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>
40 #include <linux/vs_limit.h>
42 #include <asm/tlbflush.h>
43 #include <asm/div64.h>
47 * MCD - HACK: Find somewhere to initialize this EARLY, or make this
50 nodemask_t node_online_map __read_mostly = { { [0] = 1UL } };
51 EXPORT_SYMBOL(node_online_map);
52 nodemask_t node_possible_map __read_mostly = NODE_MASK_ALL;
53 EXPORT_SYMBOL(node_possible_map);
54 unsigned long totalram_pages __read_mostly;
55 unsigned long totalhigh_pages __read_mostly;
56 unsigned long totalreserve_pages __read_mostly;
58 int percpu_pagelist_fraction;
60 static void __free_pages_ok(struct page *page, unsigned int order);
63 * results with 256, 32 in the lowmem_reserve sysctl:
64 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
65 * 1G machine -> (16M dma, 784M normal, 224M high)
66 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
67 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
68 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
70 * TBD: should special case ZONE_DMA32 machines here - in those we normally
71 * don't need any ZONE_NORMAL reservation
73 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 256, 256, 32 };
75 EXPORT_SYMBOL(totalram_pages);
78 * Used by page_zone() to look up the address of the struct zone whose
79 * id is encoded in the upper bits of page->flags
81 struct zone *zone_table[1 << ZONETABLE_SHIFT] __read_mostly;
82 EXPORT_SYMBOL(zone_table);
84 static char *zone_names[MAX_NR_ZONES] = { "DMA", "DMA32", "Normal", "HighMem" };
85 int min_free_kbytes = 1024;
87 unsigned long __meminitdata nr_kernel_pages;
88 unsigned long __meminitdata nr_all_pages;
90 #ifdef CONFIG_DEBUG_VM
91 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
95 unsigned long pfn = page_to_pfn(page);
98 seq = zone_span_seqbegin(zone);
99 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
101 else if (pfn < zone->zone_start_pfn)
103 } while (zone_span_seqretry(zone, seq));
108 static int page_is_consistent(struct zone *zone, struct page *page)
110 #ifdef CONFIG_HOLES_IN_ZONE
111 if (!pfn_valid(page_to_pfn(page)))
114 if (zone != page_zone(page))
120 * Temporary debugging check for pages not lying within a given zone.
122 static int bad_range(struct zone *zone, struct page *page)
124 if (page_outside_zone_boundaries(zone, page))
126 if (!page_is_consistent(zone, page))
133 static inline int bad_range(struct zone *zone, struct page *page)
139 static void bad_page(struct page *page)
141 printk(KERN_EMERG "Bad page state in process '%s'\n"
142 KERN_EMERG "page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d (%s)\n"
143 KERN_EMERG "Trying to fix it up, but a reboot is needed\n"
144 KERN_EMERG "Backtrace:\n",
145 current->comm, page, (int)(2*sizeof(unsigned long)),
146 (unsigned long)page->flags, page->mapping,
147 page_mapcount(page), page_count(page), print_tainted());
149 page->flags &= ~(1 << PG_lru |
159 set_page_count(page, 0);
160 reset_page_mapcount(page);
161 page->mapping = NULL;
162 add_taint(TAINT_BAD_PAGE);
166 * Higher-order pages are called "compound pages". They are structured thusly:
168 * The first PAGE_SIZE page is called the "head page".
170 * The remaining PAGE_SIZE pages are called "tail pages".
172 * All pages have PG_compound set. All pages have their ->private pointing at
173 * the head page (even the head page has this).
175 * The first tail page's ->lru.next holds the address of the compound page's
176 * put_page() function. Its ->lru.prev holds the order of allocation.
177 * This usage means that zero-order pages may not be compound.
180 static void free_compound_page(struct page *page)
182 __free_pages_ok(page, (unsigned long)page[1].lru.prev);
185 static void prep_compound_page(struct page *page, unsigned long order)
188 int nr_pages = 1 << order;
190 page[1].lru.next = (void *)free_compound_page; /* set dtor */
191 page[1].lru.prev = (void *)order;
192 for (i = 0; i < nr_pages; i++) {
193 struct page *p = page + i;
195 __SetPageCompound(p);
196 set_page_private(p, (unsigned long)page);
200 static void destroy_compound_page(struct page *page, unsigned long order)
203 int nr_pages = 1 << order;
205 if (unlikely((unsigned long)page[1].lru.prev != order))
208 for (i = 0; i < nr_pages; i++) {
209 struct page *p = page + i;
211 if (unlikely(!PageCompound(p) |
212 (page_private(p) != (unsigned long)page)))
214 __ClearPageCompound(p);
218 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
222 BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
224 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
225 * and __GFP_HIGHMEM from hard or soft interrupt context.
227 BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
228 for (i = 0; i < (1 << order); i++)
229 clear_highpage(page + i);
233 * function for dealing with page's order in buddy system.
234 * zone->lock is already acquired when we use these.
235 * So, we don't need atomic page->flags operations here.
237 static inline unsigned long page_order(struct page *page)
239 return page_private(page);
242 static inline void set_page_order(struct page *page, int order)
244 set_page_private(page, order);
245 __SetPageBuddy(page);
248 static inline void rmv_page_order(struct page *page)
250 __ClearPageBuddy(page);
251 set_page_private(page, 0);
255 * Locate the struct page for both the matching buddy in our
256 * pair (buddy1) and the combined O(n+1) page they form (page).
258 * 1) Any buddy B1 will have an order O twin B2 which satisfies
259 * the following equation:
261 * For example, if the starting buddy (buddy2) is #8 its order
263 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
265 * 2) Any buddy B will have an order O+1 parent P which
266 * satisfies the following equation:
269 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
271 static inline struct page *
272 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
274 unsigned long buddy_idx = page_idx ^ (1 << order);
276 return page + (buddy_idx - page_idx);
279 static inline unsigned long
280 __find_combined_index(unsigned long page_idx, unsigned int order)
282 return (page_idx & ~(1 << order));
286 * This function checks whether a page is free && is the buddy
287 * we can do coalesce a page and its buddy if
288 * (a) the buddy is not in a hole &&
289 * (b) the buddy is in the buddy system &&
290 * (c) a page and its buddy have the same order &&
291 * (d) a page and its buddy are in the same zone.
293 * For recording whether a page is in the buddy system, we use PG_buddy.
294 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
296 * For recording page's order, we use page_private(page).
298 static inline int page_is_buddy(struct page *page, struct page *buddy,
301 #ifdef CONFIG_HOLES_IN_ZONE
302 if (!pfn_valid(page_to_pfn(buddy)))
306 if (page_zone_id(page) != page_zone_id(buddy))
309 if (PageBuddy(buddy) && page_order(buddy) == order) {
310 BUG_ON(page_count(buddy) != 0);
317 * Freeing function for a buddy system allocator.
319 * The concept of a buddy system is to maintain direct-mapped table
320 * (containing bit values) for memory blocks of various "orders".
321 * The bottom level table contains the map for the smallest allocatable
322 * units of memory (here, pages), and each level above it describes
323 * pairs of units from the levels below, hence, "buddies".
324 * At a high level, all that happens here is marking the table entry
325 * at the bottom level available, and propagating the changes upward
326 * as necessary, plus some accounting needed to play nicely with other
327 * parts of the VM system.
328 * At each level, we keep a list of pages, which are heads of continuous
329 * free pages of length of (1 << order) and marked with PG_buddy. Page's
330 * order is recorded in page_private(page) field.
331 * So when we are allocating or freeing one, we can derive the state of the
332 * other. That is, if we allocate a small block, and both were
333 * free, the remainder of the region must be split into blocks.
334 * If a block is freed, and its buddy is also free, then this
335 * triggers coalescing into a block of larger size.
340 static inline void __free_one_page(struct page *page,
341 struct zone *zone, unsigned int order)
343 unsigned long page_idx;
344 int order_size = 1 << order;
346 if (unlikely(PageCompound(page)))
347 destroy_compound_page(page, order);
349 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
351 BUG_ON(page_idx & (order_size - 1));
352 BUG_ON(bad_range(zone, page));
354 zone->free_pages += order_size;
355 while (order < MAX_ORDER-1) {
356 unsigned long combined_idx;
357 struct free_area *area;
360 buddy = __page_find_buddy(page, page_idx, order);
361 if (!page_is_buddy(page, buddy, order))
362 break; /* Move the buddy up one level. */
364 list_del(&buddy->lru);
365 area = zone->free_area + order;
367 rmv_page_order(buddy);
368 combined_idx = __find_combined_index(page_idx, order);
369 page = page + (combined_idx - page_idx);
370 page_idx = combined_idx;
373 set_page_order(page, order);
374 list_add(&page->lru, &zone->free_area[order].free_list);
375 zone->free_area[order].nr_free++;
378 static inline int free_pages_check(struct page *page)
380 if (unlikely(page_mapcount(page) |
381 (page->mapping != NULL) |
382 (page_count(page) != 0) |
396 __ClearPageDirty(page);
398 * For now, we report if PG_reserved was found set, but do not
399 * clear it, and do not free the page. But we shall soon need
400 * to do more, for when the ZERO_PAGE count wraps negative.
402 return PageReserved(page);
406 * Frees a list of pages.
407 * Assumes all pages on list are in same zone, and of same order.
408 * count is the number of pages to free.
410 * If the zone was previously in an "all pages pinned" state then look to
411 * see if this freeing clears that state.
413 * And clear the zone's pages_scanned counter, to hold off the "all pages are
414 * pinned" detection logic.
416 static void free_pages_bulk(struct zone *zone, int count,
417 struct list_head *list, int order)
419 spin_lock(&zone->lock);
420 zone->all_unreclaimable = 0;
421 zone->pages_scanned = 0;
425 BUG_ON(list_empty(list));
426 page = list_entry(list->prev, struct page, lru);
427 /* have to delete it as __free_one_page list manipulates */
428 list_del(&page->lru);
429 __free_one_page(page, zone, order);
431 spin_unlock(&zone->lock);
434 static void free_one_page(struct zone *zone, struct page *page, int order)
437 list_add(&page->lru, &list);
438 free_pages_bulk(zone, 1, &list, order);
441 static void __free_pages_ok(struct page *page, unsigned int order)
447 if (arch_free_page(page, order))
449 if (!PageHighMem(page))
450 debug_check_no_locks_freed(page_address(page),
453 for (i = 0 ; i < (1 << order) ; ++i)
454 reserved += free_pages_check(page + i);
458 kernel_map_pages(page, 1 << order, 0);
459 local_irq_save(flags);
460 __count_vm_events(PGFREE, 1 << order);
461 free_one_page(page_zone(page), page, order);
462 local_irq_restore(flags);
466 * permit the bootmem allocator to evade page validation on high-order frees
468 void fastcall __init __free_pages_bootmem(struct page *page, unsigned int order)
471 __ClearPageReserved(page);
472 set_page_count(page, 0);
473 set_page_refcounted(page);
479 for (loop = 0; loop < BITS_PER_LONG; loop++) {
480 struct page *p = &page[loop];
482 if (loop + 1 < BITS_PER_LONG)
484 __ClearPageReserved(p);
485 set_page_count(p, 0);
488 set_page_refcounted(page);
489 __free_pages(page, order);
495 * The order of subdivision here is critical for the IO subsystem.
496 * Please do not alter this order without good reasons and regression
497 * testing. Specifically, as large blocks of memory are subdivided,
498 * the order in which smaller blocks are delivered depends on the order
499 * they're subdivided in this function. This is the primary factor
500 * influencing the order in which pages are delivered to the IO
501 * subsystem according to empirical testing, and this is also justified
502 * by considering the behavior of a buddy system containing a single
503 * large block of memory acted on by a series of small allocations.
504 * This behavior is a critical factor in sglist merging's success.
508 static inline void expand(struct zone *zone, struct page *page,
509 int low, int high, struct free_area *area)
511 unsigned long size = 1 << high;
517 BUG_ON(bad_range(zone, &page[size]));
518 list_add(&page[size].lru, &area->free_list);
520 set_page_order(&page[size], high);
525 * This page is about to be returned from the page allocator
527 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
529 if (unlikely(page_mapcount(page) |
530 (page->mapping != NULL) |
531 (page_count(page) != 0) |
547 * For now, we report if PG_reserved was found set, but do not
548 * clear it, and do not allocate the page: as a safety net.
550 if (PageReserved(page))
553 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
554 1 << PG_referenced | 1 << PG_arch_1 |
555 1 << PG_fs_misc | 1 << PG_mappedtodisk);
556 set_page_private(page, 0);
557 set_page_refcounted(page);
558 kernel_map_pages(page, 1 << order, 1);
560 if (gfp_flags & __GFP_ZERO)
561 prep_zero_page(page, order, gfp_flags);
563 if (order && (gfp_flags & __GFP_COMP))
564 prep_compound_page(page, order);
570 * Do the hard work of removing an element from the buddy allocator.
571 * Call me with the zone->lock already held.
573 static struct page *__rmqueue(struct zone *zone, unsigned int order)
575 struct free_area * area;
576 unsigned int current_order;
579 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
580 area = zone->free_area + current_order;
581 if (list_empty(&area->free_list))
584 page = list_entry(area->free_list.next, struct page, lru);
585 list_del(&page->lru);
586 rmv_page_order(page);
588 zone->free_pages -= 1UL << order;
589 expand(zone, page, order, current_order, area);
597 * Obtain a specified number of elements from the buddy allocator, all under
598 * a single hold of the lock, for efficiency. Add them to the supplied list.
599 * Returns the number of new pages which were placed at *list.
601 static int rmqueue_bulk(struct zone *zone, unsigned int order,
602 unsigned long count, struct list_head *list)
606 spin_lock(&zone->lock);
607 for (i = 0; i < count; ++i) {
608 struct page *page = __rmqueue(zone, order);
609 if (unlikely(page == NULL))
611 list_add_tail(&page->lru, list);
613 spin_unlock(&zone->lock);
619 * Called from the slab reaper to drain pagesets on a particular node that
620 * belong to the currently executing processor.
621 * Note that this function must be called with the thread pinned to
622 * a single processor.
624 void drain_node_pages(int nodeid)
629 for (z = 0; z < MAX_NR_ZONES; z++) {
630 struct zone *zone = NODE_DATA(nodeid)->node_zones + z;
631 struct per_cpu_pageset *pset;
633 pset = zone_pcp(zone, smp_processor_id());
634 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
635 struct per_cpu_pages *pcp;
639 local_irq_save(flags);
640 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
642 local_irq_restore(flags);
649 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
650 static void __drain_pages(unsigned int cpu)
656 for_each_zone(zone) {
657 struct per_cpu_pageset *pset;
659 pset = zone_pcp(zone, cpu);
660 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
661 struct per_cpu_pages *pcp;
664 local_irq_save(flags);
665 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
667 local_irq_restore(flags);
671 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
675 void mark_free_pages(struct zone *zone)
677 unsigned long zone_pfn, flags;
679 struct list_head *curr;
681 if (!zone->spanned_pages)
684 spin_lock_irqsave(&zone->lock, flags);
685 for (zone_pfn = 0; zone_pfn < zone->spanned_pages; ++zone_pfn)
686 ClearPageNosaveFree(pfn_to_page(zone_pfn + zone->zone_start_pfn));
688 for (order = MAX_ORDER - 1; order >= 0; --order)
689 list_for_each(curr, &zone->free_area[order].free_list) {
690 unsigned long start_pfn, i;
692 start_pfn = page_to_pfn(list_entry(curr, struct page, lru));
694 for (i=0; i < (1<<order); i++)
695 SetPageNosaveFree(pfn_to_page(start_pfn+i));
697 spin_unlock_irqrestore(&zone->lock, flags);
701 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
703 void drain_local_pages(void)
707 local_irq_save(flags);
708 __drain_pages(smp_processor_id());
709 local_irq_restore(flags);
711 #endif /* CONFIG_PM */
714 * Free a 0-order page
716 static void fastcall free_hot_cold_page(struct page *page, int cold)
718 struct zone *zone = page_zone(page);
719 struct per_cpu_pages *pcp;
722 if (arch_free_page(page, 0))
726 page->mapping = NULL;
727 if (free_pages_check(page))
730 kernel_map_pages(page, 1, 0);
732 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
733 local_irq_save(flags);
734 __count_vm_event(PGFREE);
735 list_add(&page->lru, &pcp->list);
737 if (pcp->count >= pcp->high) {
738 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
739 pcp->count -= pcp->batch;
741 local_irq_restore(flags);
745 void fastcall free_hot_page(struct page *page)
747 free_hot_cold_page(page, 0);
750 void fastcall free_cold_page(struct page *page)
752 free_hot_cold_page(page, 1);
756 * split_page takes a non-compound higher-order page, and splits it into
757 * n (1<<order) sub-pages: page[0..n]
758 * Each sub-page must be freed individually.
760 * Note: this is probably too low level an operation for use in drivers.
761 * Please consult with lkml before using this in your driver.
763 void split_page(struct page *page, unsigned int order)
767 BUG_ON(PageCompound(page));
768 BUG_ON(!page_count(page));
769 for (i = 1; i < (1 << order); i++)
770 set_page_refcounted(page + i);
774 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
775 * we cheat by calling it from here, in the order > 0 path. Saves a branch
778 static struct page *buffered_rmqueue(struct zonelist *zonelist,
779 struct zone *zone, int order, gfp_t gfp_flags)
783 int cold = !!(gfp_flags & __GFP_COLD);
788 if (likely(order == 0)) {
789 struct per_cpu_pages *pcp;
791 pcp = &zone_pcp(zone, cpu)->pcp[cold];
792 local_irq_save(flags);
794 pcp->count += rmqueue_bulk(zone, 0,
795 pcp->batch, &pcp->list);
796 if (unlikely(!pcp->count))
799 page = list_entry(pcp->list.next, struct page, lru);
800 list_del(&page->lru);
803 spin_lock_irqsave(&zone->lock, flags);
804 page = __rmqueue(zone, order);
805 spin_unlock(&zone->lock);
810 __count_zone_vm_events(PGALLOC, zone, 1 << order);
811 zone_statistics(zonelist, zone);
812 local_irq_restore(flags);
815 BUG_ON(bad_range(zone, page));
816 if (prep_new_page(page, order, gfp_flags))
821 local_irq_restore(flags);
826 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
827 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
828 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
829 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
830 #define ALLOC_HARDER 0x10 /* try to alloc harder */
831 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
832 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
835 * Return 1 if free pages are above 'mark'. This takes into account the order
838 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
839 int classzone_idx, int alloc_flags)
841 /* free_pages my go negative - that's OK */
842 long min = mark, free_pages = z->free_pages - (1 << order) + 1;
845 if (alloc_flags & ALLOC_HIGH)
847 if (alloc_flags & ALLOC_HARDER)
850 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
852 for (o = 0; o < order; o++) {
853 /* At the next order, this order's pages become unavailable */
854 free_pages -= z->free_area[o].nr_free << o;
856 /* Require fewer higher order pages to be free */
859 if (free_pages <= min)
866 * get_page_from_freeliest goes through the zonelist trying to allocate
870 get_page_from_freelist(gfp_t gfp_mask, unsigned int order,
871 struct zonelist *zonelist, int alloc_flags)
873 struct zone **z = zonelist->zones;
874 struct page *page = NULL;
875 int classzone_idx = zone_idx(*z);
878 * Go through the zonelist once, looking for a zone with enough free.
879 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
882 if ((alloc_flags & ALLOC_CPUSET) &&
883 !cpuset_zone_allowed(*z, gfp_mask))
886 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
888 if (alloc_flags & ALLOC_WMARK_MIN)
889 mark = (*z)->pages_min;
890 else if (alloc_flags & ALLOC_WMARK_LOW)
891 mark = (*z)->pages_low;
893 mark = (*z)->pages_high;
894 if (!zone_watermark_ok(*z, order, mark,
895 classzone_idx, alloc_flags))
896 if (!zone_reclaim_mode ||
897 !zone_reclaim(*z, gfp_mask, order))
901 page = buffered_rmqueue(zonelist, *z, order, gfp_mask);
905 } while (*(++z) != NULL);
910 * This is the 'heart' of the zoned buddy allocator.
912 struct page * fastcall
913 __alloc_pages(gfp_t gfp_mask, unsigned int order,
914 struct zonelist *zonelist)
916 const gfp_t wait = gfp_mask & __GFP_WAIT;
919 struct reclaim_state reclaim_state;
920 struct task_struct *p = current;
923 int did_some_progress;
925 might_sleep_if(wait);
928 z = zonelist->zones; /* the list of zones suitable for gfp_mask */
930 if (unlikely(*z == NULL)) {
931 /* Should this ever happen?? */
935 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
936 zonelist, ALLOC_WMARK_LOW|ALLOC_CPUSET);
941 wakeup_kswapd(*z, order);
945 * OK, we're below the kswapd watermark and have kicked background
946 * reclaim. Now things get more complex, so set up alloc_flags according
947 * to how we want to proceed.
949 * The caller may dip into page reserves a bit more if the caller
950 * cannot run direct reclaim, or if the caller has realtime scheduling
951 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
952 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
954 alloc_flags = ALLOC_WMARK_MIN;
955 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
956 alloc_flags |= ALLOC_HARDER;
957 if (gfp_mask & __GFP_HIGH)
958 alloc_flags |= ALLOC_HIGH;
960 alloc_flags |= ALLOC_CPUSET;
963 * Go through the zonelist again. Let __GFP_HIGH and allocations
964 * coming from realtime tasks go deeper into reserves.
966 * This is the last chance, in general, before the goto nopage.
967 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
968 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
970 page = get_page_from_freelist(gfp_mask, order, zonelist, alloc_flags);
974 /* This allocation should allow future memory freeing. */
976 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
977 && !in_interrupt()) {
978 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
980 /* go through the zonelist yet again, ignoring mins */
981 page = get_page_from_freelist(gfp_mask, order,
982 zonelist, ALLOC_NO_WATERMARKS);
985 if (gfp_mask & __GFP_NOFAIL) {
986 blk_congestion_wait(WRITE, HZ/50);
993 /* Atomic allocations - we can't balance anything */
1000 /* We now go into synchronous reclaim */
1001 cpuset_memory_pressure_bump();
1002 p->flags |= PF_MEMALLOC;
1003 reclaim_state.reclaimed_slab = 0;
1004 p->reclaim_state = &reclaim_state;
1006 did_some_progress = try_to_free_pages(zonelist->zones, gfp_mask);
1008 p->reclaim_state = NULL;
1009 p->flags &= ~PF_MEMALLOC;
1013 if (likely(did_some_progress)) {
1014 page = get_page_from_freelist(gfp_mask, order,
1015 zonelist, alloc_flags);
1018 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1020 * Go through the zonelist yet one more time, keep
1021 * very high watermark here, this is only to catch
1022 * a parallel oom killing, we must fail if we're still
1023 * under heavy pressure.
1025 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1026 zonelist, ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1030 out_of_memory(zonelist, gfp_mask, order);
1035 * Don't let big-order allocations loop unless the caller explicitly
1036 * requests that. Wait for some write requests to complete then retry.
1038 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
1039 * <= 3, but that may not be true in other implementations.
1042 if (!(gfp_mask & __GFP_NORETRY)) {
1043 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
1045 if (gfp_mask & __GFP_NOFAIL)
1049 blk_congestion_wait(WRITE, HZ/50);
1054 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1055 printk(KERN_WARNING "%s: page allocation failure."
1056 " order:%d, mode:0x%x\n",
1057 p->comm, order, gfp_mask);
1065 EXPORT_SYMBOL(__alloc_pages);
1068 * Common helper functions.
1070 fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1073 page = alloc_pages(gfp_mask, order);
1076 return (unsigned long) page_address(page);
1079 EXPORT_SYMBOL(__get_free_pages);
1081 fastcall unsigned long get_zeroed_page(gfp_t gfp_mask)
1086 * get_zeroed_page() returns a 32-bit address, which cannot represent
1089 BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1091 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1093 return (unsigned long) page_address(page);
1097 EXPORT_SYMBOL(get_zeroed_page);
1099 void __pagevec_free(struct pagevec *pvec)
1101 int i = pagevec_count(pvec);
1104 free_hot_cold_page(pvec->pages[i], pvec->cold);
1107 fastcall void __free_pages(struct page *page, unsigned int order)
1109 if (put_page_testzero(page)) {
1111 free_hot_page(page);
1113 __free_pages_ok(page, order);
1117 EXPORT_SYMBOL(__free_pages);
1119 fastcall void free_pages(unsigned long addr, unsigned int order)
1122 BUG_ON(!virt_addr_valid((void *)addr));
1123 __free_pages(virt_to_page((void *)addr), order);
1127 EXPORT_SYMBOL(free_pages);
1130 * Total amount of free (allocatable) RAM:
1132 unsigned int nr_free_pages(void)
1134 unsigned int sum = 0;
1138 sum += zone->free_pages;
1143 EXPORT_SYMBOL(nr_free_pages);
1146 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
1148 unsigned int i, sum = 0;
1150 for (i = 0; i < MAX_NR_ZONES; i++)
1151 sum += pgdat->node_zones[i].free_pages;
1157 static unsigned int nr_free_zone_pages(int offset)
1159 /* Just pick one node, since fallback list is circular */
1160 pg_data_t *pgdat = NODE_DATA(numa_node_id());
1161 unsigned int sum = 0;
1163 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1164 struct zone **zonep = zonelist->zones;
1167 for (zone = *zonep++; zone; zone = *zonep++) {
1168 unsigned long size = zone->present_pages;
1169 unsigned long high = zone->pages_high;
1178 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1180 unsigned int nr_free_buffer_pages(void)
1182 return nr_free_zone_pages(gfp_zone(GFP_USER));
1186 * Amount of free RAM allocatable within all zones
1188 unsigned int nr_free_pagecache_pages(void)
1190 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER));
1193 #ifdef CONFIG_HIGHMEM
1194 unsigned int nr_free_highpages (void)
1197 unsigned int pages = 0;
1199 for_each_online_pgdat(pgdat)
1200 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1207 static void show_node(struct zone *zone)
1209 printk("Node %d ", zone->zone_pgdat->node_id);
1212 #define show_node(zone) do { } while (0)
1215 void si_meminfo(struct sysinfo *val)
1217 val->totalram = totalram_pages;
1219 val->freeram = nr_free_pages();
1220 val->bufferram = nr_blockdev_pages();
1221 #ifdef CONFIG_HIGHMEM
1222 val->totalhigh = totalhigh_pages;
1223 val->freehigh = nr_free_highpages();
1228 val->mem_unit = PAGE_SIZE;
1229 if (vx_flags(VXF_VIRT_MEM, 0))
1230 vx_vsi_meminfo(val);
1233 EXPORT_SYMBOL(si_meminfo);
1236 void si_meminfo_node(struct sysinfo *val, int nid)
1238 pg_data_t *pgdat = NODE_DATA(nid);
1240 val->totalram = pgdat->node_present_pages;
1241 val->freeram = nr_free_pages_pgdat(pgdat);
1242 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1243 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1244 val->mem_unit = PAGE_SIZE;
1245 if (vx_flags(VXF_VIRT_MEM, 0))
1246 vx_vsi_meminfo(val);
1250 #define K(x) ((x) << (PAGE_SHIFT-10))
1253 * Show free area list (used inside shift_scroll-lock stuff)
1254 * We also calculate the percentage fragmentation. We do this by counting the
1255 * memory on each free list with the exception of the first item on the list.
1257 void show_free_areas(void)
1259 int cpu, temperature;
1260 unsigned long active;
1261 unsigned long inactive;
1265 for_each_zone(zone) {
1267 printk("%s per-cpu:", zone->name);
1269 if (!populated_zone(zone)) {
1275 for_each_online_cpu(cpu) {
1276 struct per_cpu_pageset *pageset;
1278 pageset = zone_pcp(zone, cpu);
1280 for (temperature = 0; temperature < 2; temperature++)
1281 printk("cpu %d %s: high %d, batch %d used:%d\n",
1283 temperature ? "cold" : "hot",
1284 pageset->pcp[temperature].high,
1285 pageset->pcp[temperature].batch,
1286 pageset->pcp[temperature].count);
1290 get_zone_counts(&active, &inactive, &free);
1292 printk("Free pages: %11ukB (%ukB HighMem)\n",
1294 K(nr_free_highpages()));
1296 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1297 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1300 global_page_state(NR_FILE_DIRTY),
1301 global_page_state(NR_WRITEBACK),
1302 global_page_state(NR_UNSTABLE_NFS),
1304 global_page_state(NR_SLAB),
1305 global_page_state(NR_FILE_MAPPED),
1306 global_page_state(NR_PAGETABLE));
1308 for_each_zone(zone) {
1320 " pages_scanned:%lu"
1321 " all_unreclaimable? %s"
1324 K(zone->free_pages),
1327 K(zone->pages_high),
1329 K(zone->nr_inactive),
1330 K(zone->present_pages),
1331 zone->pages_scanned,
1332 (zone->all_unreclaimable ? "yes" : "no")
1334 printk("lowmem_reserve[]:");
1335 for (i = 0; i < MAX_NR_ZONES; i++)
1336 printk(" %lu", zone->lowmem_reserve[i]);
1340 for_each_zone(zone) {
1341 unsigned long nr[MAX_ORDER], flags, order, total = 0;
1344 printk("%s: ", zone->name);
1345 if (!populated_zone(zone)) {
1350 spin_lock_irqsave(&zone->lock, flags);
1351 for (order = 0; order < MAX_ORDER; order++) {
1352 nr[order] = zone->free_area[order].nr_free;
1353 total += nr[order] << order;
1355 spin_unlock_irqrestore(&zone->lock, flags);
1356 for (order = 0; order < MAX_ORDER; order++)
1357 printk("%lu*%lukB ", nr[order], K(1UL) << order);
1358 printk("= %lukB\n", K(total));
1361 show_swap_cache_info();
1365 * Builds allocation fallback zone lists.
1367 * Add all populated zones of a node to the zonelist.
1369 static int __meminit build_zonelists_node(pg_data_t *pgdat,
1370 struct zonelist *zonelist, int nr_zones, int zone_type)
1374 BUG_ON(zone_type > ZONE_HIGHMEM);
1377 zone = pgdat->node_zones + zone_type;
1378 if (populated_zone(zone)) {
1379 #ifndef CONFIG_HIGHMEM
1380 BUG_ON(zone_type > ZONE_NORMAL);
1382 zonelist->zones[nr_zones++] = zone;
1383 check_highest_zone(zone_type);
1387 } while (zone_type >= 0);
1391 static inline int highest_zone(int zone_bits)
1393 int res = ZONE_NORMAL;
1394 if (zone_bits & (__force int)__GFP_HIGHMEM)
1396 if (zone_bits & (__force int)__GFP_DMA32)
1398 if (zone_bits & (__force int)__GFP_DMA)
1404 #define MAX_NODE_LOAD (num_online_nodes())
1405 static int __meminitdata node_load[MAX_NUMNODES];
1407 * find_next_best_node - find the next node that should appear in a given node's fallback list
1408 * @node: node whose fallback list we're appending
1409 * @used_node_mask: nodemask_t of already used nodes
1411 * We use a number of factors to determine which is the next node that should
1412 * appear on a given node's fallback list. The node should not have appeared
1413 * already in @node's fallback list, and it should be the next closest node
1414 * according to the distance array (which contains arbitrary distance values
1415 * from each node to each node in the system), and should also prefer nodes
1416 * with no CPUs, since presumably they'll have very little allocation pressure
1417 * on them otherwise.
1418 * It returns -1 if no node is found.
1420 static int __meminit find_next_best_node(int node, nodemask_t *used_node_mask)
1423 int min_val = INT_MAX;
1426 /* Use the local node if we haven't already */
1427 if (!node_isset(node, *used_node_mask)) {
1428 node_set(node, *used_node_mask);
1432 for_each_online_node(n) {
1435 /* Don't want a node to appear more than once */
1436 if (node_isset(n, *used_node_mask))
1439 /* Use the distance array to find the distance */
1440 val = node_distance(node, n);
1442 /* Penalize nodes under us ("prefer the next node") */
1445 /* Give preference to headless and unused nodes */
1446 tmp = node_to_cpumask(n);
1447 if (!cpus_empty(tmp))
1448 val += PENALTY_FOR_NODE_WITH_CPUS;
1450 /* Slight preference for less loaded node */
1451 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1452 val += node_load[n];
1454 if (val < min_val) {
1461 node_set(best_node, *used_node_mask);
1466 static void __meminit build_zonelists(pg_data_t *pgdat)
1468 int i, j, k, node, local_node;
1469 int prev_node, load;
1470 struct zonelist *zonelist;
1471 nodemask_t used_mask;
1473 /* initialize zonelists */
1474 for (i = 0; i < GFP_ZONETYPES; i++) {
1475 zonelist = pgdat->node_zonelists + i;
1476 zonelist->zones[0] = NULL;
1479 /* NUMA-aware ordering of nodes */
1480 local_node = pgdat->node_id;
1481 load = num_online_nodes();
1482 prev_node = local_node;
1483 nodes_clear(used_mask);
1484 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1485 int distance = node_distance(local_node, node);
1488 * If another node is sufficiently far away then it is better
1489 * to reclaim pages in a zone before going off node.
1491 if (distance > RECLAIM_DISTANCE)
1492 zone_reclaim_mode = 1;
1495 * We don't want to pressure a particular node.
1496 * So adding penalty to the first node in same
1497 * distance group to make it round-robin.
1500 if (distance != node_distance(local_node, prev_node))
1501 node_load[node] += load;
1504 for (i = 0; i < GFP_ZONETYPES; i++) {
1505 zonelist = pgdat->node_zonelists + i;
1506 for (j = 0; zonelist->zones[j] != NULL; j++);
1508 k = highest_zone(i);
1510 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1511 zonelist->zones[j] = NULL;
1516 #else /* CONFIG_NUMA */
1518 static void __meminit build_zonelists(pg_data_t *pgdat)
1520 int i, j, k, node, local_node;
1522 local_node = pgdat->node_id;
1523 for (i = 0; i < GFP_ZONETYPES; i++) {
1524 struct zonelist *zonelist;
1526 zonelist = pgdat->node_zonelists + i;
1529 k = highest_zone(i);
1530 j = build_zonelists_node(pgdat, zonelist, j, k);
1532 * Now we build the zonelist so that it contains the zones
1533 * of all the other nodes.
1534 * We don't want to pressure a particular node, so when
1535 * building the zones for node N, we make sure that the
1536 * zones coming right after the local ones are those from
1537 * node N+1 (modulo N)
1539 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1540 if (!node_online(node))
1542 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1544 for (node = 0; node < local_node; node++) {
1545 if (!node_online(node))
1547 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1550 zonelist->zones[j] = NULL;
1554 #endif /* CONFIG_NUMA */
1556 /* return values int ....just for stop_machine_run() */
1557 static int __meminit __build_all_zonelists(void *dummy)
1560 for_each_online_node(nid)
1561 build_zonelists(NODE_DATA(nid));
1565 void __meminit build_all_zonelists(void)
1567 if (system_state == SYSTEM_BOOTING) {
1568 __build_all_zonelists(0);
1569 cpuset_init_current_mems_allowed();
1571 /* we have to stop all cpus to guaranntee there is no user
1573 stop_machine_run(__build_all_zonelists, NULL, NR_CPUS);
1574 /* cpuset refresh routine should be here */
1576 vm_total_pages = nr_free_pagecache_pages();
1577 printk("Built %i zonelists. Total pages: %ld\n",
1578 num_online_nodes(), vm_total_pages);
1582 * Helper functions to size the waitqueue hash table.
1583 * Essentially these want to choose hash table sizes sufficiently
1584 * large so that collisions trying to wait on pages are rare.
1585 * But in fact, the number of active page waitqueues on typical
1586 * systems is ridiculously low, less than 200. So this is even
1587 * conservative, even though it seems large.
1589 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1590 * waitqueues, i.e. the size of the waitq table given the number of pages.
1592 #define PAGES_PER_WAITQUEUE 256
1594 #ifndef CONFIG_MEMORY_HOTPLUG
1595 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
1597 unsigned long size = 1;
1599 pages /= PAGES_PER_WAITQUEUE;
1601 while (size < pages)
1605 * Once we have dozens or even hundreds of threads sleeping
1606 * on IO we've got bigger problems than wait queue collision.
1607 * Limit the size of the wait table to a reasonable size.
1609 size = min(size, 4096UL);
1611 return max(size, 4UL);
1615 * A zone's size might be changed by hot-add, so it is not possible to determine
1616 * a suitable size for its wait_table. So we use the maximum size now.
1618 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
1620 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
1621 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
1622 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
1624 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
1625 * or more by the traditional way. (See above). It equals:
1627 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
1628 * ia64(16K page size) : = ( 8G + 4M)byte.
1629 * powerpc (64K page size) : = (32G +16M)byte.
1631 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
1638 * This is an integer logarithm so that shifts can be used later
1639 * to extract the more random high bits from the multiplicative
1640 * hash function before the remainder is taken.
1642 static inline unsigned long wait_table_bits(unsigned long size)
1647 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1649 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1650 unsigned long *zones_size, unsigned long *zholes_size)
1652 unsigned long realtotalpages, totalpages = 0;
1655 for (i = 0; i < MAX_NR_ZONES; i++)
1656 totalpages += zones_size[i];
1657 pgdat->node_spanned_pages = totalpages;
1659 realtotalpages = totalpages;
1661 for (i = 0; i < MAX_NR_ZONES; i++)
1662 realtotalpages -= zholes_size[i];
1663 pgdat->node_present_pages = realtotalpages;
1664 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1669 * Initially all pages are reserved - free ones are freed
1670 * up by free_all_bootmem() once the early boot process is
1671 * done. Non-atomic initialization, single-pass.
1673 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1674 unsigned long start_pfn)
1677 unsigned long end_pfn = start_pfn + size;
1680 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1681 if (!early_pfn_valid(pfn))
1683 if (!early_pfn_in_nid(pfn, nid))
1685 page = pfn_to_page(pfn);
1686 set_page_links(page, zone, nid, pfn);
1687 init_page_count(page);
1688 reset_page_mapcount(page);
1689 SetPageReserved(page);
1690 INIT_LIST_HEAD(&page->lru);
1691 #ifdef WANT_PAGE_VIRTUAL
1692 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1693 if (!is_highmem_idx(zone))
1694 set_page_address(page, __va(pfn << PAGE_SHIFT));
1699 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1703 for (order = 0; order < MAX_ORDER ; order++) {
1704 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1705 zone->free_area[order].nr_free = 0;
1709 #define ZONETABLE_INDEX(x, zone_nr) ((x << ZONES_SHIFT) | zone_nr)
1710 void zonetable_add(struct zone *zone, int nid, int zid, unsigned long pfn,
1713 unsigned long snum = pfn_to_section_nr(pfn);
1714 unsigned long end = pfn_to_section_nr(pfn + size);
1717 zone_table[ZONETABLE_INDEX(nid, zid)] = zone;
1719 for (; snum <= end; snum++)
1720 zone_table[ZONETABLE_INDEX(snum, zid)] = zone;
1723 #ifndef __HAVE_ARCH_MEMMAP_INIT
1724 #define memmap_init(size, nid, zone, start_pfn) \
1725 memmap_init_zone((size), (nid), (zone), (start_pfn))
1728 static int __cpuinit zone_batchsize(struct zone *zone)
1733 * The per-cpu-pages pools are set to around 1000th of the
1734 * size of the zone. But no more than 1/2 of a meg.
1736 * OK, so we don't know how big the cache is. So guess.
1738 batch = zone->present_pages / 1024;
1739 if (batch * PAGE_SIZE > 512 * 1024)
1740 batch = (512 * 1024) / PAGE_SIZE;
1741 batch /= 4; /* We effectively *= 4 below */
1746 * Clamp the batch to a 2^n - 1 value. Having a power
1747 * of 2 value was found to be more likely to have
1748 * suboptimal cache aliasing properties in some cases.
1750 * For example if 2 tasks are alternately allocating
1751 * batches of pages, one task can end up with a lot
1752 * of pages of one half of the possible page colors
1753 * and the other with pages of the other colors.
1755 batch = (1 << (fls(batch + batch/2)-1)) - 1;
1760 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
1762 struct per_cpu_pages *pcp;
1764 memset(p, 0, sizeof(*p));
1766 pcp = &p->pcp[0]; /* hot */
1768 pcp->high = 6 * batch;
1769 pcp->batch = max(1UL, 1 * batch);
1770 INIT_LIST_HEAD(&pcp->list);
1772 pcp = &p->pcp[1]; /* cold*/
1774 pcp->high = 2 * batch;
1775 pcp->batch = max(1UL, batch/2);
1776 INIT_LIST_HEAD(&pcp->list);
1780 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
1781 * to the value high for the pageset p.
1784 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
1787 struct per_cpu_pages *pcp;
1789 pcp = &p->pcp[0]; /* hot list */
1791 pcp->batch = max(1UL, high/4);
1792 if ((high/4) > (PAGE_SHIFT * 8))
1793 pcp->batch = PAGE_SHIFT * 8;
1799 * Boot pageset table. One per cpu which is going to be used for all
1800 * zones and all nodes. The parameters will be set in such a way
1801 * that an item put on a list will immediately be handed over to
1802 * the buddy list. This is safe since pageset manipulation is done
1803 * with interrupts disabled.
1805 * Some NUMA counter updates may also be caught by the boot pagesets.
1807 * The boot_pagesets must be kept even after bootup is complete for
1808 * unused processors and/or zones. They do play a role for bootstrapping
1809 * hotplugged processors.
1811 * zoneinfo_show() and maybe other functions do
1812 * not check if the processor is online before following the pageset pointer.
1813 * Other parts of the kernel may not check if the zone is available.
1815 static struct per_cpu_pageset boot_pageset[NR_CPUS];
1818 * Dynamically allocate memory for the
1819 * per cpu pageset array in struct zone.
1821 static int __cpuinit process_zones(int cpu)
1823 struct zone *zone, *dzone;
1825 for_each_zone(zone) {
1827 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
1828 GFP_KERNEL, cpu_to_node(cpu));
1829 if (!zone_pcp(zone, cpu))
1832 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
1834 if (percpu_pagelist_fraction)
1835 setup_pagelist_highmark(zone_pcp(zone, cpu),
1836 (zone->present_pages / percpu_pagelist_fraction));
1841 for_each_zone(dzone) {
1844 kfree(zone_pcp(dzone, cpu));
1845 zone_pcp(dzone, cpu) = NULL;
1850 static inline void free_zone_pagesets(int cpu)
1854 for_each_zone(zone) {
1855 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
1857 /* Free per_cpu_pageset if it is slab allocated */
1858 if (pset != &boot_pageset[cpu])
1860 zone_pcp(zone, cpu) = NULL;
1864 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
1865 unsigned long action,
1868 int cpu = (long)hcpu;
1869 int ret = NOTIFY_OK;
1872 case CPU_UP_PREPARE:
1873 if (process_zones(cpu))
1876 case CPU_UP_CANCELED:
1878 free_zone_pagesets(cpu);
1886 static struct notifier_block __cpuinitdata pageset_notifier =
1887 { &pageset_cpuup_callback, NULL, 0 };
1889 void __init setup_per_cpu_pageset(void)
1893 /* Initialize per_cpu_pageset for cpu 0.
1894 * A cpuup callback will do this for every cpu
1895 * as it comes online
1897 err = process_zones(smp_processor_id());
1899 register_cpu_notifier(&pageset_notifier);
1905 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
1908 struct pglist_data *pgdat = zone->zone_pgdat;
1912 * The per-page waitqueue mechanism uses hashed waitqueues
1915 zone->wait_table_hash_nr_entries =
1916 wait_table_hash_nr_entries(zone_size_pages);
1917 zone->wait_table_bits =
1918 wait_table_bits(zone->wait_table_hash_nr_entries);
1919 alloc_size = zone->wait_table_hash_nr_entries
1920 * sizeof(wait_queue_head_t);
1922 if (system_state == SYSTEM_BOOTING) {
1923 zone->wait_table = (wait_queue_head_t *)
1924 alloc_bootmem_node(pgdat, alloc_size);
1927 * This case means that a zone whose size was 0 gets new memory
1928 * via memory hot-add.
1929 * But it may be the case that a new node was hot-added. In
1930 * this case vmalloc() will not be able to use this new node's
1931 * memory - this wait_table must be initialized to use this new
1932 * node itself as well.
1933 * To use this new node's memory, further consideration will be
1936 zone->wait_table = (wait_queue_head_t *)vmalloc(alloc_size);
1938 if (!zone->wait_table)
1941 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
1942 init_waitqueue_head(zone->wait_table + i);
1947 static __meminit void zone_pcp_init(struct zone *zone)
1950 unsigned long batch = zone_batchsize(zone);
1952 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1954 /* Early boot. Slab allocator not functional yet */
1955 zone_pcp(zone, cpu) = &boot_pageset[cpu];
1956 setup_pageset(&boot_pageset[cpu],0);
1958 setup_pageset(zone_pcp(zone,cpu), batch);
1961 if (zone->present_pages)
1962 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
1963 zone->name, zone->present_pages, batch);
1966 __meminit int init_currently_empty_zone(struct zone *zone,
1967 unsigned long zone_start_pfn,
1970 struct pglist_data *pgdat = zone->zone_pgdat;
1972 ret = zone_wait_table_init(zone, size);
1975 pgdat->nr_zones = zone_idx(zone) + 1;
1977 zone->zone_start_pfn = zone_start_pfn;
1979 memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn);
1981 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1987 * Set up the zone data structures:
1988 * - mark all pages reserved
1989 * - mark all memory queues empty
1990 * - clear the memory bitmaps
1992 static void __meminit free_area_init_core(struct pglist_data *pgdat,
1993 unsigned long *zones_size, unsigned long *zholes_size)
1996 int nid = pgdat->node_id;
1997 unsigned long zone_start_pfn = pgdat->node_start_pfn;
2000 pgdat_resize_init(pgdat);
2001 pgdat->nr_zones = 0;
2002 init_waitqueue_head(&pgdat->kswapd_wait);
2003 pgdat->kswapd_max_order = 0;
2005 for (j = 0; j < MAX_NR_ZONES; j++) {
2006 struct zone *zone = pgdat->node_zones + j;
2007 unsigned long size, realsize;
2009 realsize = size = zones_size[j];
2011 realsize -= zholes_size[j];
2013 if (j < ZONE_HIGHMEM)
2014 nr_kernel_pages += realsize;
2015 nr_all_pages += realsize;
2017 zone->spanned_pages = size;
2018 zone->present_pages = realsize;
2020 zone->min_unmapped_ratio = (realsize*sysctl_min_unmapped_ratio)
2022 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
2024 zone->name = zone_names[j];
2025 spin_lock_init(&zone->lock);
2026 spin_lock_init(&zone->lru_lock);
2027 zone_seqlock_init(zone);
2028 zone->zone_pgdat = pgdat;
2029 zone->free_pages = 0;
2031 zone->prev_priority = DEF_PRIORITY;
2033 zone_pcp_init(zone);
2034 INIT_LIST_HEAD(&zone->active_list);
2035 INIT_LIST_HEAD(&zone->inactive_list);
2036 zone->nr_scan_active = 0;
2037 zone->nr_scan_inactive = 0;
2038 zone->nr_active = 0;
2039 zone->nr_inactive = 0;
2040 zap_zone_vm_stats(zone);
2041 atomic_set(&zone->reclaim_in_progress, 0);
2045 zonetable_add(zone, nid, j, zone_start_pfn, size);
2046 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
2048 zone_start_pfn += size;
2052 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
2054 /* Skip empty nodes */
2055 if (!pgdat->node_spanned_pages)
2058 #ifdef CONFIG_FLAT_NODE_MEM_MAP
2059 /* ia64 gets its own node_mem_map, before this, without bootmem */
2060 if (!pgdat->node_mem_map) {
2061 unsigned long size, start, end;
2065 * The zone's endpoints aren't required to be MAX_ORDER
2066 * aligned but the node_mem_map endpoints must be in order
2067 * for the buddy allocator to function correctly.
2069 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
2070 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
2071 end = ALIGN(end, MAX_ORDER_NR_PAGES);
2072 size = (end - start) * sizeof(struct page);
2073 map = alloc_remap(pgdat->node_id, size);
2075 map = alloc_bootmem_node(pgdat, size);
2076 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
2078 #ifdef CONFIG_FLATMEM
2080 * With no DISCONTIG, the global mem_map is just set as node 0's
2082 if (pgdat == NODE_DATA(0))
2083 mem_map = NODE_DATA(0)->node_mem_map;
2085 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
2088 void __meminit free_area_init_node(int nid, struct pglist_data *pgdat,
2089 unsigned long *zones_size, unsigned long node_start_pfn,
2090 unsigned long *zholes_size)
2092 pgdat->node_id = nid;
2093 pgdat->node_start_pfn = node_start_pfn;
2094 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
2096 alloc_node_mem_map(pgdat);
2098 free_area_init_core(pgdat, zones_size, zholes_size);
2101 #ifndef CONFIG_NEED_MULTIPLE_NODES
2102 static bootmem_data_t contig_bootmem_data;
2103 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
2105 EXPORT_SYMBOL(contig_page_data);
2108 void __init free_area_init(unsigned long *zones_size)
2110 free_area_init_node(0, NODE_DATA(0), zones_size,
2111 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
2114 #ifdef CONFIG_HOTPLUG_CPU
2115 static int page_alloc_cpu_notify(struct notifier_block *self,
2116 unsigned long action, void *hcpu)
2118 int cpu = (unsigned long)hcpu;
2120 if (action == CPU_DEAD) {
2121 local_irq_disable();
2123 vm_events_fold_cpu(cpu);
2125 refresh_cpu_vm_stats(cpu);
2129 #endif /* CONFIG_HOTPLUG_CPU */
2131 void __init page_alloc_init(void)
2133 hotcpu_notifier(page_alloc_cpu_notify, 0);
2137 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
2138 * or min_free_kbytes changes.
2140 static void calculate_totalreserve_pages(void)
2142 struct pglist_data *pgdat;
2143 unsigned long reserve_pages = 0;
2146 for_each_online_pgdat(pgdat) {
2147 for (i = 0; i < MAX_NR_ZONES; i++) {
2148 struct zone *zone = pgdat->node_zones + i;
2149 unsigned long max = 0;
2151 /* Find valid and maximum lowmem_reserve in the zone */
2152 for (j = i; j < MAX_NR_ZONES; j++) {
2153 if (zone->lowmem_reserve[j] > max)
2154 max = zone->lowmem_reserve[j];
2157 /* we treat pages_high as reserved pages. */
2158 max += zone->pages_high;
2160 if (max > zone->present_pages)
2161 max = zone->present_pages;
2162 reserve_pages += max;
2165 totalreserve_pages = reserve_pages;
2169 * setup_per_zone_lowmem_reserve - called whenever
2170 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
2171 * has a correct pages reserved value, so an adequate number of
2172 * pages are left in the zone after a successful __alloc_pages().
2174 static void setup_per_zone_lowmem_reserve(void)
2176 struct pglist_data *pgdat;
2179 for_each_online_pgdat(pgdat) {
2180 for (j = 0; j < MAX_NR_ZONES; j++) {
2181 struct zone *zone = pgdat->node_zones + j;
2182 unsigned long present_pages = zone->present_pages;
2184 zone->lowmem_reserve[j] = 0;
2186 for (idx = j-1; idx >= 0; idx--) {
2187 struct zone *lower_zone;
2189 if (sysctl_lowmem_reserve_ratio[idx] < 1)
2190 sysctl_lowmem_reserve_ratio[idx] = 1;
2192 lower_zone = pgdat->node_zones + idx;
2193 lower_zone->lowmem_reserve[j] = present_pages /
2194 sysctl_lowmem_reserve_ratio[idx];
2195 present_pages += lower_zone->present_pages;
2200 /* update totalreserve_pages */
2201 calculate_totalreserve_pages();
2205 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
2206 * that the pages_{min,low,high} values for each zone are set correctly
2207 * with respect to min_free_kbytes.
2209 void setup_per_zone_pages_min(void)
2211 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
2212 unsigned long lowmem_pages = 0;
2214 unsigned long flags;
2216 /* Calculate total number of !ZONE_HIGHMEM pages */
2217 for_each_zone(zone) {
2218 if (!is_highmem(zone))
2219 lowmem_pages += zone->present_pages;
2222 for_each_zone(zone) {
2225 spin_lock_irqsave(&zone->lru_lock, flags);
2226 tmp = (u64)pages_min * zone->present_pages;
2227 do_div(tmp, lowmem_pages);
2228 if (is_highmem(zone)) {
2230 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
2231 * need highmem pages, so cap pages_min to a small
2234 * The (pages_high-pages_low) and (pages_low-pages_min)
2235 * deltas controls asynch page reclaim, and so should
2236 * not be capped for highmem.
2240 min_pages = zone->present_pages / 1024;
2241 if (min_pages < SWAP_CLUSTER_MAX)
2242 min_pages = SWAP_CLUSTER_MAX;
2243 if (min_pages > 128)
2245 zone->pages_min = min_pages;
2248 * If it's a lowmem zone, reserve a number of pages
2249 * proportionate to the zone's size.
2251 zone->pages_min = tmp;
2254 zone->pages_low = zone->pages_min + (tmp >> 2);
2255 zone->pages_high = zone->pages_min + (tmp >> 1);
2256 spin_unlock_irqrestore(&zone->lru_lock, flags);
2259 /* update totalreserve_pages */
2260 calculate_totalreserve_pages();
2264 * Initialise min_free_kbytes.
2266 * For small machines we want it small (128k min). For large machines
2267 * we want it large (64MB max). But it is not linear, because network
2268 * bandwidth does not increase linearly with machine size. We use
2270 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2271 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2287 static int __init init_per_zone_pages_min(void)
2289 unsigned long lowmem_kbytes;
2291 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2293 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2294 if (min_free_kbytes < 128)
2295 min_free_kbytes = 128;
2296 if (min_free_kbytes > 65536)
2297 min_free_kbytes = 65536;
2298 setup_per_zone_pages_min();
2299 setup_per_zone_lowmem_reserve();
2302 module_init(init_per_zone_pages_min)
2305 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2306 * that we can call two helper functions whenever min_free_kbytes
2309 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2310 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2312 proc_dointvec(table, write, file, buffer, length, ppos);
2313 setup_per_zone_pages_min();
2318 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
2319 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2324 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2329 zone->min_unmapped_ratio = (zone->present_pages *
2330 sysctl_min_unmapped_ratio) / 100;
2334 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
2335 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2340 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2345 zone->min_slab_pages = (zone->present_pages *
2346 sysctl_min_slab_ratio) / 100;
2352 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2353 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2354 * whenever sysctl_lowmem_reserve_ratio changes.
2356 * The reserve ratio obviously has absolutely no relation with the
2357 * pages_min watermarks. The lowmem reserve ratio can only make sense
2358 * if in function of the boot time zone sizes.
2360 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
2361 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2363 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2364 setup_per_zone_lowmem_reserve();
2369 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
2370 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
2371 * can have before it gets flushed back to buddy allocator.
2374 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
2375 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2381 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2382 if (!write || (ret == -EINVAL))
2384 for_each_zone(zone) {
2385 for_each_online_cpu(cpu) {
2387 high = zone->present_pages / percpu_pagelist_fraction;
2388 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
2394 __initdata int hashdist = HASHDIST_DEFAULT;
2397 static int __init set_hashdist(char *str)
2401 hashdist = simple_strtoul(str, &str, 0);
2404 __setup("hashdist=", set_hashdist);
2408 * allocate a large system hash table from bootmem
2409 * - it is assumed that the hash table must contain an exact power-of-2
2410 * quantity of entries
2411 * - limit is the number of hash buckets, not the total allocation size
2413 void *__init alloc_large_system_hash(const char *tablename,
2414 unsigned long bucketsize,
2415 unsigned long numentries,
2418 unsigned int *_hash_shift,
2419 unsigned int *_hash_mask,
2420 unsigned long limit)
2422 unsigned long long max = limit;
2423 unsigned long log2qty, size;
2426 /* allow the kernel cmdline to have a say */
2428 /* round applicable memory size up to nearest megabyte */
2429 numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages;
2430 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2431 numentries >>= 20 - PAGE_SHIFT;
2432 numentries <<= 20 - PAGE_SHIFT;
2434 /* limit to 1 bucket per 2^scale bytes of low memory */
2435 if (scale > PAGE_SHIFT)
2436 numentries >>= (scale - PAGE_SHIFT);
2438 numentries <<= (PAGE_SHIFT - scale);
2440 numentries = roundup_pow_of_two(numentries);
2442 /* limit allocation size to 1/16 total memory by default */
2444 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2445 do_div(max, bucketsize);
2448 if (numentries > max)
2451 log2qty = long_log2(numentries);
2454 size = bucketsize << log2qty;
2455 if (flags & HASH_EARLY)
2456 table = alloc_bootmem(size);
2458 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
2460 unsigned long order;
2461 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
2463 table = (void*) __get_free_pages(GFP_ATOMIC, order);
2465 } while (!table && size > PAGE_SIZE && --log2qty);
2468 panic("Failed to allocate %s hash table\n", tablename);
2470 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2473 long_log2(size) - PAGE_SHIFT,
2477 *_hash_shift = log2qty;
2479 *_hash_mask = (1 << log2qty) - 1;
2484 #ifdef CONFIG_OUT_OF_LINE_PFN_TO_PAGE
2485 struct page *pfn_to_page(unsigned long pfn)
2487 return __pfn_to_page(pfn);
2489 unsigned long page_to_pfn(struct page *page)
2491 return __page_to_pfn(page);
2493 EXPORT_SYMBOL(pfn_to_page);
2494 EXPORT_SYMBOL(page_to_pfn);
2495 #endif /* CONFIG_OUT_OF_LINE_PFN_TO_PAGE */