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 totalhigh_pages __read_mostly;
55 unsigned long totalreserve_pages __read_mostly;
57 int percpu_pagelist_fraction;
59 static void __free_pages_ok(struct page *page, unsigned int order);
62 * results with 256, 32 in the lowmem_reserve sysctl:
63 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
64 * 1G machine -> (16M dma, 784M normal, 224M high)
65 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
66 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
67 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
69 * TBD: should special case ZONE_DMA32 machines here - in those we normally
70 * don't need any ZONE_NORMAL reservation
72 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 256, 256, 32 };
74 EXPORT_SYMBOL(totalram_pages);
77 * Used by page_zone() to look up the address of the struct zone whose
78 * id is encoded in the upper bits of page->flags
80 struct zone *zone_table[1 << ZONETABLE_SHIFT] __read_mostly;
81 EXPORT_SYMBOL(zone_table);
83 static char *zone_names[MAX_NR_ZONES] = { "DMA", "DMA32", "Normal", "HighMem" };
84 int min_free_kbytes = 1024;
86 unsigned long __meminitdata nr_kernel_pages;
87 unsigned long __meminitdata nr_all_pages;
89 #ifdef CONFIG_DEBUG_VM
90 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
94 unsigned long pfn = page_to_pfn(page);
97 seq = zone_span_seqbegin(zone);
98 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
100 else if (pfn < zone->zone_start_pfn)
102 } while (zone_span_seqretry(zone, seq));
107 static int page_is_consistent(struct zone *zone, struct page *page)
109 #ifdef CONFIG_HOLES_IN_ZONE
110 if (!pfn_valid(page_to_pfn(page)))
113 if (zone != page_zone(page))
119 * Temporary debugging check for pages not lying within a given zone.
121 static int bad_range(struct zone *zone, struct page *page)
123 if (page_outside_zone_boundaries(zone, page))
125 if (!page_is_consistent(zone, page))
131 static inline int bad_range(struct zone *zone, struct page *page)
137 static void bad_page(struct page *page)
139 printk(KERN_EMERG "Bad page state in process '%s'\n"
140 KERN_EMERG "page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d\n"
141 KERN_EMERG "Trying to fix it up, but a reboot is needed\n"
142 KERN_EMERG "Backtrace:\n",
143 current->comm, page, (int)(2*sizeof(unsigned long)),
144 (unsigned long)page->flags, page->mapping,
145 page_mapcount(page), page_count(page));
147 page->flags &= ~(1 << PG_lru |
157 set_page_count(page, 0);
158 reset_page_mapcount(page);
159 page->mapping = NULL;
160 add_taint(TAINT_BAD_PAGE);
164 * Higher-order pages are called "compound pages". They are structured thusly:
166 * The first PAGE_SIZE page is called the "head page".
168 * The remaining PAGE_SIZE pages are called "tail pages".
170 * All pages have PG_compound set. All pages have their ->private pointing at
171 * the head page (even the head page has this).
173 * The first tail page's ->lru.next holds the address of the compound page's
174 * put_page() function. Its ->lru.prev holds the order of allocation.
175 * This usage means that zero-order pages may not be compound.
178 static void free_compound_page(struct page *page)
180 __free_pages_ok(page, (unsigned long)page[1].lru.prev);
183 static void prep_compound_page(struct page *page, unsigned long order)
186 int nr_pages = 1 << order;
188 page[1].lru.next = (void *)free_compound_page; /* set dtor */
189 page[1].lru.prev = (void *)order;
190 for (i = 0; i < nr_pages; i++) {
191 struct page *p = page + i;
193 __SetPageCompound(p);
194 set_page_private(p, (unsigned long)page);
198 static void destroy_compound_page(struct page *page, unsigned long order)
201 int nr_pages = 1 << order;
203 if (unlikely((unsigned long)page[1].lru.prev != order))
206 for (i = 0; i < nr_pages; i++) {
207 struct page *p = page + i;
209 if (unlikely(!PageCompound(p) |
210 (page_private(p) != (unsigned long)page)))
212 __ClearPageCompound(p);
216 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
220 VM_BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
222 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
223 * and __GFP_HIGHMEM from hard or soft interrupt context.
225 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
226 for (i = 0; i < (1 << order); i++)
227 clear_highpage(page + i);
231 * function for dealing with page's order in buddy system.
232 * zone->lock is already acquired when we use these.
233 * So, we don't need atomic page->flags operations here.
235 static inline unsigned long page_order(struct page *page)
237 return page_private(page);
240 static inline void set_page_order(struct page *page, int order)
242 set_page_private(page, order);
243 __SetPageBuddy(page);
246 static inline void rmv_page_order(struct page *page)
248 __ClearPageBuddy(page);
249 set_page_private(page, 0);
253 * Locate the struct page for both the matching buddy in our
254 * pair (buddy1) and the combined O(n+1) page they form (page).
256 * 1) Any buddy B1 will have an order O twin B2 which satisfies
257 * the following equation:
259 * For example, if the starting buddy (buddy2) is #8 its order
261 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
263 * 2) Any buddy B will have an order O+1 parent P which
264 * satisfies the following equation:
267 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
269 static inline struct page *
270 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
272 unsigned long buddy_idx = page_idx ^ (1 << order);
274 return page + (buddy_idx - page_idx);
277 static inline unsigned long
278 __find_combined_index(unsigned long page_idx, unsigned int order)
280 return (page_idx & ~(1 << order));
284 * This function checks whether a page is free && is the buddy
285 * we can do coalesce a page and its buddy if
286 * (a) the buddy is not in a hole &&
287 * (b) the buddy is in the buddy system &&
288 * (c) a page and its buddy have the same order &&
289 * (d) a page and its buddy are in the same zone.
291 * For recording whether a page is in the buddy system, we use PG_buddy.
292 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
294 * For recording page's order, we use page_private(page).
296 static inline int page_is_buddy(struct page *page, struct page *buddy,
299 #ifdef CONFIG_HOLES_IN_ZONE
300 if (!pfn_valid(page_to_pfn(buddy)))
304 if (page_zone_id(page) != page_zone_id(buddy))
307 if (PageBuddy(buddy) && page_order(buddy) == order) {
308 BUG_ON(page_count(buddy) != 0);
315 * Freeing function for a buddy system allocator.
317 * The concept of a buddy system is to maintain direct-mapped table
318 * (containing bit values) for memory blocks of various "orders".
319 * The bottom level table contains the map for the smallest allocatable
320 * units of memory (here, pages), and each level above it describes
321 * pairs of units from the levels below, hence, "buddies".
322 * At a high level, all that happens here is marking the table entry
323 * at the bottom level available, and propagating the changes upward
324 * as necessary, plus some accounting needed to play nicely with other
325 * parts of the VM system.
326 * At each level, we keep a list of pages, which are heads of continuous
327 * free pages of length of (1 << order) and marked with PG_buddy. Page's
328 * order is recorded in page_private(page) field.
329 * So when we are allocating or freeing one, we can derive the state of the
330 * other. That is, if we allocate a small block, and both were
331 * free, the remainder of the region must be split into blocks.
332 * If a block is freed, and its buddy is also free, then this
333 * triggers coalescing into a block of larger size.
338 static inline void __free_one_page(struct page *page,
339 struct zone *zone, unsigned int order)
341 unsigned long page_idx;
342 int order_size = 1 << order;
344 if (unlikely(PageCompound(page)))
345 destroy_compound_page(page, order);
347 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
349 VM_BUG_ON(page_idx & (order_size - 1));
350 VM_BUG_ON(bad_range(zone, page));
352 zone->free_pages += order_size;
353 while (order < MAX_ORDER-1) {
354 unsigned long combined_idx;
355 struct free_area *area;
358 buddy = __page_find_buddy(page, page_idx, order);
359 if (!page_is_buddy(page, buddy, order))
360 break; /* Move the buddy up one level. */
362 list_del(&buddy->lru);
363 area = zone->free_area + order;
365 rmv_page_order(buddy);
366 combined_idx = __find_combined_index(page_idx, order);
367 page = page + (combined_idx - page_idx);
368 page_idx = combined_idx;
371 set_page_order(page, order);
372 list_add(&page->lru, &zone->free_area[order].free_list);
373 zone->free_area[order].nr_free++;
376 static inline int free_pages_check(struct page *page)
378 if (unlikely(page_mapcount(page) |
379 (page->mapping != NULL) |
380 (page_count(page) != 0) |
394 __ClearPageDirty(page);
396 * For now, we report if PG_reserved was found set, but do not
397 * clear it, and do not free the page. But we shall soon need
398 * to do more, for when the ZERO_PAGE count wraps negative.
400 return PageReserved(page);
404 * Frees a list of pages.
405 * Assumes all pages on list are in same zone, and of same order.
406 * count is the number of pages to free.
408 * If the zone was previously in an "all pages pinned" state then look to
409 * see if this freeing clears that state.
411 * And clear the zone's pages_scanned counter, to hold off the "all pages are
412 * pinned" detection logic.
414 static void free_pages_bulk(struct zone *zone, int count,
415 struct list_head *list, int order)
417 spin_lock(&zone->lock);
418 zone->all_unreclaimable = 0;
419 zone->pages_scanned = 0;
423 VM_BUG_ON(list_empty(list));
424 page = list_entry(list->prev, struct page, lru);
425 /* have to delete it as __free_one_page list manipulates */
426 list_del(&page->lru);
427 __free_one_page(page, zone, order);
429 spin_unlock(&zone->lock);
432 static void free_one_page(struct zone *zone, struct page *page, int order)
435 list_add(&page->lru, &list);
436 free_pages_bulk(zone, 1, &list, order);
439 static void __free_pages_ok(struct page *page, unsigned int order)
445 arch_free_page(page, order);
446 if (!PageHighMem(page))
447 debug_check_no_locks_freed(page_address(page),
450 for (i = 0 ; i < (1 << order) ; ++i)
451 reserved += free_pages_check(page + i);
455 kernel_map_pages(page, 1 << order, 0);
456 local_irq_save(flags);
457 __count_vm_events(PGFREE, 1 << order);
458 free_one_page(page_zone(page), page, order);
459 local_irq_restore(flags);
463 * permit the bootmem allocator to evade page validation on high-order frees
465 void fastcall __init __free_pages_bootmem(struct page *page, unsigned int order)
468 __ClearPageReserved(page);
469 set_page_count(page, 0);
470 set_page_refcounted(page);
476 for (loop = 0; loop < BITS_PER_LONG; loop++) {
477 struct page *p = &page[loop];
479 if (loop + 1 < BITS_PER_LONG)
481 __ClearPageReserved(p);
482 set_page_count(p, 0);
485 set_page_refcounted(page);
486 __free_pages(page, order);
492 * The order of subdivision here is critical for the IO subsystem.
493 * Please do not alter this order without good reasons and regression
494 * testing. Specifically, as large blocks of memory are subdivided,
495 * the order in which smaller blocks are delivered depends on the order
496 * they're subdivided in this function. This is the primary factor
497 * influencing the order in which pages are delivered to the IO
498 * subsystem according to empirical testing, and this is also justified
499 * by considering the behavior of a buddy system containing a single
500 * large block of memory acted on by a series of small allocations.
501 * This behavior is a critical factor in sglist merging's success.
505 static inline void expand(struct zone *zone, struct page *page,
506 int low, int high, struct free_area *area)
508 unsigned long size = 1 << high;
514 VM_BUG_ON(bad_range(zone, &page[size]));
515 list_add(&page[size].lru, &area->free_list);
517 set_page_order(&page[size], high);
522 * This page is about to be returned from the page allocator
524 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
526 if (unlikely(page_mapcount(page) |
527 (page->mapping != NULL) |
528 (page_count(page) != 0) |
544 * For now, we report if PG_reserved was found set, but do not
545 * clear it, and do not allocate the page: as a safety net.
547 if (PageReserved(page))
550 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
551 1 << PG_referenced | 1 << PG_arch_1 |
552 1 << PG_checked | 1 << PG_mappedtodisk);
553 set_page_private(page, 0);
554 set_page_refcounted(page);
555 kernel_map_pages(page, 1 << order, 1);
557 if (gfp_flags & __GFP_ZERO)
558 prep_zero_page(page, order, gfp_flags);
560 if (order && (gfp_flags & __GFP_COMP))
561 prep_compound_page(page, order);
567 * Do the hard work of removing an element from the buddy allocator.
568 * Call me with the zone->lock already held.
570 static struct page *__rmqueue(struct zone *zone, unsigned int order)
572 struct free_area * area;
573 unsigned int current_order;
576 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
577 area = zone->free_area + current_order;
578 if (list_empty(&area->free_list))
581 page = list_entry(area->free_list.next, struct page, lru);
582 list_del(&page->lru);
583 rmv_page_order(page);
585 zone->free_pages -= 1UL << order;
586 expand(zone, page, order, current_order, area);
594 * Obtain a specified number of elements from the buddy allocator, all under
595 * a single hold of the lock, for efficiency. Add them to the supplied list.
596 * Returns the number of new pages which were placed at *list.
598 static int rmqueue_bulk(struct zone *zone, unsigned int order,
599 unsigned long count, struct list_head *list)
603 spin_lock(&zone->lock);
604 for (i = 0; i < count; ++i) {
605 struct page *page = __rmqueue(zone, order);
606 if (unlikely(page == NULL))
608 list_add_tail(&page->lru, list);
610 spin_unlock(&zone->lock);
616 * Called from the slab reaper to drain pagesets on a particular node that
617 * belong to the currently executing processor.
618 * Note that this function must be called with the thread pinned to
619 * a single processor.
621 void drain_node_pages(int nodeid)
626 for (z = 0; z < MAX_NR_ZONES; z++) {
627 struct zone *zone = NODE_DATA(nodeid)->node_zones + z;
628 struct per_cpu_pageset *pset;
630 pset = zone_pcp(zone, smp_processor_id());
631 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
632 struct per_cpu_pages *pcp;
636 local_irq_save(flags);
637 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
639 local_irq_restore(flags);
646 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
647 static void __drain_pages(unsigned int cpu)
653 for_each_zone(zone) {
654 struct per_cpu_pageset *pset;
656 pset = zone_pcp(zone, cpu);
657 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
658 struct per_cpu_pages *pcp;
661 local_irq_save(flags);
662 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
664 local_irq_restore(flags);
668 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
672 void mark_free_pages(struct zone *zone)
674 unsigned long zone_pfn, flags;
676 struct list_head *curr;
678 if (!zone->spanned_pages)
681 spin_lock_irqsave(&zone->lock, flags);
682 for (zone_pfn = 0; zone_pfn < zone->spanned_pages; ++zone_pfn)
683 ClearPageNosaveFree(pfn_to_page(zone_pfn + zone->zone_start_pfn));
685 for (order = MAX_ORDER - 1; order >= 0; --order)
686 list_for_each(curr, &zone->free_area[order].free_list) {
687 unsigned long start_pfn, i;
689 start_pfn = page_to_pfn(list_entry(curr, struct page, lru));
691 for (i=0; i < (1<<order); i++)
692 SetPageNosaveFree(pfn_to_page(start_pfn+i));
694 spin_unlock_irqrestore(&zone->lock, flags);
698 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
700 void drain_local_pages(void)
704 local_irq_save(flags);
705 __drain_pages(smp_processor_id());
706 local_irq_restore(flags);
708 #endif /* CONFIG_PM */
711 * Free a 0-order page
713 static void fastcall free_hot_cold_page(struct page *page, int cold)
715 struct zone *zone = page_zone(page);
716 struct per_cpu_pages *pcp;
719 arch_free_page(page, 0);
722 page->mapping = NULL;
723 if (free_pages_check(page))
726 kernel_map_pages(page, 1, 0);
728 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
729 local_irq_save(flags);
730 __count_vm_event(PGFREE);
731 list_add(&page->lru, &pcp->list);
733 if (pcp->count >= pcp->high) {
734 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
735 pcp->count -= pcp->batch;
737 local_irq_restore(flags);
741 void fastcall free_hot_page(struct page *page)
743 free_hot_cold_page(page, 0);
746 void fastcall free_cold_page(struct page *page)
748 free_hot_cold_page(page, 1);
752 * split_page takes a non-compound higher-order page, and splits it into
753 * n (1<<order) sub-pages: page[0..n]
754 * Each sub-page must be freed individually.
756 * Note: this is probably too low level an operation for use in drivers.
757 * Please consult with lkml before using this in your driver.
759 void split_page(struct page *page, unsigned int order)
763 VM_BUG_ON(PageCompound(page));
764 VM_BUG_ON(!page_count(page));
765 for (i = 1; i < (1 << order); i++)
766 set_page_refcounted(page + i);
770 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
771 * we cheat by calling it from here, in the order > 0 path. Saves a branch
774 static struct page *buffered_rmqueue(struct zonelist *zonelist,
775 struct zone *zone, int order, gfp_t gfp_flags)
779 int cold = !!(gfp_flags & __GFP_COLD);
784 if (likely(order == 0)) {
785 struct per_cpu_pages *pcp;
787 pcp = &zone_pcp(zone, cpu)->pcp[cold];
788 local_irq_save(flags);
790 pcp->count += rmqueue_bulk(zone, 0,
791 pcp->batch, &pcp->list);
792 if (unlikely(!pcp->count))
795 page = list_entry(pcp->list.next, struct page, lru);
796 list_del(&page->lru);
799 spin_lock_irqsave(&zone->lock, flags);
800 page = __rmqueue(zone, order);
801 spin_unlock(&zone->lock);
806 __count_zone_vm_events(PGALLOC, zone, 1 << order);
807 zone_statistics(zonelist, zone);
808 local_irq_restore(flags);
811 VM_BUG_ON(bad_range(zone, page));
812 if (prep_new_page(page, order, gfp_flags))
817 local_irq_restore(flags);
822 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
823 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
824 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
825 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
826 #define ALLOC_HARDER 0x10 /* try to alloc harder */
827 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
828 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
831 * Return 1 if free pages are above 'mark'. This takes into account the order
834 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
835 int classzone_idx, int alloc_flags)
837 /* free_pages my go negative - that's OK */
838 long min = mark, free_pages = z->free_pages - (1 << order) + 1;
841 if (alloc_flags & ALLOC_HIGH)
843 if (alloc_flags & ALLOC_HARDER)
846 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
848 for (o = 0; o < order; o++) {
849 /* At the next order, this order's pages become unavailable */
850 free_pages -= z->free_area[o].nr_free << o;
852 /* Require fewer higher order pages to be free */
855 if (free_pages <= min)
862 * get_page_from_freeliest goes through the zonelist trying to allocate
866 get_page_from_freelist(gfp_t gfp_mask, unsigned int order,
867 struct zonelist *zonelist, int alloc_flags)
869 struct zone **z = zonelist->zones;
870 struct page *page = NULL;
871 int classzone_idx = zone_idx(*z);
874 * Go through the zonelist once, looking for a zone with enough free.
875 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
878 if ((alloc_flags & ALLOC_CPUSET) &&
879 !cpuset_zone_allowed(*z, gfp_mask))
882 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
884 if (alloc_flags & ALLOC_WMARK_MIN)
885 mark = (*z)->pages_min;
886 else if (alloc_flags & ALLOC_WMARK_LOW)
887 mark = (*z)->pages_low;
889 mark = (*z)->pages_high;
890 if (!zone_watermark_ok(*z, order, mark,
891 classzone_idx, alloc_flags))
892 if (!zone_reclaim_mode ||
893 !zone_reclaim(*z, gfp_mask, order))
897 page = buffered_rmqueue(zonelist, *z, order, gfp_mask);
901 } while (*(++z) != NULL);
906 * This is the 'heart' of the zoned buddy allocator.
908 struct page * fastcall
909 __alloc_pages(gfp_t gfp_mask, unsigned int order,
910 struct zonelist *zonelist)
912 const gfp_t wait = gfp_mask & __GFP_WAIT;
915 struct reclaim_state reclaim_state;
916 struct task_struct *p = current;
919 int did_some_progress;
921 might_sleep_if(wait);
924 z = zonelist->zones; /* the list of zones suitable for gfp_mask */
926 if (unlikely(*z == NULL)) {
927 /* Should this ever happen?? */
931 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
932 zonelist, ALLOC_WMARK_LOW|ALLOC_CPUSET);
937 wakeup_kswapd(*z, order);
941 * OK, we're below the kswapd watermark and have kicked background
942 * reclaim. Now things get more complex, so set up alloc_flags according
943 * to how we want to proceed.
945 * The caller may dip into page reserves a bit more if the caller
946 * cannot run direct reclaim, or if the caller has realtime scheduling
947 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
948 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
950 alloc_flags = ALLOC_WMARK_MIN;
951 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
952 alloc_flags |= ALLOC_HARDER;
953 if (gfp_mask & __GFP_HIGH)
954 alloc_flags |= ALLOC_HIGH;
956 alloc_flags |= ALLOC_CPUSET;
959 * Go through the zonelist again. Let __GFP_HIGH and allocations
960 * coming from realtime tasks go deeper into reserves.
962 * This is the last chance, in general, before the goto nopage.
963 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
964 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
966 page = get_page_from_freelist(gfp_mask, order, zonelist, alloc_flags);
970 /* This allocation should allow future memory freeing. */
972 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
973 && !in_interrupt()) {
974 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
976 /* go through the zonelist yet again, ignoring mins */
977 page = get_page_from_freelist(gfp_mask, order,
978 zonelist, ALLOC_NO_WATERMARKS);
981 if (gfp_mask & __GFP_NOFAIL) {
982 blk_congestion_wait(WRITE, HZ/50);
989 /* Atomic allocations - we can't balance anything */
996 /* We now go into synchronous reclaim */
997 cpuset_memory_pressure_bump();
998 p->flags |= PF_MEMALLOC;
999 reclaim_state.reclaimed_slab = 0;
1000 p->reclaim_state = &reclaim_state;
1002 did_some_progress = try_to_free_pages(zonelist->zones, gfp_mask);
1004 p->reclaim_state = NULL;
1005 p->flags &= ~PF_MEMALLOC;
1009 if (likely(did_some_progress)) {
1010 page = get_page_from_freelist(gfp_mask, order,
1011 zonelist, alloc_flags);
1014 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1016 * Go through the zonelist yet one more time, keep
1017 * very high watermark here, this is only to catch
1018 * a parallel oom killing, we must fail if we're still
1019 * under heavy pressure.
1021 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1022 zonelist, ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1026 out_of_memory(zonelist, gfp_mask, order);
1031 * Don't let big-order allocations loop unless the caller explicitly
1032 * requests that. Wait for some write requests to complete then retry.
1034 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
1035 * <= 3, but that may not be true in other implementations.
1038 if (!(gfp_mask & __GFP_NORETRY)) {
1039 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
1041 if (gfp_mask & __GFP_NOFAIL)
1045 blk_congestion_wait(WRITE, HZ/50);
1050 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1051 printk(KERN_WARNING "%s: page allocation failure."
1052 " order:%d, mode:0x%x\n",
1053 p->comm, order, gfp_mask);
1061 EXPORT_SYMBOL(__alloc_pages);
1064 * Common helper functions.
1066 fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1069 page = alloc_pages(gfp_mask, order);
1072 return (unsigned long) page_address(page);
1075 EXPORT_SYMBOL(__get_free_pages);
1077 fastcall unsigned long get_zeroed_page(gfp_t gfp_mask)
1082 * get_zeroed_page() returns a 32-bit address, which cannot represent
1085 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1087 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1089 return (unsigned long) page_address(page);
1093 EXPORT_SYMBOL(get_zeroed_page);
1095 void __pagevec_free(struct pagevec *pvec)
1097 int i = pagevec_count(pvec);
1100 free_hot_cold_page(pvec->pages[i], pvec->cold);
1103 fastcall void __free_pages(struct page *page, unsigned int order)
1105 if (put_page_testzero(page)) {
1107 free_hot_page(page);
1109 __free_pages_ok(page, order);
1113 EXPORT_SYMBOL(__free_pages);
1115 fastcall void free_pages(unsigned long addr, unsigned int order)
1118 VM_BUG_ON(!virt_addr_valid((void *)addr));
1119 __free_pages(virt_to_page((void *)addr), order);
1123 EXPORT_SYMBOL(free_pages);
1126 * Total amount of free (allocatable) RAM:
1128 unsigned int nr_free_pages(void)
1130 unsigned int sum = 0;
1134 sum += zone->free_pages;
1139 EXPORT_SYMBOL(nr_free_pages);
1142 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
1144 unsigned int i, sum = 0;
1146 for (i = 0; i < MAX_NR_ZONES; i++)
1147 sum += pgdat->node_zones[i].free_pages;
1153 static unsigned int nr_free_zone_pages(int offset)
1155 /* Just pick one node, since fallback list is circular */
1156 pg_data_t *pgdat = NODE_DATA(numa_node_id());
1157 unsigned int sum = 0;
1159 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1160 struct zone **zonep = zonelist->zones;
1163 for (zone = *zonep++; zone; zone = *zonep++) {
1164 unsigned long size = zone->present_pages;
1165 unsigned long high = zone->pages_high;
1174 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1176 unsigned int nr_free_buffer_pages(void)
1178 return nr_free_zone_pages(gfp_zone(GFP_USER));
1182 * Amount of free RAM allocatable within all zones
1184 unsigned int nr_free_pagecache_pages(void)
1186 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER));
1189 #ifdef CONFIG_HIGHMEM
1190 unsigned int nr_free_highpages (void)
1193 unsigned int pages = 0;
1195 for_each_online_pgdat(pgdat)
1196 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1203 static void show_node(struct zone *zone)
1205 printk("Node %d ", zone->zone_pgdat->node_id);
1208 #define show_node(zone) do { } while (0)
1211 void si_meminfo(struct sysinfo *val)
1213 val->totalram = totalram_pages;
1215 val->freeram = nr_free_pages();
1216 val->bufferram = nr_blockdev_pages();
1217 #ifdef CONFIG_HIGHMEM
1218 val->totalhigh = totalhigh_pages;
1219 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 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1237 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1238 val->mem_unit = PAGE_SIZE;
1242 #define K(x) ((x) << (PAGE_SHIFT-10))
1245 * Show free area list (used inside shift_scroll-lock stuff)
1246 * We also calculate the percentage fragmentation. We do this by counting the
1247 * memory on each free list with the exception of the first item on the list.
1249 void show_free_areas(void)
1251 int cpu, temperature;
1252 unsigned long active;
1253 unsigned long inactive;
1257 for_each_zone(zone) {
1259 printk("%s per-cpu:", zone->name);
1261 if (!populated_zone(zone)) {
1267 for_each_online_cpu(cpu) {
1268 struct per_cpu_pageset *pageset;
1270 pageset = zone_pcp(zone, cpu);
1272 for (temperature = 0; temperature < 2; temperature++)
1273 printk("cpu %d %s: high %d, batch %d used:%d\n",
1275 temperature ? "cold" : "hot",
1276 pageset->pcp[temperature].high,
1277 pageset->pcp[temperature].batch,
1278 pageset->pcp[temperature].count);
1282 get_zone_counts(&active, &inactive, &free);
1284 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1285 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1288 global_page_state(NR_FILE_DIRTY),
1289 global_page_state(NR_WRITEBACK),
1290 global_page_state(NR_UNSTABLE_NFS),
1292 global_page_state(NR_SLAB),
1293 global_page_state(NR_FILE_MAPPED),
1294 global_page_state(NR_PAGETABLE));
1296 for_each_zone(zone) {
1308 " pages_scanned:%lu"
1309 " all_unreclaimable? %s"
1312 K(zone->free_pages),
1315 K(zone->pages_high),
1317 K(zone->nr_inactive),
1318 K(zone->present_pages),
1319 zone->pages_scanned,
1320 (zone->all_unreclaimable ? "yes" : "no")
1322 printk("lowmem_reserve[]:");
1323 for (i = 0; i < MAX_NR_ZONES; i++)
1324 printk(" %lu", zone->lowmem_reserve[i]);
1328 for_each_zone(zone) {
1329 unsigned long nr[MAX_ORDER], flags, order, total = 0;
1332 printk("%s: ", zone->name);
1333 if (!populated_zone(zone)) {
1338 spin_lock_irqsave(&zone->lock, flags);
1339 for (order = 0; order < MAX_ORDER; order++) {
1340 nr[order] = zone->free_area[order].nr_free;
1341 total += nr[order] << order;
1343 spin_unlock_irqrestore(&zone->lock, flags);
1344 for (order = 0; order < MAX_ORDER; order++)
1345 printk("%lu*%lukB ", nr[order], K(1UL) << order);
1346 printk("= %lukB\n", K(total));
1349 show_swap_cache_info();
1353 * Builds allocation fallback zone lists.
1355 * Add all populated zones of a node to the zonelist.
1357 static int __meminit build_zonelists_node(pg_data_t *pgdat,
1358 struct zonelist *zonelist, int nr_zones, int zone_type)
1362 BUG_ON(zone_type > ZONE_HIGHMEM);
1365 zone = pgdat->node_zones + zone_type;
1366 if (populated_zone(zone)) {
1367 #ifndef CONFIG_HIGHMEM
1368 BUG_ON(zone_type > ZONE_NORMAL);
1370 zonelist->zones[nr_zones++] = zone;
1371 check_highest_zone(zone_type);
1375 } while (zone_type >= 0);
1379 static inline int highest_zone(int zone_bits)
1381 int res = ZONE_NORMAL;
1382 if (zone_bits & (__force int)__GFP_HIGHMEM)
1384 if (zone_bits & (__force int)__GFP_DMA32)
1386 if (zone_bits & (__force int)__GFP_DMA)
1392 #define MAX_NODE_LOAD (num_online_nodes())
1393 static int __meminitdata node_load[MAX_NUMNODES];
1395 * find_next_best_node - find the next node that should appear in a given node's fallback list
1396 * @node: node whose fallback list we're appending
1397 * @used_node_mask: nodemask_t of already used nodes
1399 * We use a number of factors to determine which is the next node that should
1400 * appear on a given node's fallback list. The node should not have appeared
1401 * already in @node's fallback list, and it should be the next closest node
1402 * according to the distance array (which contains arbitrary distance values
1403 * from each node to each node in the system), and should also prefer nodes
1404 * with no CPUs, since presumably they'll have very little allocation pressure
1405 * on them otherwise.
1406 * It returns -1 if no node is found.
1408 static int __meminit find_next_best_node(int node, nodemask_t *used_node_mask)
1411 int min_val = INT_MAX;
1414 /* Use the local node if we haven't already */
1415 if (!node_isset(node, *used_node_mask)) {
1416 node_set(node, *used_node_mask);
1420 for_each_online_node(n) {
1423 /* Don't want a node to appear more than once */
1424 if (node_isset(n, *used_node_mask))
1427 /* Use the distance array to find the distance */
1428 val = node_distance(node, n);
1430 /* Penalize nodes under us ("prefer the next node") */
1433 /* Give preference to headless and unused nodes */
1434 tmp = node_to_cpumask(n);
1435 if (!cpus_empty(tmp))
1436 val += PENALTY_FOR_NODE_WITH_CPUS;
1438 /* Slight preference for less loaded node */
1439 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1440 val += node_load[n];
1442 if (val < min_val) {
1449 node_set(best_node, *used_node_mask);
1454 static void __meminit build_zonelists(pg_data_t *pgdat)
1456 int i, j, k, node, local_node;
1457 int prev_node, load;
1458 struct zonelist *zonelist;
1459 nodemask_t used_mask;
1461 /* initialize zonelists */
1462 for (i = 0; i < GFP_ZONETYPES; i++) {
1463 zonelist = pgdat->node_zonelists + i;
1464 zonelist->zones[0] = NULL;
1467 /* NUMA-aware ordering of nodes */
1468 local_node = pgdat->node_id;
1469 load = num_online_nodes();
1470 prev_node = local_node;
1471 nodes_clear(used_mask);
1472 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1473 int distance = node_distance(local_node, node);
1476 * If another node is sufficiently far away then it is better
1477 * to reclaim pages in a zone before going off node.
1479 if (distance > RECLAIM_DISTANCE)
1480 zone_reclaim_mode = 1;
1483 * We don't want to pressure a particular node.
1484 * So adding penalty to the first node in same
1485 * distance group to make it round-robin.
1488 if (distance != node_distance(local_node, prev_node))
1489 node_load[node] += load;
1492 for (i = 0; i < GFP_ZONETYPES; i++) {
1493 zonelist = pgdat->node_zonelists + i;
1494 for (j = 0; zonelist->zones[j] != NULL; j++);
1496 k = highest_zone(i);
1498 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1499 zonelist->zones[j] = NULL;
1504 #else /* CONFIG_NUMA */
1506 static void __meminit build_zonelists(pg_data_t *pgdat)
1508 int i, j, k, node, local_node;
1510 local_node = pgdat->node_id;
1511 for (i = 0; i < GFP_ZONETYPES; i++) {
1512 struct zonelist *zonelist;
1514 zonelist = pgdat->node_zonelists + i;
1517 k = highest_zone(i);
1518 j = build_zonelists_node(pgdat, zonelist, j, k);
1520 * Now we build the zonelist so that it contains the zones
1521 * of all the other nodes.
1522 * We don't want to pressure a particular node, so when
1523 * building the zones for node N, we make sure that the
1524 * zones coming right after the local ones are those from
1525 * node N+1 (modulo N)
1527 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1528 if (!node_online(node))
1530 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1532 for (node = 0; node < local_node; node++) {
1533 if (!node_online(node))
1535 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1538 zonelist->zones[j] = NULL;
1542 #endif /* CONFIG_NUMA */
1544 /* return values int ....just for stop_machine_run() */
1545 static int __meminit __build_all_zonelists(void *dummy)
1548 for_each_online_node(nid)
1549 build_zonelists(NODE_DATA(nid));
1553 void __meminit build_all_zonelists(void)
1555 if (system_state == SYSTEM_BOOTING) {
1556 __build_all_zonelists(0);
1557 cpuset_init_current_mems_allowed();
1559 /* we have to stop all cpus to guaranntee there is no user
1561 stop_machine_run(__build_all_zonelists, NULL, NR_CPUS);
1562 /* cpuset refresh routine should be here */
1564 vm_total_pages = nr_free_pagecache_pages();
1565 printk("Built %i zonelists. Total pages: %ld\n",
1566 num_online_nodes(), vm_total_pages);
1570 * Helper functions to size the waitqueue hash table.
1571 * Essentially these want to choose hash table sizes sufficiently
1572 * large so that collisions trying to wait on pages are rare.
1573 * But in fact, the number of active page waitqueues on typical
1574 * systems is ridiculously low, less than 200. So this is even
1575 * conservative, even though it seems large.
1577 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1578 * waitqueues, i.e. the size of the waitq table given the number of pages.
1580 #define PAGES_PER_WAITQUEUE 256
1582 #ifndef CONFIG_MEMORY_HOTPLUG
1583 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
1585 unsigned long size = 1;
1587 pages /= PAGES_PER_WAITQUEUE;
1589 while (size < pages)
1593 * Once we have dozens or even hundreds of threads sleeping
1594 * on IO we've got bigger problems than wait queue collision.
1595 * Limit the size of the wait table to a reasonable size.
1597 size = min(size, 4096UL);
1599 return max(size, 4UL);
1603 * A zone's size might be changed by hot-add, so it is not possible to determine
1604 * a suitable size for its wait_table. So we use the maximum size now.
1606 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
1608 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
1609 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
1610 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
1612 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
1613 * or more by the traditional way. (See above). It equals:
1615 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
1616 * ia64(16K page size) : = ( 8G + 4M)byte.
1617 * powerpc (64K page size) : = (32G +16M)byte.
1619 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
1626 * This is an integer logarithm so that shifts can be used later
1627 * to extract the more random high bits from the multiplicative
1628 * hash function before the remainder is taken.
1630 static inline unsigned long wait_table_bits(unsigned long size)
1635 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1637 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1638 unsigned long *zones_size, unsigned long *zholes_size)
1640 unsigned long realtotalpages, totalpages = 0;
1643 for (i = 0; i < MAX_NR_ZONES; i++)
1644 totalpages += zones_size[i];
1645 pgdat->node_spanned_pages = totalpages;
1647 realtotalpages = totalpages;
1649 for (i = 0; i < MAX_NR_ZONES; i++)
1650 realtotalpages -= zholes_size[i];
1651 pgdat->node_present_pages = realtotalpages;
1652 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1657 * Initially all pages are reserved - free ones are freed
1658 * up by free_all_bootmem() once the early boot process is
1659 * done. Non-atomic initialization, single-pass.
1661 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1662 unsigned long start_pfn)
1665 unsigned long end_pfn = start_pfn + size;
1668 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1669 if (!early_pfn_valid(pfn))
1671 page = pfn_to_page(pfn);
1672 set_page_links(page, zone, nid, pfn);
1673 init_page_count(page);
1674 reset_page_mapcount(page);
1675 SetPageReserved(page);
1676 INIT_LIST_HEAD(&page->lru);
1677 #ifdef WANT_PAGE_VIRTUAL
1678 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1679 if (!is_highmem_idx(zone))
1680 set_page_address(page, __va(pfn << PAGE_SHIFT));
1685 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1689 for (order = 0; order < MAX_ORDER ; order++) {
1690 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1691 zone->free_area[order].nr_free = 0;
1695 #define ZONETABLE_INDEX(x, zone_nr) ((x << ZONES_SHIFT) | zone_nr)
1696 void zonetable_add(struct zone *zone, int nid, int zid, unsigned long pfn,
1699 unsigned long snum = pfn_to_section_nr(pfn);
1700 unsigned long end = pfn_to_section_nr(pfn + size);
1703 zone_table[ZONETABLE_INDEX(nid, zid)] = zone;
1705 for (; snum <= end; snum++)
1706 zone_table[ZONETABLE_INDEX(snum, zid)] = zone;
1709 #ifndef __HAVE_ARCH_MEMMAP_INIT
1710 #define memmap_init(size, nid, zone, start_pfn) \
1711 memmap_init_zone((size), (nid), (zone), (start_pfn))
1714 static int __cpuinit zone_batchsize(struct zone *zone)
1719 * The per-cpu-pages pools are set to around 1000th of the
1720 * size of the zone. But no more than 1/2 of a meg.
1722 * OK, so we don't know how big the cache is. So guess.
1724 batch = zone->present_pages / 1024;
1725 if (batch * PAGE_SIZE > 512 * 1024)
1726 batch = (512 * 1024) / PAGE_SIZE;
1727 batch /= 4; /* We effectively *= 4 below */
1732 * Clamp the batch to a 2^n - 1 value. Having a power
1733 * of 2 value was found to be more likely to have
1734 * suboptimal cache aliasing properties in some cases.
1736 * For example if 2 tasks are alternately allocating
1737 * batches of pages, one task can end up with a lot
1738 * of pages of one half of the possible page colors
1739 * and the other with pages of the other colors.
1741 batch = (1 << (fls(batch + batch/2)-1)) - 1;
1746 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
1748 struct per_cpu_pages *pcp;
1750 memset(p, 0, sizeof(*p));
1752 pcp = &p->pcp[0]; /* hot */
1754 pcp->high = 6 * batch;
1755 pcp->batch = max(1UL, 1 * batch);
1756 INIT_LIST_HEAD(&pcp->list);
1758 pcp = &p->pcp[1]; /* cold*/
1760 pcp->high = 2 * batch;
1761 pcp->batch = max(1UL, batch/2);
1762 INIT_LIST_HEAD(&pcp->list);
1766 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
1767 * to the value high for the pageset p.
1770 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
1773 struct per_cpu_pages *pcp;
1775 pcp = &p->pcp[0]; /* hot list */
1777 pcp->batch = max(1UL, high/4);
1778 if ((high/4) > (PAGE_SHIFT * 8))
1779 pcp->batch = PAGE_SHIFT * 8;
1785 * Boot pageset table. One per cpu which is going to be used for all
1786 * zones and all nodes. The parameters will be set in such a way
1787 * that an item put on a list will immediately be handed over to
1788 * the buddy list. This is safe since pageset manipulation is done
1789 * with interrupts disabled.
1791 * Some NUMA counter updates may also be caught by the boot pagesets.
1793 * The boot_pagesets must be kept even after bootup is complete for
1794 * unused processors and/or zones. They do play a role for bootstrapping
1795 * hotplugged processors.
1797 * zoneinfo_show() and maybe other functions do
1798 * not check if the processor is online before following the pageset pointer.
1799 * Other parts of the kernel may not check if the zone is available.
1801 static struct per_cpu_pageset boot_pageset[NR_CPUS];
1804 * Dynamically allocate memory for the
1805 * per cpu pageset array in struct zone.
1807 static int __cpuinit process_zones(int cpu)
1809 struct zone *zone, *dzone;
1811 for_each_zone(zone) {
1813 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
1814 GFP_KERNEL, cpu_to_node(cpu));
1815 if (!zone_pcp(zone, cpu))
1818 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
1820 if (percpu_pagelist_fraction)
1821 setup_pagelist_highmark(zone_pcp(zone, cpu),
1822 (zone->present_pages / percpu_pagelist_fraction));
1827 for_each_zone(dzone) {
1830 kfree(zone_pcp(dzone, cpu));
1831 zone_pcp(dzone, cpu) = NULL;
1836 static inline void free_zone_pagesets(int cpu)
1840 for_each_zone(zone) {
1841 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
1843 /* Free per_cpu_pageset if it is slab allocated */
1844 if (pset != &boot_pageset[cpu])
1846 zone_pcp(zone, cpu) = NULL;
1850 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
1851 unsigned long action,
1854 int cpu = (long)hcpu;
1855 int ret = NOTIFY_OK;
1858 case CPU_UP_PREPARE:
1859 if (process_zones(cpu))
1862 case CPU_UP_CANCELED:
1864 free_zone_pagesets(cpu);
1872 static struct notifier_block __cpuinitdata pageset_notifier =
1873 { &pageset_cpuup_callback, NULL, 0 };
1875 void __init setup_per_cpu_pageset(void)
1879 /* Initialize per_cpu_pageset for cpu 0.
1880 * A cpuup callback will do this for every cpu
1881 * as it comes online
1883 err = process_zones(smp_processor_id());
1885 register_cpu_notifier(&pageset_notifier);
1891 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
1894 struct pglist_data *pgdat = zone->zone_pgdat;
1898 * The per-page waitqueue mechanism uses hashed waitqueues
1901 zone->wait_table_hash_nr_entries =
1902 wait_table_hash_nr_entries(zone_size_pages);
1903 zone->wait_table_bits =
1904 wait_table_bits(zone->wait_table_hash_nr_entries);
1905 alloc_size = zone->wait_table_hash_nr_entries
1906 * sizeof(wait_queue_head_t);
1908 if (system_state == SYSTEM_BOOTING) {
1909 zone->wait_table = (wait_queue_head_t *)
1910 alloc_bootmem_node(pgdat, alloc_size);
1913 * This case means that a zone whose size was 0 gets new memory
1914 * via memory hot-add.
1915 * But it may be the case that a new node was hot-added. In
1916 * this case vmalloc() will not be able to use this new node's
1917 * memory - this wait_table must be initialized to use this new
1918 * node itself as well.
1919 * To use this new node's memory, further consideration will be
1922 zone->wait_table = (wait_queue_head_t *)vmalloc(alloc_size);
1924 if (!zone->wait_table)
1927 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
1928 init_waitqueue_head(zone->wait_table + i);
1933 static __meminit void zone_pcp_init(struct zone *zone)
1936 unsigned long batch = zone_batchsize(zone);
1938 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1940 /* Early boot. Slab allocator not functional yet */
1941 zone_pcp(zone, cpu) = &boot_pageset[cpu];
1942 setup_pageset(&boot_pageset[cpu],0);
1944 setup_pageset(zone_pcp(zone,cpu), batch);
1947 if (zone->present_pages)
1948 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
1949 zone->name, zone->present_pages, batch);
1952 __meminit int init_currently_empty_zone(struct zone *zone,
1953 unsigned long zone_start_pfn,
1956 struct pglist_data *pgdat = zone->zone_pgdat;
1958 ret = zone_wait_table_init(zone, size);
1961 pgdat->nr_zones = zone_idx(zone) + 1;
1963 zone->zone_start_pfn = zone_start_pfn;
1965 memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn);
1967 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1973 * Set up the zone data structures:
1974 * - mark all pages reserved
1975 * - mark all memory queues empty
1976 * - clear the memory bitmaps
1978 static void __meminit free_area_init_core(struct pglist_data *pgdat,
1979 unsigned long *zones_size, unsigned long *zholes_size)
1982 int nid = pgdat->node_id;
1983 unsigned long zone_start_pfn = pgdat->node_start_pfn;
1986 pgdat_resize_init(pgdat);
1987 pgdat->nr_zones = 0;
1988 init_waitqueue_head(&pgdat->kswapd_wait);
1989 pgdat->kswapd_max_order = 0;
1991 for (j = 0; j < MAX_NR_ZONES; j++) {
1992 struct zone *zone = pgdat->node_zones + j;
1993 unsigned long size, realsize;
1995 realsize = size = zones_size[j];
1997 realsize -= zholes_size[j];
1999 if (j < ZONE_HIGHMEM)
2000 nr_kernel_pages += realsize;
2001 nr_all_pages += realsize;
2003 zone->spanned_pages = size;
2004 zone->present_pages = realsize;
2006 zone->min_unmapped_ratio = (realsize*sysctl_min_unmapped_ratio)
2009 zone->name = zone_names[j];
2010 spin_lock_init(&zone->lock);
2011 spin_lock_init(&zone->lru_lock);
2012 zone_seqlock_init(zone);
2013 zone->zone_pgdat = pgdat;
2014 zone->free_pages = 0;
2016 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
2018 zone_pcp_init(zone);
2019 INIT_LIST_HEAD(&zone->active_list);
2020 INIT_LIST_HEAD(&zone->inactive_list);
2021 zone->nr_scan_active = 0;
2022 zone->nr_scan_inactive = 0;
2023 zone->nr_active = 0;
2024 zone->nr_inactive = 0;
2025 zap_zone_vm_stats(zone);
2026 atomic_set(&zone->reclaim_in_progress, 0);
2030 zonetable_add(zone, nid, j, zone_start_pfn, size);
2031 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
2033 zone_start_pfn += size;
2037 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
2039 /* Skip empty nodes */
2040 if (!pgdat->node_spanned_pages)
2043 #ifdef CONFIG_FLAT_NODE_MEM_MAP
2044 /* ia64 gets its own node_mem_map, before this, without bootmem */
2045 if (!pgdat->node_mem_map) {
2046 unsigned long size, start, end;
2050 * The zone's endpoints aren't required to be MAX_ORDER
2051 * aligned but the node_mem_map endpoints must be in order
2052 * for the buddy allocator to function correctly.
2054 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
2055 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
2056 end = ALIGN(end, MAX_ORDER_NR_PAGES);
2057 size = (end - start) * sizeof(struct page);
2058 map = alloc_remap(pgdat->node_id, size);
2060 map = alloc_bootmem_node(pgdat, size);
2061 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
2063 #ifdef CONFIG_FLATMEM
2065 * With no DISCONTIG, the global mem_map is just set as node 0's
2067 if (pgdat == NODE_DATA(0))
2068 mem_map = NODE_DATA(0)->node_mem_map;
2070 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
2073 void __meminit free_area_init_node(int nid, struct pglist_data *pgdat,
2074 unsigned long *zones_size, unsigned long node_start_pfn,
2075 unsigned long *zholes_size)
2077 pgdat->node_id = nid;
2078 pgdat->node_start_pfn = node_start_pfn;
2079 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
2081 alloc_node_mem_map(pgdat);
2083 free_area_init_core(pgdat, zones_size, zholes_size);
2086 #ifndef CONFIG_NEED_MULTIPLE_NODES
2087 static bootmem_data_t contig_bootmem_data;
2088 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
2090 EXPORT_SYMBOL(contig_page_data);
2093 void __init free_area_init(unsigned long *zones_size)
2095 free_area_init_node(0, NODE_DATA(0), zones_size,
2096 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
2099 #ifdef CONFIG_HOTPLUG_CPU
2100 static int page_alloc_cpu_notify(struct notifier_block *self,
2101 unsigned long action, void *hcpu)
2103 int cpu = (unsigned long)hcpu;
2105 if (action == CPU_DEAD) {
2106 local_irq_disable();
2108 vm_events_fold_cpu(cpu);
2110 refresh_cpu_vm_stats(cpu);
2114 #endif /* CONFIG_HOTPLUG_CPU */
2116 void __init page_alloc_init(void)
2118 hotcpu_notifier(page_alloc_cpu_notify, 0);
2122 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
2123 * or min_free_kbytes changes.
2125 static void calculate_totalreserve_pages(void)
2127 struct pglist_data *pgdat;
2128 unsigned long reserve_pages = 0;
2131 for_each_online_pgdat(pgdat) {
2132 for (i = 0; i < MAX_NR_ZONES; i++) {
2133 struct zone *zone = pgdat->node_zones + i;
2134 unsigned long max = 0;
2136 /* Find valid and maximum lowmem_reserve in the zone */
2137 for (j = i; j < MAX_NR_ZONES; j++) {
2138 if (zone->lowmem_reserve[j] > max)
2139 max = zone->lowmem_reserve[j];
2142 /* we treat pages_high as reserved pages. */
2143 max += zone->pages_high;
2145 if (max > zone->present_pages)
2146 max = zone->present_pages;
2147 reserve_pages += max;
2150 totalreserve_pages = reserve_pages;
2154 * setup_per_zone_lowmem_reserve - called whenever
2155 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
2156 * has a correct pages reserved value, so an adequate number of
2157 * pages are left in the zone after a successful __alloc_pages().
2159 static void setup_per_zone_lowmem_reserve(void)
2161 struct pglist_data *pgdat;
2164 for_each_online_pgdat(pgdat) {
2165 for (j = 0; j < MAX_NR_ZONES; j++) {
2166 struct zone *zone = pgdat->node_zones + j;
2167 unsigned long present_pages = zone->present_pages;
2169 zone->lowmem_reserve[j] = 0;
2171 for (idx = j-1; idx >= 0; idx--) {
2172 struct zone *lower_zone;
2174 if (sysctl_lowmem_reserve_ratio[idx] < 1)
2175 sysctl_lowmem_reserve_ratio[idx] = 1;
2177 lower_zone = pgdat->node_zones + idx;
2178 lower_zone->lowmem_reserve[j] = present_pages /
2179 sysctl_lowmem_reserve_ratio[idx];
2180 present_pages += lower_zone->present_pages;
2185 /* update totalreserve_pages */
2186 calculate_totalreserve_pages();
2190 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
2191 * that the pages_{min,low,high} values for each zone are set correctly
2192 * with respect to min_free_kbytes.
2194 void setup_per_zone_pages_min(void)
2196 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
2197 unsigned long lowmem_pages = 0;
2199 unsigned long flags;
2201 /* Calculate total number of !ZONE_HIGHMEM pages */
2202 for_each_zone(zone) {
2203 if (!is_highmem(zone))
2204 lowmem_pages += zone->present_pages;
2207 for_each_zone(zone) {
2210 spin_lock_irqsave(&zone->lru_lock, flags);
2211 tmp = (u64)pages_min * zone->present_pages;
2212 do_div(tmp, lowmem_pages);
2213 if (is_highmem(zone)) {
2215 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
2216 * need highmem pages, so cap pages_min to a small
2219 * The (pages_high-pages_low) and (pages_low-pages_min)
2220 * deltas controls asynch page reclaim, and so should
2221 * not be capped for highmem.
2225 min_pages = zone->present_pages / 1024;
2226 if (min_pages < SWAP_CLUSTER_MAX)
2227 min_pages = SWAP_CLUSTER_MAX;
2228 if (min_pages > 128)
2230 zone->pages_min = min_pages;
2233 * If it's a lowmem zone, reserve a number of pages
2234 * proportionate to the zone's size.
2236 zone->pages_min = tmp;
2239 zone->pages_low = zone->pages_min + (tmp >> 2);
2240 zone->pages_high = zone->pages_min + (tmp >> 1);
2241 spin_unlock_irqrestore(&zone->lru_lock, flags);
2244 /* update totalreserve_pages */
2245 calculate_totalreserve_pages();
2249 * Initialise min_free_kbytes.
2251 * For small machines we want it small (128k min). For large machines
2252 * we want it large (64MB max). But it is not linear, because network
2253 * bandwidth does not increase linearly with machine size. We use
2255 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2256 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2272 static int __init init_per_zone_pages_min(void)
2274 unsigned long lowmem_kbytes;
2276 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2278 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2279 if (min_free_kbytes < 128)
2280 min_free_kbytes = 128;
2281 if (min_free_kbytes > 65536)
2282 min_free_kbytes = 65536;
2283 setup_per_zone_pages_min();
2284 setup_per_zone_lowmem_reserve();
2287 module_init(init_per_zone_pages_min)
2290 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2291 * that we can call two helper functions whenever min_free_kbytes
2294 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2295 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2297 proc_dointvec(table, write, file, buffer, length, ppos);
2298 setup_per_zone_pages_min();
2303 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
2304 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2309 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2314 zone->min_unmapped_ratio = (zone->present_pages *
2315 sysctl_min_unmapped_ratio) / 100;
2321 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2322 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2323 * whenever sysctl_lowmem_reserve_ratio changes.
2325 * The reserve ratio obviously has absolutely no relation with the
2326 * pages_min watermarks. The lowmem reserve ratio can only make sense
2327 * if in function of the boot time zone sizes.
2329 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
2330 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2332 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2333 setup_per_zone_lowmem_reserve();
2338 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
2339 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
2340 * can have before it gets flushed back to buddy allocator.
2343 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
2344 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2350 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2351 if (!write || (ret == -EINVAL))
2353 for_each_zone(zone) {
2354 for_each_online_cpu(cpu) {
2356 high = zone->present_pages / percpu_pagelist_fraction;
2357 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
2363 int hashdist = HASHDIST_DEFAULT;
2366 static int __init set_hashdist(char *str)
2370 hashdist = simple_strtoul(str, &str, 0);
2373 __setup("hashdist=", set_hashdist);
2377 * allocate a large system hash table from bootmem
2378 * - it is assumed that the hash table must contain an exact power-of-2
2379 * quantity of entries
2380 * - limit is the number of hash buckets, not the total allocation size
2382 void *__init alloc_large_system_hash(const char *tablename,
2383 unsigned long bucketsize,
2384 unsigned long numentries,
2387 unsigned int *_hash_shift,
2388 unsigned int *_hash_mask,
2389 unsigned long limit)
2391 unsigned long long max = limit;
2392 unsigned long log2qty, size;
2395 /* allow the kernel cmdline to have a say */
2397 /* round applicable memory size up to nearest megabyte */
2398 numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages;
2399 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2400 numentries >>= 20 - PAGE_SHIFT;
2401 numentries <<= 20 - PAGE_SHIFT;
2403 /* limit to 1 bucket per 2^scale bytes of low memory */
2404 if (scale > PAGE_SHIFT)
2405 numentries >>= (scale - PAGE_SHIFT);
2407 numentries <<= (PAGE_SHIFT - scale);
2409 numentries = roundup_pow_of_two(numentries);
2411 /* limit allocation size to 1/16 total memory by default */
2413 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2414 do_div(max, bucketsize);
2417 if (numentries > max)
2420 log2qty = long_log2(numentries);
2423 size = bucketsize << log2qty;
2424 if (flags & HASH_EARLY)
2425 table = alloc_bootmem(size);
2427 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
2429 unsigned long order;
2430 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
2432 table = (void*) __get_free_pages(GFP_ATOMIC, order);
2434 } while (!table && size > PAGE_SIZE && --log2qty);
2437 panic("Failed to allocate %s hash table\n", tablename);
2439 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2442 long_log2(size) - PAGE_SHIFT,
2446 *_hash_shift = log2qty;
2448 *_hash_mask = (1 << log2qty) - 1;
2453 #ifdef CONFIG_OUT_OF_LINE_PFN_TO_PAGE
2454 struct page *pfn_to_page(unsigned long pfn)
2456 return __pfn_to_page(pfn);
2458 unsigned long page_to_pfn(struct page *page)
2460 return __page_to_pfn(page);
2462 EXPORT_SYMBOL(pfn_to_page);
2463 EXPORT_SYMBOL(page_to_pfn);
2464 #endif /* CONFIG_OUT_OF_LINE_PFN_TO_PAGE */