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>
43 #include <asm/tlbflush.h>
44 #include <asm/div64.h>
48 * MCD - HACK: Find somewhere to initialize this EARLY, or make this
51 nodemask_t node_online_map __read_mostly = { { [0] = 1UL } };
52 EXPORT_SYMBOL(node_online_map);
53 nodemask_t node_possible_map __read_mostly = NODE_MASK_ALL;
54 EXPORT_SYMBOL(node_possible_map);
55 unsigned long totalram_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] = {
75 #ifdef CONFIG_ZONE_DMA32
83 EXPORT_SYMBOL(totalram_pages);
86 * Used by page_zone() to look up the address of the struct zone whose
87 * id is encoded in the upper bits of page->flags
89 struct zone *zone_table[1 << ZONETABLE_SHIFT] __read_mostly;
90 EXPORT_SYMBOL(zone_table);
92 static char *zone_names[MAX_NR_ZONES] = {
94 #ifdef CONFIG_ZONE_DMA32
103 int min_free_kbytes = 1024;
105 unsigned long __meminitdata nr_kernel_pages;
106 unsigned long __meminitdata nr_all_pages;
107 static unsigned long __initdata dma_reserve;
109 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
111 * MAX_ACTIVE_REGIONS determines the maxmimum number of distinct
112 * ranges of memory (RAM) that may be registered with add_active_range().
113 * Ranges passed to add_active_range() will be merged if possible
114 * so the number of times add_active_range() can be called is
115 * related to the number of nodes and the number of holes
117 #ifdef CONFIG_MAX_ACTIVE_REGIONS
118 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
119 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
121 #if MAX_NUMNODES >= 32
122 /* If there can be many nodes, allow up to 50 holes per node */
123 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
125 /* By default, allow up to 256 distinct regions */
126 #define MAX_ACTIVE_REGIONS 256
130 struct node_active_region __initdata early_node_map[MAX_ACTIVE_REGIONS];
131 int __initdata nr_nodemap_entries;
132 unsigned long __initdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
133 unsigned long __initdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
134 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
135 unsigned long __initdata node_boundary_start_pfn[MAX_NUMNODES];
136 unsigned long __initdata node_boundary_end_pfn[MAX_NUMNODES];
137 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
138 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
140 #ifdef CONFIG_DEBUG_VM
141 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
145 unsigned long pfn = page_to_pfn(page);
148 seq = zone_span_seqbegin(zone);
149 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
151 else if (pfn < zone->zone_start_pfn)
153 } while (zone_span_seqretry(zone, seq));
158 static int page_is_consistent(struct zone *zone, struct page *page)
160 #ifdef CONFIG_HOLES_IN_ZONE
161 if (!pfn_valid(page_to_pfn(page)))
164 if (zone != page_zone(page))
170 * Temporary debugging check for pages not lying within a given zone.
172 static int bad_range(struct zone *zone, struct page *page)
174 if (page_outside_zone_boundaries(zone, page))
176 if (!page_is_consistent(zone, page))
182 static inline int bad_range(struct zone *zone, struct page *page)
188 static void bad_page(struct page *page)
190 printk(KERN_EMERG "Bad page state in process '%s'\n"
191 KERN_EMERG "page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d\n"
192 KERN_EMERG "Trying to fix it up, but a reboot is needed\n"
193 KERN_EMERG "Backtrace:\n",
194 current->comm, page, (int)(2*sizeof(unsigned long)),
195 (unsigned long)page->flags, page->mapping,
196 page_mapcount(page), page_count(page));
198 page->flags &= ~(1 << PG_lru |
208 set_page_count(page, 0);
209 reset_page_mapcount(page);
210 page->mapping = NULL;
211 add_taint(TAINT_BAD_PAGE);
215 * Higher-order pages are called "compound pages". They are structured thusly:
217 * The first PAGE_SIZE page is called the "head page".
219 * The remaining PAGE_SIZE pages are called "tail pages".
221 * All pages have PG_compound set. All pages have their ->private pointing at
222 * the head page (even the head page has this).
224 * The first tail page's ->lru.next holds the address of the compound page's
225 * put_page() function. Its ->lru.prev holds the order of allocation.
226 * This usage means that zero-order pages may not be compound.
229 static void free_compound_page(struct page *page)
231 __free_pages_ok(page, (unsigned long)page[1].lru.prev);
234 static void prep_compound_page(struct page *page, unsigned long order)
237 int nr_pages = 1 << order;
239 page[1].lru.next = (void *)free_compound_page; /* set dtor */
240 page[1].lru.prev = (void *)order;
241 for (i = 0; i < nr_pages; i++) {
242 struct page *p = page + i;
244 __SetPageCompound(p);
245 set_page_private(p, (unsigned long)page);
249 static void destroy_compound_page(struct page *page, unsigned long order)
252 int nr_pages = 1 << order;
254 if (unlikely((unsigned long)page[1].lru.prev != order))
257 for (i = 0; i < nr_pages; i++) {
258 struct page *p = page + i;
260 if (unlikely(!PageCompound(p) |
261 (page_private(p) != (unsigned long)page)))
263 __ClearPageCompound(p);
267 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
271 VM_BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
273 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
274 * and __GFP_HIGHMEM from hard or soft interrupt context.
276 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
277 for (i = 0; i < (1 << order); i++)
278 clear_highpage(page + i);
282 * function for dealing with page's order in buddy system.
283 * zone->lock is already acquired when we use these.
284 * So, we don't need atomic page->flags operations here.
286 static inline unsigned long page_order(struct page *page)
288 return page_private(page);
291 static inline void set_page_order(struct page *page, int order)
293 set_page_private(page, order);
294 __SetPageBuddy(page);
297 static inline void rmv_page_order(struct page *page)
299 __ClearPageBuddy(page);
300 set_page_private(page, 0);
304 * Locate the struct page for both the matching buddy in our
305 * pair (buddy1) and the combined O(n+1) page they form (page).
307 * 1) Any buddy B1 will have an order O twin B2 which satisfies
308 * the following equation:
310 * For example, if the starting buddy (buddy2) is #8 its order
312 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
314 * 2) Any buddy B will have an order O+1 parent P which
315 * satisfies the following equation:
318 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
320 static inline struct page *
321 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
323 unsigned long buddy_idx = page_idx ^ (1 << order);
325 return page + (buddy_idx - page_idx);
328 static inline unsigned long
329 __find_combined_index(unsigned long page_idx, unsigned int order)
331 return (page_idx & ~(1 << order));
335 * This function checks whether a page is free && is the buddy
336 * we can do coalesce a page and its buddy if
337 * (a) the buddy is not in a hole &&
338 * (b) the buddy is in the buddy system &&
339 * (c) a page and its buddy have the same order &&
340 * (d) a page and its buddy are in the same zone.
342 * For recording whether a page is in the buddy system, we use PG_buddy.
343 * Setting, clearing, and testing PG_buddy is serialized by zone->lock.
345 * For recording page's order, we use page_private(page).
347 static inline int page_is_buddy(struct page *page, struct page *buddy,
350 #ifdef CONFIG_HOLES_IN_ZONE
351 if (!pfn_valid(page_to_pfn(buddy)))
355 if (page_zone_id(page) != page_zone_id(buddy))
358 if (PageBuddy(buddy) && page_order(buddy) == order) {
359 BUG_ON(page_count(buddy) != 0);
366 * Freeing function for a buddy system allocator.
368 * The concept of a buddy system is to maintain direct-mapped table
369 * (containing bit values) for memory blocks of various "orders".
370 * The bottom level table contains the map for the smallest allocatable
371 * units of memory (here, pages), and each level above it describes
372 * pairs of units from the levels below, hence, "buddies".
373 * At a high level, all that happens here is marking the table entry
374 * at the bottom level available, and propagating the changes upward
375 * as necessary, plus some accounting needed to play nicely with other
376 * parts of the VM system.
377 * At each level, we keep a list of pages, which are heads of continuous
378 * free pages of length of (1 << order) and marked with PG_buddy. Page's
379 * order is recorded in page_private(page) field.
380 * So when we are allocating or freeing one, we can derive the state of the
381 * other. That is, if we allocate a small block, and both were
382 * free, the remainder of the region must be split into blocks.
383 * If a block is freed, and its buddy is also free, then this
384 * triggers coalescing into a block of larger size.
389 static inline void __free_one_page(struct page *page,
390 struct zone *zone, unsigned int order)
392 unsigned long page_idx;
393 int order_size = 1 << order;
395 if (unlikely(PageCompound(page)))
396 destroy_compound_page(page, order);
398 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
400 VM_BUG_ON(page_idx & (order_size - 1));
401 VM_BUG_ON(bad_range(zone, page));
403 zone->free_pages += order_size;
404 while (order < MAX_ORDER-1) {
405 unsigned long combined_idx;
406 struct free_area *area;
409 buddy = __page_find_buddy(page, page_idx, order);
410 if (!page_is_buddy(page, buddy, order))
411 break; /* Move the buddy up one level. */
413 list_del(&buddy->lru);
414 area = zone->free_area + order;
416 rmv_page_order(buddy);
417 combined_idx = __find_combined_index(page_idx, order);
418 page = page + (combined_idx - page_idx);
419 page_idx = combined_idx;
422 set_page_order(page, order);
423 list_add(&page->lru, &zone->free_area[order].free_list);
424 zone->free_area[order].nr_free++;
427 static inline int free_pages_check(struct page *page)
429 if (unlikely(page_mapcount(page) |
430 (page->mapping != NULL) |
431 (page_count(page) != 0) |
445 __ClearPageDirty(page);
447 * For now, we report if PG_reserved was found set, but do not
448 * clear it, and do not free the page. But we shall soon need
449 * to do more, for when the ZERO_PAGE count wraps negative.
451 return PageReserved(page);
455 * Frees a list of pages.
456 * Assumes all pages on list are in same zone, and of same order.
457 * count is the number of pages to free.
459 * If the zone was previously in an "all pages pinned" state then look to
460 * see if this freeing clears that state.
462 * And clear the zone's pages_scanned counter, to hold off the "all pages are
463 * pinned" detection logic.
465 static void free_pages_bulk(struct zone *zone, int count,
466 struct list_head *list, int order)
468 spin_lock(&zone->lock);
469 zone->all_unreclaimable = 0;
470 zone->pages_scanned = 0;
474 VM_BUG_ON(list_empty(list));
475 page = list_entry(list->prev, struct page, lru);
476 /* have to delete it as __free_one_page list manipulates */
477 list_del(&page->lru);
478 __free_one_page(page, zone, order);
480 spin_unlock(&zone->lock);
483 static void free_one_page(struct zone *zone, struct page *page, int order)
485 spin_lock(&zone->lock);
486 zone->all_unreclaimable = 0;
487 zone->pages_scanned = 0;
488 __free_one_page(page, zone ,order);
489 spin_unlock(&zone->lock);
492 static void __free_pages_ok(struct page *page, unsigned int order)
498 arch_free_page(page, order);
499 if (!PageHighMem(page))
500 debug_check_no_locks_freed(page_address(page),
503 for (i = 0 ; i < (1 << order) ; ++i)
504 reserved += free_pages_check(page + i);
508 kernel_map_pages(page, 1 << order, 0);
509 local_irq_save(flags);
510 __count_vm_events(PGFREE, 1 << order);
511 free_one_page(page_zone(page), page, order);
512 local_irq_restore(flags);
516 * permit the bootmem allocator to evade page validation on high-order frees
518 void fastcall __init __free_pages_bootmem(struct page *page, unsigned int order)
521 __ClearPageReserved(page);
522 set_page_count(page, 0);
523 set_page_refcounted(page);
529 for (loop = 0; loop < BITS_PER_LONG; loop++) {
530 struct page *p = &page[loop];
532 if (loop + 1 < BITS_PER_LONG)
534 __ClearPageReserved(p);
535 set_page_count(p, 0);
538 set_page_refcounted(page);
539 __free_pages(page, order);
545 * The order of subdivision here is critical for the IO subsystem.
546 * Please do not alter this order without good reasons and regression
547 * testing. Specifically, as large blocks of memory are subdivided,
548 * the order in which smaller blocks are delivered depends on the order
549 * they're subdivided in this function. This is the primary factor
550 * influencing the order in which pages are delivered to the IO
551 * subsystem according to empirical testing, and this is also justified
552 * by considering the behavior of a buddy system containing a single
553 * large block of memory acted on by a series of small allocations.
554 * This behavior is a critical factor in sglist merging's success.
558 static inline void expand(struct zone *zone, struct page *page,
559 int low, int high, struct free_area *area)
561 unsigned long size = 1 << high;
567 VM_BUG_ON(bad_range(zone, &page[size]));
568 list_add(&page[size].lru, &area->free_list);
570 set_page_order(&page[size], high);
575 * This page is about to be returned from the page allocator
577 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
579 if (unlikely(page_mapcount(page) |
580 (page->mapping != NULL) |
581 (page_count(page) != 0) |
597 * For now, we report if PG_reserved was found set, but do not
598 * clear it, and do not allocate the page: as a safety net.
600 if (PageReserved(page))
603 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
604 1 << PG_referenced | 1 << PG_arch_1 |
605 1 << PG_checked | 1 << PG_mappedtodisk);
606 set_page_private(page, 0);
607 set_page_refcounted(page);
608 kernel_map_pages(page, 1 << order, 1);
610 if (gfp_flags & __GFP_ZERO)
611 prep_zero_page(page, order, gfp_flags);
613 if (order && (gfp_flags & __GFP_COMP))
614 prep_compound_page(page, order);
620 * Do the hard work of removing an element from the buddy allocator.
621 * Call me with the zone->lock already held.
623 static struct page *__rmqueue(struct zone *zone, unsigned int order)
625 struct free_area * area;
626 unsigned int current_order;
629 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
630 area = zone->free_area + current_order;
631 if (list_empty(&area->free_list))
634 page = list_entry(area->free_list.next, struct page, lru);
635 list_del(&page->lru);
636 rmv_page_order(page);
638 zone->free_pages -= 1UL << order;
639 expand(zone, page, order, current_order, area);
647 * Obtain a specified number of elements from the buddy allocator, all under
648 * a single hold of the lock, for efficiency. Add them to the supplied list.
649 * Returns the number of new pages which were placed at *list.
651 static int rmqueue_bulk(struct zone *zone, unsigned int order,
652 unsigned long count, struct list_head *list)
656 spin_lock(&zone->lock);
657 for (i = 0; i < count; ++i) {
658 struct page *page = __rmqueue(zone, order);
659 if (unlikely(page == NULL))
661 list_add_tail(&page->lru, list);
663 spin_unlock(&zone->lock);
669 * Called from the slab reaper to drain pagesets on a particular node that
670 * belongs to the currently executing processor.
671 * Note that this function must be called with the thread pinned to
672 * a single processor.
674 void drain_node_pages(int nodeid)
680 for (z = 0; z < MAX_NR_ZONES; z++) {
681 struct zone *zone = NODE_DATA(nodeid)->node_zones + z;
682 struct per_cpu_pageset *pset;
684 if (!populated_zone(zone))
687 pset = zone_pcp(zone, smp_processor_id());
688 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
689 struct per_cpu_pages *pcp;
693 local_irq_save(flags);
694 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
696 local_irq_restore(flags);
703 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
704 static void __drain_pages(unsigned int cpu)
710 for_each_zone(zone) {
711 struct per_cpu_pageset *pset;
713 pset = zone_pcp(zone, cpu);
714 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
715 struct per_cpu_pages *pcp;
718 local_irq_save(flags);
719 free_pages_bulk(zone, pcp->count, &pcp->list, 0);
721 local_irq_restore(flags);
725 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
729 void mark_free_pages(struct zone *zone)
731 unsigned long pfn, max_zone_pfn;
734 struct list_head *curr;
736 if (!zone->spanned_pages)
739 spin_lock_irqsave(&zone->lock, flags);
741 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
742 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
743 if (pfn_valid(pfn)) {
744 struct page *page = pfn_to_page(pfn);
746 if (!PageNosave(page))
747 ClearPageNosaveFree(page);
750 for (order = MAX_ORDER - 1; order >= 0; --order)
751 list_for_each(curr, &zone->free_area[order].free_list) {
754 pfn = page_to_pfn(list_entry(curr, struct page, lru));
755 for (i = 0; i < (1UL << order); i++)
756 SetPageNosaveFree(pfn_to_page(pfn + i));
759 spin_unlock_irqrestore(&zone->lock, flags);
763 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
765 void drain_local_pages(void)
769 local_irq_save(flags);
770 __drain_pages(smp_processor_id());
771 local_irq_restore(flags);
773 #endif /* CONFIG_PM */
776 * Free a 0-order page
778 static void fastcall free_hot_cold_page(struct page *page, int cold)
780 struct zone *zone = page_zone(page);
781 struct per_cpu_pages *pcp;
784 arch_free_page(page, 0);
787 page->mapping = NULL;
788 if (free_pages_check(page))
791 kernel_map_pages(page, 1, 0);
793 pcp = &zone_pcp(zone, get_cpu())->pcp[cold];
794 local_irq_save(flags);
795 __count_vm_event(PGFREE);
796 list_add(&page->lru, &pcp->list);
798 if (pcp->count >= pcp->high) {
799 free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
800 pcp->count -= pcp->batch;
802 local_irq_restore(flags);
806 void fastcall free_hot_page(struct page *page)
808 free_hot_cold_page(page, 0);
811 void fastcall free_cold_page(struct page *page)
813 free_hot_cold_page(page, 1);
817 * split_page takes a non-compound higher-order page, and splits it into
818 * n (1<<order) sub-pages: page[0..n]
819 * Each sub-page must be freed individually.
821 * Note: this is probably too low level an operation for use in drivers.
822 * Please consult with lkml before using this in your driver.
824 void split_page(struct page *page, unsigned int order)
828 VM_BUG_ON(PageCompound(page));
829 VM_BUG_ON(!page_count(page));
830 for (i = 1; i < (1 << order); i++)
831 set_page_refcounted(page + i);
835 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
836 * we cheat by calling it from here, in the order > 0 path. Saves a branch
839 static struct page *buffered_rmqueue(struct zonelist *zonelist,
840 struct zone *zone, int order, gfp_t gfp_flags)
844 int cold = !!(gfp_flags & __GFP_COLD);
849 if (likely(order == 0)) {
850 struct per_cpu_pages *pcp;
852 pcp = &zone_pcp(zone, cpu)->pcp[cold];
853 local_irq_save(flags);
855 pcp->count += rmqueue_bulk(zone, 0,
856 pcp->batch, &pcp->list);
857 if (unlikely(!pcp->count))
860 page = list_entry(pcp->list.next, struct page, lru);
861 list_del(&page->lru);
864 spin_lock_irqsave(&zone->lock, flags);
865 page = __rmqueue(zone, order);
866 spin_unlock(&zone->lock);
871 __count_zone_vm_events(PGALLOC, zone, 1 << order);
872 zone_statistics(zonelist, zone);
873 local_irq_restore(flags);
876 VM_BUG_ON(bad_range(zone, page));
877 if (prep_new_page(page, order, gfp_flags))
882 local_irq_restore(flags);
887 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */
888 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */
889 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */
890 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */
891 #define ALLOC_HARDER 0x10 /* try to alloc harder */
892 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
893 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
896 * Return 1 if free pages are above 'mark'. This takes into account the order
899 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
900 int classzone_idx, int alloc_flags)
902 /* free_pages my go negative - that's OK */
903 long min = mark, free_pages = z->free_pages - (1 << order) + 1;
906 if (alloc_flags & ALLOC_HIGH)
908 if (alloc_flags & ALLOC_HARDER)
911 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
913 for (o = 0; o < order; o++) {
914 /* At the next order, this order's pages become unavailable */
915 free_pages -= z->free_area[o].nr_free << o;
917 /* Require fewer higher order pages to be free */
920 if (free_pages <= min)
927 * get_page_from_freeliest goes through the zonelist trying to allocate
931 get_page_from_freelist(gfp_t gfp_mask, unsigned int order,
932 struct zonelist *zonelist, int alloc_flags)
934 struct zone **z = zonelist->zones;
935 struct page *page = NULL;
936 int classzone_idx = zone_idx(*z);
940 * Go through the zonelist once, looking for a zone with enough free.
941 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
945 if (unlikely((gfp_mask & __GFP_THISNODE) &&
946 zone->zone_pgdat != zonelist->zones[0]->zone_pgdat))
948 if ((alloc_flags & ALLOC_CPUSET) &&
949 !cpuset_zone_allowed(zone, gfp_mask))
952 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
954 if (alloc_flags & ALLOC_WMARK_MIN)
955 mark = zone->pages_min;
956 else if (alloc_flags & ALLOC_WMARK_LOW)
957 mark = zone->pages_low;
959 mark = zone->pages_high;
960 if (!zone_watermark_ok(zone , order, mark,
961 classzone_idx, alloc_flags))
962 if (!zone_reclaim_mode ||
963 !zone_reclaim(zone, gfp_mask, order))
967 page = buffered_rmqueue(zonelist, zone, order, gfp_mask);
971 } while (*(++z) != NULL);
976 * This is the 'heart' of the zoned buddy allocator.
978 struct page * fastcall
979 __alloc_pages(gfp_t gfp_mask, unsigned int order,
980 struct zonelist *zonelist)
982 const gfp_t wait = gfp_mask & __GFP_WAIT;
985 struct reclaim_state reclaim_state;
986 struct task_struct *p = current;
989 int did_some_progress;
991 might_sleep_if(wait);
994 z = zonelist->zones; /* the list of zones suitable for gfp_mask */
996 if (unlikely(*z == NULL)) {
997 /* Should this ever happen?? */
1001 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1002 zonelist, ALLOC_WMARK_LOW|ALLOC_CPUSET);
1007 wakeup_kswapd(*z, order);
1011 * OK, we're below the kswapd watermark and have kicked background
1012 * reclaim. Now things get more complex, so set up alloc_flags according
1013 * to how we want to proceed.
1015 * The caller may dip into page reserves a bit more if the caller
1016 * cannot run direct reclaim, or if the caller has realtime scheduling
1017 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
1018 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
1020 alloc_flags = ALLOC_WMARK_MIN;
1021 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait)
1022 alloc_flags |= ALLOC_HARDER;
1023 if (gfp_mask & __GFP_HIGH)
1024 alloc_flags |= ALLOC_HIGH;
1026 alloc_flags |= ALLOC_CPUSET;
1029 * Go through the zonelist again. Let __GFP_HIGH and allocations
1030 * coming from realtime tasks go deeper into reserves.
1032 * This is the last chance, in general, before the goto nopage.
1033 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
1034 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1036 page = get_page_from_freelist(gfp_mask, order, zonelist, alloc_flags);
1040 /* This allocation should allow future memory freeing. */
1042 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
1043 && !in_interrupt()) {
1044 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
1046 /* go through the zonelist yet again, ignoring mins */
1047 page = get_page_from_freelist(gfp_mask, order,
1048 zonelist, ALLOC_NO_WATERMARKS);
1051 if (gfp_mask & __GFP_NOFAIL) {
1052 blk_congestion_wait(WRITE, HZ/50);
1059 /* Atomic allocations - we can't balance anything */
1066 /* We now go into synchronous reclaim */
1067 cpuset_memory_pressure_bump();
1068 p->flags |= PF_MEMALLOC;
1069 reclaim_state.reclaimed_slab = 0;
1070 p->reclaim_state = &reclaim_state;
1072 did_some_progress = try_to_free_pages(zonelist->zones, gfp_mask);
1074 p->reclaim_state = NULL;
1075 p->flags &= ~PF_MEMALLOC;
1079 if (likely(did_some_progress)) {
1080 page = get_page_from_freelist(gfp_mask, order,
1081 zonelist, alloc_flags);
1084 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
1086 * Go through the zonelist yet one more time, keep
1087 * very high watermark here, this is only to catch
1088 * a parallel oom killing, we must fail if we're still
1089 * under heavy pressure.
1091 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
1092 zonelist, ALLOC_WMARK_HIGH|ALLOC_CPUSET);
1096 out_of_memory(zonelist, gfp_mask, order);
1101 * Don't let big-order allocations loop unless the caller explicitly
1102 * requests that. Wait for some write requests to complete then retry.
1104 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
1105 * <= 3, but that may not be true in other implementations.
1108 if (!(gfp_mask & __GFP_NORETRY)) {
1109 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
1111 if (gfp_mask & __GFP_NOFAIL)
1115 blk_congestion_wait(WRITE, HZ/50);
1120 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
1121 printk(KERN_WARNING "%s: page allocation failure."
1122 " order:%d, mode:0x%x\n",
1123 p->comm, order, gfp_mask);
1131 EXPORT_SYMBOL(__alloc_pages);
1134 * Common helper functions.
1136 fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
1139 page = alloc_pages(gfp_mask, order);
1142 return (unsigned long) page_address(page);
1145 EXPORT_SYMBOL(__get_free_pages);
1147 fastcall unsigned long get_zeroed_page(gfp_t gfp_mask)
1152 * get_zeroed_page() returns a 32-bit address, which cannot represent
1155 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
1157 page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
1159 return (unsigned long) page_address(page);
1163 EXPORT_SYMBOL(get_zeroed_page);
1165 void __pagevec_free(struct pagevec *pvec)
1167 int i = pagevec_count(pvec);
1170 free_hot_cold_page(pvec->pages[i], pvec->cold);
1173 fastcall void __free_pages(struct page *page, unsigned int order)
1175 if (put_page_testzero(page)) {
1177 free_hot_page(page);
1179 __free_pages_ok(page, order);
1183 EXPORT_SYMBOL(__free_pages);
1185 fastcall void free_pages(unsigned long addr, unsigned int order)
1188 VM_BUG_ON(!virt_addr_valid((void *)addr));
1189 __free_pages(virt_to_page((void *)addr), order);
1193 EXPORT_SYMBOL(free_pages);
1196 * Total amount of free (allocatable) RAM:
1198 unsigned int nr_free_pages(void)
1200 unsigned int sum = 0;
1204 sum += zone->free_pages;
1209 EXPORT_SYMBOL(nr_free_pages);
1212 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
1214 unsigned int sum = 0;
1217 for (i = 0; i < MAX_NR_ZONES; i++)
1218 sum += pgdat->node_zones[i].free_pages;
1224 static unsigned int nr_free_zone_pages(int offset)
1226 /* Just pick one node, since fallback list is circular */
1227 pg_data_t *pgdat = NODE_DATA(numa_node_id());
1228 unsigned int sum = 0;
1230 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1231 struct zone **zonep = zonelist->zones;
1234 for (zone = *zonep++; zone; zone = *zonep++) {
1235 unsigned long size = zone->present_pages;
1236 unsigned long high = zone->pages_high;
1245 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1247 unsigned int nr_free_buffer_pages(void)
1249 return nr_free_zone_pages(gfp_zone(GFP_USER));
1253 * Amount of free RAM allocatable within all zones
1255 unsigned int nr_free_pagecache_pages(void)
1257 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER));
1260 static void show_node(struct zone *zone)
1262 printk("Node %ld ", zone_to_nid(zone));
1265 #define show_node(zone) do { } while (0)
1268 void si_meminfo(struct sysinfo *val)
1270 val->totalram = totalram_pages;
1272 val->freeram = nr_free_pages();
1273 val->bufferram = nr_blockdev_pages();
1274 val->totalhigh = totalhigh_pages;
1275 val->freehigh = nr_free_highpages();
1276 val->mem_unit = PAGE_SIZE;
1279 EXPORT_SYMBOL(si_meminfo);
1282 void si_meminfo_node(struct sysinfo *val, int nid)
1284 pg_data_t *pgdat = NODE_DATA(nid);
1286 val->totalram = pgdat->node_present_pages;
1287 val->freeram = nr_free_pages_pgdat(pgdat);
1288 #ifdef CONFIG_HIGHMEM
1289 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1290 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1295 val->mem_unit = PAGE_SIZE;
1299 #define K(x) ((x) << (PAGE_SHIFT-10))
1302 * Show free area list (used inside shift_scroll-lock stuff)
1303 * We also calculate the percentage fragmentation. We do this by counting the
1304 * memory on each free list with the exception of the first item on the list.
1306 void show_free_areas(void)
1309 unsigned long active;
1310 unsigned long inactive;
1314 for_each_zone(zone) {
1315 if (!populated_zone(zone))
1319 printk("%s per-cpu:\n", zone->name);
1321 for_each_online_cpu(cpu) {
1322 struct per_cpu_pageset *pageset;
1324 pageset = zone_pcp(zone, cpu);
1326 printk("CPU %4d: Hot: hi:%5d, btch:%4d usd:%4d "
1327 "Cold: hi:%5d, btch:%4d usd:%4d\n",
1328 cpu, pageset->pcp[0].high,
1329 pageset->pcp[0].batch, pageset->pcp[0].count,
1330 pageset->pcp[1].high, pageset->pcp[1].batch,
1331 pageset->pcp[1].count);
1335 get_zone_counts(&active, &inactive, &free);
1337 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1338 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1341 global_page_state(NR_FILE_DIRTY),
1342 global_page_state(NR_WRITEBACK),
1343 global_page_state(NR_UNSTABLE_NFS),
1345 global_page_state(NR_SLAB_RECLAIMABLE) +
1346 global_page_state(NR_SLAB_UNRECLAIMABLE),
1347 global_page_state(NR_FILE_MAPPED),
1348 global_page_state(NR_PAGETABLE));
1350 for_each_zone(zone) {
1353 if (!populated_zone(zone))
1365 " pages_scanned:%lu"
1366 " all_unreclaimable? %s"
1369 K(zone->free_pages),
1372 K(zone->pages_high),
1374 K(zone->nr_inactive),
1375 K(zone->present_pages),
1376 zone->pages_scanned,
1377 (zone->all_unreclaimable ? "yes" : "no")
1379 printk("lowmem_reserve[]:");
1380 for (i = 0; i < MAX_NR_ZONES; i++)
1381 printk(" %lu", zone->lowmem_reserve[i]);
1385 for_each_zone(zone) {
1386 unsigned long nr[MAX_ORDER], flags, order, total = 0;
1388 if (!populated_zone(zone))
1392 printk("%s: ", zone->name);
1394 spin_lock_irqsave(&zone->lock, flags);
1395 for (order = 0; order < MAX_ORDER; order++) {
1396 nr[order] = zone->free_area[order].nr_free;
1397 total += nr[order] << order;
1399 spin_unlock_irqrestore(&zone->lock, flags);
1400 for (order = 0; order < MAX_ORDER; order++)
1401 printk("%lu*%lukB ", nr[order], K(1UL) << order);
1402 printk("= %lukB\n", K(total));
1405 show_swap_cache_info();
1409 * Builds allocation fallback zone lists.
1411 * Add all populated zones of a node to the zonelist.
1413 static int __meminit build_zonelists_node(pg_data_t *pgdat,
1414 struct zonelist *zonelist, int nr_zones, enum zone_type zone_type)
1418 BUG_ON(zone_type >= MAX_NR_ZONES);
1423 zone = pgdat->node_zones + zone_type;
1424 if (populated_zone(zone)) {
1425 zonelist->zones[nr_zones++] = zone;
1426 check_highest_zone(zone_type);
1429 } while (zone_type);
1434 #define MAX_NODE_LOAD (num_online_nodes())
1435 static int __meminitdata node_load[MAX_NUMNODES];
1437 * find_next_best_node - find the next node that should appear in a given node's fallback list
1438 * @node: node whose fallback list we're appending
1439 * @used_node_mask: nodemask_t of already used nodes
1441 * We use a number of factors to determine which is the next node that should
1442 * appear on a given node's fallback list. The node should not have appeared
1443 * already in @node's fallback list, and it should be the next closest node
1444 * according to the distance array (which contains arbitrary distance values
1445 * from each node to each node in the system), and should also prefer nodes
1446 * with no CPUs, since presumably they'll have very little allocation pressure
1447 * on them otherwise.
1448 * It returns -1 if no node is found.
1450 static int __meminit find_next_best_node(int node, nodemask_t *used_node_mask)
1453 int min_val = INT_MAX;
1456 /* Use the local node if we haven't already */
1457 if (!node_isset(node, *used_node_mask)) {
1458 node_set(node, *used_node_mask);
1462 for_each_online_node(n) {
1465 /* Don't want a node to appear more than once */
1466 if (node_isset(n, *used_node_mask))
1469 /* Use the distance array to find the distance */
1470 val = node_distance(node, n);
1472 /* Penalize nodes under us ("prefer the next node") */
1475 /* Give preference to headless and unused nodes */
1476 tmp = node_to_cpumask(n);
1477 if (!cpus_empty(tmp))
1478 val += PENALTY_FOR_NODE_WITH_CPUS;
1480 /* Slight preference for less loaded node */
1481 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1482 val += node_load[n];
1484 if (val < min_val) {
1491 node_set(best_node, *used_node_mask);
1496 static void __meminit build_zonelists(pg_data_t *pgdat)
1498 int j, node, local_node;
1500 int prev_node, load;
1501 struct zonelist *zonelist;
1502 nodemask_t used_mask;
1504 /* initialize zonelists */
1505 for (i = 0; i < MAX_NR_ZONES; i++) {
1506 zonelist = pgdat->node_zonelists + i;
1507 zonelist->zones[0] = NULL;
1510 /* NUMA-aware ordering of nodes */
1511 local_node = pgdat->node_id;
1512 load = num_online_nodes();
1513 prev_node = local_node;
1514 nodes_clear(used_mask);
1515 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1516 int distance = node_distance(local_node, node);
1519 * If another node is sufficiently far away then it is better
1520 * to reclaim pages in a zone before going off node.
1522 if (distance > RECLAIM_DISTANCE)
1523 zone_reclaim_mode = 1;
1526 * We don't want to pressure a particular node.
1527 * So adding penalty to the first node in same
1528 * distance group to make it round-robin.
1531 if (distance != node_distance(local_node, prev_node))
1532 node_load[node] += load;
1535 for (i = 0; i < MAX_NR_ZONES; i++) {
1536 zonelist = pgdat->node_zonelists + i;
1537 for (j = 0; zonelist->zones[j] != NULL; j++);
1539 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
1540 zonelist->zones[j] = NULL;
1545 #else /* CONFIG_NUMA */
1547 static void __meminit build_zonelists(pg_data_t *pgdat)
1549 int node, local_node;
1552 local_node = pgdat->node_id;
1553 for (i = 0; i < MAX_NR_ZONES; i++) {
1554 struct zonelist *zonelist;
1556 zonelist = pgdat->node_zonelists + i;
1558 j = build_zonelists_node(pgdat, zonelist, 0, i);
1560 * Now we build the zonelist so that it contains the zones
1561 * of all the other nodes.
1562 * We don't want to pressure a particular node, so when
1563 * building the zones for node N, we make sure that the
1564 * zones coming right after the local ones are those from
1565 * node N+1 (modulo N)
1567 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1568 if (!node_online(node))
1570 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
1572 for (node = 0; node < local_node; node++) {
1573 if (!node_online(node))
1575 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i);
1578 zonelist->zones[j] = NULL;
1582 #endif /* CONFIG_NUMA */
1584 /* return values int ....just for stop_machine_run() */
1585 static int __meminit __build_all_zonelists(void *dummy)
1588 for_each_online_node(nid)
1589 build_zonelists(NODE_DATA(nid));
1593 void __meminit build_all_zonelists(void)
1595 if (system_state == SYSTEM_BOOTING) {
1596 __build_all_zonelists(0);
1597 cpuset_init_current_mems_allowed();
1599 /* we have to stop all cpus to guaranntee there is no user
1601 stop_machine_run(__build_all_zonelists, NULL, NR_CPUS);
1602 /* cpuset refresh routine should be here */
1604 vm_total_pages = nr_free_pagecache_pages();
1605 printk("Built %i zonelists. Total pages: %ld\n",
1606 num_online_nodes(), vm_total_pages);
1610 * Helper functions to size the waitqueue hash table.
1611 * Essentially these want to choose hash table sizes sufficiently
1612 * large so that collisions trying to wait on pages are rare.
1613 * But in fact, the number of active page waitqueues on typical
1614 * systems is ridiculously low, less than 200. So this is even
1615 * conservative, even though it seems large.
1617 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1618 * waitqueues, i.e. the size of the waitq table given the number of pages.
1620 #define PAGES_PER_WAITQUEUE 256
1622 #ifndef CONFIG_MEMORY_HOTPLUG
1623 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
1625 unsigned long size = 1;
1627 pages /= PAGES_PER_WAITQUEUE;
1629 while (size < pages)
1633 * Once we have dozens or even hundreds of threads sleeping
1634 * on IO we've got bigger problems than wait queue collision.
1635 * Limit the size of the wait table to a reasonable size.
1637 size = min(size, 4096UL);
1639 return max(size, 4UL);
1643 * A zone's size might be changed by hot-add, so it is not possible to determine
1644 * a suitable size for its wait_table. So we use the maximum size now.
1646 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
1648 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
1649 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
1650 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
1652 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
1653 * or more by the traditional way. (See above). It equals:
1655 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
1656 * ia64(16K page size) : = ( 8G + 4M)byte.
1657 * powerpc (64K page size) : = (32G +16M)byte.
1659 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
1666 * This is an integer logarithm so that shifts can be used later
1667 * to extract the more random high bits from the multiplicative
1668 * hash function before the remainder is taken.
1670 static inline unsigned long wait_table_bits(unsigned long size)
1675 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1678 * Initially all pages are reserved - free ones are freed
1679 * up by free_all_bootmem() once the early boot process is
1680 * done. Non-atomic initialization, single-pass.
1682 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1683 unsigned long start_pfn)
1686 unsigned long end_pfn = start_pfn + size;
1689 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1690 if (!early_pfn_valid(pfn))
1692 page = pfn_to_page(pfn);
1693 set_page_links(page, zone, nid, pfn);
1694 init_page_count(page);
1695 reset_page_mapcount(page);
1696 SetPageReserved(page);
1697 INIT_LIST_HEAD(&page->lru);
1698 #ifdef WANT_PAGE_VIRTUAL
1699 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1700 if (!is_highmem_idx(zone))
1701 set_page_address(page, __va(pfn << PAGE_SHIFT));
1706 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1710 for (order = 0; order < MAX_ORDER ; order++) {
1711 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1712 zone->free_area[order].nr_free = 0;
1716 #define ZONETABLE_INDEX(x, zone_nr) ((x << ZONES_SHIFT) | zone_nr)
1717 void zonetable_add(struct zone *zone, int nid, enum zone_type zid,
1718 unsigned long pfn, unsigned long size)
1720 unsigned long snum = pfn_to_section_nr(pfn);
1721 unsigned long end = pfn_to_section_nr(pfn + size);
1724 zone_table[ZONETABLE_INDEX(nid, zid)] = zone;
1726 for (; snum <= end; snum++)
1727 zone_table[ZONETABLE_INDEX(snum, zid)] = zone;
1730 #ifndef __HAVE_ARCH_MEMMAP_INIT
1731 #define memmap_init(size, nid, zone, start_pfn) \
1732 memmap_init_zone((size), (nid), (zone), (start_pfn))
1735 static int __cpuinit zone_batchsize(struct zone *zone)
1740 * The per-cpu-pages pools are set to around 1000th of the
1741 * size of the zone. But no more than 1/2 of a meg.
1743 * OK, so we don't know how big the cache is. So guess.
1745 batch = zone->present_pages / 1024;
1746 if (batch * PAGE_SIZE > 512 * 1024)
1747 batch = (512 * 1024) / PAGE_SIZE;
1748 batch /= 4; /* We effectively *= 4 below */
1753 * Clamp the batch to a 2^n - 1 value. Having a power
1754 * of 2 value was found to be more likely to have
1755 * suboptimal cache aliasing properties in some cases.
1757 * For example if 2 tasks are alternately allocating
1758 * batches of pages, one task can end up with a lot
1759 * of pages of one half of the possible page colors
1760 * and the other with pages of the other colors.
1762 batch = (1 << (fls(batch + batch/2)-1)) - 1;
1767 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
1769 struct per_cpu_pages *pcp;
1771 memset(p, 0, sizeof(*p));
1773 pcp = &p->pcp[0]; /* hot */
1775 pcp->high = 6 * batch;
1776 pcp->batch = max(1UL, 1 * batch);
1777 INIT_LIST_HEAD(&pcp->list);
1779 pcp = &p->pcp[1]; /* cold*/
1781 pcp->high = 2 * batch;
1782 pcp->batch = max(1UL, batch/2);
1783 INIT_LIST_HEAD(&pcp->list);
1787 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
1788 * to the value high for the pageset p.
1791 static void setup_pagelist_highmark(struct per_cpu_pageset *p,
1794 struct per_cpu_pages *pcp;
1796 pcp = &p->pcp[0]; /* hot list */
1798 pcp->batch = max(1UL, high/4);
1799 if ((high/4) > (PAGE_SHIFT * 8))
1800 pcp->batch = PAGE_SHIFT * 8;
1806 * Boot pageset table. One per cpu which is going to be used for all
1807 * zones and all nodes. The parameters will be set in such a way
1808 * that an item put on a list will immediately be handed over to
1809 * the buddy list. This is safe since pageset manipulation is done
1810 * with interrupts disabled.
1812 * Some NUMA counter updates may also be caught by the boot pagesets.
1814 * The boot_pagesets must be kept even after bootup is complete for
1815 * unused processors and/or zones. They do play a role for bootstrapping
1816 * hotplugged processors.
1818 * zoneinfo_show() and maybe other functions do
1819 * not check if the processor is online before following the pageset pointer.
1820 * Other parts of the kernel may not check if the zone is available.
1822 static struct per_cpu_pageset boot_pageset[NR_CPUS];
1825 * Dynamically allocate memory for the
1826 * per cpu pageset array in struct zone.
1828 static int __cpuinit process_zones(int cpu)
1830 struct zone *zone, *dzone;
1832 for_each_zone(zone) {
1834 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset),
1835 GFP_KERNEL, cpu_to_node(cpu));
1836 if (!zone_pcp(zone, cpu))
1839 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone));
1841 if (percpu_pagelist_fraction)
1842 setup_pagelist_highmark(zone_pcp(zone, cpu),
1843 (zone->present_pages / percpu_pagelist_fraction));
1848 for_each_zone(dzone) {
1851 kfree(zone_pcp(dzone, cpu));
1852 zone_pcp(dzone, cpu) = NULL;
1857 static inline void free_zone_pagesets(int cpu)
1861 for_each_zone(zone) {
1862 struct per_cpu_pageset *pset = zone_pcp(zone, cpu);
1864 /* Free per_cpu_pageset if it is slab allocated */
1865 if (pset != &boot_pageset[cpu])
1867 zone_pcp(zone, cpu) = NULL;
1871 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb,
1872 unsigned long action,
1875 int cpu = (long)hcpu;
1876 int ret = NOTIFY_OK;
1879 case CPU_UP_PREPARE:
1880 if (process_zones(cpu))
1883 case CPU_UP_CANCELED:
1885 free_zone_pagesets(cpu);
1893 static struct notifier_block __cpuinitdata pageset_notifier =
1894 { &pageset_cpuup_callback, NULL, 0 };
1896 void __init setup_per_cpu_pageset(void)
1900 /* Initialize per_cpu_pageset for cpu 0.
1901 * A cpuup callback will do this for every cpu
1902 * as it comes online
1904 err = process_zones(smp_processor_id());
1906 register_cpu_notifier(&pageset_notifier);
1912 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
1915 struct pglist_data *pgdat = zone->zone_pgdat;
1919 * The per-page waitqueue mechanism uses hashed waitqueues
1922 zone->wait_table_hash_nr_entries =
1923 wait_table_hash_nr_entries(zone_size_pages);
1924 zone->wait_table_bits =
1925 wait_table_bits(zone->wait_table_hash_nr_entries);
1926 alloc_size = zone->wait_table_hash_nr_entries
1927 * sizeof(wait_queue_head_t);
1929 if (system_state == SYSTEM_BOOTING) {
1930 zone->wait_table = (wait_queue_head_t *)
1931 alloc_bootmem_node(pgdat, alloc_size);
1934 * This case means that a zone whose size was 0 gets new memory
1935 * via memory hot-add.
1936 * But it may be the case that a new node was hot-added. In
1937 * this case vmalloc() will not be able to use this new node's
1938 * memory - this wait_table must be initialized to use this new
1939 * node itself as well.
1940 * To use this new node's memory, further consideration will be
1943 zone->wait_table = (wait_queue_head_t *)vmalloc(alloc_size);
1945 if (!zone->wait_table)
1948 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
1949 init_waitqueue_head(zone->wait_table + i);
1954 static __meminit void zone_pcp_init(struct zone *zone)
1957 unsigned long batch = zone_batchsize(zone);
1959 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1961 /* Early boot. Slab allocator not functional yet */
1962 zone_pcp(zone, cpu) = &boot_pageset[cpu];
1963 setup_pageset(&boot_pageset[cpu],0);
1965 setup_pageset(zone_pcp(zone,cpu), batch);
1968 if (zone->present_pages)
1969 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
1970 zone->name, zone->present_pages, batch);
1973 __meminit int init_currently_empty_zone(struct zone *zone,
1974 unsigned long zone_start_pfn,
1977 struct pglist_data *pgdat = zone->zone_pgdat;
1979 ret = zone_wait_table_init(zone, size);
1982 pgdat->nr_zones = zone_idx(zone) + 1;
1984 zone->zone_start_pfn = zone_start_pfn;
1986 memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn);
1988 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1993 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
1995 * Basic iterator support. Return the first range of PFNs for a node
1996 * Note: nid == MAX_NUMNODES returns first region regardless of node
1998 static int __init first_active_region_index_in_nid(int nid)
2002 for (i = 0; i < nr_nodemap_entries; i++)
2003 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
2010 * Basic iterator support. Return the next active range of PFNs for a node
2011 * Note: nid == MAX_NUMNODES returns next region regardles of node
2013 static int __init next_active_region_index_in_nid(int index, int nid)
2015 for (index = index + 1; index < nr_nodemap_entries; index++)
2016 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
2022 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
2024 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
2025 * Architectures may implement their own version but if add_active_range()
2026 * was used and there are no special requirements, this is a convenient
2029 int __init early_pfn_to_nid(unsigned long pfn)
2033 for (i = 0; i < nr_nodemap_entries; i++) {
2034 unsigned long start_pfn = early_node_map[i].start_pfn;
2035 unsigned long end_pfn = early_node_map[i].end_pfn;
2037 if (start_pfn <= pfn && pfn < end_pfn)
2038 return early_node_map[i].nid;
2043 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
2045 /* Basic iterator support to walk early_node_map[] */
2046 #define for_each_active_range_index_in_nid(i, nid) \
2047 for (i = first_active_region_index_in_nid(nid); i != -1; \
2048 i = next_active_region_index_in_nid(i, nid))
2051 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
2052 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed
2053 * @max_low_pfn: The highest PFN that till be passed to free_bootmem_node
2055 * If an architecture guarantees that all ranges registered with
2056 * add_active_ranges() contain no holes and may be freed, this
2057 * this function may be used instead of calling free_bootmem() manually.
2059 void __init free_bootmem_with_active_regions(int nid,
2060 unsigned long max_low_pfn)
2064 for_each_active_range_index_in_nid(i, nid) {
2065 unsigned long size_pages = 0;
2066 unsigned long end_pfn = early_node_map[i].end_pfn;
2068 if (early_node_map[i].start_pfn >= max_low_pfn)
2071 if (end_pfn > max_low_pfn)
2072 end_pfn = max_low_pfn;
2074 size_pages = end_pfn - early_node_map[i].start_pfn;
2075 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
2076 PFN_PHYS(early_node_map[i].start_pfn),
2077 size_pages << PAGE_SHIFT);
2082 * sparse_memory_present_with_active_regions - Call memory_present for each active range
2083 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used
2085 * If an architecture guarantees that all ranges registered with
2086 * add_active_ranges() contain no holes and may be freed, this
2087 * this function may be used instead of calling memory_present() manually.
2089 void __init sparse_memory_present_with_active_regions(int nid)
2093 for_each_active_range_index_in_nid(i, nid)
2094 memory_present(early_node_map[i].nid,
2095 early_node_map[i].start_pfn,
2096 early_node_map[i].end_pfn);
2100 * push_node_boundaries - Push node boundaries to at least the requested boundary
2101 * @nid: The nid of the node to push the boundary for
2102 * @start_pfn: The start pfn of the node
2103 * @end_pfn: The end pfn of the node
2105 * In reserve-based hot-add, mem_map is allocated that is unused until hotadd
2106 * time. Specifically, on x86_64, SRAT will report ranges that can potentially
2107 * be hotplugged even though no physical memory exists. This function allows
2108 * an arch to push out the node boundaries so mem_map is allocated that can
2111 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
2112 void __init push_node_boundaries(unsigned int nid,
2113 unsigned long start_pfn, unsigned long end_pfn)
2115 printk(KERN_DEBUG "Entering push_node_boundaries(%u, %lu, %lu)\n",
2116 nid, start_pfn, end_pfn);
2118 /* Initialise the boundary for this node if necessary */
2119 if (node_boundary_end_pfn[nid] == 0)
2120 node_boundary_start_pfn[nid] = -1UL;
2122 /* Update the boundaries */
2123 if (node_boundary_start_pfn[nid] > start_pfn)
2124 node_boundary_start_pfn[nid] = start_pfn;
2125 if (node_boundary_end_pfn[nid] < end_pfn)
2126 node_boundary_end_pfn[nid] = end_pfn;
2129 /* If necessary, push the node boundary out for reserve hotadd */
2130 static void __init account_node_boundary(unsigned int nid,
2131 unsigned long *start_pfn, unsigned long *end_pfn)
2133 printk(KERN_DEBUG "Entering account_node_boundary(%u, %lu, %lu)\n",
2134 nid, *start_pfn, *end_pfn);
2136 /* Return if boundary information has not been provided */
2137 if (node_boundary_end_pfn[nid] == 0)
2140 /* Check the boundaries and update if necessary */
2141 if (node_boundary_start_pfn[nid] < *start_pfn)
2142 *start_pfn = node_boundary_start_pfn[nid];
2143 if (node_boundary_end_pfn[nid] > *end_pfn)
2144 *end_pfn = node_boundary_end_pfn[nid];
2147 void __init push_node_boundaries(unsigned int nid,
2148 unsigned long start_pfn, unsigned long end_pfn) {}
2150 static void __init account_node_boundary(unsigned int nid,
2151 unsigned long *start_pfn, unsigned long *end_pfn) {}
2156 * get_pfn_range_for_nid - Return the start and end page frames for a node
2157 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned
2158 * @start_pfn: Passed by reference. On return, it will have the node start_pfn
2159 * @end_pfn: Passed by reference. On return, it will have the node end_pfn
2161 * It returns the start and end page frame of a node based on information
2162 * provided by an arch calling add_active_range(). If called for a node
2163 * with no available memory, a warning is printed and the start and end
2166 void __init get_pfn_range_for_nid(unsigned int nid,
2167 unsigned long *start_pfn, unsigned long *end_pfn)
2173 for_each_active_range_index_in_nid(i, nid) {
2174 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
2175 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
2178 if (*start_pfn == -1UL) {
2179 printk(KERN_WARNING "Node %u active with no memory\n", nid);
2183 /* Push the node boundaries out if requested */
2184 account_node_boundary(nid, start_pfn, end_pfn);
2188 * Return the number of pages a zone spans in a node, including holes
2189 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
2191 unsigned long __init zone_spanned_pages_in_node(int nid,
2192 unsigned long zone_type,
2193 unsigned long *ignored)
2195 unsigned long node_start_pfn, node_end_pfn;
2196 unsigned long zone_start_pfn, zone_end_pfn;
2198 /* Get the start and end of the node and zone */
2199 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
2200 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
2201 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
2203 /* Check that this node has pages within the zone's required range */
2204 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
2207 /* Move the zone boundaries inside the node if necessary */
2208 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
2209 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
2211 /* Return the spanned pages */
2212 return zone_end_pfn - zone_start_pfn;
2216 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
2217 * then all holes in the requested range will be accounted for
2219 unsigned long __init __absent_pages_in_range(int nid,
2220 unsigned long range_start_pfn,
2221 unsigned long range_end_pfn)
2224 unsigned long prev_end_pfn = 0, hole_pages = 0;
2225 unsigned long start_pfn;
2227 /* Find the end_pfn of the first active range of pfns in the node */
2228 i = first_active_region_index_in_nid(nid);
2232 /* Account for ranges before physical memory on this node */
2233 if (early_node_map[i].start_pfn > range_start_pfn)
2234 hole_pages = early_node_map[i].start_pfn - range_start_pfn;
2236 prev_end_pfn = early_node_map[i].start_pfn;
2238 /* Find all holes for the zone within the node */
2239 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
2241 /* No need to continue if prev_end_pfn is outside the zone */
2242 if (prev_end_pfn >= range_end_pfn)
2245 /* Make sure the end of the zone is not within the hole */
2246 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
2247 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
2249 /* Update the hole size cound and move on */
2250 if (start_pfn > range_start_pfn) {
2251 BUG_ON(prev_end_pfn > start_pfn);
2252 hole_pages += start_pfn - prev_end_pfn;
2254 prev_end_pfn = early_node_map[i].end_pfn;
2257 /* Account for ranges past physical memory on this node */
2258 if (range_end_pfn > prev_end_pfn)
2259 hole_pages = range_end_pfn -
2260 max(range_start_pfn, prev_end_pfn);
2266 * absent_pages_in_range - Return number of page frames in holes within a range
2267 * @start_pfn: The start PFN to start searching for holes
2268 * @end_pfn: The end PFN to stop searching for holes
2270 * It returns the number of pages frames in memory holes within a range
2272 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
2273 unsigned long end_pfn)
2275 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
2278 /* Return the number of page frames in holes in a zone on a node */
2279 unsigned long __init zone_absent_pages_in_node(int nid,
2280 unsigned long zone_type,
2281 unsigned long *ignored)
2283 unsigned long node_start_pfn, node_end_pfn;
2284 unsigned long zone_start_pfn, zone_end_pfn;
2286 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
2287 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
2289 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
2292 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
2295 /* Return the zone index a PFN is in */
2296 int memmap_zone_idx(struct page *lmem_map)
2299 unsigned long phys_addr = virt_to_phys(lmem_map);
2300 unsigned long pfn = phys_addr >> PAGE_SHIFT;
2302 for (i = 0; i < MAX_NR_ZONES; i++)
2303 if (pfn < arch_zone_highest_possible_pfn[i])
2309 static inline unsigned long zone_spanned_pages_in_node(int nid,
2310 unsigned long zone_type,
2311 unsigned long *zones_size)
2313 return zones_size[zone_type];
2316 static inline unsigned long zone_absent_pages_in_node(int nid,
2317 unsigned long zone_type,
2318 unsigned long *zholes_size)
2323 return zholes_size[zone_type];
2326 static inline int memmap_zone_idx(struct page *lmem_map)
2328 return MAX_NR_ZONES;
2332 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
2333 unsigned long *zones_size, unsigned long *zholes_size)
2335 unsigned long realtotalpages, totalpages = 0;
2338 for (i = 0; i < MAX_NR_ZONES; i++)
2339 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
2341 pgdat->node_spanned_pages = totalpages;
2343 realtotalpages = totalpages;
2344 for (i = 0; i < MAX_NR_ZONES; i++)
2346 zone_absent_pages_in_node(pgdat->node_id, i,
2348 pgdat->node_present_pages = realtotalpages;
2349 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
2354 * Set up the zone data structures:
2355 * - mark all pages reserved
2356 * - mark all memory queues empty
2357 * - clear the memory bitmaps
2359 static void __meminit free_area_init_core(struct pglist_data *pgdat,
2360 unsigned long *zones_size, unsigned long *zholes_size)
2363 int nid = pgdat->node_id;
2364 unsigned long zone_start_pfn = pgdat->node_start_pfn;
2367 pgdat_resize_init(pgdat);
2368 pgdat->nr_zones = 0;
2369 init_waitqueue_head(&pgdat->kswapd_wait);
2370 pgdat->kswapd_max_order = 0;
2372 for (j = 0; j < MAX_NR_ZONES; j++) {
2373 struct zone *zone = pgdat->node_zones + j;
2374 unsigned long size, realsize, memmap_pages;
2376 size = zone_spanned_pages_in_node(nid, j, zones_size);
2377 realsize = size - zone_absent_pages_in_node(nid, j,
2381 * Adjust realsize so that it accounts for how much memory
2382 * is used by this zone for memmap. This affects the watermark
2383 * and per-cpu initialisations
2385 memmap_pages = (size * sizeof(struct page)) >> PAGE_SHIFT;
2386 if (realsize >= memmap_pages) {
2387 realsize -= memmap_pages;
2389 " %s zone: %lu pages used for memmap\n",
2390 zone_names[j], memmap_pages);
2393 " %s zone: %lu pages exceeds realsize %lu\n",
2394 zone_names[j], memmap_pages, realsize);
2396 /* Account for reserved DMA pages */
2397 if (j == ZONE_DMA && realsize > dma_reserve) {
2398 realsize -= dma_reserve;
2399 printk(KERN_DEBUG " DMA zone: %lu pages reserved\n",
2403 if (!is_highmem_idx(j))
2404 nr_kernel_pages += realsize;
2405 nr_all_pages += realsize;
2407 zone->spanned_pages = size;
2408 zone->present_pages = realsize;
2410 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
2412 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
2414 zone->name = zone_names[j];
2415 spin_lock_init(&zone->lock);
2416 spin_lock_init(&zone->lru_lock);
2417 zone_seqlock_init(zone);
2418 zone->zone_pgdat = pgdat;
2419 zone->free_pages = 0;
2421 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
2423 zone_pcp_init(zone);
2424 INIT_LIST_HEAD(&zone->active_list);
2425 INIT_LIST_HEAD(&zone->inactive_list);
2426 zone->nr_scan_active = 0;
2427 zone->nr_scan_inactive = 0;
2428 zone->nr_active = 0;
2429 zone->nr_inactive = 0;
2430 zap_zone_vm_stats(zone);
2431 atomic_set(&zone->reclaim_in_progress, 0);
2435 zonetable_add(zone, nid, j, zone_start_pfn, size);
2436 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
2438 zone_start_pfn += size;
2442 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
2444 /* Skip empty nodes */
2445 if (!pgdat->node_spanned_pages)
2448 #ifdef CONFIG_FLAT_NODE_MEM_MAP
2449 /* ia64 gets its own node_mem_map, before this, without bootmem */
2450 if (!pgdat->node_mem_map) {
2451 unsigned long size, start, end;
2455 * The zone's endpoints aren't required to be MAX_ORDER
2456 * aligned but the node_mem_map endpoints must be in order
2457 * for the buddy allocator to function correctly.
2459 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
2460 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
2461 end = ALIGN(end, MAX_ORDER_NR_PAGES);
2462 size = (end - start) * sizeof(struct page);
2463 map = alloc_remap(pgdat->node_id, size);
2465 map = alloc_bootmem_node(pgdat, size);
2466 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
2468 #ifdef CONFIG_FLATMEM
2470 * With no DISCONTIG, the global mem_map is just set as node 0's
2472 if (pgdat == NODE_DATA(0)) {
2473 mem_map = NODE_DATA(0)->node_mem_map;
2474 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
2475 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
2476 mem_map -= pgdat->node_start_pfn;
2477 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
2480 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
2483 void __meminit free_area_init_node(int nid, struct pglist_data *pgdat,
2484 unsigned long *zones_size, unsigned long node_start_pfn,
2485 unsigned long *zholes_size)
2487 pgdat->node_id = nid;
2488 pgdat->node_start_pfn = node_start_pfn;
2489 calculate_node_totalpages(pgdat, zones_size, zholes_size);
2491 alloc_node_mem_map(pgdat);
2493 free_area_init_core(pgdat, zones_size, zholes_size);
2496 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
2498 * add_active_range - Register a range of PFNs backed by physical memory
2499 * @nid: The node ID the range resides on
2500 * @start_pfn: The start PFN of the available physical memory
2501 * @end_pfn: The end PFN of the available physical memory
2503 * These ranges are stored in an early_node_map[] and later used by
2504 * free_area_init_nodes() to calculate zone sizes and holes. If the
2505 * range spans a memory hole, it is up to the architecture to ensure
2506 * the memory is not freed by the bootmem allocator. If possible
2507 * the range being registered will be merged with existing ranges.
2509 void __init add_active_range(unsigned int nid, unsigned long start_pfn,
2510 unsigned long end_pfn)
2514 printk(KERN_DEBUG "Entering add_active_range(%d, %lu, %lu) "
2515 "%d entries of %d used\n",
2516 nid, start_pfn, end_pfn,
2517 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
2519 /* Merge with existing active regions if possible */
2520 for (i = 0; i < nr_nodemap_entries; i++) {
2521 if (early_node_map[i].nid != nid)
2524 /* Skip if an existing region covers this new one */
2525 if (start_pfn >= early_node_map[i].start_pfn &&
2526 end_pfn <= early_node_map[i].end_pfn)
2529 /* Merge forward if suitable */
2530 if (start_pfn <= early_node_map[i].end_pfn &&
2531 end_pfn > early_node_map[i].end_pfn) {
2532 early_node_map[i].end_pfn = end_pfn;
2536 /* Merge backward if suitable */
2537 if (start_pfn < early_node_map[i].end_pfn &&
2538 end_pfn >= early_node_map[i].start_pfn) {
2539 early_node_map[i].start_pfn = start_pfn;
2544 /* Check that early_node_map is large enough */
2545 if (i >= MAX_ACTIVE_REGIONS) {
2546 printk(KERN_CRIT "More than %d memory regions, truncating\n",
2547 MAX_ACTIVE_REGIONS);
2551 early_node_map[i].nid = nid;
2552 early_node_map[i].start_pfn = start_pfn;
2553 early_node_map[i].end_pfn = end_pfn;
2554 nr_nodemap_entries = i + 1;
2558 * shrink_active_range - Shrink an existing registered range of PFNs
2559 * @nid: The node id the range is on that should be shrunk
2560 * @old_end_pfn: The old end PFN of the range
2561 * @new_end_pfn: The new PFN of the range
2563 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
2564 * The map is kept at the end physical page range that has already been
2565 * registered with add_active_range(). This function allows an arch to shrink
2566 * an existing registered range.
2568 void __init shrink_active_range(unsigned int nid, unsigned long old_end_pfn,
2569 unsigned long new_end_pfn)
2573 /* Find the old active region end and shrink */
2574 for_each_active_range_index_in_nid(i, nid)
2575 if (early_node_map[i].end_pfn == old_end_pfn) {
2576 early_node_map[i].end_pfn = new_end_pfn;
2582 * remove_all_active_ranges - Remove all currently registered regions
2583 * During discovery, it may be found that a table like SRAT is invalid
2584 * and an alternative discovery method must be used. This function removes
2585 * all currently registered regions.
2587 void __init remove_all_active_ranges()
2589 memset(early_node_map, 0, sizeof(early_node_map));
2590 nr_nodemap_entries = 0;
2591 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
2592 memset(node_boundary_start_pfn, 0, sizeof(node_boundary_start_pfn));
2593 memset(node_boundary_end_pfn, 0, sizeof(node_boundary_end_pfn));
2594 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */
2597 /* Compare two active node_active_regions */
2598 static int __init cmp_node_active_region(const void *a, const void *b)
2600 struct node_active_region *arange = (struct node_active_region *)a;
2601 struct node_active_region *brange = (struct node_active_region *)b;
2603 /* Done this way to avoid overflows */
2604 if (arange->start_pfn > brange->start_pfn)
2606 if (arange->start_pfn < brange->start_pfn)
2612 /* sort the node_map by start_pfn */
2613 static void __init sort_node_map(void)
2615 sort(early_node_map, (size_t)nr_nodemap_entries,
2616 sizeof(struct node_active_region),
2617 cmp_node_active_region, NULL);
2620 /* Find the lowest pfn for a node. This depends on a sorted early_node_map */
2621 unsigned long __init find_min_pfn_for_node(unsigned long nid)
2625 /* Assuming a sorted map, the first range found has the starting pfn */
2626 for_each_active_range_index_in_nid(i, nid)
2627 return early_node_map[i].start_pfn;
2629 printk(KERN_WARNING "Could not find start_pfn for node %lu\n", nid);
2634 * find_min_pfn_with_active_regions - Find the minimum PFN registered
2636 * It returns the minimum PFN based on information provided via
2637 * add_active_range()
2639 unsigned long __init find_min_pfn_with_active_regions(void)
2641 return find_min_pfn_for_node(MAX_NUMNODES);
2645 * find_max_pfn_with_active_regions - Find the maximum PFN registered
2647 * It returns the maximum PFN based on information provided via
2648 * add_active_range()
2650 unsigned long __init find_max_pfn_with_active_regions(void)
2653 unsigned long max_pfn = 0;
2655 for (i = 0; i < nr_nodemap_entries; i++)
2656 max_pfn = max(max_pfn, early_node_map[i].end_pfn);
2662 * free_area_init_nodes - Initialise all pg_data_t and zone data
2663 * @arch_max_dma_pfn: The maximum PFN usable for ZONE_DMA
2664 * @arch_max_dma32_pfn: The maximum PFN usable for ZONE_DMA32
2665 * @arch_max_low_pfn: The maximum PFN usable for ZONE_NORMAL
2666 * @arch_max_high_pfn: The maximum PFN usable for ZONE_HIGHMEM
2668 * This will call free_area_init_node() for each active node in the system.
2669 * Using the page ranges provided by add_active_range(), the size of each
2670 * zone in each node and their holes is calculated. If the maximum PFN
2671 * between two adjacent zones match, it is assumed that the zone is empty.
2672 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
2673 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
2674 * starts where the previous one ended. For example, ZONE_DMA32 starts
2675 * at arch_max_dma_pfn.
2677 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
2682 /* Record where the zone boundaries are */
2683 memset(arch_zone_lowest_possible_pfn, 0,
2684 sizeof(arch_zone_lowest_possible_pfn));
2685 memset(arch_zone_highest_possible_pfn, 0,
2686 sizeof(arch_zone_highest_possible_pfn));
2687 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
2688 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
2689 for (i = 1; i < MAX_NR_ZONES; i++) {
2690 arch_zone_lowest_possible_pfn[i] =
2691 arch_zone_highest_possible_pfn[i-1];
2692 arch_zone_highest_possible_pfn[i] =
2693 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
2696 /* Regions in the early_node_map can be in any order */
2699 /* Print out the zone ranges */
2700 printk("Zone PFN ranges:\n");
2701 for (i = 0; i < MAX_NR_ZONES; i++)
2702 printk(" %-8s %8lu -> %8lu\n",
2704 arch_zone_lowest_possible_pfn[i],
2705 arch_zone_highest_possible_pfn[i]);
2707 /* Print out the early_node_map[] */
2708 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
2709 for (i = 0; i < nr_nodemap_entries; i++)
2710 printk(" %3d: %8lu -> %8lu\n", early_node_map[i].nid,
2711 early_node_map[i].start_pfn,
2712 early_node_map[i].end_pfn);
2714 /* Initialise every node */
2715 for_each_online_node(nid) {
2716 pg_data_t *pgdat = NODE_DATA(nid);
2717 free_area_init_node(nid, pgdat, NULL,
2718 find_min_pfn_for_node(nid), NULL);
2721 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
2724 * set_dma_reserve - Account the specified number of pages reserved in ZONE_DMA
2725 * @new_dma_reserve - The number of pages to mark reserved
2727 * The per-cpu batchsize and zone watermarks are determined by present_pages.
2728 * In the DMA zone, a significant percentage may be consumed by kernel image
2729 * and other unfreeable allocations which can skew the watermarks badly. This
2730 * function may optionally be used to account for unfreeable pages in
2731 * ZONE_DMA. The effect will be lower watermarks and smaller per-cpu batchsize
2733 void __init set_dma_reserve(unsigned long new_dma_reserve)
2735 dma_reserve = new_dma_reserve;
2738 #ifndef CONFIG_NEED_MULTIPLE_NODES
2739 static bootmem_data_t contig_bootmem_data;
2740 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
2742 EXPORT_SYMBOL(contig_page_data);
2745 void __init free_area_init(unsigned long *zones_size)
2747 free_area_init_node(0, NODE_DATA(0), zones_size,
2748 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
2751 #ifdef CONFIG_HOTPLUG_CPU
2752 static int page_alloc_cpu_notify(struct notifier_block *self,
2753 unsigned long action, void *hcpu)
2755 int cpu = (unsigned long)hcpu;
2757 if (action == CPU_DEAD) {
2758 local_irq_disable();
2760 vm_events_fold_cpu(cpu);
2762 refresh_cpu_vm_stats(cpu);
2766 #endif /* CONFIG_HOTPLUG_CPU */
2768 void __init page_alloc_init(void)
2770 hotcpu_notifier(page_alloc_cpu_notify, 0);
2774 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
2775 * or min_free_kbytes changes.
2777 static void calculate_totalreserve_pages(void)
2779 struct pglist_data *pgdat;
2780 unsigned long reserve_pages = 0;
2781 enum zone_type i, j;
2783 for_each_online_pgdat(pgdat) {
2784 for (i = 0; i < MAX_NR_ZONES; i++) {
2785 struct zone *zone = pgdat->node_zones + i;
2786 unsigned long max = 0;
2788 /* Find valid and maximum lowmem_reserve in the zone */
2789 for (j = i; j < MAX_NR_ZONES; j++) {
2790 if (zone->lowmem_reserve[j] > max)
2791 max = zone->lowmem_reserve[j];
2794 /* we treat pages_high as reserved pages. */
2795 max += zone->pages_high;
2797 if (max > zone->present_pages)
2798 max = zone->present_pages;
2799 reserve_pages += max;
2802 totalreserve_pages = reserve_pages;
2806 * setup_per_zone_lowmem_reserve - called whenever
2807 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
2808 * has a correct pages reserved value, so an adequate number of
2809 * pages are left in the zone after a successful __alloc_pages().
2811 static void setup_per_zone_lowmem_reserve(void)
2813 struct pglist_data *pgdat;
2814 enum zone_type j, idx;
2816 for_each_online_pgdat(pgdat) {
2817 for (j = 0; j < MAX_NR_ZONES; j++) {
2818 struct zone *zone = pgdat->node_zones + j;
2819 unsigned long present_pages = zone->present_pages;
2821 zone->lowmem_reserve[j] = 0;
2825 struct zone *lower_zone;
2829 if (sysctl_lowmem_reserve_ratio[idx] < 1)
2830 sysctl_lowmem_reserve_ratio[idx] = 1;
2832 lower_zone = pgdat->node_zones + idx;
2833 lower_zone->lowmem_reserve[j] = present_pages /
2834 sysctl_lowmem_reserve_ratio[idx];
2835 present_pages += lower_zone->present_pages;
2840 /* update totalreserve_pages */
2841 calculate_totalreserve_pages();
2845 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
2846 * that the pages_{min,low,high} values for each zone are set correctly
2847 * with respect to min_free_kbytes.
2849 void setup_per_zone_pages_min(void)
2851 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
2852 unsigned long lowmem_pages = 0;
2854 unsigned long flags;
2856 /* Calculate total number of !ZONE_HIGHMEM pages */
2857 for_each_zone(zone) {
2858 if (!is_highmem(zone))
2859 lowmem_pages += zone->present_pages;
2862 for_each_zone(zone) {
2865 spin_lock_irqsave(&zone->lru_lock, flags);
2866 tmp = (u64)pages_min * zone->present_pages;
2867 do_div(tmp, lowmem_pages);
2868 if (is_highmem(zone)) {
2870 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
2871 * need highmem pages, so cap pages_min to a small
2874 * The (pages_high-pages_low) and (pages_low-pages_min)
2875 * deltas controls asynch page reclaim, and so should
2876 * not be capped for highmem.
2880 min_pages = zone->present_pages / 1024;
2881 if (min_pages < SWAP_CLUSTER_MAX)
2882 min_pages = SWAP_CLUSTER_MAX;
2883 if (min_pages > 128)
2885 zone->pages_min = min_pages;
2888 * If it's a lowmem zone, reserve a number of pages
2889 * proportionate to the zone's size.
2891 zone->pages_min = tmp;
2894 zone->pages_low = zone->pages_min + (tmp >> 2);
2895 zone->pages_high = zone->pages_min + (tmp >> 1);
2896 spin_unlock_irqrestore(&zone->lru_lock, flags);
2899 /* update totalreserve_pages */
2900 calculate_totalreserve_pages();
2904 * Initialise min_free_kbytes.
2906 * For small machines we want it small (128k min). For large machines
2907 * we want it large (64MB max). But it is not linear, because network
2908 * bandwidth does not increase linearly with machine size. We use
2910 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2911 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
2927 static int __init init_per_zone_pages_min(void)
2929 unsigned long lowmem_kbytes;
2931 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2933 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2934 if (min_free_kbytes < 128)
2935 min_free_kbytes = 128;
2936 if (min_free_kbytes > 65536)
2937 min_free_kbytes = 65536;
2938 setup_per_zone_pages_min();
2939 setup_per_zone_lowmem_reserve();
2942 module_init(init_per_zone_pages_min)
2945 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2946 * that we can call two helper functions whenever min_free_kbytes
2949 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2950 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2952 proc_dointvec(table, write, file, buffer, length, ppos);
2953 setup_per_zone_pages_min();
2958 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
2959 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2964 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2969 zone->min_unmapped_pages = (zone->present_pages *
2970 sysctl_min_unmapped_ratio) / 100;
2974 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
2975 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2980 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2985 zone->min_slab_pages = (zone->present_pages *
2986 sysctl_min_slab_ratio) / 100;
2992 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2993 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2994 * whenever sysctl_lowmem_reserve_ratio changes.
2996 * The reserve ratio obviously has absolutely no relation with the
2997 * pages_min watermarks. The lowmem reserve ratio can only make sense
2998 * if in function of the boot time zone sizes.
3000 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
3001 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
3003 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
3004 setup_per_zone_lowmem_reserve();
3009 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
3010 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
3011 * can have before it gets flushed back to buddy allocator.
3014 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
3015 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
3021 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos);
3022 if (!write || (ret == -EINVAL))
3024 for_each_zone(zone) {
3025 for_each_online_cpu(cpu) {
3027 high = zone->present_pages / percpu_pagelist_fraction;
3028 setup_pagelist_highmark(zone_pcp(zone, cpu), high);
3034 int hashdist = HASHDIST_DEFAULT;
3037 static int __init set_hashdist(char *str)
3041 hashdist = simple_strtoul(str, &str, 0);
3044 __setup("hashdist=", set_hashdist);
3048 * allocate a large system hash table from bootmem
3049 * - it is assumed that the hash table must contain an exact power-of-2
3050 * quantity of entries
3051 * - limit is the number of hash buckets, not the total allocation size
3053 void *__init alloc_large_system_hash(const char *tablename,
3054 unsigned long bucketsize,
3055 unsigned long numentries,
3058 unsigned int *_hash_shift,
3059 unsigned int *_hash_mask,
3060 unsigned long limit)
3062 unsigned long long max = limit;
3063 unsigned long log2qty, size;
3066 /* allow the kernel cmdline to have a say */
3068 /* round applicable memory size up to nearest megabyte */
3069 numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages;
3070 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
3071 numentries >>= 20 - PAGE_SHIFT;
3072 numentries <<= 20 - PAGE_SHIFT;
3074 /* limit to 1 bucket per 2^scale bytes of low memory */
3075 if (scale > PAGE_SHIFT)
3076 numentries >>= (scale - PAGE_SHIFT);
3078 numentries <<= (PAGE_SHIFT - scale);
3080 numentries = roundup_pow_of_two(numentries);
3082 /* limit allocation size to 1/16 total memory by default */
3084 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
3085 do_div(max, bucketsize);
3088 if (numentries > max)
3091 log2qty = long_log2(numentries);
3094 size = bucketsize << log2qty;
3095 if (flags & HASH_EARLY)
3096 table = alloc_bootmem(size);
3098 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
3100 unsigned long order;
3101 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
3103 table = (void*) __get_free_pages(GFP_ATOMIC, order);
3105 } while (!table && size > PAGE_SIZE && --log2qty);
3108 panic("Failed to allocate %s hash table\n", tablename);
3110 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
3113 long_log2(size) - PAGE_SHIFT,
3117 *_hash_shift = log2qty;
3119 *_hash_mask = (1 << log2qty) - 1;
3124 #ifdef CONFIG_OUT_OF_LINE_PFN_TO_PAGE
3125 struct page *pfn_to_page(unsigned long pfn)
3127 return __pfn_to_page(pfn);
3129 unsigned long page_to_pfn(struct page *page)
3131 return __page_to_pfn(page);
3133 EXPORT_SYMBOL(pfn_to_page);
3134 EXPORT_SYMBOL(page_to_pfn);
3135 #endif /* CONFIG_OUT_OF_LINE_PFN_TO_PAGE */