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/nodemask.h>
37 #include <linux/vmalloc.h>
39 #include <asm/tlbflush.h>
43 * MCD - HACK: Find somewhere to initialize this EARLY, or make this
46 nodemask_t node_online_map __read_mostly = { { [0] = 1UL } };
47 EXPORT_SYMBOL(node_online_map);
48 nodemask_t node_possible_map __read_mostly = NODE_MASK_ALL;
49 EXPORT_SYMBOL(node_possible_map);
50 struct pglist_data *pgdat_list __read_mostly;
51 unsigned long totalram_pages __read_mostly;
52 unsigned long totalhigh_pages __read_mostly;
56 * results with 256, 32 in the lowmem_reserve sysctl:
57 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
58 * 1G machine -> (16M dma, 784M normal, 224M high)
59 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
60 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
61 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
63 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 256, 32 };
65 EXPORT_SYMBOL(totalram_pages);
66 EXPORT_SYMBOL(nr_swap_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)
83 if (page_to_pfn(page) >= zone->zone_start_pfn + zone->spanned_pages)
85 if (page_to_pfn(page) < zone->zone_start_pfn)
91 static int page_is_consistent(struct zone *zone, struct page *page)
93 #ifdef CONFIG_HOLES_IN_ZONE
94 if (!pfn_valid(page_to_pfn(page)))
97 if (zone != page_zone(page))
103 * Temporary debugging check for pages not lying within a given zone.
105 static int bad_range(struct zone *zone, struct page *page)
107 if (page_outside_zone_boundaries(zone, page))
109 if (!page_is_consistent(zone, page))
115 static void bad_page(const char *function, struct page *page)
117 printk(KERN_EMERG "Bad page state at %s (in process '%s', page %p)\n",
118 function, current->comm, page);
119 printk(KERN_EMERG "flags:0x%0*lx mapping:%p mapcount:%d count:%d\n",
120 (int)(2*sizeof(page_flags_t)), (unsigned long)page->flags,
121 page->mapping, page_mapcount(page), page_count(page));
122 printk(KERN_EMERG "Backtrace:\n");
124 printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n");
125 page->flags &= ~(1 << PG_lru |
135 set_page_count(page, 0);
136 reset_page_mapcount(page);
137 page->mapping = NULL;
138 add_taint(TAINT_BAD_PAGE);
141 #ifndef CONFIG_HUGETLB_PAGE
142 #define prep_compound_page(page, order) do { } while (0)
143 #define destroy_compound_page(page, order) do { } while (0)
146 * Higher-order pages are called "compound pages". They are structured thusly:
148 * The first PAGE_SIZE page is called the "head page".
150 * The remaining PAGE_SIZE pages are called "tail pages".
152 * All pages have PG_compound set. All pages have their ->private pointing at
153 * the head page (even the head page has this).
155 * The first tail page's ->mapping, if non-zero, holds the address of the
156 * compound page's put_page() function.
158 * The order of the allocation is stored in the first tail page's ->index
159 * This is only for debug at present. This usage means that zero-order pages
160 * may not be compound.
162 static void prep_compound_page(struct page *page, unsigned long order)
165 int nr_pages = 1 << order;
167 page[1].mapping = NULL;
168 page[1].index = order;
169 for (i = 0; i < nr_pages; i++) {
170 struct page *p = page + i;
173 set_page_private(p, (unsigned long)page);
177 static void destroy_compound_page(struct page *page, unsigned long order)
180 int nr_pages = 1 << order;
182 if (!PageCompound(page))
185 if (page[1].index != order)
186 bad_page(__FUNCTION__, page);
188 for (i = 0; i < nr_pages; i++) {
189 struct page *p = page + i;
191 if (!PageCompound(p))
192 bad_page(__FUNCTION__, page);
193 if (page_private(p) != (unsigned long)page)
194 bad_page(__FUNCTION__, page);
195 ClearPageCompound(p);
198 #endif /* CONFIG_HUGETLB_PAGE */
201 * function for dealing with page's order in buddy system.
202 * zone->lock is already acquired when we use these.
203 * So, we don't need atomic page->flags operations here.
205 static inline unsigned long page_order(struct page *page) {
206 return page_private(page);
209 static inline void set_page_order(struct page *page, int order) {
210 set_page_private(page, order);
211 __SetPagePrivate(page);
214 static inline void rmv_page_order(struct page *page)
216 __ClearPagePrivate(page);
217 set_page_private(page, 0);
221 * Locate the struct page for both the matching buddy in our
222 * pair (buddy1) and the combined O(n+1) page they form (page).
224 * 1) Any buddy B1 will have an order O twin B2 which satisfies
225 * the following equation:
227 * For example, if the starting buddy (buddy2) is #8 its order
229 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
231 * 2) Any buddy B will have an order O+1 parent P which
232 * satisfies the following equation:
235 * Assumption: *_mem_map is contigious at least up to MAX_ORDER
237 static inline struct page *
238 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
240 unsigned long buddy_idx = page_idx ^ (1 << order);
242 return page + (buddy_idx - page_idx);
245 static inline unsigned long
246 __find_combined_index(unsigned long page_idx, unsigned int order)
248 return (page_idx & ~(1 << order));
252 * This function checks whether a page is free && is the buddy
253 * we can do coalesce a page and its buddy if
254 * (a) the buddy is free &&
255 * (b) the buddy is on the buddy system &&
256 * (c) a page and its buddy have the same order.
257 * for recording page's order, we use page_private(page) and PG_private.
260 static inline int page_is_buddy(struct page *page, int order)
262 if (PagePrivate(page) &&
263 (page_order(page) == order) &&
264 page_count(page) == 0)
270 * Freeing function for a buddy system allocator.
272 * The concept of a buddy system is to maintain direct-mapped table
273 * (containing bit values) for memory blocks of various "orders".
274 * The bottom level table contains the map for the smallest allocatable
275 * units of memory (here, pages), and each level above it describes
276 * pairs of units from the levels below, hence, "buddies".
277 * At a high level, all that happens here is marking the table entry
278 * at the bottom level available, and propagating the changes upward
279 * as necessary, plus some accounting needed to play nicely with other
280 * parts of the VM system.
281 * At each level, we keep a list of pages, which are heads of continuous
282 * free pages of length of (1 << order) and marked with PG_Private.Page's
283 * order is recorded in page_private(page) field.
284 * So when we are allocating or freeing one, we can derive the state of the
285 * other. That is, if we allocate a small block, and both were
286 * free, the remainder of the region must be split into blocks.
287 * If a block is freed, and its buddy is also free, then this
288 * triggers coalescing into a block of larger size.
293 static inline void __free_pages_bulk (struct page *page,
294 struct zone *zone, unsigned int order)
296 unsigned long page_idx;
297 int order_size = 1 << order;
300 destroy_compound_page(page, order);
302 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
304 BUG_ON(page_idx & (order_size - 1));
305 BUG_ON(bad_range(zone, page));
307 zone->free_pages += order_size;
308 while (order < MAX_ORDER-1) {
309 unsigned long combined_idx;
310 struct free_area *area;
313 combined_idx = __find_combined_index(page_idx, order);
314 buddy = __page_find_buddy(page, page_idx, order);
316 if (bad_range(zone, buddy))
318 if (!page_is_buddy(buddy, order))
319 break; /* Move the buddy up one level. */
320 list_del(&buddy->lru);
321 area = zone->free_area + order;
323 rmv_page_order(buddy);
324 page = page + (combined_idx - page_idx);
325 page_idx = combined_idx;
328 set_page_order(page, order);
329 list_add(&page->lru, &zone->free_area[order].free_list);
330 zone->free_area[order].nr_free++;
333 static inline void free_pages_check(const char *function, struct page *page)
335 if ( page_mapcount(page) ||
336 page->mapping != NULL ||
337 page_count(page) != 0 ||
348 bad_page(function, page);
350 __ClearPageDirty(page);
354 * Frees a list of pages.
355 * Assumes all pages on list are in same zone, and of same order.
356 * count is the number of pages to free.
358 * If the zone was previously in an "all pages pinned" state then look to
359 * see if this freeing clears that state.
361 * And clear the zone's pages_scanned counter, to hold off the "all pages are
362 * pinned" detection logic.
365 free_pages_bulk(struct zone *zone, int count,
366 struct list_head *list, unsigned int order)
369 struct page *page = NULL;
372 spin_lock_irqsave(&zone->lock, flags);
373 zone->all_unreclaimable = 0;
374 zone->pages_scanned = 0;
375 while (!list_empty(list) && count--) {
376 page = list_entry(list->prev, struct page, lru);
377 /* have to delete it as __free_pages_bulk list manipulates */
378 list_del(&page->lru);
379 __free_pages_bulk(page, zone, order);
382 spin_unlock_irqrestore(&zone->lock, flags);
386 void __free_pages_ok(struct page *page, unsigned int order)
391 arch_free_page(page, order);
393 mod_page_state(pgfree, 1 << order);
397 for (i = 1 ; i < (1 << order) ; ++i)
398 __put_page(page + i);
401 for (i = 0 ; i < (1 << order) ; ++i)
402 free_pages_check(__FUNCTION__, page + i);
403 list_add(&page->lru, &list);
404 kernel_map_pages(page, 1<<order, 0);
405 free_pages_bulk(page_zone(page), 1, &list, order);
410 * The order of subdivision here is critical for the IO subsystem.
411 * Please do not alter this order without good reasons and regression
412 * testing. Specifically, as large blocks of memory are subdivided,
413 * the order in which smaller blocks are delivered depends on the order
414 * they're subdivided in this function. This is the primary factor
415 * influencing the order in which pages are delivered to the IO
416 * subsystem according to empirical testing, and this is also justified
417 * by considering the behavior of a buddy system containing a single
418 * large block of memory acted on by a series of small allocations.
419 * This behavior is a critical factor in sglist merging's success.
423 static inline struct page *
424 expand(struct zone *zone, struct page *page,
425 int low, int high, struct free_area *area)
427 unsigned long size = 1 << high;
433 BUG_ON(bad_range(zone, &page[size]));
434 list_add(&page[size].lru, &area->free_list);
436 set_page_order(&page[size], high);
441 void set_page_refs(struct page *page, int order)
444 set_page_count(page, 1);
449 * We need to reference all the pages for this order, otherwise if
450 * anyone accesses one of the pages with (get/put) it will be freed.
451 * - eg: access_process_vm()
453 for (i = 0; i < (1 << order); i++)
454 set_page_count(page + i, 1);
455 #endif /* CONFIG_MMU */
459 * This page is about to be returned from the page allocator
461 static void prep_new_page(struct page *page, int order)
463 if ( page_mapcount(page) ||
464 page->mapping != NULL ||
465 page_count(page) != 0 ||
477 bad_page(__FUNCTION__, page);
479 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
480 1 << PG_referenced | 1 << PG_arch_1 |
481 1 << PG_checked | 1 << PG_mappedtodisk);
482 set_page_private(page, 0);
483 set_page_refs(page, order);
484 kernel_map_pages(page, 1 << order, 1);
488 * Do the hard work of removing an element from the buddy allocator.
489 * Call me with the zone->lock already held.
491 static struct page *__rmqueue(struct zone *zone, unsigned int order)
493 struct free_area * area;
494 unsigned int current_order;
497 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
498 area = zone->free_area + current_order;
499 if (list_empty(&area->free_list))
502 page = list_entry(area->free_list.next, struct page, lru);
503 list_del(&page->lru);
504 rmv_page_order(page);
506 zone->free_pages -= 1UL << order;
507 return expand(zone, page, order, current_order, area);
514 * Obtain a specified number of elements from the buddy allocator, all under
515 * a single hold of the lock, for efficiency. Add them to the supplied list.
516 * Returns the number of new pages which were placed at *list.
518 static int rmqueue_bulk(struct zone *zone, unsigned int order,
519 unsigned long count, struct list_head *list)
526 spin_lock_irqsave(&zone->lock, flags);
527 for (i = 0; i < count; ++i) {
528 page = __rmqueue(zone, order);
532 list_add_tail(&page->lru, list);
534 spin_unlock_irqrestore(&zone->lock, flags);
539 /* Called from the slab reaper to drain remote pagesets */
540 void drain_remote_pages(void)
546 local_irq_save(flags);
547 for_each_zone(zone) {
548 struct per_cpu_pageset *pset;
550 /* Do not drain local pagesets */
551 if (zone->zone_pgdat->node_id == numa_node_id())
554 pset = zone->pageset[smp_processor_id()];
555 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
556 struct per_cpu_pages *pcp;
560 pcp->count -= free_pages_bulk(zone, pcp->count,
564 local_irq_restore(flags);
568 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
569 static void __drain_pages(unsigned int cpu)
574 for_each_zone(zone) {
575 struct per_cpu_pageset *pset;
577 pset = zone_pcp(zone, cpu);
578 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
579 struct per_cpu_pages *pcp;
582 pcp->count -= free_pages_bulk(zone, pcp->count,
587 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
591 void mark_free_pages(struct zone *zone)
593 unsigned long zone_pfn, flags;
595 struct list_head *curr;
597 if (!zone->spanned_pages)
600 spin_lock_irqsave(&zone->lock, flags);
601 for (zone_pfn = 0; zone_pfn < zone->spanned_pages; ++zone_pfn)
602 ClearPageNosaveFree(pfn_to_page(zone_pfn + zone->zone_start_pfn));
604 for (order = MAX_ORDER - 1; order >= 0; --order)
605 list_for_each(curr, &zone->free_area[order].free_list) {
606 unsigned long start_pfn, i;
608 start_pfn = page_to_pfn(list_entry(curr, struct page, lru));
610 for (i=0; i < (1<<order); i++)
611 SetPageNosaveFree(pfn_to_page(start_pfn+i));
613 spin_unlock_irqrestore(&zone->lock, flags);
617 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
619 void drain_local_pages(void)
623 local_irq_save(flags);
624 __drain_pages(smp_processor_id());
625 local_irq_restore(flags);
627 #endif /* CONFIG_PM */
629 static void zone_statistics(struct zonelist *zonelist, struct zone *z)
634 pg_data_t *pg = z->zone_pgdat;
635 pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
636 struct per_cpu_pageset *p;
638 local_irq_save(flags);
639 cpu = smp_processor_id();
645 zone_pcp(zonelist->zones[0], cpu)->numa_foreign++;
647 if (pg == NODE_DATA(numa_node_id()))
651 local_irq_restore(flags);
656 * Free a 0-order page
658 static void FASTCALL(free_hot_cold_page(struct page *page, int cold));
659 static void fastcall free_hot_cold_page(struct page *page, int cold)
661 struct zone *zone = page_zone(page);
662 struct per_cpu_pages *pcp;
665 arch_free_page(page, 0);
667 kernel_map_pages(page, 1, 0);
668 inc_page_state(pgfree);
670 page->mapping = NULL;
671 free_pages_check(__FUNCTION__, page);
672 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
673 local_irq_save(flags);
674 list_add(&page->lru, &pcp->list);
676 if (pcp->count >= pcp->high)
677 pcp->count -= free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
678 local_irq_restore(flags);
682 void fastcall free_hot_page(struct page *page)
684 free_hot_cold_page(page, 0);
687 void fastcall free_cold_page(struct page *page)
689 free_hot_cold_page(page, 1);
692 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
696 BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
697 for(i = 0; i < (1 << order); i++)
698 clear_highpage(page + i);
702 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
703 * we cheat by calling it from here, in the order > 0 path. Saves a branch
707 buffered_rmqueue(struct zone *zone, int order, gfp_t gfp_flags)
710 struct page *page = NULL;
711 int cold = !!(gfp_flags & __GFP_COLD);
714 struct per_cpu_pages *pcp;
716 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
717 local_irq_save(flags);
718 if (pcp->count <= pcp->low)
719 pcp->count += rmqueue_bulk(zone, 0,
720 pcp->batch, &pcp->list);
722 page = list_entry(pcp->list.next, struct page, lru);
723 list_del(&page->lru);
726 local_irq_restore(flags);
731 spin_lock_irqsave(&zone->lock, flags);
732 page = __rmqueue(zone, order);
733 spin_unlock_irqrestore(&zone->lock, flags);
737 BUG_ON(bad_range(zone, page));
738 mod_page_state_zone(zone, pgalloc, 1 << order);
739 prep_new_page(page, order);
741 if (gfp_flags & __GFP_ZERO)
742 prep_zero_page(page, order, gfp_flags);
744 if (order && (gfp_flags & __GFP_COMP))
745 prep_compound_page(page, order);
751 * Return 1 if free pages are above 'mark'. This takes into account the order
754 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
755 int classzone_idx, int can_try_harder, gfp_t gfp_high)
757 /* free_pages my go negative - that's OK */
758 long min = mark, free_pages = z->free_pages - (1 << order) + 1;
766 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
768 for (o = 0; o < order; o++) {
769 /* At the next order, this order's pages become unavailable */
770 free_pages -= z->free_area[o].nr_free << o;
772 /* Require fewer higher order pages to be free */
775 if (free_pages <= min)
782 should_reclaim_zone(struct zone *z, gfp_t gfp_mask)
784 if (!z->reclaim_pages)
786 if (gfp_mask & __GFP_NORECLAIM)
792 * This is the 'heart' of the zoned buddy allocator.
794 struct page * fastcall
795 __alloc_pages(gfp_t gfp_mask, unsigned int order,
796 struct zonelist *zonelist)
798 const gfp_t wait = gfp_mask & __GFP_WAIT;
799 struct zone **zones, *z;
801 struct reclaim_state reclaim_state;
802 struct task_struct *p = current;
807 int did_some_progress;
809 might_sleep_if(wait);
812 * The caller may dip into page reserves a bit more if the caller
813 * cannot run direct reclaim, or is the caller has realtime scheduling
816 can_try_harder = (unlikely(rt_task(p)) && !in_interrupt()) || !wait;
818 zones = zonelist->zones; /* the list of zones suitable for gfp_mask */
820 if (unlikely(zones[0] == NULL)) {
821 /* Should this ever happen?? */
825 classzone_idx = zone_idx(zones[0]);
829 * Go through the zonelist once, looking for a zone with enough free.
830 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
832 for (i = 0; (z = zones[i]) != NULL; i++) {
833 int do_reclaim = should_reclaim_zone(z, gfp_mask);
835 if (!cpuset_zone_allowed(z, __GFP_HARDWALL))
839 * If the zone is to attempt early page reclaim then this loop
840 * will try to reclaim pages and check the watermark a second
841 * time before giving up and falling back to the next zone.
844 if (!zone_watermark_ok(z, order, z->pages_low,
845 classzone_idx, 0, 0)) {
849 zone_reclaim(z, gfp_mask, order);
850 /* Only try reclaim once */
852 goto zone_reclaim_retry;
856 page = buffered_rmqueue(z, order, gfp_mask);
861 for (i = 0; (z = zones[i]) != NULL; i++)
862 wakeup_kswapd(z, order);
865 * Go through the zonelist again. Let __GFP_HIGH and allocations
866 * coming from realtime tasks to go deeper into reserves
868 * This is the last chance, in general, before the goto nopage.
869 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
870 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
872 for (i = 0; (z = zones[i]) != NULL; i++) {
873 if (!zone_watermark_ok(z, order, z->pages_min,
874 classzone_idx, can_try_harder,
875 gfp_mask & __GFP_HIGH))
878 if (wait && !cpuset_zone_allowed(z, gfp_mask))
881 page = buffered_rmqueue(z, order, gfp_mask);
886 /* This allocation should allow future memory freeing. */
888 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
889 && !in_interrupt()) {
890 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
891 /* go through the zonelist yet again, ignoring mins */
892 for (i = 0; (z = zones[i]) != NULL; i++) {
893 if (!cpuset_zone_allowed(z, gfp_mask))
895 page = buffered_rmqueue(z, order, gfp_mask);
903 /* Atomic allocations - we can't balance anything */
910 /* We now go into synchronous reclaim */
911 p->flags |= PF_MEMALLOC;
912 reclaim_state.reclaimed_slab = 0;
913 p->reclaim_state = &reclaim_state;
915 did_some_progress = try_to_free_pages(zones, gfp_mask);
917 p->reclaim_state = NULL;
918 p->flags &= ~PF_MEMALLOC;
922 if (likely(did_some_progress)) {
923 for (i = 0; (z = zones[i]) != NULL; i++) {
924 if (!zone_watermark_ok(z, order, z->pages_min,
925 classzone_idx, can_try_harder,
926 gfp_mask & __GFP_HIGH))
929 if (!cpuset_zone_allowed(z, gfp_mask))
932 page = buffered_rmqueue(z, order, gfp_mask);
936 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
938 * Go through the zonelist yet one more time, keep
939 * very high watermark here, this is only to catch
940 * a parallel oom killing, we must fail if we're still
941 * under heavy pressure.
943 for (i = 0; (z = zones[i]) != NULL; i++) {
944 if (!zone_watermark_ok(z, order, z->pages_high,
945 classzone_idx, 0, 0))
948 if (!cpuset_zone_allowed(z, __GFP_HARDWALL))
951 page = buffered_rmqueue(z, order, gfp_mask);
956 out_of_memory(gfp_mask, order);
961 * Don't let big-order allocations loop unless the caller explicitly
962 * requests that. Wait for some write requests to complete then retry.
964 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
965 * <= 3, but that may not be true in other implementations.
968 if (!(gfp_mask & __GFP_NORETRY)) {
969 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
971 if (gfp_mask & __GFP_NOFAIL)
975 blk_congestion_wait(WRITE, HZ/50);
980 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
981 printk(KERN_WARNING "%s: page allocation failure."
982 " order:%d, mode:0x%x\n",
983 p->comm, order, gfp_mask);
989 zone_statistics(zonelist, z);
993 EXPORT_SYMBOL(__alloc_pages);
996 * Common helper functions.
998 fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1001 page = alloc_pages(gfp_mask, order);
1004 return (unsigned long) page_address(page);
1007 EXPORT_SYMBOL(__get_free_pages);
1009 fastcall unsigned long get_zeroed_page(gfp_t gfp_mask)
1014 * get_zeroed_page() returns a 32-bit address, which cannot represent
1017 BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1019 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1021 return (unsigned long) page_address(page);
1025 EXPORT_SYMBOL(get_zeroed_page);
1027 void __pagevec_free(struct pagevec *pvec)
1029 int i = pagevec_count(pvec);
1032 free_hot_cold_page(pvec->pages[i], pvec->cold);
1035 fastcall void __free_pages(struct page *page, unsigned int order)
1037 if (put_page_testzero(page)) {
1039 free_hot_page(page);
1041 __free_pages_ok(page, order);
1045 EXPORT_SYMBOL(__free_pages);
1047 fastcall void free_pages(unsigned long addr, unsigned int order)
1050 BUG_ON(!virt_addr_valid((void *)addr));
1051 __free_pages(virt_to_page((void *)addr), order);
1055 EXPORT_SYMBOL(free_pages);
1058 * Total amount of free (allocatable) RAM:
1060 unsigned int nr_free_pages(void)
1062 unsigned int sum = 0;
1066 sum += zone->free_pages;
1071 EXPORT_SYMBOL(nr_free_pages);
1074 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
1076 unsigned int i, sum = 0;
1078 for (i = 0; i < MAX_NR_ZONES; i++)
1079 sum += pgdat->node_zones[i].free_pages;
1085 static unsigned int nr_free_zone_pages(int offset)
1087 /* Just pick one node, since fallback list is circular */
1088 pg_data_t *pgdat = NODE_DATA(numa_node_id());
1089 unsigned int sum = 0;
1091 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1092 struct zone **zonep = zonelist->zones;
1095 for (zone = *zonep++; zone; zone = *zonep++) {
1096 unsigned long size = zone->present_pages;
1097 unsigned long high = zone->pages_high;
1106 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1108 unsigned int nr_free_buffer_pages(void)
1110 return nr_free_zone_pages(gfp_zone(GFP_USER));
1114 * Amount of free RAM allocatable within all zones
1116 unsigned int nr_free_pagecache_pages(void)
1118 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER));
1121 #ifdef CONFIG_HIGHMEM
1122 unsigned int nr_free_highpages (void)
1125 unsigned int pages = 0;
1127 for_each_pgdat(pgdat)
1128 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1135 static void show_node(struct zone *zone)
1137 printk("Node %d ", zone->zone_pgdat->node_id);
1140 #define show_node(zone) do { } while (0)
1144 * Accumulate the page_state information across all CPUs.
1145 * The result is unavoidably approximate - it can change
1146 * during and after execution of this function.
1148 static DEFINE_PER_CPU(struct page_state, page_states) = {0};
1150 atomic_t nr_pagecache = ATOMIC_INIT(0);
1151 EXPORT_SYMBOL(nr_pagecache);
1153 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
1156 void __get_page_state(struct page_state *ret, int nr, cpumask_t *cpumask)
1160 memset(ret, 0, sizeof(*ret));
1161 cpus_and(*cpumask, *cpumask, cpu_online_map);
1163 cpu = first_cpu(*cpumask);
1164 while (cpu < NR_CPUS) {
1165 unsigned long *in, *out, off;
1167 in = (unsigned long *)&per_cpu(page_states, cpu);
1169 cpu = next_cpu(cpu, *cpumask);
1172 prefetch(&per_cpu(page_states, cpu));
1174 out = (unsigned long *)ret;
1175 for (off = 0; off < nr; off++)
1180 void get_page_state_node(struct page_state *ret, int node)
1183 cpumask_t mask = node_to_cpumask(node);
1185 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
1186 nr /= sizeof(unsigned long);
1188 __get_page_state(ret, nr+1, &mask);
1191 void get_page_state(struct page_state *ret)
1194 cpumask_t mask = CPU_MASK_ALL;
1196 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
1197 nr /= sizeof(unsigned long);
1199 __get_page_state(ret, nr + 1, &mask);
1202 void get_full_page_state(struct page_state *ret)
1204 cpumask_t mask = CPU_MASK_ALL;
1206 __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long), &mask);
1209 unsigned long __read_page_state(unsigned long offset)
1211 unsigned long ret = 0;
1214 for_each_online_cpu(cpu) {
1217 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
1218 ret += *((unsigned long *)in);
1223 void __mod_page_state(unsigned long offset, unsigned long delta)
1225 unsigned long flags;
1228 local_irq_save(flags);
1229 ptr = &__get_cpu_var(page_states);
1230 *(unsigned long*)(ptr + offset) += delta;
1231 local_irq_restore(flags);
1234 EXPORT_SYMBOL(__mod_page_state);
1236 void __get_zone_counts(unsigned long *active, unsigned long *inactive,
1237 unsigned long *free, struct pglist_data *pgdat)
1239 struct zone *zones = pgdat->node_zones;
1245 for (i = 0; i < MAX_NR_ZONES; i++) {
1246 *active += zones[i].nr_active;
1247 *inactive += zones[i].nr_inactive;
1248 *free += zones[i].free_pages;
1252 void get_zone_counts(unsigned long *active,
1253 unsigned long *inactive, unsigned long *free)
1255 struct pglist_data *pgdat;
1260 for_each_pgdat(pgdat) {
1261 unsigned long l, m, n;
1262 __get_zone_counts(&l, &m, &n, pgdat);
1269 void si_meminfo(struct sysinfo *val)
1271 val->totalram = totalram_pages;
1273 val->freeram = nr_free_pages();
1274 val->bufferram = nr_blockdev_pages();
1275 #ifdef CONFIG_HIGHMEM
1276 val->totalhigh = totalhigh_pages;
1277 val->freehigh = nr_free_highpages();
1282 val->mem_unit = PAGE_SIZE;
1285 EXPORT_SYMBOL(si_meminfo);
1288 void si_meminfo_node(struct sysinfo *val, int nid)
1290 pg_data_t *pgdat = NODE_DATA(nid);
1292 val->totalram = pgdat->node_present_pages;
1293 val->freeram = nr_free_pages_pgdat(pgdat);
1294 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1295 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1296 val->mem_unit = PAGE_SIZE;
1300 #define K(x) ((x) << (PAGE_SHIFT-10))
1303 * Show free area list (used inside shift_scroll-lock stuff)
1304 * We also calculate the percentage fragmentation. We do this by counting the
1305 * memory on each free list with the exception of the first item on the list.
1307 void show_free_areas(void)
1309 struct page_state ps;
1310 int cpu, temperature;
1311 unsigned long active;
1312 unsigned long inactive;
1316 for_each_zone(zone) {
1318 printk("%s per-cpu:", zone->name);
1320 if (!zone->present_pages) {
1326 for (cpu = 0; cpu < NR_CPUS; ++cpu) {
1327 struct per_cpu_pageset *pageset;
1329 if (!cpu_possible(cpu))
1332 pageset = zone_pcp(zone, cpu);
1334 for (temperature = 0; temperature < 2; temperature++)
1335 printk("cpu %d %s: low %d, high %d, batch %d used:%d\n",
1337 temperature ? "cold" : "hot",
1338 pageset->pcp[temperature].low,
1339 pageset->pcp[temperature].high,
1340 pageset->pcp[temperature].batch,
1341 pageset->pcp[temperature].count);
1345 get_page_state(&ps);
1346 get_zone_counts(&active, &inactive, &free);
1348 printk("Free pages: %11ukB (%ukB HighMem)\n",
1350 K(nr_free_highpages()));
1352 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1353 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1362 ps.nr_page_table_pages);
1364 for_each_zone(zone) {
1376 " pages_scanned:%lu"
1377 " all_unreclaimable? %s"
1380 K(zone->free_pages),
1383 K(zone->pages_high),
1385 K(zone->nr_inactive),
1386 K(zone->present_pages),
1387 zone->pages_scanned,
1388 (zone->all_unreclaimable ? "yes" : "no")
1390 printk("lowmem_reserve[]:");
1391 for (i = 0; i < MAX_NR_ZONES; i++)
1392 printk(" %lu", zone->lowmem_reserve[i]);
1396 for_each_zone(zone) {
1397 unsigned long nr, flags, order, total = 0;
1400 printk("%s: ", zone->name);
1401 if (!zone->present_pages) {
1406 spin_lock_irqsave(&zone->lock, flags);
1407 for (order = 0; order < MAX_ORDER; order++) {
1408 nr = zone->free_area[order].nr_free;
1409 total += nr << order;
1410 printk("%lu*%lukB ", nr, K(1UL) << order);
1412 spin_unlock_irqrestore(&zone->lock, flags);
1413 printk("= %lukB\n", K(total));
1416 show_swap_cache_info();
1420 * Builds allocation fallback zone lists.
1422 static int __init build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, int j, int k)
1429 zone = pgdat->node_zones + ZONE_HIGHMEM;
1430 if (zone->present_pages) {
1431 #ifndef CONFIG_HIGHMEM
1434 zonelist->zones[j++] = zone;
1437 zone = pgdat->node_zones + ZONE_NORMAL;
1438 if (zone->present_pages)
1439 zonelist->zones[j++] = zone;
1441 zone = pgdat->node_zones + ZONE_DMA;
1442 if (zone->present_pages)
1443 zonelist->zones[j++] = zone;
1449 static inline int highest_zone(int zone_bits)
1451 int res = ZONE_NORMAL;
1452 if (zone_bits & (__force int)__GFP_HIGHMEM)
1454 if (zone_bits & (__force int)__GFP_DMA)
1460 #define MAX_NODE_LOAD (num_online_nodes())
1461 static int __initdata node_load[MAX_NUMNODES];
1463 * find_next_best_node - find the next node that should appear in a given node's fallback list
1464 * @node: node whose fallback list we're appending
1465 * @used_node_mask: nodemask_t of already used nodes
1467 * We use a number of factors to determine which is the next node that should
1468 * appear on a given node's fallback list. The node should not have appeared
1469 * already in @node's fallback list, and it should be the next closest node
1470 * according to the distance array (which contains arbitrary distance values
1471 * from each node to each node in the system), and should also prefer nodes
1472 * with no CPUs, since presumably they'll have very little allocation pressure
1473 * on them otherwise.
1474 * It returns -1 if no node is found.
1476 static int __init find_next_best_node(int node, nodemask_t *used_node_mask)
1479 int min_val = INT_MAX;
1482 for_each_online_node(i) {
1485 /* Start from local node */
1486 n = (node+i) % num_online_nodes();
1488 /* Don't want a node to appear more than once */
1489 if (node_isset(n, *used_node_mask))
1492 /* Use the local node if we haven't already */
1493 if (!node_isset(node, *used_node_mask)) {
1498 /* Use the distance array to find the distance */
1499 val = node_distance(node, n);
1501 /* Give preference to headless and unused nodes */
1502 tmp = node_to_cpumask(n);
1503 if (!cpus_empty(tmp))
1504 val += PENALTY_FOR_NODE_WITH_CPUS;
1506 /* Slight preference for less loaded node */
1507 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1508 val += node_load[n];
1510 if (val < min_val) {
1517 node_set(best_node, *used_node_mask);
1522 static void __init build_zonelists(pg_data_t *pgdat)
1524 int i, j, k, node, local_node;
1525 int prev_node, load;
1526 struct zonelist *zonelist;
1527 nodemask_t used_mask;
1529 /* initialize zonelists */
1530 for (i = 0; i < GFP_ZONETYPES; i++) {
1531 zonelist = pgdat->node_zonelists + i;
1532 zonelist->zones[0] = NULL;
1535 /* NUMA-aware ordering of nodes */
1536 local_node = pgdat->node_id;
1537 load = num_online_nodes();
1538 prev_node = local_node;
1539 nodes_clear(used_mask);
1540 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1542 * We don't want to pressure a particular node.
1543 * So adding penalty to the first node in same
1544 * distance group to make it round-robin.
1546 if (node_distance(local_node, node) !=
1547 node_distance(local_node, prev_node))
1548 node_load[node] += load;
1551 for (i = 0; i < GFP_ZONETYPES; i++) {
1552 zonelist = pgdat->node_zonelists + i;
1553 for (j = 0; zonelist->zones[j] != NULL; j++);
1555 k = highest_zone(i);
1557 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1558 zonelist->zones[j] = NULL;
1563 #else /* CONFIG_NUMA */
1565 static void __init build_zonelists(pg_data_t *pgdat)
1567 int i, j, k, node, local_node;
1569 local_node = pgdat->node_id;
1570 for (i = 0; i < GFP_ZONETYPES; i++) {
1571 struct zonelist *zonelist;
1573 zonelist = pgdat->node_zonelists + i;
1576 k = highest_zone(i);
1577 j = build_zonelists_node(pgdat, zonelist, j, k);
1579 * Now we build the zonelist so that it contains the zones
1580 * of all the other nodes.
1581 * We don't want to pressure a particular node, so when
1582 * building the zones for node N, we make sure that the
1583 * zones coming right after the local ones are those from
1584 * node N+1 (modulo N)
1586 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1587 if (!node_online(node))
1589 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1591 for (node = 0; node < local_node; node++) {
1592 if (!node_online(node))
1594 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1597 zonelist->zones[j] = NULL;
1601 #endif /* CONFIG_NUMA */
1603 void __init build_all_zonelists(void)
1607 for_each_online_node(i)
1608 build_zonelists(NODE_DATA(i));
1609 printk("Built %i zonelists\n", num_online_nodes());
1610 cpuset_init_current_mems_allowed();
1614 * Helper functions to size the waitqueue hash table.
1615 * Essentially these want to choose hash table sizes sufficiently
1616 * large so that collisions trying to wait on pages are rare.
1617 * But in fact, the number of active page waitqueues on typical
1618 * systems is ridiculously low, less than 200. So this is even
1619 * conservative, even though it seems large.
1621 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1622 * waitqueues, i.e. the size of the waitq table given the number of pages.
1624 #define PAGES_PER_WAITQUEUE 256
1626 static inline unsigned long wait_table_size(unsigned long pages)
1628 unsigned long size = 1;
1630 pages /= PAGES_PER_WAITQUEUE;
1632 while (size < pages)
1636 * Once we have dozens or even hundreds of threads sleeping
1637 * on IO we've got bigger problems than wait queue collision.
1638 * Limit the size of the wait table to a reasonable size.
1640 size = min(size, 4096UL);
1642 return max(size, 4UL);
1646 * This is an integer logarithm so that shifts can be used later
1647 * to extract the more random high bits from the multiplicative
1648 * hash function before the remainder is taken.
1650 static inline unsigned long wait_table_bits(unsigned long size)
1655 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1657 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1658 unsigned long *zones_size, unsigned long *zholes_size)
1660 unsigned long realtotalpages, totalpages = 0;
1663 for (i = 0; i < MAX_NR_ZONES; i++)
1664 totalpages += zones_size[i];
1665 pgdat->node_spanned_pages = totalpages;
1667 realtotalpages = totalpages;
1669 for (i = 0; i < MAX_NR_ZONES; i++)
1670 realtotalpages -= zholes_size[i];
1671 pgdat->node_present_pages = realtotalpages;
1672 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1677 * Initially all pages are reserved - free ones are freed
1678 * up by free_all_bootmem() once the early boot process is
1679 * done. Non-atomic initialization, single-pass.
1681 void __init memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1682 unsigned long start_pfn)
1685 unsigned long end_pfn = start_pfn + size;
1688 for (pfn = start_pfn; pfn < end_pfn; pfn++, page++) {
1689 if (!early_pfn_valid(pfn))
1691 if (!early_pfn_in_nid(pfn, nid))
1693 page = pfn_to_page(pfn);
1694 set_page_links(page, zone, nid, pfn);
1695 set_page_count(page, 1);
1696 reset_page_mapcount(page);
1697 SetPageReserved(page);
1698 INIT_LIST_HEAD(&page->lru);
1699 #ifdef WANT_PAGE_VIRTUAL
1700 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1701 if (!is_highmem_idx(zone))
1702 set_page_address(page, __va(pfn << PAGE_SHIFT));
1707 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1711 for (order = 0; order < MAX_ORDER ; order++) {
1712 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1713 zone->free_area[order].nr_free = 0;
1717 #define ZONETABLE_INDEX(x, zone_nr) ((x << ZONES_SHIFT) | zone_nr)
1718 void zonetable_add(struct zone *zone, int nid, int zid, unsigned long pfn,
1721 unsigned long snum = pfn_to_section_nr(pfn);
1722 unsigned long end = pfn_to_section_nr(pfn + size);
1725 zone_table[ZONETABLE_INDEX(nid, zid)] = zone;
1727 for (; snum <= end; snum++)
1728 zone_table[ZONETABLE_INDEX(snum, zid)] = zone;
1731 #ifndef __HAVE_ARCH_MEMMAP_INIT
1732 #define memmap_init(size, nid, zone, start_pfn) \
1733 memmap_init_zone((size), (nid), (zone), (start_pfn))
1736 static int __devinit zone_batchsize(struct zone *zone)
1741 * The per-cpu-pages pools are set to around 1000th of the
1742 * size of the zone. But no more than 1/2 of a meg.
1744 * OK, so we don't know how big the cache is. So guess.
1746 batch = zone->present_pages / 1024;
1747 if (batch * PAGE_SIZE > 512 * 1024)
1748 batch = (512 * 1024) / PAGE_SIZE;
1749 batch /= 4; /* We effectively *= 4 below */
1754 * We will be trying to allcoate bigger chunks of contiguous
1755 * memory of the order of fls(batch). This should result in
1756 * better cache coloring.
1758 * A sanity check also to ensure that batch is still in limits.
1760 batch = (1 << fls(batch + batch/2));
1762 if (fls(batch) >= (PAGE_SHIFT + MAX_ORDER - 2))
1763 batch = PAGE_SHIFT + ((MAX_ORDER - 1 - PAGE_SHIFT)/2);
1768 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
1770 struct per_cpu_pages *pcp;
1772 memset(p, 0, sizeof(*p));
1774 pcp = &p->pcp[0]; /* hot */
1777 pcp->high = 6 * batch;
1778 pcp->batch = max(1UL, 1 * batch);
1779 INIT_LIST_HEAD(&pcp->list);
1781 pcp = &p->pcp[1]; /* cold*/
1784 pcp->high = 2 * batch;
1785 pcp->batch = max(1UL, batch/2);
1786 INIT_LIST_HEAD(&pcp->list);
1791 * Boot pageset table. One per cpu which is going to be used for all
1792 * zones and all nodes. The parameters will be set in such a way
1793 * that an item put on a list will immediately be handed over to
1794 * the buddy list. This is safe since pageset manipulation is done
1795 * with interrupts disabled.
1797 * Some NUMA counter updates may also be caught by the boot pagesets.
1799 * The boot_pagesets must be kept even after bootup is complete for
1800 * unused processors and/or zones. They do play a role for bootstrapping
1801 * hotplugged processors.
1803 * zoneinfo_show() and maybe other functions do
1804 * not check if the processor is online before following the pageset pointer.
1805 * Other parts of the kernel may not check if the zone is available.
1807 static struct per_cpu_pageset
1808 boot_pageset[NR_CPUS];
1811 * Dynamically allocate memory for the
1812 * per cpu pageset array in struct zone.
1814 static int __devinit process_zones(int cpu)
1816 struct zone *zone, *dzone;
1818 for_each_zone(zone) {
1820 zone->pageset[cpu] = kmalloc_node(sizeof(struct per_cpu_pageset),
1821 GFP_KERNEL, cpu_to_node(cpu));
1822 if (!zone->pageset[cpu])
1825 setup_pageset(zone->pageset[cpu], zone_batchsize(zone));
1830 for_each_zone(dzone) {
1833 kfree(dzone->pageset[cpu]);
1834 dzone->pageset[cpu] = NULL;
1839 static inline void free_zone_pagesets(int cpu)
1844 for_each_zone(zone) {
1845 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
1847 zone_pcp(zone, cpu) = NULL;
1853 static int __devinit pageset_cpuup_callback(struct notifier_block *nfb,
1854 unsigned long action,
1857 int cpu = (long)hcpu;
1858 int ret = NOTIFY_OK;
1861 case CPU_UP_PREPARE:
1862 if (process_zones(cpu))
1865 #ifdef CONFIG_HOTPLUG_CPU
1867 free_zone_pagesets(cpu);
1876 static struct notifier_block pageset_notifier =
1877 { &pageset_cpuup_callback, NULL, 0 };
1879 void __init setup_per_cpu_pageset()
1883 /* Initialize per_cpu_pageset for cpu 0.
1884 * A cpuup callback will do this for every cpu
1885 * as it comes online
1887 err = process_zones(smp_processor_id());
1889 register_cpu_notifier(&pageset_notifier);
1895 void zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
1898 struct pglist_data *pgdat = zone->zone_pgdat;
1901 * The per-page waitqueue mechanism uses hashed waitqueues
1904 zone->wait_table_size = wait_table_size(zone_size_pages);
1905 zone->wait_table_bits = wait_table_bits(zone->wait_table_size);
1906 zone->wait_table = (wait_queue_head_t *)
1907 alloc_bootmem_node(pgdat, zone->wait_table_size
1908 * sizeof(wait_queue_head_t));
1910 for(i = 0; i < zone->wait_table_size; ++i)
1911 init_waitqueue_head(zone->wait_table + i);
1914 static __devinit void zone_pcp_init(struct zone *zone)
1917 unsigned long batch = zone_batchsize(zone);
1919 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1921 /* Early boot. Slab allocator not functional yet */
1922 zone->pageset[cpu] = &boot_pageset[cpu];
1923 setup_pageset(&boot_pageset[cpu],0);
1925 setup_pageset(zone_pcp(zone,cpu), batch);
1928 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
1929 zone->name, zone->present_pages, batch);
1932 static __devinit void init_currently_empty_zone(struct zone *zone,
1933 unsigned long zone_start_pfn, unsigned long size)
1935 struct pglist_data *pgdat = zone->zone_pgdat;
1937 zone_wait_table_init(zone, size);
1938 pgdat->nr_zones = zone_idx(zone) + 1;
1940 zone->zone_mem_map = pfn_to_page(zone_start_pfn);
1941 zone->zone_start_pfn = zone_start_pfn;
1943 memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn);
1945 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1949 * Set up the zone data structures:
1950 * - mark all pages reserved
1951 * - mark all memory queues empty
1952 * - clear the memory bitmaps
1954 static void __init free_area_init_core(struct pglist_data *pgdat,
1955 unsigned long *zones_size, unsigned long *zholes_size)
1958 int nid = pgdat->node_id;
1959 unsigned long zone_start_pfn = pgdat->node_start_pfn;
1961 pgdat->nr_zones = 0;
1962 init_waitqueue_head(&pgdat->kswapd_wait);
1963 pgdat->kswapd_max_order = 0;
1965 for (j = 0; j < MAX_NR_ZONES; j++) {
1966 struct zone *zone = pgdat->node_zones + j;
1967 unsigned long size, realsize;
1969 realsize = size = zones_size[j];
1971 realsize -= zholes_size[j];
1973 if (j == ZONE_DMA || j == ZONE_NORMAL)
1974 nr_kernel_pages += realsize;
1975 nr_all_pages += realsize;
1977 zone->spanned_pages = size;
1978 zone->present_pages = realsize;
1979 zone->name = zone_names[j];
1980 spin_lock_init(&zone->lock);
1981 spin_lock_init(&zone->lru_lock);
1982 zone->zone_pgdat = pgdat;
1983 zone->free_pages = 0;
1985 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
1987 zone_pcp_init(zone);
1988 INIT_LIST_HEAD(&zone->active_list);
1989 INIT_LIST_HEAD(&zone->inactive_list);
1990 zone->nr_scan_active = 0;
1991 zone->nr_scan_inactive = 0;
1992 zone->nr_active = 0;
1993 zone->nr_inactive = 0;
1994 atomic_set(&zone->reclaim_in_progress, 0);
1998 zonetable_add(zone, nid, j, zone_start_pfn, size);
1999 init_currently_empty_zone(zone, zone_start_pfn, size);
2000 zone_start_pfn += size;
2004 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
2006 /* Skip empty nodes */
2007 if (!pgdat->node_spanned_pages)
2010 #ifdef CONFIG_FLAT_NODE_MEM_MAP
2011 /* ia64 gets its own node_mem_map, before this, without bootmem */
2012 if (!pgdat->node_mem_map) {
2016 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
2017 map = alloc_remap(pgdat->node_id, size);
2019 map = alloc_bootmem_node(pgdat, size);
2020 pgdat->node_mem_map = map;
2022 #ifdef CONFIG_FLATMEM
2024 * With no DISCONTIG, the global mem_map is just set as node 0's
2026 if (pgdat == NODE_DATA(0))
2027 mem_map = NODE_DATA(0)->node_mem_map;
2029 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
2032 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
2033 unsigned long *zones_size, unsigned long node_start_pfn,
2034 unsigned long *zholes_size)
2036 pgdat->node_id = nid;
2037 pgdat->node_start_pfn = node_start_pfn;
2038 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
2040 alloc_node_mem_map(pgdat);
2042 free_area_init_core(pgdat, zones_size, zholes_size);
2045 #ifndef CONFIG_NEED_MULTIPLE_NODES
2046 static bootmem_data_t contig_bootmem_data;
2047 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
2049 EXPORT_SYMBOL(contig_page_data);
2052 void __init free_area_init(unsigned long *zones_size)
2054 free_area_init_node(0, NODE_DATA(0), zones_size,
2055 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
2058 #ifdef CONFIG_PROC_FS
2060 #include <linux/seq_file.h>
2062 static void *frag_start(struct seq_file *m, loff_t *pos)
2067 for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
2073 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
2075 pg_data_t *pgdat = (pg_data_t *)arg;
2078 return pgdat->pgdat_next;
2081 static void frag_stop(struct seq_file *m, void *arg)
2086 * This walks the free areas for each zone.
2088 static int frag_show(struct seq_file *m, void *arg)
2090 pg_data_t *pgdat = (pg_data_t *)arg;
2092 struct zone *node_zones = pgdat->node_zones;
2093 unsigned long flags;
2096 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2097 if (!zone->present_pages)
2100 spin_lock_irqsave(&zone->lock, flags);
2101 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
2102 for (order = 0; order < MAX_ORDER; ++order)
2103 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
2104 spin_unlock_irqrestore(&zone->lock, flags);
2110 struct seq_operations fragmentation_op = {
2111 .start = frag_start,
2118 * Output information about zones in @pgdat.
2120 static int zoneinfo_show(struct seq_file *m, void *arg)
2122 pg_data_t *pgdat = arg;
2124 struct zone *node_zones = pgdat->node_zones;
2125 unsigned long flags;
2127 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; zone++) {
2130 if (!zone->present_pages)
2133 spin_lock_irqsave(&zone->lock, flags);
2134 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
2142 "\n scanned %lu (a: %lu i: %lu)"
2151 zone->pages_scanned,
2152 zone->nr_scan_active, zone->nr_scan_inactive,
2153 zone->spanned_pages,
2154 zone->present_pages);
2156 "\n protection: (%lu",
2157 zone->lowmem_reserve[0]);
2158 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
2159 seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
2163 for (i = 0; i < ARRAY_SIZE(zone->pageset); i++) {
2164 struct per_cpu_pageset *pageset;
2167 pageset = zone_pcp(zone, i);
2168 for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2169 if (pageset->pcp[j].count)
2172 if (j == ARRAY_SIZE(pageset->pcp))
2174 for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
2176 "\n cpu: %i pcp: %i"
2182 pageset->pcp[j].count,
2183 pageset->pcp[j].low,
2184 pageset->pcp[j].high,
2185 pageset->pcp[j].batch);
2191 "\n numa_foreign: %lu"
2192 "\n interleave_hit: %lu"
2193 "\n local_node: %lu"
2194 "\n other_node: %lu",
2197 pageset->numa_foreign,
2198 pageset->interleave_hit,
2199 pageset->local_node,
2200 pageset->other_node);
2204 "\n all_unreclaimable: %u"
2205 "\n prev_priority: %i"
2206 "\n temp_priority: %i"
2207 "\n start_pfn: %lu",
2208 zone->all_unreclaimable,
2209 zone->prev_priority,
2210 zone->temp_priority,
2211 zone->zone_start_pfn);
2212 spin_unlock_irqrestore(&zone->lock, flags);
2218 struct seq_operations zoneinfo_op = {
2219 .start = frag_start, /* iterate over all zones. The same as in
2223 .show = zoneinfo_show,
2226 static char *vmstat_text[] = {
2230 "nr_page_table_pages",
2255 "pgscan_kswapd_high",
2256 "pgscan_kswapd_normal",
2258 "pgscan_kswapd_dma",
2259 "pgscan_direct_high",
2260 "pgscan_direct_normal",
2261 "pgscan_direct_dma",
2266 "kswapd_inodesteal",
2274 static void *vmstat_start(struct seq_file *m, loff_t *pos)
2276 struct page_state *ps;
2278 if (*pos >= ARRAY_SIZE(vmstat_text))
2281 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
2284 return ERR_PTR(-ENOMEM);
2285 get_full_page_state(ps);
2286 ps->pgpgin /= 2; /* sectors -> kbytes */
2288 return (unsigned long *)ps + *pos;
2291 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
2294 if (*pos >= ARRAY_SIZE(vmstat_text))
2296 return (unsigned long *)m->private + *pos;
2299 static int vmstat_show(struct seq_file *m, void *arg)
2301 unsigned long *l = arg;
2302 unsigned long off = l - (unsigned long *)m->private;
2304 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
2308 static void vmstat_stop(struct seq_file *m, void *arg)
2314 struct seq_operations vmstat_op = {
2315 .start = vmstat_start,
2316 .next = vmstat_next,
2317 .stop = vmstat_stop,
2318 .show = vmstat_show,
2321 #endif /* CONFIG_PROC_FS */
2323 #ifdef CONFIG_HOTPLUG_CPU
2324 static int page_alloc_cpu_notify(struct notifier_block *self,
2325 unsigned long action, void *hcpu)
2327 int cpu = (unsigned long)hcpu;
2329 unsigned long *src, *dest;
2331 if (action == CPU_DEAD) {
2334 /* Drain local pagecache count. */
2335 count = &per_cpu(nr_pagecache_local, cpu);
2336 atomic_add(*count, &nr_pagecache);
2338 local_irq_disable();
2341 /* Add dead cpu's page_states to our own. */
2342 dest = (unsigned long *)&__get_cpu_var(page_states);
2343 src = (unsigned long *)&per_cpu(page_states, cpu);
2345 for (i = 0; i < sizeof(struct page_state)/sizeof(unsigned long);
2355 #endif /* CONFIG_HOTPLUG_CPU */
2357 void __init page_alloc_init(void)
2359 hotcpu_notifier(page_alloc_cpu_notify, 0);
2363 * setup_per_zone_lowmem_reserve - called whenever
2364 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
2365 * has a correct pages reserved value, so an adequate number of
2366 * pages are left in the zone after a successful __alloc_pages().
2368 static void setup_per_zone_lowmem_reserve(void)
2370 struct pglist_data *pgdat;
2373 for_each_pgdat(pgdat) {
2374 for (j = 0; j < MAX_NR_ZONES; j++) {
2375 struct zone *zone = pgdat->node_zones + j;
2376 unsigned long present_pages = zone->present_pages;
2378 zone->lowmem_reserve[j] = 0;
2380 for (idx = j-1; idx >= 0; idx--) {
2381 struct zone *lower_zone;
2383 if (sysctl_lowmem_reserve_ratio[idx] < 1)
2384 sysctl_lowmem_reserve_ratio[idx] = 1;
2386 lower_zone = pgdat->node_zones + idx;
2387 lower_zone->lowmem_reserve[j] = present_pages /
2388 sysctl_lowmem_reserve_ratio[idx];
2389 present_pages += lower_zone->present_pages;
2396 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
2397 * that the pages_{min,low,high} values for each zone are set correctly
2398 * with respect to min_free_kbytes.
2400 static void setup_per_zone_pages_min(void)
2402 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
2403 unsigned long lowmem_pages = 0;
2405 unsigned long flags;
2407 /* Calculate total number of !ZONE_HIGHMEM pages */
2408 for_each_zone(zone) {
2409 if (!is_highmem(zone))
2410 lowmem_pages += zone->present_pages;
2413 for_each_zone(zone) {
2414 spin_lock_irqsave(&zone->lru_lock, flags);
2415 if (is_highmem(zone)) {
2417 * Often, highmem doesn't need to reserve any pages.
2418 * But the pages_min/low/high values are also used for
2419 * batching up page reclaim activity so we need a
2420 * decent value here.
2424 min_pages = zone->present_pages / 1024;
2425 if (min_pages < SWAP_CLUSTER_MAX)
2426 min_pages = SWAP_CLUSTER_MAX;
2427 if (min_pages > 128)
2429 zone->pages_min = min_pages;
2431 /* if it's a lowmem zone, reserve a number of pages
2432 * proportionate to the zone's size.
2434 zone->pages_min = (pages_min * zone->present_pages) /
2439 * When interpreting these watermarks, just keep in mind that:
2440 * zone->pages_min == (zone->pages_min * 4) / 4;
2442 zone->pages_low = (zone->pages_min * 5) / 4;
2443 zone->pages_high = (zone->pages_min * 6) / 4;
2444 spin_unlock_irqrestore(&zone->lru_lock, flags);
2449 * Initialise min_free_kbytes.
2451 * For small machines we want it small (128k min). For large machines
2452 * we want it large (64MB max). But it is not linear, because network
2453 * bandwidth does not increase linearly with machine size. We use
2455 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2456 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2472 static int __init init_per_zone_pages_min(void)
2474 unsigned long lowmem_kbytes;
2476 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2478 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2479 if (min_free_kbytes < 128)
2480 min_free_kbytes = 128;
2481 if (min_free_kbytes > 65536)
2482 min_free_kbytes = 65536;
2483 setup_per_zone_pages_min();
2484 setup_per_zone_lowmem_reserve();
2487 module_init(init_per_zone_pages_min)
2490 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2491 * that we can call two helper functions whenever min_free_kbytes
2494 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2495 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2497 proc_dointvec(table, write, file, buffer, length, ppos);
2498 setup_per_zone_pages_min();
2503 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2504 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2505 * whenever sysctl_lowmem_reserve_ratio changes.
2507 * The reserve ratio obviously has absolutely no relation with the
2508 * pages_min watermarks. The lowmem reserve ratio can only make sense
2509 * if in function of the boot time zone sizes.
2511 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
2512 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2514 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2515 setup_per_zone_lowmem_reserve();
2519 __initdata int hashdist = HASHDIST_DEFAULT;
2522 static int __init set_hashdist(char *str)
2526 hashdist = simple_strtoul(str, &str, 0);
2529 __setup("hashdist=", set_hashdist);
2533 * allocate a large system hash table from bootmem
2534 * - it is assumed that the hash table must contain an exact power-of-2
2535 * quantity of entries
2536 * - limit is the number of hash buckets, not the total allocation size
2538 void *__init alloc_large_system_hash(const char *tablename,
2539 unsigned long bucketsize,
2540 unsigned long numentries,
2543 unsigned int *_hash_shift,
2544 unsigned int *_hash_mask,
2545 unsigned long limit)
2547 unsigned long long max = limit;
2548 unsigned long log2qty, size;
2551 /* allow the kernel cmdline to have a say */
2553 /* round applicable memory size up to nearest megabyte */
2554 numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages;
2555 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2556 numentries >>= 20 - PAGE_SHIFT;
2557 numentries <<= 20 - PAGE_SHIFT;
2559 /* limit to 1 bucket per 2^scale bytes of low memory */
2560 if (scale > PAGE_SHIFT)
2561 numentries >>= (scale - PAGE_SHIFT);
2563 numentries <<= (PAGE_SHIFT - scale);
2565 /* rounded up to nearest power of 2 in size */
2566 numentries = 1UL << (long_log2(numentries) + 1);
2568 /* limit allocation size to 1/16 total memory by default */
2570 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2571 do_div(max, bucketsize);
2574 if (numentries > max)
2577 log2qty = long_log2(numentries);
2580 size = bucketsize << log2qty;
2581 if (flags & HASH_EARLY)
2582 table = alloc_bootmem(size);
2584 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
2586 unsigned long order;
2587 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
2589 table = (void*) __get_free_pages(GFP_ATOMIC, order);
2591 } while (!table && size > PAGE_SIZE && --log2qty);
2594 panic("Failed to allocate %s hash table\n", tablename);
2596 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2599 long_log2(size) - PAGE_SHIFT,
2603 *_hash_shift = log2qty;
2605 *_hash_mask = (1 << log2qty) - 1;