBuddy(buddy) && page_order(buddy) == order) {
325 BUG_ON(page_count(buddy) != 0);
332 * Freeing function for a buddy system allocator.
334 * The concept of a buddy system is to maintain direct-mapped table
335 * (containing bit values) for memory blocks of various "orders".
336 * The bottom level table contains the map for the smallest allocatable
337 * units of memory (here, pages), and each level above it describes
338 * pairs of units from the levels below, hence, "buddies".
339 * At a high level, all that happens here is marking the table entry
340 * at the bottom level available, and propagating the changes upward
341 * as necessary, plus some accounting needed to play nicely with other
342 * parts of the VM system.
343 * At each level, we keep a list of pages, which are heads of continuous
344 * free pages of length of (1 << order) and marked with PG_buddy. Page's
345 * order is recorded in page_private(page) field.
346 * So when we are allocating or freeing one, we can derive the state of the
347 * other. That is, if we allocate a small block, and both were
348 * free, the remainder of the region must be split into blocks.
349 * If a block is freed, and its buddy is also free, then this
350 * triggers coalescing into a block of larger size.
355 static inline void __free_one_page(struct page *page,
356 struct zone *zone, unsigned int order)
358 unsigned long page_idx;
359 int order_size = 1 << order;
361 if (unlikely(PageCompound(page)))
362 destroy_compound_page(page, order);
364 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
366 VM_BUG_ON(page_idx & (order_size - 1));
367 VM_BUG_ON(bad_range(zone, page));
369 zone->free_pages += order_size;
370 while (order < MAX_ORDER-1) {
371 unsigned long combined_idx;
372 struct free_area *area;
375 buddy = __page_find_buddy(page, page_idx, order);
376 if (!page_is_buddy(page, buddy, order))
377 break; /* Move the buddy up one level. */
379 list_del(&buddy->lru);
380 area = zone->free_area + order;
382 rmv_page_order(buddy);
383 combined_idx = __find_combined_index(page_idx, order);
384 page = page + (combined_idx - page_idx);
385 page_idx = combined_idx;
388 set_page_order(page, order);
389 list_add(&page->lru, &zone->free_area[order].free_list);
390 zone->free_area[order].nr_free++;
393 static inline int free_pages_check(struct page *page)
395 if (unlikely(page_mapcount(page) |
396 (page->mapping != NULL) |
397 (page_count(page) != 0) |
411 __ClearPageDirty(page);
413 * For now, we report if PG_reserved was found set, but do not
414 * clear it, and do not free the page. But we shall soon need
415 * to do more, for when the ZERO_PAGE count wraps negative.
417 return PageReserved(page);
421 * Frees a list of pages.
422 * Assumes all pages on list are in same zone, and of same order.
423 * count is the number of pages to free.
425 * If the zone was previously in an "all pages pinned" state then look to
426 * see if this freeing clears that state.
428 * And clear the zone's pages_scanned counter, to hold off the "all pages are
429 * pinned" detection logic.
431 static void free_pages_bulk(struct zone *zone, int count,
432 struct list_head *list, int order)
434 spin_lock(&zone->lock);
435 zone->all_unreclaimable = 0;
436 zone->pages_scanned = 0;
440 VM_BUG_ON(list_empty(list));
441 page = list_entry(list->prev, struct page, lru);
442 /* have to delete it as __free_one_page list manipulates */
443 list_del(&page->lru);
444 __free_one_page(page, zone, order);
446 spin_unlock(&zone->lock);
449 static void free_one_page(struct zone *zone, struct page *page, int order)
451 spin_lock(&zone->lock);
452 zone->all_unreclaimable = 0;
453 zone->pages_scanned = 0;
454 __free_one_page(page, zone ,order);
455 spin_unlock(&zone->lock);
458 static void __free_pages_ok(struct page *page, unsigned int order)
464 arch_free_page(page, order);
465 if (!PageHighMem(page))
466 debug_check_no_locks_freed(page_address(page),
469 for (i = 0 ; i < (1 << order) ; ++i)
470 reserved += free_pages_check(page + i);
474 kernel_map_pages(page, 1 << order, 0);
475 local_irq_save(flags);
476 __count_vm_events(PGFREE, 1 << order);
477 free_one_page(page_zone(page), page, order);
478 local_irq_restore(flags);
482 * permit the bootmem allocator to evade page validation on high-order frees
484 void fastcall __init __free_pages_bootmem(struct page *page, unsigned int order)
487 __ClearPageReserved(page);
488 set_page_count(page, 0);
489 set_page_refcounted(page);
495 for (loop = 0; loop < BITS_PER_LONG; loop++) {
496 struct page *p = &page[loop];
498 if (loop + 1 < BITS_PER_LONG)
500 __ClearPageReserved(p);
501 set_page_count(p, 0);
504 set_page_refcounted(page);
505 __free_pages(page, order);
511 * The order of subdivision here is critical for the IO subsystem.
512 * Please do not alter this order without good reasons and regression
513 * testing. Specifically, as large blocks of memory are subdivided,
514 * the order in which smaller blocks are delivered depends on the order
515 * they're subdivided in this function. This is the primary factor
516 * influencing the order in which pages are delivered to the IO
517 * subsystem according to empirical testing, and this is also justified
518 * by considering the behavior of a buddy system containing a single
519 * large block of memory acted on by a series of small allocations.
520 * This behavior is a critical factor in sglist merging's success.
524 static inline void expand(struct zone *zone, struct page *page,
525 int low, int high, struct free_area *area)
527 unsigned long size = 1 << high;
533 VM_BUG_ON(bad_range(zone, &page[size]));
534 list_add(&page[size].lru, &area->free_list);
536 set_page_order(&page[size], high);
541 * This page is about to be returned from the page allocator
543 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
545 if (unlikely(page_mapcount(page) |
546 (page->mapping != NULL) |
547 (page_count(page) != 0) |
563 * For now, we report if PG_reserved was found set, but do not
564 * clear it, and do not allocate the page: as a safety net.
566 if (PageReserved(page))
569 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
570 1 << PG_referenced | 1 << PG_arch_1 |
571 1 << PG_checked | 1 << PG_mappedtodisk);
572 set_page_private(page, 0);
573 set_page_refcounted(page);
574 kernel_map_pages(page, 1 << order, 1);
576 if (gfp_flags & __GFP_ZERO)
577 prep_zero_page(page, order, gfp_flags);
579 if (order && (gfp_flags & __GFP_COMP))
580 prep_compound_page(page, order);
586 * Do the hard work of removing an element from the buddy allocator.
587 * Call me with the zone->lock already held.
589 static struct page *__rmqueue(struct zone *zone, unsigned int order)
591 struct free_area * area;
592 unsigned int current_order;
595 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
596 area = zone->free_area + current_order;
597 if (list_empty(&area->free_list))
600 page = list_entry(area->free_list.next, struct page, lru);
601 list_del(&page->lru);
602 rmv_page_order(page);
604 zone->free_pages -= 1UL << order;
605 expand(zone, page, order, current_order, area);
613 * Obtain a specified number of elements from the buddy allocator, all under
614 * a single hold of the lock, for efficiency. Add them to the supplied list.
615 * Returns the number of new pages which were placed at *list.
617 static int rmqueue_bulk(struct zone *zone, unsigned int order,
618 unsigned long count, struct list_head *list)
622 spin_lock(&zone->lock);
623 for (i = 0; i < count; ++i) {
624 struct page *page = __rmqueue(zone, order);
625 if (unlikely(page == NULL))
627 list_add_tail(&page->lru, list);
629 spin_unlock(&zone->lock);
635 * Called from the slab reaper to drain pagesets on a particular node that
636 * belongs to the currently executing processor.
637 * Note that this function must be called with the thread pinned to
638 * a single processor.
640 void drain_node_pages(int nodeid)
646 for (z = 0; z < MAX_NR_ZONES; z++) {
647 struct zone *zone = NODE_DATA(nodeid)->node_zones + z;
648 struct per_cpu_pageset *pset;
650 if (!populated_zone(zone))
653 pset = zone_pcp(zone, smp_processor_id());
654 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
655 struct per_cpu_pages *pcp;
659 local_irq_save(flags);
660 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
662 local_irq_restore(flags);
669 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
670 static void __drain_pages(unsigned int cpu)
676 for_each_zone(zone) {
677 struct per_cpu_pageset *pset;
679 pset = zone_pcp(zone, cpu);
680 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
681 struct per_cpu_pages *pcp;
684 local_irq_save(flags);
685 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
687 local_irq_restore(flags);
691 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
695 void mark_free_pages(struct zone *zone)
697 unsigned long pfn, max_zone_pfn;
700 struct list_head *curr;
702 if (!zone->spanned_pages)
705 spin_lock_irqsave(&zone->lock, flags);
707 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
708 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
709 if (pfn_valid(pfn)) {
710 struct page *page = pfn_to_page(pfn);
712 if (!PageNosave(page))
713 ClearPageNosaveFree(page);
716 for (order = MAX_ORDER - 1; order >= 0; --order)
717 list_for_each(curr, &zone->free_area[order].free_list) {
720 pfn = page_to_pfn(list_entry(curr, struct page, lru));
721 for (i = 0; i < (1UL << order); i++)
722 SetPageNosaveFree(pfn_to_page(pfn + i));
725 spin_unlock_irqrestore(&zone->lock, flags);
729 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
731 void drain_local_pages(void)
735 local_irq_save(flags);
736 __drain_pages(smp_processor_id());
737 local_irq_restore(flags);
739 #endif /* CONFIG_PM */
742 * Free a 0-order page
744 static void fastcall free_hot_cold_page(struct page *page, int cold)
746 struct zone *zone = page_zone(page);
747 struct per_cpu_pages *pcp;
750 arch_free_page(page, 0);
753 page->mapping = NULL;
754 if (free_pages_check(page))
757 kernel_map_pages(page, 1, 0);
759 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
760 local_irq_save(flags);
761 __count_vm_event(PGFREE);
762 list_add(&page->lru, &pcp->list);
764 if (pcp->count >= pcp->high) {
765 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
766 pcp->count -= pcp->batch;
768 local_irq_restore(flags);
772 void fastcall free_hot_page(struct page *page)
774 free_hot_cold_page(page, 0);
777 void fastcall free_cold_page(struct page *page)
779 free_hot_cold_page(page, 1);
783 * split_page takes a non-compound higher-order page, and splits it into
784 * n (1<<order) sub-pages: page[0..n]
785 * Each sub-page must be freed individually.
787 * Note: this is probably too low level an operation for use in drivers.
788 * Please consult with lkml before using this in your driver.
790 void split_page(struct page *page, unsigned int order)
794 VM_BUG_ON(PageCompound(page));
795 VM_BUG_ON(!page_count(page));
796 for (i = 1; i < (1 << order); i++)
797 set_page_refcounted(page + i);
801 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
802 * we cheat by calling it from here, in the order > 0 path. Saves a branch
805 static struct page *buffered_rmqueue(struct zonelist *zonelist,
806 struct zone *zone, int order, gfp_t gfp_flags)
810 int cold = !!(gfp_flags & __GFP_COLD);
815 if (likely(order == 0)) {
816 struct per_cpu_pages *pcp;
818 pcp = &zone_pcp(zone, cpu)->pcp[cold];
819 local_irq_save(flags);
821 pcp->count += rmqueue_bulk(zone, 0,
822 pcp->batch, &pcp->list);
823 if (unlikely(!pcp->count))
826 page = list_entry(pcp->list.next, struct page, lru);
827 list_del(&page->lru);
830 spin_lock_irqsave(&zone->lock, flags);
831 page = __rmqueue(zone, order);
832 spin_unlock(&zone->lock);
837 __count_zone_vm_events(PGALLOC, zone, 1 << order);
838 zone_statistics(zonelist, zone);
839 local_irq_restore(flags);
842 VM_BUG_ON(bad_range(zone, page));
843 if (prep_new_page(page, order, gfp_flags))
848 local_irq_restore(flags);
853 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
854 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
855 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
856 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
857 #define ALLOC_HARDER 0x10 /* try to alloc harder */
858 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
859 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
862 * Return 1 if free pages are above 'mark'. This takes into account the order
865 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
866 int classzone_idx, int alloc_flags)
868 /* free_pages my go negative - that's OK */
869 long min = mark, free_pages = z->free_pages - (1 << order) + 1;
872 if (alloc_flags & ALLOC_HIGH)
874 if (alloc_flags & ALLOC_HARDER)
877 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
879 for (o = 0; o < order; o++) {
880 /* At the next order, this order's pages become unavailable */
881 free_pages -= z->free_area[o].nr_free << o;
883 /* Require fewer higher order pages to be free */
886 if (free_pages <= min)
893 * get_page_from_freeliest goes through the zonelist trying to allocate
897 get_page_from_freelist(gfp_t gfp_mask, unsigned int order,
898 struct zonelist *zonelist, int alloc_flags)
900 struct zone **z = zonelist->zones;
901 struct page *page = NULL;
902 int classzone_idx = zone_idx(*z);
906 * Go through the zonelist once, looking for a zone with enough free.
907 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
911 if (unlikely((gfp_mask & __GFP_THISNODE) &&
912 zone->zone_pgdat != zonelist->zones[0]->zone_pgdat))
914 if ((alloc_flags & ALLOC_CPUSET) &&
915 !cpuset_zone_allowed(zone, gfp_mask))
918 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
920 if (alloc_flags & ALLOC_WMARK_MIN)
921 mark = zone->pages_min;
922 else if (alloc_flags & ALLOC_WMARK_LOW)
923 mark = zone->pages_low;
925 mark = zone->pages_high;
926 if (!zone_watermark_ok(zone , order, mark,
927 classzone_idx, alloc_flags))
928 if (!zone_reclaim_mode ||
929 !zone_reclaim(zone, gfp_mask, order))
933 page = buffered_rmqueue(zonelist, zone, order, gfp_mask);
937 } while (*(++z) != NULL);
942 * This is the 'heart' of the zoned buddy allocator.
944 struct page * fastcall
945 __alloc_pages(gfp_t gfp_mask, unsigned int order,
946 struct zonelist *zonelist)
948 const gfp_t wait = gfp_mask & __GFP_WAIT;
951 struct reclaim_state reclaim_state;
952 struct task_struct *p = current;
955 int did_some_progress;
957 might_sleep_if(wait);
960 z = zonelist->zones; /* the list of zones suitable for gfp_mask */
962 if (unlikely(*z == NULL)) {
963 /* Should this ever happen?? */
967 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
968 zonelist, ALLOC_WMARK_LOW|ALLOC_CPUSET);
973 wakeup_kswapd(*z, order);
977 * OK, we're below the kswapd watermark and have kicked background
978 * reclaim. Now things get more complex, so set up alloc_flags according
979 * to how we want to proceed.
981 * The caller may dip into page reserves a bit more if the caller
982 * cannot run direct reclaim, or if the caller has realtime scheduling
983 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
984 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
986 alloc_flags = ALLOC_WMARK_MIN;
987 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
988 alloc_flags |= ALLOC_HARDER;
989 if (gfp_mask & __GFP_HIGH)
990 alloc_flags |= ALLOC_HIGH;
992 alloc_flags |= ALLOC_CPUSET;
995 * Go through the zonelist again. Let __GFP_HIGH and allocations
996 * coming from realtime tasks go deeper into reserves.
998 * This is the last chance, in general, before the goto nopage.
999 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1000 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1002 page = get_page_from_freelist(gfp_mask, order, zonelist, alloc_flags);
1006 /* This allocation should allow future memory freeing. */
1008 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
1009 && !in_interrupt()) {
1010 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
1012 /* go through the zonelist yet again, ignoring mins */
1013 page = get_page_from_freelist(gfp_mask, order,
1014 zonelist, ALLOC_NO_WATERMARKS);
1017 if (gfp_mask & __GFP_NOFAIL) {
1018 blk_congestion_wait(WRITE, HZ/50);
1025 /* Atomic allocations - we can't balance anything */
1032 /* We now go into synchronous reclaim */
1033 cpuset_memory_pressure_bump();
1034 p->flags |= PF_MEMALLOC;
1035 reclaim_state.reclaimed_slab = 0;
1036 p->reclaim_state = &reclaim_state;
1038 did_some_progress = try_to_free_pages(zonelist->zones, gfp_mask);
1040 p->reclaim_state = NULL;
1041 p->flags &= ~PF_MEMALLOC;
1045 if (likely(did_some_progress)) {
1046 page = get_page_from_freelist(gfp_mask, order,
1047 zonelist, alloc_flags);
1050 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1052 * Go through the zonelist yet one more time, keep
1053 * very high watermark here, this is only to catch
1054 * a parallel oom killing, we must fail if we're still
1055 * under heavy pressure.
1057 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1058 zonelist, ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1062 out_of_memory(zonelist, gfp_mask, order);
1067 * Don't let big-order allocations loop unless the caller explicitly
1068 * requests that. Wait for some write requests to complete then retry.
1070 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
1071 * <= 3, but that may not be true in other implementations.
1074 if (!(gfp_mask & __GFP_NORETRY)) {
1075 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
1077 if (gfp_mask & __GFP_NOFAIL)
1081 blk_congestion_wait(WRITE, HZ/50);
1086 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1087 printk(KERN_WARNING "%s: page allocation failure."
1088 " order:%d, mode:0x%x\n",
1089 p->comm, order, gfp_mask);
1097 EXPORT_SYMBOL(__alloc_pages);
1100 * Common helper functions.
1102 fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1105 page = alloc_pages(gfp_mask, order);
1108 return (unsigned long) page_address(page);
1111 EXPORT_SYMBOL(__get_free_pages);
1113 fastcall unsigned long get_zeroed_page(gfp_t gfp_mask)
1118 * get_zeroed_page() returns a 32-bit address, which cannot represent
1121 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1123 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1125 return (unsigned long) page_address(page);
1129 EXPORT_SYMBOL(get_zeroed_page);
1131 void __pagevec_free(struct pagevec *pvec)
1133 int i = pagevec_count(pvec);
1136 free_hot_cold_page(pvec->pages[i], pvec->cold);
1139 fastcall void __free_pages(struct page *page, unsigned int order)
1141 if (put_page_testzero(page)) {
1143 free_hot_page(page);
1145 __free_pages_ok(page, order);
1149 EXPORT_SYMBOL(__free_pages);
1151 fastcall void free_pages(unsigned long addr, unsigned int order)
1154 VM_BUG_ON(!virt_addr_valid((void *)addr));
1155 __free_pages(virt_to_page((void *)addr), order);
1159 EXPORT_SYMBOL(free_pages);
1162 * Total amount of free (allocatable) RAM:
1164 unsigned int nr_free_pages(void)
1166 unsigned int sum = 0;
1170 sum += zone->free_pages;
1175 EXPORT_SYMBOL(nr_free_pages);
1178 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
1180 unsigned int sum = 0;
1183 for (i = 0; i < MAX_NR_ZONES; i++)
1184 sum += pgdat->node_zones[i].free_pages;
1190 static unsigned int nr_free_zone_pages(int offset)
1192 /* Just pick one node, since fallback list is circular */
1193 pg_data_t *pgdat = NODE_DATA(numa_node_id());
1194 unsigned int sum = 0;
1196 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1197 struct zone **zonep = zonelist->zones;
1200 for (zone = *zonep++; zone; zone = *zonep++) {
1201 unsigned long size = zone->present_pages;
1202 unsigned long high = zone->pages_high;
1211 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1213 unsigned int nr_free_buffer_pages(void)
1215 return nr_free_zone_pages(gfp_zone(GFP_USER));
1219 * Amount of free RAM allocatable within all zones
1221 unsigned int nr_free_pagecache_pages(void)
1223 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER));
1226 static void show_node(struct zone *zone)
1228 printk("Node %ld ", zone_to_nid(zone));
1231 #define show_node(zone) do { } while (0)
1234 void si_meminfo(struct sysinfo *val)
1236 val->totalram = totalram_pages;
1238 val->freeram = nr_free_pages();
1239 val->bufferram = nr_blockdev_pages();
1240 val->totalhigh = totalhigh_pages;
1241 val->freehigh = nr_free_highpages();
1242 val->mem_unit = PAGE_SIZE;
1245 EXPORT_SYMBOL(si_meminfo);
1248 void si_meminfo_node(struct sysinfo *val, int nid)
1250 pg_data_t *pgdat = NODE_DATA(nid);
1252 val->totalram = pgdat->node_present_pages;
1253 val->freeram = nr_free_pages_pgdat(pgdat);
1254 #ifdef CONFIG_HIGHMEM
1255 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1256 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1261 val->mem_unit = PAGE_SIZE;
1265 #define K(x) ((x) << (PAGE_SHIFT-10))
1268 * Show free area list (used inside shift_scroll-lock stuff)
1269 * We also calculate the percentage fragmentation. We do this by counting the
1270 * memory on each free list with the exception of the first item on the list.
1272 void show_free_areas(void)
1274 int cpu, temperature;
1275 unsigned long active;
1276 unsigned long inactive;
1280 for_each_zone(zone) {
1282 printk("%s per-cpu:", zone->name);
1284 if (!populated_zone(zone)) {
1290 for_each_online_cpu(cpu) {
1291 struct per_cpu_pageset *pageset;
1293 pageset = zone_pcp(zone, cpu);
1295 for (temperature = 0; temperature < 2; temperature++)
1296 printk("cpu %d %s: high %d, batch %d used:%d\n",
1298 temperature ? "cold" : "hot",
1299 pageset->pcp[temperature].high,
1300 pageset->pcp[temperature].batch,
1301 pageset->pcp[temperature].count);
1305 get_zone_counts(&active, &inactive, &free);
1307 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1308 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1311 global_page_state(NR_FILE_DIRTY),
1312 global_page_state(NR_WRITEBACK),
1313 global_page_state(NR_UNSTABLE_NFS),
1315 global_page_state(NR_SLAB_RECLAIMABLE) +
1316 global_page_state(NR_SLAB_UNRECLAIMABLE),
1317 global_page_state(NR_FILE_MAPPED),
1318 global_page_state(NR_PAGETABLE));
1320 for_each_zone(zone) {
1332 " pages_scanned:%lu"
1333 " all_unreclaimable? %s"
1336 K(zone->free_pages),
1339 K(zone->pages_high),
1341 K(zone->nr_inactive),
1342 K(zone->present_pages),
1343 zone->pages_scanned,
1344 (zone->all_unreclaimable ? "yes" : "no")
1346 printk("lowmem_reserve[]:");
1347 for (i = 0; i < MAX_NR_ZONES; i++)
1348 printk(" %lu", zone->lowmem_reserve[i]);
1352 for_each_zone(zone) {
1353 unsigned long nr[MAX_ORDER], flags, order, total = 0;
1356 printk("%s: ", zone->name);
1357 if (!populated_zone(zone)) {
1362 spin_lock_irqsave(&zone->lock, flags);
1363 for (order = 0; order < MAX_ORDER; order++) {
1364 nr[order] = zone->free_area[order].nr_free;
1365 total += nr[order] << order;
1367 spin_unlock_irqrestore(&zone->lock, flags);
1368 for (order = 0; order < MAX_ORDER; order++)
1369 printk("%lu*%lukB ", nr[order], K(1UL) << order);
1370 printk("= %lukB\n", K(total));
1373 show_swap_cache_info();
1377 * Builds allocation fallback zone lists.
1379 * Add all populated zones of a node to the zonelist.
1381 static int __meminit build_zonelists_node(pg_data_t *pgdat,
1382 struct zonelist *zonelist, int nr_zones, enum zone_type zone_type)
1386 BUG_ON(zone_type >= MAX_NR_ZONES);
1391 zone = pgdat->node_zones + zone_type;
1392 if (populated_zone(zone)) {
1393 zonelist->zones[nr_zones++] = zone;
1394 check_highest_zone(zone_type);
1397 } while (zone_type);
1402 #define MAX_NODE_LOAD (num_online_nodes())
1403 static int __meminitdata node_load[MAX_NUMNODES];
1405 * find_next_best_node - find the next node that should appear in a given node's fallback list
1406 * @node: node whose fallback list we're appending
1407 * @used_node_mask: nodemask_t of already used nodes
1409 * We use a number of factors to determine which is the next node that should
1410 * appear on a given node's fallback list. The node should not have appeared
1411 * already in @node's fallback list, and it should be the next closest node
1412 * according to the distance array (which contains arbitrary distance values
1413 * from each node to each node in the system), and should also prefer nodes
1414 * with no CPUs, since presumably they'll have very little allocation pressure
1415 * on them otherwise.
1416 * It returns -1 if no node is found.
1418 static int __meminit find_next_best_node(int node, nodemask_t *used_node_mask)
1421 int min_val = INT_MAX;
1424 /* Use the local node if we haven't already */
1425 if (!node_isset(node, *used_node_mask)) {
1426 node_set(node, *used_node_mask);
1430 for_each_online_node(n) {
1433 /* Don't want a node to appear more than once */
1434 if (node_isset(n, *used_node_mask))
1437 /* Use the distance array to find the distance */
1438 val = node_distance(node, n);
1440 /* Penalize nodes under us ("prefer the next node") */
1443 /* Give preference to headless and unused nodes */
1444 tmp = node_to_cpumask(n);
1445 if (!cpus_empty(tmp))
1446 val += PENALTY_FOR_NODE_WITH_CPUS;
1448 /* Slight preference for less loaded node */
1449 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1450 val += node_load[n];
1452 if (val < min_val) {
1459 node_set(best_node, *used_node_mask);
1464 static void __meminit build_zonelists(pg_data_t *pgdat)
1466 int j, node, local_node;
1468 int prev_node, load;
1469 struct zonelist *zonelist;
1470 nodemask_t used_mask;
1472 /* initialize zonelists */
1473 for (i = 0; i < MAX_NR_ZONES; i++) {
1474 zonelist = pgdat->node_zonelists + i;
1475 zonelist->zones[0] = NULL;
1478 /* NUMA-aware ordering of nodes */
1479 local_node = pgdat->node_id;
1480 load = num_online_nodes();
1481 prev_node = local_node;
1482 nodes_clear(used_mask);
1483 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1484 int distance = node_distance(local_node, node);
1487 * If another node is sufficiently far away then it is better
1488 * to reclaim pages in a zone before going off node.
1490 if (distance > RECLAIM_DISTANCE)
1491 zone_reclaim_mode = 1;
1494 * We don't want to pressure a particular node.
1495 * So adding penalty to the first node in same
1496 * distance group to make it round-robin.
1499 if (distance != node_distance(local_node, prev_node))
1500 node_load[node] += load;
1503 for (i = 0; i < MAX_NR_ZONES; i++) {
1504 zonelist = pgdat->node_zonelists + i;
1505 for (j = 0; zonelist->zones[j] != NULL; j++);
1507 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
1508 zonelist->zones[j] = NULL;
1513 #else /* CONFIG_NUMA */
1515 static void __meminit build_zonelists(pg_data_t *pgdat)
1517 int node, local_node;
1520 local_node = pgdat->node_id;
1521 for (i = 0; i < MAX_NR_ZONES; i++) {
1522 struct zonelist *zonelist;
1524 zonelist = pgdat->node_zonelists + i;
1526 j = build_zonelists_node(pgdat, zonelist, 0, i);
1528 * Now we build the zonelist so that it contains the zones
1529 * of all the other nodes.
1530 * We don't want to pressure a particular node, so when
1531 * building the zones for node N, we make sure that the
1532 * zones coming right after the local ones are those from
1533 * node N+1 (modulo N)
1535 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1536 if (!node_online(node))
1538 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
1540 for (node = 0; node < local_node; node++) {
1541 if (!node_online(node))
1543 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
1546 zonelist->zones[j] = NULL;
1550 #endif /* CONFIG_NUMA */
1552 /* return values int ....just for stop_machine_run() */
1553 static int __meminit __build_all_zonelists(void *dummy)
1556 for_each_online_node(nid)
1557 build_zonelists(NODE_DATA(nid));
1561 void __meminit build_all_zonelists(void)
1563 if (system_state == SYSTEM_BOOTING) {
1564 __build_all_zonelists(0);
1565 cpuset_init_current_mems_allowed();
1567 /* we have to stop all cpus to guaranntee there is no user
1569 stop_machine_run(__build_all_zonelists, NULL, NR_CPUS);
1570 /* cpuset refresh routine should be here */
1572 vm_total_pages = nr_free_pagecache_pages();
1573 printk("Built %i zonelists. Total pages: %ld\n",
1574 num_online_nodes(), vm_total_pages);
1578 * Helper functions to size the waitqueue hash table.
1579 * Essentially these want to choose hash table sizes sufficiently
1580 * large so that collisions trying to wait on pages are rare.
1581 * But in fact, the number of active page waitqueues on typical
1582 * systems is ridiculously low, less than 200. So this is even
1583 * conservative, even though it seems large.
1585 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1586 * waitqueues, i.e. the size of the waitq table given the number of pages.
1588 #define PAGES_PER_WAITQUEUE 256
1590 #ifndef CONFIG_MEMORY_HOTPLUG
1591 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
1593 unsigned long size = 1;
1595 pages /= PAGES_PER_WAITQUEUE;
1597 while (size < pages)
1601 * Once we have dozens or even hundreds of threads sleeping
1602 * on IO we've got bigger problems than wait queue collision.
1603 * Limit the size of the wait table to a reasonable size.
1605 size = min(size, 4096UL);
1607 return max(size, 4UL);
1611 * A zone's size might be changed by hot-add, so it is not possible to determine
1612 * a suitable size for its wait_table. So we use the maximum size now.
1614 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
1616 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
1617 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
1618 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
1620 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
1621 * or more by the traditional way. (See above). It equals:
1623 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
1624 * ia64(16K page size) : = ( 8G + 4M)byte.
1625 * powerpc (64K page size) : = (32G +16M)byte.
1627 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
1634 * This is an integer logarithm so that shifts can be used later
1635 * to extract the more random high bits from the multiplicative
1636 * hash function before the remainder is taken.
1638 static inline unsigned long wait_table_bits(unsigned long size)
1643 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1645 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1646 unsigned long *zones_size, unsigned long *zholes_size)
1648 unsigned long realtotalpages, totalpages = 0;
1651 for (i = 0; i < MAX_NR_ZONES; i++)
1652 totalpages += zones_size[i];
1653 pgdat->node_spanned_pages = totalpages;
1655 realtotalpages = totalpages;
1657 for (i = 0; i < MAX_NR_ZONES; i++)
1658 realtotalpages -= zholes_size[i];
1659 pgdat->node_present_pages = realtotalpages;
1660 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1665 * Initially all pages are reserved - free ones are freed
1666 * up by free_all_bootmem() once the early boot process is
1667 * done. Non-atomic initialization, single-pass.
1669 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1670 unsigned long start_pfn)
1673 unsigned long end_pfn = start_pfn + size;
1676 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1677 if (!early_pfn_valid(pfn))
1679 page = pfn_to_page(pfn);
1680 set_page_links(page, zone, nid, pfn);
1681 init_page_count(page);
1682 reset_page_mapcount(page);
1683 SetPageReserved(page);
1684 INIT_LIST_HEAD(&page->lru);
1685 #ifdef WANT_PAGE_VIRTUAL
1686 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1687 if (!is_highmem_idx(zone))
1688 set_page_address(page, __va(pfn << PAGE_SHIFT));
1693 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1697 for (order = 0; order < MAX_ORDER ; order++) {
1698 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1699 zone->free_area[order].nr_free = 0;
1703 #define ZONETABLE_INDEX(x, zone_nr) ((x << ZONES_SHIFT) | zone_nr)
1704 void zonetable_add(struct zone *zone, int nid, enum zone_type zid,
1705 unsigned long pfn, unsigned long size)
1707 unsigned long snum = pfn_to_section_nr(pfn);
1708 unsigned long end = pfn_to_section_nr(pfn + size);
1711 zone_table[ZONETABLE_INDEX(nid, zid)] = zone;
1713 for (; snum <= end; snum++)
1714 zone_table[ZONETABLE_INDEX(snum, zid)] = zone;
1717 #ifndef __HAVE_ARCH_MEMMAP_INIT
1718 #define memmap_init(size, nid, zone, start_pfn) \
1719 memmap_init_zone((size), (nid), (zone), (start_pfn))
1722 static int __cpuinit zone_batchsize(struct zone *zone)
1727 * The per-cpu-pages pools are set to around 1000th of the
1728 * size of the zone. But no more than 1/2 of a meg.
1730 * OK, so we don't know how big the cache is. So guess.
1732 batch = zone->present_pages / 1024;
1733 if (batch * PAGE_SIZE > 512 * 1024)
1734 batch = (512 * 1024) / PAGE_SIZE;
1735 batch /= 4; /* We effectively *= 4 below */
1740 * Clamp the batch to a 2^n - 1 value. Having a power
1741 * of 2 value was found to be more likely to have
1742 * suboptimal cache aliasing properties in some cases.
1744 * For example if 2 tasks are alternately allocating
1745 * batches of pages, one task can end up with a lot
1746 * of pages of one half of the possible page colors
1747 * and the other with pages of the other colors.
1749 batch = (1 << (fls(batch + batch/2)-1)) - 1;
1754 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
1756 struct per_cpu_pages *pcp;
1758 memset(p, 0, sizeof(*p));
1760 pcp = &p->pcp[0]; /* hot */
1762 pcp->high = 6 * batch;
1763 pcp->batch = max(1UL, 1 * batch);
1764 INIT_LIST_HEAD(&pcp->list);
1766 pcp = &p->pcp[1]; /* cold*/
1768 pcp->high = 2 * batch;
1769 pcp->batch = max(1UL, batch/2);
1770 INIT_LIST_HEAD(&pcp->list);
1774 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
1775 * to the value high for the pageset p.
1778 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
1781 struct per_cpu_pages *pcp;
1783 pcp = &p->pcp[0]; /* hot list */
1785 pcp->batch = max(1UL, high/4);
1786 if ((high/4) > (PAGE_SHIFT * 8))
1787 pcp->batch = PAGE_SHIFT * 8;
1793 * Boot pageset table. One per cpu which is going to be used for all
1794 * zones and all nodes. The parameters will be set in such a way
1795 * that an item put on a list will immediately be handed over to
1796 * the buddy list. This is safe since pageset manipulation is done
1797 * with interrupts disabled.
1799 * Some NUMA counter updates may also be caught by the boot pagesets.
1801 * The boot_pagesets must be kept even after bootup is complete for
1802 * unused processors and/or zones. They do play a role for bootstrapping
1803 * hotplugged processors.
1805 * zoneinfo_show() and maybe other functions do
1806 * not check if the processor is online before following the pageset pointer.
1807 * Other parts of the kernel may not check if the zone is available.
1809 static struct per_cpu_pageset boot_pageset[NR_CPUS];
1812 * Dynamically allocate memory for the
1813 * per cpu pageset array in struct zone.
1815 static int __cpuinit process_zones(int cpu)
1817 struct zone *zone, *dzone;
1819 for_each_zone(zone) {
1821 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
1822 GFP_KERNEL, cpu_to_node(cpu));
1823 if (!zone_pcp(zone, cpu))
1826 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
1828 if (percpu_pagelist_fraction)
1829 setup_pagelist_highmark(zone_pcp(zone, cpu),
1830 (zone->present_pages / percpu_pagelist_fraction));
1835 for_each_zone(dzone) {
1838 kfree(zone_pcp(dzone, cpu));
1839 zone_pcp(dzone, cpu) = NULL;
1844 static inline void free_zone_pagesets(int cpu)
1848 for_each_zone(zone) {
1849 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
1851 /* Free per_cpu_pageset if it is slab allocated */
1852 if (pset != &boot_pageset[cpu])
1854 zone_pcp(zone, cpu) = NULL;
1858 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
1859 unsigned long action,
1862 int cpu = (long)hcpu;
1863 int ret = NOTIFY_OK;
1866 case CPU_UP_PREPARE:
1867 if (process_zones(cpu))
1870 case CPU_UP_CANCELED:
1872 free_zone_pagesets(cpu);
1880 static struct notifier_block __cpuinitdata pageset_notifier =
1881 { &pageset_cpuup_callback, NULL, 0 };
1883 void __init setup_per_cpu_pageset(void)
1887 /* Initialize per_cpu_pageset for cpu 0.
1888 * A cpuup callback will do this for every cpu
1889 * as it comes online
1891 err = process_zones(smp_processor_id());
1893 register_cpu_notifier(&pageset_notifier);
1899 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
1902 struct pglist_data *pgdat = zone->zone_pgdat;
1906 * The per-page waitqueue mechanism uses hashed waitqueues
1909 zone->wait_table_hash_nr_entries =
1910 wait_table_hash_nr_entries(zone_size_pages);
1911 zone->wait_table_bits =
1912 wait_table_bits(zone->wait_table_hash_nr_entries);
1913 alloc_size = zone->wait_table_hash_nr_entries
1914 * sizeof(wait_queue_head_t);
1916 if (system_state == SYSTEM_BOOTING) {
1917 zone->wait_table = (wait_queue_head_t *)
1918 alloc_bootmem_node(pgdat, alloc_size);
1921 * This case means that a zone whose size was 0 gets new memory
1922 * via memory hot-add.
1923 * But it may be the case that a new node was hot-added. In
1924 * this case vmalloc() will not be able to use this new node's
1925 * memory - this wait_table must be initialized to use this new
1926 * node itself as well.
1927 * To use this new node's memory, further consideration will be
1930 zone->wait_table = (wait_queue_head_t *)vmalloc(alloc_size);
1932 if (!zone->wait_table)
1935 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
1936 init_waitqueue_head(zone->wait_table + i);
1941 static __meminit void zone_pcp_init(struct zone *zone)
1944 unsigned long batch = zone_batchsize(zone);
1946 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1948 /* Early boot. Slab allocator not functional yet */
1949 zone_pcp(zone, cpu) = &boot_pageset[cpu];
1950 setup_pageset(&boot_pageset[cpu],0);
1952 setup_pageset(zone_pcp(zone,cpu), batch);
1955 if (zone->present_pages)
1956 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
1957 zone->name, zone->present_pages, batch);
1960 __meminit int init_currently_empty_zone(struct zone *zone,
1961 unsigned long zone_start_pfn,
1964 struct pglist_data *pgdat = zone->zone_pgdat;
1966 ret = zone_wait_table_init(zone, size);
1969 pgdat->nr_zones = zone_idx(zone) + 1;
1971 zone->zone_start_pfn = zone_start_pfn;
1973 memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn);
1975 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1981 * Set up the zone data structures:
1982 * - mark all pages reserved
1983 * - mark all memory queues empty
1984 * - clear the memory bitmaps
1986 static void __meminit free_area_init_core(struct pglist_data *pgdat,
1987 unsigned long *zones_size, unsigned long *zholes_size)
1990 int nid = pgdat->node_id;
1991 unsigned long zone_start_pfn = pgdat->node_start_pfn;
1994 pgdat_resize_init(pgdat);
1995 pgdat->nr_zones = 0;
1996 init_waitqueue_head(&pgdat->kswapd_wait);
1997 pgdat->kswapd_max_order = 0;
1999 for (j = 0; j < MAX_NR_ZONES; j++) {
2000 struct zone *zone = pgdat->node_zones + j;
2001 unsigned long size, realsize;
2003 realsize = size = zones_size[j];
2005 realsize -= zholes_size[j];
2007 if (!is_highmem_idx(j))
2008 nr_kernel_pages += realsize;
2009 nr_all_pages += realsize;
2011 zone->spanned_pages = size;
2012 zone->present_pages = realsize;
2014 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
2016 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
2018 zone->name = zone_names[j];
2019 spin_lock_init(&zone->lock);
2020 spin_lock_init(&zone->lru_lock);
2021 zone_seqlock_init(zone);
2022 zone->zone_pgdat = pgdat;
2023 zone->free_pages = 0;
2025 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
2027 zone_pcp_init(zone);
2028 INIT_LIST_HEAD(&zone->active_list);
2029 INIT_LIST_HEAD(&zone->inactive_list);
2030 zone->nr_scan_active = 0;
2031 zone->nr_scan_inactive = 0;
2032 zone->nr_active = 0;
2033 zone->nr_inactive = 0;
2034 zap_zone_vm_stats(zone);
2035 atomic_set(&zone->reclaim_in_progress, 0);
2039 zonetable_add(zone, nid, j, zone_start_pfn, size);
2040 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
2042 zone_start_pfn += size;
2046 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
2048 /* Skip empty nodes */
2049 if (!pgdat->node_spanned_pages)
2052 #ifdef CONFIG_FLAT_NODE_MEM_MAP
2053 /* ia64 gets its own node_mem_map, before this, without bootmem */
2054 if (!pgdat->node_mem_map) {
2055 unsigned long size, start, end;
2059 * The zone's endpoints aren't required to be MAX_ORDER
2060 * aligned but the node_mem_map endpoints must be in order
2061 * for the buddy allocator to function correctly.
2063 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
2064 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
2065 end = ALIGN(end, MAX_ORDER_NR_PAGES);
2066 size = (end - start) * sizeof(struct page);
2067 map = alloc_remap(pgdat->node_id, size);
2069 map = alloc_bootmem_node(pgdat, size);
2070 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
2072 #ifdef CONFIG_FLATMEM
2074 * With no DISCONTIG, the global mem_map is just set as node 0's
2076 if (pgdat == NODE_DATA(0))
2077 mem_map = NODE_DATA(0)->node_mem_map;
2079 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
2082 void __meminit free_area_init_node(int nid, struct pglist_data *pgdat,
2083 unsigned long *zones_size, unsigned long node_start_pfn,
2084 unsigned long *zholes_size)
2086 pgdat->node_id = nid;
2087 pgdat->node_start_pfn = node_start_pfn;
2088 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
2090 alloc_node_mem_map(pgdat);
2092 free_area_init_core(pgdat, zones_size, zholes_size);
2095 #ifndef CONFIG_NEED_MULTIPLE_NODES
2096 static bootmem_data_t contig_bootmem_data;
2097 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
2099 EXPORT_SYMBOL(contig_page_data);
2102 void __init free_area_init(unsigned long *zones_size)
2104 free_area_init_node(0, NODE_DATA(0), zones_size,
2105 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
2108 #ifdef CONFIG_HOTPLUG_CPU
2109 static int page_alloc_cpu_notify(struct notifier_block *self,
2110 unsigned long action, void *hcpu)
2112 int cpu = (unsigned long)hcpu;
2114 if (action == CPU_DEAD) {
2115 local_irq_disable();
2117 vm_events_fold_cpu(cpu);
2119 refresh_cpu_vm_stats(cpu);
2123 #endif /* CONFIG_HOTPLUG_CPU */
2125 void __init page_alloc_init(void)
2127 hotcpu_notifier(page_alloc_cpu_notify, 0);
2131 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
2132 * or min_free_kbytes changes.
2134 static void calculate_totalreserve_pages(void)
2136 struct pglist_data *pgdat;
2137 unsigned long reserve_pages = 0;
2138 enum zone_type i, j;
2140 for_each_online_pgdat(pgdat) {
2141 for (i = 0; i < MAX_NR_ZONES; i++) {
2142 struct zone *zone = pgdat->node_zones + i;
2143 unsigned long max = 0;
2145 /* Find valid and maximum lowmem_reserve in the zone */
2146 for (j = i; j < MAX_NR_ZONES; j++) {
2147 if (zone->lowmem_reserve[j] > max)
2148 max = zone->lowmem_reserve[j];
2151 /* we treat pages_high as reserved pages. */
2152 max += zone->pages_high;
2154 if (max > zone->present_pages)
2155 max = zone->present_pages;
2156 reserve_pages += max;
2159 totalreserve_pages = reserve_pages;
2163 * setup_per_zone_lowmem_reserve - called whenever
2164 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
2165 * has a correct pages reserved value, so an adequate number of
2166 * pages are left in the zone after a successful __alloc_pages().
2168 static void setup_per_zone_lowmem_reserve(void)
2170 struct pglist_data *pgdat;
2171 enum zone_type j, idx;
2173 for_each_online_pgdat(pgdat) {
2174 for (j = 0; j < MAX_NR_ZONES; j++) {
2175 struct zone *zone = pgdat->node_zones + j;
2176 unsigned long present_pages = zone->present_pages;
2178 zone->lowmem_reserve[j] = 0;
2182 struct zone *lower_zone;
2186 if (sysctl_lowmem_reserve_ratio[idx] < 1)
2187 sysctl_lowmem_reserve_ratio[idx] = 1;
2189 lower_zone = pgdat->node_zones + idx;
2190 lower_zone->lowmem_reserve[j] = present_pages /
2191 sysctl_lowmem_reserve_ratio[idx];
2192 present_pages += lower_zone->present_pages;
2197 /* update totalreserve_pages */
2198 calculate_totalreserve_pages();
2202 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
2203 * that the pages_{min,low,high} values for each zone are set correctly
2204 * with respect to min_free_kbytes.
2206 void setup_per_zone_pages_min(void)
2208 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
2209 unsigned long lowmem_pages = 0;
2211 unsigned long flags;
2213 /* Calculate total number of !ZONE_HIGHMEM pages */
2214 for_each_zone(zone) {
2215 if (!is_highmem(zone))
2216 lowmem_pages += zone->present_pages;
2219 for_each_zone(zone) {
2222 spin_lock_irqsave(&zone->lru_lock, flags);
2223 tmp = (u64)pages_min * zone->present_pages;
2224 do_div(tmp, lowmem_pages);
2225 if (is_highmem(zone)) {
2227 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
2228 * need highmem pages, so cap pages_min to a small
2231 * The (pages_high-pages_low) and (pages_low-pages_min)
2232 * deltas controls asynch page reclaim, and so should
2233 * not be capped for highmem.
2237 min_pages = zone->present_pages / 1024;
2238 if (min_pages < SWAP_CLUSTER_MAX)
2239 min_pages = SWAP_CLUSTER_MAX;
2240 if (min_pages > 128)
2242 zone->pages_min = min_pages;
2245 * If it's a lowmem zone, reserve a number of pages
2246 * proportionate to the zone's size.
2248 zone->pages_min = tmp;
2251 zone->pages_low = zone->pages_min + (tmp >> 2);
2252 zone->pages_high = zone->pages_min + (tmp >> 1);
2253 spin_unlock_irqrestore(&zone->lru_lock, flags);
2256 /* update totalreserve_pages */
2257 calculate_totalreserve_pages();
2261 * Initialise min_free_kbytes.
2263 * For small machines we want it small (128k min). For large machines
2264 * we want it large (64MB max). But it is not linear, because network
2265 * bandwidth does not increase linearly with machine size. We use
2267 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2268 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2284 static int __init init_per_zone_pages_min(void)
2286 unsigned long lowmem_kbytes;
2288 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2290 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2291 if (min_free_kbytes < 128)
2292 min_free_kbytes = 128;
2293 if (min_free_kbytes > 65536)
2294 min_free_kbytes = 65536;
2295 setup_per_zone_pages_min();
2296 setup_per_zone_lowmem_reserve();
2299 module_init(init_per_zone_pages_min)
2302 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2303 * that we can call two helper functions whenever min_free_kbytes
2306 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2307 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2309 proc_dointvec(table, write, file, buffer, length, ppos);
2310 setup_per_zone_pages_min();
2315 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
2316 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2321 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2326 zone->min_unmapped_pages = (zone->present_pages *
2327 sysctl_min_unmapped_ratio) / 100;
2331 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
2332 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2337 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2342 zone->min_slab_pages = (zone->present_pages *
2343 sysctl_min_slab_ratio) / 100;
2349 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2350 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2351 * whenever sysctl_lowmem_reserve_ratio changes.
2353 * The reserve ratio obviously has absolutely no relation with the
2354 * pages_min watermarks. The lowmem reserve ratio can only make sense
2355 * if in function of the boot time zone sizes.
2357 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
2358 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2360 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2361 setup_per_zone_lowmem_reserve();
2366 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
2367 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
2368 * can have before it gets flushed back to buddy allocator.
2371 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
2372 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2378 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2379 if (!write || (ret == -EINVAL))
2381 for_each_zone(zone) {
2382 for_each_online_cpu(cpu) {
2384 high = zone->present_pages / percpu_pagelist_fraction;
2385 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
2391 int hashdist = HASHDIST_DEFAULT;
2394 static int __init set_hashdist(char *str)
2398 hashdist = simple_strtoul(str, &str, 0);
2401 __setup("hashdist=", set_hashdist);
2405 * allocate a large system hash table from bootmem
2406 * - it is assumed that the hash table must contain an exact power-of-2
2407 * quantity of entries
2408 * - limit is the number of hash buckets, not the total allocation size
2410 void *__init alloc_large_system_hash(const char *tablename,
2411 unsigned long bucketsize,
2412 unsigned long numentries,
2415 unsigned int *_hash_shift,
2416 unsigned int *_hash_mask,
2417 unsigned long limit)
2419 unsigned long long max = limit;
2420 unsigned long log2qty, size;
2423 /* allow the kernel cmdline to have a say */
2425 /* round applicable memory size up to nearest megabyte */
2426 numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages;
2427 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2428 numentries >>= 20 - PAGE_SHIFT;
2429 numentries <<= 20 - PAGE_SHIFT;
2431 /* limit to 1 bucket per 2^scale bytes of low memory */
2432 if (scale > PAGE_SHIFT)
2433 numentries >>= (scale - PAGE_SHIFT);
2435 numentries <<= (PAGE_SHIFT - scale);
2437 numentries = roundup_pow_of_two(numentries);
2439 /* limit allocation size to 1/16 total memory by default */
2441 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2442 do_div(max, bucketsize);
2445 if (numentries > max)
2448 log2qty = long_log2(numentries);
2451 size = bucketsize << log2qty;
2452 if (flags & HASH_EARLY)
2453 table = alloc_bootmem(size);
2455 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
2457 unsigned long order;
2458 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
2460 table = (void*) __get_free_pages(GFP_ATOMIC, order);
2462 } while (!table && size > PAGE_SIZE && --log2qty);
2465 panic("Failed to allocate %s hash table\n", tablename);
2467 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2470 long_log2(size) - PAGE_SHIFT,
2474 *_hash_shift = log2qty;
2476 *_hash_mask = (1 << log2qty) - 1;
2481 #ifdef CONFIG_OUT_OF_LINE_PFN_TO_PAGE
2482 struct page *pfn_to_page(unsigned long pfn)
2484 return __pfn_to_page(pfn);
2486 unsigned long page_to_pfn(struct page *page)
2488 return __page_to_pfn(page);
2490 EXPORT_SYMBOL(pfn_to_page);
2491 EXPORT_SYMBOL(page_to_pfn);
2492 #endif /* CONFIG_OUT_OF_LINE_PFN_TO_PAGE */