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 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 256, 32 };
66 EXPORT_SYMBOL(totalram_pages);
69 * Used by page_zone() to look up the address of the struct zone whose
70 * id is encoded in the upper bits of page->flags
72 struct zone *zone_table[1 << ZONETABLE_SHIFT] __read_mostly;
73 EXPORT_SYMBOL(zone_table);
75 static char *zone_names[MAX_NR_ZONES] = { "DMA", "Normal", "HighMem" };
76 int min_free_kbytes = 1024;
78 unsigned long __initdata nr_kernel_pages;
79 unsigned long __initdata nr_all_pages;
81 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
85 unsigned long pfn = page_to_pfn(page);
88 seq = zone_span_seqbegin(zone);
89 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
91 else if (pfn < zone->zone_start_pfn)
93 } while (zone_span_seqretry(zone, seq));
98 static int page_is_consistent(struct zone *zone, struct page *page)
100 #ifdef CONFIG_HOLES_IN_ZONE
101 if (!pfn_valid(page_to_pfn(page)))
104 if (zone != page_zone(page))
110 * Temporary debugging check for pages not lying within a given zone.
112 static int bad_range(struct zone *zone, struct page *page)
114 if (page_outside_zone_boundaries(zone, page))
116 if (!page_is_consistent(zone, page))
122 static void bad_page(const char *function, struct page *page)
124 printk(KERN_EMERG "Bad page state at %s (in process '%s', page %p)\n",
125 function, current->comm, page);
126 printk(KERN_EMERG "flags:0x%0*lx mapping:%p mapcount:%d count:%d\n",
127 (int)(2*sizeof(page_flags_t)), (unsigned long)page->flags,
128 page->mapping, page_mapcount(page), page_count(page));
129 printk(KERN_EMERG "Backtrace:\n");
131 printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n");
132 page->flags &= ~(1 << PG_lru |
142 set_page_count(page, 0);
143 reset_page_mapcount(page);
144 page->mapping = NULL;
145 add_taint(TAINT_BAD_PAGE);
148 #ifndef CONFIG_HUGETLB_PAGE
149 #define prep_compound_page(page, order) do { } while (0)
150 #define destroy_compound_page(page, order) do { } while (0)
153 * Higher-order pages are called "compound pages". They are structured thusly:
155 * The first PAGE_SIZE page is called the "head page".
157 * The remaining PAGE_SIZE pages are called "tail pages".
159 * All pages have PG_compound set. All pages have their ->private pointing at
160 * the head page (even the head page has this).
162 * The first tail page's ->mapping, if non-zero, holds the address of the
163 * compound page's put_page() function.
165 * The order of the allocation is stored in the first tail page's ->index
166 * This is only for debug at present. This usage means that zero-order pages
167 * may not be compound.
169 static void prep_compound_page(struct page *page, unsigned long order)
172 int nr_pages = 1 << order;
174 page[1].mapping = NULL;
175 page[1].index = order;
176 for (i = 0; i < nr_pages; i++) {
177 struct page *p = page + i;
180 set_page_private(p, (unsigned long)page);
184 static void destroy_compound_page(struct page *page, unsigned long order)
187 int nr_pages = 1 << order;
189 if (!PageCompound(page))
192 if (page[1].index != order)
193 bad_page(__FUNCTION__, page);
195 for (i = 0; i < nr_pages; i++) {
196 struct page *p = page + i;
198 if (!PageCompound(p))
199 bad_page(__FUNCTION__, page);
200 if (page_private(p) != (unsigned long)page)
201 bad_page(__FUNCTION__, page);
202 ClearPageCompound(p);
205 #endif /* CONFIG_HUGETLB_PAGE */
208 * function for dealing with page's order in buddy system.
209 * zone->lock is already acquired when we use these.
210 * So, we don't need atomic page->flags operations here.
212 static inline unsigned long page_order(struct page *page) {
213 return page_private(page);
216 static inline void set_page_order(struct page *page, int order) {
217 set_page_private(page, order);
218 __SetPagePrivate(page);
221 static inline void rmv_page_order(struct page *page)
223 __ClearPagePrivate(page);
224 set_page_private(page, 0);
228 * Locate the struct page for both the matching buddy in our
229 * pair (buddy1) and the combined O(n+1) page they form (page).
231 * 1) Any buddy B1 will have an order O twin B2 which satisfies
232 * the following equation:
234 * For example, if the starting buddy (buddy2) is #8 its order
236 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
238 * 2) Any buddy B will have an order O+1 parent P which
239 * satisfies the following equation:
242 * Assumption: *_mem_map is contigious at least up to MAX_ORDER
244 static inline struct page *
245 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
247 unsigned long buddy_idx = page_idx ^ (1 << order);
249 return page + (buddy_idx - page_idx);
252 static inline unsigned long
253 __find_combined_index(unsigned long page_idx, unsigned int order)
255 return (page_idx & ~(1 << order));
259 * This function checks whether a page is free && is the buddy
260 * we can do coalesce a page and its buddy if
261 * (a) the buddy is free &&
262 * (b) the buddy is on the buddy system &&
263 * (c) a page and its buddy have the same order.
264 * for recording page's order, we use page_private(page) and PG_private.
267 static inline int page_is_buddy(struct page *page, int order)
269 if (PagePrivate(page) &&
270 (page_order(page) == order) &&
271 page_count(page) == 0)
277 * Freeing function for a buddy system allocator.
279 * The concept of a buddy system is to maintain direct-mapped table
280 * (containing bit values) for memory blocks of various "orders".
281 * The bottom level table contains the map for the smallest allocatable
282 * units of memory (here, pages), and each level above it describes
283 * pairs of units from the levels below, hence, "buddies".
284 * At a high level, all that happens here is marking the table entry
285 * at the bottom level available, and propagating the changes upward
286 * as necessary, plus some accounting needed to play nicely with other
287 * parts of the VM system.
288 * At each level, we keep a list of pages, which are heads of continuous
289 * free pages of length of (1 << order) and marked with PG_Private.Page's
290 * order is recorded in page_private(page) field.
291 * So when we are allocating or freeing one, we can derive the state of the
292 * other. That is, if we allocate a small block, and both were
293 * free, the remainder of the region must be split into blocks.
294 * If a block is freed, and its buddy is also free, then this
295 * triggers coalescing into a block of larger size.
300 static inline void __free_pages_bulk (struct page *page,
301 struct zone *zone, unsigned int order)
303 unsigned long page_idx;
304 int order_size = 1 << order;
307 destroy_compound_page(page, order);
309 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
311 BUG_ON(page_idx & (order_size - 1));
312 BUG_ON(bad_range(zone, page));
314 zone->free_pages += order_size;
315 while (order < MAX_ORDER-1) {
316 unsigned long combined_idx;
317 struct free_area *area;
320 combined_idx = __find_combined_index(page_idx, order);
321 buddy = __page_find_buddy(page, page_idx, order);
323 if (bad_range(zone, buddy))
325 if (!page_is_buddy(buddy, order))
326 break; /* Move the buddy up one level. */
327 list_del(&buddy->lru);
328 area = zone->free_area + order;
330 rmv_page_order(buddy);
331 page = page + (combined_idx - page_idx);
332 page_idx = combined_idx;
335 set_page_order(page, order);
336 list_add(&page->lru, &zone->free_area[order].free_list);
337 zone->free_area[order].nr_free++;
340 static inline void free_pages_check(const char *function, struct page *page)
342 if ( page_mapcount(page) ||
343 page->mapping != NULL ||
344 page_count(page) != 0 ||
355 bad_page(function, page);
357 __ClearPageDirty(page);
361 * Frees a list of pages.
362 * Assumes all pages on list are in same zone, and of same order.
363 * count is the number of pages to free.
365 * If the zone was previously in an "all pages pinned" state then look to
366 * see if this freeing clears that state.
368 * And clear the zone's pages_scanned counter, to hold off the "all pages are
369 * pinned" detection logic.
372 free_pages_bulk(struct zone *zone, int count,
373 struct list_head *list, unsigned int order)
376 struct page *page = NULL;
379 spin_lock_irqsave(&zone->lock, flags);
380 zone->all_unreclaimable = 0;
381 zone->pages_scanned = 0;
382 while (!list_empty(list) && count--) {
383 page = list_entry(list->prev, struct page, lru);
384 /* have to delete it as __free_pages_bulk list manipulates */
385 list_del(&page->lru);
386 __free_pages_bulk(page, zone, order);
389 spin_unlock_irqrestore(&zone->lock, flags);
393 void __free_pages_ok(struct page *page, unsigned int order)
398 arch_free_page(page, order);
400 mod_page_state(pgfree, 1 << order);
404 for (i = 1 ; i < (1 << order) ; ++i)
405 __put_page(page + i);
408 for (i = 0 ; i < (1 << order) ; ++i)
409 free_pages_check(__FUNCTION__, page + i);
410 list_add(&page->lru, &list);
411 kernel_map_pages(page, 1<<order, 0);
412 free_pages_bulk(page_zone(page), 1, &list, order);
417 * The order of subdivision here is critical for the IO subsystem.
418 * Please do not alter this order without good reasons and regression
419 * testing. Specifically, as large blocks of memory are subdivided,
420 * the order in which smaller blocks are delivered depends on the order
421 * they're subdivided in this function. This is the primary factor
422 * influencing the order in which pages are delivered to the IO
423 * subsystem according to empirical testing, and this is also justified
424 * by considering the behavior of a buddy system containing a single
425 * large block of memory acted on by a series of small allocations.
426 * This behavior is a critical factor in sglist merging's success.
430 static inline struct page *
431 expand(struct zone *zone, struct page *page,
432 int low, int high, struct free_area *area)
434 unsigned long size = 1 << high;
440 BUG_ON(bad_range(zone, &page[size]));
441 list_add(&page[size].lru, &area->free_list);
443 set_page_order(&page[size], high);
448 void set_page_refs(struct page *page, int order)
451 set_page_count(page, 1);
456 * We need to reference all the pages for this order, otherwise if
457 * anyone accesses one of the pages with (get/put) it will be freed.
458 * - eg: access_process_vm()
460 for (i = 0; i < (1 << order); i++)
461 set_page_count(page + i, 1);
462 #endif /* CONFIG_MMU */
466 * This page is about to be returned from the page allocator
468 static void prep_new_page(struct page *page, int order)
470 if ( page_mapcount(page) ||
471 page->mapping != NULL ||
472 page_count(page) != 0 ||
484 bad_page(__FUNCTION__, page);
486 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
487 1 << PG_referenced | 1 << PG_arch_1 |
488 1 << PG_checked | 1 << PG_mappedtodisk);
489 set_page_private(page, 0);
490 set_page_refs(page, order);
491 kernel_map_pages(page, 1 << order, 1);
495 * Do the hard work of removing an element from the buddy allocator.
496 * Call me with the zone->lock already held.
498 static struct page *__rmqueue(struct zone *zone, unsigned int order)
500 struct free_area * area;
501 unsigned int current_order;
504 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
505 area = zone->free_area + current_order;
506 if (list_empty(&area->free_list))
509 page = list_entry(area->free_list.next, struct page, lru);
510 list_del(&page->lru);
511 rmv_page_order(page);
513 zone->free_pages -= 1UL << order;
514 return expand(zone, page, order, current_order, area);
521 * Obtain a specified number of elements from the buddy allocator, all under
522 * a single hold of the lock, for efficiency. Add them to the supplied list.
523 * Returns the number of new pages which were placed at *list.
525 static int rmqueue_bulk(struct zone *zone, unsigned int order,
526 unsigned long count, struct list_head *list)
533 spin_lock_irqsave(&zone->lock, flags);
534 for (i = 0; i < count; ++i) {
535 page = __rmqueue(zone, order);
539 list_add_tail(&page->lru, list);
541 spin_unlock_irqrestore(&zone->lock, flags);
546 /* Called from the slab reaper to drain remote pagesets */
547 void drain_remote_pages(void)
553 local_irq_save(flags);
554 for_each_zone(zone) {
555 struct per_cpu_pageset *pset;
557 /* Do not drain local pagesets */
558 if (zone->zone_pgdat->node_id == numa_node_id())
561 pset = zone->pageset[smp_processor_id()];
562 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
563 struct per_cpu_pages *pcp;
567 pcp->count -= free_pages_bulk(zone, pcp->count,
571 local_irq_restore(flags);
575 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
576 static void __drain_pages(unsigned int cpu)
581 for_each_zone(zone) {
582 struct per_cpu_pageset *pset;
584 pset = zone_pcp(zone, cpu);
585 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
586 struct per_cpu_pages *pcp;
589 pcp->count -= free_pages_bulk(zone, pcp->count,
594 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
598 void mark_free_pages(struct zone *zone)
600 unsigned long zone_pfn, flags;
602 struct list_head *curr;
604 if (!zone->spanned_pages)
607 spin_lock_irqsave(&zone->lock, flags);
608 for (zone_pfn = 0; zone_pfn < zone->spanned_pages; ++zone_pfn)
609 ClearPageNosaveFree(pfn_to_page(zone_pfn + zone->zone_start_pfn));
611 for (order = MAX_ORDER - 1; order >= 0; --order)
612 list_for_each(curr, &zone->free_area[order].free_list) {
613 unsigned long start_pfn, i;
615 start_pfn = page_to_pfn(list_entry(curr, struct page, lru));
617 for (i=0; i < (1<<order); i++)
618 SetPageNosaveFree(pfn_to_page(start_pfn+i));
620 spin_unlock_irqrestore(&zone->lock, flags);
624 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
626 void drain_local_pages(void)
630 local_irq_save(flags);
631 __drain_pages(smp_processor_id());
632 local_irq_restore(flags);
634 #endif /* CONFIG_PM */
636 static void zone_statistics(struct zonelist *zonelist, struct zone *z)
641 pg_data_t *pg = z->zone_pgdat;
642 pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
643 struct per_cpu_pageset *p;
645 local_irq_save(flags);
646 cpu = smp_processor_id();
652 zone_pcp(zonelist->zones[0], cpu)->numa_foreign++;
654 if (pg == NODE_DATA(numa_node_id()))
658 local_irq_restore(flags);
663 * Free a 0-order page
665 static void FASTCALL(free_hot_cold_page(struct page *page, int cold));
666 static void fastcall free_hot_cold_page(struct page *page, int cold)
668 struct zone *zone = page_zone(page);
669 struct per_cpu_pages *pcp;
672 arch_free_page(page, 0);
674 kernel_map_pages(page, 1, 0);
675 inc_page_state(pgfree);
677 page->mapping = NULL;
678 free_pages_check(__FUNCTION__, page);
679 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
680 local_irq_save(flags);
681 list_add(&page->lru, &pcp->list);
683 if (pcp->count >= pcp->high)
684 pcp->count -= free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
685 local_irq_restore(flags);
689 void fastcall free_hot_page(struct page *page)
691 free_hot_cold_page(page, 0);
694 void fastcall free_cold_page(struct page *page)
696 free_hot_cold_page(page, 1);
699 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
703 BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
704 for(i = 0; i < (1 << order); i++)
705 clear_highpage(page + i);
709 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
710 * we cheat by calling it from here, in the order > 0 path. Saves a branch
714 buffered_rmqueue(struct zone *zone, int order, gfp_t gfp_flags)
717 struct page *page = NULL;
718 int cold = !!(gfp_flags & __GFP_COLD);
721 struct per_cpu_pages *pcp;
723 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
724 local_irq_save(flags);
725 if (pcp->count <= pcp->low)
726 pcp->count += rmqueue_bulk(zone, 0,
727 pcp->batch, &pcp->list);
729 page = list_entry(pcp->list.next, struct page, lru);
730 list_del(&page->lru);
733 local_irq_restore(flags);
736 spin_lock_irqsave(&zone->lock, flags);
737 page = __rmqueue(zone, order);
738 spin_unlock_irqrestore(&zone->lock, flags);
742 BUG_ON(bad_range(zone, page));
743 mod_page_state_zone(zone, pgalloc, 1 << order);
744 prep_new_page(page, order);
746 if (gfp_flags & __GFP_ZERO)
747 prep_zero_page(page, order, gfp_flags);
749 if (order && (gfp_flags & __GFP_COMP))
750 prep_compound_page(page, order);
755 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
756 #define ALLOC_HARDER 0x02 /* try to alloc harder */
757 #define ALLOC_HIGH 0x04 /* __GFP_HIGH set */
758 #define ALLOC_CPUSET 0x08 /* check for correct cpuset */
761 * Return 1 if free pages are above 'mark'. This takes into account the order
764 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
765 int classzone_idx, int alloc_flags)
767 /* free_pages my go negative - that's OK */
768 long min = mark, free_pages = z->free_pages - (1 << order) + 1;
771 if (alloc_flags & ALLOC_HIGH)
773 if (alloc_flags & ALLOC_HARDER)
776 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
778 for (o = 0; o < order; o++) {
779 /* At the next order, this order's pages become unavailable */
780 free_pages -= z->free_area[o].nr_free << o;
782 /* Require fewer higher order pages to be free */
785 if (free_pages <= min)
792 * get_page_from_freeliest goes through the zonelist trying to allocate
796 get_page_from_freelist(gfp_t gfp_mask, unsigned int order,
797 struct zonelist *zonelist, int alloc_flags)
799 struct zone **z = zonelist->zones;
800 struct page *page = NULL;
801 int classzone_idx = zone_idx(*z);
804 * Go through the zonelist once, looking for a zone with enough free.
805 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
808 if ((alloc_flags & ALLOC_CPUSET) &&
809 !cpuset_zone_allowed(*z, gfp_mask))
812 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
813 if (!zone_watermark_ok(*z, order, (*z)->pages_low,
814 classzone_idx, alloc_flags))
818 page = buffered_rmqueue(*z, order, gfp_mask);
820 zone_statistics(zonelist, *z);
823 } while (*(++z) != NULL);
828 * This is the 'heart' of the zoned buddy allocator.
830 struct page * fastcall
831 __alloc_pages(gfp_t gfp_mask, unsigned int order,
832 struct zonelist *zonelist)
834 const gfp_t wait = gfp_mask & __GFP_WAIT;
837 struct reclaim_state reclaim_state;
838 struct task_struct *p = current;
841 int did_some_progress;
843 might_sleep_if(wait);
845 z = zonelist->zones; /* the list of zones suitable for gfp_mask */
847 if (unlikely(*z == NULL)) {
848 /* Should this ever happen?? */
852 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
853 zonelist, ALLOC_CPUSET);
858 wakeup_kswapd(*z, order);
862 * OK, we're below the kswapd watermark and have kicked background
863 * reclaim. Now things get more complex, so set up alloc_flags according
864 * to how we want to proceed.
866 * The caller may dip into page reserves a bit more if the caller
867 * cannot run direct reclaim, or if the caller has realtime scheduling
871 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
872 alloc_flags |= ALLOC_HARDER;
873 if (gfp_mask & __GFP_HIGH)
874 alloc_flags |= ALLOC_HIGH;
876 alloc_flags |= ALLOC_CPUSET;
879 * Go through the zonelist again. Let __GFP_HIGH and allocations
880 * coming from realtime tasks go deeper into reserves.
882 * This is the last chance, in general, before the goto nopage.
883 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
884 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
886 page = get_page_from_freelist(gfp_mask, order, zonelist, alloc_flags);
890 /* This allocation should allow future memory freeing. */
892 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
893 && !in_interrupt()) {
894 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
896 /* go through the zonelist yet again, ignoring mins */
897 page = get_page_from_freelist(gfp_mask, order,
898 zonelist, ALLOC_NO_WATERMARKS|ALLOC_CPUSET);
901 if (gfp_mask & __GFP_NOFAIL) {
902 blk_congestion_wait(WRITE, HZ/50);
909 /* Atomic allocations - we can't balance anything */
916 /* We now go into synchronous reclaim */
917 p->flags |= PF_MEMALLOC;
918 reclaim_state.reclaimed_slab = 0;
919 p->reclaim_state = &reclaim_state;
921 did_some_progress = try_to_free_pages(zonelist->zones, gfp_mask);
923 p->reclaim_state = NULL;
924 p->flags &= ~PF_MEMALLOC;
928 if (likely(did_some_progress)) {
929 page = get_page_from_freelist(gfp_mask, order,
930 zonelist, alloc_flags);
933 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
935 * Go through the zonelist yet one more time, keep
936 * very high watermark here, this is only to catch
937 * a parallel oom killing, we must fail if we're still
938 * under heavy pressure.
940 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
941 zonelist, ALLOC_CPUSET);
945 out_of_memory(gfp_mask, order);
950 * Don't let big-order allocations loop unless the caller explicitly
951 * requests that. Wait for some write requests to complete then retry.
953 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
954 * <= 3, but that may not be true in other implementations.
957 if (!(gfp_mask & __GFP_NORETRY)) {
958 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
960 if (gfp_mask & __GFP_NOFAIL)
964 blk_congestion_wait(WRITE, HZ/50);
969 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
970 printk(KERN_WARNING "%s: page allocation failure."
971 " order:%d, mode:0x%x\n",
972 p->comm, order, gfp_mask);
980 EXPORT_SYMBOL(__alloc_pages);
983 * Common helper functions.
985 fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
988 page = alloc_pages(gfp_mask, order);
991 return (unsigned long) page_address(page);
994 EXPORT_SYMBOL(__get_free_pages);
996 fastcall unsigned long get_zeroed_page(gfp_t gfp_mask)
1001 * get_zeroed_page() returns a 32-bit address, which cannot represent
1004 BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1006 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1008 return (unsigned long) page_address(page);
1012 EXPORT_SYMBOL(get_zeroed_page);
1014 void __pagevec_free(struct pagevec *pvec)
1016 int i = pagevec_count(pvec);
1019 free_hot_cold_page(pvec->pages[i], pvec->cold);
1022 fastcall void __free_pages(struct page *page, unsigned int order)
1024 if (put_page_testzero(page)) {
1026 free_hot_page(page);
1028 __free_pages_ok(page, order);
1032 EXPORT_SYMBOL(__free_pages);
1034 fastcall void free_pages(unsigned long addr, unsigned int order)
1037 BUG_ON(!virt_addr_valid((void *)addr));
1038 __free_pages(virt_to_page((void *)addr), order);
1042 EXPORT_SYMBOL(free_pages);
1045 * Total amount of free (allocatable) RAM:
1047 unsigned int nr_free_pages(void)
1049 unsigned int sum = 0;
1053 sum += zone->free_pages;
1058 EXPORT_SYMBOL(nr_free_pages);
1061 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
1063 unsigned int i, sum = 0;
1065 for (i = 0; i < MAX_NR_ZONES; i++)
1066 sum += pgdat->node_zones[i].free_pages;
1072 static unsigned int nr_free_zone_pages(int offset)
1074 /* Just pick one node, since fallback list is circular */
1075 pg_data_t *pgdat = NODE_DATA(numa_node_id());
1076 unsigned int sum = 0;
1078 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1079 struct zone **zonep = zonelist->zones;
1082 for (zone = *zonep++; zone; zone = *zonep++) {
1083 unsigned long size = zone->present_pages;
1084 unsigned long high = zone->pages_high;
1093 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1095 unsigned int nr_free_buffer_pages(void)
1097 return nr_free_zone_pages(gfp_zone(GFP_USER));
1101 * Amount of free RAM allocatable within all zones
1103 unsigned int nr_free_pagecache_pages(void)
1105 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER));
1108 #ifdef CONFIG_HIGHMEM
1109 unsigned int nr_free_highpages (void)
1112 unsigned int pages = 0;
1114 for_each_pgdat(pgdat)
1115 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1122 static void show_node(struct zone *zone)
1124 printk("Node %d ", zone->zone_pgdat->node_id);
1127 #define show_node(zone) do { } while (0)
1131 * Accumulate the page_state information across all CPUs.
1132 * The result is unavoidably approximate - it can change
1133 * during and after execution of this function.
1135 static DEFINE_PER_CPU(struct page_state, page_states) = {0};
1137 atomic_t nr_pagecache = ATOMIC_INIT(0);
1138 EXPORT_SYMBOL(nr_pagecache);
1140 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
1143 void __get_page_state(struct page_state *ret, int nr, cpumask_t *cpumask)
1147 memset(ret, 0, sizeof(*ret));
1148 cpus_and(*cpumask, *cpumask, cpu_online_map);
1150 cpu = first_cpu(*cpumask);
1151 while (cpu < NR_CPUS) {
1152 unsigned long *in, *out, off;
1154 in = (unsigned long *)&per_cpu(page_states, cpu);
1156 cpu = next_cpu(cpu, *cpumask);
1159 prefetch(&per_cpu(page_states, cpu));
1161 out = (unsigned long *)ret;
1162 for (off = 0; off < nr; off++)
1167 void get_page_state_node(struct page_state *ret, int node)
1170 cpumask_t mask = node_to_cpumask(node);
1172 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
1173 nr /= sizeof(unsigned long);
1175 __get_page_state(ret, nr+1, &mask);
1178 void get_page_state(struct page_state *ret)
1181 cpumask_t mask = CPU_MASK_ALL;
1183 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
1184 nr /= sizeof(unsigned long);
1186 __get_page_state(ret, nr + 1, &mask);
1189 void get_full_page_state(struct page_state *ret)
1191 cpumask_t mask = CPU_MASK_ALL;
1193 __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long), &mask);
1196 unsigned long __read_page_state(unsigned long offset)
1198 unsigned long ret = 0;
1201 for_each_online_cpu(cpu) {
1204 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
1205 ret += *((unsigned long *)in);
1210 void __mod_page_state(unsigned long offset, unsigned long delta)
1212 unsigned long flags;
1215 local_irq_save(flags);
1216 ptr = &__get_cpu_var(page_states);
1217 *(unsigned long*)(ptr + offset) += delta;
1218 local_irq_restore(flags);
1221 EXPORT_SYMBOL(__mod_page_state);
1223 void __get_zone_counts(unsigned long *active, unsigned long *inactive,
1224 unsigned long *free, struct pglist_data *pgdat)
1226 struct zone *zones = pgdat->node_zones;
1232 for (i = 0; i < MAX_NR_ZONES; i++) {
1233 *active += zones[i].nr_active;
1234 *inactive += zones[i].nr_inactive;
1235 *free += zones[i].free_pages;
1239 void get_zone_counts(unsigned long *active,
1240 unsigned long *inactive, unsigned long *free)
1242 struct pglist_data *pgdat;
1247 for_each_pgdat(pgdat) {
1248 unsigned long l, m, n;
1249 __get_zone_counts(&l, &m, &n, pgdat);
1256 void si_meminfo(struct sysinfo *val)
1258 val->totalram = totalram_pages;
1260 val->freeram = nr_free_pages();
1261 val->bufferram = nr_blockdev_pages();
1262 #ifdef CONFIG_HIGHMEM
1263 val->totalhigh = totalhigh_pages;
1264 val->freehigh = nr_free_highpages();
1269 val->mem_unit = PAGE_SIZE;
1272 EXPORT_SYMBOL(si_meminfo);
1275 void si_meminfo_node(struct sysinfo *val, int nid)
1277 pg_data_t *pgdat = NODE_DATA(nid);
1279 val->totalram = pgdat->node_present_pages;
1280 val->freeram = nr_free_pages_pgdat(pgdat);
1281 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1282 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1283 val->mem_unit = PAGE_SIZE;
1287 #define K(x) ((x) << (PAGE_SHIFT-10))
1290 * Show free area list (used inside shift_scroll-lock stuff)
1291 * We also calculate the percentage fragmentation. We do this by counting the
1292 * memory on each free list with the exception of the first item on the list.
1294 void show_free_areas(void)
1296 struct page_state ps;
1297 int cpu, temperature;
1298 unsigned long active;
1299 unsigned long inactive;
1303 for_each_zone(zone) {
1305 printk("%s per-cpu:", zone->name);
1307 if (!zone->present_pages) {
1313 for_each_online_cpu(cpu) {
1314 struct per_cpu_pageset *pageset;
1316 pageset = zone_pcp(zone, cpu);
1318 for (temperature = 0; temperature < 2; temperature++)
1319 printk("cpu %d %s: low %d, high %d, batch %d used:%d\n",
1321 temperature ? "cold" : "hot",
1322 pageset->pcp[temperature].low,
1323 pageset->pcp[temperature].high,
1324 pageset->pcp[temperature].batch,
1325 pageset->pcp[temperature].count);
1329 get_page_state(&ps);
1330 get_zone_counts(&active, &inactive, &free);
1332 printk("Free pages: %11ukB (%ukB HighMem)\n",
1334 K(nr_free_highpages()));
1336 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1337 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1346 ps.nr_page_table_pages);
1348 for_each_zone(zone) {
1360 " pages_scanned:%lu"
1361 " all_unreclaimable? %s"
1364 K(zone->free_pages),
1367 K(zone->pages_high),
1369 K(zone->nr_inactive),
1370 K(zone->present_pages),
1371 zone->pages_scanned,
1372 (zone->all_unreclaimable ? "yes" : "no")
1374 printk("lowmem_reserve[]:");
1375 for (i = 0; i < MAX_NR_ZONES; i++)
1376 printk(" %lu", zone->lowmem_reserve[i]);
1380 for_each_zone(zone) {
1381 unsigned long nr, flags, order, total = 0;
1384 printk("%s: ", zone->name);
1385 if (!zone->present_pages) {
1390 spin_lock_irqsave(&zone->lock, flags);
1391 for (order = 0; order < MAX_ORDER; order++) {
1392 nr = zone->free_area[order].nr_free;
1393 total += nr << order;
1394 printk("%lu*%lukB ", nr, K(1UL) << order);
1396 spin_unlock_irqrestore(&zone->lock, flags);
1397 printk("= %lukB\n", K(total));
1400 show_swap_cache_info();
1404 * Builds allocation fallback zone lists.
1406 static int __init build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, int j, int k)
1413 zone = pgdat->node_zones + ZONE_HIGHMEM;
1414 if (zone->present_pages) {
1415 #ifndef CONFIG_HIGHMEM
1418 zonelist->zones[j++] = zone;
1421 zone = pgdat->node_zones + ZONE_NORMAL;
1422 if (zone->present_pages)
1423 zonelist->zones[j++] = zone;
1425 zone = pgdat->node_zones + ZONE_DMA;
1426 if (zone->present_pages)
1427 zonelist->zones[j++] = zone;
1433 static inline int highest_zone(int zone_bits)
1435 int res = ZONE_NORMAL;
1436 if (zone_bits & (__force int)__GFP_HIGHMEM)
1438 if (zone_bits & (__force int)__GFP_DMA)
1444 #define MAX_NODE_LOAD (num_online_nodes())
1445 static int __initdata node_load[MAX_NUMNODES];
1447 * find_next_best_node - find the next node that should appear in a given node's fallback list
1448 * @node: node whose fallback list we're appending
1449 * @used_node_mask: nodemask_t of already used nodes
1451 * We use a number of factors to determine which is the next node that should
1452 * appear on a given node's fallback list. The node should not have appeared
1453 * already in @node's fallback list, and it should be the next closest node
1454 * according to the distance array (which contains arbitrary distance values
1455 * from each node to each node in the system), and should also prefer nodes
1456 * with no CPUs, since presumably they'll have very little allocation pressure
1457 * on them otherwise.
1458 * It returns -1 if no node is found.
1460 static int __init find_next_best_node(int node, nodemask_t *used_node_mask)
1463 int min_val = INT_MAX;
1466 for_each_online_node(i) {
1469 /* Start from local node */
1470 n = (node+i) % num_online_nodes();
1472 /* Don't want a node to appear more than once */
1473 if (node_isset(n, *used_node_mask))
1476 /* Use the local node if we haven't already */
1477 if (!node_isset(node, *used_node_mask)) {
1482 /* Use the distance array to find the distance */
1483 val = node_distance(node, n);
1485 /* Give preference to headless and unused nodes */
1486 tmp = node_to_cpumask(n);
1487 if (!cpus_empty(tmp))
1488 val += PENALTY_FOR_NODE_WITH_CPUS;
1490 /* Slight preference for less loaded node */
1491 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1492 val += node_load[n];
1494 if (val < min_val) {
1501 node_set(best_node, *used_node_mask);
1506 static void __init build_zonelists(pg_data_t *pgdat)
1508 int i, j, k, node, local_node;
1509 int prev_node, load;
1510 struct zonelist *zonelist;
1511 nodemask_t used_mask;
1513 /* initialize zonelists */
1514 for (i = 0; i < GFP_ZONETYPES; i++) {
1515 zonelist = pgdat->node_zonelists + i;
1516 zonelist->zones[0] = NULL;
1519 /* NUMA-aware ordering of nodes */
1520 local_node = pgdat->node_id;
1521 load = num_online_nodes();
1522 prev_node = local_node;
1523 nodes_clear(used_mask);
1524 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1526 * We don't want to pressure a particular node.
1527 * So adding penalty to the first node in same
1528 * distance group to make it round-robin.
1530 if (node_distance(local_node, node) !=
1531 node_distance(local_node, prev_node))
1532 node_load[node] += load;
1535 for (i = 0; i < GFP_ZONETYPES; i++) {
1536 zonelist = pgdat->node_zonelists + i;
1537 for (j = 0; zonelist->zones[j] != NULL; j++);
1539 k = highest_zone(i);
1541 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1542 zonelist->zones[j] = NULL;
1547 #else /* CONFIG_NUMA */
1549 static void __init build_zonelists(pg_data_t *pgdat)
1551 int i, j, k, node, local_node;
1553 local_node = pgdat->node_id;
1554 for (i = 0; i < GFP_ZONETYPES; i++) {
1555 struct zonelist *zonelist;
1557 zonelist = pgdat->node_zonelists + i;
1560 k = highest_zone(i);
1561 j = build_zonelists_node(pgdat, zonelist, j, k);
1563 * Now we build the zonelist so that it contains the zones
1564 * of all the other nodes.
1565 * We don't want to pressure a particular node, so when
1566 * building the zones for node N, we make sure that the
1567 * zones coming right after the local ones are those from
1568 * node N+1 (modulo N)
1570 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1571 if (!node_online(node))
1573 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1575 for (node = 0; node < local_node; node++) {
1576 if (!node_online(node))
1578 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1581 zonelist->zones[j] = NULL;
1585 #endif /* CONFIG_NUMA */
1587 void __init build_all_zonelists(void)
1591 for_each_online_node(i)
1592 build_zonelists(NODE_DATA(i));
1593 printk("Built %i zonelists\n", num_online_nodes());
1594 cpuset_init_current_mems_allowed();
1598 * Helper functions to size the waitqueue hash table.
1599 * Essentially these want to choose hash table sizes sufficiently
1600 * large so that collisions trying to wait on pages are rare.
1601 * But in fact, the number of active page waitqueues on typical
1602 * systems is ridiculously low, less than 200. So this is even
1603 * conservative, even though it seems large.
1605 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1606 * waitqueues, i.e. the size of the waitq table given the number of pages.
1608 #define PAGES_PER_WAITQUEUE 256
1610 static inline unsigned long wait_table_size(unsigned long pages)
1612 unsigned long size = 1;
1614 pages /= PAGES_PER_WAITQUEUE;
1616 while (size < pages)
1620 * Once we have dozens or even hundreds of threads sleeping
1621 * on IO we've got bigger problems than wait queue collision.
1622 * Limit the size of the wait table to a reasonable size.
1624 size = min(size, 4096UL);
1626 return max(size, 4UL);
1630 * This is an integer logarithm so that shifts can be used later
1631 * to extract the more random high bits from the multiplicative
1632 * hash function before the remainder is taken.
1634 static inline unsigned long wait_table_bits(unsigned long size)
1639 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1641 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1642 unsigned long *zones_size, unsigned long *zholes_size)
1644 unsigned long realtotalpages, totalpages = 0;
1647 for (i = 0; i < MAX_NR_ZONES; i++)
1648 totalpages += zones_size[i];
1649 pgdat->node_spanned_pages = totalpages;
1651 realtotalpages = totalpages;
1653 for (i = 0; i < MAX_NR_ZONES; i++)
1654 realtotalpages -= zholes_size[i];
1655 pgdat->node_present_pages = realtotalpages;
1656 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1661 * Initially all pages are reserved - free ones are freed
1662 * up by free_all_bootmem() once the early boot process is
1663 * done. Non-atomic initialization, single-pass.
1665 void __devinit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1666 unsigned long start_pfn)
1669 unsigned long end_pfn = start_pfn + size;
1672 for (pfn = start_pfn; pfn < end_pfn; pfn++, page++) {
1673 if (!early_pfn_valid(pfn))
1675 if (!early_pfn_in_nid(pfn, nid))
1677 page = pfn_to_page(pfn);
1678 set_page_links(page, zone, nid, pfn);
1679 set_page_count(page, 1);
1680 reset_page_mapcount(page);
1681 SetPageReserved(page);
1682 INIT_LIST_HEAD(&page->lru);
1683 #ifdef WANT_PAGE_VIRTUAL
1684 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1685 if (!is_highmem_idx(zone))
1686 set_page_address(page, __va(pfn << PAGE_SHIFT));
1691 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1695 for (order = 0; order < MAX_ORDER ; order++) {
1696 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1697 zone->free_area[order].nr_free = 0;
1701 #define ZONETABLE_INDEX(x, zone_nr) ((x << ZONES_SHIFT) | zone_nr)
1702 void zonetable_add(struct zone *zone, int nid, int zid, unsigned long pfn,
1705 unsigned long snum = pfn_to_section_nr(pfn);
1706 unsigned long end = pfn_to_section_nr(pfn + size);
1709 zone_table[ZONETABLE_INDEX(nid, zid)] = zone;
1711 for (; snum <= end; snum++)
1712 zone_table[ZONETABLE_INDEX(snum, zid)] = zone;
1715 #ifndef __HAVE_ARCH_MEMMAP_INIT
1716 #define memmap_init(size, nid, zone, start_pfn) \
1717 memmap_init_zone((size), (nid), (zone), (start_pfn))
1720 static int __devinit zone_batchsize(struct zone *zone)
1725 * The per-cpu-pages pools are set to around 1000th of the
1726 * size of the zone. But no more than 1/2 of a meg.
1728 * OK, so we don't know how big the cache is. So guess.
1730 batch = zone->present_pages / 1024;
1731 if (batch * PAGE_SIZE > 512 * 1024)
1732 batch = (512 * 1024) / PAGE_SIZE;
1733 batch /= 4; /* We effectively *= 4 below */
1738 * We will be trying to allcoate bigger chunks of contiguous
1739 * memory of the order of fls(batch). This should result in
1740 * better cache coloring.
1742 * A sanity check also to ensure that batch is still in limits.
1744 batch = (1 << fls(batch + batch/2));
1746 if (fls(batch) >= (PAGE_SHIFT + MAX_ORDER - 2))
1747 batch = PAGE_SHIFT + ((MAX_ORDER - 1 - PAGE_SHIFT)/2);
1752 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
1754 struct per_cpu_pages *pcp;
1756 memset(p, 0, sizeof(*p));
1758 pcp = &p->pcp[0]; /* hot */
1761 pcp->high = 6 * batch;
1762 pcp->batch = max(1UL, 1 * batch);
1763 INIT_LIST_HEAD(&pcp->list);
1765 pcp = &p->pcp[1]; /* cold*/
1768 pcp->high = 2 * batch;
1769 pcp->batch = max(1UL, batch/2);
1770 INIT_LIST_HEAD(&pcp->list);
1775 * Boot pageset table. One per cpu which is going to be used for all
1776 * zones and all nodes. The parameters will be set in such a way
1777 * that an item put on a list will immediately be handed over to
1778 * the buddy list. This is safe since pageset manipulation is done
1779 * with interrupts disabled.
1781 * Some NUMA counter updates may also be caught by the boot pagesets.
1783 * The boot_pagesets must be kept even after bootup is complete for
1784 * unused processors and/or zones. They do play a role for bootstrapping
1785 * hotplugged processors.
1787 * zoneinfo_show() and maybe other functions do
1788 * not check if the processor is online before following the pageset pointer.
1789 * Other parts of the kernel may not check if the zone is available.
1791 static struct per_cpu_pageset
1792 boot_pageset[NR_CPUS];
1795 * Dynamically allocate memory for the
1796 * per cpu pageset array in struct zone.
1798 static int __devinit process_zones(int cpu)
1800 struct zone *zone, *dzone;
1802 for_each_zone(zone) {
1804 zone->pageset[cpu] = kmalloc_node(sizeof(struct per_cpu_pageset),
1805 GFP_KERNEL, cpu_to_node(cpu));
1806 if (!zone->pageset[cpu])
1809 setup_pageset(zone->pageset[cpu], zone_batchsize(zone));
1814 for_each_zone(dzone) {
1817 kfree(dzone->pageset[cpu]);
1818 dzone->pageset[cpu] = NULL;
1823 static inline void free_zone_pagesets(int cpu)
1828 for_each_zone(zone) {
1829 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
1831 zone_pcp(zone, cpu) = NULL;
1837 static int __devinit pageset_cpuup_callback(struct notifier_block *nfb,
1838 unsigned long action,
1841 int cpu = (long)hcpu;
1842 int ret = NOTIFY_OK;
1845 case CPU_UP_PREPARE:
1846 if (process_zones(cpu))
1849 #ifdef CONFIG_HOTPLUG_CPU
1851 free_zone_pagesets(cpu);
1860 static struct notifier_block pageset_notifier =
1861 { &pageset_cpuup_callback, NULL, 0 };
1863 void __init setup_per_cpu_pageset()
1867 /* Initialize per_cpu_pageset for cpu 0.
1868 * A cpuup callback will do this for every cpu
1869 * as it comes online
1871 err = process_zones(smp_processor_id());
1873 register_cpu_notifier(&pageset_notifier);
1879 void zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
1882 struct pglist_data *pgdat = zone->zone_pgdat;
1885 * The per-page waitqueue mechanism uses hashed waitqueues
1888 zone->wait_table_size = wait_table_size(zone_size_pages);
1889 zone->wait_table_bits = wait_table_bits(zone->wait_table_size);
1890 zone->wait_table = (wait_queue_head_t *)
1891 alloc_bootmem_node(pgdat, zone->wait_table_size
1892 * sizeof(wait_queue_head_t));
1894 for(i = 0; i < zone->wait_table_size; ++i)
1895 init_waitqueue_head(zone->wait_table + i);
1898 static __devinit void zone_pcp_init(struct zone *zone)
1901 unsigned long batch = zone_batchsize(zone);
1903 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1905 /* Early boot. Slab allocator not functional yet */
1906 zone->pageset[cpu] = &boot_pageset[cpu];
1907 setup_pageset(&boot_pageset[cpu],0);
1909 setup_pageset(zone_pcp(zone,cpu), batch);
1912 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
1913 zone->name, zone->present_pages, batch);
1916 static __devinit void init_currently_empty_zone(struct zone *zone,
1917 unsigned long zone_start_pfn, unsigned long size)
1919 struct pglist_data *pgdat = zone->zone_pgdat;
1921 zone_wait_table_init(zone, size);
1922 pgdat->nr_zones = zone_idx(zone) + 1;
1924 zone->zone_mem_map = pfn_to_page(zone_start_pfn);
1925 zone->zone_start_pfn = zone_start_pfn;
1927 memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn);
1929 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1933 * Set up the zone data structures:
1934 * - mark all pages reserved
1935 * - mark all memory queues empty
1936 * - clear the memory bitmaps
1938 static void __init free_area_init_core(struct pglist_data *pgdat,
1939 unsigned long *zones_size, unsigned long *zholes_size)
1942 int nid = pgdat->node_id;
1943 unsigned long zone_start_pfn = pgdat->node_start_pfn;
1945 pgdat_resize_init(pgdat);
1946 pgdat->nr_zones = 0;
1947 init_waitqueue_head(&pgdat->kswapd_wait);
1948 pgdat->kswapd_max_order = 0;
1950 for (j = 0; j < MAX_NR_ZONES; j++) {
1951 struct zone *zone = pgdat->node_zones + j;
1952 unsigned long size, realsize;
1954 realsize = size = zones_size[j];
1956 realsize -= zholes_size[j];
1958 if (j == ZONE_DMA || j == ZONE_NORMAL)
1959 nr_kernel_pages += realsize;
1960 nr_all_pages += realsize;
1962 zone->spanned_pages = size;
1963 zone->present_pages = realsize;
1964 zone->name = zone_names[j];
1965 spin_lock_init(&zone->lock);
1966 spin_lock_init(&zone->lru_lock);
1967 zone_seqlock_init(zone);
1968 zone->zone_pgdat = pgdat;
1969 zone->free_pages = 0;
1971 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
1973 zone_pcp_init(zone);
1974 INIT_LIST_HEAD(&zone->active_list);
1975 INIT_LIST_HEAD(&zone->inactive_list);
1976 zone->nr_scan_active = 0;
1977 zone->nr_scan_inactive = 0;
1978 zone->nr_active = 0;
1979 zone->nr_inactive = 0;
1980 atomic_set(&zone->reclaim_in_progress, 0);
1984 zonetable_add(zone, nid, j, zone_start_pfn, size);
1985 init_currently_empty_zone(zone, zone_start_pfn, size);
1986 zone_start_pfn += size;
1990 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
1992 /* Skip empty nodes */
1993 if (!pgdat->node_spanned_pages)
1996 #ifdef CONFIG_FLAT_NODE_MEM_MAP
1997 /* ia64 gets its own node_mem_map, before this, without bootmem */
1998 if (!pgdat->node_mem_map) {
2002 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
2003 map = alloc_remap(pgdat->node_id, size);
2005 map = alloc_bootmem_node(pgdat, size);
2006 pgdat->node_mem_map = map;
2008 #ifdef CONFIG_FLATMEM
2010 * With no DISCONTIG, the global mem_map is just set as node 0's
2012 if (pgdat == NODE_DATA(0))
2013 mem_map = NODE_DATA(0)->node_mem_map;
2015 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
2018 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
2019 unsigned long *zones_size, unsigned long node_start_pfn,
2020 unsigned long *zholes_size)
2022 pgdat->node_id = nid;
2023 pgdat->node_start_pfn = node_start_pfn;
2024 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
2026 alloc_node_mem_map(pgdat);
2028 free_area_init_core(pgdat, zones_size, zholes_size);
2031 #ifndef CONFIG_NEED_MULTIPLE_NODES
2032 static bootmem_data_t contig_bootmem_data;
2033 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
2035 EXPORT_SYMBOL(contig_page_data);
2038 void __init free_area_init(unsigned long *zones_size)
2040 free_area_init_node(0, NODE_DATA(0), zones_size,
2041 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
2044 #ifdef CONFIG_PROC_FS
2046 #include <linux/seq_file.h>
2048 static void *frag_start(struct seq_file *m, loff_t *pos)
2053 for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
2059 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
2061 pg_data_t *pgdat = (pg_data_t *)arg;
2064 return pgdat->pgdat_next;
2067 static void frag_stop(struct seq_file *m, void *arg)
2072 * This walks the free areas for each zone.
2074 static int frag_show(struct seq_file *m, void *arg)
2076 pg_data_t *pgdat = (pg_data_t *)arg;
2078 struct zone *node_zones = pgdat->node_zones;
2079 unsigned long flags;
2082 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2083 if (!zone->present_pages)
2086 spin_lock_irqsave(&zone->lock, flags);
2087 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
2088 for (order = 0; order < MAX_ORDER; ++order)
2089 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
2090 spin_unlock_irqrestore(&zone->lock, flags);
2096 struct seq_operations fragmentation_op = {
2097 .start = frag_start,
2104 * Output information about zones in @pgdat.
2106 static int zoneinfo_show(struct seq_file *m, void *arg)
2108 pg_data_t *pgdat = arg;
2110 struct zone *node_zones = pgdat->node_zones;
2111 unsigned long flags;
2113 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; zone++) {
2116 if (!zone->present_pages)
2119 spin_lock_irqsave(&zone->lock, flags);
2120 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
2128 "\n scanned %lu (a: %lu i: %lu)"
2137 zone->pages_scanned,
2138 zone->nr_scan_active, zone->nr_scan_inactive,
2139 zone->spanned_pages,
2140 zone->present_pages);
2142 "\n protection: (%lu",
2143 zone->lowmem_reserve[0]);
2144 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
2145 seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
2149 for (i = 0; i < ARRAY_SIZE(zone->pageset); i++) {
2150 struct per_cpu_pageset *pageset;
2153 pageset = zone_pcp(zone, i);
2154 for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2155 if (pageset->pcp[j].count)
2158 if (j == ARRAY_SIZE(pageset->pcp))
2160 for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2162 "\n cpu: %i pcp: %i"
2168 pageset->pcp[j].count,
2169 pageset->pcp[j].low,
2170 pageset->pcp[j].high,
2171 pageset->pcp[j].batch);
2177 "\n numa_foreign: %lu"
2178 "\n interleave_hit: %lu"
2179 "\n local_node: %lu"
2180 "\n other_node: %lu",
2183 pageset->numa_foreign,
2184 pageset->interleave_hit,
2185 pageset->local_node,
2186 pageset->other_node);
2190 "\n all_unreclaimable: %u"
2191 "\n prev_priority: %i"
2192 "\n temp_priority: %i"
2193 "\n start_pfn: %lu",
2194 zone->all_unreclaimable,
2195 zone->prev_priority,
2196 zone->temp_priority,
2197 zone->zone_start_pfn);
2198 spin_unlock_irqrestore(&zone->lock, flags);
2204 struct seq_operations zoneinfo_op = {
2205 .start = frag_start, /* iterate over all zones. The same as in
2209 .show = zoneinfo_show,
2212 static char *vmstat_text[] = {
2216 "nr_page_table_pages",
2241 "pgscan_kswapd_high",
2242 "pgscan_kswapd_normal",
2244 "pgscan_kswapd_dma",
2245 "pgscan_direct_high",
2246 "pgscan_direct_normal",
2247 "pgscan_direct_dma",
2252 "kswapd_inodesteal",
2260 static void *vmstat_start(struct seq_file *m, loff_t *pos)
2262 struct page_state *ps;
2264 if (*pos >= ARRAY_SIZE(vmstat_text))
2267 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
2270 return ERR_PTR(-ENOMEM);
2271 get_full_page_state(ps);
2272 ps->pgpgin /= 2; /* sectors -> kbytes */
2274 return (unsigned long *)ps + *pos;
2277 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
2280 if (*pos >= ARRAY_SIZE(vmstat_text))
2282 return (unsigned long *)m->private + *pos;
2285 static int vmstat_show(struct seq_file *m, void *arg)
2287 unsigned long *l = arg;
2288 unsigned long off = l - (unsigned long *)m->private;
2290 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
2294 static void vmstat_stop(struct seq_file *m, void *arg)
2300 struct seq_operations vmstat_op = {
2301 .start = vmstat_start,
2302 .next = vmstat_next,
2303 .stop = vmstat_stop,
2304 .show = vmstat_show,
2307 #endif /* CONFIG_PROC_FS */
2309 #ifdef CONFIG_HOTPLUG_CPU
2310 static int page_alloc_cpu_notify(struct notifier_block *self,
2311 unsigned long action, void *hcpu)
2313 int cpu = (unsigned long)hcpu;
2315 unsigned long *src, *dest;
2317 if (action == CPU_DEAD) {
2320 /* Drain local pagecache count. */
2321 count = &per_cpu(nr_pagecache_local, cpu);
2322 atomic_add(*count, &nr_pagecache);
2324 local_irq_disable();
2327 /* Add dead cpu's page_states to our own. */
2328 dest = (unsigned long *)&__get_cpu_var(page_states);
2329 src = (unsigned long *)&per_cpu(page_states, cpu);
2331 for (i = 0; i < sizeof(struct page_state)/sizeof(unsigned long);
2341 #endif /* CONFIG_HOTPLUG_CPU */
2343 void __init page_alloc_init(void)
2345 hotcpu_notifier(page_alloc_cpu_notify, 0);
2349 * setup_per_zone_lowmem_reserve - called whenever
2350 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
2351 * has a correct pages reserved value, so an adequate number of
2352 * pages are left in the zone after a successful __alloc_pages().
2354 static void setup_per_zone_lowmem_reserve(void)
2356 struct pglist_data *pgdat;
2359 for_each_pgdat(pgdat) {
2360 for (j = 0; j < MAX_NR_ZONES; j++) {
2361 struct zone *zone = pgdat->node_zones + j;
2362 unsigned long present_pages = zone->present_pages;
2364 zone->lowmem_reserve[j] = 0;
2366 for (idx = j-1; idx >= 0; idx--) {
2367 struct zone *lower_zone;
2369 if (sysctl_lowmem_reserve_ratio[idx] < 1)
2370 sysctl_lowmem_reserve_ratio[idx] = 1;
2372 lower_zone = pgdat->node_zones + idx;
2373 lower_zone->lowmem_reserve[j] = present_pages /
2374 sysctl_lowmem_reserve_ratio[idx];
2375 present_pages += lower_zone->present_pages;
2382 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
2383 * that the pages_{min,low,high} values for each zone are set correctly
2384 * with respect to min_free_kbytes.
2386 void setup_per_zone_pages_min(void)
2388 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
2389 unsigned long lowmem_pages = 0;
2391 unsigned long flags;
2393 /* Calculate total number of !ZONE_HIGHMEM pages */
2394 for_each_zone(zone) {
2395 if (!is_highmem(zone))
2396 lowmem_pages += zone->present_pages;
2399 for_each_zone(zone) {
2400 spin_lock_irqsave(&zone->lru_lock, flags);
2401 if (is_highmem(zone)) {
2403 * Often, highmem doesn't need to reserve any pages.
2404 * But the pages_min/low/high values are also used for
2405 * batching up page reclaim activity so we need a
2406 * decent value here.
2410 min_pages = zone->present_pages / 1024;
2411 if (min_pages < SWAP_CLUSTER_MAX)
2412 min_pages = SWAP_CLUSTER_MAX;
2413 if (min_pages > 128)
2415 zone->pages_min = min_pages;
2417 /* if it's a lowmem zone, reserve a number of pages
2418 * proportionate to the zone's size.
2420 zone->pages_min = (pages_min * zone->present_pages) /
2425 * When interpreting these watermarks, just keep in mind that:
2426 * zone->pages_min == (zone->pages_min * 4) / 4;
2428 zone->pages_low = (zone->pages_min * 5) / 4;
2429 zone->pages_high = (zone->pages_min * 6) / 4;
2430 spin_unlock_irqrestore(&zone->lru_lock, flags);
2435 * Initialise min_free_kbytes.
2437 * For small machines we want it small (128k min). For large machines
2438 * we want it large (64MB max). But it is not linear, because network
2439 * bandwidth does not increase linearly with machine size. We use
2441 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2442 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2458 static int __init init_per_zone_pages_min(void)
2460 unsigned long lowmem_kbytes;
2462 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2464 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2465 if (min_free_kbytes < 128)
2466 min_free_kbytes = 128;
2467 if (min_free_kbytes > 65536)
2468 min_free_kbytes = 65536;
2469 setup_per_zone_pages_min();
2470 setup_per_zone_lowmem_reserve();
2473 module_init(init_per_zone_pages_min)
2476 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2477 * that we can call two helper functions whenever min_free_kbytes
2480 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2481 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2483 proc_dointvec(table, write, file, buffer, length, ppos);
2484 setup_per_zone_pages_min();
2489 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2490 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2491 * whenever sysctl_lowmem_reserve_ratio changes.
2493 * The reserve ratio obviously has absolutely no relation with the
2494 * pages_min watermarks. The lowmem reserve ratio can only make sense
2495 * if in function of the boot time zone sizes.
2497 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
2498 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2500 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2501 setup_per_zone_lowmem_reserve();
2505 __initdata int hashdist = HASHDIST_DEFAULT;
2508 static int __init set_hashdist(char *str)
2512 hashdist = simple_strtoul(str, &str, 0);
2515 __setup("hashdist=", set_hashdist);
2519 * allocate a large system hash table from bootmem
2520 * - it is assumed that the hash table must contain an exact power-of-2
2521 * quantity of entries
2522 * - limit is the number of hash buckets, not the total allocation size
2524 void *__init alloc_large_system_hash(const char *tablename,
2525 unsigned long bucketsize,
2526 unsigned long numentries,
2529 unsigned int *_hash_shift,
2530 unsigned int *_hash_mask,
2531 unsigned long limit)
2533 unsigned long long max = limit;
2534 unsigned long log2qty, size;
2537 /* allow the kernel cmdline to have a say */
2539 /* round applicable memory size up to nearest megabyte */
2540 numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages;
2541 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2542 numentries >>= 20 - PAGE_SHIFT;
2543 numentries <<= 20 - PAGE_SHIFT;
2545 /* limit to 1 bucket per 2^scale bytes of low memory */
2546 if (scale > PAGE_SHIFT)
2547 numentries >>= (scale - PAGE_SHIFT);
2549 numentries <<= (PAGE_SHIFT - scale);
2551 /* rounded up to nearest power of 2 in size */
2552 numentries = 1UL << (long_log2(numentries) + 1);
2554 /* limit allocation size to 1/16 total memory by default */
2556 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2557 do_div(max, bucketsize);
2560 if (numentries > max)
2563 log2qty = long_log2(numentries);
2566 size = bucketsize << log2qty;
2567 if (flags & HASH_EARLY)
2568 table = alloc_bootmem(size);
2570 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
2572 unsigned long order;
2573 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
2575 table = (void*) __get_free_pages(GFP_ATOMIC, order);
2577 } while (!table && size > PAGE_SIZE && --log2qty);
2580 panic("Failed to allocate %s hash table\n", tablename);
2582 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2585 long_log2(size) - PAGE_SHIFT,
2589 *_hash_shift = log2qty;
2591 *_hash_mask = (1 << log2qty) - 1;