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)
452 list_add(&page->lru, &list);
453 free_pages_bulk(zone, 1, &list, order);
456 static void __free_pages_ok(struct page *page, unsigned int order)
462 arch_free_page(page, order);
463 if (!PageHighMem(page))
464 debug_check_no_locks_freed(page_address(page),
467 for (i = 0 ; i < (1 << order) ; ++i)
468 reserved += free_pages_check(page + i);
472 kernel_map_pages(page, 1 << order, 0);
473 local_irq_save(flags);
474 __count_vm_events(PGFREE, 1 << order);
475 free_one_page(page_zone(page), page, order);
476 local_irq_restore(flags);
480 * permit the bootmem allocator to evade page validation on high-order frees
482 void fastcall __init __free_pages_bootmem(struct page *page, unsigned int order)
485 __ClearPageReserved(page);
486 set_page_count(page, 0);
487 set_page_refcounted(page);
493 for (loop = 0; loop < BITS_PER_LONG; loop++) {
494 struct page *p = &page[loop];
496 if (loop + 1 < BITS_PER_LONG)
498 __ClearPageReserved(p);
499 set_page_count(p, 0);
502 set_page_refcounted(page);
503 __free_pages(page, order);
509 * The order of subdivision here is critical for the IO subsystem.
510 * Please do not alter this order without good reasons and regression
511 * testing. Specifically, as large blocks of memory are subdivided,
512 * the order in which smaller blocks are delivered depends on the order
513 * they're subdivided in this function. This is the primary factor
514 * influencing the order in which pages are delivered to the IO
515 * subsystem according to empirical testing, and this is also justified
516 * by considering the behavior of a buddy system containing a single
517 * large block of memory acted on by a series of small allocations.
518 * This behavior is a critical factor in sglist merging's success.
522 static inline void expand(struct zone *zone, struct page *page,
523 int low, int high, struct free_area *area)
525 unsigned long size = 1 << high;
531 VM_BUG_ON(bad_range(zone, &page[size]));
532 list_add(&page[size].lru, &area->free_list);
534 set_page_order(&page[size], high);
539 * This page is about to be returned from the page allocator
541 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
543 if (unlikely(page_mapcount(page) |
544 (page->mapping != NULL) |
545 (page_count(page) != 0) |
561 * For now, we report if PG_reserved was found set, but do not
562 * clear it, and do not allocate the page: as a safety net.
564 if (PageReserved(page))
567 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
568 1 << PG_referenced | 1 << PG_arch_1 |
569 1 << PG_checked | 1 << PG_mappedtodisk);
570 set_page_private(page, 0);
571 set_page_refcounted(page);
572 kernel_map_pages(page, 1 << order, 1);
574 if (gfp_flags & __GFP_ZERO)
575 prep_zero_page(page, order, gfp_flags);
577 if (order && (gfp_flags & __GFP_COMP))
578 prep_compound_page(page, order);
584 * Do the hard work of removing an element from the buddy allocator.
585 * Call me with the zone->lock already held.
587 static struct page *__rmqueue(struct zone *zone, unsigned int order)
589 struct free_area * area;
590 unsigned int current_order;
593 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
594 area = zone->free_area + current_order;
595 if (list_empty(&area->free_list))
598 page = list_entry(area->free_list.next, struct page, lru);
599 list_del(&page->lru);
600 rmv_page_order(page);
602 zone->free_pages -= 1UL << order;
603 expand(zone, page, order, current_order, area);
611 * Obtain a specified number of elements from the buddy allocator, all under
612 * a single hold of the lock, for efficiency. Add them to the supplied list.
613 * Returns the number of new pages which were placed at *list.
615 static int rmqueue_bulk(struct zone *zone, unsigned int order,
616 unsigned long count, struct list_head *list)
620 spin_lock(&zone->lock);
621 for (i = 0; i < count; ++i) {
622 struct page *page = __rmqueue(zone, order);
623 if (unlikely(page == NULL))
625 list_add_tail(&page->lru, list);
627 spin_unlock(&zone->lock);
633 * Called from the slab reaper to drain pagesets on a particular node that
634 * belong to the currently executing processor.
635 * Note that this function must be called with the thread pinned to
636 * a single processor.
638 void drain_node_pages(int nodeid)
644 for (z = 0; z < MAX_NR_ZONES; z++) {
645 struct zone *zone = NODE_DATA(nodeid)->node_zones + z;
646 struct per_cpu_pageset *pset;
648 pset = zone_pcp(zone, smp_processor_id());
649 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
650 struct per_cpu_pages *pcp;
654 local_irq_save(flags);
655 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
657 local_irq_restore(flags);
664 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
665 static void __drain_pages(unsigned int cpu)
671 for_each_zone(zone) {
672 struct per_cpu_pageset *pset;
674 pset = zone_pcp(zone, cpu);
675 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
676 struct per_cpu_pages *pcp;
679 local_irq_save(flags);
680 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
682 local_irq_restore(flags);
686 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
690 void mark_free_pages(struct zone *zone)
692 unsigned long zone_pfn, flags;
694 struct list_head *curr;
696 if (!zone->spanned_pages)
699 spin_lock_irqsave(&zone->lock, flags);
700 for (zone_pfn = 0; zone_pfn < zone->spanned_pages; ++zone_pfn)
701 ClearPageNosaveFree(pfn_to_page(zone_pfn + zone->zone_start_pfn));
703 for (order = MAX_ORDER - 1; order >= 0; --order)
704 list_for_each(curr, &zone->free_area[order].free_list) {
705 unsigned long start_pfn, i;
707 start_pfn = page_to_pfn(list_entry(curr, struct page, lru));
709 for (i=0; i < (1<<order); i++)
710 SetPageNosaveFree(pfn_to_page(start_pfn+i));
712 spin_unlock_irqrestore(&zone->lock, flags);
716 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
718 void drain_local_pages(void)
722 local_irq_save(flags);
723 __drain_pages(smp_processor_id());
724 local_irq_restore(flags);
726 #endif /* CONFIG_PM */
729 * Free a 0-order page
731 static void fastcall free_hot_cold_page(struct page *page, int cold)
733 struct zone *zone = page_zone(page);
734 struct per_cpu_pages *pcp;
737 arch_free_page(page, 0);
740 page->mapping = NULL;
741 if (free_pages_check(page))
744 kernel_map_pages(page, 1, 0);
746 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
747 local_irq_save(flags);
748 __count_vm_event(PGFREE);
749 list_add(&page->lru, &pcp->list);
751 if (pcp->count >= pcp->high) {
752 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
753 pcp->count -= pcp->batch;
755 local_irq_restore(flags);
759 void fastcall free_hot_page(struct page *page)
761 free_hot_cold_page(page, 0);
764 void fastcall free_cold_page(struct page *page)
766 free_hot_cold_page(page, 1);
770 * split_page takes a non-compound higher-order page, and splits it into
771 * n (1<<order) sub-pages: page[0..n]
772 * Each sub-page must be freed individually.
774 * Note: this is probably too low level an operation for use in drivers.
775 * Please consult with lkml before using this in your driver.
777 void split_page(struct page *page, unsigned int order)
781 VM_BUG_ON(PageCompound(page));
782 VM_BUG_ON(!page_count(page));
783 for (i = 1; i < (1 << order); i++)
784 set_page_refcounted(page + i);
788 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
789 * we cheat by calling it from here, in the order > 0 path. Saves a branch
792 static struct page *buffered_rmqueue(struct zonelist *zonelist,
793 struct zone *zone, int order, gfp_t gfp_flags)
797 int cold = !!(gfp_flags & __GFP_COLD);
802 if (likely(order == 0)) {
803 struct per_cpu_pages *pcp;
805 pcp = &zone_pcp(zone, cpu)->pcp[cold];
806 local_irq_save(flags);
808 pcp->count += rmqueue_bulk(zone, 0,
809 pcp->batch, &pcp->list);
810 if (unlikely(!pcp->count))
813 page = list_entry(pcp->list.next, struct page, lru);
814 list_del(&page->lru);
817 spin_lock_irqsave(&zone->lock, flags);
818 page = __rmqueue(zone, order);
819 spin_unlock(&zone->lock);
824 __count_zone_vm_events(PGALLOC, zone, 1 << order);
825 zone_statistics(zonelist, zone);
826 local_irq_restore(flags);
829 VM_BUG_ON(bad_range(zone, page));
830 if (prep_new_page(page, order, gfp_flags))
835 local_irq_restore(flags);
840 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
841 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
842 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
843 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
844 #define ALLOC_HARDER 0x10 /* try to alloc harder */
845 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
846 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
849 * Return 1 if free pages are above 'mark'. This takes into account the order
852 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
853 int classzone_idx, int alloc_flags)
855 /* free_pages my go negative - that's OK */
856 long min = mark, free_pages = z->free_pages - (1 << order) + 1;
859 if (alloc_flags & ALLOC_HIGH)
861 if (alloc_flags & ALLOC_HARDER)
864 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
866 for (o = 0; o < order; o++) {
867 /* At the next order, this order's pages become unavailable */
868 free_pages -= z->free_area[o].nr_free << o;
870 /* Require fewer higher order pages to be free */
873 if (free_pages <= min)
880 * get_page_from_freeliest goes through the zonelist trying to allocate
884 get_page_from_freelist(gfp_t gfp_mask, unsigned int order,
885 struct zonelist *zonelist, int alloc_flags)
887 struct zone **z = zonelist->zones;
888 struct page *page = NULL;
889 int classzone_idx = zone_idx(*z);
892 * Go through the zonelist once, looking for a zone with enough free.
893 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
896 if ((alloc_flags & ALLOC_CPUSET) &&
897 !cpuset_zone_allowed(*z, gfp_mask))
900 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
902 if (alloc_flags & ALLOC_WMARK_MIN)
903 mark = (*z)->pages_min;
904 else if (alloc_flags & ALLOC_WMARK_LOW)
905 mark = (*z)->pages_low;
907 mark = (*z)->pages_high;
908 if (!zone_watermark_ok(*z, order, mark,
909 classzone_idx, alloc_flags))
910 if (!zone_reclaim_mode ||
911 !zone_reclaim(*z, gfp_mask, order))
915 page = buffered_rmqueue(zonelist, *z, order, gfp_mask);
919 } while (*(++z) != NULL);
924 * This is the 'heart' of the zoned buddy allocator.
926 struct page * fastcall
927 __alloc_pages(gfp_t gfp_mask, unsigned int order,
928 struct zonelist *zonelist)
930 const gfp_t wait = gfp_mask & __GFP_WAIT;
933 struct reclaim_state reclaim_state;
934 struct task_struct *p = current;
937 int did_some_progress;
939 might_sleep_if(wait);
942 z = zonelist->zones; /* the list of zones suitable for gfp_mask */
944 if (unlikely(*z == NULL)) {
945 /* Should this ever happen?? */
949 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
950 zonelist, ALLOC_WMARK_LOW|ALLOC_CPUSET);
955 wakeup_kswapd(*z, order);
959 * OK, we're below the kswapd watermark and have kicked background
960 * reclaim. Now things get more complex, so set up alloc_flags according
961 * to how we want to proceed.
963 * The caller may dip into page reserves a bit more if the caller
964 * cannot run direct reclaim, or if the caller has realtime scheduling
965 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
966 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
968 alloc_flags = ALLOC_WMARK_MIN;
969 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
970 alloc_flags |= ALLOC_HARDER;
971 if (gfp_mask & __GFP_HIGH)
972 alloc_flags |= ALLOC_HIGH;
974 alloc_flags |= ALLOC_CPUSET;
977 * Go through the zonelist again. Let __GFP_HIGH and allocations
978 * coming from realtime tasks go deeper into reserves.
980 * This is the last chance, in general, before the goto nopage.
981 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
982 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
984 page = get_page_from_freelist(gfp_mask, order, zonelist, alloc_flags);
988 /* This allocation should allow future memory freeing. */
990 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
991 && !in_interrupt()) {
992 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
994 /* go through the zonelist yet again, ignoring mins */
995 page = get_page_from_freelist(gfp_mask, order,
996 zonelist, ALLOC_NO_WATERMARKS);
999 if (gfp_mask & __GFP_NOFAIL) {
1000 blk_congestion_wait(WRITE, HZ/50);
1007 /* Atomic allocations - we can't balance anything */
1014 /* We now go into synchronous reclaim */
1015 cpuset_memory_pressure_bump();
1016 p->flags |= PF_MEMALLOC;
1017 reclaim_state.reclaimed_slab = 0;
1018 p->reclaim_state = &reclaim_state;
1020 did_some_progress = try_to_free_pages(zonelist->zones, gfp_mask);
1022 p->reclaim_state = NULL;
1023 p->flags &= ~PF_MEMALLOC;
1027 if (likely(did_some_progress)) {
1028 page = get_page_from_freelist(gfp_mask, order,
1029 zonelist, alloc_flags);
1032 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1034 * Go through the zonelist yet one more time, keep
1035 * very high watermark here, this is only to catch
1036 * a parallel oom killing, we must fail if we're still
1037 * under heavy pressure.
1039 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1040 zonelist, ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1044 out_of_memory(zonelist, gfp_mask, order);
1049 * Don't let big-order allocations loop unless the caller explicitly
1050 * requests that. Wait for some write requests to complete then retry.
1052 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
1053 * <= 3, but that may not be true in other implementations.
1056 if (!(gfp_mask & __GFP_NORETRY)) {
1057 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
1059 if (gfp_mask & __GFP_NOFAIL)
1063 blk_congestion_wait(WRITE, HZ/50);
1068 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1069 printk(KERN_WARNING "%s: page allocation failure."
1070 " order:%d, mode:0x%x\n",
1071 p->comm, order, gfp_mask);
1079 EXPORT_SYMBOL(__alloc_pages);
1082 * Common helper functions.
1084 fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1087 page = alloc_pages(gfp_mask, order);
1090 return (unsigned long) page_address(page);
1093 EXPORT_SYMBOL(__get_free_pages);
1095 fastcall unsigned long get_zeroed_page(gfp_t gfp_mask)
1100 * get_zeroed_page() returns a 32-bit address, which cannot represent
1103 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1105 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1107 return (unsigned long) page_address(page);
1111 EXPORT_SYMBOL(get_zeroed_page);
1113 void __pagevec_free(struct pagevec *pvec)
1115 int i = pagevec_count(pvec);
1118 free_hot_cold_page(pvec->pages[i], pvec->cold);
1121 fastcall void __free_pages(struct page *page, unsigned int order)
1123 if (put_page_testzero(page)) {
1125 free_hot_page(page);
1127 __free_pages_ok(page, order);
1131 EXPORT_SYMBOL(__free_pages);
1133 fastcall void free_pages(unsigned long addr, unsigned int order)
1136 VM_BUG_ON(!virt_addr_valid((void *)addr));
1137 __free_pages(virt_to_page((void *)addr), order);
1141 EXPORT_SYMBOL(free_pages);
1144 * Total amount of free (allocatable) RAM:
1146 unsigned int nr_free_pages(void)
1148 unsigned int sum = 0;
1152 sum += zone->free_pages;
1157 EXPORT_SYMBOL(nr_free_pages);
1160 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
1162 unsigned int sum = 0;
1165 for (i = 0; i < MAX_NR_ZONES; i++)
1166 sum += pgdat->node_zones[i].free_pages;
1172 static unsigned int nr_free_zone_pages(int offset)
1174 /* Just pick one node, since fallback list is circular */
1175 pg_data_t *pgdat = NODE_DATA(numa_node_id());
1176 unsigned int sum = 0;
1178 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1179 struct zone **zonep = zonelist->zones;
1182 for (zone = *zonep++; zone; zone = *zonep++) {
1183 unsigned long size = zone->present_pages;
1184 unsigned long high = zone->pages_high;
1193 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1195 unsigned int nr_free_buffer_pages(void)
1197 return nr_free_zone_pages(gfp_zone(GFP_USER));
1201 * Amount of free RAM allocatable within all zones
1203 unsigned int nr_free_pagecache_pages(void)
1205 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER));
1208 static void show_node(struct zone *zone)
1210 printk("Node %d ", zone->zone_pgdat->node_id);
1213 #define show_node(zone) do { } while (0)
1216 void si_meminfo(struct sysinfo *val)
1218 val->totalram = totalram_pages;
1220 val->freeram = nr_free_pages();
1221 val->bufferram = nr_blockdev_pages();
1222 val->totalhigh = totalhigh_pages;
1223 val->freehigh = nr_free_highpages();
1224 val->mem_unit = PAGE_SIZE;
1227 EXPORT_SYMBOL(si_meminfo);
1230 void si_meminfo_node(struct sysinfo *val, int nid)
1232 pg_data_t *pgdat = NODE_DATA(nid);
1234 val->totalram = pgdat->node_present_pages;
1235 val->freeram = nr_free_pages_pgdat(pgdat);
1236 #ifdef CONFIG_HIGHMEM
1237 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1238 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1243 val->mem_unit = PAGE_SIZE;
1247 #define K(x) ((x) << (PAGE_SHIFT-10))
1250 * Show free area list (used inside shift_scroll-lock stuff)
1251 * We also calculate the percentage fragmentation. We do this by counting the
1252 * memory on each free list with the exception of the first item on the list.
1254 void show_free_areas(void)
1256 int cpu, temperature;
1257 unsigned long active;
1258 unsigned long inactive;
1262 for_each_zone(zone) {
1264 printk("%s per-cpu:", zone->name);
1266 if (!populated_zone(zone)) {
1272 for_each_online_cpu(cpu) {
1273 struct per_cpu_pageset *pageset;
1275 pageset = zone_pcp(zone, cpu);
1277 for (temperature = 0; temperature < 2; temperature++)
1278 printk("cpu %d %s: high %d, batch %d used:%d\n",
1280 temperature ? "cold" : "hot",
1281 pageset->pcp[temperature].high,
1282 pageset->pcp[temperature].batch,
1283 pageset->pcp[temperature].count);
1287 get_zone_counts(&active, &inactive, &free);
1289 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1290 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1293 global_page_state(NR_FILE_DIRTY),
1294 global_page_state(NR_WRITEBACK),
1295 global_page_state(NR_UNSTABLE_NFS),
1297 global_page_state(NR_SLAB),
1298 global_page_state(NR_FILE_MAPPED),
1299 global_page_state(NR_PAGETABLE));
1301 for_each_zone(zone) {
1313 " pages_scanned:%lu"
1314 " all_unreclaimable? %s"
1317 K(zone->free_pages),
1320 K(zone->pages_high),
1322 K(zone->nr_inactive),
1323 K(zone->present_pages),
1324 zone->pages_scanned,
1325 (zone->all_unreclaimable ? "yes" : "no")
1327 printk("lowmem_reserve[]:");
1328 for (i = 0; i < MAX_NR_ZONES; i++)
1329 printk(" %lu", zone->lowmem_reserve[i]);
1333 for_each_zone(zone) {
1334 unsigned long nr[MAX_ORDER], flags, order, total = 0;
1337 printk("%s: ", zone->name);
1338 if (!populated_zone(zone)) {
1343 spin_lock_irqsave(&zone->lock, flags);
1344 for (order = 0; order < MAX_ORDER; order++) {
1345 nr[order] = zone->free_area[order].nr_free;
1346 total += nr[order] << order;
1348 spin_unlock_irqrestore(&zone->lock, flags);
1349 for (order = 0; order < MAX_ORDER; order++)
1350 printk("%lu*%lukB ", nr[order], K(1UL) << order);
1351 printk("= %lukB\n", K(total));
1354 show_swap_cache_info();
1358 * Builds allocation fallback zone lists.
1360 * Add all populated zones of a node to the zonelist.
1362 static int __meminit build_zonelists_node(pg_data_t *pgdat,
1363 struct zonelist *zonelist, int nr_zones, enum zone_type zone_type)
1367 BUG_ON(zone_type >= MAX_NR_ZONES);
1372 zone = pgdat->node_zones + zone_type;
1373 if (populated_zone(zone)) {
1374 zonelist->zones[nr_zones++] = zone;
1375 check_highest_zone(zone_type);
1378 } while (zone_type);
1383 #define MAX_NODE_LOAD (num_online_nodes())
1384 static int __meminitdata node_load[MAX_NUMNODES];
1386 * find_next_best_node - find the next node that should appear in a given node's fallback list
1387 * @node: node whose fallback list we're appending
1388 * @used_node_mask: nodemask_t of already used nodes
1390 * We use a number of factors to determine which is the next node that should
1391 * appear on a given node's fallback list. The node should not have appeared
1392 * already in @node's fallback list, and it should be the next closest node
1393 * according to the distance array (which contains arbitrary distance values
1394 * from each node to each node in the system), and should also prefer nodes
1395 * with no CPUs, since presumably they'll have very little allocation pressure
1396 * on them otherwise.
1397 * It returns -1 if no node is found.
1399 static int __meminit find_next_best_node(int node, nodemask_t *used_node_mask)
1402 int min_val = INT_MAX;
1405 /* Use the local node if we haven't already */
1406 if (!node_isset(node, *used_node_mask)) {
1407 node_set(node, *used_node_mask);
1411 for_each_online_node(n) {
1414 /* Don't want a node to appear more than once */
1415 if (node_isset(n, *used_node_mask))
1418 /* Use the distance array to find the distance */
1419 val = node_distance(node, n);
1421 /* Penalize nodes under us ("prefer the next node") */
1424 /* Give preference to headless and unused nodes */
1425 tmp = node_to_cpumask(n);
1426 if (!cpus_empty(tmp))
1427 val += PENALTY_FOR_NODE_WITH_CPUS;
1429 /* Slight preference for less loaded node */
1430 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1431 val += node_load[n];
1433 if (val < min_val) {
1440 node_set(best_node, *used_node_mask);
1445 static void __meminit build_zonelists(pg_data_t *pgdat)
1447 int j, node, local_node;
1449 int prev_node, load;
1450 struct zonelist *zonelist;
1451 nodemask_t used_mask;
1453 /* initialize zonelists */
1454 for (i = 0; i < MAX_NR_ZONES; i++) {
1455 zonelist = pgdat->node_zonelists + i;
1456 zonelist->zones[0] = NULL;
1459 /* NUMA-aware ordering of nodes */
1460 local_node = pgdat->node_id;
1461 load = num_online_nodes();
1462 prev_node = local_node;
1463 nodes_clear(used_mask);
1464 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1465 int distance = node_distance(local_node, node);
1468 * If another node is sufficiently far away then it is better
1469 * to reclaim pages in a zone before going off node.
1471 if (distance > RECLAIM_DISTANCE)
1472 zone_reclaim_mode = 1;
1475 * We don't want to pressure a particular node.
1476 * So adding penalty to the first node in same
1477 * distance group to make it round-robin.
1480 if (distance != node_distance(local_node, prev_node))
1481 node_load[node] += load;
1484 for (i = 0; i < MAX_NR_ZONES; i++) {
1485 zonelist = pgdat->node_zonelists + i;
1486 for (j = 0; zonelist->zones[j] != NULL; j++);
1488 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
1489 zonelist->zones[j] = NULL;
1494 #else /* CONFIG_NUMA */
1496 static void __meminit build_zonelists(pg_data_t *pgdat)
1498 int node, local_node;
1501 local_node = pgdat->node_id;
1502 for (i = 0; i < MAX_NR_ZONES; i++) {
1503 struct zonelist *zonelist;
1505 zonelist = pgdat->node_zonelists + i;
1507 j = build_zonelists_node(pgdat, zonelist, 0, i);
1509 * Now we build the zonelist so that it contains the zones
1510 * of all the other nodes.
1511 * We don't want to pressure a particular node, so when
1512 * building the zones for node N, we make sure that the
1513 * zones coming right after the local ones are those from
1514 * node N+1 (modulo N)
1516 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1517 if (!node_online(node))
1519 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
1521 for (node = 0; node < local_node; node++) {
1522 if (!node_online(node))
1524 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
1527 zonelist->zones[j] = NULL;
1531 #endif /* CONFIG_NUMA */
1533 /* return values int ....just for stop_machine_run() */
1534 static int __meminit __build_all_zonelists(void *dummy)
1537 for_each_online_node(nid)
1538 build_zonelists(NODE_DATA(nid));
1542 void __meminit build_all_zonelists(void)
1544 if (system_state == SYSTEM_BOOTING) {
1545 __build_all_zonelists(0);
1546 cpuset_init_current_mems_allowed();
1548 /* we have to stop all cpus to guaranntee there is no user
1550 stop_machine_run(__build_all_zonelists, NULL, NR_CPUS);
1551 /* cpuset refresh routine should be here */
1553 vm_total_pages = nr_free_pagecache_pages();
1554 printk("Built %i zonelists. Total pages: %ld\n",
1555 num_online_nodes(), vm_total_pages);
1559 * Helper functions to size the waitqueue hash table.
1560 * Essentially these want to choose hash table sizes sufficiently
1561 * large so that collisions trying to wait on pages are rare.
1562 * But in fact, the number of active page waitqueues on typical
1563 * systems is ridiculously low, less than 200. So this is even
1564 * conservative, even though it seems large.
1566 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1567 * waitqueues, i.e. the size of the waitq table given the number of pages.
1569 #define PAGES_PER_WAITQUEUE 256
1571 #ifndef CONFIG_MEMORY_HOTPLUG
1572 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
1574 unsigned long size = 1;
1576 pages /= PAGES_PER_WAITQUEUE;
1578 while (size < pages)
1582 * Once we have dozens or even hundreds of threads sleeping
1583 * on IO we've got bigger problems than wait queue collision.
1584 * Limit the size of the wait table to a reasonable size.
1586 size = min(size, 4096UL);
1588 return max(size, 4UL);
1592 * A zone's size might be changed by hot-add, so it is not possible to determine
1593 * a suitable size for its wait_table. So we use the maximum size now.
1595 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
1597 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
1598 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
1599 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
1601 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
1602 * or more by the traditional way. (See above). It equals:
1604 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
1605 * ia64(16K page size) : = ( 8G + 4M)byte.
1606 * powerpc (64K page size) : = (32G +16M)byte.
1608 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
1615 * This is an integer logarithm so that shifts can be used later
1616 * to extract the more random high bits from the multiplicative
1617 * hash function before the remainder is taken.
1619 static inline unsigned long wait_table_bits(unsigned long size)
1624 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1626 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1627 unsigned long *zones_size, unsigned long *zholes_size)
1629 unsigned long realtotalpages, totalpages = 0;
1632 for (i = 0; i < MAX_NR_ZONES; i++)
1633 totalpages += zones_size[i];
1634 pgdat->node_spanned_pages = totalpages;
1636 realtotalpages = totalpages;
1638 for (i = 0; i < MAX_NR_ZONES; i++)
1639 realtotalpages -= zholes_size[i];
1640 pgdat->node_present_pages = realtotalpages;
1641 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1646 * Initially all pages are reserved - free ones are freed
1647 * up by free_all_bootmem() once the early boot process is
1648 * done. Non-atomic initialization, single-pass.
1650 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1651 unsigned long start_pfn)
1654 unsigned long end_pfn = start_pfn + size;
1657 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1658 if (!early_pfn_valid(pfn))
1660 page = pfn_to_page(pfn);
1661 set_page_links(page, zone, nid, pfn);
1662 init_page_count(page);
1663 reset_page_mapcount(page);
1664 SetPageReserved(page);
1665 INIT_LIST_HEAD(&page->lru);
1666 #ifdef WANT_PAGE_VIRTUAL
1667 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1668 if (!is_highmem_idx(zone))
1669 set_page_address(page, __va(pfn << PAGE_SHIFT));
1674 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1678 for (order = 0; order < MAX_ORDER ; order++) {
1679 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1680 zone->free_area[order].nr_free = 0;
1684 #define ZONETABLE_INDEX(x, zone_nr) ((x << ZONES_SHIFT) | zone_nr)
1685 void zonetable_add(struct zone *zone, int nid, enum zone_type zid,
1686 unsigned long pfn, unsigned long size)
1688 unsigned long snum = pfn_to_section_nr(pfn);
1689 unsigned long end = pfn_to_section_nr(pfn + size);
1692 zone_table[ZONETABLE_INDEX(nid, zid)] = zone;
1694 for (; snum <= end; snum++)
1695 zone_table[ZONETABLE_INDEX(snum, zid)] = zone;
1698 #ifndef __HAVE_ARCH_MEMMAP_INIT
1699 #define memmap_init(size, nid, zone, start_pfn) \
1700 memmap_init_zone((size), (nid), (zone), (start_pfn))
1703 static int __cpuinit zone_batchsize(struct zone *zone)
1708 * The per-cpu-pages pools are set to around 1000th of the
1709 * size of the zone. But no more than 1/2 of a meg.
1711 * OK, so we don't know how big the cache is. So guess.
1713 batch = zone->present_pages / 1024;
1714 if (batch * PAGE_SIZE > 512 * 1024)
1715 batch = (512 * 1024) / PAGE_SIZE;
1716 batch /= 4; /* We effectively *= 4 below */
1721 * Clamp the batch to a 2^n - 1 value. Having a power
1722 * of 2 value was found to be more likely to have
1723 * suboptimal cache aliasing properties in some cases.
1725 * For example if 2 tasks are alternately allocating
1726 * batches of pages, one task can end up with a lot
1727 * of pages of one half of the possible page colors
1728 * and the other with pages of the other colors.
1730 batch = (1 << (fls(batch + batch/2)-1)) - 1;
1735 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
1737 struct per_cpu_pages *pcp;
1739 memset(p, 0, sizeof(*p));
1741 pcp = &p->pcp[0]; /* hot */
1743 pcp->high = 6 * batch;
1744 pcp->batch = max(1UL, 1 * batch);
1745 INIT_LIST_HEAD(&pcp->list);
1747 pcp = &p->pcp[1]; /* cold*/
1749 pcp->high = 2 * batch;
1750 pcp->batch = max(1UL, batch/2);
1751 INIT_LIST_HEAD(&pcp->list);
1755 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
1756 * to the value high for the pageset p.
1759 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
1762 struct per_cpu_pages *pcp;
1764 pcp = &p->pcp[0]; /* hot list */
1766 pcp->batch = max(1UL, high/4);
1767 if ((high/4) > (PAGE_SHIFT * 8))
1768 pcp->batch = PAGE_SHIFT * 8;
1774 * Boot pageset table. One per cpu which is going to be used for all
1775 * zones and all nodes. The parameters will be set in such a way
1776 * that an item put on a list will immediately be handed over to
1777 * the buddy list. This is safe since pageset manipulation is done
1778 * with interrupts disabled.
1780 * Some NUMA counter updates may also be caught by the boot pagesets.
1782 * The boot_pagesets must be kept even after bootup is complete for
1783 * unused processors and/or zones. They do play a role for bootstrapping
1784 * hotplugged processors.
1786 * zoneinfo_show() and maybe other functions do
1787 * not check if the processor is online before following the pageset pointer.
1788 * Other parts of the kernel may not check if the zone is available.
1790 static struct per_cpu_pageset boot_pageset[NR_CPUS];
1793 * Dynamically allocate memory for the
1794 * per cpu pageset array in struct zone.
1796 static int __cpuinit process_zones(int cpu)
1798 struct zone *zone, *dzone;
1800 for_each_zone(zone) {
1802 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
1803 GFP_KERNEL, cpu_to_node(cpu));
1804 if (!zone_pcp(zone, cpu))
1807 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
1809 if (percpu_pagelist_fraction)
1810 setup_pagelist_highmark(zone_pcp(zone, cpu),
1811 (zone->present_pages / percpu_pagelist_fraction));
1816 for_each_zone(dzone) {
1819 kfree(zone_pcp(dzone, cpu));
1820 zone_pcp(dzone, cpu) = NULL;
1825 static inline void free_zone_pagesets(int cpu)
1829 for_each_zone(zone) {
1830 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
1832 /* Free per_cpu_pageset if it is slab allocated */
1833 if (pset != &boot_pageset[cpu])
1835 zone_pcp(zone, cpu) = NULL;
1839 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
1840 unsigned long action,
1843 int cpu = (long)hcpu;
1844 int ret = NOTIFY_OK;
1847 case CPU_UP_PREPARE:
1848 if (process_zones(cpu))
1851 case CPU_UP_CANCELED:
1853 free_zone_pagesets(cpu);
1861 static struct notifier_block __cpuinitdata pageset_notifier =
1862 { &pageset_cpuup_callback, NULL, 0 };
1864 void __init setup_per_cpu_pageset(void)
1868 /* Initialize per_cpu_pageset for cpu 0.
1869 * A cpuup callback will do this for every cpu
1870 * as it comes online
1872 err = process_zones(smp_processor_id());
1874 register_cpu_notifier(&pageset_notifier);
1880 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
1883 struct pglist_data *pgdat = zone->zone_pgdat;
1887 * The per-page waitqueue mechanism uses hashed waitqueues
1890 zone->wait_table_hash_nr_entries =
1891 wait_table_hash_nr_entries(zone_size_pages);
1892 zone->wait_table_bits =
1893 wait_table_bits(zone->wait_table_hash_nr_entries);
1894 alloc_size = zone->wait_table_hash_nr_entries
1895 * sizeof(wait_queue_head_t);
1897 if (system_state == SYSTEM_BOOTING) {
1898 zone->wait_table = (wait_queue_head_t *)
1899 alloc_bootmem_node(pgdat, alloc_size);
1902 * This case means that a zone whose size was 0 gets new memory
1903 * via memory hot-add.
1904 * But it may be the case that a new node was hot-added. In
1905 * this case vmalloc() will not be able to use this new node's
1906 * memory - this wait_table must be initialized to use this new
1907 * node itself as well.
1908 * To use this new node's memory, further consideration will be
1911 zone->wait_table = (wait_queue_head_t *)vmalloc(alloc_size);
1913 if (!zone->wait_table)
1916 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
1917 init_waitqueue_head(zone->wait_table + i);
1922 static __meminit void zone_pcp_init(struct zone *zone)
1925 unsigned long batch = zone_batchsize(zone);
1927 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1929 /* Early boot. Slab allocator not functional yet */
1930 zone_pcp(zone, cpu) = &boot_pageset[cpu];
1931 setup_pageset(&boot_pageset[cpu],0);
1933 setup_pageset(zone_pcp(zone,cpu), batch);
1936 if (zone->present_pages)
1937 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
1938 zone->name, zone->present_pages, batch);
1941 __meminit int init_currently_empty_zone(struct zone *zone,
1942 unsigned long zone_start_pfn,
1945 struct pglist_data *pgdat = zone->zone_pgdat;
1947 ret = zone_wait_table_init(zone, size);
1950 pgdat->nr_zones = zone_idx(zone) + 1;
1952 zone->zone_start_pfn = zone_start_pfn;
1954 memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn);
1956 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1962 * Set up the zone data structures:
1963 * - mark all pages reserved
1964 * - mark all memory queues empty
1965 * - clear the memory bitmaps
1967 static void __meminit free_area_init_core(struct pglist_data *pgdat,
1968 unsigned long *zones_size, unsigned long *zholes_size)
1971 int nid = pgdat->node_id;
1972 unsigned long zone_start_pfn = pgdat->node_start_pfn;
1975 pgdat_resize_init(pgdat);
1976 pgdat->nr_zones = 0;
1977 init_waitqueue_head(&pgdat->kswapd_wait);
1978 pgdat->kswapd_max_order = 0;
1980 for (j = 0; j < MAX_NR_ZONES; j++) {
1981 struct zone *zone = pgdat->node_zones + j;
1982 unsigned long size, realsize;
1984 realsize = size = zones_size[j];
1986 realsize -= zholes_size[j];
1988 if (!is_highmem_idx(j))
1989 nr_kernel_pages += realsize;
1990 nr_all_pages += realsize;
1992 zone->spanned_pages = size;
1993 zone->present_pages = realsize;
1995 zone->min_unmapped_ratio = (realsize*sysctl_min_unmapped_ratio)
1998 zone->name = zone_names[j];
1999 spin_lock_init(&zone->lock);
2000 spin_lock_init(&zone->lru_lock);
2001 zone_seqlock_init(zone);
2002 zone->zone_pgdat = pgdat;
2003 zone->free_pages = 0;
2005 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
2007 zone_pcp_init(zone);
2008 INIT_LIST_HEAD(&zone->active_list);
2009 INIT_LIST_HEAD(&zone->inactive_list);
2010 zone->nr_scan_active = 0;
2011 zone->nr_scan_inactive = 0;
2012 zone->nr_active = 0;
2013 zone->nr_inactive = 0;
2014 zap_zone_vm_stats(zone);
2015 atomic_set(&zone->reclaim_in_progress, 0);
2019 zonetable_add(zone, nid, j, zone_start_pfn, size);
2020 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
2022 zone_start_pfn += size;
2026 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
2028 /* Skip empty nodes */
2029 if (!pgdat->node_spanned_pages)
2032 #ifdef CONFIG_FLAT_NODE_MEM_MAP
2033 /* ia64 gets its own node_mem_map, before this, without bootmem */
2034 if (!pgdat->node_mem_map) {
2035 unsigned long size, start, end;
2039 * The zone's endpoints aren't required to be MAX_ORDER
2040 * aligned but the node_mem_map endpoints must be in order
2041 * for the buddy allocator to function correctly.
2043 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
2044 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
2045 end = ALIGN(end, MAX_ORDER_NR_PAGES);
2046 size = (end - start) * sizeof(struct page);
2047 map = alloc_remap(pgdat->node_id, size);
2049 map = alloc_bootmem_node(pgdat, size);
2050 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
2052 #ifdef CONFIG_FLATMEM
2054 * With no DISCONTIG, the global mem_map is just set as node 0's
2056 if (pgdat == NODE_DATA(0))
2057 mem_map = NODE_DATA(0)->node_mem_map;
2059 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
2062 void __meminit free_area_init_node(int nid, struct pglist_data *pgdat,
2063 unsigned long *zones_size, unsigned long node_start_pfn,
2064 unsigned long *zholes_size)
2066 pgdat->node_id = nid;
2067 pgdat->node_start_pfn = node_start_pfn;
2068 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
2070 alloc_node_mem_map(pgdat);
2072 free_area_init_core(pgdat, zones_size, zholes_size);
2075 #ifndef CONFIG_NEED_MULTIPLE_NODES
2076 static bootmem_data_t contig_bootmem_data;
2077 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
2079 EXPORT_SYMBOL(contig_page_data);
2082 void __init free_area_init(unsigned long *zones_size)
2084 free_area_init_node(0, NODE_DATA(0), zones_size,
2085 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
2088 #ifdef CONFIG_HOTPLUG_CPU
2089 static int page_alloc_cpu_notify(struct notifier_block *self,
2090 unsigned long action, void *hcpu)
2092 int cpu = (unsigned long)hcpu;
2094 if (action == CPU_DEAD) {
2095 local_irq_disable();
2097 vm_events_fold_cpu(cpu);
2099 refresh_cpu_vm_stats(cpu);
2103 #endif /* CONFIG_HOTPLUG_CPU */
2105 void __init page_alloc_init(void)
2107 hotcpu_notifier(page_alloc_cpu_notify, 0);
2111 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
2112 * or min_free_kbytes changes.
2114 static void calculate_totalreserve_pages(void)
2116 struct pglist_data *pgdat;
2117 unsigned long reserve_pages = 0;
2118 enum zone_type i, j;
2120 for_each_online_pgdat(pgdat) {
2121 for (i = 0; i < MAX_NR_ZONES; i++) {
2122 struct zone *zone = pgdat->node_zones + i;
2123 unsigned long max = 0;
2125 /* Find valid and maximum lowmem_reserve in the zone */
2126 for (j = i; j < MAX_NR_ZONES; j++) {
2127 if (zone->lowmem_reserve[j] > max)
2128 max = zone->lowmem_reserve[j];
2131 /* we treat pages_high as reserved pages. */
2132 max += zone->pages_high;
2134 if (max > zone->present_pages)
2135 max = zone->present_pages;
2136 reserve_pages += max;
2139 totalreserve_pages = reserve_pages;
2143 * setup_per_zone_lowmem_reserve - called whenever
2144 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
2145 * has a correct pages reserved value, so an adequate number of
2146 * pages are left in the zone after a successful __alloc_pages().
2148 static void setup_per_zone_lowmem_reserve(void)
2150 struct pglist_data *pgdat;
2151 enum zone_type j, idx;
2153 for_each_online_pgdat(pgdat) {
2154 for (j = 0; j < MAX_NR_ZONES; j++) {
2155 struct zone *zone = pgdat->node_zones + j;
2156 unsigned long present_pages = zone->present_pages;
2158 zone->lowmem_reserve[j] = 0;
2162 struct zone *lower_zone;
2166 if (sysctl_lowmem_reserve_ratio[idx] < 1)
2167 sysctl_lowmem_reserve_ratio[idx] = 1;
2169 lower_zone = pgdat->node_zones + idx;
2170 lower_zone->lowmem_reserve[j] = present_pages /
2171 sysctl_lowmem_reserve_ratio[idx];
2172 present_pages += lower_zone->present_pages;
2177 /* update totalreserve_pages */
2178 calculate_totalreserve_pages();
2182 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
2183 * that the pages_{min,low,high} values for each zone are set correctly
2184 * with respect to min_free_kbytes.
2186 void setup_per_zone_pages_min(void)
2188 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
2189 unsigned long lowmem_pages = 0;
2191 unsigned long flags;
2193 /* Calculate total number of !ZONE_HIGHMEM pages */
2194 for_each_zone(zone) {
2195 if (!is_highmem(zone))
2196 lowmem_pages += zone->present_pages;
2199 for_each_zone(zone) {
2202 spin_lock_irqsave(&zone->lru_lock, flags);
2203 tmp = (u64)pages_min * zone->present_pages;
2204 do_div(tmp, lowmem_pages);
2205 if (is_highmem(zone)) {
2207 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
2208 * need highmem pages, so cap pages_min to a small
2211 * The (pages_high-pages_low) and (pages_low-pages_min)
2212 * deltas controls asynch page reclaim, and so should
2213 * not be capped for highmem.
2217 min_pages = zone->present_pages / 1024;
2218 if (min_pages < SWAP_CLUSTER_MAX)
2219 min_pages = SWAP_CLUSTER_MAX;
2220 if (min_pages > 128)
2222 zone->pages_min = min_pages;
2225 * If it's a lowmem zone, reserve a number of pages
2226 * proportionate to the zone's size.
2228 zone->pages_min = tmp;
2231 zone->pages_low = zone->pages_min + (tmp >> 2);
2232 zone->pages_high = zone->pages_min + (tmp >> 1);
2233 spin_unlock_irqrestore(&zone->lru_lock, flags);
2236 /* update totalreserve_pages */
2237 calculate_totalreserve_pages();
2241 * Initialise min_free_kbytes.
2243 * For small machines we want it small (128k min). For large machines
2244 * we want it large (64MB max). But it is not linear, because network
2245 * bandwidth does not increase linearly with machine size. We use
2247 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2248 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2264 static int __init init_per_zone_pages_min(void)
2266 unsigned long lowmem_kbytes;
2268 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2270 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2271 if (min_free_kbytes < 128)
2272 min_free_kbytes = 128;
2273 if (min_free_kbytes > 65536)
2274 min_free_kbytes = 65536;
2275 setup_per_zone_pages_min();
2276 setup_per_zone_lowmem_reserve();
2279 module_init(init_per_zone_pages_min)
2282 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2283 * that we can call two helper functions whenever min_free_kbytes
2286 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2287 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2289 proc_dointvec(table, write, file, buffer, length, ppos);
2290 setup_per_zone_pages_min();
2295 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
2296 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2301 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2306 zone->min_unmapped_ratio = (zone->present_pages *
2307 sysctl_min_unmapped_ratio) / 100;
2313 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2314 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2315 * whenever sysctl_lowmem_reserve_ratio changes.
2317 * The reserve ratio obviously has absolutely no relation with the
2318 * pages_min watermarks. The lowmem reserve ratio can only make sense
2319 * if in function of the boot time zone sizes.
2321 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
2322 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2324 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2325 setup_per_zone_lowmem_reserve();
2330 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
2331 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
2332 * can have before it gets flushed back to buddy allocator.
2335 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
2336 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2342 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2343 if (!write || (ret == -EINVAL))
2345 for_each_zone(zone) {
2346 for_each_online_cpu(cpu) {
2348 high = zone->present_pages / percpu_pagelist_fraction;
2349 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
2355 int hashdist = HASHDIST_DEFAULT;
2358 static int __init set_hashdist(char *str)
2362 hashdist = simple_strtoul(str, &str, 0);
2365 __setup("hashdist=", set_hashdist);
2369 * allocate a large system hash table from bootmem
2370 * - it is assumed that the hash table must contain an exact power-of-2
2371 * quantity of entries
2372 * - limit is the number of hash buckets, not the total allocation size
2374 void *__init alloc_large_system_hash(const char *tablename,
2375 unsigned long bucketsize,
2376 unsigned long numentries,
2379 unsigned int *_hash_shift,
2380 unsigned int *_hash_mask,
2381 unsigned long limit)
2383 unsigned long long max = limit;
2384 unsigned long log2qty, size;
2387 /* allow the kernel cmdline to have a say */
2389 /* round applicable memory size up to nearest megabyte */
2390 numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages;
2391 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2392 numentries >>= 20 - PAGE_SHIFT;
2393 numentries <<= 20 - PAGE_SHIFT;
2395 /* limit to 1 bucket per 2^scale bytes of low memory */
2396 if (scale > PAGE_SHIFT)
2397 numentries >>= (scale - PAGE_SHIFT);
2399 numentries <<= (PAGE_SHIFT - scale);
2401 numentries = roundup_pow_of_two(numentries);
2403 /* limit allocation size to 1/16 total memory by default */
2405 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2406 do_div(max, bucketsize);
2409 if (numentries > max)
2412 log2qty = long_log2(numentries);
2415 size = bucketsize << log2qty;
2416 if (flags & HASH_EARLY)
2417 table = alloc_bootmem(size);
2419 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
2421 unsigned long order;
2422 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
2424 table = (void*) __get_free_pages(GFP_ATOMIC, order);
2426 } while (!table && size > PAGE_SIZE && --log2qty);
2429 panic("Failed to allocate %s hash table\n", tablename);
2431 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2434 long_log2(size) - PAGE_SHIFT,
2438 *_hash_shift = log2qty;
2440 *_hash_mask = (1 << log2qty) - 1;
2445 #ifdef CONFIG_OUT_OF_LINE_PFN_TO_PAGE
2446 struct page *pfn_to_page(unsigned long pfn)
2448 return __pfn_to_page(pfn);
2450 unsigned long page_to_pfn(struct page *page)
2452 return __page_to_pfn(page);
2454 EXPORT_SYMBOL(pfn_to_page);
2455 EXPORT_SYMBOL(page_to_pfn);
2456 #endif /* CONFIG_OUT_OF_LINE_PFN_TO_PAGE */