2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/bootmem.h>
23 #include <linux/compiler.h>
24 #include <linux/kernel.h>
25 #include <linux/module.h>
26 #include <linux/suspend.h>
27 #include <linux/pagevec.h>
28 #include <linux/blkdev.h>
29 #include <linux/slab.h>
30 #include <linux/notifier.h>
31 #include <linux/topology.h>
32 #include <linux/sysctl.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/memory_hotplug.h>
36 #include <linux/nodemask.h>
37 #include <linux/vmalloc.h>
38 #include <linux/mempolicy.h>
39 #include <linux/stop_machine.h>
41 #include <asm/tlbflush.h>
42 #include <asm/div64.h>
46 * MCD - HACK: Find somewhere to initialize this EARLY, or make this
49 nodemask_t node_online_map __read_mostly = { { [0] = 1UL } };
50 EXPORT_SYMBOL(node_online_map);
51 nodemask_t node_possible_map __read_mostly = NODE_MASK_ALL;
52 EXPORT_SYMBOL(node_possible_map);
53 unsigned long totalram_pages __read_mostly;
54 unsigned long totalreserve_pages __read_mostly;
56 int percpu_pagelist_fraction;
58 static void __free_pages_ok(struct page *page, unsigned int order);
61 * results with 256, 32 in the lowmem_reserve sysctl:
62 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
63 * 1G machine -> (16M dma, 784M normal, 224M high)
64 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
65 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
66 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
68 * TBD: should special case ZONE_DMA32 machines here - in those we normally
69 * don't need any ZONE_NORMAL reservation
71 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 256, 256, 32 };
73 EXPORT_SYMBOL(totalram_pages);
76 * Used by page_zone() to look up the address of the struct zone whose
77 * id is encoded in the upper bits of page->flags
79 struct zone *zone_table[1 << ZONETABLE_SHIFT] __read_mostly;
80 EXPORT_SYMBOL(zone_table);
82 static char *zone_names[MAX_NR_ZONES] = { "DMA", "DMA32", "Normal", "HighMem" };
83 int min_free_kbytes = 1024;
85 unsigned long __meminitdata nr_kernel_pages;
86 unsigned long __meminitdata nr_all_pages;
88 #ifdef CONFIG_DEBUG_VM
89 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
93 unsigned long pfn = page_to_pfn(page);
96 seq = zone_span_seqbegin(zone);
97 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
99 else if (pfn < zone->zone_start_pfn)
101 } while (zone_span_seqretry(zone, seq));
106 static int page_is_consistent(struct zone *zone, struct page *page)
108 #ifdef CONFIG_HOLES_IN_ZONE
109 if (!pfn_valid(page_to_pfn(page)))
112 if (zone != page_zone(page))
118 * Temporary debugging check for pages not lying within a given zone.
120 static int bad_range(struct zone *zone, struct page *page)
122 if (page_outside_zone_boundaries(zone, page))
124 if (!page_is_consistent(zone, page))
130 static inline int bad_range(struct zone *zone, struct page *page)
136 static void bad_page(struct page *page)
138 printk(KERN_EMERG "Bad page state in process '%s'\n"
139 KERN_EMERG "page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d\n"
140 KERN_EMERG "Trying to fix it up, but a reboot is needed\n"
141 KERN_EMERG "Backtrace:\n",
142 current->comm, page, (int)(2*sizeof(unsigned long)),
143 (unsigned long)page->flags, page->mapping,
144 page_mapcount(page), page_count(page));
146 page->flags &= ~(1 << PG_lru |
156 set_page_count(page, 0);
157 reset_page_mapcount(page);
158 page->mapping = NULL;
159 add_taint(TAINT_BAD_PAGE);
163 * Higher-order pages are called "compound pages". They are structured thusly:
165 * The first PAGE_SIZE page is called the "head page".
167 * The remaining PAGE_SIZE pages are called "tail pages".
169 * All pages have PG_compound set. All pages have their ->private pointing at
170 * the head page (even the head page has this).
172 * The first tail page's ->lru.next holds the address of the compound page's
173 * put_page() function. Its ->lru.prev holds the order of allocation.
174 * This usage means that zero-order pages may not be compound.
177 static void free_compound_page(struct page *page)
179 __free_pages_ok(page, (unsigned long)page[1].lru.prev);
182 static void prep_compound_page(struct page *page, unsigned long order)
185 int nr_pages = 1 << order;
187 page[1].lru.next = (void *)free_compound_page; /* set dtor */
188 page[1].lru.prev = (void *)order;
189 for (i = 0; i < nr_pages; i++) {
190 struct page *p = page + i;
192 __SetPageCompound(p);
193 set_page_private(p, (unsigned long)page);
197 static void destroy_compound_page(struct page *page, unsigned long order)
200 int nr_pages = 1 << order;
202 if (unlikely((unsigned long)page[1].lru.prev != order))
205 for (i = 0; i < nr_pages; i++) {
206 struct page *p = page + i;
208 if (unlikely(!PageCompound(p) |
209 (page_private(p) != (unsigned long)page)))
211 __ClearPageCompound(p);
215 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
219 VM_BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
221 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
222 * and __GFP_HIGHMEM from hard or soft interrupt context.
224 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
225 for (i = 0; i < (1 << order); i++)
226 clear_highpage(page + i);
230 * function for dealing with page's order in buddy system.
231 * zone->lock is already acquired when we use these.
232 * So, we don't need atomic page->flags operations here.
234 static inline unsigned long page_order(struct page *page)
236 return page_private(page);
239 static inline void set_page_order(struct page *page, int order)
241 set_page_private(page, order);
242 __SetPageBuddy(page);
245 static inline void rmv_page_order(struct page *page)
247 __ClearPageBuddy(page);
248 set_page_private(page, 0);
252 * Locate the struct page for both the matching buddy in our
253 * pair (buddy1) and the combined O(n+1) page they form (page).
255 * 1) Any buddy B1 will have an order O twin B2 which satisfies
256 * the following equation:
258 * For example, if the starting buddy (buddy2) is #8 its order
260 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
262 * 2) Any buddy B will have an order O+1 parent P which
263 * satisfies the following equation:
266 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
268 static inline struct page *
269 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
271 unsigned long buddy_idx = page_idx ^ (1 << order);
273 return page + (buddy_idx - page_idx);
276 static inline unsigned long
277 __find_combined_index(unsigned long page_idx, unsigned int order)
279 return (page_idx & ~(1 << order));
283 * This function checks whether a page is free && is the buddy
284 * we can do coalesce a page and its buddy if
285 * (a) the buddy is not in a hole &&
286 * (b) the buddy is in the buddy system &&
287 * (c) a page and its buddy have the same order &&
288 * (d) a page and its buddy are in the same zone.
290 * For recording whether a page is in the buddy system, we use PG_buddy.
291 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
293 * For recording page's order, we use page_private(page).
295 static inline int page_is_buddy(struct page *page, struct page *buddy,
298 #ifdef CONFIG_HOLES_IN_ZONE
299 if (!pfn_valid(page_to_pfn(buddy)))
303 if (page_zone_id(page) != page_zone_id(buddy))
306 if (PageBuddy(buddy) && page_order(buddy) == order) {
307 BUG_ON(page_count(buddy) != 0);
314 * Freeing function for a buddy system allocator.
316 * The concept of a buddy system is to maintain direct-mapped table
317 * (containing bit values) for memory blocks of various "orders".
318 * The bottom level table contains the map for the smallest allocatable
319 * units of memory (here, pages), and each level above it describes
320 * pairs of units from the levels below, hence, "buddies".
321 * At a high level, all that happens here is marking the table entry
322 * at the bottom level available, and propagating the changes upward
323 * as necessary, plus some accounting needed to play nicely with other
324 * parts of the VM system.
325 * At each level, we keep a list of pages, which are heads of continuous
326 * free pages of length of (1 << order) and marked with PG_buddy. Page's
327 * order is recorded in page_private(page) field.
328 * So when we are allocating or freeing one, we can derive the state of the
329 * other. That is, if we allocate a small block, and both were
330 * free, the remainder of the region must be split into blocks.
331 * If a block is freed, and its buddy is also free, then this
332 * triggers coalescing into a block of larger size.
337 static inline void __free_one_page(struct page *page,
338 struct zone *zone, unsigned int order)
340 unsigned long page_idx;
341 int order_size = 1 << order;
343 if (unlikely(PageCompound(page)))
344 destroy_compound_page(page, order);
346 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
348 VM_BUG_ON(page_idx & (order_size - 1));
349 VM_BUG_ON(bad_range(zone, page));
351 zone->free_pages += order_size;
352 while (order < MAX_ORDER-1) {
353 unsigned long combined_idx;
354 struct free_area *area;
357 buddy = __page_find_buddy(page, page_idx, order);
358 if (!page_is_buddy(page, buddy, order))
359 break; /* Move the buddy up one level. */
361 list_del(&buddy->lru);
362 area = zone->free_area + order;
364 rmv_page_order(buddy);
365 combined_idx = __find_combined_index(page_idx, order);
366 page = page + (combined_idx - page_idx);
367 page_idx = combined_idx;
370 set_page_order(page, order);
371 list_add(&page->lru, &zone->free_area[order].free_list);
372 zone->free_area[order].nr_free++;
375 static inline int free_pages_check(struct page *page)
377 if (unlikely(page_mapcount(page) |
378 (page->mapping != NULL) |
379 (page_count(page) != 0) |
393 __ClearPageDirty(page);
395 * For now, we report if PG_reserved was found set, but do not
396 * clear it, and do not free the page. But we shall soon need
397 * to do more, for when the ZERO_PAGE count wraps negative.
399 return PageReserved(page);
403 * Frees a list of pages.
404 * Assumes all pages on list are in same zone, and of same order.
405 * count is the number of pages to free.
407 * If the zone was previously in an "all pages pinned" state then look to
408 * see if this freeing clears that state.
410 * And clear the zone's pages_scanned counter, to hold off the "all pages are
411 * pinned" detection logic.
413 static void free_pages_bulk(struct zone *zone, int count,
414 struct list_head *list, int order)
416 spin_lock(&zone->lock);
417 zone->all_unreclaimable = 0;
418 zone->pages_scanned = 0;
422 VM_BUG_ON(list_empty(list));
423 page = list_entry(list->prev, struct page, lru);
424 /* have to delete it as __free_one_page list manipulates */
425 list_del(&page->lru);
426 __free_one_page(page, zone, order);
428 spin_unlock(&zone->lock);
431 static void free_one_page(struct zone *zone, struct page *page, int order)
434 list_add(&page->lru, &list);
435 free_pages_bulk(zone, 1, &list, order);
438 static void __free_pages_ok(struct page *page, unsigned int order)
444 arch_free_page(page, order);
445 if (!PageHighMem(page))
446 debug_check_no_locks_freed(page_address(page),
449 for (i = 0 ; i < (1 << order) ; ++i)
450 reserved += free_pages_check(page + i);
454 kernel_map_pages(page, 1 << order, 0);
455 local_irq_save(flags);
456 __count_vm_events(PGFREE, 1 << order);
457 free_one_page(page_zone(page), page, order);
458 local_irq_restore(flags);
462 * permit the bootmem allocator to evade page validation on high-order frees
464 void fastcall __init __free_pages_bootmem(struct page *page, unsigned int order)
467 __ClearPageReserved(page);
468 set_page_count(page, 0);
469 set_page_refcounted(page);
475 for (loop = 0; loop < BITS_PER_LONG; loop++) {
476 struct page *p = &page[loop];
478 if (loop + 1 < BITS_PER_LONG)
480 __ClearPageReserved(p);
481 set_page_count(p, 0);
484 set_page_refcounted(page);
485 __free_pages(page, order);
491 * The order of subdivision here is critical for the IO subsystem.
492 * Please do not alter this order without good reasons and regression
493 * testing. Specifically, as large blocks of memory are subdivided,
494 * the order in which smaller blocks are delivered depends on the order
495 * they're subdivided in this function. This is the primary factor
496 * influencing the order in which pages are delivered to the IO
497 * subsystem according to empirical testing, and this is also justified
498 * by considering the behavior of a buddy system containing a single
499 * large block of memory acted on by a series of small allocations.
500 * This behavior is a critical factor in sglist merging's success.
504 static inline void expand(struct zone *zone, struct page *page,
505 int low, int high, struct free_area *area)
507 unsigned long size = 1 << high;
513 VM_BUG_ON(bad_range(zone, &page[size]));
514 list_add(&page[size].lru, &area->free_list);
516 set_page_order(&page[size], high);
521 * This page is about to be returned from the page allocator
523 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
525 if (unlikely(page_mapcount(page) |
526 (page->mapping != NULL) |
527 (page_count(page) != 0) |
543 * For now, we report if PG_reserved was found set, but do not
544 * clear it, and do not allocate the page: as a safety net.
546 if (PageReserved(page))
549 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
550 1 << PG_referenced | 1 << PG_arch_1 |
551 1 << PG_checked | 1 << PG_mappedtodisk);
552 set_page_private(page, 0);
553 set_page_refcounted(page);
554 kernel_map_pages(page, 1 << order, 1);
556 if (gfp_flags & __GFP_ZERO)
557 prep_zero_page(page, order, gfp_flags);
559 if (order && (gfp_flags & __GFP_COMP))
560 prep_compound_page(page, order);
566 * Do the hard work of removing an element from the buddy allocator.
567 * Call me with the zone->lock already held.
569 static struct page *__rmqueue(struct zone *zone, unsigned int order)
571 struct free_area * area;
572 unsigned int current_order;
575 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
576 area = zone->free_area + current_order;
577 if (list_empty(&area->free_list))
580 page = list_entry(area->free_list.next, struct page, lru);
581 list_del(&page->lru);
582 rmv_page_order(page);
584 zone->free_pages -= 1UL << order;
585 expand(zone, page, order, current_order, area);
593 * Obtain a specified number of elements from the buddy allocator, all under
594 * a single hold of the lock, for efficiency. Add them to the supplied list.
595 * Returns the number of new pages which were placed at *list.
597 static int rmqueue_bulk(struct zone *zone, unsigned int order,
598 unsigned long count, struct list_head *list)
602 spin_lock(&zone->lock);
603 for (i = 0; i < count; ++i) {
604 struct page *page = __rmqueue(zone, order);
605 if (unlikely(page == NULL))
607 list_add_tail(&page->lru, list);
609 spin_unlock(&zone->lock);
615 * Called from the slab reaper to drain pagesets on a particular node that
616 * belong to the currently executing processor.
617 * Note that this function must be called with the thread pinned to
618 * a single processor.
620 void drain_node_pages(int nodeid)
625 for (z = 0; z < MAX_NR_ZONES; z++) {
626 struct zone *zone = NODE_DATA(nodeid)->node_zones + z;
627 struct per_cpu_pageset *pset;
629 pset = zone_pcp(zone, smp_processor_id());
630 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
631 struct per_cpu_pages *pcp;
635 local_irq_save(flags);
636 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
638 local_irq_restore(flags);
645 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
646 static void __drain_pages(unsigned int cpu)
652 for_each_zone(zone) {
653 struct per_cpu_pageset *pset;
655 pset = zone_pcp(zone, cpu);
656 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
657 struct per_cpu_pages *pcp;
660 local_irq_save(flags);
661 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
663 local_irq_restore(flags);
667 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
671 void mark_free_pages(struct zone *zone)
673 unsigned long zone_pfn, flags;
675 struct list_head *curr;
677 if (!zone->spanned_pages)
680 spin_lock_irqsave(&zone->lock, flags);
681 for (zone_pfn = 0; zone_pfn < zone->spanned_pages; ++zone_pfn)
682 ClearPageNosaveFree(pfn_to_page(zone_pfn + zone->zone_start_pfn));
684 for (order = MAX_ORDER - 1; order >= 0; --order)
685 list_for_each(curr, &zone->free_area[order].free_list) {
686 unsigned long start_pfn, i;
688 start_pfn = page_to_pfn(list_entry(curr, struct page, lru));
690 for (i=0; i < (1<<order); i++)
691 SetPageNosaveFree(pfn_to_page(start_pfn+i));
693 spin_unlock_irqrestore(&zone->lock, flags);
697 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
699 void drain_local_pages(void)
703 local_irq_save(flags);
704 __drain_pages(smp_processor_id());
705 local_irq_restore(flags);
707 #endif /* CONFIG_PM */
710 * Free a 0-order page
712 static void fastcall free_hot_cold_page(struct page *page, int cold)
714 struct zone *zone = page_zone(page);
715 struct per_cpu_pages *pcp;
718 arch_free_page(page, 0);
721 page->mapping = NULL;
722 if (free_pages_check(page))
725 kernel_map_pages(page, 1, 0);
727 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
728 local_irq_save(flags);
729 __count_vm_event(PGFREE);
730 list_add(&page->lru, &pcp->list);
732 if (pcp->count >= pcp->high) {
733 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
734 pcp->count -= pcp->batch;
736 local_irq_restore(flags);
740 void fastcall free_hot_page(struct page *page)
742 free_hot_cold_page(page, 0);
745 void fastcall free_cold_page(struct page *page)
747 free_hot_cold_page(page, 1);
751 * split_page takes a non-compound higher-order page, and splits it into
752 * n (1<<order) sub-pages: page[0..n]
753 * Each sub-page must be freed individually.
755 * Note: this is probably too low level an operation for use in drivers.
756 * Please consult with lkml before using this in your driver.
758 void split_page(struct page *page, unsigned int order)
762 VM_BUG_ON(PageCompound(page));
763 VM_BUG_ON(!page_count(page));
764 for (i = 1; i < (1 << order); i++)
765 set_page_refcounted(page + i);
769 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
770 * we cheat by calling it from here, in the order > 0 path. Saves a branch
773 static struct page *buffered_rmqueue(struct zonelist *zonelist,
774 struct zone *zone, int order, gfp_t gfp_flags)
778 int cold = !!(gfp_flags & __GFP_COLD);
783 if (likely(order == 0)) {
784 struct per_cpu_pages *pcp;
786 pcp = &zone_pcp(zone, cpu)->pcp[cold];
787 local_irq_save(flags);
789 pcp->count += rmqueue_bulk(zone, 0,
790 pcp->batch, &pcp->list);
791 if (unlikely(!pcp->count))
794 page = list_entry(pcp->list.next, struct page, lru);
795 list_del(&page->lru);
798 spin_lock_irqsave(&zone->lock, flags);
799 page = __rmqueue(zone, order);
800 spin_unlock(&zone->lock);
805 __count_zone_vm_events(PGALLOC, zone, 1 << order);
806 zone_statistics(zonelist, zone);
807 local_irq_restore(flags);
810 VM_BUG_ON(bad_range(zone, page));
811 if (prep_new_page(page, order, gfp_flags))
816 local_irq_restore(flags);
821 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
822 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
823 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
824 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
825 #define ALLOC_HARDER 0x10 /* try to alloc harder */
826 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
827 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
830 * Return 1 if free pages are above 'mark'. This takes into account the order
833 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
834 int classzone_idx, int alloc_flags)
836 /* free_pages my go negative - that's OK */
837 long min = mark, free_pages = z->free_pages - (1 << order) + 1;
840 if (alloc_flags & ALLOC_HIGH)
842 if (alloc_flags & ALLOC_HARDER)
845 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
847 for (o = 0; o < order; o++) {
848 /* At the next order, this order's pages become unavailable */
849 free_pages -= z->free_area[o].nr_free << o;
851 /* Require fewer higher order pages to be free */
854 if (free_pages <= min)
861 * get_page_from_freeliest goes through the zonelist trying to allocate
865 get_page_from_freelist(gfp_t gfp_mask, unsigned int order,
866 struct zonelist *zonelist, int alloc_flags)
868 struct zone **z = zonelist->zones;
869 struct page *page = NULL;
870 int classzone_idx = zone_idx(*z);
873 * Go through the zonelist once, looking for a zone with enough free.
874 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
877 if ((alloc_flags & ALLOC_CPUSET) &&
878 !cpuset_zone_allowed(*z, gfp_mask))
881 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
883 if (alloc_flags & ALLOC_WMARK_MIN)
884 mark = (*z)->pages_min;
885 else if (alloc_flags & ALLOC_WMARK_LOW)
886 mark = (*z)->pages_low;
888 mark = (*z)->pages_high;
889 if (!zone_watermark_ok(*z, order, mark,
890 classzone_idx, alloc_flags))
891 if (!zone_reclaim_mode ||
892 !zone_reclaim(*z, gfp_mask, order))
896 page = buffered_rmqueue(zonelist, *z, order, gfp_mask);
900 } while (*(++z) != NULL);
905 * This is the 'heart' of the zoned buddy allocator.
907 struct page * fastcall
908 __alloc_pages(gfp_t gfp_mask, unsigned int order,
909 struct zonelist *zonelist)
911 const gfp_t wait = gfp_mask & __GFP_WAIT;
914 struct reclaim_state reclaim_state;
915 struct task_struct *p = current;
918 int did_some_progress;
920 might_sleep_if(wait);
923 z = zonelist->zones; /* the list of zones suitable for gfp_mask */
925 if (unlikely(*z == NULL)) {
926 /* Should this ever happen?? */
930 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
931 zonelist, ALLOC_WMARK_LOW|ALLOC_CPUSET);
936 wakeup_kswapd(*z, order);
940 * OK, we're below the kswapd watermark and have kicked background
941 * reclaim. Now things get more complex, so set up alloc_flags according
942 * to how we want to proceed.
944 * The caller may dip into page reserves a bit more if the caller
945 * cannot run direct reclaim, or if the caller has realtime scheduling
946 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
947 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
949 alloc_flags = ALLOC_WMARK_MIN;
950 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
951 alloc_flags |= ALLOC_HARDER;
952 if (gfp_mask & __GFP_HIGH)
953 alloc_flags |= ALLOC_HIGH;
955 alloc_flags |= ALLOC_CPUSET;
958 * Go through the zonelist again. Let __GFP_HIGH and allocations
959 * coming from realtime tasks go deeper into reserves.
961 * This is the last chance, in general, before the goto nopage.
962 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
963 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
965 page = get_page_from_freelist(gfp_mask, order, zonelist, alloc_flags);
969 /* This allocation should allow future memory freeing. */
971 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
972 && !in_interrupt()) {
973 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
975 /* go through the zonelist yet again, ignoring mins */
976 page = get_page_from_freelist(gfp_mask, order,
977 zonelist, ALLOC_NO_WATERMARKS);
980 if (gfp_mask & __GFP_NOFAIL) {
981 blk_congestion_wait(WRITE, HZ/50);
988 /* Atomic allocations - we can't balance anything */
995 /* We now go into synchronous reclaim */
996 cpuset_memory_pressure_bump();
997 p->flags |= PF_MEMALLOC;
998 reclaim_state.reclaimed_slab = 0;
999 p->reclaim_state = &reclaim_state;
1001 did_some_progress = try_to_free_pages(zonelist->zones, gfp_mask);
1003 p->reclaim_state = NULL;
1004 p->flags &= ~PF_MEMALLOC;
1008 if (likely(did_some_progress)) {
1009 page = get_page_from_freelist(gfp_mask, order,
1010 zonelist, alloc_flags);
1013 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1015 * Go through the zonelist yet one more time, keep
1016 * very high watermark here, this is only to catch
1017 * a parallel oom killing, we must fail if we're still
1018 * under heavy pressure.
1020 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1021 zonelist, ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1025 out_of_memory(zonelist, gfp_mask, order);
1030 * Don't let big-order allocations loop unless the caller explicitly
1031 * requests that. Wait for some write requests to complete then retry.
1033 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
1034 * <= 3, but that may not be true in other implementations.
1037 if (!(gfp_mask & __GFP_NORETRY)) {
1038 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
1040 if (gfp_mask & __GFP_NOFAIL)
1044 blk_congestion_wait(WRITE, HZ/50);
1049 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1050 printk(KERN_WARNING "%s: page allocation failure."
1051 " order:%d, mode:0x%x\n",
1052 p->comm, order, gfp_mask);
1060 EXPORT_SYMBOL(__alloc_pages);
1063 * Common helper functions.
1065 fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1068 page = alloc_pages(gfp_mask, order);
1071 return (unsigned long) page_address(page);
1074 EXPORT_SYMBOL(__get_free_pages);
1076 fastcall unsigned long get_zeroed_page(gfp_t gfp_mask)
1081 * get_zeroed_page() returns a 32-bit address, which cannot represent
1084 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1086 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1088 return (unsigned long) page_address(page);
1092 EXPORT_SYMBOL(get_zeroed_page);
1094 void __pagevec_free(struct pagevec *pvec)
1096 int i = pagevec_count(pvec);
1099 free_hot_cold_page(pvec->pages[i], pvec->cold);
1102 fastcall void __free_pages(struct page *page, unsigned int order)
1104 if (put_page_testzero(page)) {
1106 free_hot_page(page);
1108 __free_pages_ok(page, order);
1112 EXPORT_SYMBOL(__free_pages);
1114 fastcall void free_pages(unsigned long addr, unsigned int order)
1117 VM_BUG_ON(!virt_addr_valid((void *)addr));
1118 __free_pages(virt_to_page((void *)addr), order);
1122 EXPORT_SYMBOL(free_pages);
1125 * Total amount of free (allocatable) RAM:
1127 unsigned int nr_free_pages(void)
1129 unsigned int sum = 0;
1133 sum += zone->free_pages;
1138 EXPORT_SYMBOL(nr_free_pages);
1141 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
1143 unsigned int i, sum = 0;
1145 for (i = 0; i < MAX_NR_ZONES; i++)
1146 sum += pgdat->node_zones[i].free_pages;
1152 static unsigned int nr_free_zone_pages(int offset)
1154 /* Just pick one node, since fallback list is circular */
1155 pg_data_t *pgdat = NODE_DATA(numa_node_id());
1156 unsigned int sum = 0;
1158 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1159 struct zone **zonep = zonelist->zones;
1162 for (zone = *zonep++; zone; zone = *zonep++) {
1163 unsigned long size = zone->present_pages;
1164 unsigned long high = zone->pages_high;
1173 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1175 unsigned int nr_free_buffer_pages(void)
1177 return nr_free_zone_pages(gfp_zone(GFP_USER));
1181 * Amount of free RAM allocatable within all zones
1183 unsigned int nr_free_pagecache_pages(void)
1185 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER));
1188 static void show_node(struct zone *zone)
1190 printk("Node %d ", zone->zone_pgdat->node_id);
1193 #define show_node(zone) do { } while (0)
1196 void si_meminfo(struct sysinfo *val)
1198 val->totalram = totalram_pages;
1200 val->freeram = nr_free_pages();
1201 val->bufferram = nr_blockdev_pages();
1202 #ifdef CONFIG_HIGHMEM
1203 val->totalhigh = totalhigh_pages;
1204 val->freehigh = nr_free_highpages();
1209 val->mem_unit = PAGE_SIZE;
1212 EXPORT_SYMBOL(si_meminfo);
1215 void si_meminfo_node(struct sysinfo *val, int nid)
1217 pg_data_t *pgdat = NODE_DATA(nid);
1219 val->totalram = pgdat->node_present_pages;
1220 val->freeram = nr_free_pages_pgdat(pgdat);
1221 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1222 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1223 val->mem_unit = PAGE_SIZE;
1227 #define K(x) ((x) << (PAGE_SHIFT-10))
1230 * Show free area list (used inside shift_scroll-lock stuff)
1231 * We also calculate the percentage fragmentation. We do this by counting the
1232 * memory on each free list with the exception of the first item on the list.
1234 void show_free_areas(void)
1236 int cpu, temperature;
1237 unsigned long active;
1238 unsigned long inactive;
1242 for_each_zone(zone) {
1244 printk("%s per-cpu:", zone->name);
1246 if (!populated_zone(zone)) {
1252 for_each_online_cpu(cpu) {
1253 struct per_cpu_pageset *pageset;
1255 pageset = zone_pcp(zone, cpu);
1257 for (temperature = 0; temperature < 2; temperature++)
1258 printk("cpu %d %s: high %d, batch %d used:%d\n",
1260 temperature ? "cold" : "hot",
1261 pageset->pcp[temperature].high,
1262 pageset->pcp[temperature].batch,
1263 pageset->pcp[temperature].count);
1267 get_zone_counts(&active, &inactive, &free);
1269 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1270 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1273 global_page_state(NR_FILE_DIRTY),
1274 global_page_state(NR_WRITEBACK),
1275 global_page_state(NR_UNSTABLE_NFS),
1277 global_page_state(NR_SLAB),
1278 global_page_state(NR_FILE_MAPPED),
1279 global_page_state(NR_PAGETABLE));
1281 for_each_zone(zone) {
1293 " pages_scanned:%lu"
1294 " all_unreclaimable? %s"
1297 K(zone->free_pages),
1300 K(zone->pages_high),
1302 K(zone->nr_inactive),
1303 K(zone->present_pages),
1304 zone->pages_scanned,
1305 (zone->all_unreclaimable ? "yes" : "no")
1307 printk("lowmem_reserve[]:");
1308 for (i = 0; i < MAX_NR_ZONES; i++)
1309 printk(" %lu", zone->lowmem_reserve[i]);
1313 for_each_zone(zone) {
1314 unsigned long nr[MAX_ORDER], flags, order, total = 0;
1317 printk("%s: ", zone->name);
1318 if (!populated_zone(zone)) {
1323 spin_lock_irqsave(&zone->lock, flags);
1324 for (order = 0; order < MAX_ORDER; order++) {
1325 nr[order] = zone->free_area[order].nr_free;
1326 total += nr[order] << order;
1328 spin_unlock_irqrestore(&zone->lock, flags);
1329 for (order = 0; order < MAX_ORDER; order++)
1330 printk("%lu*%lukB ", nr[order], K(1UL) << order);
1331 printk("= %lukB\n", K(total));
1334 show_swap_cache_info();
1338 * Builds allocation fallback zone lists.
1340 * Add all populated zones of a node to the zonelist.
1342 static int __meminit build_zonelists_node(pg_data_t *pgdat,
1343 struct zonelist *zonelist, int nr_zones, int zone_type)
1347 BUG_ON(zone_type > ZONE_HIGHMEM);
1350 zone = pgdat->node_zones + zone_type;
1351 if (populated_zone(zone)) {
1352 #ifndef CONFIG_HIGHMEM
1353 BUG_ON(zone_type > ZONE_NORMAL);
1355 zonelist->zones[nr_zones++] = zone;
1356 check_highest_zone(zone_type);
1360 } while (zone_type >= 0);
1364 static inline int highest_zone(int zone_bits)
1366 int res = ZONE_NORMAL;
1367 if (zone_bits & (__force int)__GFP_HIGHMEM)
1369 if (zone_bits & (__force int)__GFP_DMA32)
1371 if (zone_bits & (__force int)__GFP_DMA)
1377 #define MAX_NODE_LOAD (num_online_nodes())
1378 static int __meminitdata node_load[MAX_NUMNODES];
1380 * find_next_best_node - find the next node that should appear in a given node's fallback list
1381 * @node: node whose fallback list we're appending
1382 * @used_node_mask: nodemask_t of already used nodes
1384 * We use a number of factors to determine which is the next node that should
1385 * appear on a given node's fallback list. The node should not have appeared
1386 * already in @node's fallback list, and it should be the next closest node
1387 * according to the distance array (which contains arbitrary distance values
1388 * from each node to each node in the system), and should also prefer nodes
1389 * with no CPUs, since presumably they'll have very little allocation pressure
1390 * on them otherwise.
1391 * It returns -1 if no node is found.
1393 static int __meminit find_next_best_node(int node, nodemask_t *used_node_mask)
1396 int min_val = INT_MAX;
1399 /* Use the local node if we haven't already */
1400 if (!node_isset(node, *used_node_mask)) {
1401 node_set(node, *used_node_mask);
1405 for_each_online_node(n) {
1408 /* Don't want a node to appear more than once */
1409 if (node_isset(n, *used_node_mask))
1412 /* Use the distance array to find the distance */
1413 val = node_distance(node, n);
1415 /* Penalize nodes under us ("prefer the next node") */
1418 /* Give preference to headless and unused nodes */
1419 tmp = node_to_cpumask(n);
1420 if (!cpus_empty(tmp))
1421 val += PENALTY_FOR_NODE_WITH_CPUS;
1423 /* Slight preference for less loaded node */
1424 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1425 val += node_load[n];
1427 if (val < min_val) {
1434 node_set(best_node, *used_node_mask);
1439 static void __meminit build_zonelists(pg_data_t *pgdat)
1441 int i, j, k, node, local_node;
1442 int prev_node, load;
1443 struct zonelist *zonelist;
1444 nodemask_t used_mask;
1446 /* initialize zonelists */
1447 for (i = 0; i < GFP_ZONETYPES; i++) {
1448 zonelist = pgdat->node_zonelists + i;
1449 zonelist->zones[0] = NULL;
1452 /* NUMA-aware ordering of nodes */
1453 local_node = pgdat->node_id;
1454 load = num_online_nodes();
1455 prev_node = local_node;
1456 nodes_clear(used_mask);
1457 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1458 int distance = node_distance(local_node, node);
1461 * If another node is sufficiently far away then it is better
1462 * to reclaim pages in a zone before going off node.
1464 if (distance > RECLAIM_DISTANCE)
1465 zone_reclaim_mode = 1;
1468 * We don't want to pressure a particular node.
1469 * So adding penalty to the first node in same
1470 * distance group to make it round-robin.
1473 if (distance != node_distance(local_node, prev_node))
1474 node_load[node] += load;
1477 for (i = 0; i < GFP_ZONETYPES; i++) {
1478 zonelist = pgdat->node_zonelists + i;
1479 for (j = 0; zonelist->zones[j] != NULL; j++);
1481 k = highest_zone(i);
1483 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1484 zonelist->zones[j] = NULL;
1489 #else /* CONFIG_NUMA */
1491 static void __meminit build_zonelists(pg_data_t *pgdat)
1493 int i, j, k, node, local_node;
1495 local_node = pgdat->node_id;
1496 for (i = 0; i < GFP_ZONETYPES; i++) {
1497 struct zonelist *zonelist;
1499 zonelist = pgdat->node_zonelists + i;
1502 k = highest_zone(i);
1503 j = build_zonelists_node(pgdat, zonelist, j, k);
1505 * Now we build the zonelist so that it contains the zones
1506 * of all the other nodes.
1507 * We don't want to pressure a particular node, so when
1508 * building the zones for node N, we make sure that the
1509 * zones coming right after the local ones are those from
1510 * node N+1 (modulo N)
1512 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1513 if (!node_online(node))
1515 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1517 for (node = 0; node < local_node; node++) {
1518 if (!node_online(node))
1520 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1523 zonelist->zones[j] = NULL;
1527 #endif /* CONFIG_NUMA */
1529 /* return values int ....just for stop_machine_run() */
1530 static int __meminit __build_all_zonelists(void *dummy)
1533 for_each_online_node(nid)
1534 build_zonelists(NODE_DATA(nid));
1538 void __meminit build_all_zonelists(void)
1540 if (system_state == SYSTEM_BOOTING) {
1541 __build_all_zonelists(0);
1542 cpuset_init_current_mems_allowed();
1544 /* we have to stop all cpus to guaranntee there is no user
1546 stop_machine_run(__build_all_zonelists, NULL, NR_CPUS);
1547 /* cpuset refresh routine should be here */
1549 vm_total_pages = nr_free_pagecache_pages();
1550 printk("Built %i zonelists. Total pages: %ld\n",
1551 num_online_nodes(), vm_total_pages);
1555 * Helper functions to size the waitqueue hash table.
1556 * Essentially these want to choose hash table sizes sufficiently
1557 * large so that collisions trying to wait on pages are rare.
1558 * But in fact, the number of active page waitqueues on typical
1559 * systems is ridiculously low, less than 200. So this is even
1560 * conservative, even though it seems large.
1562 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1563 * waitqueues, i.e. the size of the waitq table given the number of pages.
1565 #define PAGES_PER_WAITQUEUE 256
1567 #ifndef CONFIG_MEMORY_HOTPLUG
1568 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
1570 unsigned long size = 1;
1572 pages /= PAGES_PER_WAITQUEUE;
1574 while (size < pages)
1578 * Once we have dozens or even hundreds of threads sleeping
1579 * on IO we've got bigger problems than wait queue collision.
1580 * Limit the size of the wait table to a reasonable size.
1582 size = min(size, 4096UL);
1584 return max(size, 4UL);
1588 * A zone's size might be changed by hot-add, so it is not possible to determine
1589 * a suitable size for its wait_table. So we use the maximum size now.
1591 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
1593 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
1594 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
1595 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
1597 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
1598 * or more by the traditional way. (See above). It equals:
1600 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
1601 * ia64(16K page size) : = ( 8G + 4M)byte.
1602 * powerpc (64K page size) : = (32G +16M)byte.
1604 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
1611 * This is an integer logarithm so that shifts can be used later
1612 * to extract the more random high bits from the multiplicative
1613 * hash function before the remainder is taken.
1615 static inline unsigned long wait_table_bits(unsigned long size)
1620 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1622 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1623 unsigned long *zones_size, unsigned long *zholes_size)
1625 unsigned long realtotalpages, totalpages = 0;
1628 for (i = 0; i < MAX_NR_ZONES; i++)
1629 totalpages += zones_size[i];
1630 pgdat->node_spanned_pages = totalpages;
1632 realtotalpages = totalpages;
1634 for (i = 0; i < MAX_NR_ZONES; i++)
1635 realtotalpages -= zholes_size[i];
1636 pgdat->node_present_pages = realtotalpages;
1637 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1642 * Initially all pages are reserved - free ones are freed
1643 * up by free_all_bootmem() once the early boot process is
1644 * done. Non-atomic initialization, single-pass.
1646 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1647 unsigned long start_pfn)
1650 unsigned long end_pfn = start_pfn + size;
1653 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1654 if (!early_pfn_valid(pfn))
1656 page = pfn_to_page(pfn);
1657 set_page_links(page, zone, nid, pfn);
1658 init_page_count(page);
1659 reset_page_mapcount(page);
1660 SetPageReserved(page);
1661 INIT_LIST_HEAD(&page->lru);
1662 #ifdef WANT_PAGE_VIRTUAL
1663 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1664 if (!is_highmem_idx(zone))
1665 set_page_address(page, __va(pfn << PAGE_SHIFT));
1670 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1674 for (order = 0; order < MAX_ORDER ; order++) {
1675 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1676 zone->free_area[order].nr_free = 0;
1680 #define ZONETABLE_INDEX(x, zone_nr) ((x << ZONES_SHIFT) | zone_nr)
1681 void zonetable_add(struct zone *zone, int nid, int zid, unsigned long pfn,
1684 unsigned long snum = pfn_to_section_nr(pfn);
1685 unsigned long end = pfn_to_section_nr(pfn + size);
1688 zone_table[ZONETABLE_INDEX(nid, zid)] = zone;
1690 for (; snum <= end; snum++)
1691 zone_table[ZONETABLE_INDEX(snum, zid)] = zone;
1694 #ifndef __HAVE_ARCH_MEMMAP_INIT
1695 #define memmap_init(size, nid, zone, start_pfn) \
1696 memmap_init_zone((size), (nid), (zone), (start_pfn))
1699 static int __cpuinit zone_batchsize(struct zone *zone)
1704 * The per-cpu-pages pools are set to around 1000th of the
1705 * size of the zone. But no more than 1/2 of a meg.
1707 * OK, so we don't know how big the cache is. So guess.
1709 batch = zone->present_pages / 1024;
1710 if (batch * PAGE_SIZE > 512 * 1024)
1711 batch = (512 * 1024) / PAGE_SIZE;
1712 batch /= 4; /* We effectively *= 4 below */
1717 * Clamp the batch to a 2^n - 1 value. Having a power
1718 * of 2 value was found to be more likely to have
1719 * suboptimal cache aliasing properties in some cases.
1721 * For example if 2 tasks are alternately allocating
1722 * batches of pages, one task can end up with a lot
1723 * of pages of one half of the possible page colors
1724 * and the other with pages of the other colors.
1726 batch = (1 << (fls(batch + batch/2)-1)) - 1;
1731 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
1733 struct per_cpu_pages *pcp;
1735 memset(p, 0, sizeof(*p));
1737 pcp = &p->pcp[0]; /* hot */
1739 pcp->high = 6 * batch;
1740 pcp->batch = max(1UL, 1 * batch);
1741 INIT_LIST_HEAD(&pcp->list);
1743 pcp = &p->pcp[1]; /* cold*/
1745 pcp->high = 2 * batch;
1746 pcp->batch = max(1UL, batch/2);
1747 INIT_LIST_HEAD(&pcp->list);
1751 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
1752 * to the value high for the pageset p.
1755 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
1758 struct per_cpu_pages *pcp;
1760 pcp = &p->pcp[0]; /* hot list */
1762 pcp->batch = max(1UL, high/4);
1763 if ((high/4) > (PAGE_SHIFT * 8))
1764 pcp->batch = PAGE_SHIFT * 8;
1770 * Boot pageset table. One per cpu which is going to be used for all
1771 * zones and all nodes. The parameters will be set in such a way
1772 * that an item put on a list will immediately be handed over to
1773 * the buddy list. This is safe since pageset manipulation is done
1774 * with interrupts disabled.
1776 * Some NUMA counter updates may also be caught by the boot pagesets.
1778 * The boot_pagesets must be kept even after bootup is complete for
1779 * unused processors and/or zones. They do play a role for bootstrapping
1780 * hotplugged processors.
1782 * zoneinfo_show() and maybe other functions do
1783 * not check if the processor is online before following the pageset pointer.
1784 * Other parts of the kernel may not check if the zone is available.
1786 static struct per_cpu_pageset boot_pageset[NR_CPUS];
1789 * Dynamically allocate memory for the
1790 * per cpu pageset array in struct zone.
1792 static int __cpuinit process_zones(int cpu)
1794 struct zone *zone, *dzone;
1796 for_each_zone(zone) {
1798 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
1799 GFP_KERNEL, cpu_to_node(cpu));
1800 if (!zone_pcp(zone, cpu))
1803 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
1805 if (percpu_pagelist_fraction)
1806 setup_pagelist_highmark(zone_pcp(zone, cpu),
1807 (zone->present_pages / percpu_pagelist_fraction));
1812 for_each_zone(dzone) {
1815 kfree(zone_pcp(dzone, cpu));
1816 zone_pcp(dzone, cpu) = NULL;
1821 static inline void free_zone_pagesets(int cpu)
1825 for_each_zone(zone) {
1826 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
1828 /* Free per_cpu_pageset if it is slab allocated */
1829 if (pset != &boot_pageset[cpu])
1831 zone_pcp(zone, cpu) = NULL;
1835 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
1836 unsigned long action,
1839 int cpu = (long)hcpu;
1840 int ret = NOTIFY_OK;
1843 case CPU_UP_PREPARE:
1844 if (process_zones(cpu))
1847 case CPU_UP_CANCELED:
1849 free_zone_pagesets(cpu);
1857 static struct notifier_block __cpuinitdata pageset_notifier =
1858 { &pageset_cpuup_callback, NULL, 0 };
1860 void __init setup_per_cpu_pageset(void)
1864 /* Initialize per_cpu_pageset for cpu 0.
1865 * A cpuup callback will do this for every cpu
1866 * as it comes online
1868 err = process_zones(smp_processor_id());
1870 register_cpu_notifier(&pageset_notifier);
1876 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
1879 struct pglist_data *pgdat = zone->zone_pgdat;
1883 * The per-page waitqueue mechanism uses hashed waitqueues
1886 zone->wait_table_hash_nr_entries =
1887 wait_table_hash_nr_entries(zone_size_pages);
1888 zone->wait_table_bits =
1889 wait_table_bits(zone->wait_table_hash_nr_entries);
1890 alloc_size = zone->wait_table_hash_nr_entries
1891 * sizeof(wait_queue_head_t);
1893 if (system_state == SYSTEM_BOOTING) {
1894 zone->wait_table = (wait_queue_head_t *)
1895 alloc_bootmem_node(pgdat, alloc_size);
1898 * This case means that a zone whose size was 0 gets new memory
1899 * via memory hot-add.
1900 * But it may be the case that a new node was hot-added. In
1901 * this case vmalloc() will not be able to use this new node's
1902 * memory - this wait_table must be initialized to use this new
1903 * node itself as well.
1904 * To use this new node's memory, further consideration will be
1907 zone->wait_table = (wait_queue_head_t *)vmalloc(alloc_size);
1909 if (!zone->wait_table)
1912 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
1913 init_waitqueue_head(zone->wait_table + i);
1918 static __meminit void zone_pcp_init(struct zone *zone)
1921 unsigned long batch = zone_batchsize(zone);
1923 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1925 /* Early boot. Slab allocator not functional yet */
1926 zone_pcp(zone, cpu) = &boot_pageset[cpu];
1927 setup_pageset(&boot_pageset[cpu],0);
1929 setup_pageset(zone_pcp(zone,cpu), batch);
1932 if (zone->present_pages)
1933 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
1934 zone->name, zone->present_pages, batch);
1937 __meminit int init_currently_empty_zone(struct zone *zone,
1938 unsigned long zone_start_pfn,
1941 struct pglist_data *pgdat = zone->zone_pgdat;
1943 ret = zone_wait_table_init(zone, size);
1946 pgdat->nr_zones = zone_idx(zone) + 1;
1948 zone->zone_start_pfn = zone_start_pfn;
1950 memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn);
1952 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1958 * Set up the zone data structures:
1959 * - mark all pages reserved
1960 * - mark all memory queues empty
1961 * - clear the memory bitmaps
1963 static void __meminit free_area_init_core(struct pglist_data *pgdat,
1964 unsigned long *zones_size, unsigned long *zholes_size)
1967 int nid = pgdat->node_id;
1968 unsigned long zone_start_pfn = pgdat->node_start_pfn;
1971 pgdat_resize_init(pgdat);
1972 pgdat->nr_zones = 0;
1973 init_waitqueue_head(&pgdat->kswapd_wait);
1974 pgdat->kswapd_max_order = 0;
1976 for (j = 0; j < MAX_NR_ZONES; j++) {
1977 struct zone *zone = pgdat->node_zones + j;
1978 unsigned long size, realsize;
1980 realsize = size = zones_size[j];
1982 realsize -= zholes_size[j];
1984 if (j < ZONE_HIGHMEM)
1985 nr_kernel_pages += realsize;
1986 nr_all_pages += realsize;
1988 zone->spanned_pages = size;
1989 zone->present_pages = realsize;
1991 zone->min_unmapped_ratio = (realsize*sysctl_min_unmapped_ratio)
1994 zone->name = zone_names[j];
1995 spin_lock_init(&zone->lock);
1996 spin_lock_init(&zone->lru_lock);
1997 zone_seqlock_init(zone);
1998 zone->zone_pgdat = pgdat;
1999 zone->free_pages = 0;
2001 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
2003 zone_pcp_init(zone);
2004 INIT_LIST_HEAD(&zone->active_list);
2005 INIT_LIST_HEAD(&zone->inactive_list);
2006 zone->nr_scan_active = 0;
2007 zone->nr_scan_inactive = 0;
2008 zone->nr_active = 0;
2009 zone->nr_inactive = 0;
2010 zap_zone_vm_stats(zone);
2011 atomic_set(&zone->reclaim_in_progress, 0);
2015 zonetable_add(zone, nid, j, zone_start_pfn, size);
2016 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
2018 zone_start_pfn += size;
2022 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
2024 /* Skip empty nodes */
2025 if (!pgdat->node_spanned_pages)
2028 #ifdef CONFIG_FLAT_NODE_MEM_MAP
2029 /* ia64 gets its own node_mem_map, before this, without bootmem */
2030 if (!pgdat->node_mem_map) {
2031 unsigned long size, start, end;
2035 * The zone's endpoints aren't required to be MAX_ORDER
2036 * aligned but the node_mem_map endpoints must be in order
2037 * for the buddy allocator to function correctly.
2039 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
2040 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
2041 end = ALIGN(end, MAX_ORDER_NR_PAGES);
2042 size = (end - start) * sizeof(struct page);
2043 map = alloc_remap(pgdat->node_id, size);
2045 map = alloc_bootmem_node(pgdat, size);
2046 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
2048 #ifdef CONFIG_FLATMEM
2050 * With no DISCONTIG, the global mem_map is just set as node 0's
2052 if (pgdat == NODE_DATA(0))
2053 mem_map = NODE_DATA(0)->node_mem_map;
2055 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
2058 void __meminit free_area_init_node(int nid, struct pglist_data *pgdat,
2059 unsigned long *zones_size, unsigned long node_start_pfn,
2060 unsigned long *zholes_size)
2062 pgdat->node_id = nid;
2063 pgdat->node_start_pfn = node_start_pfn;
2064 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
2066 alloc_node_mem_map(pgdat);
2068 free_area_init_core(pgdat, zones_size, zholes_size);
2071 #ifndef CONFIG_NEED_MULTIPLE_NODES
2072 static bootmem_data_t contig_bootmem_data;
2073 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
2075 EXPORT_SYMBOL(contig_page_data);
2078 void __init free_area_init(unsigned long *zones_size)
2080 free_area_init_node(0, NODE_DATA(0), zones_size,
2081 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
2084 #ifdef CONFIG_HOTPLUG_CPU
2085 static int page_alloc_cpu_notify(struct notifier_block *self,
2086 unsigned long action, void *hcpu)
2088 int cpu = (unsigned long)hcpu;
2090 if (action == CPU_DEAD) {
2091 local_irq_disable();
2093 vm_events_fold_cpu(cpu);
2095 refresh_cpu_vm_stats(cpu);
2099 #endif /* CONFIG_HOTPLUG_CPU */
2101 void __init page_alloc_init(void)
2103 hotcpu_notifier(page_alloc_cpu_notify, 0);
2107 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
2108 * or min_free_kbytes changes.
2110 static void calculate_totalreserve_pages(void)
2112 struct pglist_data *pgdat;
2113 unsigned long reserve_pages = 0;
2116 for_each_online_pgdat(pgdat) {
2117 for (i = 0; i < MAX_NR_ZONES; i++) {
2118 struct zone *zone = pgdat->node_zones + i;
2119 unsigned long max = 0;
2121 /* Find valid and maximum lowmem_reserve in the zone */
2122 for (j = i; j < MAX_NR_ZONES; j++) {
2123 if (zone->lowmem_reserve[j] > max)
2124 max = zone->lowmem_reserve[j];
2127 /* we treat pages_high as reserved pages. */
2128 max += zone->pages_high;
2130 if (max > zone->present_pages)
2131 max = zone->present_pages;
2132 reserve_pages += max;
2135 totalreserve_pages = reserve_pages;
2139 * setup_per_zone_lowmem_reserve - called whenever
2140 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
2141 * has a correct pages reserved value, so an adequate number of
2142 * pages are left in the zone after a successful __alloc_pages().
2144 static void setup_per_zone_lowmem_reserve(void)
2146 struct pglist_data *pgdat;
2149 for_each_online_pgdat(pgdat) {
2150 for (j = 0; j < MAX_NR_ZONES; j++) {
2151 struct zone *zone = pgdat->node_zones + j;
2152 unsigned long present_pages = zone->present_pages;
2154 zone->lowmem_reserve[j] = 0;
2156 for (idx = j-1; idx >= 0; idx--) {
2157 struct zone *lower_zone;
2159 if (sysctl_lowmem_reserve_ratio[idx] < 1)
2160 sysctl_lowmem_reserve_ratio[idx] = 1;
2162 lower_zone = pgdat->node_zones + idx;
2163 lower_zone->lowmem_reserve[j] = present_pages /
2164 sysctl_lowmem_reserve_ratio[idx];
2165 present_pages += lower_zone->present_pages;
2170 /* update totalreserve_pages */
2171 calculate_totalreserve_pages();
2175 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
2176 * that the pages_{min,low,high} values for each zone are set correctly
2177 * with respect to min_free_kbytes.
2179 void setup_per_zone_pages_min(void)
2181 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
2182 unsigned long lowmem_pages = 0;
2184 unsigned long flags;
2186 /* Calculate total number of !ZONE_HIGHMEM pages */
2187 for_each_zone(zone) {
2188 if (!is_highmem(zone))
2189 lowmem_pages += zone->present_pages;
2192 for_each_zone(zone) {
2195 spin_lock_irqsave(&zone->lru_lock, flags);
2196 tmp = (u64)pages_min * zone->present_pages;
2197 do_div(tmp, lowmem_pages);
2198 if (is_highmem(zone)) {
2200 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
2201 * need highmem pages, so cap pages_min to a small
2204 * The (pages_high-pages_low) and (pages_low-pages_min)
2205 * deltas controls asynch page reclaim, and so should
2206 * not be capped for highmem.
2210 min_pages = zone->present_pages / 1024;
2211 if (min_pages < SWAP_CLUSTER_MAX)
2212 min_pages = SWAP_CLUSTER_MAX;
2213 if (min_pages > 128)
2215 zone->pages_min = min_pages;
2218 * If it's a lowmem zone, reserve a number of pages
2219 * proportionate to the zone's size.
2221 zone->pages_min = tmp;
2224 zone->pages_low = zone->pages_min + (tmp >> 2);
2225 zone->pages_high = zone->pages_min + (tmp >> 1);
2226 spin_unlock_irqrestore(&zone->lru_lock, flags);
2229 /* update totalreserve_pages */
2230 calculate_totalreserve_pages();
2234 * Initialise min_free_kbytes.
2236 * For small machines we want it small (128k min). For large machines
2237 * we want it large (64MB max). But it is not linear, because network
2238 * bandwidth does not increase linearly with machine size. We use
2240 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2241 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2257 static int __init init_per_zone_pages_min(void)
2259 unsigned long lowmem_kbytes;
2261 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2263 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2264 if (min_free_kbytes < 128)
2265 min_free_kbytes = 128;
2266 if (min_free_kbytes > 65536)
2267 min_free_kbytes = 65536;
2268 setup_per_zone_pages_min();
2269 setup_per_zone_lowmem_reserve();
2272 module_init(init_per_zone_pages_min)
2275 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2276 * that we can call two helper functions whenever min_free_kbytes
2279 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2280 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2282 proc_dointvec(table, write, file, buffer, length, ppos);
2283 setup_per_zone_pages_min();
2288 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
2289 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2294 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2299 zone->min_unmapped_ratio = (zone->present_pages *
2300 sysctl_min_unmapped_ratio) / 100;
2306 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2307 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2308 * whenever sysctl_lowmem_reserve_ratio changes.
2310 * The reserve ratio obviously has absolutely no relation with the
2311 * pages_min watermarks. The lowmem reserve ratio can only make sense
2312 * if in function of the boot time zone sizes.
2314 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
2315 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2317 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2318 setup_per_zone_lowmem_reserve();
2323 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
2324 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
2325 * can have before it gets flushed back to buddy allocator.
2328 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
2329 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2335 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2336 if (!write || (ret == -EINVAL))
2338 for_each_zone(zone) {
2339 for_each_online_cpu(cpu) {
2341 high = zone->present_pages / percpu_pagelist_fraction;
2342 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
2348 int hashdist = HASHDIST_DEFAULT;
2351 static int __init set_hashdist(char *str)
2355 hashdist = simple_strtoul(str, &str, 0);
2358 __setup("hashdist=", set_hashdist);
2362 * allocate a large system hash table from bootmem
2363 * - it is assumed that the hash table must contain an exact power-of-2
2364 * quantity of entries
2365 * - limit is the number of hash buckets, not the total allocation size
2367 void *__init alloc_large_system_hash(const char *tablename,
2368 unsigned long bucketsize,
2369 unsigned long numentries,
2372 unsigned int *_hash_shift,
2373 unsigned int *_hash_mask,
2374 unsigned long limit)
2376 unsigned long long max = limit;
2377 unsigned long log2qty, size;
2380 /* allow the kernel cmdline to have a say */
2382 /* round applicable memory size up to nearest megabyte */
2383 numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages;
2384 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2385 numentries >>= 20 - PAGE_SHIFT;
2386 numentries <<= 20 - PAGE_SHIFT;
2388 /* limit to 1 bucket per 2^scale bytes of low memory */
2389 if (scale > PAGE_SHIFT)
2390 numentries >>= (scale - PAGE_SHIFT);
2392 numentries <<= (PAGE_SHIFT - scale);
2394 numentries = roundup_pow_of_two(numentries);
2396 /* limit allocation size to 1/16 total memory by default */
2398 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2399 do_div(max, bucketsize);
2402 if (numentries > max)
2405 log2qty = long_log2(numentries);
2408 size = bucketsize << log2qty;
2409 if (flags & HASH_EARLY)
2410 table = alloc_bootmem(size);
2412 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
2414 unsigned long order;
2415 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
2417 table = (void*) __get_free_pages(GFP_ATOMIC, order);
2419 } while (!table && size > PAGE_SIZE && --log2qty);
2422 panic("Failed to allocate %s hash table\n", tablename);
2424 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2427 long_log2(size) - PAGE_SHIFT,
2431 *_hash_shift = log2qty;
2433 *_hash_mask = (1 << log2qty) - 1;
2438 #ifdef CONFIG_OUT_OF_LINE_PFN_TO_PAGE
2439 struct page *pfn_to_page(unsigned long pfn)
2441 return __pfn_to_page(pfn);
2443 unsigned long page_to_pfn(struct page *page)
2445 return __page_to_pfn(page);
2447 EXPORT_SYMBOL(pfn_to_page);
2448 EXPORT_SYMBOL(page_to_pfn);
2449 #endif /* CONFIG_OUT_OF_LINE_PFN_TO_PAGE */