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/sort.h>
41 #include <linux/pfn.h>
42 #include <linux/backing-dev.h>
44 #include <asm/tlbflush.h>
45 #include <asm/div64.h>
49 * MCD - HACK: Find somewhere to initialize this EARLY, or make this
52 nodemask_t node_online_map __read_mostly = { { [0] = 1UL } };
53 EXPORT_SYMBOL(node_online_map);
54 nodemask_t node_possible_map __read_mostly = NODE_MASK_ALL;
55 EXPORT_SYMBOL(node_possible_map);
56 unsigned long totalram_pages __read_mostly;
57 unsigned long totalreserve_pages __read_mostly;
59 int percpu_pagelist_fraction;
61 static void __free_pages_ok(struct page *page, unsigned int order);
64 * results with 256, 32 in the lowmem_reserve sysctl:
65 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
66 * 1G machine -> (16M dma, 784M normal, 224M high)
67 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
68 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
69 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
71 * TBD: should special case ZONE_DMA32 machines here - in those we normally
72 * don't need any ZONE_NORMAL reservation
74 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
76 #ifdef CONFIG_ZONE_DMA32
84 EXPORT_SYMBOL(totalram_pages);
86 static char *zone_names[MAX_NR_ZONES] = {
88 #ifdef CONFIG_ZONE_DMA32
97 int min_free_kbytes = 1024;
99 unsigned long __meminitdata nr_kernel_pages;
100 unsigned long __meminitdata nr_all_pages;
101 static unsigned long __initdata dma_reserve;
103 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
105 * MAX_ACTIVE_REGIONS determines the maxmimum number of distinct
106 * ranges of memory (RAM) that may be registered with add_active_range().
107 * Ranges passed to add_active_range() will be merged if possible
108 * so the number of times add_active_range() can be called is
109 * related to the number of nodes and the number of holes
111 #ifdef CONFIG_MAX_ACTIVE_REGIONS
112 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
113 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
115 #if MAX_NUMNODES >= 32
116 /* If there can be many nodes, allow up to 50 holes per node */
117 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
119 /* By default, allow up to 256 distinct regions */
120 #define MAX_ACTIVE_REGIONS 256
124 struct node_active_region __initdata early_node_map[MAX_ACTIVE_REGIONS];
125 int __initdata nr_nodemap_entries;
126 unsigned long __initdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
127 unsigned long __initdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
128 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
129 unsigned long __initdata node_boundary_start_pfn[MAX_NUMNODES];
130 unsigned long __initdata node_boundary_end_pfn[MAX_NUMNODES];
131 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
132 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
134 #ifdef CONFIG_DEBUG_VM
135 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
139 unsigned long pfn = page_to_pfn(page);
142 seq = zone_span_seqbegin(zone);
143 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
145 else if (pfn < zone->zone_start_pfn)
147 } while (zone_span_seqretry(zone, seq));
152 static int page_is_consistent(struct zone *zone, struct page *page)
154 #ifdef CONFIG_HOLES_IN_ZONE
155 if (!pfn_valid(page_to_pfn(page)))
158 if (zone != page_zone(page))
164 * Temporary debugging check for pages not lying within a given zone.
166 static int bad_range(struct zone *zone, struct page *page)
168 if (page_outside_zone_boundaries(zone, page))
170 if (!page_is_consistent(zone, page))
176 static inline int bad_range(struct zone *zone, struct page *page)
182 static void bad_page(struct page *page)
184 printk(KERN_EMERG "Bad page state in process '%s'\n"
185 KERN_EMERG "page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d\n"
186 KERN_EMERG "Trying to fix it up, but a reboot is needed\n"
187 KERN_EMERG "Backtrace:\n",
188 current->comm, page, (int)(2*sizeof(unsigned long)),
189 (unsigned long)page->flags, page->mapping,
190 page_mapcount(page), page_count(page));
192 page->flags &= ~(1 << PG_lru |
202 set_page_count(page, 0);
203 reset_page_mapcount(page);
204 page->mapping = NULL;
205 add_taint(TAINT_BAD_PAGE);
209 * Higher-order pages are called "compound pages". They are structured thusly:
211 * The first PAGE_SIZE page is called the "head page".
213 * The remaining PAGE_SIZE pages are called "tail pages".
215 * All pages have PG_compound set. All pages have their ->private pointing at
216 * the head page (even the head page has this).
218 * The first tail page's ->lru.next holds the address of the compound page's
219 * put_page() function. Its ->lru.prev holds the order of allocation.
220 * This usage means that zero-order pages may not be compound.
223 static void free_compound_page(struct page *page)
225 __free_pages_ok(page, (unsigned long)page[1].lru.prev);
228 static void prep_compound_page(struct page *page, unsigned long order)
231 int nr_pages = 1 << order;
233 page[1].lru.next = (void *)free_compound_page; /* set dtor */
234 page[1].lru.prev = (void *)order;
235 for (i = 0; i < nr_pages; i++) {
236 struct page *p = page + i;
238 __SetPageCompound(p);
239 set_page_private(p, (unsigned long)page);
243 static void destroy_compound_page(struct page *page, unsigned long order)
246 int nr_pages = 1 << order;
248 if (unlikely((unsigned long)page[1].lru.prev != order))
251 for (i = 0; i < nr_pages; i++) {
252 struct page *p = page + i;
254 if (unlikely(!PageCompound(p) |
255 (page_private(p) != (unsigned long)page)))
257 __ClearPageCompound(p);
261 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
265 VM_BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
267 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
268 * and __GFP_HIGHMEM from hard or soft interrupt context.
270 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
271 for (i = 0; i < (1 << order); i++)
272 clear_highpage(page + i);
276 * function for dealing with page's order in buddy system.
277 * zone->lock is already acquired when we use these.
278 * So, we don't need atomic page->flags operations here.
280 static inline unsigned long page_order(struct page *page)
282 return page_private(page);
285 static inline void set_page_order(struct page *page, int order)
287 set_page_private(page, order);
288 __SetPageBuddy(page);
291 static inline void rmv_page_order(struct page *page)
293 __ClearPageBuddy(page);
294 set_page_private(page, 0);
298 * Locate the struct page for both the matching buddy in our
299 * pair (buddy1) and the combined O(n+1) page they form (page).
301 * 1) Any buddy B1 will have an order O twin B2 which satisfies
302 * the following equation:
304 * For example, if the starting buddy (buddy2) is #8 its order
306 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
308 * 2) Any buddy B will have an order O+1 parent P which
309 * satisfies the following equation:
312 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
314 static inline struct page *
315 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
317 unsigned long buddy_idx = page_idx ^ (1 << order);
319 return page + (buddy_idx - page_idx);
322 static inline unsigned long
323 __find_combined_index(unsigned long page_idx, unsigned int order)
325 return (page_idx & ~(1 << order));
329 * This function checks whether a page is free && is the buddy
330 * we can do coalesce a page and its buddy if
331 * (a) the buddy is not in a hole &&
332 * (b) the buddy is in the buddy system &&
333 * (c) a page and its buddy have the same order &&
334 * (d) a page and its buddy are in the same zone.
336 * For recording whether a page is in the buddy system, we use PG_buddy.
337 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
339 * For recording page's order, we use page_private(page).
341 static inline int page_is_buddy(struct page *page, struct page *buddy,
344 #ifdef CONFIG_HOLES_IN_ZONE
345 if (!pfn_valid(page_to_pfn(buddy)))
349 if (page_zone_id(page) != page_zone_id(buddy))
352 if (PageBuddy(buddy) && page_order(buddy) == order) {
353 BUG_ON(page_count(buddy) != 0);
360 * Freeing function for a buddy system allocator.
362 * The concept of a buddy system is to maintain direct-mapped table
363 * (containing bit values) for memory blocks of various "orders".
364 * The bottom level table contains the map for the smallest allocatable
365 * units of memory (here, pages), and each level above it describes
366 * pairs of units from the levels below, hence, "buddies".
367 * At a high level, all that happens here is marking the table entry
368 * at the bottom level available, and propagating the changes upward
369 * as necessary, plus some accounting needed to play nicely with other
370 * parts of the VM system.
371 * At each level, we keep a list of pages, which are heads of continuous
372 * free pages of length of (1 << order) and marked with PG_buddy. Page's
373 * order is recorded in page_private(page) field.
374 * So when we are allocating or freeing one, we can derive the state of the
375 * other. That is, if we allocate a small block, and both were
376 * free, the remainder of the region must be split into blocks.
377 * If a block is freed, and its buddy is also free, then this
378 * triggers coalescing into a block of larger size.
383 static inline void __free_one_page(struct page *page,
384 struct zone *zone, unsigned int order)
386 unsigned long page_idx;
387 int order_size = 1 << order;
389 if (unlikely(PageCompound(page)))
390 destroy_compound_page(page, order);
392 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
394 VM_BUG_ON(page_idx & (order_size - 1));
395 VM_BUG_ON(bad_range(zone, page));
397 zone->free_pages += order_size;
398 while (order < MAX_ORDER-1) {
399 unsigned long combined_idx;
400 struct free_area *area;
403 buddy = __page_find_buddy(page, page_idx, order);
404 if (!page_is_buddy(page, buddy, order))
405 break; /* Move the buddy up one level. */
407 list_del(&buddy->lru);
408 area = zone->free_area + order;
410 rmv_page_order(buddy);
411 combined_idx = __find_combined_index(page_idx, order);
412 page = page + (combined_idx - page_idx);
413 page_idx = combined_idx;
416 set_page_order(page, order);
417 list_add(&page->lru, &zone->free_area[order].free_list);
418 zone->free_area[order].nr_free++;
421 static inline int free_pages_check(struct page *page)
423 if (unlikely(page_mapcount(page) |
424 (page->mapping != NULL) |
425 (page_count(page) != 0) |
439 __ClearPageDirty(page);
441 * For now, we report if PG_reserved was found set, but do not
442 * clear it, and do not free the page. But we shall soon need
443 * to do more, for when the ZERO_PAGE count wraps negative.
445 return PageReserved(page);
449 * Frees a list of pages.
450 * Assumes all pages on list are in same zone, and of same order.
451 * count is the number of pages to free.
453 * If the zone was previously in an "all pages pinned" state then look to
454 * see if this freeing clears that state.
456 * And clear the zone's pages_scanned counter, to hold off the "all pages are
457 * pinned" detection logic.
459 static void free_pages_bulk(struct zone *zone, int count,
460 struct list_head *list, int order)
462 spin_lock(&zone->lock);
463 zone->all_unreclaimable = 0;
464 zone->pages_scanned = 0;
468 VM_BUG_ON(list_empty(list));
469 page = list_entry(list->prev, struct page, lru);
470 /* have to delete it as __free_one_page list manipulates */
471 list_del(&page->lru);
472 __free_one_page(page, zone, order);
474 spin_unlock(&zone->lock);
477 static void free_one_page(struct zone *zone, struct page *page, int order)
479 spin_lock(&zone->lock);
480 zone->all_unreclaimable = 0;
481 zone->pages_scanned = 0;
482 __free_one_page(page, zone, order);
483 spin_unlock(&zone->lock);
486 static void __free_pages_ok(struct page *page, unsigned int order)
492 for (i = 0 ; i < (1 << order) ; ++i)
493 reserved += free_pages_check(page + i);
497 if (!PageHighMem(page))
498 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
499 arch_free_page(page, order);
500 kernel_map_pages(page, 1 << order, 0);
502 local_irq_save(flags);
503 __count_vm_events(PGFREE, 1 << order);
504 free_one_page(page_zone(page), page, order);
505 local_irq_restore(flags);
509 * permit the bootmem allocator to evade page validation on high-order frees
511 void fastcall __init __free_pages_bootmem(struct page *page, unsigned int order)
514 __ClearPageReserved(page);
515 set_page_count(page, 0);
516 set_page_refcounted(page);
522 for (loop = 0; loop < BITS_PER_LONG; loop++) {
523 struct page *p = &page[loop];
525 if (loop + 1 < BITS_PER_LONG)
527 __ClearPageReserved(p);
528 set_page_count(p, 0);
531 set_page_refcounted(page);
532 __free_pages(page, order);
538 * The order of subdivision here is critical for the IO subsystem.
539 * Please do not alter this order without good reasons and regression
540 * testing. Specifically, as large blocks of memory are subdivided,
541 * the order in which smaller blocks are delivered depends on the order
542 * they're subdivided in this function. This is the primary factor
543 * influencing the order in which pages are delivered to the IO
544 * subsystem according to empirical testing, and this is also justified
545 * by considering the behavior of a buddy system containing a single
546 * large block of memory acted on by a series of small allocations.
547 * This behavior is a critical factor in sglist merging's success.
551 static inline void expand(struct zone *zone, struct page *page,
552 int low, int high, struct free_area *area)
554 unsigned long size = 1 << high;
560 VM_BUG_ON(bad_range(zone, &page[size]));
561 list_add(&page[size].lru, &area->free_list);
563 set_page_order(&page[size], high);
568 * This page is about to be returned from the page allocator
570 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
572 if (unlikely(page_mapcount(page) |
573 (page->mapping != NULL) |
574 (page_count(page) != 0) |
590 * For now, we report if PG_reserved was found set, but do not
591 * clear it, and do not allocate the page: as a safety net.
593 if (PageReserved(page))
596 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
597 1 << PG_referenced | 1 << PG_arch_1 |
598 1 << PG_checked | 1 << PG_mappedtodisk);
599 set_page_private(page, 0);
600 set_page_refcounted(page);
602 arch_alloc_page(page, order);
603 kernel_map_pages(page, 1 << order, 1);
605 if (gfp_flags & __GFP_ZERO)
606 prep_zero_page(page, order, gfp_flags);
608 if (order && (gfp_flags & __GFP_COMP))
609 prep_compound_page(page, order);
615 * Do the hard work of removing an element from the buddy allocator.
616 * Call me with the zone->lock already held.
618 static struct page *__rmqueue(struct zone *zone, unsigned int order)
620 struct free_area * area;
621 unsigned int current_order;
624 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
625 area = zone->free_area + current_order;
626 if (list_empty(&area->free_list))
629 page = list_entry(area->free_list.next, struct page, lru);
630 list_del(&page->lru);
631 rmv_page_order(page);
633 zone->free_pages -= 1UL << order;
634 expand(zone, page, order, current_order, area);
642 * Obtain a specified number of elements from the buddy allocator, all under
643 * a single hold of the lock, for efficiency. Add them to the supplied list.
644 * Returns the number of new pages which were placed at *list.
646 static int rmqueue_bulk(struct zone *zone, unsigned int order,
647 unsigned long count, struct list_head *list)
651 spin_lock(&zone->lock);
652 for (i = 0; i < count; ++i) {
653 struct page *page = __rmqueue(zone, order);
654 if (unlikely(page == NULL))
656 list_add_tail(&page->lru, list);
658 spin_unlock(&zone->lock);
664 * Called from the slab reaper to drain pagesets on a particular node that
665 * belongs to the currently executing processor.
666 * Note that this function must be called with the thread pinned to
667 * a single processor.
669 void drain_node_pages(int nodeid)
675 for (z = 0; z < MAX_NR_ZONES; z++) {
676 struct zone *zone = NODE_DATA(nodeid)->node_zones + z;
677 struct per_cpu_pageset *pset;
679 if (!populated_zone(zone))
682 pset = zone_pcp(zone, smp_processor_id());
683 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
684 struct per_cpu_pages *pcp;
688 local_irq_save(flags);
689 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
691 local_irq_restore(flags);
698 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
699 static void __drain_pages(unsigned int cpu)
705 for_each_zone(zone) {
706 struct per_cpu_pageset *pset;
708 pset = zone_pcp(zone, cpu);
709 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
710 struct per_cpu_pages *pcp;
713 local_irq_save(flags);
714 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
716 local_irq_restore(flags);
720 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
724 void mark_free_pages(struct zone *zone)
726 unsigned long pfn, max_zone_pfn;
729 struct list_head *curr;
731 if (!zone->spanned_pages)
734 spin_lock_irqsave(&zone->lock, flags);
736 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
737 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
738 if (pfn_valid(pfn)) {
739 struct page *page = pfn_to_page(pfn);
741 if (!PageNosave(page))
742 ClearPageNosaveFree(page);
745 for (order = MAX_ORDER - 1; order >= 0; --order)
746 list_for_each(curr, &zone->free_area[order].free_list) {
749 pfn = page_to_pfn(list_entry(curr, struct page, lru));
750 for (i = 0; i < (1UL << order); i++)
751 SetPageNosaveFree(pfn_to_page(pfn + i));
754 spin_unlock_irqrestore(&zone->lock, flags);
758 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
760 void drain_local_pages(void)
764 local_irq_save(flags);
765 __drain_pages(smp_processor_id());
766 local_irq_restore(flags);
768 #endif /* CONFIG_PM */
771 * Free a 0-order page
773 static void fastcall free_hot_cold_page(struct page *page, int cold)
775 struct zone *zone = page_zone(page);
776 struct per_cpu_pages *pcp;
780 page->mapping = NULL;
781 if (free_pages_check(page))
784 if (!PageHighMem(page))
785 debug_check_no_locks_freed(page_address(page), PAGE_SIZE);
786 arch_free_page(page, 0);
787 kernel_map_pages(page, 1, 0);
789 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
790 local_irq_save(flags);
791 __count_vm_event(PGFREE);
792 list_add(&page->lru, &pcp->list);
794 if (pcp->count >= pcp->high) {
795 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
796 pcp->count -= pcp->batch;
798 local_irq_restore(flags);
802 void fastcall free_hot_page(struct page *page)
804 free_hot_cold_page(page, 0);
807 void fastcall free_cold_page(struct page *page)
809 free_hot_cold_page(page, 1);
813 * split_page takes a non-compound higher-order page, and splits it into
814 * n (1<<order) sub-pages: page[0..n]
815 * Each sub-page must be freed individually.
817 * Note: this is probably too low level an operation for use in drivers.
818 * Please consult with lkml before using this in your driver.
820 void split_page(struct page *page, unsigned int order)
824 VM_BUG_ON(PageCompound(page));
825 VM_BUG_ON(!page_count(page));
826 for (i = 1; i < (1 << order); i++)
827 set_page_refcounted(page + i);
831 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
832 * we cheat by calling it from here, in the order > 0 path. Saves a branch
835 static struct page *buffered_rmqueue(struct zonelist *zonelist,
836 struct zone *zone, int order, gfp_t gfp_flags)
840 int cold = !!(gfp_flags & __GFP_COLD);
845 if (likely(order == 0)) {
846 struct per_cpu_pages *pcp;
848 pcp = &zone_pcp(zone, cpu)->pcp[cold];
849 local_irq_save(flags);
851 pcp->count = rmqueue_bulk(zone, 0,
852 pcp->batch, &pcp->list);
853 if (unlikely(!pcp->count))
856 page = list_entry(pcp->list.next, struct page, lru);
857 list_del(&page->lru);
860 spin_lock_irqsave(&zone->lock, flags);
861 page = __rmqueue(zone, order);
862 spin_unlock(&zone->lock);
867 __count_zone_vm_events(PGALLOC, zone, 1 << order);
868 zone_statistics(zonelist, zone);
869 local_irq_restore(flags);
872 VM_BUG_ON(bad_range(zone, page));
873 if (prep_new_page(page, order, gfp_flags))
878 local_irq_restore(flags);
883 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
884 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
885 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
886 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
887 #define ALLOC_HARDER 0x10 /* try to alloc harder */
888 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
889 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
892 * Return 1 if free pages are above 'mark'. This takes into account the order
895 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
896 int classzone_idx, int alloc_flags)
898 /* free_pages my go negative - that's OK */
899 unsigned long min = mark;
900 long free_pages = z->free_pages - (1 << order) + 1;
903 if (alloc_flags & ALLOC_HIGH)
905 if (alloc_flags & ALLOC_HARDER)
908 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
910 for (o = 0; o < order; o++) {
911 /* At the next order, this order's pages become unavailable */
912 free_pages -= z->free_area[o].nr_free << o;
914 /* Require fewer higher order pages to be free */
917 if (free_pages <= min)
925 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
926 * skip over zones that are not allowed by the cpuset, or that have
927 * been recently (in last second) found to be nearly full. See further
928 * comments in mmzone.h. Reduces cache footprint of zonelist scans
929 * that have to skip over alot of full or unallowed zones.
931 * If the zonelist cache is present in the passed in zonelist, then
932 * returns a pointer to the allowed node mask (either the current
933 * tasks mems_allowed, or node_online_map.)
935 * If the zonelist cache is not available for this zonelist, does
936 * nothing and returns NULL.
938 * If the fullzones BITMAP in the zonelist cache is stale (more than
939 * a second since last zap'd) then we zap it out (clear its bits.)
941 * We hold off even calling zlc_setup, until after we've checked the
942 * first zone in the zonelist, on the theory that most allocations will
943 * be satisfied from that first zone, so best to examine that zone as
946 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
948 struct zonelist_cache *zlc; /* cached zonelist speedup info */
949 nodemask_t *allowednodes; /* zonelist_cache approximation */
951 zlc = zonelist->zlcache_ptr;
955 if (jiffies - zlc->last_full_zap > 1 * HZ) {
956 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
957 zlc->last_full_zap = jiffies;
960 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
961 &cpuset_current_mems_allowed :
967 * Given 'z' scanning a zonelist, run a couple of quick checks to see
968 * if it is worth looking at further for free memory:
969 * 1) Check that the zone isn't thought to be full (doesn't have its
970 * bit set in the zonelist_cache fullzones BITMAP).
971 * 2) Check that the zones node (obtained from the zonelist_cache
972 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
973 * Return true (non-zero) if zone is worth looking at further, or
974 * else return false (zero) if it is not.
976 * This check -ignores- the distinction between various watermarks,
977 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
978 * found to be full for any variation of these watermarks, it will
979 * be considered full for up to one second by all requests, unless
980 * we are so low on memory on all allowed nodes that we are forced
981 * into the second scan of the zonelist.
983 * In the second scan we ignore this zonelist cache and exactly
984 * apply the watermarks to all zones, even it is slower to do so.
985 * We are low on memory in the second scan, and should leave no stone
986 * unturned looking for a free page.
988 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zone **z,
989 nodemask_t *allowednodes)
991 struct zonelist_cache *zlc; /* cached zonelist speedup info */
992 int i; /* index of *z in zonelist zones */
993 int n; /* node that zone *z is on */
995 zlc = zonelist->zlcache_ptr;
999 i = z - zonelist->zones;
1002 /* This zone is worth trying if it is allowed but not full */
1003 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1007 * Given 'z' scanning a zonelist, set the corresponding bit in
1008 * zlc->fullzones, so that subsequent attempts to allocate a page
1009 * from that zone don't waste time re-examining it.
1011 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zone **z)
1013 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1014 int i; /* index of *z in zonelist zones */
1016 zlc = zonelist->zlcache_ptr;
1020 i = z - zonelist->zones;
1022 set_bit(i, zlc->fullzones);
1025 #else /* CONFIG_NUMA */
1027 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1032 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zone **z,
1033 nodemask_t *allowednodes)
1038 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zone **z)
1041 #endif /* CONFIG_NUMA */
1044 * get_page_from_freelist goes through the zonelist trying to allocate
1047 static struct page *
1048 get_page_from_freelist(gfp_t gfp_mask, unsigned int order,
1049 struct zonelist *zonelist, int alloc_flags)
1052 struct page *page = NULL;
1053 int classzone_idx = zone_idx(zonelist->zones[0]);
1055 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1056 int zlc_active = 0; /* set if using zonelist_cache */
1057 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1061 * Scan zonelist, looking for a zone with enough free.
1062 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1064 z = zonelist->zones;
1067 if (NUMA_BUILD && zlc_active &&
1068 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1071 if (unlikely(NUMA_BUILD && (gfp_mask & __GFP_THISNODE) &&
1072 zone->zone_pgdat != zonelist->zones[0]->zone_pgdat))
1074 if ((alloc_flags & ALLOC_CPUSET) &&
1075 !cpuset_zone_allowed(zone, gfp_mask))
1078 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1080 if (alloc_flags & ALLOC_WMARK_MIN)
1081 mark = zone->pages_min;
1082 else if (alloc_flags & ALLOC_WMARK_LOW)
1083 mark = zone->pages_low;
1085 mark = zone->pages_high;
1086 if (!zone_watermark_ok(zone, order, mark,
1087 classzone_idx, alloc_flags)) {
1088 if (!zone_reclaim_mode ||
1089 !zone_reclaim(zone, gfp_mask, order))
1090 goto this_zone_full;
1094 page = buffered_rmqueue(zonelist, zone, order, gfp_mask);
1099 zlc_mark_zone_full(zonelist, z);
1101 if (NUMA_BUILD && !did_zlc_setup) {
1102 /* we do zlc_setup after the first zone is tried */
1103 allowednodes = zlc_setup(zonelist, alloc_flags);
1107 } while (*(++z) != NULL);
1109 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1110 /* Disable zlc cache for second zonelist scan */
1118 * This is the 'heart' of the zoned buddy allocator.
1120 struct page * fastcall
1121 __alloc_pages(gfp_t gfp_mask, unsigned int order,
1122 struct zonelist *zonelist)
1124 const gfp_t wait = gfp_mask & __GFP_WAIT;
1127 struct reclaim_state reclaim_state;
1128 struct task_struct *p = current;
1131 int did_some_progress;
1133 might_sleep_if(wait);
1136 z = zonelist->zones; /* the list of zones suitable for gfp_mask */
1138 if (unlikely(*z == NULL)) {
1139 /* Should this ever happen?? */
1143 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1144 zonelist, ALLOC_WMARK_LOW|ALLOC_CPUSET);
1148 for (z = zonelist->zones; *z; z++)
1149 wakeup_kswapd(*z, order);
1152 * OK, we're below the kswapd watermark and have kicked background
1153 * reclaim. Now things get more complex, so set up alloc_flags according
1154 * to how we want to proceed.
1156 * The caller may dip into page reserves a bit more if the caller
1157 * cannot run direct reclaim, or if the caller has realtime scheduling
1158 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1159 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1161 alloc_flags = ALLOC_WMARK_MIN;
1162 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
1163 alloc_flags |= ALLOC_HARDER;
1164 if (gfp_mask & __GFP_HIGH)
1165 alloc_flags |= ALLOC_HIGH;
1167 alloc_flags |= ALLOC_CPUSET;
1170 * Go through the zonelist again. Let __GFP_HIGH and allocations
1171 * coming from realtime tasks go deeper into reserves.
1173 * This is the last chance, in general, before the goto nopage.
1174 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1175 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1177 page = get_page_from_freelist(gfp_mask, order, zonelist, alloc_flags);
1181 /* This allocation should allow future memory freeing. */
1184 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
1185 && !in_interrupt()) {
1186 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
1188 /* go through the zonelist yet again, ignoring mins */
1189 page = get_page_from_freelist(gfp_mask, order,
1190 zonelist, ALLOC_NO_WATERMARKS);
1193 if (gfp_mask & __GFP_NOFAIL) {
1194 congestion_wait(WRITE, HZ/50);
1201 /* Atomic allocations - we can't balance anything */
1207 /* We now go into synchronous reclaim */
1208 cpuset_memory_pressure_bump();
1209 p->flags |= PF_MEMALLOC;
1210 reclaim_state.reclaimed_slab = 0;
1211 p->reclaim_state = &reclaim_state;
1213 did_some_progress = try_to_free_pages(zonelist->zones, gfp_mask);
1215 p->reclaim_state = NULL;
1216 p->flags &= ~PF_MEMALLOC;
1220 if (likely(did_some_progress)) {
1221 page = get_page_from_freelist(gfp_mask, order,
1222 zonelist, alloc_flags);
1225 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1227 * Go through the zonelist yet one more time, keep
1228 * very high watermark here, this is only to catch
1229 * a parallel oom killing, we must fail if we're still
1230 * under heavy pressure.
1232 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1233 zonelist, ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1237 out_of_memory(zonelist, gfp_mask, order);
1242 * Don't let big-order allocations loop unless the caller explicitly
1243 * requests that. Wait for some write requests to complete then retry.
1245 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
1246 * <= 3, but that may not be true in other implementations.
1249 if (!(gfp_mask & __GFP_NORETRY)) {
1250 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
1252 if (gfp_mask & __GFP_NOFAIL)
1256 congestion_wait(WRITE, HZ/50);
1261 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1262 printk(KERN_WARNING "%s: page allocation failure."
1263 " order:%d, mode:0x%x\n",
1264 p->comm, order, gfp_mask);
1272 EXPORT_SYMBOL(__alloc_pages);
1275 * Common helper functions.
1277 fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1280 page = alloc_pages(gfp_mask, order);
1283 return (unsigned long) page_address(page);
1286 EXPORT_SYMBOL(__get_free_pages);
1288 fastcall unsigned long get_zeroed_page(gfp_t gfp_mask)
1293 * get_zeroed_page() returns a 32-bit address, which cannot represent
1296 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1298 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1300 return (unsigned long) page_address(page);
1304 EXPORT_SYMBOL(get_zeroed_page);
1306 void __pagevec_free(struct pagevec *pvec)
1308 int i = pagevec_count(pvec);
1311 free_hot_cold_page(pvec->pages[i], pvec->cold);
1314 fastcall void __free_pages(struct page *page, unsigned int order)
1316 if (put_page_testzero(page)) {
1318 free_hot_page(page);
1320 __free_pages_ok(page, order);
1324 EXPORT_SYMBOL(__free_pages);
1326 fastcall void free_pages(unsigned long addr, unsigned int order)
1329 VM_BUG_ON(!virt_addr_valid((void *)addr));
1330 __free_pages(virt_to_page((void *)addr), order);
1334 EXPORT_SYMBOL(free_pages);
1337 * Total amount of free (allocatable) RAM:
1339 unsigned int nr_free_pages(void)
1341 unsigned int sum = 0;
1345 sum += zone->free_pages;
1350 EXPORT_SYMBOL(nr_free_pages);
1353 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
1355 unsigned int sum = 0;
1358 for (i = 0; i < MAX_NR_ZONES; i++)
1359 sum += pgdat->node_zones[i].free_pages;
1365 static unsigned int nr_free_zone_pages(int offset)
1367 /* Just pick one node, since fallback list is circular */
1368 pg_data_t *pgdat = NODE_DATA(numa_node_id());
1369 unsigned int sum = 0;
1371 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1372 struct zone **zonep = zonelist->zones;
1375 for (zone = *zonep++; zone; zone = *zonep++) {
1376 unsigned long size = zone->present_pages;
1377 unsigned long high = zone->pages_high;
1386 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1388 unsigned int nr_free_buffer_pages(void)
1390 return nr_free_zone_pages(gfp_zone(GFP_USER));
1394 * Amount of free RAM allocatable within all zones
1396 unsigned int nr_free_pagecache_pages(void)
1398 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER));
1401 static inline void show_node(struct zone *zone)
1404 printk("Node %ld ", zone_to_nid(zone));
1407 void si_meminfo(struct sysinfo *val)
1409 val->totalram = totalram_pages;
1411 val->freeram = nr_free_pages();
1412 val->bufferram = nr_blockdev_pages();
1413 val->totalhigh = totalhigh_pages;
1414 val->freehigh = nr_free_highpages();
1415 val->mem_unit = PAGE_SIZE;
1418 EXPORT_SYMBOL(si_meminfo);
1421 void si_meminfo_node(struct sysinfo *val, int nid)
1423 pg_data_t *pgdat = NODE_DATA(nid);
1425 val->totalram = pgdat->node_present_pages;
1426 val->freeram = nr_free_pages_pgdat(pgdat);
1427 #ifdef CONFIG_HIGHMEM
1428 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1429 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1434 val->mem_unit = PAGE_SIZE;
1438 #define K(x) ((x) << (PAGE_SHIFT-10))
1441 * Show free area list (used inside shift_scroll-lock stuff)
1442 * We also calculate the percentage fragmentation. We do this by counting the
1443 * memory on each free list with the exception of the first item on the list.
1445 void show_free_areas(void)
1448 unsigned long active;
1449 unsigned long inactive;
1453 for_each_zone(zone) {
1454 if (!populated_zone(zone))
1458 printk("%s per-cpu:\n", zone->name);
1460 for_each_online_cpu(cpu) {
1461 struct per_cpu_pageset *pageset;
1463 pageset = zone_pcp(zone, cpu);
1465 printk("CPU %4d: Hot: hi:%5d, btch:%4d usd:%4d "
1466 "Cold: hi:%5d, btch:%4d usd:%4d\n",
1467 cpu, pageset->pcp[0].high,
1468 pageset->pcp[0].batch, pageset->pcp[0].count,
1469 pageset->pcp[1].high, pageset->pcp[1].batch,
1470 pageset->pcp[1].count);
1474 get_zone_counts(&active, &inactive, &free);
1476 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1477 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1480 global_page_state(NR_FILE_DIRTY),
1481 global_page_state(NR_WRITEBACK),
1482 global_page_state(NR_UNSTABLE_NFS),
1484 global_page_state(NR_SLAB_RECLAIMABLE) +
1485 global_page_state(NR_SLAB_UNRECLAIMABLE),
1486 global_page_state(NR_FILE_MAPPED),
1487 global_page_state(NR_PAGETABLE));
1489 for_each_zone(zone) {
1492 if (!populated_zone(zone))
1504 " pages_scanned:%lu"
1505 " all_unreclaimable? %s"
1508 K(zone->free_pages),
1511 K(zone->pages_high),
1513 K(zone->nr_inactive),
1514 K(zone->present_pages),
1515 zone->pages_scanned,
1516 (zone->all_unreclaimable ? "yes" : "no")
1518 printk("lowmem_reserve[]:");
1519 for (i = 0; i < MAX_NR_ZONES; i++)
1520 printk(" %lu", zone->lowmem_reserve[i]);
1524 for_each_zone(zone) {
1525 unsigned long nr[MAX_ORDER], flags, order, total = 0;
1527 if (!populated_zone(zone))
1531 printk("%s: ", zone->name);
1533 spin_lock_irqsave(&zone->lock, flags);
1534 for (order = 0; order < MAX_ORDER; order++) {
1535 nr[order] = zone->free_area[order].nr_free;
1536 total += nr[order] << order;
1538 spin_unlock_irqrestore(&zone->lock, flags);
1539 for (order = 0; order < MAX_ORDER; order++)
1540 printk("%lu*%lukB ", nr[order], K(1UL) << order);
1541 printk("= %lukB\n", K(total));
1544 show_swap_cache_info();
1548 * Builds allocation fallback zone lists.
1550 * Add all populated zones of a node to the zonelist.
1552 static int __meminit build_zonelists_node(pg_data_t *pgdat,
1553 struct zonelist *zonelist, int nr_zones, enum zone_type zone_type)
1557 BUG_ON(zone_type >= MAX_NR_ZONES);
1562 zone = pgdat->node_zones + zone_type;
1563 if (populated_zone(zone)) {
1564 zonelist->zones[nr_zones++] = zone;
1565 check_highest_zone(zone_type);
1568 } while (zone_type);
1573 #define MAX_NODE_LOAD (num_online_nodes())
1574 static int __meminitdata node_load[MAX_NUMNODES];
1576 * find_next_best_node - find the next node that should appear in a given node's fallback list
1577 * @node: node whose fallback list we're appending
1578 * @used_node_mask: nodemask_t of already used nodes
1580 * We use a number of factors to determine which is the next node that should
1581 * appear on a given node's fallback list. The node should not have appeared
1582 * already in @node's fallback list, and it should be the next closest node
1583 * according to the distance array (which contains arbitrary distance values
1584 * from each node to each node in the system), and should also prefer nodes
1585 * with no CPUs, since presumably they'll have very little allocation pressure
1586 * on them otherwise.
1587 * It returns -1 if no node is found.
1589 static int __meminit find_next_best_node(int node, nodemask_t *used_node_mask)
1592 int min_val = INT_MAX;
1595 /* Use the local node if we haven't already */
1596 if (!node_isset(node, *used_node_mask)) {
1597 node_set(node, *used_node_mask);
1601 for_each_online_node(n) {
1604 /* Don't want a node to appear more than once */
1605 if (node_isset(n, *used_node_mask))
1608 /* Use the distance array to find the distance */
1609 val = node_distance(node, n);
1611 /* Penalize nodes under us ("prefer the next node") */
1614 /* Give preference to headless and unused nodes */
1615 tmp = node_to_cpumask(n);
1616 if (!cpus_empty(tmp))
1617 val += PENALTY_FOR_NODE_WITH_CPUS;
1619 /* Slight preference for less loaded node */
1620 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1621 val += node_load[n];
1623 if (val < min_val) {
1630 node_set(best_node, *used_node_mask);
1635 static void __meminit build_zonelists(pg_data_t *pgdat)
1637 int j, node, local_node;
1639 int prev_node, load;
1640 struct zonelist *zonelist;
1641 nodemask_t used_mask;
1643 /* initialize zonelists */
1644 for (i = 0; i < MAX_NR_ZONES; i++) {
1645 zonelist = pgdat->node_zonelists + i;
1646 zonelist->zones[0] = NULL;
1649 /* NUMA-aware ordering of nodes */
1650 local_node = pgdat->node_id;
1651 load = num_online_nodes();
1652 prev_node = local_node;
1653 nodes_clear(used_mask);
1654 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1655 int distance = node_distance(local_node, node);
1658 * If another node is sufficiently far away then it is better
1659 * to reclaim pages in a zone before going off node.
1661 if (distance > RECLAIM_DISTANCE)
1662 zone_reclaim_mode = 1;
1665 * We don't want to pressure a particular node.
1666 * So adding penalty to the first node in same
1667 * distance group to make it round-robin.
1670 if (distance != node_distance(local_node, prev_node))
1671 node_load[node] += load;
1674 for (i = 0; i < MAX_NR_ZONES; i++) {
1675 zonelist = pgdat->node_zonelists + i;
1676 for (j = 0; zonelist->zones[j] != NULL; j++);
1678 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
1679 zonelist->zones[j] = NULL;
1684 /* Construct the zonelist performance cache - see further mmzone.h */
1685 static void __meminit build_zonelist_cache(pg_data_t *pgdat)
1689 for (i = 0; i < MAX_NR_ZONES; i++) {
1690 struct zonelist *zonelist;
1691 struct zonelist_cache *zlc;
1694 zonelist = pgdat->node_zonelists + i;
1695 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
1696 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1697 for (z = zonelist->zones; *z; z++)
1698 zlc->z_to_n[z - zonelist->zones] = zone_to_nid(*z);
1702 #else /* CONFIG_NUMA */
1704 static void __meminit build_zonelists(pg_data_t *pgdat)
1706 int node, local_node;
1709 local_node = pgdat->node_id;
1710 for (i = 0; i < MAX_NR_ZONES; i++) {
1711 struct zonelist *zonelist;
1713 zonelist = pgdat->node_zonelists + i;
1715 j = build_zonelists_node(pgdat, zonelist, 0, i);
1717 * Now we build the zonelist so that it contains the zones
1718 * of all the other nodes.
1719 * We don't want to pressure a particular node, so when
1720 * building the zones for node N, we make sure that the
1721 * zones coming right after the local ones are those from
1722 * node N+1 (modulo N)
1724 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1725 if (!node_online(node))
1727 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
1729 for (node = 0; node < local_node; node++) {
1730 if (!node_online(node))
1732 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
1735 zonelist->zones[j] = NULL;
1739 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
1740 static void __meminit build_zonelist_cache(pg_data_t *pgdat)
1744 for (i = 0; i < MAX_NR_ZONES; i++)
1745 pgdat->node_zonelists[i].zlcache_ptr = NULL;
1748 #endif /* CONFIG_NUMA */
1750 /* return values int ....just for stop_machine_run() */
1751 static int __meminit __build_all_zonelists(void *dummy)
1755 for_each_online_node(nid) {
1756 build_zonelists(NODE_DATA(nid));
1757 build_zonelist_cache(NODE_DATA(nid));
1762 void __meminit build_all_zonelists(void)
1764 if (system_state == SYSTEM_BOOTING) {
1765 __build_all_zonelists(NULL);
1766 cpuset_init_current_mems_allowed();
1768 /* we have to stop all cpus to guaranntee there is no user
1770 stop_machine_run(__build_all_zonelists, NULL, NR_CPUS);
1771 /* cpuset refresh routine should be here */
1773 vm_total_pages = nr_free_pagecache_pages();
1774 printk("Built %i zonelists. Total pages: %ld\n",
1775 num_online_nodes(), vm_total_pages);
1779 * Helper functions to size the waitqueue hash table.
1780 * Essentially these want to choose hash table sizes sufficiently
1781 * large so that collisions trying to wait on pages are rare.
1782 * But in fact, the number of active page waitqueues on typical
1783 * systems is ridiculously low, less than 200. So this is even
1784 * conservative, even though it seems large.
1786 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1787 * waitqueues, i.e. the size of the waitq table given the number of pages.
1789 #define PAGES_PER_WAITQUEUE 256
1791 #ifndef CONFIG_MEMORY_HOTPLUG
1792 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
1794 unsigned long size = 1;
1796 pages /= PAGES_PER_WAITQUEUE;
1798 while (size < pages)
1802 * Once we have dozens or even hundreds of threads sleeping
1803 * on IO we've got bigger problems than wait queue collision.
1804 * Limit the size of the wait table to a reasonable size.
1806 size = min(size, 4096UL);
1808 return max(size, 4UL);
1812 * A zone's size might be changed by hot-add, so it is not possible to determine
1813 * a suitable size for its wait_table. So we use the maximum size now.
1815 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
1817 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
1818 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
1819 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
1821 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
1822 * or more by the traditional way. (See above). It equals:
1824 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
1825 * ia64(16K page size) : = ( 8G + 4M)byte.
1826 * powerpc (64K page size) : = (32G +16M)byte.
1828 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
1835 * This is an integer logarithm so that shifts can be used later
1836 * to extract the more random high bits from the multiplicative
1837 * hash function before the remainder is taken.
1839 static inline unsigned long wait_table_bits(unsigned long size)
1844 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1847 * Initially all pages are reserved - free ones are freed
1848 * up by free_all_bootmem() once the early boot process is
1849 * done. Non-atomic initialization, single-pass.
1851 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1852 unsigned long start_pfn)
1855 unsigned long end_pfn = start_pfn + size;
1858 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1859 if (!early_pfn_valid(pfn))
1861 if (!early_pfn_in_nid(pfn, nid))
1863 page = pfn_to_page(pfn);
1864 set_page_links(page, zone, nid, pfn);
1865 init_page_count(page);
1866 reset_page_mapcount(page);
1867 SetPageReserved(page);
1868 INIT_LIST_HEAD(&page->lru);
1869 #ifdef WANT_PAGE_VIRTUAL
1870 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1871 if (!is_highmem_idx(zone))
1872 set_page_address(page, __va(pfn << PAGE_SHIFT));
1877 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1881 for (order = 0; order < MAX_ORDER ; order++) {
1882 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1883 zone->free_area[order].nr_free = 0;
1887 #ifndef __HAVE_ARCH_MEMMAP_INIT
1888 #define memmap_init(size, nid, zone, start_pfn) \
1889 memmap_init_zone((size), (nid), (zone), (start_pfn))
1892 static int __cpuinit zone_batchsize(struct zone *zone)
1897 * The per-cpu-pages pools are set to around 1000th of the
1898 * size of the zone. But no more than 1/2 of a meg.
1900 * OK, so we don't know how big the cache is. So guess.
1902 batch = zone->present_pages / 1024;
1903 if (batch * PAGE_SIZE > 512 * 1024)
1904 batch = (512 * 1024) / PAGE_SIZE;
1905 batch /= 4; /* We effectively *= 4 below */
1910 * Clamp the batch to a 2^n - 1 value. Having a power
1911 * of 2 value was found to be more likely to have
1912 * suboptimal cache aliasing properties in some cases.
1914 * For example if 2 tasks are alternately allocating
1915 * batches of pages, one task can end up with a lot
1916 * of pages of one half of the possible page colors
1917 * and the other with pages of the other colors.
1919 batch = (1 << (fls(batch + batch/2)-1)) - 1;
1924 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
1926 struct per_cpu_pages *pcp;
1928 memset(p, 0, sizeof(*p));
1930 pcp = &p->pcp[0]; /* hot */
1932 pcp->high = 6 * batch;
1933 pcp->batch = max(1UL, 1 * batch);
1934 INIT_LIST_HEAD(&pcp->list);
1936 pcp = &p->pcp[1]; /* cold*/
1938 pcp->high = 2 * batch;
1939 pcp->batch = max(1UL, batch/2);
1940 INIT_LIST_HEAD(&pcp->list);
1944 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
1945 * to the value high for the pageset p.
1948 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
1951 struct per_cpu_pages *pcp;
1953 pcp = &p->pcp[0]; /* hot list */
1955 pcp->batch = max(1UL, high/4);
1956 if ((high/4) > (PAGE_SHIFT * 8))
1957 pcp->batch = PAGE_SHIFT * 8;
1963 * Boot pageset table. One per cpu which is going to be used for all
1964 * zones and all nodes. The parameters will be set in such a way
1965 * that an item put on a list will immediately be handed over to
1966 * the buddy list. This is safe since pageset manipulation is done
1967 * with interrupts disabled.
1969 * Some NUMA counter updates may also be caught by the boot pagesets.
1971 * The boot_pagesets must be kept even after bootup is complete for
1972 * unused processors and/or zones. They do play a role for bootstrapping
1973 * hotplugged processors.
1975 * zoneinfo_show() and maybe other functions do
1976 * not check if the processor is online before following the pageset pointer.
1977 * Other parts of the kernel may not check if the zone is available.
1979 static struct per_cpu_pageset boot_pageset[NR_CPUS];
1982 * Dynamically allocate memory for the
1983 * per cpu pageset array in struct zone.
1985 static int __cpuinit process_zones(int cpu)
1987 struct zone *zone, *dzone;
1989 for_each_zone(zone) {
1991 if (!populated_zone(zone))
1994 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
1995 GFP_KERNEL, cpu_to_node(cpu));
1996 if (!zone_pcp(zone, cpu))
1999 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
2001 if (percpu_pagelist_fraction)
2002 setup_pagelist_highmark(zone_pcp(zone, cpu),
2003 (zone->present_pages / percpu_pagelist_fraction));
2008 for_each_zone(dzone) {
2011 kfree(zone_pcp(dzone, cpu));
2012 zone_pcp(dzone, cpu) = NULL;
2017 static inline void free_zone_pagesets(int cpu)
2021 for_each_zone(zone) {
2022 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
2024 /* Free per_cpu_pageset if it is slab allocated */
2025 if (pset != &boot_pageset[cpu])
2027 zone_pcp(zone, cpu) = NULL;
2031 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
2032 unsigned long action,
2035 int cpu = (long)hcpu;
2036 int ret = NOTIFY_OK;
2039 case CPU_UP_PREPARE:
2040 if (process_zones(cpu))
2043 case CPU_UP_CANCELED:
2045 free_zone_pagesets(cpu);
2053 static struct notifier_block __cpuinitdata pageset_notifier =
2054 { &pageset_cpuup_callback, NULL, 0 };
2056 void __init setup_per_cpu_pageset(void)
2060 /* Initialize per_cpu_pageset for cpu 0.
2061 * A cpuup callback will do this for every cpu
2062 * as it comes online
2064 err = process_zones(smp_processor_id());
2066 register_cpu_notifier(&pageset_notifier);
2072 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
2075 struct pglist_data *pgdat = zone->zone_pgdat;
2079 * The per-page waitqueue mechanism uses hashed waitqueues
2082 zone->wait_table_hash_nr_entries =
2083 wait_table_hash_nr_entries(zone_size_pages);
2084 zone->wait_table_bits =
2085 wait_table_bits(zone->wait_table_hash_nr_entries);
2086 alloc_size = zone->wait_table_hash_nr_entries
2087 * sizeof(wait_queue_head_t);
2089 if (system_state == SYSTEM_BOOTING) {
2090 zone->wait_table = (wait_queue_head_t *)
2091 alloc_bootmem_node(pgdat, alloc_size);
2094 * This case means that a zone whose size was 0 gets new memory
2095 * via memory hot-add.
2096 * But it may be the case that a new node was hot-added. In
2097 * this case vmalloc() will not be able to use this new node's
2098 * memory - this wait_table must be initialized to use this new
2099 * node itself as well.
2100 * To use this new node's memory, further consideration will be
2103 zone->wait_table = (wait_queue_head_t *)vmalloc(alloc_size);
2105 if (!zone->wait_table)
2108 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
2109 init_waitqueue_head(zone->wait_table + i);
2114 static __meminit void zone_pcp_init(struct zone *zone)
2117 unsigned long batch = zone_batchsize(zone);
2119 for (cpu = 0; cpu < NR_CPUS; cpu++) {
2121 /* Early boot. Slab allocator not functional yet */
2122 zone_pcp(zone, cpu) = &boot_pageset[cpu];
2123 setup_pageset(&boot_pageset[cpu],0);
2125 setup_pageset(zone_pcp(zone,cpu), batch);
2128 if (zone->present_pages)
2129 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
2130 zone->name, zone->present_pages, batch);
2133 __meminit int init_currently_empty_zone(struct zone *zone,
2134 unsigned long zone_start_pfn,
2137 struct pglist_data *pgdat = zone->zone_pgdat;
2139 ret = zone_wait_table_init(zone, size);
2142 pgdat->nr_zones = zone_idx(zone) + 1;
2144 zone->zone_start_pfn = zone_start_pfn;
2146 memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn);
2148 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
2153 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
2155 * Basic iterator support. Return the first range of PFNs for a node
2156 * Note: nid == MAX_NUMNODES returns first region regardless of node
2158 static int __init first_active_region_index_in_nid(int nid)
2162 for (i = 0; i < nr_nodemap_entries; i++)
2163 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
2170 * Basic iterator support. Return the next active range of PFNs for a node
2171 * Note: nid == MAX_NUMNODES returns next region regardles of node
2173 static int __init next_active_region_index_in_nid(int index, int nid)
2175 for (index = index + 1; index < nr_nodemap_entries; index++)
2176 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
2182 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
2184 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
2185 * Architectures may implement their own version but if add_active_range()
2186 * was used and there are no special requirements, this is a convenient
2189 int __init early_pfn_to_nid(unsigned long pfn)
2193 for (i = 0; i < nr_nodemap_entries; i++) {
2194 unsigned long start_pfn = early_node_map[i].start_pfn;
2195 unsigned long end_pfn = early_node_map[i].end_pfn;
2197 if (start_pfn <= pfn && pfn < end_pfn)
2198 return early_node_map[i].nid;
2203 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
2205 /* Basic iterator support to walk early_node_map[] */
2206 #define for_each_active_range_index_in_nid(i, nid) \
2207 for (i = first_active_region_index_in_nid(nid); i != -1; \
2208 i = next_active_region_index_in_nid(i, nid))
2211 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
2212 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
2213 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
2215 * If an architecture guarantees that all ranges registered with
2216 * add_active_ranges() contain no holes and may be freed, this
2217 * this function may be used instead of calling free_bootmem() manually.
2219 void __init free_bootmem_with_active_regions(int nid,
2220 unsigned long max_low_pfn)
2224 for_each_active_range_index_in_nid(i, nid) {
2225 unsigned long size_pages = 0;
2226 unsigned long end_pfn = early_node_map[i].end_pfn;
2228 if (early_node_map[i].start_pfn >= max_low_pfn)
2231 if (end_pfn > max_low_pfn)
2232 end_pfn = max_low_pfn;
2234 size_pages = end_pfn - early_node_map[i].start_pfn;
2235 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
2236 PFN_PHYS(early_node_map[i].start_pfn),
2237 size_pages << PAGE_SHIFT);
2242 * sparse_memory_present_with_active_regions - Call memory_present for each active range
2243 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
2245 * If an architecture guarantees that all ranges registered with
2246 * add_active_ranges() contain no holes and may be freed, this
2247 * function may be used instead of calling memory_present() manually.
2249 void __init sparse_memory_present_with_active_regions(int nid)
2253 for_each_active_range_index_in_nid(i, nid)
2254 memory_present(early_node_map[i].nid,
2255 early_node_map[i].start_pfn,
2256 early_node_map[i].end_pfn);
2260 * push_node_boundaries - Push node boundaries to at least the requested boundary
2261 * @nid: The nid of the node to push the boundary for
2262 * @start_pfn: The start pfn of the node
2263 * @end_pfn: The end pfn of the node
2265 * In reserve-based hot-add, mem_map is allocated that is unused until hotadd
2266 * time. Specifically, on x86_64, SRAT will report ranges that can potentially
2267 * be hotplugged even though no physical memory exists. This function allows
2268 * an arch to push out the node boundaries so mem_map is allocated that can
2271 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
2272 void __init push_node_boundaries(unsigned int nid,
2273 unsigned long start_pfn, unsigned long end_pfn)
2275 printk(KERN_DEBUG "Entering push_node_boundaries(%u, %lu, %lu)\n",
2276 nid, start_pfn, end_pfn);
2278 /* Initialise the boundary for this node if necessary */
2279 if (node_boundary_end_pfn[nid] == 0)
2280 node_boundary_start_pfn[nid] = -1UL;
2282 /* Update the boundaries */
2283 if (node_boundary_start_pfn[nid] > start_pfn)
2284 node_boundary_start_pfn[nid] = start_pfn;
2285 if (node_boundary_end_pfn[nid] < end_pfn)
2286 node_boundary_end_pfn[nid] = end_pfn;
2289 /* If necessary, push the node boundary out for reserve hotadd */
2290 static void __init account_node_boundary(unsigned int nid,
2291 unsigned long *start_pfn, unsigned long *end_pfn)
2293 printk(KERN_DEBUG "Entering account_node_boundary(%u, %lu, %lu)\n",
2294 nid, *start_pfn, *end_pfn);
2296 /* Return if boundary information has not been provided */
2297 if (node_boundary_end_pfn[nid] == 0)
2300 /* Check the boundaries and update if necessary */
2301 if (node_boundary_start_pfn[nid] < *start_pfn)
2302 *start_pfn = node_boundary_start_pfn[nid];
2303 if (node_boundary_end_pfn[nid] > *end_pfn)
2304 *end_pfn = node_boundary_end_pfn[nid];
2307 void __init push_node_boundaries(unsigned int nid,
2308 unsigned long start_pfn, unsigned long end_pfn) {}
2310 static void __init account_node_boundary(unsigned int nid,
2311 unsigned long *start_pfn, unsigned long *end_pfn) {}
2316 * get_pfn_range_for_nid - Return the start and end page frames for a node
2317 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
2318 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
2319 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
2321 * It returns the start and end page frame of a node based on information
2322 * provided by an arch calling add_active_range(). If called for a node
2323 * with no available memory, a warning is printed and the start and end
2326 void __init get_pfn_range_for_nid(unsigned int nid,
2327 unsigned long *start_pfn, unsigned long *end_pfn)
2333 for_each_active_range_index_in_nid(i, nid) {
2334 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
2335 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
2338 if (*start_pfn == -1UL) {
2339 printk(KERN_WARNING "Node %u active with no memory\n", nid);
2343 /* Push the node boundaries out if requested */
2344 account_node_boundary(nid, start_pfn, end_pfn);
2348 * Return the number of pages a zone spans in a node, including holes
2349 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
2351 unsigned long __init zone_spanned_pages_in_node(int nid,
2352 unsigned long zone_type,
2353 unsigned long *ignored)
2355 unsigned long node_start_pfn, node_end_pfn;
2356 unsigned long zone_start_pfn, zone_end_pfn;
2358 /* Get the start and end of the node and zone */
2359 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
2360 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
2361 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
2363 /* Check that this node has pages within the zone's required range */
2364 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
2367 /* Move the zone boundaries inside the node if necessary */
2368 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
2369 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
2371 /* Return the spanned pages */
2372 return zone_end_pfn - zone_start_pfn;
2376 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
2377 * then all holes in the requested range will be accounted for.
2379 unsigned long __init __absent_pages_in_range(int nid,
2380 unsigned long range_start_pfn,
2381 unsigned long range_end_pfn)
2384 unsigned long prev_end_pfn = 0, hole_pages = 0;
2385 unsigned long start_pfn;
2387 /* Find the end_pfn of the first active range of pfns in the node */
2388 i = first_active_region_index_in_nid(nid);
2392 /* Account for ranges before physical memory on this node */
2393 if (early_node_map[i].start_pfn > range_start_pfn)
2394 hole_pages = early_node_map[i].start_pfn - range_start_pfn;
2396 prev_end_pfn = early_node_map[i].start_pfn;
2398 /* Find all holes for the zone within the node */
2399 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
2401 /* No need to continue if prev_end_pfn is outside the zone */
2402 if (prev_end_pfn >= range_end_pfn)
2405 /* Make sure the end of the zone is not within the hole */
2406 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
2407 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
2409 /* Update the hole size cound and move on */
2410 if (start_pfn > range_start_pfn) {
2411 BUG_ON(prev_end_pfn > start_pfn);
2412 hole_pages += start_pfn - prev_end_pfn;
2414 prev_end_pfn = early_node_map[i].end_pfn;
2417 /* Account for ranges past physical memory on this node */
2418 if (range_end_pfn > prev_end_pfn)
2419 hole_pages += range_end_pfn -
2420 max(range_start_pfn, prev_end_pfn);
2426 * absent_pages_in_range - Return number of page frames in holes within a range
2427 * @start_pfn: The start PFN to start searching for holes
2428 * @end_pfn: The end PFN to stop searching for holes
2430 * It returns the number of pages frames in memory holes within a range.
2432 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
2433 unsigned long end_pfn)
2435 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
2438 /* Return the number of page frames in holes in a zone on a node */
2439 unsigned long __init zone_absent_pages_in_node(int nid,
2440 unsigned long zone_type,
2441 unsigned long *ignored)
2443 unsigned long node_start_pfn, node_end_pfn;
2444 unsigned long zone_start_pfn, zone_end_pfn;
2446 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
2447 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
2449 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
2452 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
2456 static inline unsigned long zone_spanned_pages_in_node(int nid,
2457 unsigned long zone_type,
2458 unsigned long *zones_size)
2460 return zones_size[zone_type];
2463 static inline unsigned long zone_absent_pages_in_node(int nid,
2464 unsigned long zone_type,
2465 unsigned long *zholes_size)
2470 return zholes_size[zone_type];
2475 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
2476 unsigned long *zones_size, unsigned long *zholes_size)
2478 unsigned long realtotalpages, totalpages = 0;
2481 for (i = 0; i < MAX_NR_ZONES; i++)
2482 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
2484 pgdat->node_spanned_pages = totalpages;
2486 realtotalpages = totalpages;
2487 for (i = 0; i < MAX_NR_ZONES; i++)
2489 zone_absent_pages_in_node(pgdat->node_id, i,
2491 pgdat->node_present_pages = realtotalpages;
2492 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
2497 * Set up the zone data structures:
2498 * - mark all pages reserved
2499 * - mark all memory queues empty
2500 * - clear the memory bitmaps
2502 static void __meminit free_area_init_core(struct pglist_data *pgdat,
2503 unsigned long *zones_size, unsigned long *zholes_size)
2506 int nid = pgdat->node_id;
2507 unsigned long zone_start_pfn = pgdat->node_start_pfn;
2510 pgdat_resize_init(pgdat);
2511 pgdat->nr_zones = 0;
2512 init_waitqueue_head(&pgdat->kswapd_wait);
2513 pgdat->kswapd_max_order = 0;
2515 for (j = 0; j < MAX_NR_ZONES; j++) {
2516 struct zone *zone = pgdat->node_zones + j;
2517 unsigned long size, realsize, memmap_pages;
2519 size = zone_spanned_pages_in_node(nid, j, zones_size);
2520 realsize = size - zone_absent_pages_in_node(nid, j,
2524 * Adjust realsize so that it accounts for how much memory
2525 * is used by this zone for memmap. This affects the watermark
2526 * and per-cpu initialisations
2528 memmap_pages = (size * sizeof(struct page)) >> PAGE_SHIFT;
2529 if (realsize >= memmap_pages) {
2530 realsize -= memmap_pages;
2532 " %s zone: %lu pages used for memmap\n",
2533 zone_names[j], memmap_pages);
2536 " %s zone: %lu pages exceeds realsize %lu\n",
2537 zone_names[j], memmap_pages, realsize);
2539 /* Account for reserved DMA pages */
2540 if (j == ZONE_DMA && realsize > dma_reserve) {
2541 realsize -= dma_reserve;
2542 printk(KERN_DEBUG " DMA zone: %lu pages reserved\n",
2546 if (!is_highmem_idx(j))
2547 nr_kernel_pages += realsize;
2548 nr_all_pages += realsize;
2550 zone->spanned_pages = size;
2551 zone->present_pages = realsize;
2554 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
2556 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
2558 zone->name = zone_names[j];
2559 spin_lock_init(&zone->lock);
2560 spin_lock_init(&zone->lru_lock);
2561 zone_seqlock_init(zone);
2562 zone->zone_pgdat = pgdat;
2563 zone->free_pages = 0;
2565 zone->prev_priority = DEF_PRIORITY;
2567 zone_pcp_init(zone);
2568 INIT_LIST_HEAD(&zone->active_list);
2569 INIT_LIST_HEAD(&zone->inactive_list);
2570 zone->nr_scan_active = 0;
2571 zone->nr_scan_inactive = 0;
2572 zone->nr_active = 0;
2573 zone->nr_inactive = 0;
2574 zap_zone_vm_stats(zone);
2575 atomic_set(&zone->reclaim_in_progress, 0);
2579 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
2581 zone_start_pfn += size;
2585 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
2587 /* Skip empty nodes */
2588 if (!pgdat->node_spanned_pages)
2591 #ifdef CONFIG_FLAT_NODE_MEM_MAP
2592 /* ia64 gets its own node_mem_map, before this, without bootmem */
2593 if (!pgdat->node_mem_map) {
2594 unsigned long size, start, end;
2598 * The zone's endpoints aren't required to be MAX_ORDER
2599 * aligned but the node_mem_map endpoints must be in order
2600 * for the buddy allocator to function correctly.
2602 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
2603 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
2604 end = ALIGN(end, MAX_ORDER_NR_PAGES);
2605 size = (end - start) * sizeof(struct page);
2606 map = alloc_remap(pgdat->node_id, size);
2608 map = alloc_bootmem_node(pgdat, size);
2609 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
2611 #ifdef CONFIG_FLATMEM
2613 * With no DISCONTIG, the global mem_map is just set as node 0's
2615 if (pgdat == NODE_DATA(0)) {
2616 mem_map = NODE_DATA(0)->node_mem_map;
2617 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
2618 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
2619 mem_map -= pgdat->node_start_pfn;
2620 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
2623 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
2626 void __meminit free_area_init_node(int nid, struct pglist_data *pgdat,
2627 unsigned long *zones_size, unsigned long node_start_pfn,
2628 unsigned long *zholes_size)
2630 pgdat->node_id = nid;
2631 pgdat->node_start_pfn = node_start_pfn;
2632 calculate_node_totalpages(pgdat, zones_size, zholes_size);
2634 alloc_node_mem_map(pgdat);
2636 free_area_init_core(pgdat, zones_size, zholes_size);
2639 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
2641 * add_active_range - Register a range of PFNs backed by physical memory
2642 * @nid: The node ID the range resides on
2643 * @start_pfn: The start PFN of the available physical memory
2644 * @end_pfn: The end PFN of the available physical memory
2646 * These ranges are stored in an early_node_map[] and later used by
2647 * free_area_init_nodes() to calculate zone sizes and holes. If the
2648 * range spans a memory hole, it is up to the architecture to ensure
2649 * the memory is not freed by the bootmem allocator. If possible
2650 * the range being registered will be merged with existing ranges.
2652 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
2653 unsigned long end_pfn)
2657 printk(KERN_DEBUG "Entering add_active_range(%d, %lu, %lu) "
2658 "%d entries of %d used\n",
2659 nid, start_pfn, end_pfn,
2660 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
2662 /* Merge with existing active regions if possible */
2663 for (i = 0; i < nr_nodemap_entries; i++) {
2664 if (early_node_map[i].nid != nid)
2667 /* Skip if an existing region covers this new one */
2668 if (start_pfn >= early_node_map[i].start_pfn &&
2669 end_pfn <= early_node_map[i].end_pfn)
2672 /* Merge forward if suitable */
2673 if (start_pfn <= early_node_map[i].end_pfn &&
2674 end_pfn > early_node_map[i].end_pfn) {
2675 early_node_map[i].end_pfn = end_pfn;
2679 /* Merge backward if suitable */
2680 if (start_pfn < early_node_map[i].end_pfn &&
2681 end_pfn >= early_node_map[i].start_pfn) {
2682 early_node_map[i].start_pfn = start_pfn;
2687 /* Check that early_node_map is large enough */
2688 if (i >= MAX_ACTIVE_REGIONS) {
2689 printk(KERN_CRIT "More than %d memory regions, truncating\n",
2690 MAX_ACTIVE_REGIONS);
2694 early_node_map[i].nid = nid;
2695 early_node_map[i].start_pfn = start_pfn;
2696 early_node_map[i].end_pfn = end_pfn;
2697 nr_nodemap_entries = i + 1;
2701 * shrink_active_range - Shrink an existing registered range of PFNs
2702 * @nid: The node id the range is on that should be shrunk
2703 * @old_end_pfn: The old end PFN of the range
2704 * @new_end_pfn: The new PFN of the range
2706 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
2707 * The map is kept at the end physical page range that has already been
2708 * registered with add_active_range(). This function allows an arch to shrink
2709 * an existing registered range.
2711 void __init shrink_active_range(unsigned int nid, unsigned long old_end_pfn,
2712 unsigned long new_end_pfn)
2716 /* Find the old active region end and shrink */
2717 for_each_active_range_index_in_nid(i, nid)
2718 if (early_node_map[i].end_pfn == old_end_pfn) {
2719 early_node_map[i].end_pfn = new_end_pfn;
2725 * remove_all_active_ranges - Remove all currently registered regions
2727 * During discovery, it may be found that a table like SRAT is invalid
2728 * and an alternative discovery method must be used. This function removes
2729 * all currently registered regions.
2731 void __init remove_all_active_ranges(void)
2733 memset(early_node_map, 0, sizeof(early_node_map));
2734 nr_nodemap_entries = 0;
2735 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
2736 memset(node_boundary_start_pfn, 0, sizeof(node_boundary_start_pfn));
2737 memset(node_boundary_end_pfn, 0, sizeof(node_boundary_end_pfn));
2738 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
2741 /* Compare two active node_active_regions */
2742 static int __init cmp_node_active_region(const void *a, const void *b)
2744 struct node_active_region *arange = (struct node_active_region *)a;
2745 struct node_active_region *brange = (struct node_active_region *)b;
2747 /* Done this way to avoid overflows */
2748 if (arange->start_pfn > brange->start_pfn)
2750 if (arange->start_pfn < brange->start_pfn)
2756 /* sort the node_map by start_pfn */
2757 static void __init sort_node_map(void)
2759 sort(early_node_map, (size_t)nr_nodemap_entries,
2760 sizeof(struct node_active_region),
2761 cmp_node_active_region, NULL);
2764 /* Find the lowest pfn for a node. This depends on a sorted early_node_map */
2765 unsigned long __init find_min_pfn_for_node(unsigned long nid)
2769 /* Regions in the early_node_map can be in any order */
2772 /* Assuming a sorted map, the first range found has the starting pfn */
2773 for_each_active_range_index_in_nid(i, nid)
2774 return early_node_map[i].start_pfn;
2776 printk(KERN_WARNING "Could not find start_pfn for node %lu\n", nid);
2781 * find_min_pfn_with_active_regions - Find the minimum PFN registered
2783 * It returns the minimum PFN based on information provided via
2784 * add_active_range().
2786 unsigned long __init find_min_pfn_with_active_regions(void)
2788 return find_min_pfn_for_node(MAX_NUMNODES);
2792 * find_max_pfn_with_active_regions - Find the maximum PFN registered
2794 * It returns the maximum PFN based on information provided via
2795 * add_active_range().
2797 unsigned long __init find_max_pfn_with_active_regions(void)
2800 unsigned long max_pfn = 0;
2802 for (i = 0; i < nr_nodemap_entries; i++)
2803 max_pfn = max(max_pfn, early_node_map[i].end_pfn);
2809 * free_area_init_nodes - Initialise all pg_data_t and zone data
2810 * @max_zone_pfn: an array of max PFNs for each zone
2812 * This will call free_area_init_node() for each active node in the system.
2813 * Using the page ranges provided by add_active_range(), the size of each
2814 * zone in each node and their holes is calculated. If the maximum PFN
2815 * between two adjacent zones match, it is assumed that the zone is empty.
2816 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
2817 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
2818 * starts where the previous one ended. For example, ZONE_DMA32 starts
2819 * at arch_max_dma_pfn.
2821 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
2826 /* Record where the zone boundaries are */
2827 memset(arch_zone_lowest_possible_pfn, 0,
2828 sizeof(arch_zone_lowest_possible_pfn));
2829 memset(arch_zone_highest_possible_pfn, 0,
2830 sizeof(arch_zone_highest_possible_pfn));
2831 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
2832 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
2833 for (i = 1; i < MAX_NR_ZONES; i++) {
2834 arch_zone_lowest_possible_pfn[i] =
2835 arch_zone_highest_possible_pfn[i-1];
2836 arch_zone_highest_possible_pfn[i] =
2837 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
2840 /* Print out the zone ranges */
2841 printk("Zone PFN ranges:\n");
2842 for (i = 0; i < MAX_NR_ZONES; i++)
2843 printk(" %-8s %8lu -> %8lu\n",
2845 arch_zone_lowest_possible_pfn[i],
2846 arch_zone_highest_possible_pfn[i]);
2848 /* Print out the early_node_map[] */
2849 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
2850 for (i = 0; i < nr_nodemap_entries; i++)
2851 printk(" %3d: %8lu -> %8lu\n", early_node_map[i].nid,
2852 early_node_map[i].start_pfn,
2853 early_node_map[i].end_pfn);
2855 /* Initialise every node */
2856 for_each_online_node(nid) {
2857 pg_data_t *pgdat = NODE_DATA(nid);
2858 free_area_init_node(nid, pgdat, NULL,
2859 find_min_pfn_for_node(nid), NULL);
2862 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
2865 * set_dma_reserve - set the specified number of pages reserved in the first zone
2866 * @new_dma_reserve: The number of pages to mark reserved
2868 * The per-cpu batchsize and zone watermarks are determined by present_pages.
2869 * In the DMA zone, a significant percentage may be consumed by kernel image
2870 * and other unfreeable allocations which can skew the watermarks badly. This
2871 * function may optionally be used to account for unfreeable pages in the
2872 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
2873 * smaller per-cpu batchsize.
2875 void __init set_dma_reserve(unsigned long new_dma_reserve)
2877 dma_reserve = new_dma_reserve;
2880 #ifndef CONFIG_NEED_MULTIPLE_NODES
2881 static bootmem_data_t contig_bootmem_data;
2882 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
2884 EXPORT_SYMBOL(contig_page_data);
2887 void __init free_area_init(unsigned long *zones_size)
2889 free_area_init_node(0, NODE_DATA(0), zones_size,
2890 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
2893 #ifdef CONFIG_HOTPLUG_CPU
2894 static int page_alloc_cpu_notify(struct notifier_block *self,
2895 unsigned long action, void *hcpu)
2897 int cpu = (unsigned long)hcpu;
2899 if (action == CPU_DEAD) {
2900 local_irq_disable();
2902 vm_events_fold_cpu(cpu);
2904 refresh_cpu_vm_stats(cpu);
2908 #endif /* CONFIG_HOTPLUG_CPU */
2910 void __init page_alloc_init(void)
2912 hotcpu_notifier(page_alloc_cpu_notify, 0);
2916 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
2917 * or min_free_kbytes changes.
2919 static void calculate_totalreserve_pages(void)
2921 struct pglist_data *pgdat;
2922 unsigned long reserve_pages = 0;
2923 enum zone_type i, j;
2925 for_each_online_pgdat(pgdat) {
2926 for (i = 0; i < MAX_NR_ZONES; i++) {
2927 struct zone *zone = pgdat->node_zones + i;
2928 unsigned long max = 0;
2930 /* Find valid and maximum lowmem_reserve in the zone */
2931 for (j = i; j < MAX_NR_ZONES; j++) {
2932 if (zone->lowmem_reserve[j] > max)
2933 max = zone->lowmem_reserve[j];
2936 /* we treat pages_high as reserved pages. */
2937 max += zone->pages_high;
2939 if (max > zone->present_pages)
2940 max = zone->present_pages;
2941 reserve_pages += max;
2944 totalreserve_pages = reserve_pages;
2948 * setup_per_zone_lowmem_reserve - called whenever
2949 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
2950 * has a correct pages reserved value, so an adequate number of
2951 * pages are left in the zone after a successful __alloc_pages().
2953 static void setup_per_zone_lowmem_reserve(void)
2955 struct pglist_data *pgdat;
2956 enum zone_type j, idx;
2958 for_each_online_pgdat(pgdat) {
2959 for (j = 0; j < MAX_NR_ZONES; j++) {
2960 struct zone *zone = pgdat->node_zones + j;
2961 unsigned long present_pages = zone->present_pages;
2963 zone->lowmem_reserve[j] = 0;
2967 struct zone *lower_zone;
2971 if (sysctl_lowmem_reserve_ratio[idx] < 1)
2972 sysctl_lowmem_reserve_ratio[idx] = 1;
2974 lower_zone = pgdat->node_zones + idx;
2975 lower_zone->lowmem_reserve[j] = present_pages /
2976 sysctl_lowmem_reserve_ratio[idx];
2977 present_pages += lower_zone->present_pages;
2982 /* update totalreserve_pages */
2983 calculate_totalreserve_pages();
2987 * setup_per_zone_pages_min - called when min_free_kbytes changes.
2989 * Ensures that the pages_{min,low,high} values for each zone are set correctly
2990 * with respect to min_free_kbytes.
2992 void setup_per_zone_pages_min(void)
2994 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
2995 unsigned long lowmem_pages = 0;
2997 unsigned long flags;
2999 /* Calculate total number of !ZONE_HIGHMEM pages */
3000 for_each_zone(zone) {
3001 if (!is_highmem(zone))
3002 lowmem_pages += zone->present_pages;
3005 for_each_zone(zone) {
3008 spin_lock_irqsave(&zone->lru_lock, flags);
3009 tmp = (u64)pages_min * zone->present_pages;
3010 do_div(tmp, lowmem_pages);
3011 if (is_highmem(zone)) {
3013 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
3014 * need highmem pages, so cap pages_min to a small
3017 * The (pages_high-pages_low) and (pages_low-pages_min)
3018 * deltas controls asynch page reclaim, and so should
3019 * not be capped for highmem.
3023 min_pages = zone->present_pages / 1024;
3024 if (min_pages < SWAP_CLUSTER_MAX)
3025 min_pages = SWAP_CLUSTER_MAX;
3026 if (min_pages > 128)
3028 zone->pages_min = min_pages;
3031 * If it's a lowmem zone, reserve a number of pages
3032 * proportionate to the zone's size.
3034 zone->pages_min = tmp;
3037 zone->pages_low = zone->pages_min + (tmp >> 2);
3038 zone->pages_high = zone->pages_min + (tmp >> 1);
3039 spin_unlock_irqrestore(&zone->lru_lock, flags);
3042 /* update totalreserve_pages */
3043 calculate_totalreserve_pages();
3047 * Initialise min_free_kbytes.
3049 * For small machines we want it small (128k min). For large machines
3050 * we want it large (64MB max). But it is not linear, because network
3051 * bandwidth does not increase linearly with machine size. We use
3053 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
3054 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
3070 static int __init init_per_zone_pages_min(void)
3072 unsigned long lowmem_kbytes;
3074 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
3076 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
3077 if (min_free_kbytes < 128)
3078 min_free_kbytes = 128;
3079 if (min_free_kbytes > 65536)
3080 min_free_kbytes = 65536;
3081 setup_per_zone_pages_min();
3082 setup_per_zone_lowmem_reserve();
3085 module_init(init_per_zone_pages_min)
3088 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
3089 * that we can call two helper functions whenever min_free_kbytes
3092 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
3093 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
3095 proc_dointvec(table, write, file, buffer, length, ppos);
3096 setup_per_zone_pages_min();
3101 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
3102 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
3107 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
3112 zone->min_unmapped_pages = (zone->present_pages *
3113 sysctl_min_unmapped_ratio) / 100;
3117 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
3118 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
3123 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
3128 zone->min_slab_pages = (zone->present_pages *
3129 sysctl_min_slab_ratio) / 100;
3135 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
3136 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
3137 * whenever sysctl_lowmem_reserve_ratio changes.
3139 * The reserve ratio obviously has absolutely no relation with the
3140 * pages_min watermarks. The lowmem reserve ratio can only make sense
3141 * if in function of the boot time zone sizes.
3143 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
3144 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
3146 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
3147 setup_per_zone_lowmem_reserve();
3152 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
3153 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
3154 * can have before it gets flushed back to buddy allocator.
3157 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
3158 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
3164 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
3165 if (!write || (ret == -EINVAL))
3167 for_each_zone(zone) {
3168 for_each_online_cpu(cpu) {
3170 high = zone->present_pages / percpu_pagelist_fraction;
3171 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
3177 int hashdist = HASHDIST_DEFAULT;
3180 static int __init set_hashdist(char *str)
3184 hashdist = simple_strtoul(str, &str, 0);
3187 __setup("hashdist=", set_hashdist);
3191 * allocate a large system hash table from bootmem
3192 * - it is assumed that the hash table must contain an exact power-of-2
3193 * quantity of entries
3194 * - limit is the number of hash buckets, not the total allocation size
3196 void *__init alloc_large_system_hash(const char *tablename,
3197 unsigned long bucketsize,
3198 unsigned long numentries,
3201 unsigned int *_hash_shift,
3202 unsigned int *_hash_mask,
3203 unsigned long limit)
3205 unsigned long long max = limit;
3206 unsigned long log2qty, size;
3209 /* allow the kernel cmdline to have a say */
3211 /* round applicable memory size up to nearest megabyte */
3212 numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages;
3213 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
3214 numentries >>= 20 - PAGE_SHIFT;
3215 numentries <<= 20 - PAGE_SHIFT;
3217 /* limit to 1 bucket per 2^scale bytes of low memory */
3218 if (scale > PAGE_SHIFT)
3219 numentries >>= (scale - PAGE_SHIFT);
3221 numentries <<= (PAGE_SHIFT - scale);
3223 numentries = roundup_pow_of_two(numentries);
3225 /* limit allocation size to 1/16 total memory by default */
3227 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
3228 do_div(max, bucketsize);
3231 if (numentries > max)
3234 log2qty = long_log2(numentries);
3237 size = bucketsize << log2qty;
3238 if (flags & HASH_EARLY)
3239 table = alloc_bootmem(size);
3241 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
3243 unsigned long order;
3244 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
3246 table = (void*) __get_free_pages(GFP_ATOMIC, order);
3248 } while (!table && size > PAGE_SIZE && --log2qty);
3251 panic("Failed to allocate %s hash table\n", tablename);
3253 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
3256 long_log2(size) - PAGE_SHIFT,
3260 *_hash_shift = log2qty;
3262 *_hash_mask = (1 << log2qty) - 1;
3267 #ifdef CONFIG_OUT_OF_LINE_PFN_TO_PAGE
3268 struct page *pfn_to_page(unsigned long pfn)
3270 return __pfn_to_page(pfn);
3272 unsigned long page_to_pfn(struct page *page)
3274 return __page_to_pfn(page);
3276 EXPORT_SYMBOL(pfn_to_page);
3277 EXPORT_SYMBOL(page_to_pfn);
3278 #endif /* CONFIG_OUT_OF_LINE_PFN_TO_PAGE */
3280 #if MAX_NUMNODES > 1
3282 * Find the highest possible node id.
3284 int highest_possible_node_id(void)
3287 unsigned int highest = 0;
3289 for_each_node_mask(node, node_possible_map)
3293 EXPORT_SYMBOL(highest_possible_node_id);