2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/config.h>
18 #include <linux/stddef.h>
20 #include <linux/swap.h>
21 #include <linux/interrupt.h>
22 #include <linux/pagemap.h>
23 #include <linux/bootmem.h>
24 #include <linux/compiler.h>
25 #include <linux/kernel.h>
26 #include <linux/module.h>
27 #include <linux/suspend.h>
28 #include <linux/pagevec.h>
29 #include <linux/blkdev.h>
30 #include <linux/slab.h>
31 #include <linux/notifier.h>
32 #include <linux/topology.h>
33 #include <linux/sysctl.h>
34 #include <linux/cpu.h>
35 #include <linux/cpuset.h>
36 #include <linux/memory_hotplug.h>
37 #include <linux/nodemask.h>
38 #include <linux/vmalloc.h>
40 #include <asm/tlbflush.h>
44 * MCD - HACK: Find somewhere to initialize this EARLY, or make this
47 nodemask_t node_online_map __read_mostly = { { [0] = 1UL } };
48 EXPORT_SYMBOL(node_online_map);
49 nodemask_t node_possible_map __read_mostly = NODE_MASK_ALL;
50 EXPORT_SYMBOL(node_possible_map);
51 struct pglist_data *pgdat_list __read_mostly;
52 unsigned long totalram_pages __read_mostly;
53 unsigned long totalhigh_pages __read_mostly;
57 * results with 256, 32 in the lowmem_reserve sysctl:
58 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
59 * 1G machine -> (16M dma, 784M normal, 224M high)
60 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
61 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
62 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
64 * TBD: should special case ZONE_DMA32 machines here - in those we normally
65 * don't need any ZONE_NORMAL reservation
67 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 256, 256, 32 };
69 EXPORT_SYMBOL(totalram_pages);
72 * Used by page_zone() to look up the address of the struct zone whose
73 * id is encoded in the upper bits of page->flags
75 struct zone *zone_table[1 << ZONETABLE_SHIFT] __read_mostly;
76 EXPORT_SYMBOL(zone_table);
78 static char *zone_names[MAX_NR_ZONES] = { "DMA", "DMA32", "Normal", "HighMem" };
79 int min_free_kbytes = 1024;
81 unsigned long __initdata nr_kernel_pages;
82 unsigned long __initdata nr_all_pages;
84 #ifdef CONFIG_DEBUG_VM
85 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
89 unsigned long pfn = page_to_pfn(page);
92 seq = zone_span_seqbegin(zone);
93 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
95 else if (pfn < zone->zone_start_pfn)
97 } while (zone_span_seqretry(zone, seq));
102 static int page_is_consistent(struct zone *zone, struct page *page)
104 #ifdef CONFIG_HOLES_IN_ZONE
105 if (!pfn_valid(page_to_pfn(page)))
108 if (zone != page_zone(page))
114 * Temporary debugging check for pages not lying within a given zone.
116 static int bad_range(struct zone *zone, struct page *page)
118 if (page_outside_zone_boundaries(zone, page))
120 if (!page_is_consistent(zone, page))
127 static inline int bad_range(struct zone *zone, struct page *page)
133 static void bad_page(const char *function, struct page *page)
135 printk(KERN_EMERG "Bad page state at %s (in process '%s', page %p)\n",
136 function, current->comm, page);
137 printk(KERN_EMERG "flags:0x%0*lx mapping:%p mapcount:%d count:%d\n",
138 (int)(2*sizeof(unsigned long)), (unsigned long)page->flags,
139 page->mapping, page_mapcount(page), page_count(page));
140 printk(KERN_EMERG "Backtrace:\n");
142 printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n");
143 page->flags &= ~(1 << PG_lru |
152 set_page_count(page, 0);
153 reset_page_mapcount(page);
154 page->mapping = NULL;
155 add_taint(TAINT_BAD_PAGE);
159 * Higher-order pages are called "compound pages". They are structured thusly:
161 * The first PAGE_SIZE page is called the "head page".
163 * The remaining PAGE_SIZE pages are called "tail pages".
165 * All pages have PG_compound set. All pages have their ->private pointing at
166 * the head page (even the head page has this).
168 * The first tail page's ->mapping, if non-zero, holds the address of the
169 * compound page's put_page() function.
171 * The order of the allocation is stored in the first tail page's ->index
172 * This is only for debug at present. This usage means that zero-order pages
173 * may not be compound.
175 static void prep_compound_page(struct page *page, unsigned long order)
178 int nr_pages = 1 << order;
180 page[1].mapping = NULL;
181 page[1].index = order;
182 for (i = 0; i < nr_pages; i++) {
183 struct page *p = page + i;
186 set_page_private(p, (unsigned long)page);
190 static void destroy_compound_page(struct page *page, unsigned long order)
193 int nr_pages = 1 << order;
195 if (!PageCompound(page))
198 if (page[1].index != order)
199 bad_page(__FUNCTION__, page);
201 for (i = 0; i < nr_pages; i++) {
202 struct page *p = page + i;
204 if (!PageCompound(p))
205 bad_page(__FUNCTION__, page);
206 if (page_private(p) != (unsigned long)page)
207 bad_page(__FUNCTION__, page);
208 ClearPageCompound(p);
213 * function for dealing with page's order in buddy system.
214 * zone->lock is already acquired when we use these.
215 * So, we don't need atomic page->flags operations here.
217 static inline unsigned long page_order(struct page *page) {
218 return page_private(page);
221 static inline void set_page_order(struct page *page, int order) {
222 set_page_private(page, order);
223 __SetPagePrivate(page);
226 static inline void rmv_page_order(struct page *page)
228 __ClearPagePrivate(page);
229 set_page_private(page, 0);
233 * Locate the struct page for both the matching buddy in our
234 * pair (buddy1) and the combined O(n+1) page they form (page).
236 * 1) Any buddy B1 will have an order O twin B2 which satisfies
237 * the following equation:
239 * For example, if the starting buddy (buddy2) is #8 its order
241 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
243 * 2) Any buddy B will have an order O+1 parent P which
244 * satisfies the following equation:
247 * Assumption: *_mem_map is contigious at least up to MAX_ORDER
249 static inline struct page *
250 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
252 unsigned long buddy_idx = page_idx ^ (1 << order);
254 return page + (buddy_idx - page_idx);
257 static inline unsigned long
258 __find_combined_index(unsigned long page_idx, unsigned int order)
260 return (page_idx & ~(1 << order));
264 * This function checks whether a page is free && is the buddy
265 * we can do coalesce a page and its buddy if
266 * (a) the buddy is not in a hole &&
267 * (b) the buddy is free &&
268 * (c) the buddy is on the buddy system &&
269 * (d) a page and its buddy have the same order.
270 * for recording page's order, we use page_private(page) and PG_private.
273 static inline int page_is_buddy(struct page *page, int order)
275 #ifdef CONFIG_HOLES_IN_ZONE
276 if (!pfn_valid(page_to_pfn(page)))
280 if (PagePrivate(page) &&
281 (page_order(page) == order) &&
282 page_count(page) == 0)
288 * Freeing function for a buddy system allocator.
290 * The concept of a buddy system is to maintain direct-mapped table
291 * (containing bit values) for memory blocks of various "orders".
292 * The bottom level table contains the map for the smallest allocatable
293 * units of memory (here, pages), and each level above it describes
294 * pairs of units from the levels below, hence, "buddies".
295 * At a high level, all that happens here is marking the table entry
296 * at the bottom level available, and propagating the changes upward
297 * as necessary, plus some accounting needed to play nicely with other
298 * parts of the VM system.
299 * At each level, we keep a list of pages, which are heads of continuous
300 * free pages of length of (1 << order) and marked with PG_Private.Page's
301 * order is recorded in page_private(page) field.
302 * So when we are allocating or freeing one, we can derive the state of the
303 * other. That is, if we allocate a small block, and both were
304 * free, the remainder of the region must be split into blocks.
305 * If a block is freed, and its buddy is also free, then this
306 * triggers coalescing into a block of larger size.
311 static inline void __free_pages_bulk (struct page *page,
312 struct zone *zone, unsigned int order)
314 unsigned long page_idx;
315 int order_size = 1 << order;
318 destroy_compound_page(page, order);
320 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
322 BUG_ON(page_idx & (order_size - 1));
323 BUG_ON(bad_range(zone, page));
325 zone->free_pages += order_size;
326 while (order < MAX_ORDER-1) {
327 unsigned long combined_idx;
328 struct free_area *area;
331 buddy = __page_find_buddy(page, page_idx, order);
332 if (!page_is_buddy(buddy, order))
333 break; /* Move the buddy up one level. */
335 list_del(&buddy->lru);
336 area = zone->free_area + order;
338 rmv_page_order(buddy);
339 combined_idx = __find_combined_index(page_idx, order);
340 page = page + (combined_idx - page_idx);
341 page_idx = combined_idx;
344 set_page_order(page, order);
345 list_add(&page->lru, &zone->free_area[order].free_list);
346 zone->free_area[order].nr_free++;
349 static inline int free_pages_check(const char *function, struct page *page)
351 if (unlikely(page_mapcount(page) |
352 (page->mapping != NULL) |
353 (page_count(page) != 0) |
363 1 << PG_reserved ))))
364 bad_page(function, page);
366 __ClearPageDirty(page);
368 * For now, we report if PG_reserved was found set, but do not
369 * clear it, and do not free the page. But we shall soon need
370 * to do more, for when the ZERO_PAGE count wraps negative.
372 return PageReserved(page);
376 * Frees a list of pages.
377 * Assumes all pages on list are in same zone, and of same order.
378 * count is the number of pages to free.
380 * If the zone was previously in an "all pages pinned" state then look to
381 * see if this freeing clears that state.
383 * And clear the zone's pages_scanned counter, to hold off the "all pages are
384 * pinned" detection logic.
387 free_pages_bulk(struct zone *zone, int count,
388 struct list_head *list, unsigned int order)
390 struct page *page = NULL;
393 spin_lock(&zone->lock);
394 zone->all_unreclaimable = 0;
395 zone->pages_scanned = 0;
396 while (!list_empty(list) && count--) {
397 page = list_entry(list->prev, struct page, lru);
398 /* have to delete it as __free_pages_bulk list manipulates */
399 list_del(&page->lru);
400 __free_pages_bulk(page, zone, order);
403 spin_unlock(&zone->lock);
407 void __free_pages_ok(struct page *page, unsigned int order)
414 arch_free_page(page, order);
418 for (i = 1 ; i < (1 << order) ; ++i)
419 __put_page(page + i);
422 for (i = 0 ; i < (1 << order) ; ++i)
423 reserved += free_pages_check(__FUNCTION__, page + i);
427 list_add(&page->lru, &list);
428 mod_page_state(pgfree, 1 << order);
429 kernel_map_pages(page, 1<<order, 0);
430 local_irq_save(flags);
431 free_pages_bulk(page_zone(page), 1, &list, order);
432 local_irq_restore(flags);
437 * The order of subdivision here is critical for the IO subsystem.
438 * Please do not alter this order without good reasons and regression
439 * testing. Specifically, as large blocks of memory are subdivided,
440 * the order in which smaller blocks are delivered depends on the order
441 * they're subdivided in this function. This is the primary factor
442 * influencing the order in which pages are delivered to the IO
443 * subsystem according to empirical testing, and this is also justified
444 * by considering the behavior of a buddy system containing a single
445 * large block of memory acted on by a series of small allocations.
446 * This behavior is a critical factor in sglist merging's success.
450 static inline struct page *
451 expand(struct zone *zone, struct page *page,
452 int low, int high, struct free_area *area)
454 unsigned long size = 1 << high;
460 BUG_ON(bad_range(zone, &page[size]));
461 list_add(&page[size].lru, &area->free_list);
463 set_page_order(&page[size], high);
469 * This page is about to be returned from the page allocator
471 static int prep_new_page(struct page *page, int order)
473 if (unlikely(page_mapcount(page) |
474 (page->mapping != NULL) |
475 (page_count(page) != 0) |
486 1 << PG_reserved ))))
487 bad_page(__FUNCTION__, page);
490 * For now, we report if PG_reserved was found set, but do not
491 * clear it, and do not allocate the page: as a safety net.
493 if (PageReserved(page))
496 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
497 1 << PG_referenced | 1 << PG_arch_1 |
498 1 << PG_checked | 1 << PG_mappedtodisk);
499 set_page_private(page, 0);
500 set_page_refs(page, order);
501 kernel_map_pages(page, 1 << order, 1);
506 * Do the hard work of removing an element from the buddy allocator.
507 * Call me with the zone->lock already held.
509 static struct page *__rmqueue(struct zone *zone, unsigned int order)
511 struct free_area * area;
512 unsigned int current_order;
515 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
516 area = zone->free_area + current_order;
517 if (list_empty(&area->free_list))
520 page = list_entry(area->free_list.next, struct page, lru);
521 list_del(&page->lru);
522 rmv_page_order(page);
524 zone->free_pages -= 1UL << order;
525 return expand(zone, page, order, current_order, area);
532 * Obtain a specified number of elements from the buddy allocator, all under
533 * a single hold of the lock, for efficiency. Add them to the supplied list.
534 * Returns the number of new pages which were placed at *list.
536 static int rmqueue_bulk(struct zone *zone, unsigned int order,
537 unsigned long count, struct list_head *list)
543 spin_lock(&zone->lock);
544 for (i = 0; i < count; ++i) {
545 page = __rmqueue(zone, order);
549 list_add_tail(&page->lru, list);
551 spin_unlock(&zone->lock);
556 /* Called from the slab reaper to drain remote pagesets */
557 void drain_remote_pages(void)
563 local_irq_save(flags);
564 for_each_zone(zone) {
565 struct per_cpu_pageset *pset;
567 /* Do not drain local pagesets */
568 if (zone->zone_pgdat->node_id == numa_node_id())
571 pset = zone->pageset[smp_processor_id()];
572 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
573 struct per_cpu_pages *pcp;
577 pcp->count -= free_pages_bulk(zone, pcp->count,
581 local_irq_restore(flags);
585 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
586 static void __drain_pages(unsigned int cpu)
592 for_each_zone(zone) {
593 struct per_cpu_pageset *pset;
595 pset = zone_pcp(zone, cpu);
596 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
597 struct per_cpu_pages *pcp;
600 local_irq_save(flags);
601 pcp->count -= free_pages_bulk(zone, pcp->count,
603 local_irq_restore(flags);
607 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
611 void mark_free_pages(struct zone *zone)
613 unsigned long zone_pfn, flags;
615 struct list_head *curr;
617 if (!zone->spanned_pages)
620 spin_lock_irqsave(&zone->lock, flags);
621 for (zone_pfn = 0; zone_pfn < zone->spanned_pages; ++zone_pfn)
622 ClearPageNosaveFree(pfn_to_page(zone_pfn + zone->zone_start_pfn));
624 for (order = MAX_ORDER - 1; order >= 0; --order)
625 list_for_each(curr, &zone->free_area[order].free_list) {
626 unsigned long start_pfn, i;
628 start_pfn = page_to_pfn(list_entry(curr, struct page, lru));
630 for (i=0; i < (1<<order); i++)
631 SetPageNosaveFree(pfn_to_page(start_pfn+i));
633 spin_unlock_irqrestore(&zone->lock, flags);
637 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
639 void drain_local_pages(void)
643 local_irq_save(flags);
644 __drain_pages(smp_processor_id());
645 local_irq_restore(flags);
647 #endif /* CONFIG_PM */
649 static void zone_statistics(struct zonelist *zonelist, struct zone *z)
654 pg_data_t *pg = z->zone_pgdat;
655 pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
656 struct per_cpu_pageset *p;
658 local_irq_save(flags);
659 cpu = smp_processor_id();
665 zone_pcp(zonelist->zones[0], cpu)->numa_foreign++;
667 if (pg == NODE_DATA(numa_node_id()))
671 local_irq_restore(flags);
676 * Free a 0-order page
678 static void FASTCALL(free_hot_cold_page(struct page *page, int cold));
679 static void fastcall free_hot_cold_page(struct page *page, int cold)
681 struct zone *zone = page_zone(page);
682 struct per_cpu_pages *pcp;
685 arch_free_page(page, 0);
688 page->mapping = NULL;
689 if (free_pages_check(__FUNCTION__, page))
692 inc_page_state(pgfree);
693 kernel_map_pages(page, 1, 0);
695 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
696 local_irq_save(flags);
697 list_add(&page->lru, &pcp->list);
699 if (pcp->count >= pcp->high)
700 pcp->count -= free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
701 local_irq_restore(flags);
705 void fastcall free_hot_page(struct page *page)
707 free_hot_cold_page(page, 0);
710 void fastcall free_cold_page(struct page *page)
712 free_hot_cold_page(page, 1);
715 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
719 BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
720 for(i = 0; i < (1 << order); i++)
721 clear_highpage(page + i);
725 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
726 * we cheat by calling it from here, in the order > 0 path. Saves a branch
730 buffered_rmqueue(struct zone *zone, int order, gfp_t gfp_flags)
734 int cold = !!(gfp_flags & __GFP_COLD);
738 struct per_cpu_pages *pcp;
741 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
742 local_irq_save(flags);
744 pcp->count += rmqueue_bulk(zone, 0,
745 pcp->batch, &pcp->list);
746 if (likely(pcp->count)) {
747 page = list_entry(pcp->list.next, struct page, lru);
748 list_del(&page->lru);
751 local_irq_restore(flags);
754 spin_lock_irqsave(&zone->lock, flags);
755 page = __rmqueue(zone, order);
756 spin_unlock_irqrestore(&zone->lock, flags);
760 BUG_ON(bad_range(zone, page));
761 mod_page_state_zone(zone, pgalloc, 1 << order);
762 if (prep_new_page(page, order))
765 if (gfp_flags & __GFP_ZERO)
766 prep_zero_page(page, order, gfp_flags);
768 if (order && (gfp_flags & __GFP_COMP))
769 prep_compound_page(page, order);
774 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
775 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
776 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
777 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
778 #define ALLOC_HARDER 0x10 /* try to alloc harder */
779 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
780 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
783 * Return 1 if free pages are above 'mark'. This takes into account the order
786 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
787 int classzone_idx, int alloc_flags)
789 /* free_pages my go negative - that's OK */
790 long min = mark, free_pages = z->free_pages - (1 << order) + 1;
793 if (alloc_flags & ALLOC_HIGH)
795 if (alloc_flags & ALLOC_HARDER)
798 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
800 for (o = 0; o < order; o++) {
801 /* At the next order, this order's pages become unavailable */
802 free_pages -= z->free_area[o].nr_free << o;
804 /* Require fewer higher order pages to be free */
807 if (free_pages <= min)
814 * get_page_from_freeliest goes through the zonelist trying to allocate
818 get_page_from_freelist(gfp_t gfp_mask, unsigned int order,
819 struct zonelist *zonelist, int alloc_flags)
821 struct zone **z = zonelist->zones;
822 struct page *page = NULL;
823 int classzone_idx = zone_idx(*z);
826 * Go through the zonelist once, looking for a zone with enough free.
827 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
830 if ((alloc_flags & ALLOC_CPUSET) &&
831 !cpuset_zone_allowed(*z, gfp_mask))
834 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
836 if (alloc_flags & ALLOC_WMARK_MIN)
837 mark = (*z)->pages_min;
838 else if (alloc_flags & ALLOC_WMARK_LOW)
839 mark = (*z)->pages_low;
841 mark = (*z)->pages_high;
842 if (!zone_watermark_ok(*z, order, mark,
843 classzone_idx, alloc_flags))
847 page = buffered_rmqueue(*z, order, gfp_mask);
849 zone_statistics(zonelist, *z);
852 } while (*(++z) != NULL);
857 * This is the 'heart' of the zoned buddy allocator.
859 struct page * fastcall
860 __alloc_pages(gfp_t gfp_mask, unsigned int order,
861 struct zonelist *zonelist)
863 const gfp_t wait = gfp_mask & __GFP_WAIT;
866 struct reclaim_state reclaim_state;
867 struct task_struct *p = current;
870 int did_some_progress;
872 might_sleep_if(wait);
875 z = zonelist->zones; /* the list of zones suitable for gfp_mask */
877 if (unlikely(*z == NULL)) {
878 /* Should this ever happen?? */
882 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
883 zonelist, ALLOC_WMARK_LOW|ALLOC_CPUSET);
888 wakeup_kswapd(*z, order);
892 * OK, we're below the kswapd watermark and have kicked background
893 * reclaim. Now things get more complex, so set up alloc_flags according
894 * to how we want to proceed.
896 * The caller may dip into page reserves a bit more if the caller
897 * cannot run direct reclaim, or if the caller has realtime scheduling
900 alloc_flags = ALLOC_WMARK_MIN;
901 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
902 alloc_flags |= ALLOC_HARDER;
903 if (gfp_mask & __GFP_HIGH)
904 alloc_flags |= ALLOC_HIGH;
905 alloc_flags |= ALLOC_CPUSET;
908 * Go through the zonelist again. Let __GFP_HIGH and allocations
909 * coming from realtime tasks go deeper into reserves.
911 * This is the last chance, in general, before the goto nopage.
912 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
913 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
915 page = get_page_from_freelist(gfp_mask, order, zonelist, alloc_flags);
919 /* This allocation should allow future memory freeing. */
921 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
922 && !in_interrupt()) {
923 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
925 /* go through the zonelist yet again, ignoring mins */
926 page = get_page_from_freelist(gfp_mask, order,
927 zonelist, ALLOC_NO_WATERMARKS);
930 if (gfp_mask & __GFP_NOFAIL) {
931 blk_congestion_wait(WRITE, HZ/50);
938 /* Atomic allocations - we can't balance anything */
945 /* We now go into synchronous reclaim */
946 p->flags |= PF_MEMALLOC;
947 reclaim_state.reclaimed_slab = 0;
948 p->reclaim_state = &reclaim_state;
950 did_some_progress = try_to_free_pages(zonelist->zones, gfp_mask);
952 p->reclaim_state = NULL;
953 p->flags &= ~PF_MEMALLOC;
957 if (likely(did_some_progress)) {
958 page = get_page_from_freelist(gfp_mask, order,
959 zonelist, alloc_flags);
962 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
964 * Go through the zonelist yet one more time, keep
965 * very high watermark here, this is only to catch
966 * a parallel oom killing, we must fail if we're still
967 * under heavy pressure.
969 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
970 zonelist, ALLOC_WMARK_HIGH|ALLOC_CPUSET);
974 out_of_memory(gfp_mask, order);
979 * Don't let big-order allocations loop unless the caller explicitly
980 * requests that. Wait for some write requests to complete then retry.
982 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
983 * <= 3, but that may not be true in other implementations.
986 if (!(gfp_mask & __GFP_NORETRY)) {
987 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
989 if (gfp_mask & __GFP_NOFAIL)
993 blk_congestion_wait(WRITE, HZ/50);
998 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
999 printk(KERN_WARNING "%s: page allocation failure."
1000 " order:%d, mode:0x%x\n",
1001 p->comm, order, gfp_mask);
1009 EXPORT_SYMBOL(__alloc_pages);
1012 * Common helper functions.
1014 fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1017 page = alloc_pages(gfp_mask, order);
1020 return (unsigned long) page_address(page);
1023 EXPORT_SYMBOL(__get_free_pages);
1025 fastcall unsigned long get_zeroed_page(gfp_t gfp_mask)
1030 * get_zeroed_page() returns a 32-bit address, which cannot represent
1033 BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1035 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1037 return (unsigned long) page_address(page);
1041 EXPORT_SYMBOL(get_zeroed_page);
1043 void __pagevec_free(struct pagevec *pvec)
1045 int i = pagevec_count(pvec);
1048 free_hot_cold_page(pvec->pages[i], pvec->cold);
1051 fastcall void __free_pages(struct page *page, unsigned int order)
1053 if (put_page_testzero(page)) {
1055 free_hot_page(page);
1057 __free_pages_ok(page, order);
1061 EXPORT_SYMBOL(__free_pages);
1063 fastcall void free_pages(unsigned long addr, unsigned int order)
1066 BUG_ON(!virt_addr_valid((void *)addr));
1067 __free_pages(virt_to_page((void *)addr), order);
1071 EXPORT_SYMBOL(free_pages);
1074 * Total amount of free (allocatable) RAM:
1076 unsigned int nr_free_pages(void)
1078 unsigned int sum = 0;
1082 sum += zone->free_pages;
1087 EXPORT_SYMBOL(nr_free_pages);
1090 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
1092 unsigned int i, sum = 0;
1094 for (i = 0; i < MAX_NR_ZONES; i++)
1095 sum += pgdat->node_zones[i].free_pages;
1101 static unsigned int nr_free_zone_pages(int offset)
1103 /* Just pick one node, since fallback list is circular */
1104 pg_data_t *pgdat = NODE_DATA(numa_node_id());
1105 unsigned int sum = 0;
1107 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1108 struct zone **zonep = zonelist->zones;
1111 for (zone = *zonep++; zone; zone = *zonep++) {
1112 unsigned long size = zone->present_pages;
1113 unsigned long high = zone->pages_high;
1122 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1124 unsigned int nr_free_buffer_pages(void)
1126 return nr_free_zone_pages(gfp_zone(GFP_USER));
1130 * Amount of free RAM allocatable within all zones
1132 unsigned int nr_free_pagecache_pages(void)
1134 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER));
1137 #ifdef CONFIG_HIGHMEM
1138 unsigned int nr_free_highpages (void)
1141 unsigned int pages = 0;
1143 for_each_pgdat(pgdat)
1144 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1151 static void show_node(struct zone *zone)
1153 printk("Node %d ", zone->zone_pgdat->node_id);
1156 #define show_node(zone) do { } while (0)
1160 * Accumulate the page_state information across all CPUs.
1161 * The result is unavoidably approximate - it can change
1162 * during and after execution of this function.
1164 static DEFINE_PER_CPU(struct page_state, page_states) = {0};
1166 atomic_t nr_pagecache = ATOMIC_INIT(0);
1167 EXPORT_SYMBOL(nr_pagecache);
1169 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
1172 static void __get_page_state(struct page_state *ret, int nr, cpumask_t *cpumask)
1176 memset(ret, 0, sizeof(*ret));
1178 cpu = first_cpu(*cpumask);
1179 while (cpu < NR_CPUS) {
1180 unsigned long *in, *out, off;
1182 in = (unsigned long *)&per_cpu(page_states, cpu);
1184 cpu = next_cpu(cpu, *cpumask);
1187 prefetch(&per_cpu(page_states, cpu));
1189 out = (unsigned long *)ret;
1190 for (off = 0; off < nr; off++)
1195 void get_page_state_node(struct page_state *ret, int node)
1198 cpumask_t mask = node_to_cpumask(node);
1200 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
1201 nr /= sizeof(unsigned long);
1203 __get_page_state(ret, nr+1, &mask);
1206 void get_page_state(struct page_state *ret)
1209 cpumask_t mask = CPU_MASK_ALL;
1211 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
1212 nr /= sizeof(unsigned long);
1214 __get_page_state(ret, nr + 1, &mask);
1217 void get_full_page_state(struct page_state *ret)
1219 cpumask_t mask = CPU_MASK_ALL;
1221 __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long), &mask);
1224 unsigned long __read_page_state(unsigned long offset)
1226 unsigned long ret = 0;
1232 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
1233 ret += *((unsigned long *)in);
1238 void __mod_page_state(unsigned long offset, unsigned long delta)
1240 unsigned long flags;
1243 local_irq_save(flags);
1244 ptr = &__get_cpu_var(page_states);
1245 *(unsigned long*)(ptr + offset) += delta;
1246 local_irq_restore(flags);
1249 EXPORT_SYMBOL(__mod_page_state);
1251 void __get_zone_counts(unsigned long *active, unsigned long *inactive,
1252 unsigned long *free, struct pglist_data *pgdat)
1254 struct zone *zones = pgdat->node_zones;
1260 for (i = 0; i < MAX_NR_ZONES; i++) {
1261 *active += zones[i].nr_active;
1262 *inactive += zones[i].nr_inactive;
1263 *free += zones[i].free_pages;
1267 void get_zone_counts(unsigned long *active,
1268 unsigned long *inactive, unsigned long *free)
1270 struct pglist_data *pgdat;
1275 for_each_pgdat(pgdat) {
1276 unsigned long l, m, n;
1277 __get_zone_counts(&l, &m, &n, pgdat);
1284 void si_meminfo(struct sysinfo *val)
1286 val->totalram = totalram_pages;
1288 val->freeram = nr_free_pages();
1289 val->bufferram = nr_blockdev_pages();
1290 #ifdef CONFIG_HIGHMEM
1291 val->totalhigh = totalhigh_pages;
1292 val->freehigh = nr_free_highpages();
1297 val->mem_unit = PAGE_SIZE;
1300 EXPORT_SYMBOL(si_meminfo);
1303 void si_meminfo_node(struct sysinfo *val, int nid)
1305 pg_data_t *pgdat = NODE_DATA(nid);
1307 val->totalram = pgdat->node_present_pages;
1308 val->freeram = nr_free_pages_pgdat(pgdat);
1309 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1310 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1311 val->mem_unit = PAGE_SIZE;
1315 #define K(x) ((x) << (PAGE_SHIFT-10))
1318 * Show free area list (used inside shift_scroll-lock stuff)
1319 * We also calculate the percentage fragmentation. We do this by counting the
1320 * memory on each free list with the exception of the first item on the list.
1322 void show_free_areas(void)
1324 struct page_state ps;
1325 int cpu, temperature;
1326 unsigned long active;
1327 unsigned long inactive;
1331 for_each_zone(zone) {
1333 printk("%s per-cpu:", zone->name);
1335 if (!zone->present_pages) {
1341 for_each_online_cpu(cpu) {
1342 struct per_cpu_pageset *pageset;
1344 pageset = zone_pcp(zone, cpu);
1346 for (temperature = 0; temperature < 2; temperature++)
1347 printk("cpu %d %s: high %d, batch %d used:%d\n",
1349 temperature ? "cold" : "hot",
1350 pageset->pcp[temperature].high,
1351 pageset->pcp[temperature].batch,
1352 pageset->pcp[temperature].count);
1356 get_page_state(&ps);
1357 get_zone_counts(&active, &inactive, &free);
1359 printk("Free pages: %11ukB (%ukB HighMem)\n",
1361 K(nr_free_highpages()));
1363 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1364 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1373 ps.nr_page_table_pages);
1375 for_each_zone(zone) {
1387 " pages_scanned:%lu"
1388 " all_unreclaimable? %s"
1391 K(zone->free_pages),
1394 K(zone->pages_high),
1396 K(zone->nr_inactive),
1397 K(zone->present_pages),
1398 zone->pages_scanned,
1399 (zone->all_unreclaimable ? "yes" : "no")
1401 printk("lowmem_reserve[]:");
1402 for (i = 0; i < MAX_NR_ZONES; i++)
1403 printk(" %lu", zone->lowmem_reserve[i]);
1407 for_each_zone(zone) {
1408 unsigned long nr, flags, order, total = 0;
1411 printk("%s: ", zone->name);
1412 if (!zone->present_pages) {
1417 spin_lock_irqsave(&zone->lock, flags);
1418 for (order = 0; order < MAX_ORDER; order++) {
1419 nr = zone->free_area[order].nr_free;
1420 total += nr << order;
1421 printk("%lu*%lukB ", nr, K(1UL) << order);
1423 spin_unlock_irqrestore(&zone->lock, flags);
1424 printk("= %lukB\n", K(total));
1427 show_swap_cache_info();
1431 * Builds allocation fallback zone lists.
1433 static int __init build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, int j, int k)
1440 zone = pgdat->node_zones + ZONE_HIGHMEM;
1441 if (zone->present_pages) {
1442 #ifndef CONFIG_HIGHMEM
1445 zonelist->zones[j++] = zone;
1448 zone = pgdat->node_zones + ZONE_NORMAL;
1449 if (zone->present_pages)
1450 zonelist->zones[j++] = zone;
1452 zone = pgdat->node_zones + ZONE_DMA32;
1453 if (zone->present_pages)
1454 zonelist->zones[j++] = zone;
1456 zone = pgdat->node_zones + ZONE_DMA;
1457 if (zone->present_pages)
1458 zonelist->zones[j++] = zone;
1464 static inline int highest_zone(int zone_bits)
1466 int res = ZONE_NORMAL;
1467 if (zone_bits & (__force int)__GFP_HIGHMEM)
1469 if (zone_bits & (__force int)__GFP_DMA32)
1471 if (zone_bits & (__force int)__GFP_DMA)
1477 #define MAX_NODE_LOAD (num_online_nodes())
1478 static int __initdata node_load[MAX_NUMNODES];
1480 * find_next_best_node - find the next node that should appear in a given node's fallback list
1481 * @node: node whose fallback list we're appending
1482 * @used_node_mask: nodemask_t of already used nodes
1484 * We use a number of factors to determine which is the next node that should
1485 * appear on a given node's fallback list. The node should not have appeared
1486 * already in @node's fallback list, and it should be the next closest node
1487 * according to the distance array (which contains arbitrary distance values
1488 * from each node to each node in the system), and should also prefer nodes
1489 * with no CPUs, since presumably they'll have very little allocation pressure
1490 * on them otherwise.
1491 * It returns -1 if no node is found.
1493 static int __init find_next_best_node(int node, nodemask_t *used_node_mask)
1496 int min_val = INT_MAX;
1499 for_each_online_node(i) {
1502 /* Start from local node */
1503 n = (node+i) % num_online_nodes();
1505 /* Don't want a node to appear more than once */
1506 if (node_isset(n, *used_node_mask))
1509 /* Use the local node if we haven't already */
1510 if (!node_isset(node, *used_node_mask)) {
1515 /* Use the distance array to find the distance */
1516 val = node_distance(node, n);
1518 /* Give preference to headless and unused nodes */
1519 tmp = node_to_cpumask(n);
1520 if (!cpus_empty(tmp))
1521 val += PENALTY_FOR_NODE_WITH_CPUS;
1523 /* Slight preference for less loaded node */
1524 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1525 val += node_load[n];
1527 if (val < min_val) {
1534 node_set(best_node, *used_node_mask);
1539 static void __init build_zonelists(pg_data_t *pgdat)
1541 int i, j, k, node, local_node;
1542 int prev_node, load;
1543 struct zonelist *zonelist;
1544 nodemask_t used_mask;
1546 /* initialize zonelists */
1547 for (i = 0; i < GFP_ZONETYPES; i++) {
1548 zonelist = pgdat->node_zonelists + i;
1549 zonelist->zones[0] = NULL;
1552 /* NUMA-aware ordering of nodes */
1553 local_node = pgdat->node_id;
1554 load = num_online_nodes();
1555 prev_node = local_node;
1556 nodes_clear(used_mask);
1557 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1559 * We don't want to pressure a particular node.
1560 * So adding penalty to the first node in same
1561 * distance group to make it round-robin.
1563 if (node_distance(local_node, node) !=
1564 node_distance(local_node, prev_node))
1565 node_load[node] += load;
1568 for (i = 0; i < GFP_ZONETYPES; i++) {
1569 zonelist = pgdat->node_zonelists + i;
1570 for (j = 0; zonelist->zones[j] != NULL; j++);
1572 k = highest_zone(i);
1574 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1575 zonelist->zones[j] = NULL;
1580 #else /* CONFIG_NUMA */
1582 static void __init build_zonelists(pg_data_t *pgdat)
1584 int i, j, k, node, local_node;
1586 local_node = pgdat->node_id;
1587 for (i = 0; i < GFP_ZONETYPES; i++) {
1588 struct zonelist *zonelist;
1590 zonelist = pgdat->node_zonelists + i;
1593 k = highest_zone(i);
1594 j = build_zonelists_node(pgdat, zonelist, j, k);
1596 * Now we build the zonelist so that it contains the zones
1597 * of all the other nodes.
1598 * We don't want to pressure a particular node, so when
1599 * building the zones for node N, we make sure that the
1600 * zones coming right after the local ones are those from
1601 * node N+1 (modulo N)
1603 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1604 if (!node_online(node))
1606 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1608 for (node = 0; node < local_node; node++) {
1609 if (!node_online(node))
1611 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1614 zonelist->zones[j] = NULL;
1618 #endif /* CONFIG_NUMA */
1620 void __init build_all_zonelists(void)
1624 for_each_online_node(i)
1625 build_zonelists(NODE_DATA(i));
1626 printk("Built %i zonelists\n", num_online_nodes());
1627 cpuset_init_current_mems_allowed();
1631 * Helper functions to size the waitqueue hash table.
1632 * Essentially these want to choose hash table sizes sufficiently
1633 * large so that collisions trying to wait on pages are rare.
1634 * But in fact, the number of active page waitqueues on typical
1635 * systems is ridiculously low, less than 200. So this is even
1636 * conservative, even though it seems large.
1638 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1639 * waitqueues, i.e. the size of the waitq table given the number of pages.
1641 #define PAGES_PER_WAITQUEUE 256
1643 static inline unsigned long wait_table_size(unsigned long pages)
1645 unsigned long size = 1;
1647 pages /= PAGES_PER_WAITQUEUE;
1649 while (size < pages)
1653 * Once we have dozens or even hundreds of threads sleeping
1654 * on IO we've got bigger problems than wait queue collision.
1655 * Limit the size of the wait table to a reasonable size.
1657 size = min(size, 4096UL);
1659 return max(size, 4UL);
1663 * This is an integer logarithm so that shifts can be used later
1664 * to extract the more random high bits from the multiplicative
1665 * hash function before the remainder is taken.
1667 static inline unsigned long wait_table_bits(unsigned long size)
1672 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1674 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1675 unsigned long *zones_size, unsigned long *zholes_size)
1677 unsigned long realtotalpages, totalpages = 0;
1680 for (i = 0; i < MAX_NR_ZONES; i++)
1681 totalpages += zones_size[i];
1682 pgdat->node_spanned_pages = totalpages;
1684 realtotalpages = totalpages;
1686 for (i = 0; i < MAX_NR_ZONES; i++)
1687 realtotalpages -= zholes_size[i];
1688 pgdat->node_present_pages = realtotalpages;
1689 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1694 * Initially all pages are reserved - free ones are freed
1695 * up by free_all_bootmem() once the early boot process is
1696 * done. Non-atomic initialization, single-pass.
1698 void __devinit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1699 unsigned long start_pfn)
1702 unsigned long end_pfn = start_pfn + size;
1705 for (pfn = start_pfn; pfn < end_pfn; pfn++, page++) {
1706 if (!early_pfn_valid(pfn))
1708 page = pfn_to_page(pfn);
1709 set_page_links(page, zone, nid, pfn);
1710 set_page_count(page, 1);
1711 reset_page_mapcount(page);
1712 SetPageReserved(page);
1713 INIT_LIST_HEAD(&page->lru);
1714 #ifdef WANT_PAGE_VIRTUAL
1715 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1716 if (!is_highmem_idx(zone))
1717 set_page_address(page, __va(pfn << PAGE_SHIFT));
1722 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1726 for (order = 0; order < MAX_ORDER ; order++) {
1727 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1728 zone->free_area[order].nr_free = 0;
1732 #define ZONETABLE_INDEX(x, zone_nr) ((x << ZONES_SHIFT) | zone_nr)
1733 void zonetable_add(struct zone *zone, int nid, int zid, unsigned long pfn,
1736 unsigned long snum = pfn_to_section_nr(pfn);
1737 unsigned long end = pfn_to_section_nr(pfn + size);
1740 zone_table[ZONETABLE_INDEX(nid, zid)] = zone;
1742 for (; snum <= end; snum++)
1743 zone_table[ZONETABLE_INDEX(snum, zid)] = zone;
1746 #ifndef __HAVE_ARCH_MEMMAP_INIT
1747 #define memmap_init(size, nid, zone, start_pfn) \
1748 memmap_init_zone((size), (nid), (zone), (start_pfn))
1751 static int __devinit zone_batchsize(struct zone *zone)
1756 * The per-cpu-pages pools are set to around 1000th of the
1757 * size of the zone. But no more than 1/2 of a meg.
1759 * OK, so we don't know how big the cache is. So guess.
1761 batch = zone->present_pages / 1024;
1762 if (batch * PAGE_SIZE > 512 * 1024)
1763 batch = (512 * 1024) / PAGE_SIZE;
1764 batch /= 4; /* We effectively *= 4 below */
1769 * Clamp the batch to a 2^n - 1 value. Having a power
1770 * of 2 value was found to be more likely to have
1771 * suboptimal cache aliasing properties in some cases.
1773 * For example if 2 tasks are alternately allocating
1774 * batches of pages, one task can end up with a lot
1775 * of pages of one half of the possible page colors
1776 * and the other with pages of the other colors.
1778 batch = (1 << (fls(batch + batch/2)-1)) - 1;
1783 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
1785 struct per_cpu_pages *pcp;
1787 memset(p, 0, sizeof(*p));
1789 pcp = &p->pcp[0]; /* hot */
1791 pcp->high = 6 * batch;
1792 pcp->batch = max(1UL, 1 * batch);
1793 INIT_LIST_HEAD(&pcp->list);
1795 pcp = &p->pcp[1]; /* cold*/
1797 pcp->high = 2 * batch;
1798 pcp->batch = max(1UL, batch/2);
1799 INIT_LIST_HEAD(&pcp->list);
1804 * Boot pageset table. One per cpu which is going to be used for all
1805 * zones and all nodes. The parameters will be set in such a way
1806 * that an item put on a list will immediately be handed over to
1807 * the buddy list. This is safe since pageset manipulation is done
1808 * with interrupts disabled.
1810 * Some NUMA counter updates may also be caught by the boot pagesets.
1812 * The boot_pagesets must be kept even after bootup is complete for
1813 * unused processors and/or zones. They do play a role for bootstrapping
1814 * hotplugged processors.
1816 * zoneinfo_show() and maybe other functions do
1817 * not check if the processor is online before following the pageset pointer.
1818 * Other parts of the kernel may not check if the zone is available.
1820 static struct per_cpu_pageset
1821 boot_pageset[NR_CPUS];
1824 * Dynamically allocate memory for the
1825 * per cpu pageset array in struct zone.
1827 static int __devinit process_zones(int cpu)
1829 struct zone *zone, *dzone;
1831 for_each_zone(zone) {
1833 zone->pageset[cpu] = kmalloc_node(sizeof(struct per_cpu_pageset),
1834 GFP_KERNEL, cpu_to_node(cpu));
1835 if (!zone->pageset[cpu])
1838 setup_pageset(zone->pageset[cpu], zone_batchsize(zone));
1843 for_each_zone(dzone) {
1846 kfree(dzone->pageset[cpu]);
1847 dzone->pageset[cpu] = NULL;
1852 static inline void free_zone_pagesets(int cpu)
1857 for_each_zone(zone) {
1858 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
1860 zone_pcp(zone, cpu) = NULL;
1866 static int __devinit pageset_cpuup_callback(struct notifier_block *nfb,
1867 unsigned long action,
1870 int cpu = (long)hcpu;
1871 int ret = NOTIFY_OK;
1874 case CPU_UP_PREPARE:
1875 if (process_zones(cpu))
1878 case CPU_UP_CANCELED:
1880 free_zone_pagesets(cpu);
1888 static struct notifier_block pageset_notifier =
1889 { &pageset_cpuup_callback, NULL, 0 };
1891 void __init setup_per_cpu_pageset(void)
1895 /* Initialize per_cpu_pageset for cpu 0.
1896 * A cpuup callback will do this for every cpu
1897 * as it comes online
1899 err = process_zones(smp_processor_id());
1901 register_cpu_notifier(&pageset_notifier);
1907 void zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
1910 struct pglist_data *pgdat = zone->zone_pgdat;
1913 * The per-page waitqueue mechanism uses hashed waitqueues
1916 zone->wait_table_size = wait_table_size(zone_size_pages);
1917 zone->wait_table_bits = wait_table_bits(zone->wait_table_size);
1918 zone->wait_table = (wait_queue_head_t *)
1919 alloc_bootmem_node(pgdat, zone->wait_table_size
1920 * sizeof(wait_queue_head_t));
1922 for(i = 0; i < zone->wait_table_size; ++i)
1923 init_waitqueue_head(zone->wait_table + i);
1926 static __devinit void zone_pcp_init(struct zone *zone)
1929 unsigned long batch = zone_batchsize(zone);
1931 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1933 /* Early boot. Slab allocator not functional yet */
1934 zone->pageset[cpu] = &boot_pageset[cpu];
1935 setup_pageset(&boot_pageset[cpu],0);
1937 setup_pageset(zone_pcp(zone,cpu), batch);
1940 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
1941 zone->name, zone->present_pages, batch);
1944 static __devinit void init_currently_empty_zone(struct zone *zone,
1945 unsigned long zone_start_pfn, unsigned long size)
1947 struct pglist_data *pgdat = zone->zone_pgdat;
1949 zone_wait_table_init(zone, size);
1950 pgdat->nr_zones = zone_idx(zone) + 1;
1952 zone->zone_mem_map = pfn_to_page(zone_start_pfn);
1953 zone->zone_start_pfn = zone_start_pfn;
1955 memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn);
1957 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1961 * Set up the zone data structures:
1962 * - mark all pages reserved
1963 * - mark all memory queues empty
1964 * - clear the memory bitmaps
1966 static void __init free_area_init_core(struct pglist_data *pgdat,
1967 unsigned long *zones_size, unsigned long *zholes_size)
1970 int nid = pgdat->node_id;
1971 unsigned long zone_start_pfn = pgdat->node_start_pfn;
1973 pgdat_resize_init(pgdat);
1974 pgdat->nr_zones = 0;
1975 init_waitqueue_head(&pgdat->kswapd_wait);
1976 pgdat->kswapd_max_order = 0;
1978 for (j = 0; j < MAX_NR_ZONES; j++) {
1979 struct zone *zone = pgdat->node_zones + j;
1980 unsigned long size, realsize;
1982 realsize = size = zones_size[j];
1984 realsize -= zholes_size[j];
1986 if (j < ZONE_HIGHMEM)
1987 nr_kernel_pages += realsize;
1988 nr_all_pages += realsize;
1990 zone->spanned_pages = size;
1991 zone->present_pages = realsize;
1992 zone->name = zone_names[j];
1993 spin_lock_init(&zone->lock);
1994 spin_lock_init(&zone->lru_lock);
1995 zone_seqlock_init(zone);
1996 zone->zone_pgdat = pgdat;
1997 zone->free_pages = 0;
1999 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
2001 zone_pcp_init(zone);
2002 INIT_LIST_HEAD(&zone->active_list);
2003 INIT_LIST_HEAD(&zone->inactive_list);
2004 zone->nr_scan_active = 0;
2005 zone->nr_scan_inactive = 0;
2006 zone->nr_active = 0;
2007 zone->nr_inactive = 0;
2008 atomic_set(&zone->reclaim_in_progress, 0);
2012 zonetable_add(zone, nid, j, zone_start_pfn, size);
2013 init_currently_empty_zone(zone, zone_start_pfn, size);
2014 zone_start_pfn += size;
2018 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
2020 /* Skip empty nodes */
2021 if (!pgdat->node_spanned_pages)
2024 #ifdef CONFIG_FLAT_NODE_MEM_MAP
2025 /* ia64 gets its own node_mem_map, before this, without bootmem */
2026 if (!pgdat->node_mem_map) {
2030 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
2031 map = alloc_remap(pgdat->node_id, size);
2033 map = alloc_bootmem_node(pgdat, size);
2034 pgdat->node_mem_map = map;
2036 #ifdef CONFIG_FLATMEM
2038 * With no DISCONTIG, the global mem_map is just set as node 0's
2040 if (pgdat == NODE_DATA(0))
2041 mem_map = NODE_DATA(0)->node_mem_map;
2043 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
2046 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
2047 unsigned long *zones_size, unsigned long node_start_pfn,
2048 unsigned long *zholes_size)
2050 pgdat->node_id = nid;
2051 pgdat->node_start_pfn = node_start_pfn;
2052 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
2054 alloc_node_mem_map(pgdat);
2056 free_area_init_core(pgdat, zones_size, zholes_size);
2059 #ifndef CONFIG_NEED_MULTIPLE_NODES
2060 static bootmem_data_t contig_bootmem_data;
2061 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
2063 EXPORT_SYMBOL(contig_page_data);
2066 void __init free_area_init(unsigned long *zones_size)
2068 free_area_init_node(0, NODE_DATA(0), zones_size,
2069 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
2072 #ifdef CONFIG_PROC_FS
2074 #include <linux/seq_file.h>
2076 static void *frag_start(struct seq_file *m, loff_t *pos)
2081 for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
2087 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
2089 pg_data_t *pgdat = (pg_data_t *)arg;
2092 return pgdat->pgdat_next;
2095 static void frag_stop(struct seq_file *m, void *arg)
2100 * This walks the free areas for each zone.
2102 static int frag_show(struct seq_file *m, void *arg)
2104 pg_data_t *pgdat = (pg_data_t *)arg;
2106 struct zone *node_zones = pgdat->node_zones;
2107 unsigned long flags;
2110 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2111 if (!zone->present_pages)
2114 spin_lock_irqsave(&zone->lock, flags);
2115 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
2116 for (order = 0; order < MAX_ORDER; ++order)
2117 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
2118 spin_unlock_irqrestore(&zone->lock, flags);
2124 struct seq_operations fragmentation_op = {
2125 .start = frag_start,
2132 * Output information about zones in @pgdat.
2134 static int zoneinfo_show(struct seq_file *m, void *arg)
2136 pg_data_t *pgdat = arg;
2138 struct zone *node_zones = pgdat->node_zones;
2139 unsigned long flags;
2141 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; zone++) {
2144 if (!zone->present_pages)
2147 spin_lock_irqsave(&zone->lock, flags);
2148 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
2156 "\n scanned %lu (a: %lu i: %lu)"
2165 zone->pages_scanned,
2166 zone->nr_scan_active, zone->nr_scan_inactive,
2167 zone->spanned_pages,
2168 zone->present_pages);
2170 "\n protection: (%lu",
2171 zone->lowmem_reserve[0]);
2172 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
2173 seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
2177 for (i = 0; i < ARRAY_SIZE(zone->pageset); i++) {
2178 struct per_cpu_pageset *pageset;
2181 pageset = zone_pcp(zone, i);
2182 for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2183 if (pageset->pcp[j].count)
2186 if (j == ARRAY_SIZE(pageset->pcp))
2188 for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2190 "\n cpu: %i pcp: %i"
2195 pageset->pcp[j].count,
2196 pageset->pcp[j].high,
2197 pageset->pcp[j].batch);
2203 "\n numa_foreign: %lu"
2204 "\n interleave_hit: %lu"
2205 "\n local_node: %lu"
2206 "\n other_node: %lu",
2209 pageset->numa_foreign,
2210 pageset->interleave_hit,
2211 pageset->local_node,
2212 pageset->other_node);
2216 "\n all_unreclaimable: %u"
2217 "\n prev_priority: %i"
2218 "\n temp_priority: %i"
2219 "\n start_pfn: %lu",
2220 zone->all_unreclaimable,
2221 zone->prev_priority,
2222 zone->temp_priority,
2223 zone->zone_start_pfn);
2224 spin_unlock_irqrestore(&zone->lock, flags);
2230 struct seq_operations zoneinfo_op = {
2231 .start = frag_start, /* iterate over all zones. The same as in
2235 .show = zoneinfo_show,
2238 static char *vmstat_text[] = {
2242 "nr_page_table_pages",
2267 "pgscan_kswapd_high",
2268 "pgscan_kswapd_normal",
2270 "pgscan_kswapd_dma",
2271 "pgscan_direct_high",
2272 "pgscan_direct_normal",
2273 "pgscan_direct_dma",
2278 "kswapd_inodesteal",
2286 static void *vmstat_start(struct seq_file *m, loff_t *pos)
2288 struct page_state *ps;
2290 if (*pos >= ARRAY_SIZE(vmstat_text))
2293 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
2296 return ERR_PTR(-ENOMEM);
2297 get_full_page_state(ps);
2298 ps->pgpgin /= 2; /* sectors -> kbytes */
2300 return (unsigned long *)ps + *pos;
2303 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
2306 if (*pos >= ARRAY_SIZE(vmstat_text))
2308 return (unsigned long *)m->private + *pos;
2311 static int vmstat_show(struct seq_file *m, void *arg)
2313 unsigned long *l = arg;
2314 unsigned long off = l - (unsigned long *)m->private;
2316 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
2320 static void vmstat_stop(struct seq_file *m, void *arg)
2326 struct seq_operations vmstat_op = {
2327 .start = vmstat_start,
2328 .next = vmstat_next,
2329 .stop = vmstat_stop,
2330 .show = vmstat_show,
2333 #endif /* CONFIG_PROC_FS */
2335 #ifdef CONFIG_HOTPLUG_CPU
2336 static int page_alloc_cpu_notify(struct notifier_block *self,
2337 unsigned long action, void *hcpu)
2339 int cpu = (unsigned long)hcpu;
2341 unsigned long *src, *dest;
2343 if (action == CPU_DEAD) {
2346 /* Drain local pagecache count. */
2347 count = &per_cpu(nr_pagecache_local, cpu);
2348 atomic_add(*count, &nr_pagecache);
2350 local_irq_disable();
2353 /* Add dead cpu's page_states to our own. */
2354 dest = (unsigned long *)&__get_cpu_var(page_states);
2355 src = (unsigned long *)&per_cpu(page_states, cpu);
2357 for (i = 0; i < sizeof(struct page_state)/sizeof(unsigned long);
2367 #endif /* CONFIG_HOTPLUG_CPU */
2369 void __init page_alloc_init(void)
2371 hotcpu_notifier(page_alloc_cpu_notify, 0);
2375 * setup_per_zone_lowmem_reserve - called whenever
2376 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
2377 * has a correct pages reserved value, so an adequate number of
2378 * pages are left in the zone after a successful __alloc_pages().
2380 static void setup_per_zone_lowmem_reserve(void)
2382 struct pglist_data *pgdat;
2385 for_each_pgdat(pgdat) {
2386 for (j = 0; j < MAX_NR_ZONES; j++) {
2387 struct zone *zone = pgdat->node_zones + j;
2388 unsigned long present_pages = zone->present_pages;
2390 zone->lowmem_reserve[j] = 0;
2392 for (idx = j-1; idx >= 0; idx--) {
2393 struct zone *lower_zone;
2395 if (sysctl_lowmem_reserve_ratio[idx] < 1)
2396 sysctl_lowmem_reserve_ratio[idx] = 1;
2398 lower_zone = pgdat->node_zones + idx;
2399 lower_zone->lowmem_reserve[j] = present_pages /
2400 sysctl_lowmem_reserve_ratio[idx];
2401 present_pages += lower_zone->present_pages;
2408 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
2409 * that the pages_{min,low,high} values for each zone are set correctly
2410 * with respect to min_free_kbytes.
2412 void setup_per_zone_pages_min(void)
2414 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
2415 unsigned long lowmem_pages = 0;
2417 unsigned long flags;
2419 /* Calculate total number of !ZONE_HIGHMEM pages */
2420 for_each_zone(zone) {
2421 if (!is_highmem(zone))
2422 lowmem_pages += zone->present_pages;
2425 for_each_zone(zone) {
2427 spin_lock_irqsave(&zone->lru_lock, flags);
2428 tmp = (pages_min * zone->present_pages) / lowmem_pages;
2429 if (is_highmem(zone)) {
2431 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
2432 * need highmem pages, so cap pages_min to a small
2435 * The (pages_high-pages_low) and (pages_low-pages_min)
2436 * deltas controls asynch page reclaim, and so should
2437 * not be capped for highmem.
2441 min_pages = zone->present_pages / 1024;
2442 if (min_pages < SWAP_CLUSTER_MAX)
2443 min_pages = SWAP_CLUSTER_MAX;
2444 if (min_pages > 128)
2446 zone->pages_min = min_pages;
2449 * If it's a lowmem zone, reserve a number of pages
2450 * proportionate to the zone's size.
2452 zone->pages_min = tmp;
2455 zone->pages_low = zone->pages_min + tmp / 4;
2456 zone->pages_high = zone->pages_min + tmp / 2;
2457 spin_unlock_irqrestore(&zone->lru_lock, flags);
2462 * Initialise min_free_kbytes.
2464 * For small machines we want it small (128k min). For large machines
2465 * we want it large (64MB max). But it is not linear, because network
2466 * bandwidth does not increase linearly with machine size. We use
2468 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2469 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2485 static int __init init_per_zone_pages_min(void)
2487 unsigned long lowmem_kbytes;
2489 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2491 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2492 if (min_free_kbytes < 128)
2493 min_free_kbytes = 128;
2494 if (min_free_kbytes > 65536)
2495 min_free_kbytes = 65536;
2496 setup_per_zone_pages_min();
2497 setup_per_zone_lowmem_reserve();
2500 module_init(init_per_zone_pages_min)
2503 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2504 * that we can call two helper functions whenever min_free_kbytes
2507 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2508 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2510 proc_dointvec(table, write, file, buffer, length, ppos);
2511 setup_per_zone_pages_min();
2516 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2517 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2518 * whenever sysctl_lowmem_reserve_ratio changes.
2520 * The reserve ratio obviously has absolutely no relation with the
2521 * pages_min watermarks. The lowmem reserve ratio can only make sense
2522 * if in function of the boot time zone sizes.
2524 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
2525 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2527 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2528 setup_per_zone_lowmem_reserve();
2532 __initdata int hashdist = HASHDIST_DEFAULT;
2535 static int __init set_hashdist(char *str)
2539 hashdist = simple_strtoul(str, &str, 0);
2542 __setup("hashdist=", set_hashdist);
2546 * allocate a large system hash table from bootmem
2547 * - it is assumed that the hash table must contain an exact power-of-2
2548 * quantity of entries
2549 * - limit is the number of hash buckets, not the total allocation size
2551 void *__init alloc_large_system_hash(const char *tablename,
2552 unsigned long bucketsize,
2553 unsigned long numentries,
2556 unsigned int *_hash_shift,
2557 unsigned int *_hash_mask,
2558 unsigned long limit)
2560 unsigned long long max = limit;
2561 unsigned long log2qty, size;
2564 /* allow the kernel cmdline to have a say */
2566 /* round applicable memory size up to nearest megabyte */
2567 numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages;
2568 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2569 numentries >>= 20 - PAGE_SHIFT;
2570 numentries <<= 20 - PAGE_SHIFT;
2572 /* limit to 1 bucket per 2^scale bytes of low memory */
2573 if (scale > PAGE_SHIFT)
2574 numentries >>= (scale - PAGE_SHIFT);
2576 numentries <<= (PAGE_SHIFT - scale);
2578 /* rounded up to nearest power of 2 in size */
2579 numentries = 1UL << (long_log2(numentries) + 1);
2581 /* limit allocation size to 1/16 total memory by default */
2583 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2584 do_div(max, bucketsize);
2587 if (numentries > max)
2590 log2qty = long_log2(numentries);
2593 size = bucketsize << log2qty;
2594 if (flags & HASH_EARLY)
2595 table = alloc_bootmem(size);
2597 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
2599 unsigned long order;
2600 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
2602 table = (void*) __get_free_pages(GFP_ATOMIC, order);
2604 } while (!table && size > PAGE_SIZE && --log2qty);
2607 panic("Failed to allocate %s hash table\n", tablename);
2609 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2612 long_log2(size) - PAGE_SHIFT,
2616 *_hash_shift = log2qty;
2618 *_hash_mask = (1 << log2qty) - 1;