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>
40 #include <linux/vs_base.h>
41 #include <linux/vs_limit.h>
43 #include <asm/tlbflush.h>
44 #include <asm/div64.h>
48 * MCD - HACK: Find somewhere to initialize this EARLY, or make this
51 nodemask_t node_online_map __read_mostly = { { [0] = 1UL } };
52 EXPORT_SYMBOL(node_online_map);
53 nodemask_t node_possible_map __read_mostly = NODE_MASK_ALL;
54 EXPORT_SYMBOL(node_possible_map);
55 unsigned long totalram_pages __read_mostly;
56 unsigned long totalhigh_pages __read_mostly;
57 unsigned long totalreserve_pages __read_mostly;
59 int percpu_pagelist_fraction;
61 static void __free_pages_ok(struct page *page, unsigned int order);
64 * results with 256, 32 in the lowmem_reserve sysctl:
65 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
66 * 1G machine -> (16M dma, 784M normal, 224M high)
67 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
68 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
69 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
71 * TBD: should special case ZONE_DMA32 machines here - in those we normally
72 * don't need any ZONE_NORMAL reservation
74 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 256, 256, 32 };
76 EXPORT_SYMBOL(totalram_pages);
79 * Used by page_zone() to look up the address of the struct zone whose
80 * id is encoded in the upper bits of page->flags
82 struct zone *zone_table[1 << ZONETABLE_SHIFT] __read_mostly;
83 EXPORT_SYMBOL(zone_table);
85 static char *zone_names[MAX_NR_ZONES] = { "DMA", "DMA32", "Normal", "HighMem" };
86 int min_free_kbytes = 1024;
88 unsigned long __meminitdata nr_kernel_pages;
89 unsigned long __meminitdata nr_all_pages;
91 #ifdef CONFIG_DEBUG_VM
92 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
96 unsigned long pfn = page_to_pfn(page);
99 seq = zone_span_seqbegin(zone);
100 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
102 else if (pfn < zone->zone_start_pfn)
104 } while (zone_span_seqretry(zone, seq));
109 static int page_is_consistent(struct zone *zone, struct page *page)
111 #ifdef CONFIG_HOLES_IN_ZONE
112 if (!pfn_valid(page_to_pfn(page)))
115 if (zone != page_zone(page))
121 * Temporary debugging check for pages not lying within a given zone.
123 static int bad_range(struct zone *zone, struct page *page)
125 if (page_outside_zone_boundaries(zone, page))
127 if (!page_is_consistent(zone, page))
134 static inline int bad_range(struct zone *zone, struct page *page)
140 static void bad_page(struct page *page)
142 printk(KERN_EMERG "Bad page state in process '%s'\n"
143 KERN_EMERG "page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d (%s)\n"
144 KERN_EMERG "Trying to fix it up, but a reboot is needed\n"
145 KERN_EMERG "Backtrace:\n",
146 current->comm, page, (int)(2*sizeof(unsigned long)),
147 (unsigned long)page->flags, page->mapping,
148 page_mapcount(page), page_count(page), print_tainted());
150 page->flags &= ~(1 << PG_lru |
160 set_page_count(page, 0);
161 reset_page_mapcount(page);
162 page->mapping = NULL;
163 add_taint(TAINT_BAD_PAGE);
167 * Higher-order pages are called "compound pages". They are structured thusly:
169 * The first PAGE_SIZE page is called the "head page".
171 * The remaining PAGE_SIZE pages are called "tail pages".
173 * All pages have PG_compound set. All pages have their ->private pointing at
174 * the head page (even the head page has this).
176 * The first tail page's ->lru.next holds the address of the compound page's
177 * put_page() function. Its ->lru.prev holds the order of allocation.
178 * This usage means that zero-order pages may not be compound.
181 static void free_compound_page(struct page *page)
183 __free_pages_ok(page, (unsigned long)page[1].lru.prev);
186 static void prep_compound_page(struct page *page, unsigned long order)
189 int nr_pages = 1 << order;
191 page[1].lru.next = (void *)free_compound_page; /* set dtor */
192 page[1].lru.prev = (void *)order;
193 for (i = 0; i < nr_pages; i++) {
194 struct page *p = page + i;
196 __SetPageCompound(p);
197 set_page_private(p, (unsigned long)page);
201 static void destroy_compound_page(struct page *page, unsigned long order)
204 int nr_pages = 1 << order;
206 if (unlikely((unsigned long)page[1].lru.prev != order))
209 for (i = 0; i < nr_pages; i++) {
210 struct page *p = page + i;
212 if (unlikely(!PageCompound(p) |
213 (page_private(p) != (unsigned long)page)))
215 __ClearPageCompound(p);
219 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
223 BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
225 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
226 * and __GFP_HIGHMEM from hard or soft interrupt context.
228 BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
229 for (i = 0; i < (1 << order); i++)
230 clear_highpage(page + i);
234 * function for dealing with page's order in buddy system.
235 * zone->lock is already acquired when we use these.
236 * So, we don't need atomic page->flags operations here.
238 static inline unsigned long page_order(struct page *page)
240 return page_private(page);
243 static inline void set_page_order(struct page *page, int order)
245 set_page_private(page, order);
246 __SetPageBuddy(page);
249 static inline void rmv_page_order(struct page *page)
251 __ClearPageBuddy(page);
252 set_page_private(page, 0);
256 * Locate the struct page for both the matching buddy in our
257 * pair (buddy1) and the combined O(n+1) page they form (page).
259 * 1) Any buddy B1 will have an order O twin B2 which satisfies
260 * the following equation:
262 * For example, if the starting buddy (buddy2) is #8 its order
264 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
266 * 2) Any buddy B will have an order O+1 parent P which
267 * satisfies the following equation:
270 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
272 static inline struct page *
273 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
275 unsigned long buddy_idx = page_idx ^ (1 << order);
277 return page + (buddy_idx - page_idx);
280 static inline unsigned long
281 __find_combined_index(unsigned long page_idx, unsigned int order)
283 return (page_idx & ~(1 << order));
287 * This function checks whether a page is free && is the buddy
288 * we can do coalesce a page and its buddy if
289 * (a) the buddy is not in a hole &&
290 * (b) the buddy is in the buddy system &&
291 * (c) a page and its buddy have the same order &&
292 * (d) a page and its buddy are in the same zone.
294 * For recording whether a page is in the buddy system, we use PG_buddy.
295 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
297 * For recording page's order, we use page_private(page).
299 static inline int page_is_buddy(struct page *page, struct page *buddy,
302 #ifdef CONFIG_HOLES_IN_ZONE
303 if (!pfn_valid(page_to_pfn(buddy)))
307 if (page_zone_id(page) != page_zone_id(buddy))
310 if (PageBuddy(buddy) && page_order(buddy) == order) {
311 BUG_ON(page_count(buddy) != 0);
318 * Freeing function for a buddy system allocator.
320 * The concept of a buddy system is to maintain direct-mapped table
321 * (containing bit values) for memory blocks of various "orders".
322 * The bottom level table contains the map for the smallest allocatable
323 * units of memory (here, pages), and each level above it describes
324 * pairs of units from the levels below, hence, "buddies".
325 * At a high level, all that happens here is marking the table entry
326 * at the bottom level available, and propagating the changes upward
327 * as necessary, plus some accounting needed to play nicely with other
328 * parts of the VM system.
329 * At each level, we keep a list of pages, which are heads of continuous
330 * free pages of length of (1 << order) and marked with PG_buddy. Page's
331 * order is recorded in page_private(page) field.
332 * So when we are allocating or freeing one, we can derive the state of the
333 * other. That is, if we allocate a small block, and both were
334 * free, the remainder of the region must be split into blocks.
335 * If a block is freed, and its buddy is also free, then this
336 * triggers coalescing into a block of larger size.
341 static inline void __free_one_page(struct page *page,
342 struct zone *zone, unsigned int order)
344 unsigned long page_idx;
345 int order_size = 1 << order;
347 if (unlikely(PageCompound(page)))
348 destroy_compound_page(page, order);
350 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
352 BUG_ON(page_idx & (order_size - 1));
353 BUG_ON(bad_range(zone, page));
355 zone->free_pages += order_size;
356 while (order < MAX_ORDER-1) {
357 unsigned long combined_idx;
358 struct free_area *area;
361 buddy = __page_find_buddy(page, page_idx, order);
362 if (!page_is_buddy(page, buddy, order))
363 break; /* Move the buddy up one level. */
365 list_del(&buddy->lru);
366 area = zone->free_area + order;
368 rmv_page_order(buddy);
369 combined_idx = __find_combined_index(page_idx, order);
370 page = page + (combined_idx - page_idx);
371 page_idx = combined_idx;
374 set_page_order(page, order);
375 list_add(&page->lru, &zone->free_area[order].free_list);
376 zone->free_area[order].nr_free++;
379 static inline int free_pages_check(struct page *page)
381 if (unlikely(page_mapcount(page) |
382 (page->mapping != NULL) |
383 (page_count(page) != 0) |
397 __ClearPageDirty(page);
399 * For now, we report if PG_reserved was found set, but do not
400 * clear it, and do not free the page. But we shall soon need
401 * to do more, for when the ZERO_PAGE count wraps negative.
403 return PageReserved(page);
407 * Frees a list of pages.
408 * Assumes all pages on list are in same zone, and of same order.
409 * count is the number of pages to free.
411 * If the zone was previously in an "all pages pinned" state then look to
412 * see if this freeing clears that state.
414 * And clear the zone's pages_scanned counter, to hold off the "all pages are
415 * pinned" detection logic.
417 static void free_pages_bulk(struct zone *zone, int count,
418 struct list_head *list, int order)
420 spin_lock(&zone->lock);
421 zone->all_unreclaimable = 0;
422 zone->pages_scanned = 0;
426 BUG_ON(list_empty(list));
427 page = list_entry(list->prev, struct page, lru);
428 /* have to delete it as __free_one_page list manipulates */
429 list_del(&page->lru);
430 __free_one_page(page, zone, order);
432 spin_unlock(&zone->lock);
435 static void free_one_page(struct zone *zone, struct page *page, int order)
438 list_add(&page->lru, &list);
439 free_pages_bulk(zone, 1, &list, order);
442 static void __free_pages_ok(struct page *page, unsigned int order)
448 if (arch_free_page(page, order))
450 if (!PageHighMem(page))
451 debug_check_no_locks_freed(page_address(page),
454 for (i = 0 ; i < (1 << order) ; ++i)
455 reserved += free_pages_check(page + i);
459 kernel_map_pages(page, 1 << order, 0);
460 local_irq_save(flags);
461 __count_vm_events(PGFREE, 1 << order);
462 free_one_page(page_zone(page), page, order);
463 local_irq_restore(flags);
467 * permit the bootmem allocator to evade page validation on high-order frees
469 void fastcall __init __free_pages_bootmem(struct page *page, unsigned int order)
472 __ClearPageReserved(page);
473 set_page_count(page, 0);
474 set_page_refcounted(page);
480 for (loop = 0; loop < BITS_PER_LONG; loop++) {
481 struct page *p = &page[loop];
483 if (loop + 1 < BITS_PER_LONG)
485 __ClearPageReserved(p);
486 set_page_count(p, 0);
489 set_page_refcounted(page);
490 __free_pages(page, order);
496 * The order of subdivision here is critical for the IO subsystem.
497 * Please do not alter this order without good reasons and regression
498 * testing. Specifically, as large blocks of memory are subdivided,
499 * the order in which smaller blocks are delivered depends on the order
500 * they're subdivided in this function. This is the primary factor
501 * influencing the order in which pages are delivered to the IO
502 * subsystem according to empirical testing, and this is also justified
503 * by considering the behavior of a buddy system containing a single
504 * large block of memory acted on by a series of small allocations.
505 * This behavior is a critical factor in sglist merging's success.
509 static inline void expand(struct zone *zone, struct page *page,
510 int low, int high, struct free_area *area)
512 unsigned long size = 1 << high;
518 BUG_ON(bad_range(zone, &page[size]));
519 list_add(&page[size].lru, &area->free_list);
521 set_page_order(&page[size], high);
526 * This page is about to be returned from the page allocator
528 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
530 if (unlikely(page_mapcount(page) |
531 (page->mapping != NULL) |
532 (page_count(page) != 0) |
548 * For now, we report if PG_reserved was found set, but do not
549 * clear it, and do not allocate the page: as a safety net.
551 if (PageReserved(page))
554 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
555 1 << PG_referenced | 1 << PG_arch_1 |
556 1 << PG_fs_misc | 1 << PG_mappedtodisk);
557 set_page_private(page, 0);
558 set_page_refcounted(page);
559 kernel_map_pages(page, 1 << order, 1);
561 if (gfp_flags & __GFP_ZERO)
562 prep_zero_page(page, order, gfp_flags);
564 if (order && (gfp_flags & __GFP_COMP))
565 prep_compound_page(page, order);
571 * Do the hard work of removing an element from the buddy allocator.
572 * Call me with the zone->lock already held.
574 static struct page *__rmqueue(struct zone *zone, unsigned int order)
576 struct free_area * area;
577 unsigned int current_order;
580 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
581 area = zone->free_area + current_order;
582 if (list_empty(&area->free_list))
585 page = list_entry(area->free_list.next, struct page, lru);
586 list_del(&page->lru);
587 rmv_page_order(page);
589 zone->free_pages -= 1UL << order;
590 expand(zone, page, order, current_order, area);
598 * Obtain a specified number of elements from the buddy allocator, all under
599 * a single hold of the lock, for efficiency. Add them to the supplied list.
600 * Returns the number of new pages which were placed at *list.
602 static int rmqueue_bulk(struct zone *zone, unsigned int order,
603 unsigned long count, struct list_head *list)
607 spin_lock(&zone->lock);
608 for (i = 0; i < count; ++i) {
609 struct page *page = __rmqueue(zone, order);
610 if (unlikely(page == NULL))
612 list_add_tail(&page->lru, list);
614 spin_unlock(&zone->lock);
620 * Called from the slab reaper to drain pagesets on a particular node that
621 * belong to the currently executing processor.
622 * Note that this function must be called with the thread pinned to
623 * a single processor.
625 void drain_node_pages(int nodeid)
630 for (z = 0; z < MAX_NR_ZONES; z++) {
631 struct zone *zone = NODE_DATA(nodeid)->node_zones + z;
632 struct per_cpu_pageset *pset;
634 pset = zone_pcp(zone, smp_processor_id());
635 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
636 struct per_cpu_pages *pcp;
640 local_irq_save(flags);
641 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
643 local_irq_restore(flags);
650 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
651 static void __drain_pages(unsigned int cpu)
657 for_each_zone(zone) {
658 struct per_cpu_pageset *pset;
660 pset = zone_pcp(zone, cpu);
661 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
662 struct per_cpu_pages *pcp;
665 local_irq_save(flags);
666 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
668 local_irq_restore(flags);
672 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
676 void mark_free_pages(struct zone *zone)
678 unsigned long zone_pfn, flags;
680 struct list_head *curr;
682 if (!zone->spanned_pages)
685 spin_lock_irqsave(&zone->lock, flags);
686 for (zone_pfn = 0; zone_pfn < zone->spanned_pages; ++zone_pfn)
687 ClearPageNosaveFree(pfn_to_page(zone_pfn + zone->zone_start_pfn));
689 for (order = MAX_ORDER - 1; order >= 0; --order)
690 list_for_each(curr, &zone->free_area[order].free_list) {
691 unsigned long start_pfn, i;
693 start_pfn = page_to_pfn(list_entry(curr, struct page, lru));
695 for (i=0; i < (1<<order); i++)
696 SetPageNosaveFree(pfn_to_page(start_pfn+i));
698 spin_unlock_irqrestore(&zone->lock, flags);
702 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
704 void drain_local_pages(void)
708 local_irq_save(flags);
709 __drain_pages(smp_processor_id());
710 local_irq_restore(flags);
712 #endif /* CONFIG_PM */
715 * Free a 0-order page
717 static void fastcall free_hot_cold_page(struct page *page, int cold)
719 struct zone *zone = page_zone(page);
720 struct per_cpu_pages *pcp;
723 if (arch_free_page(page, 0))
727 page->mapping = NULL;
728 if (free_pages_check(page))
731 kernel_map_pages(page, 1, 0);
733 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
734 local_irq_save(flags);
735 __count_vm_event(PGFREE);
736 list_add(&page->lru, &pcp->list);
738 if (pcp->count >= pcp->high) {
739 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
740 pcp->count -= pcp->batch;
742 local_irq_restore(flags);
746 void fastcall free_hot_page(struct page *page)
748 free_hot_cold_page(page, 0);
751 void fastcall free_cold_page(struct page *page)
753 free_hot_cold_page(page, 1);
757 * split_page takes a non-compound higher-order page, and splits it into
758 * n (1<<order) sub-pages: page[0..n]
759 * Each sub-page must be freed individually.
761 * Note: this is probably too low level an operation for use in drivers.
762 * Please consult with lkml before using this in your driver.
764 void split_page(struct page *page, unsigned int order)
768 BUG_ON(PageCompound(page));
769 BUG_ON(!page_count(page));
770 for (i = 1; i < (1 << order); i++)
771 set_page_refcounted(page + i);
775 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
776 * we cheat by calling it from here, in the order > 0 path. Saves a branch
779 static struct page *buffered_rmqueue(struct zonelist *zonelist,
780 struct zone *zone, int order, gfp_t gfp_flags)
784 int cold = !!(gfp_flags & __GFP_COLD);
789 if (likely(order == 0)) {
790 struct per_cpu_pages *pcp;
792 pcp = &zone_pcp(zone, cpu)->pcp[cold];
793 local_irq_save(flags);
795 pcp->count += rmqueue_bulk(zone, 0,
796 pcp->batch, &pcp->list);
797 if (unlikely(!pcp->count))
800 page = list_entry(pcp->list.next, struct page, lru);
801 list_del(&page->lru);
804 spin_lock_irqsave(&zone->lock, flags);
805 page = __rmqueue(zone, order);
806 spin_unlock(&zone->lock);
811 __count_zone_vm_events(PGALLOC, zone, 1 << order);
812 zone_statistics(zonelist, zone);
813 local_irq_restore(flags);
816 BUG_ON(bad_range(zone, page));
817 if (prep_new_page(page, order, gfp_flags))
822 local_irq_restore(flags);
827 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
828 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
829 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
830 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
831 #define ALLOC_HARDER 0x10 /* try to alloc harder */
832 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
833 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
836 * Return 1 if free pages are above 'mark'. This takes into account the order
839 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
840 int classzone_idx, int alloc_flags)
842 /* free_pages my go negative - that's OK */
843 long min = mark, free_pages = z->free_pages - (1 << order) + 1;
846 if (alloc_flags & ALLOC_HIGH)
848 if (alloc_flags & ALLOC_HARDER)
851 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
853 for (o = 0; o < order; o++) {
854 /* At the next order, this order's pages become unavailable */
855 free_pages -= z->free_area[o].nr_free << o;
857 /* Require fewer higher order pages to be free */
860 if (free_pages <= min)
867 * get_page_from_freeliest goes through the zonelist trying to allocate
871 get_page_from_freelist(gfp_t gfp_mask, unsigned int order,
872 struct zonelist *zonelist, int alloc_flags)
874 struct zone **z = zonelist->zones;
875 struct page *page = NULL;
876 int classzone_idx = zone_idx(*z);
879 * Go through the zonelist once, looking for a zone with enough free.
880 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
883 if ((alloc_flags & ALLOC_CPUSET) &&
884 !cpuset_zone_allowed(*z, gfp_mask))
887 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
889 if (alloc_flags & ALLOC_WMARK_MIN)
890 mark = (*z)->pages_min;
891 else if (alloc_flags & ALLOC_WMARK_LOW)
892 mark = (*z)->pages_low;
894 mark = (*z)->pages_high;
895 if (!zone_watermark_ok(*z, order, mark,
896 classzone_idx, alloc_flags))
897 if (!zone_reclaim_mode ||
898 !zone_reclaim(*z, gfp_mask, order))
902 page = buffered_rmqueue(zonelist, *z, order, gfp_mask);
906 } while (*(++z) != NULL);
911 * This is the 'heart' of the zoned buddy allocator.
913 struct page * fastcall
914 __alloc_pages(gfp_t gfp_mask, unsigned int order,
915 struct zonelist *zonelist)
917 const gfp_t wait = gfp_mask & __GFP_WAIT;
920 struct reclaim_state reclaim_state;
921 struct task_struct *p = current;
924 int did_some_progress;
926 might_sleep_if(wait);
929 z = zonelist->zones; /* the list of zones suitable for gfp_mask */
931 if (unlikely(*z == NULL)) {
932 /* Should this ever happen?? */
936 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
937 zonelist, ALLOC_WMARK_LOW|ALLOC_CPUSET);
942 wakeup_kswapd(*z, order);
946 * OK, we're below the kswapd watermark and have kicked background
947 * reclaim. Now things get more complex, so set up alloc_flags according
948 * to how we want to proceed.
950 * The caller may dip into page reserves a bit more if the caller
951 * cannot run direct reclaim, or if the caller has realtime scheduling
952 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
953 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
955 alloc_flags = ALLOC_WMARK_MIN;
956 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
957 alloc_flags |= ALLOC_HARDER;
958 if (gfp_mask & __GFP_HIGH)
959 alloc_flags |= ALLOC_HIGH;
961 alloc_flags |= ALLOC_CPUSET;
964 * Go through the zonelist again. Let __GFP_HIGH and allocations
965 * coming from realtime tasks go deeper into reserves.
967 * This is the last chance, in general, before the goto nopage.
968 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
969 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
971 page = get_page_from_freelist(gfp_mask, order, zonelist, alloc_flags);
975 /* This allocation should allow future memory freeing. */
977 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
978 && !in_interrupt()) {
979 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
981 /* go through the zonelist yet again, ignoring mins */
982 page = get_page_from_freelist(gfp_mask, order,
983 zonelist, ALLOC_NO_WATERMARKS);
986 if (gfp_mask & __GFP_NOFAIL) {
987 blk_congestion_wait(WRITE, HZ/50);
994 /* Atomic allocations - we can't balance anything */
1001 /* We now go into synchronous reclaim */
1002 cpuset_memory_pressure_bump();
1003 p->flags |= PF_MEMALLOC;
1004 reclaim_state.reclaimed_slab = 0;
1005 p->reclaim_state = &reclaim_state;
1007 did_some_progress = try_to_free_pages(zonelist->zones, gfp_mask);
1009 p->reclaim_state = NULL;
1010 p->flags &= ~PF_MEMALLOC;
1014 if (likely(did_some_progress)) {
1015 page = get_page_from_freelist(gfp_mask, order,
1016 zonelist, alloc_flags);
1019 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1021 * Go through the zonelist yet one more time, keep
1022 * very high watermark here, this is only to catch
1023 * a parallel oom killing, we must fail if we're still
1024 * under heavy pressure.
1026 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1027 zonelist, ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1031 out_of_memory(zonelist, gfp_mask, order);
1036 * Don't let big-order allocations loop unless the caller explicitly
1037 * requests that. Wait for some write requests to complete then retry.
1039 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
1040 * <= 3, but that may not be true in other implementations.
1043 if (!(gfp_mask & __GFP_NORETRY)) {
1044 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
1046 if (gfp_mask & __GFP_NOFAIL)
1050 blk_congestion_wait(WRITE, HZ/50);
1055 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1056 printk(KERN_WARNING "%s: page allocation failure."
1057 " order:%d, mode:0x%x\n",
1058 p->comm, order, gfp_mask);
1066 EXPORT_SYMBOL(__alloc_pages);
1069 * Common helper functions.
1071 fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1074 page = alloc_pages(gfp_mask, order);
1077 return (unsigned long) page_address(page);
1080 EXPORT_SYMBOL(__get_free_pages);
1082 fastcall unsigned long get_zeroed_page(gfp_t gfp_mask)
1087 * get_zeroed_page() returns a 32-bit address, which cannot represent
1090 BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1092 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1094 return (unsigned long) page_address(page);
1098 EXPORT_SYMBOL(get_zeroed_page);
1100 void __pagevec_free(struct pagevec *pvec)
1102 int i = pagevec_count(pvec);
1105 free_hot_cold_page(pvec->pages[i], pvec->cold);
1108 fastcall void __free_pages(struct page *page, unsigned int order)
1110 if (put_page_testzero(page)) {
1112 free_hot_page(page);
1114 __free_pages_ok(page, order);
1118 EXPORT_SYMBOL(__free_pages);
1120 fastcall void free_pages(unsigned long addr, unsigned int order)
1123 BUG_ON(!virt_addr_valid((void *)addr));
1124 __free_pages(virt_to_page((void *)addr), order);
1128 EXPORT_SYMBOL(free_pages);
1131 * Total amount of free (allocatable) RAM:
1133 unsigned int nr_free_pages(void)
1135 unsigned int sum = 0;
1139 sum += zone->free_pages;
1144 EXPORT_SYMBOL(nr_free_pages);
1147 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
1149 unsigned int i, sum = 0;
1151 for (i = 0; i < MAX_NR_ZONES; i++)
1152 sum += pgdat->node_zones[i].free_pages;
1158 static unsigned int nr_free_zone_pages(int offset)
1160 /* Just pick one node, since fallback list is circular */
1161 pg_data_t *pgdat = NODE_DATA(numa_node_id());
1162 unsigned int sum = 0;
1164 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1165 struct zone **zonep = zonelist->zones;
1168 for (zone = *zonep++; zone; zone = *zonep++) {
1169 unsigned long size = zone->present_pages;
1170 unsigned long high = zone->pages_high;
1179 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1181 unsigned int nr_free_buffer_pages(void)
1183 return nr_free_zone_pages(gfp_zone(GFP_USER));
1187 * Amount of free RAM allocatable within all zones
1189 unsigned int nr_free_pagecache_pages(void)
1191 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER));
1194 #ifdef CONFIG_HIGHMEM
1195 unsigned int nr_free_highpages (void)
1198 unsigned int pages = 0;
1200 for_each_online_pgdat(pgdat)
1201 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
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 #ifdef CONFIG_HIGHMEM
1223 val->totalhigh = totalhigh_pages;
1224 val->freehigh = nr_free_highpages();
1229 val->mem_unit = PAGE_SIZE;
1230 if (vx_flags(VXF_VIRT_MEM, 0))
1231 vx_vsi_meminfo(val);
1234 EXPORT_SYMBOL(si_meminfo);
1237 void si_meminfo_node(struct sysinfo *val, int nid)
1239 pg_data_t *pgdat = NODE_DATA(nid);
1241 val->totalram = pgdat->node_present_pages;
1242 val->freeram = nr_free_pages_pgdat(pgdat);
1243 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1244 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1245 val->mem_unit = PAGE_SIZE;
1246 if (vx_flags(VXF_VIRT_MEM, 0))
1247 vx_vsi_meminfo(val);
1251 #define K(x) ((x) << (PAGE_SHIFT-10))
1254 * Show free area list (used inside shift_scroll-lock stuff)
1255 * We also calculate the percentage fragmentation. We do this by counting the
1256 * memory on each free list with the exception of the first item on the list.
1258 void show_free_areas(void)
1260 int cpu, temperature;
1261 unsigned long active;
1262 unsigned long inactive;
1266 for_each_zone(zone) {
1268 printk("%s per-cpu:", zone->name);
1270 if (!populated_zone(zone)) {
1276 for_each_online_cpu(cpu) {
1277 struct per_cpu_pageset *pageset;
1279 pageset = zone_pcp(zone, cpu);
1281 for (temperature = 0; temperature < 2; temperature++)
1282 printk("cpu %d %s: high %d, batch %d used:%d\n",
1284 temperature ? "cold" : "hot",
1285 pageset->pcp[temperature].high,
1286 pageset->pcp[temperature].batch,
1287 pageset->pcp[temperature].count);
1291 get_zone_counts(&active, &inactive, &free);
1293 printk("Free pages: %11ukB (%ukB HighMem)\n",
1295 K(nr_free_highpages()));
1297 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1298 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1301 global_page_state(NR_FILE_DIRTY),
1302 global_page_state(NR_WRITEBACK),
1303 global_page_state(NR_UNSTABLE_NFS),
1305 global_page_state(NR_SLAB),
1306 global_page_state(NR_FILE_MAPPED),
1307 global_page_state(NR_PAGETABLE));
1309 for_each_zone(zone) {
1321 " pages_scanned:%lu"
1322 " all_unreclaimable? %s"
1325 K(zone->free_pages),
1328 K(zone->pages_high),
1330 K(zone->nr_inactive),
1331 K(zone->present_pages),
1332 zone->pages_scanned,
1333 (zone->all_unreclaimable ? "yes" : "no")
1335 printk("lowmem_reserve[]:");
1336 for (i = 0; i < MAX_NR_ZONES; i++)
1337 printk(" %lu", zone->lowmem_reserve[i]);
1341 for_each_zone(zone) {
1342 unsigned long nr[MAX_ORDER], flags, order, total = 0;
1345 printk("%s: ", zone->name);
1346 if (!populated_zone(zone)) {
1351 spin_lock_irqsave(&zone->lock, flags);
1352 for (order = 0; order < MAX_ORDER; order++) {
1353 nr[order] = zone->free_area[order].nr_free;
1354 total += nr[order] << order;
1356 spin_unlock_irqrestore(&zone->lock, flags);
1357 for (order = 0; order < MAX_ORDER; order++)
1358 printk("%lu*%lukB ", nr[order], K(1UL) << order);
1359 printk("= %lukB\n", K(total));
1362 show_swap_cache_info();
1366 * Builds allocation fallback zone lists.
1368 * Add all populated zones of a node to the zonelist.
1370 static int __meminit build_zonelists_node(pg_data_t *pgdat,
1371 struct zonelist *zonelist, int nr_zones, int zone_type)
1375 BUG_ON(zone_type > ZONE_HIGHMEM);
1378 zone = pgdat->node_zones + zone_type;
1379 if (populated_zone(zone)) {
1380 #ifndef CONFIG_HIGHMEM
1381 BUG_ON(zone_type > ZONE_NORMAL);
1383 zonelist->zones[nr_zones++] = zone;
1384 check_highest_zone(zone_type);
1388 } while (zone_type >= 0);
1392 static inline int highest_zone(int zone_bits)
1394 int res = ZONE_NORMAL;
1395 if (zone_bits & (__force int)__GFP_HIGHMEM)
1397 if (zone_bits & (__force int)__GFP_DMA32)
1399 if (zone_bits & (__force int)__GFP_DMA)
1405 #define MAX_NODE_LOAD (num_online_nodes())
1406 static int __meminitdata node_load[MAX_NUMNODES];
1408 * find_next_best_node - find the next node that should appear in a given node's fallback list
1409 * @node: node whose fallback list we're appending
1410 * @used_node_mask: nodemask_t of already used nodes
1412 * We use a number of factors to determine which is the next node that should
1413 * appear on a given node's fallback list. The node should not have appeared
1414 * already in @node's fallback list, and it should be the next closest node
1415 * according to the distance array (which contains arbitrary distance values
1416 * from each node to each node in the system), and should also prefer nodes
1417 * with no CPUs, since presumably they'll have very little allocation pressure
1418 * on them otherwise.
1419 * It returns -1 if no node is found.
1421 static int __meminit find_next_best_node(int node, nodemask_t *used_node_mask)
1424 int min_val = INT_MAX;
1427 /* Use the local node if we haven't already */
1428 if (!node_isset(node, *used_node_mask)) {
1429 node_set(node, *used_node_mask);
1433 for_each_online_node(n) {
1436 /* Don't want a node to appear more than once */
1437 if (node_isset(n, *used_node_mask))
1440 /* Use the distance array to find the distance */
1441 val = node_distance(node, n);
1443 /* Penalize nodes under us ("prefer the next node") */
1446 /* Give preference to headless and unused nodes */
1447 tmp = node_to_cpumask(n);
1448 if (!cpus_empty(tmp))
1449 val += PENALTY_FOR_NODE_WITH_CPUS;
1451 /* Slight preference for less loaded node */
1452 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1453 val += node_load[n];
1455 if (val < min_val) {
1462 node_set(best_node, *used_node_mask);
1467 static void __meminit build_zonelists(pg_data_t *pgdat)
1469 int i, j, k, node, local_node;
1470 int prev_node, load;
1471 struct zonelist *zonelist;
1472 nodemask_t used_mask;
1474 /* initialize zonelists */
1475 for (i = 0; i < GFP_ZONETYPES; i++) {
1476 zonelist = pgdat->node_zonelists + i;
1477 zonelist->zones[0] = NULL;
1480 /* NUMA-aware ordering of nodes */
1481 local_node = pgdat->node_id;
1482 load = num_online_nodes();
1483 prev_node = local_node;
1484 nodes_clear(used_mask);
1485 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1486 int distance = node_distance(local_node, node);
1489 * If another node is sufficiently far away then it is better
1490 * to reclaim pages in a zone before going off node.
1492 if (distance > RECLAIM_DISTANCE)
1493 zone_reclaim_mode = 1;
1496 * We don't want to pressure a particular node.
1497 * So adding penalty to the first node in same
1498 * distance group to make it round-robin.
1501 if (distance != node_distance(local_node, prev_node))
1502 node_load[node] += load;
1505 for (i = 0; i < GFP_ZONETYPES; i++) {
1506 zonelist = pgdat->node_zonelists + i;
1507 for (j = 0; zonelist->zones[j] != NULL; j++);
1509 k = highest_zone(i);
1511 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1512 zonelist->zones[j] = NULL;
1517 #else /* CONFIG_NUMA */
1519 static void __meminit build_zonelists(pg_data_t *pgdat)
1521 int i, j, k, node, local_node;
1523 local_node = pgdat->node_id;
1524 for (i = 0; i < GFP_ZONETYPES; i++) {
1525 struct zonelist *zonelist;
1527 zonelist = pgdat->node_zonelists + i;
1530 k = highest_zone(i);
1531 j = build_zonelists_node(pgdat, zonelist, j, k);
1533 * Now we build the zonelist so that it contains the zones
1534 * of all the other nodes.
1535 * We don't want to pressure a particular node, so when
1536 * building the zones for node N, we make sure that the
1537 * zones coming right after the local ones are those from
1538 * node N+1 (modulo N)
1540 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1541 if (!node_online(node))
1543 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1545 for (node = 0; node < local_node; node++) {
1546 if (!node_online(node))
1548 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1551 zonelist->zones[j] = NULL;
1555 #endif /* CONFIG_NUMA */
1557 /* return values int ....just for stop_machine_run() */
1558 static int __meminit __build_all_zonelists(void *dummy)
1561 for_each_online_node(nid)
1562 build_zonelists(NODE_DATA(nid));
1566 void __meminit build_all_zonelists(void)
1568 if (system_state == SYSTEM_BOOTING) {
1569 __build_all_zonelists(0);
1570 cpuset_init_current_mems_allowed();
1572 /* we have to stop all cpus to guaranntee there is no user
1574 stop_machine_run(__build_all_zonelists, NULL, NR_CPUS);
1575 /* cpuset refresh routine should be here */
1577 vm_total_pages = nr_free_pagecache_pages();
1578 printk("Built %i zonelists. Total pages: %ld\n",
1579 num_online_nodes(), vm_total_pages);
1583 * Helper functions to size the waitqueue hash table.
1584 * Essentially these want to choose hash table sizes sufficiently
1585 * large so that collisions trying to wait on pages are rare.
1586 * But in fact, the number of active page waitqueues on typical
1587 * systems is ridiculously low, less than 200. So this is even
1588 * conservative, even though it seems large.
1590 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1591 * waitqueues, i.e. the size of the waitq table given the number of pages.
1593 #define PAGES_PER_WAITQUEUE 256
1595 #ifndef CONFIG_MEMORY_HOTPLUG
1596 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
1598 unsigned long size = 1;
1600 pages /= PAGES_PER_WAITQUEUE;
1602 while (size < pages)
1606 * Once we have dozens or even hundreds of threads sleeping
1607 * on IO we've got bigger problems than wait queue collision.
1608 * Limit the size of the wait table to a reasonable size.
1610 size = min(size, 4096UL);
1612 return max(size, 4UL);
1616 * A zone's size might be changed by hot-add, so it is not possible to determine
1617 * a suitable size for its wait_table. So we use the maximum size now.
1619 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
1621 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
1622 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
1623 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
1625 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
1626 * or more by the traditional way. (See above). It equals:
1628 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
1629 * ia64(16K page size) : = ( 8G + 4M)byte.
1630 * powerpc (64K page size) : = (32G +16M)byte.
1632 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
1639 * This is an integer logarithm so that shifts can be used later
1640 * to extract the more random high bits from the multiplicative
1641 * hash function before the remainder is taken.
1643 static inline unsigned long wait_table_bits(unsigned long size)
1648 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1650 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1651 unsigned long *zones_size, unsigned long *zholes_size)
1653 unsigned long realtotalpages, totalpages = 0;
1656 for (i = 0; i < MAX_NR_ZONES; i++)
1657 totalpages += zones_size[i];
1658 pgdat->node_spanned_pages = totalpages;
1660 realtotalpages = totalpages;
1662 for (i = 0; i < MAX_NR_ZONES; i++)
1663 realtotalpages -= zholes_size[i];
1664 pgdat->node_present_pages = realtotalpages;
1665 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1670 * Initially all pages are reserved - free ones are freed
1671 * up by free_all_bootmem() once the early boot process is
1672 * done. Non-atomic initialization, single-pass.
1674 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1675 unsigned long start_pfn)
1678 unsigned long end_pfn = start_pfn + size;
1681 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1682 if (!early_pfn_valid(pfn))
1684 if (!early_pfn_in_nid(pfn, nid))
1686 page = pfn_to_page(pfn);
1687 set_page_links(page, zone, nid, pfn);
1688 init_page_count(page);
1689 reset_page_mapcount(page);
1690 SetPageReserved(page);
1691 INIT_LIST_HEAD(&page->lru);
1692 #ifdef WANT_PAGE_VIRTUAL
1693 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1694 if (!is_highmem_idx(zone))
1695 set_page_address(page, __va(pfn << PAGE_SHIFT));
1700 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1704 for (order = 0; order < MAX_ORDER ; order++) {
1705 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1706 zone->free_area[order].nr_free = 0;
1710 #define ZONETABLE_INDEX(x, zone_nr) ((x << ZONES_SHIFT) | zone_nr)
1711 void zonetable_add(struct zone *zone, int nid, int zid, unsigned long pfn,
1714 unsigned long snum = pfn_to_section_nr(pfn);
1715 unsigned long end = pfn_to_section_nr(pfn + size);
1718 zone_table[ZONETABLE_INDEX(nid, zid)] = zone;
1720 for (; snum <= end; snum++)
1721 zone_table[ZONETABLE_INDEX(snum, zid)] = zone;
1724 #ifndef __HAVE_ARCH_MEMMAP_INIT
1725 #define memmap_init(size, nid, zone, start_pfn) \
1726 memmap_init_zone((size), (nid), (zone), (start_pfn))
1729 static int __cpuinit zone_batchsize(struct zone *zone)
1734 * The per-cpu-pages pools are set to around 1000th of the
1735 * size of the zone. But no more than 1/2 of a meg.
1737 * OK, so we don't know how big the cache is. So guess.
1739 batch = zone->present_pages / 1024;
1740 if (batch * PAGE_SIZE > 512 * 1024)
1741 batch = (512 * 1024) / PAGE_SIZE;
1742 batch /= 4; /* We effectively *= 4 below */
1747 * Clamp the batch to a 2^n - 1 value. Having a power
1748 * of 2 value was found to be more likely to have
1749 * suboptimal cache aliasing properties in some cases.
1751 * For example if 2 tasks are alternately allocating
1752 * batches of pages, one task can end up with a lot
1753 * of pages of one half of the possible page colors
1754 * and the other with pages of the other colors.
1756 batch = (1 << (fls(batch + batch/2)-1)) - 1;
1761 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
1763 struct per_cpu_pages *pcp;
1765 memset(p, 0, sizeof(*p));
1767 pcp = &p->pcp[0]; /* hot */
1769 pcp->high = 6 * batch;
1770 pcp->batch = max(1UL, 1 * batch);
1771 INIT_LIST_HEAD(&pcp->list);
1773 pcp = &p->pcp[1]; /* cold*/
1775 pcp->high = 2 * batch;
1776 pcp->batch = max(1UL, batch/2);
1777 INIT_LIST_HEAD(&pcp->list);
1781 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
1782 * to the value high for the pageset p.
1785 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
1788 struct per_cpu_pages *pcp;
1790 pcp = &p->pcp[0]; /* hot list */
1792 pcp->batch = max(1UL, high/4);
1793 if ((high/4) > (PAGE_SHIFT * 8))
1794 pcp->batch = PAGE_SHIFT * 8;
1800 * Boot pageset table. One per cpu which is going to be used for all
1801 * zones and all nodes. The parameters will be set in such a way
1802 * that an item put on a list will immediately be handed over to
1803 * the buddy list. This is safe since pageset manipulation is done
1804 * with interrupts disabled.
1806 * Some NUMA counter updates may also be caught by the boot pagesets.
1808 * The boot_pagesets must be kept even after bootup is complete for
1809 * unused processors and/or zones. They do play a role for bootstrapping
1810 * hotplugged processors.
1812 * zoneinfo_show() and maybe other functions do
1813 * not check if the processor is online before following the pageset pointer.
1814 * Other parts of the kernel may not check if the zone is available.
1816 static struct per_cpu_pageset boot_pageset[NR_CPUS];
1819 * Dynamically allocate memory for the
1820 * per cpu pageset array in struct zone.
1822 static int __cpuinit process_zones(int cpu)
1824 struct zone *zone, *dzone;
1826 for_each_zone(zone) {
1828 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
1829 GFP_KERNEL, cpu_to_node(cpu));
1830 if (!zone_pcp(zone, cpu))
1833 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
1835 if (percpu_pagelist_fraction)
1836 setup_pagelist_highmark(zone_pcp(zone, cpu),
1837 (zone->present_pages / percpu_pagelist_fraction));
1842 for_each_zone(dzone) {
1845 kfree(zone_pcp(dzone, cpu));
1846 zone_pcp(dzone, cpu) = NULL;
1851 static inline void free_zone_pagesets(int cpu)
1855 for_each_zone(zone) {
1856 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
1858 /* Free per_cpu_pageset if it is slab allocated */
1859 if (pset != &boot_pageset[cpu])
1861 zone_pcp(zone, cpu) = NULL;
1865 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
1866 unsigned long action,
1869 int cpu = (long)hcpu;
1870 int ret = NOTIFY_OK;
1873 case CPU_UP_PREPARE:
1874 if (process_zones(cpu))
1877 case CPU_UP_CANCELED:
1879 free_zone_pagesets(cpu);
1887 static struct notifier_block __cpuinitdata pageset_notifier =
1888 { &pageset_cpuup_callback, NULL, 0 };
1890 void __init setup_per_cpu_pageset(void)
1894 /* Initialize per_cpu_pageset for cpu 0.
1895 * A cpuup callback will do this for every cpu
1896 * as it comes online
1898 err = process_zones(smp_processor_id());
1900 register_cpu_notifier(&pageset_notifier);
1906 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
1909 struct pglist_data *pgdat = zone->zone_pgdat;
1913 * The per-page waitqueue mechanism uses hashed waitqueues
1916 zone->wait_table_hash_nr_entries =
1917 wait_table_hash_nr_entries(zone_size_pages);
1918 zone->wait_table_bits =
1919 wait_table_bits(zone->wait_table_hash_nr_entries);
1920 alloc_size = zone->wait_table_hash_nr_entries
1921 * sizeof(wait_queue_head_t);
1923 if (system_state == SYSTEM_BOOTING) {
1924 zone->wait_table = (wait_queue_head_t *)
1925 alloc_bootmem_node(pgdat, alloc_size);
1928 * This case means that a zone whose size was 0 gets new memory
1929 * via memory hot-add.
1930 * But it may be the case that a new node was hot-added. In
1931 * this case vmalloc() will not be able to use this new node's
1932 * memory - this wait_table must be initialized to use this new
1933 * node itself as well.
1934 * To use this new node's memory, further consideration will be
1937 zone->wait_table = (wait_queue_head_t *)vmalloc(alloc_size);
1939 if (!zone->wait_table)
1942 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
1943 init_waitqueue_head(zone->wait_table + i);
1948 static __meminit void zone_pcp_init(struct zone *zone)
1951 unsigned long batch = zone_batchsize(zone);
1953 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1955 /* Early boot. Slab allocator not functional yet */
1956 zone_pcp(zone, cpu) = &boot_pageset[cpu];
1957 setup_pageset(&boot_pageset[cpu],0);
1959 setup_pageset(zone_pcp(zone,cpu), batch);
1962 if (zone->present_pages)
1963 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
1964 zone->name, zone->present_pages, batch);
1967 __meminit int init_currently_empty_zone(struct zone *zone,
1968 unsigned long zone_start_pfn,
1971 struct pglist_data *pgdat = zone->zone_pgdat;
1973 ret = zone_wait_table_init(zone, size);
1976 pgdat->nr_zones = zone_idx(zone) + 1;
1978 zone->zone_start_pfn = zone_start_pfn;
1980 memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn);
1982 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1988 * Set up the zone data structures:
1989 * - mark all pages reserved
1990 * - mark all memory queues empty
1991 * - clear the memory bitmaps
1993 static void __meminit free_area_init_core(struct pglist_data *pgdat,
1994 unsigned long *zones_size, unsigned long *zholes_size)
1997 int nid = pgdat->node_id;
1998 unsigned long zone_start_pfn = pgdat->node_start_pfn;
2001 pgdat_resize_init(pgdat);
2002 pgdat->nr_zones = 0;
2003 init_waitqueue_head(&pgdat->kswapd_wait);
2004 pgdat->kswapd_max_order = 0;
2006 for (j = 0; j < MAX_NR_ZONES; j++) {
2007 struct zone *zone = pgdat->node_zones + j;
2008 unsigned long size, realsize;
2010 realsize = size = zones_size[j];
2012 realsize -= zholes_size[j];
2014 if (j < ZONE_HIGHMEM)
2015 nr_kernel_pages += realsize;
2016 nr_all_pages += realsize;
2018 zone->spanned_pages = size;
2019 zone->present_pages = realsize;
2021 zone->min_unmapped_ratio = (realsize*sysctl_min_unmapped_ratio)
2023 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
2025 zone->name = zone_names[j];
2026 spin_lock_init(&zone->lock);
2027 spin_lock_init(&zone->lru_lock);
2028 zone_seqlock_init(zone);
2029 zone->zone_pgdat = pgdat;
2030 zone->free_pages = 0;
2032 zone->prev_priority = DEF_PRIORITY;
2034 zone_pcp_init(zone);
2035 INIT_LIST_HEAD(&zone->active_list);
2036 INIT_LIST_HEAD(&zone->inactive_list);
2037 zone->nr_scan_active = 0;
2038 zone->nr_scan_inactive = 0;
2039 zone->nr_active = 0;
2040 zone->nr_inactive = 0;
2041 zap_zone_vm_stats(zone);
2042 atomic_set(&zone->reclaim_in_progress, 0);
2046 zonetable_add(zone, nid, j, zone_start_pfn, size);
2047 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
2049 zone_start_pfn += size;
2053 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
2055 /* Skip empty nodes */
2056 if (!pgdat->node_spanned_pages)
2059 #ifdef CONFIG_FLAT_NODE_MEM_MAP
2060 /* ia64 gets its own node_mem_map, before this, without bootmem */
2061 if (!pgdat->node_mem_map) {
2062 unsigned long size, start, end;
2066 * The zone's endpoints aren't required to be MAX_ORDER
2067 * aligned but the node_mem_map endpoints must be in order
2068 * for the buddy allocator to function correctly.
2070 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
2071 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
2072 end = ALIGN(end, MAX_ORDER_NR_PAGES);
2073 size = (end - start) * sizeof(struct page);
2074 map = alloc_remap(pgdat->node_id, size);
2076 map = alloc_bootmem_node(pgdat, size);
2077 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
2079 #ifdef CONFIG_FLATMEM
2081 * With no DISCONTIG, the global mem_map is just set as node 0's
2083 if (pgdat == NODE_DATA(0))
2084 mem_map = NODE_DATA(0)->node_mem_map;
2086 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
2089 void __meminit free_area_init_node(int nid, struct pglist_data *pgdat,
2090 unsigned long *zones_size, unsigned long node_start_pfn,
2091 unsigned long *zholes_size)
2093 pgdat->node_id = nid;
2094 pgdat->node_start_pfn = node_start_pfn;
2095 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
2097 alloc_node_mem_map(pgdat);
2099 free_area_init_core(pgdat, zones_size, zholes_size);
2102 #ifndef CONFIG_NEED_MULTIPLE_NODES
2103 static bootmem_data_t contig_bootmem_data;
2104 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
2106 EXPORT_SYMBOL(contig_page_data);
2109 void __init free_area_init(unsigned long *zones_size)
2111 free_area_init_node(0, NODE_DATA(0), zones_size,
2112 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
2115 #ifdef CONFIG_HOTPLUG_CPU
2116 static int page_alloc_cpu_notify(struct notifier_block *self,
2117 unsigned long action, void *hcpu)
2119 int cpu = (unsigned long)hcpu;
2121 if (action == CPU_DEAD) {
2122 local_irq_disable();
2124 vm_events_fold_cpu(cpu);
2126 refresh_cpu_vm_stats(cpu);
2130 #endif /* CONFIG_HOTPLUG_CPU */
2132 void __init page_alloc_init(void)
2134 hotcpu_notifier(page_alloc_cpu_notify, 0);
2138 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
2139 * or min_free_kbytes changes.
2141 static void calculate_totalreserve_pages(void)
2143 struct pglist_data *pgdat;
2144 unsigned long reserve_pages = 0;
2147 for_each_online_pgdat(pgdat) {
2148 for (i = 0; i < MAX_NR_ZONES; i++) {
2149 struct zone *zone = pgdat->node_zones + i;
2150 unsigned long max = 0;
2152 /* Find valid and maximum lowmem_reserve in the zone */
2153 for (j = i; j < MAX_NR_ZONES; j++) {
2154 if (zone->lowmem_reserve[j] > max)
2155 max = zone->lowmem_reserve[j];
2158 /* we treat pages_high as reserved pages. */
2159 max += zone->pages_high;
2161 if (max > zone->present_pages)
2162 max = zone->present_pages;
2163 reserve_pages += max;
2166 totalreserve_pages = reserve_pages;
2170 * setup_per_zone_lowmem_reserve - called whenever
2171 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
2172 * has a correct pages reserved value, so an adequate number of
2173 * pages are left in the zone after a successful __alloc_pages().
2175 static void setup_per_zone_lowmem_reserve(void)
2177 struct pglist_data *pgdat;
2180 for_each_online_pgdat(pgdat) {
2181 for (j = 0; j < MAX_NR_ZONES; j++) {
2182 struct zone *zone = pgdat->node_zones + j;
2183 unsigned long present_pages = zone->present_pages;
2185 zone->lowmem_reserve[j] = 0;
2187 for (idx = j-1; idx >= 0; idx--) {
2188 struct zone *lower_zone;
2190 if (sysctl_lowmem_reserve_ratio[idx] < 1)
2191 sysctl_lowmem_reserve_ratio[idx] = 1;
2193 lower_zone = pgdat->node_zones + idx;
2194 lower_zone->lowmem_reserve[j] = present_pages /
2195 sysctl_lowmem_reserve_ratio[idx];
2196 present_pages += lower_zone->present_pages;
2201 /* update totalreserve_pages */
2202 calculate_totalreserve_pages();
2206 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
2207 * that the pages_{min,low,high} values for each zone are set correctly
2208 * with respect to min_free_kbytes.
2210 void setup_per_zone_pages_min(void)
2212 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
2213 unsigned long lowmem_pages = 0;
2215 unsigned long flags;
2217 /* Calculate total number of !ZONE_HIGHMEM pages */
2218 for_each_zone(zone) {
2219 if (!is_highmem(zone))
2220 lowmem_pages += zone->present_pages;
2223 for_each_zone(zone) {
2226 spin_lock_irqsave(&zone->lru_lock, flags);
2227 tmp = (u64)pages_min * zone->present_pages;
2228 do_div(tmp, lowmem_pages);
2229 if (is_highmem(zone)) {
2231 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
2232 * need highmem pages, so cap pages_min to a small
2235 * The (pages_high-pages_low) and (pages_low-pages_min)
2236 * deltas controls asynch page reclaim, and so should
2237 * not be capped for highmem.
2241 min_pages = zone->present_pages / 1024;
2242 if (min_pages < SWAP_CLUSTER_MAX)
2243 min_pages = SWAP_CLUSTER_MAX;
2244 if (min_pages > 128)
2246 zone->pages_min = min_pages;
2249 * If it's a lowmem zone, reserve a number of pages
2250 * proportionate to the zone's size.
2252 zone->pages_min = tmp;
2255 zone->pages_low = zone->pages_min + (tmp >> 2);
2256 zone->pages_high = zone->pages_min + (tmp >> 1);
2257 spin_unlock_irqrestore(&zone->lru_lock, flags);
2260 /* update totalreserve_pages */
2261 calculate_totalreserve_pages();
2265 * Initialise min_free_kbytes.
2267 * For small machines we want it small (128k min). For large machines
2268 * we want it large (64MB max). But it is not linear, because network
2269 * bandwidth does not increase linearly with machine size. We use
2271 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2272 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2288 static int __init init_per_zone_pages_min(void)
2290 unsigned long lowmem_kbytes;
2292 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2294 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2295 if (min_free_kbytes < 128)
2296 min_free_kbytes = 128;
2297 if (min_free_kbytes > 65536)
2298 min_free_kbytes = 65536;
2299 setup_per_zone_pages_min();
2300 setup_per_zone_lowmem_reserve();
2303 module_init(init_per_zone_pages_min)
2306 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2307 * that we can call two helper functions whenever min_free_kbytes
2310 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2311 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2313 proc_dointvec(table, write, file, buffer, length, ppos);
2314 setup_per_zone_pages_min();
2319 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
2320 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2325 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2330 zone->min_unmapped_ratio = (zone->present_pages *
2331 sysctl_min_unmapped_ratio) / 100;
2335 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
2336 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2341 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2346 zone->min_slab_pages = (zone->present_pages *
2347 sysctl_min_slab_ratio) / 100;
2353 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2354 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2355 * whenever sysctl_lowmem_reserve_ratio changes.
2357 * The reserve ratio obviously has absolutely no relation with the
2358 * pages_min watermarks. The lowmem reserve ratio can only make sense
2359 * if in function of the boot time zone sizes.
2361 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
2362 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2364 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2365 setup_per_zone_lowmem_reserve();
2370 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
2371 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
2372 * can have before it gets flushed back to buddy allocator.
2375 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
2376 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2382 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2383 if (!write || (ret == -EINVAL))
2385 for_each_zone(zone) {
2386 for_each_online_cpu(cpu) {
2388 high = zone->present_pages / percpu_pagelist_fraction;
2389 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
2395 __initdata int hashdist = HASHDIST_DEFAULT;
2398 static int __init set_hashdist(char *str)
2402 hashdist = simple_strtoul(str, &str, 0);
2405 __setup("hashdist=", set_hashdist);
2409 * allocate a large system hash table from bootmem
2410 * - it is assumed that the hash table must contain an exact power-of-2
2411 * quantity of entries
2412 * - limit is the number of hash buckets, not the total allocation size
2414 void *__init alloc_large_system_hash(const char *tablename,
2415 unsigned long bucketsize,
2416 unsigned long numentries,
2419 unsigned int *_hash_shift,
2420 unsigned int *_hash_mask,
2421 unsigned long limit)
2423 unsigned long long max = limit;
2424 unsigned long log2qty, size;
2427 /* allow the kernel cmdline to have a say */
2429 /* round applicable memory size up to nearest megabyte */
2430 numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages;
2431 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2432 numentries >>= 20 - PAGE_SHIFT;
2433 numentries <<= 20 - PAGE_SHIFT;
2435 /* limit to 1 bucket per 2^scale bytes of low memory */
2436 if (scale > PAGE_SHIFT)
2437 numentries >>= (scale - PAGE_SHIFT);
2439 numentries <<= (PAGE_SHIFT - scale);
2441 numentries = roundup_pow_of_two(numentries);
2443 /* limit allocation size to 1/16 total memory by default */
2445 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2446 do_div(max, bucketsize);
2449 if (numentries > max)
2452 log2qty = long_log2(numentries);
2455 size = bucketsize << log2qty;
2456 if (flags & HASH_EARLY)
2457 table = alloc_bootmem(size);
2459 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
2461 unsigned long order;
2462 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
2464 table = (void*) __get_free_pages(GFP_ATOMIC, order);
2466 } while (!table && size > PAGE_SIZE && --log2qty);
2469 panic("Failed to allocate %s hash table\n", tablename);
2471 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2474 long_log2(size) - PAGE_SHIFT,
2478 *_hash_shift = log2qty;
2480 *_hash_mask = (1 << log2qty) - 1;
2485 #ifdef CONFIG_OUT_OF_LINE_PFN_TO_PAGE
2486 struct page *pfn_to_page(unsigned long pfn)
2488 return __pfn_to_page(pfn);
2490 unsigned long page_to_pfn(struct page *page)
2492 return __page_to_pfn(page);
2494 EXPORT_SYMBOL(pfn_to_page);
2495 EXPORT_SYMBOL(page_to_pfn);
2496 #endif /* CONFIG_OUT_OF_LINE_PFN_TO_PAGE */