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);
67 EXPORT_SYMBOL(nr_swap_pages);
70 * Used by page_zone() to look up the address of the struct zone whose
71 * id is encoded in the upper bits of page->flags
73 struct zone *zone_table[1 << ZONETABLE_SHIFT] __read_mostly;
74 EXPORT_SYMBOL(zone_table);
76 static char *zone_names[MAX_NR_ZONES] = { "DMA", "Normal", "HighMem" };
77 int min_free_kbytes = 1024;
79 unsigned long __initdata nr_kernel_pages;
80 unsigned long __initdata nr_all_pages;
82 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
86 unsigned long pfn = page_to_pfn(page);
89 seq = zone_span_seqbegin(zone);
90 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
92 else if (pfn < zone->zone_start_pfn)
94 } while (zone_span_seqretry(zone, seq));
99 static int page_is_consistent(struct zone *zone, struct page *page)
101 #ifdef CONFIG_HOLES_IN_ZONE
102 if (!pfn_valid(page_to_pfn(page)))
105 if (zone != page_zone(page))
111 * Temporary debugging check for pages not lying within a given zone.
113 static int bad_range(struct zone *zone, struct page *page)
115 if (page_outside_zone_boundaries(zone, page))
117 if (!page_is_consistent(zone, page))
123 static void bad_page(const char *function, struct page *page)
125 printk(KERN_EMERG "Bad page state at %s (in process '%s', page %p)\n",
126 function, current->comm, page);
127 printk(KERN_EMERG "flags:0x%0*lx mapping:%p mapcount:%d count:%d\n",
128 (int)(2*sizeof(page_flags_t)), (unsigned long)page->flags,
129 page->mapping, page_mapcount(page), page_count(page));
130 printk(KERN_EMERG "Backtrace:\n");
132 printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n");
133 page->flags &= ~(1 << PG_lru |
143 set_page_count(page, 0);
144 reset_page_mapcount(page);
145 page->mapping = NULL;
146 add_taint(TAINT_BAD_PAGE);
149 #ifndef CONFIG_HUGETLB_PAGE
150 #define prep_compound_page(page, order) do { } while (0)
151 #define destroy_compound_page(page, order) do { } while (0)
154 * Higher-order pages are called "compound pages". They are structured thusly:
156 * The first PAGE_SIZE page is called the "head page".
158 * The remaining PAGE_SIZE pages are called "tail pages".
160 * All pages have PG_compound set. All pages have their ->private pointing at
161 * the head page (even the head page has this).
163 * The first tail page's ->mapping, if non-zero, holds the address of the
164 * compound page's put_page() function.
166 * The order of the allocation is stored in the first tail page's ->index
167 * This is only for debug at present. This usage means that zero-order pages
168 * may not be compound.
170 static void prep_compound_page(struct page *page, unsigned long order)
173 int nr_pages = 1 << order;
175 page[1].mapping = NULL;
176 page[1].index = order;
177 for (i = 0; i < nr_pages; i++) {
178 struct page *p = page + i;
181 set_page_private(p, (unsigned long)page);
185 static void destroy_compound_page(struct page *page, unsigned long order)
188 int nr_pages = 1 << order;
190 if (!PageCompound(page))
193 if (page[1].index != order)
194 bad_page(__FUNCTION__, page);
196 for (i = 0; i < nr_pages; i++) {
197 struct page *p = page + i;
199 if (!PageCompound(p))
200 bad_page(__FUNCTION__, page);
201 if (page_private(p) != (unsigned long)page)
202 bad_page(__FUNCTION__, page);
203 ClearPageCompound(p);
206 #endif /* CONFIG_HUGETLB_PAGE */
209 * function for dealing with page's order in buddy system.
210 * zone->lock is already acquired when we use these.
211 * So, we don't need atomic page->flags operations here.
213 static inline unsigned long page_order(struct page *page) {
214 return page_private(page);
217 static inline void set_page_order(struct page *page, int order) {
218 set_page_private(page, order);
219 __SetPagePrivate(page);
222 static inline void rmv_page_order(struct page *page)
224 __ClearPagePrivate(page);
225 set_page_private(page, 0);
229 * Locate the struct page for both the matching buddy in our
230 * pair (buddy1) and the combined O(n+1) page they form (page).
232 * 1) Any buddy B1 will have an order O twin B2 which satisfies
233 * the following equation:
235 * For example, if the starting buddy (buddy2) is #8 its order
237 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
239 * 2) Any buddy B will have an order O+1 parent P which
240 * satisfies the following equation:
243 * Assumption: *_mem_map is contigious at least up to MAX_ORDER
245 static inline struct page *
246 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
248 unsigned long buddy_idx = page_idx ^ (1 << order);
250 return page + (buddy_idx - page_idx);
253 static inline unsigned long
254 __find_combined_index(unsigned long page_idx, unsigned int order)
256 return (page_idx & ~(1 << order));
260 * This function checks whether a page is free && is the buddy
261 * we can do coalesce a page and its buddy if
262 * (a) the buddy is free &&
263 * (b) the buddy is on the buddy system &&
264 * (c) a page and its buddy have the same order.
265 * for recording page's order, we use page_private(page) and PG_private.
268 static inline int page_is_buddy(struct page *page, int order)
270 if (PagePrivate(page) &&
271 (page_order(page) == order) &&
272 page_count(page) == 0)
278 * Freeing function for a buddy system allocator.
280 * The concept of a buddy system is to maintain direct-mapped table
281 * (containing bit values) for memory blocks of various "orders".
282 * The bottom level table contains the map for the smallest allocatable
283 * units of memory (here, pages), and each level above it describes
284 * pairs of units from the levels below, hence, "buddies".
285 * At a high level, all that happens here is marking the table entry
286 * at the bottom level available, and propagating the changes upward
287 * as necessary, plus some accounting needed to play nicely with other
288 * parts of the VM system.
289 * At each level, we keep a list of pages, which are heads of continuous
290 * free pages of length of (1 << order) and marked with PG_Private.Page's
291 * order is recorded in page_private(page) field.
292 * So when we are allocating or freeing one, we can derive the state of the
293 * other. That is, if we allocate a small block, and both were
294 * free, the remainder of the region must be split into blocks.
295 * If a block is freed, and its buddy is also free, then this
296 * triggers coalescing into a block of larger size.
301 static inline void __free_pages_bulk (struct page *page,
302 struct zone *zone, unsigned int order)
304 unsigned long page_idx;
305 int order_size = 1 << order;
308 destroy_compound_page(page, order);
310 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
312 BUG_ON(page_idx & (order_size - 1));
313 BUG_ON(bad_range(zone, page));
315 zone->free_pages += order_size;
316 while (order < MAX_ORDER-1) {
317 unsigned long combined_idx;
318 struct free_area *area;
321 combined_idx = __find_combined_index(page_idx, order);
322 buddy = __page_find_buddy(page, page_idx, order);
324 if (bad_range(zone, buddy))
326 if (!page_is_buddy(buddy, order))
327 break; /* Move the buddy up one level. */
328 list_del(&buddy->lru);
329 area = zone->free_area + order;
331 rmv_page_order(buddy);
332 page = page + (combined_idx - page_idx);
333 page_idx = combined_idx;
336 set_page_order(page, order);
337 list_add(&page->lru, &zone->free_area[order].free_list);
338 zone->free_area[order].nr_free++;
341 static inline void free_pages_check(const char *function, struct page *page)
343 if ( page_mapcount(page) ||
344 page->mapping != NULL ||
345 page_count(page) != 0 ||
356 bad_page(function, page);
358 __ClearPageDirty(page);
362 * Frees a list of pages.
363 * Assumes all pages on list are in same zone, and of same order.
364 * count is the number of pages to free.
366 * If the zone was previously in an "all pages pinned" state then look to
367 * see if this freeing clears that state.
369 * And clear the zone's pages_scanned counter, to hold off the "all pages are
370 * pinned" detection logic.
373 free_pages_bulk(struct zone *zone, int count,
374 struct list_head *list, unsigned int order)
377 struct page *page = NULL;
380 spin_lock_irqsave(&zone->lock, flags);
381 zone->all_unreclaimable = 0;
382 zone->pages_scanned = 0;
383 while (!list_empty(list) && count--) {
384 page = list_entry(list->prev, struct page, lru);
385 /* have to delete it as __free_pages_bulk list manipulates */
386 list_del(&page->lru);
387 __free_pages_bulk(page, zone, order);
390 spin_unlock_irqrestore(&zone->lock, flags);
394 void __free_pages_ok(struct page *page, unsigned int order)
399 arch_free_page(page, order);
401 mod_page_state(pgfree, 1 << order);
405 for (i = 1 ; i < (1 << order) ; ++i)
406 __put_page(page + i);
409 for (i = 0 ; i < (1 << order) ; ++i)
410 free_pages_check(__FUNCTION__, page + i);
411 list_add(&page->lru, &list);
412 kernel_map_pages(page, 1<<order, 0);
413 free_pages_bulk(page_zone(page), 1, &list, order);
418 * The order of subdivision here is critical for the IO subsystem.
419 * Please do not alter this order without good reasons and regression
420 * testing. Specifically, as large blocks of memory are subdivided,
421 * the order in which smaller blocks are delivered depends on the order
422 * they're subdivided in this function. This is the primary factor
423 * influencing the order in which pages are delivered to the IO
424 * subsystem according to empirical testing, and this is also justified
425 * by considering the behavior of a buddy system containing a single
426 * large block of memory acted on by a series of small allocations.
427 * This behavior is a critical factor in sglist merging's success.
431 static inline struct page *
432 expand(struct zone *zone, struct page *page,
433 int low, int high, struct free_area *area)
435 unsigned long size = 1 << high;
441 BUG_ON(bad_range(zone, &page[size]));
442 list_add(&page[size].lru, &area->free_list);
444 set_page_order(&page[size], high);
449 void set_page_refs(struct page *page, int order)
452 set_page_count(page, 1);
457 * We need to reference all the pages for this order, otherwise if
458 * anyone accesses one of the pages with (get/put) it will be freed.
459 * - eg: access_process_vm()
461 for (i = 0; i < (1 << order); i++)
462 set_page_count(page + i, 1);
463 #endif /* CONFIG_MMU */
467 * This page is about to be returned from the page allocator
469 static void prep_new_page(struct page *page, int order)
471 if ( page_mapcount(page) ||
472 page->mapping != NULL ||
473 page_count(page) != 0 ||
485 bad_page(__FUNCTION__, page);
487 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
488 1 << PG_referenced | 1 << PG_arch_1 |
489 1 << PG_checked | 1 << PG_mappedtodisk);
490 set_page_private(page, 0);
491 set_page_refs(page, order);
492 kernel_map_pages(page, 1 << order, 1);
496 * Do the hard work of removing an element from the buddy allocator.
497 * Call me with the zone->lock already held.
499 static struct page *__rmqueue(struct zone *zone, unsigned int order)
501 struct free_area * area;
502 unsigned int current_order;
505 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
506 area = zone->free_area + current_order;
507 if (list_empty(&area->free_list))
510 page = list_entry(area->free_list.next, struct page, lru);
511 list_del(&page->lru);
512 rmv_page_order(page);
514 zone->free_pages -= 1UL << order;
515 return expand(zone, page, order, current_order, area);
522 * Obtain a specified number of elements from the buddy allocator, all under
523 * a single hold of the lock, for efficiency. Add them to the supplied list.
524 * Returns the number of new pages which were placed at *list.
526 static int rmqueue_bulk(struct zone *zone, unsigned int order,
527 unsigned long count, struct list_head *list)
534 spin_lock_irqsave(&zone->lock, flags);
535 for (i = 0; i < count; ++i) {
536 page = __rmqueue(zone, order);
540 list_add_tail(&page->lru, list);
542 spin_unlock_irqrestore(&zone->lock, flags);
547 /* Called from the slab reaper to drain remote pagesets */
548 void drain_remote_pages(void)
554 local_irq_save(flags);
555 for_each_zone(zone) {
556 struct per_cpu_pageset *pset;
558 /* Do not drain local pagesets */
559 if (zone->zone_pgdat->node_id == numa_node_id())
562 pset = zone->pageset[smp_processor_id()];
563 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
564 struct per_cpu_pages *pcp;
568 pcp->count -= free_pages_bulk(zone, pcp->count,
572 local_irq_restore(flags);
576 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
577 static void __drain_pages(unsigned int cpu)
582 for_each_zone(zone) {
583 struct per_cpu_pageset *pset;
585 pset = zone_pcp(zone, cpu);
586 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
587 struct per_cpu_pages *pcp;
590 pcp->count -= free_pages_bulk(zone, pcp->count,
595 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
599 void mark_free_pages(struct zone *zone)
601 unsigned long zone_pfn, flags;
603 struct list_head *curr;
605 if (!zone->spanned_pages)
608 spin_lock_irqsave(&zone->lock, flags);
609 for (zone_pfn = 0; zone_pfn < zone->spanned_pages; ++zone_pfn)
610 ClearPageNosaveFree(pfn_to_page(zone_pfn + zone->zone_start_pfn));
612 for (order = MAX_ORDER - 1; order >= 0; --order)
613 list_for_each(curr, &zone->free_area[order].free_list) {
614 unsigned long start_pfn, i;
616 start_pfn = page_to_pfn(list_entry(curr, struct page, lru));
618 for (i=0; i < (1<<order); i++)
619 SetPageNosaveFree(pfn_to_page(start_pfn+i));
621 spin_unlock_irqrestore(&zone->lock, flags);
625 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
627 void drain_local_pages(void)
631 local_irq_save(flags);
632 __drain_pages(smp_processor_id());
633 local_irq_restore(flags);
635 #endif /* CONFIG_PM */
637 static void zone_statistics(struct zonelist *zonelist, struct zone *z)
642 pg_data_t *pg = z->zone_pgdat;
643 pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
644 struct per_cpu_pageset *p;
646 local_irq_save(flags);
647 cpu = smp_processor_id();
653 zone_pcp(zonelist->zones[0], cpu)->numa_foreign++;
655 if (pg == NODE_DATA(numa_node_id()))
659 local_irq_restore(flags);
664 * Free a 0-order page
666 static void FASTCALL(free_hot_cold_page(struct page *page, int cold));
667 static void fastcall free_hot_cold_page(struct page *page, int cold)
669 struct zone *zone = page_zone(page);
670 struct per_cpu_pages *pcp;
673 arch_free_page(page, 0);
675 kernel_map_pages(page, 1, 0);
676 inc_page_state(pgfree);
678 page->mapping = NULL;
679 free_pages_check(__FUNCTION__, page);
680 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
681 local_irq_save(flags);
682 list_add(&page->lru, &pcp->list);
684 if (pcp->count >= pcp->high)
685 pcp->count -= free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
686 local_irq_restore(flags);
690 void fastcall free_hot_page(struct page *page)
692 free_hot_cold_page(page, 0);
695 void fastcall free_cold_page(struct page *page)
697 free_hot_cold_page(page, 1);
700 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
704 BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
705 for(i = 0; i < (1 << order); i++)
706 clear_highpage(page + i);
710 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
711 * we cheat by calling it from here, in the order > 0 path. Saves a branch
715 buffered_rmqueue(struct zone *zone, int order, gfp_t gfp_flags)
718 struct page *page = NULL;
719 int cold = !!(gfp_flags & __GFP_COLD);
722 struct per_cpu_pages *pcp;
724 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
725 local_irq_save(flags);
726 if (pcp->count <= pcp->low)
727 pcp->count += rmqueue_bulk(zone, 0,
728 pcp->batch, &pcp->list);
730 page = list_entry(pcp->list.next, struct page, lru);
731 list_del(&page->lru);
734 local_irq_restore(flags);
739 spin_lock_irqsave(&zone->lock, flags);
740 page = __rmqueue(zone, order);
741 spin_unlock_irqrestore(&zone->lock, flags);
745 BUG_ON(bad_range(zone, page));
746 mod_page_state_zone(zone, pgalloc, 1 << order);
747 prep_new_page(page, order);
749 if (gfp_flags & __GFP_ZERO)
750 prep_zero_page(page, order, gfp_flags);
752 if (order && (gfp_flags & __GFP_COMP))
753 prep_compound_page(page, order);
759 * Return 1 if free pages are above 'mark'. This takes into account the order
762 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
763 int classzone_idx, int can_try_harder, gfp_t gfp_high)
765 /* free_pages my go negative - that's OK */
766 long min = mark, free_pages = z->free_pages - (1 << order) + 1;
774 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
776 for (o = 0; o < order; o++) {
777 /* At the next order, this order's pages become unavailable */
778 free_pages -= z->free_area[o].nr_free << o;
780 /* Require fewer higher order pages to be free */
783 if (free_pages <= min)
790 should_reclaim_zone(struct zone *z, gfp_t gfp_mask)
792 if (!z->reclaim_pages)
794 if (gfp_mask & __GFP_NORECLAIM)
800 * This is the 'heart' of the zoned buddy allocator.
802 struct page * fastcall
803 __alloc_pages(gfp_t gfp_mask, unsigned int order,
804 struct zonelist *zonelist)
806 const gfp_t wait = gfp_mask & __GFP_WAIT;
807 struct zone **zones, *z;
809 struct reclaim_state reclaim_state;
810 struct task_struct *p = current;
815 int did_some_progress;
817 might_sleep_if(wait);
820 * The caller may dip into page reserves a bit more if the caller
821 * cannot run direct reclaim, or is the caller has realtime scheduling
824 can_try_harder = (unlikely(rt_task(p)) && !in_interrupt()) || !wait;
826 zones = zonelist->zones; /* the list of zones suitable for gfp_mask */
828 if (unlikely(zones[0] == NULL)) {
829 /* Should this ever happen?? */
833 classzone_idx = zone_idx(zones[0]);
837 * Go through the zonelist once, looking for a zone with enough free.
838 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
840 for (i = 0; (z = zones[i]) != NULL; i++) {
841 int do_reclaim = should_reclaim_zone(z, gfp_mask);
843 if (!cpuset_zone_allowed(z, __GFP_HARDWALL))
847 * If the zone is to attempt early page reclaim then this loop
848 * will try to reclaim pages and check the watermark a second
849 * time before giving up and falling back to the next zone.
852 if (!zone_watermark_ok(z, order, z->pages_low,
853 classzone_idx, 0, 0)) {
857 zone_reclaim(z, gfp_mask, order);
858 /* Only try reclaim once */
860 goto zone_reclaim_retry;
864 page = buffered_rmqueue(z, order, gfp_mask);
869 for (i = 0; (z = zones[i]) != NULL; i++)
870 wakeup_kswapd(z, order);
873 * Go through the zonelist again. Let __GFP_HIGH and allocations
874 * coming from realtime tasks to go deeper into reserves
876 * This is the last chance, in general, before the goto nopage.
877 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
878 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
880 for (i = 0; (z = zones[i]) != NULL; i++) {
881 if (!zone_watermark_ok(z, order, z->pages_min,
882 classzone_idx, can_try_harder,
883 gfp_mask & __GFP_HIGH))
886 if (wait && !cpuset_zone_allowed(z, gfp_mask))
889 page = buffered_rmqueue(z, order, gfp_mask);
894 /* This allocation should allow future memory freeing. */
896 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
897 && !in_interrupt()) {
898 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
899 /* go through the zonelist yet again, ignoring mins */
900 for (i = 0; (z = zones[i]) != NULL; i++) {
901 if (!cpuset_zone_allowed(z, gfp_mask))
903 page = buffered_rmqueue(z, order, gfp_mask);
911 /* Atomic allocations - we can't balance anything */
918 /* We now go into synchronous reclaim */
919 p->flags |= PF_MEMALLOC;
920 reclaim_state.reclaimed_slab = 0;
921 p->reclaim_state = &reclaim_state;
923 did_some_progress = try_to_free_pages(zones, gfp_mask);
925 p->reclaim_state = NULL;
926 p->flags &= ~PF_MEMALLOC;
930 if (likely(did_some_progress)) {
931 for (i = 0; (z = zones[i]) != NULL; i++) {
932 if (!zone_watermark_ok(z, order, z->pages_min,
933 classzone_idx, can_try_harder,
934 gfp_mask & __GFP_HIGH))
937 if (!cpuset_zone_allowed(z, gfp_mask))
940 page = buffered_rmqueue(z, order, gfp_mask);
944 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
946 * Go through the zonelist yet one more time, keep
947 * very high watermark here, this is only to catch
948 * a parallel oom killing, we must fail if we're still
949 * under heavy pressure.
951 for (i = 0; (z = zones[i]) != NULL; i++) {
952 if (!zone_watermark_ok(z, order, z->pages_high,
953 classzone_idx, 0, 0))
956 if (!cpuset_zone_allowed(z, __GFP_HARDWALL))
959 page = buffered_rmqueue(z, order, gfp_mask);
964 out_of_memory(gfp_mask, order);
969 * Don't let big-order allocations loop unless the caller explicitly
970 * requests that. Wait for some write requests to complete then retry.
972 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
973 * <= 3, but that may not be true in other implementations.
976 if (!(gfp_mask & __GFP_NORETRY)) {
977 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
979 if (gfp_mask & __GFP_NOFAIL)
983 blk_congestion_wait(WRITE, HZ/50);
988 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
989 printk(KERN_WARNING "%s: page allocation failure."
990 " order:%d, mode:0x%x\n",
991 p->comm, order, gfp_mask);
997 zone_statistics(zonelist, z);
1001 EXPORT_SYMBOL(__alloc_pages);
1004 * Common helper functions.
1006 fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1009 page = alloc_pages(gfp_mask, order);
1012 return (unsigned long) page_address(page);
1015 EXPORT_SYMBOL(__get_free_pages);
1017 fastcall unsigned long get_zeroed_page(gfp_t gfp_mask)
1022 * get_zeroed_page() returns a 32-bit address, which cannot represent
1025 BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1027 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1029 return (unsigned long) page_address(page);
1033 EXPORT_SYMBOL(get_zeroed_page);
1035 void __pagevec_free(struct pagevec *pvec)
1037 int i = pagevec_count(pvec);
1040 free_hot_cold_page(pvec->pages[i], pvec->cold);
1043 fastcall void __free_pages(struct page *page, unsigned int order)
1045 if (put_page_testzero(page)) {
1047 free_hot_page(page);
1049 __free_pages_ok(page, order);
1053 EXPORT_SYMBOL(__free_pages);
1055 fastcall void free_pages(unsigned long addr, unsigned int order)
1058 BUG_ON(!virt_addr_valid((void *)addr));
1059 __free_pages(virt_to_page((void *)addr), order);
1063 EXPORT_SYMBOL(free_pages);
1066 * Total amount of free (allocatable) RAM:
1068 unsigned int nr_free_pages(void)
1070 unsigned int sum = 0;
1074 sum += zone->free_pages;
1079 EXPORT_SYMBOL(nr_free_pages);
1082 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
1084 unsigned int i, sum = 0;
1086 for (i = 0; i < MAX_NR_ZONES; i++)
1087 sum += pgdat->node_zones[i].free_pages;
1093 static unsigned int nr_free_zone_pages(int offset)
1095 /* Just pick one node, since fallback list is circular */
1096 pg_data_t *pgdat = NODE_DATA(numa_node_id());
1097 unsigned int sum = 0;
1099 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1100 struct zone **zonep = zonelist->zones;
1103 for (zone = *zonep++; zone; zone = *zonep++) {
1104 unsigned long size = zone->present_pages;
1105 unsigned long high = zone->pages_high;
1114 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1116 unsigned int nr_free_buffer_pages(void)
1118 return nr_free_zone_pages(gfp_zone(GFP_USER));
1122 * Amount of free RAM allocatable within all zones
1124 unsigned int nr_free_pagecache_pages(void)
1126 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER));
1129 #ifdef CONFIG_HIGHMEM
1130 unsigned int nr_free_highpages (void)
1133 unsigned int pages = 0;
1135 for_each_pgdat(pgdat)
1136 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1143 static void show_node(struct zone *zone)
1145 printk("Node %d ", zone->zone_pgdat->node_id);
1148 #define show_node(zone) do { } while (0)
1152 * Accumulate the page_state information across all CPUs.
1153 * The result is unavoidably approximate - it can change
1154 * during and after execution of this function.
1156 static DEFINE_PER_CPU(struct page_state, page_states) = {0};
1158 atomic_t nr_pagecache = ATOMIC_INIT(0);
1159 EXPORT_SYMBOL(nr_pagecache);
1161 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
1164 void __get_page_state(struct page_state *ret, int nr, cpumask_t *cpumask)
1168 memset(ret, 0, sizeof(*ret));
1169 cpus_and(*cpumask, *cpumask, cpu_online_map);
1171 cpu = first_cpu(*cpumask);
1172 while (cpu < NR_CPUS) {
1173 unsigned long *in, *out, off;
1175 in = (unsigned long *)&per_cpu(page_states, cpu);
1177 cpu = next_cpu(cpu, *cpumask);
1180 prefetch(&per_cpu(page_states, cpu));
1182 out = (unsigned long *)ret;
1183 for (off = 0; off < nr; off++)
1188 void get_page_state_node(struct page_state *ret, int node)
1191 cpumask_t mask = node_to_cpumask(node);
1193 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
1194 nr /= sizeof(unsigned long);
1196 __get_page_state(ret, nr+1, &mask);
1199 void get_page_state(struct page_state *ret)
1202 cpumask_t mask = CPU_MASK_ALL;
1204 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
1205 nr /= sizeof(unsigned long);
1207 __get_page_state(ret, nr + 1, &mask);
1210 void get_full_page_state(struct page_state *ret)
1212 cpumask_t mask = CPU_MASK_ALL;
1214 __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long), &mask);
1217 unsigned long __read_page_state(unsigned long offset)
1219 unsigned long ret = 0;
1222 for_each_online_cpu(cpu) {
1225 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
1226 ret += *((unsigned long *)in);
1231 void __mod_page_state(unsigned long offset, unsigned long delta)
1233 unsigned long flags;
1236 local_irq_save(flags);
1237 ptr = &__get_cpu_var(page_states);
1238 *(unsigned long*)(ptr + offset) += delta;
1239 local_irq_restore(flags);
1242 EXPORT_SYMBOL(__mod_page_state);
1244 void __get_zone_counts(unsigned long *active, unsigned long *inactive,
1245 unsigned long *free, struct pglist_data *pgdat)
1247 struct zone *zones = pgdat->node_zones;
1253 for (i = 0; i < MAX_NR_ZONES; i++) {
1254 *active += zones[i].nr_active;
1255 *inactive += zones[i].nr_inactive;
1256 *free += zones[i].free_pages;
1260 void get_zone_counts(unsigned long *active,
1261 unsigned long *inactive, unsigned long *free)
1263 struct pglist_data *pgdat;
1268 for_each_pgdat(pgdat) {
1269 unsigned long l, m, n;
1270 __get_zone_counts(&l, &m, &n, pgdat);
1277 void si_meminfo(struct sysinfo *val)
1279 val->totalram = totalram_pages;
1281 val->freeram = nr_free_pages();
1282 val->bufferram = nr_blockdev_pages();
1283 #ifdef CONFIG_HIGHMEM
1284 val->totalhigh = totalhigh_pages;
1285 val->freehigh = nr_free_highpages();
1290 val->mem_unit = PAGE_SIZE;
1293 EXPORT_SYMBOL(si_meminfo);
1296 void si_meminfo_node(struct sysinfo *val, int nid)
1298 pg_data_t *pgdat = NODE_DATA(nid);
1300 val->totalram = pgdat->node_present_pages;
1301 val->freeram = nr_free_pages_pgdat(pgdat);
1302 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1303 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1304 val->mem_unit = PAGE_SIZE;
1308 #define K(x) ((x) << (PAGE_SHIFT-10))
1311 * Show free area list (used inside shift_scroll-lock stuff)
1312 * We also calculate the percentage fragmentation. We do this by counting the
1313 * memory on each free list with the exception of the first item on the list.
1315 void show_free_areas(void)
1317 struct page_state ps;
1318 int cpu, temperature;
1319 unsigned long active;
1320 unsigned long inactive;
1324 for_each_zone(zone) {
1326 printk("%s per-cpu:", zone->name);
1328 if (!zone->present_pages) {
1334 for (cpu = 0; cpu < NR_CPUS; ++cpu) {
1335 struct per_cpu_pageset *pageset;
1337 if (!cpu_possible(cpu))
1340 pageset = zone_pcp(zone, cpu);
1342 for (temperature = 0; temperature < 2; temperature++)
1343 printk("cpu %d %s: low %d, high %d, batch %d used:%d\n",
1345 temperature ? "cold" : "hot",
1346 pageset->pcp[temperature].low,
1347 pageset->pcp[temperature].high,
1348 pageset->pcp[temperature].batch,
1349 pageset->pcp[temperature].count);
1353 get_page_state(&ps);
1354 get_zone_counts(&active, &inactive, &free);
1356 printk("Free pages: %11ukB (%ukB HighMem)\n",
1358 K(nr_free_highpages()));
1360 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1361 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1370 ps.nr_page_table_pages);
1372 for_each_zone(zone) {
1384 " pages_scanned:%lu"
1385 " all_unreclaimable? %s"
1388 K(zone->free_pages),
1391 K(zone->pages_high),
1393 K(zone->nr_inactive),
1394 K(zone->present_pages),
1395 zone->pages_scanned,
1396 (zone->all_unreclaimable ? "yes" : "no")
1398 printk("lowmem_reserve[]:");
1399 for (i = 0; i < MAX_NR_ZONES; i++)
1400 printk(" %lu", zone->lowmem_reserve[i]);
1404 for_each_zone(zone) {
1405 unsigned long nr, flags, order, total = 0;
1408 printk("%s: ", zone->name);
1409 if (!zone->present_pages) {
1414 spin_lock_irqsave(&zone->lock, flags);
1415 for (order = 0; order < MAX_ORDER; order++) {
1416 nr = zone->free_area[order].nr_free;
1417 total += nr << order;
1418 printk("%lu*%lukB ", nr, K(1UL) << order);
1420 spin_unlock_irqrestore(&zone->lock, flags);
1421 printk("= %lukB\n", K(total));
1424 show_swap_cache_info();
1428 * Builds allocation fallback zone lists.
1430 static int __init build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, int j, int k)
1437 zone = pgdat->node_zones + ZONE_HIGHMEM;
1438 if (zone->present_pages) {
1439 #ifndef CONFIG_HIGHMEM
1442 zonelist->zones[j++] = zone;
1445 zone = pgdat->node_zones + ZONE_NORMAL;
1446 if (zone->present_pages)
1447 zonelist->zones[j++] = zone;
1449 zone = pgdat->node_zones + ZONE_DMA;
1450 if (zone->present_pages)
1451 zonelist->zones[j++] = zone;
1457 static inline int highest_zone(int zone_bits)
1459 int res = ZONE_NORMAL;
1460 if (zone_bits & (__force int)__GFP_HIGHMEM)
1462 if (zone_bits & (__force int)__GFP_DMA)
1468 #define MAX_NODE_LOAD (num_online_nodes())
1469 static int __initdata node_load[MAX_NUMNODES];
1471 * find_next_best_node - find the next node that should appear in a given node's fallback list
1472 * @node: node whose fallback list we're appending
1473 * @used_node_mask: nodemask_t of already used nodes
1475 * We use a number of factors to determine which is the next node that should
1476 * appear on a given node's fallback list. The node should not have appeared
1477 * already in @node's fallback list, and it should be the next closest node
1478 * according to the distance array (which contains arbitrary distance values
1479 * from each node to each node in the system), and should also prefer nodes
1480 * with no CPUs, since presumably they'll have very little allocation pressure
1481 * on them otherwise.
1482 * It returns -1 if no node is found.
1484 static int __init find_next_best_node(int node, nodemask_t *used_node_mask)
1487 int min_val = INT_MAX;
1490 for_each_online_node(i) {
1493 /* Start from local node */
1494 n = (node+i) % num_online_nodes();
1496 /* Don't want a node to appear more than once */
1497 if (node_isset(n, *used_node_mask))
1500 /* Use the local node if we haven't already */
1501 if (!node_isset(node, *used_node_mask)) {
1506 /* Use the distance array to find the distance */
1507 val = node_distance(node, n);
1509 /* Give preference to headless and unused nodes */
1510 tmp = node_to_cpumask(n);
1511 if (!cpus_empty(tmp))
1512 val += PENALTY_FOR_NODE_WITH_CPUS;
1514 /* Slight preference for less loaded node */
1515 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1516 val += node_load[n];
1518 if (val < min_val) {
1525 node_set(best_node, *used_node_mask);
1530 static void __init build_zonelists(pg_data_t *pgdat)
1532 int i, j, k, node, local_node;
1533 int prev_node, load;
1534 struct zonelist *zonelist;
1535 nodemask_t used_mask;
1537 /* initialize zonelists */
1538 for (i = 0; i < GFP_ZONETYPES; i++) {
1539 zonelist = pgdat->node_zonelists + i;
1540 zonelist->zones[0] = NULL;
1543 /* NUMA-aware ordering of nodes */
1544 local_node = pgdat->node_id;
1545 load = num_online_nodes();
1546 prev_node = local_node;
1547 nodes_clear(used_mask);
1548 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1550 * We don't want to pressure a particular node.
1551 * So adding penalty to the first node in same
1552 * distance group to make it round-robin.
1554 if (node_distance(local_node, node) !=
1555 node_distance(local_node, prev_node))
1556 node_load[node] += load;
1559 for (i = 0; i < GFP_ZONETYPES; i++) {
1560 zonelist = pgdat->node_zonelists + i;
1561 for (j = 0; zonelist->zones[j] != NULL; j++);
1563 k = highest_zone(i);
1565 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1566 zonelist->zones[j] = NULL;
1571 #else /* CONFIG_NUMA */
1573 static void __init build_zonelists(pg_data_t *pgdat)
1575 int i, j, k, node, local_node;
1577 local_node = pgdat->node_id;
1578 for (i = 0; i < GFP_ZONETYPES; i++) {
1579 struct zonelist *zonelist;
1581 zonelist = pgdat->node_zonelists + i;
1584 k = highest_zone(i);
1585 j = build_zonelists_node(pgdat, zonelist, j, k);
1587 * Now we build the zonelist so that it contains the zones
1588 * of all the other nodes.
1589 * We don't want to pressure a particular node, so when
1590 * building the zones for node N, we make sure that the
1591 * zones coming right after the local ones are those from
1592 * node N+1 (modulo N)
1594 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1595 if (!node_online(node))
1597 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1599 for (node = 0; node < local_node; node++) {
1600 if (!node_online(node))
1602 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1605 zonelist->zones[j] = NULL;
1609 #endif /* CONFIG_NUMA */
1611 void __init build_all_zonelists(void)
1615 for_each_online_node(i)
1616 build_zonelists(NODE_DATA(i));
1617 printk("Built %i zonelists\n", num_online_nodes());
1618 cpuset_init_current_mems_allowed();
1622 * Helper functions to size the waitqueue hash table.
1623 * Essentially these want to choose hash table sizes sufficiently
1624 * large so that collisions trying to wait on pages are rare.
1625 * But in fact, the number of active page waitqueues on typical
1626 * systems is ridiculously low, less than 200. So this is even
1627 * conservative, even though it seems large.
1629 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1630 * waitqueues, i.e. the size of the waitq table given the number of pages.
1632 #define PAGES_PER_WAITQUEUE 256
1634 static inline unsigned long wait_table_size(unsigned long pages)
1636 unsigned long size = 1;
1638 pages /= PAGES_PER_WAITQUEUE;
1640 while (size < pages)
1644 * Once we have dozens or even hundreds of threads sleeping
1645 * on IO we've got bigger problems than wait queue collision.
1646 * Limit the size of the wait table to a reasonable size.
1648 size = min(size, 4096UL);
1650 return max(size, 4UL);
1654 * This is an integer logarithm so that shifts can be used later
1655 * to extract the more random high bits from the multiplicative
1656 * hash function before the remainder is taken.
1658 static inline unsigned long wait_table_bits(unsigned long size)
1663 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1665 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1666 unsigned long *zones_size, unsigned long *zholes_size)
1668 unsigned long realtotalpages, totalpages = 0;
1671 for (i = 0; i < MAX_NR_ZONES; i++)
1672 totalpages += zones_size[i];
1673 pgdat->node_spanned_pages = totalpages;
1675 realtotalpages = totalpages;
1677 for (i = 0; i < MAX_NR_ZONES; i++)
1678 realtotalpages -= zholes_size[i];
1679 pgdat->node_present_pages = realtotalpages;
1680 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1685 * Initially all pages are reserved - free ones are freed
1686 * up by free_all_bootmem() once the early boot process is
1687 * done. Non-atomic initialization, single-pass.
1689 void __init memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1690 unsigned long start_pfn)
1693 unsigned long end_pfn = start_pfn + size;
1696 for (pfn = start_pfn; pfn < end_pfn; pfn++, page++) {
1697 if (!early_pfn_valid(pfn))
1699 if (!early_pfn_in_nid(pfn, nid))
1701 page = pfn_to_page(pfn);
1702 set_page_links(page, zone, nid, pfn);
1703 set_page_count(page, 1);
1704 reset_page_mapcount(page);
1705 SetPageReserved(page);
1706 INIT_LIST_HEAD(&page->lru);
1707 #ifdef WANT_PAGE_VIRTUAL
1708 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1709 if (!is_highmem_idx(zone))
1710 set_page_address(page, __va(pfn << PAGE_SHIFT));
1715 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1719 for (order = 0; order < MAX_ORDER ; order++) {
1720 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1721 zone->free_area[order].nr_free = 0;
1725 #define ZONETABLE_INDEX(x, zone_nr) ((x << ZONES_SHIFT) | zone_nr)
1726 void zonetable_add(struct zone *zone, int nid, int zid, unsigned long pfn,
1729 unsigned long snum = pfn_to_section_nr(pfn);
1730 unsigned long end = pfn_to_section_nr(pfn + size);
1733 zone_table[ZONETABLE_INDEX(nid, zid)] = zone;
1735 for (; snum <= end; snum++)
1736 zone_table[ZONETABLE_INDEX(snum, zid)] = zone;
1739 #ifndef __HAVE_ARCH_MEMMAP_INIT
1740 #define memmap_init(size, nid, zone, start_pfn) \
1741 memmap_init_zone((size), (nid), (zone), (start_pfn))
1744 static int __devinit zone_batchsize(struct zone *zone)
1749 * The per-cpu-pages pools are set to around 1000th of the
1750 * size of the zone. But no more than 1/2 of a meg.
1752 * OK, so we don't know how big the cache is. So guess.
1754 batch = zone->present_pages / 1024;
1755 if (batch * PAGE_SIZE > 512 * 1024)
1756 batch = (512 * 1024) / PAGE_SIZE;
1757 batch /= 4; /* We effectively *= 4 below */
1762 * We will be trying to allcoate bigger chunks of contiguous
1763 * memory of the order of fls(batch). This should result in
1764 * better cache coloring.
1766 * A sanity check also to ensure that batch is still in limits.
1768 batch = (1 << fls(batch + batch/2));
1770 if (fls(batch) >= (PAGE_SHIFT + MAX_ORDER - 2))
1771 batch = PAGE_SHIFT + ((MAX_ORDER - 1 - PAGE_SHIFT)/2);
1776 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
1778 struct per_cpu_pages *pcp;
1780 memset(p, 0, sizeof(*p));
1782 pcp = &p->pcp[0]; /* hot */
1785 pcp->high = 6 * batch;
1786 pcp->batch = max(1UL, 1 * batch);
1787 INIT_LIST_HEAD(&pcp->list);
1789 pcp = &p->pcp[1]; /* cold*/
1792 pcp->high = 2 * batch;
1793 pcp->batch = max(1UL, batch/2);
1794 INIT_LIST_HEAD(&pcp->list);
1799 * Boot pageset table. One per cpu which is going to be used for all
1800 * zones and all nodes. The parameters will be set in such a way
1801 * that an item put on a list will immediately be handed over to
1802 * the buddy list. This is safe since pageset manipulation is done
1803 * with interrupts disabled.
1805 * Some NUMA counter updates may also be caught by the boot pagesets.
1807 * The boot_pagesets must be kept even after bootup is complete for
1808 * unused processors and/or zones. They do play a role for bootstrapping
1809 * hotplugged processors.
1811 * zoneinfo_show() and maybe other functions do
1812 * not check if the processor is online before following the pageset pointer.
1813 * Other parts of the kernel may not check if the zone is available.
1815 static struct per_cpu_pageset
1816 boot_pageset[NR_CPUS];
1819 * Dynamically allocate memory for the
1820 * per cpu pageset array in struct zone.
1822 static int __devinit process_zones(int cpu)
1824 struct zone *zone, *dzone;
1826 for_each_zone(zone) {
1828 zone->pageset[cpu] = kmalloc_node(sizeof(struct per_cpu_pageset),
1829 GFP_KERNEL, cpu_to_node(cpu));
1830 if (!zone->pageset[cpu])
1833 setup_pageset(zone->pageset[cpu], zone_batchsize(zone));
1838 for_each_zone(dzone) {
1841 kfree(dzone->pageset[cpu]);
1842 dzone->pageset[cpu] = NULL;
1847 static inline void free_zone_pagesets(int cpu)
1852 for_each_zone(zone) {
1853 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
1855 zone_pcp(zone, cpu) = NULL;
1861 static int __devinit pageset_cpuup_callback(struct notifier_block *nfb,
1862 unsigned long action,
1865 int cpu = (long)hcpu;
1866 int ret = NOTIFY_OK;
1869 case CPU_UP_PREPARE:
1870 if (process_zones(cpu))
1873 #ifdef CONFIG_HOTPLUG_CPU
1875 free_zone_pagesets(cpu);
1884 static struct notifier_block pageset_notifier =
1885 { &pageset_cpuup_callback, NULL, 0 };
1887 void __init setup_per_cpu_pageset()
1891 /* Initialize per_cpu_pageset for cpu 0.
1892 * A cpuup callback will do this for every cpu
1893 * as it comes online
1895 err = process_zones(smp_processor_id());
1897 register_cpu_notifier(&pageset_notifier);
1903 void zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
1906 struct pglist_data *pgdat = zone->zone_pgdat;
1909 * The per-page waitqueue mechanism uses hashed waitqueues
1912 zone->wait_table_size = wait_table_size(zone_size_pages);
1913 zone->wait_table_bits = wait_table_bits(zone->wait_table_size);
1914 zone->wait_table = (wait_queue_head_t *)
1915 alloc_bootmem_node(pgdat, zone->wait_table_size
1916 * sizeof(wait_queue_head_t));
1918 for(i = 0; i < zone->wait_table_size; ++i)
1919 init_waitqueue_head(zone->wait_table + i);
1922 static __devinit void zone_pcp_init(struct zone *zone)
1925 unsigned long batch = zone_batchsize(zone);
1927 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1929 /* Early boot. Slab allocator not functional yet */
1930 zone->pageset[cpu] = &boot_pageset[cpu];
1931 setup_pageset(&boot_pageset[cpu],0);
1933 setup_pageset(zone_pcp(zone,cpu), batch);
1936 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
1937 zone->name, zone->present_pages, batch);
1940 static __devinit void init_currently_empty_zone(struct zone *zone,
1941 unsigned long zone_start_pfn, unsigned long size)
1943 struct pglist_data *pgdat = zone->zone_pgdat;
1945 zone_wait_table_init(zone, size);
1946 pgdat->nr_zones = zone_idx(zone) + 1;
1948 zone->zone_mem_map = pfn_to_page(zone_start_pfn);
1949 zone->zone_start_pfn = zone_start_pfn;
1951 memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn);
1953 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1957 * Set up the zone data structures:
1958 * - mark all pages reserved
1959 * - mark all memory queues empty
1960 * - clear the memory bitmaps
1962 static void __init free_area_init_core(struct pglist_data *pgdat,
1963 unsigned long *zones_size, unsigned long *zholes_size)
1966 int nid = pgdat->node_id;
1967 unsigned long zone_start_pfn = pgdat->node_start_pfn;
1969 pgdat_resize_init(pgdat);
1970 pgdat->nr_zones = 0;
1971 init_waitqueue_head(&pgdat->kswapd_wait);
1972 pgdat->kswapd_max_order = 0;
1974 for (j = 0; j < MAX_NR_ZONES; j++) {
1975 struct zone *zone = pgdat->node_zones + j;
1976 unsigned long size, realsize;
1978 realsize = size = zones_size[j];
1980 realsize -= zholes_size[j];
1982 if (j == ZONE_DMA || j == ZONE_NORMAL)
1983 nr_kernel_pages += realsize;
1984 nr_all_pages += realsize;
1986 zone->spanned_pages = size;
1987 zone->present_pages = realsize;
1988 zone->name = zone_names[j];
1989 spin_lock_init(&zone->lock);
1990 spin_lock_init(&zone->lru_lock);
1991 zone_seqlock_init(zone);
1992 zone->zone_pgdat = pgdat;
1993 zone->free_pages = 0;
1995 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
1997 zone_pcp_init(zone);
1998 INIT_LIST_HEAD(&zone->active_list);
1999 INIT_LIST_HEAD(&zone->inactive_list);
2000 zone->nr_scan_active = 0;
2001 zone->nr_scan_inactive = 0;
2002 zone->nr_active = 0;
2003 zone->nr_inactive = 0;
2004 atomic_set(&zone->reclaim_in_progress, 0);
2008 zonetable_add(zone, nid, j, zone_start_pfn, size);
2009 init_currently_empty_zone(zone, zone_start_pfn, size);
2010 zone_start_pfn += size;
2014 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
2016 /* Skip empty nodes */
2017 if (!pgdat->node_spanned_pages)
2020 #ifdef CONFIG_FLAT_NODE_MEM_MAP
2021 /* ia64 gets its own node_mem_map, before this, without bootmem */
2022 if (!pgdat->node_mem_map) {
2026 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
2027 map = alloc_remap(pgdat->node_id, size);
2029 map = alloc_bootmem_node(pgdat, size);
2030 pgdat->node_mem_map = map;
2032 #ifdef CONFIG_FLATMEM
2034 * With no DISCONTIG, the global mem_map is just set as node 0's
2036 if (pgdat == NODE_DATA(0))
2037 mem_map = NODE_DATA(0)->node_mem_map;
2039 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
2042 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
2043 unsigned long *zones_size, unsigned long node_start_pfn,
2044 unsigned long *zholes_size)
2046 pgdat->node_id = nid;
2047 pgdat->node_start_pfn = node_start_pfn;
2048 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
2050 alloc_node_mem_map(pgdat);
2052 free_area_init_core(pgdat, zones_size, zholes_size);
2055 #ifndef CONFIG_NEED_MULTIPLE_NODES
2056 static bootmem_data_t contig_bootmem_data;
2057 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
2059 EXPORT_SYMBOL(contig_page_data);
2062 void __init free_area_init(unsigned long *zones_size)
2064 free_area_init_node(0, NODE_DATA(0), zones_size,
2065 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
2068 #ifdef CONFIG_PROC_FS
2070 #include <linux/seq_file.h>
2072 static void *frag_start(struct seq_file *m, loff_t *pos)
2077 for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
2083 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
2085 pg_data_t *pgdat = (pg_data_t *)arg;
2088 return pgdat->pgdat_next;
2091 static void frag_stop(struct seq_file *m, void *arg)
2096 * This walks the free areas for each zone.
2098 static int frag_show(struct seq_file *m, void *arg)
2100 pg_data_t *pgdat = (pg_data_t *)arg;
2102 struct zone *node_zones = pgdat->node_zones;
2103 unsigned long flags;
2106 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2107 if (!zone->present_pages)
2110 spin_lock_irqsave(&zone->lock, flags);
2111 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
2112 for (order = 0; order < MAX_ORDER; ++order)
2113 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
2114 spin_unlock_irqrestore(&zone->lock, flags);
2120 struct seq_operations fragmentation_op = {
2121 .start = frag_start,
2128 * Output information about zones in @pgdat.
2130 static int zoneinfo_show(struct seq_file *m, void *arg)
2132 pg_data_t *pgdat = arg;
2134 struct zone *node_zones = pgdat->node_zones;
2135 unsigned long flags;
2137 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; zone++) {
2140 if (!zone->present_pages)
2143 spin_lock_irqsave(&zone->lock, flags);
2144 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
2152 "\n scanned %lu (a: %lu i: %lu)"
2161 zone->pages_scanned,
2162 zone->nr_scan_active, zone->nr_scan_inactive,
2163 zone->spanned_pages,
2164 zone->present_pages);
2166 "\n protection: (%lu",
2167 zone->lowmem_reserve[0]);
2168 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
2169 seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
2173 for (i = 0; i < ARRAY_SIZE(zone->pageset); i++) {
2174 struct per_cpu_pageset *pageset;
2177 pageset = zone_pcp(zone, i);
2178 for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2179 if (pageset->pcp[j].count)
2182 if (j == ARRAY_SIZE(pageset->pcp))
2184 for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2186 "\n cpu: %i pcp: %i"
2192 pageset->pcp[j].count,
2193 pageset->pcp[j].low,
2194 pageset->pcp[j].high,
2195 pageset->pcp[j].batch);
2201 "\n numa_foreign: %lu"
2202 "\n interleave_hit: %lu"
2203 "\n local_node: %lu"
2204 "\n other_node: %lu",
2207 pageset->numa_foreign,
2208 pageset->interleave_hit,
2209 pageset->local_node,
2210 pageset->other_node);
2214 "\n all_unreclaimable: %u"
2215 "\n prev_priority: %i"
2216 "\n temp_priority: %i"
2217 "\n start_pfn: %lu",
2218 zone->all_unreclaimable,
2219 zone->prev_priority,
2220 zone->temp_priority,
2221 zone->zone_start_pfn);
2222 spin_unlock_irqrestore(&zone->lock, flags);
2228 struct seq_operations zoneinfo_op = {
2229 .start = frag_start, /* iterate over all zones. The same as in
2233 .show = zoneinfo_show,
2236 static char *vmstat_text[] = {
2240 "nr_page_table_pages",
2265 "pgscan_kswapd_high",
2266 "pgscan_kswapd_normal",
2268 "pgscan_kswapd_dma",
2269 "pgscan_direct_high",
2270 "pgscan_direct_normal",
2271 "pgscan_direct_dma",
2276 "kswapd_inodesteal",
2284 static void *vmstat_start(struct seq_file *m, loff_t *pos)
2286 struct page_state *ps;
2288 if (*pos >= ARRAY_SIZE(vmstat_text))
2291 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
2294 return ERR_PTR(-ENOMEM);
2295 get_full_page_state(ps);
2296 ps->pgpgin /= 2; /* sectors -> kbytes */
2298 return (unsigned long *)ps + *pos;
2301 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
2304 if (*pos >= ARRAY_SIZE(vmstat_text))
2306 return (unsigned long *)m->private + *pos;
2309 static int vmstat_show(struct seq_file *m, void *arg)
2311 unsigned long *l = arg;
2312 unsigned long off = l - (unsigned long *)m->private;
2314 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
2318 static void vmstat_stop(struct seq_file *m, void *arg)
2324 struct seq_operations vmstat_op = {
2325 .start = vmstat_start,
2326 .next = vmstat_next,
2327 .stop = vmstat_stop,
2328 .show = vmstat_show,
2331 #endif /* CONFIG_PROC_FS */
2333 #ifdef CONFIG_HOTPLUG_CPU
2334 static int page_alloc_cpu_notify(struct notifier_block *self,
2335 unsigned long action, void *hcpu)
2337 int cpu = (unsigned long)hcpu;
2339 unsigned long *src, *dest;
2341 if (action == CPU_DEAD) {
2344 /* Drain local pagecache count. */
2345 count = &per_cpu(nr_pagecache_local, cpu);
2346 atomic_add(*count, &nr_pagecache);
2348 local_irq_disable();
2351 /* Add dead cpu's page_states to our own. */
2352 dest = (unsigned long *)&__get_cpu_var(page_states);
2353 src = (unsigned long *)&per_cpu(page_states, cpu);
2355 for (i = 0; i < sizeof(struct page_state)/sizeof(unsigned long);
2365 #endif /* CONFIG_HOTPLUG_CPU */
2367 void __init page_alloc_init(void)
2369 hotcpu_notifier(page_alloc_cpu_notify, 0);
2373 * setup_per_zone_lowmem_reserve - called whenever
2374 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
2375 * has a correct pages reserved value, so an adequate number of
2376 * pages are left in the zone after a successful __alloc_pages().
2378 static void setup_per_zone_lowmem_reserve(void)
2380 struct pglist_data *pgdat;
2383 for_each_pgdat(pgdat) {
2384 for (j = 0; j < MAX_NR_ZONES; j++) {
2385 struct zone *zone = pgdat->node_zones + j;
2386 unsigned long present_pages = zone->present_pages;
2388 zone->lowmem_reserve[j] = 0;
2390 for (idx = j-1; idx >= 0; idx--) {
2391 struct zone *lower_zone;
2393 if (sysctl_lowmem_reserve_ratio[idx] < 1)
2394 sysctl_lowmem_reserve_ratio[idx] = 1;
2396 lower_zone = pgdat->node_zones + idx;
2397 lower_zone->lowmem_reserve[j] = present_pages /
2398 sysctl_lowmem_reserve_ratio[idx];
2399 present_pages += lower_zone->present_pages;
2406 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
2407 * that the pages_{min,low,high} values for each zone are set correctly
2408 * with respect to min_free_kbytes.
2410 static void setup_per_zone_pages_min(void)
2412 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
2413 unsigned long lowmem_pages = 0;
2415 unsigned long flags;
2417 /* Calculate total number of !ZONE_HIGHMEM pages */
2418 for_each_zone(zone) {
2419 if (!is_highmem(zone))
2420 lowmem_pages += zone->present_pages;
2423 for_each_zone(zone) {
2424 spin_lock_irqsave(&zone->lru_lock, flags);
2425 if (is_highmem(zone)) {
2427 * Often, highmem doesn't need to reserve any pages.
2428 * But the pages_min/low/high values are also used for
2429 * batching up page reclaim activity so we need a
2430 * decent value here.
2434 min_pages = zone->present_pages / 1024;
2435 if (min_pages < SWAP_CLUSTER_MAX)
2436 min_pages = SWAP_CLUSTER_MAX;
2437 if (min_pages > 128)
2439 zone->pages_min = min_pages;
2441 /* if it's a lowmem zone, reserve a number of pages
2442 * proportionate to the zone's size.
2444 zone->pages_min = (pages_min * zone->present_pages) /
2449 * When interpreting these watermarks, just keep in mind that:
2450 * zone->pages_min == (zone->pages_min * 4) / 4;
2452 zone->pages_low = (zone->pages_min * 5) / 4;
2453 zone->pages_high = (zone->pages_min * 6) / 4;
2454 spin_unlock_irqrestore(&zone->lru_lock, flags);
2459 * Initialise min_free_kbytes.
2461 * For small machines we want it small (128k min). For large machines
2462 * we want it large (64MB max). But it is not linear, because network
2463 * bandwidth does not increase linearly with machine size. We use
2465 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2466 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2482 static int __init init_per_zone_pages_min(void)
2484 unsigned long lowmem_kbytes;
2486 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2488 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2489 if (min_free_kbytes < 128)
2490 min_free_kbytes = 128;
2491 if (min_free_kbytes > 65536)
2492 min_free_kbytes = 65536;
2493 setup_per_zone_pages_min();
2494 setup_per_zone_lowmem_reserve();
2497 module_init(init_per_zone_pages_min)
2500 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2501 * that we can call two helper functions whenever min_free_kbytes
2504 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2505 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2507 proc_dointvec(table, write, file, buffer, length, ppos);
2508 setup_per_zone_pages_min();
2513 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2514 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2515 * whenever sysctl_lowmem_reserve_ratio changes.
2517 * The reserve ratio obviously has absolutely no relation with the
2518 * pages_min watermarks. The lowmem reserve ratio can only make sense
2519 * if in function of the boot time zone sizes.
2521 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
2522 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2524 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2525 setup_per_zone_lowmem_reserve();
2529 __initdata int hashdist = HASHDIST_DEFAULT;
2532 static int __init set_hashdist(char *str)
2536 hashdist = simple_strtoul(str, &str, 0);
2539 __setup("hashdist=", set_hashdist);
2543 * allocate a large system hash table from bootmem
2544 * - it is assumed that the hash table must contain an exact power-of-2
2545 * quantity of entries
2546 * - limit is the number of hash buckets, not the total allocation size
2548 void *__init alloc_large_system_hash(const char *tablename,
2549 unsigned long bucketsize,
2550 unsigned long numentries,
2553 unsigned int *_hash_shift,
2554 unsigned int *_hash_mask,
2555 unsigned long limit)
2557 unsigned long long max = limit;
2558 unsigned long log2qty, size;
2561 /* allow the kernel cmdline to have a say */
2563 /* round applicable memory size up to nearest megabyte */
2564 numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages;
2565 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2566 numentries >>= 20 - PAGE_SHIFT;
2567 numentries <<= 20 - PAGE_SHIFT;
2569 /* limit to 1 bucket per 2^scale bytes of low memory */
2570 if (scale > PAGE_SHIFT)
2571 numentries >>= (scale - PAGE_SHIFT);
2573 numentries <<= (PAGE_SHIFT - scale);
2575 /* rounded up to nearest power of 2 in size */
2576 numentries = 1UL << (long_log2(numentries) + 1);
2578 /* limit allocation size to 1/16 total memory by default */
2580 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2581 do_div(max, bucketsize);
2584 if (numentries > max)
2587 log2qty = long_log2(numentries);
2590 size = bucketsize << log2qty;
2591 if (flags & HASH_EARLY)
2592 table = alloc_bootmem(size);
2594 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
2596 unsigned long order;
2597 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
2599 table = (void*) __get_free_pages(GFP_ATOMIC, order);
2601 } while (!table && size > PAGE_SIZE && --log2qty);
2604 panic("Failed to allocate %s hash table\n", tablename);
2606 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2609 long_log2(size) - PAGE_SHIFT,
2613 *_hash_shift = log2qty;
2615 *_hash_mask = (1 << log2qty) - 1;