2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/config.h>
18 #include <linux/stddef.h>
20 #include <linux/swap.h>
21 #include <linux/interrupt.h>
22 #include <linux/pagemap.h>
23 #include <linux/bootmem.h>
24 #include <linux/compiler.h>
25 #include <linux/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/vs_base.h>
35 #include <linux/vs_limit.h>
36 #include <linux/ckrm_mem_inline.h>
38 #include <asm/tlbflush.h>
40 DECLARE_BITMAP(node_online_map, MAX_NUMNODES);
41 struct pglist_data *pgdat_list;
42 unsigned long totalram_pages;
43 unsigned long totalhigh_pages;
46 int sysctl_lower_zone_protection = 0;
48 EXPORT_SYMBOL(totalram_pages);
49 EXPORT_SYMBOL(nr_swap_pages);
51 #ifdef CONFIG_CRASH_DUMP_MODULE
52 /* This symbol has to be exported to use 'for_each_pgdat' macro by modules. */
53 EXPORT_SYMBOL(pgdat_list);
57 * Used by page_zone() to look up the address of the struct zone whose
58 * id is encoded in the upper bits of page->flags
60 struct zone *zone_table[1 << (ZONES_SHIFT + NODES_SHIFT)];
61 EXPORT_SYMBOL(zone_table);
63 static char *zone_names[MAX_NR_ZONES] = { "DMA", "Normal", "HighMem" };
64 int min_free_kbytes = 1024;
66 static unsigned long __initdata nr_kernel_pages;
67 static unsigned long __initdata nr_all_pages;
70 * Temporary debugging check for pages not lying within a given zone.
72 static int bad_range(struct zone *zone, struct page *page)
74 if (page_to_pfn(page) >= zone->zone_start_pfn + zone->spanned_pages)
76 if (page_to_pfn(page) < zone->zone_start_pfn)
78 if (zone != page_zone(page))
83 static void bad_page(const char *function, struct page *page)
85 printk(KERN_EMERG "Bad page state at %s (in process '%s', page %p)\n",
86 function, current->comm, page);
87 printk(KERN_EMERG "flags:0x%08lx mapping:%p mapcount:%d count:%d\n",
88 (unsigned long)page->flags, page->mapping,
89 (int)page->mapcount, page_count(page));
90 printk(KERN_EMERG "Backtrace:\n");
92 printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n");
93 page->flags &= ~(1 << PG_private |
102 set_page_count(page, 0);
103 page->mapping = NULL;
107 #if !defined(CONFIG_HUGETLB_PAGE) && !defined(CONFIG_CRASH_DUMP) \
108 && !defined(CONFIG_CRASH_DUMP_MODULE)
109 #define prep_compound_page(page, order) do { } while (0)
110 #define destroy_compound_page(page, order) do { } while (0)
113 * Higher-order pages are called "compound pages". They are structured thusly:
115 * The first PAGE_SIZE page is called the "head page".
117 * The remaining PAGE_SIZE pages are called "tail pages".
119 * All pages have PG_compound set. All pages have their ->private pointing at
120 * the head page (even the head page has this).
122 * The first tail page's ->mapping, if non-zero, holds the address of the
123 * compound page's put_page() function.
125 * The order of the allocation is stored in the first tail page's ->index
126 * This is only for debug at present. This usage means that zero-order pages
127 * may not be compound.
129 static void prep_compound_page(struct page *page, unsigned long order)
132 int nr_pages = 1 << order;
134 page[1].mapping = NULL;
135 page[1].index = order;
136 for (i = 0; i < nr_pages; i++) {
137 struct page *p = page + i;
140 p->private = (unsigned long)page;
144 static void destroy_compound_page(struct page *page, unsigned long order)
147 int nr_pages = 1 << order;
149 if (!PageCompound(page))
152 if (page[1].index != order)
153 bad_page(__FUNCTION__, page);
155 for (i = 0; i < nr_pages; i++) {
156 struct page *p = page + i;
158 if (!PageCompound(p))
159 bad_page(__FUNCTION__, page);
160 if (p->private != (unsigned long)page)
161 bad_page(__FUNCTION__, page);
162 ClearPageCompound(p);
165 #endif /* CONFIG_HUGETLB_PAGE */
168 * Freeing function for a buddy system allocator.
170 * The concept of a buddy system is to maintain direct-mapped table
171 * (containing bit values) for memory blocks of various "orders".
172 * The bottom level table contains the map for the smallest allocatable
173 * units of memory (here, pages), and each level above it describes
174 * pairs of units from the levels below, hence, "buddies".
175 * At a high level, all that happens here is marking the table entry
176 * at the bottom level available, and propagating the changes upward
177 * as necessary, plus some accounting needed to play nicely with other
178 * parts of the VM system.
179 * At each level, we keep one bit for each pair of blocks, which
180 * is set to 1 iff only one of the pair is allocated. So when we
181 * are allocating or freeing one, we can derive the state of the
182 * other. That is, if we allocate a small block, and both were
183 * free, the remainder of the region must be split into blocks.
184 * If a block is freed, and its buddy is also free, then this
185 * triggers coalescing into a block of larger size.
190 static inline void __free_pages_bulk (struct page *page, struct page *base,
191 struct zone *zone, struct free_area *area, unsigned int order)
193 unsigned long page_idx, index, mask;
196 destroy_compound_page(page, order);
197 mask = (~0UL) << order;
198 page_idx = page - base;
199 if (page_idx & ~mask)
201 index = page_idx >> (1 + order);
203 zone->free_pages += 1 << order;
204 while (order < MAX_ORDER-1) {
205 struct page *buddy1, *buddy2;
207 BUG_ON(area >= zone->free_area + MAX_ORDER);
208 if (!__test_and_change_bit(index, area->map))
210 * the buddy page is still allocated.
214 /* Move the buddy up one level. */
215 buddy1 = base + (page_idx ^ (1 << order));
216 buddy2 = base + page_idx;
217 BUG_ON(bad_range(zone, buddy1));
218 BUG_ON(bad_range(zone, buddy2));
219 list_del(&buddy1->lru);
226 list_add(&(base + page_idx)->lru, &area->free_list);
229 static inline void free_pages_check(const char *function, struct page *page)
231 if ( page_mapped(page) ||
232 page->mapping != NULL ||
233 page_count(page) != 0 ||
244 1 << PG_writeback )))
245 bad_page(function, page);
247 ClearPageDirty(page);
251 * Frees a list of pages.
252 * Assumes all pages on list are in same zone, and of same order.
253 * count is the number of pages to free, or 0 for all on the list.
255 * If the zone was previously in an "all pages pinned" state then look to
256 * see if this freeing clears that state.
258 * And clear the zone's pages_scanned counter, to hold off the "all pages are
259 * pinned" detection logic.
262 free_pages_bulk(struct zone *zone, int count,
263 struct list_head *list, unsigned int order)
266 struct free_area *area;
267 struct page *base, *page = NULL;
270 base = zone->zone_mem_map;
271 area = zone->free_area + order;
272 spin_lock_irqsave(&zone->lock, flags);
273 zone->all_unreclaimable = 0;
274 zone->pages_scanned = 0;
275 while (!list_empty(list) && count--) {
276 page = list_entry(list->prev, struct page, lru);
277 /* have to delete it as __free_pages_bulk list manipulates */
278 list_del(&page->lru);
279 __free_pages_bulk(page, base, zone, area, order);
280 ckrm_clear_page_class(page);
283 spin_unlock_irqrestore(&zone->lock, flags);
287 void __free_pages_ok(struct page *page, unsigned int order)
292 arch_free_page(page, order);
294 mod_page_state(pgfree, 1 << order);
295 for (i = 0 ; i < (1 << order) ; ++i)
296 free_pages_check(__FUNCTION__, page + i);
297 list_add(&page->lru, &list);
298 kernel_map_pages(page, 1<<order, 0);
299 free_pages_bulk(page_zone(page), 1, &list, order);
302 #define MARK_USED(index, order, area) \
303 __change_bit((index) >> (1+(order)), (area)->map)
306 * The order of subdivision here is critical for the IO subsystem.
307 * Please do not alter this order without good reasons and regression
308 * testing. Specifically, as large blocks of memory are subdivided,
309 * the order in which smaller blocks are delivered depends on the order
310 * they're subdivided in this function. This is the primary factor
311 * influencing the order in which pages are delivered to the IO
312 * subsystem according to empirical testing, and this is also justified
313 * by considering the behavior of a buddy system containing a single
314 * large block of memory acted on by a series of small allocations.
315 * This behavior is a critical factor in sglist merging's success.
319 static inline struct page *
320 expand(struct zone *zone, struct page *page,
321 unsigned long index, int low, int high, struct free_area *area)
323 unsigned long size = 1 << high;
329 BUG_ON(bad_range(zone, &page[size]));
330 list_add(&page[size].lru, &area->free_list);
331 MARK_USED(index + size, high, area);
336 static inline void set_page_refs(struct page *page, int order)
339 set_page_count(page, 1);
344 * We need to reference all the pages for this order, otherwise if
345 * anyone accesses one of the pages with (get/put) it will be freed.
347 for (i = 0; i < (1 << order); i++)
348 set_page_count(page+i, 1);
349 #endif /* CONFIG_MMU */
353 * This page is about to be returned from the page allocator
355 static void prep_new_page(struct page *page, int order)
357 if (page->mapping || page_mapped(page) ||
368 1 << PG_writeback )))
369 bad_page(__FUNCTION__, page);
371 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
372 1 << PG_referenced | 1 << PG_arch_1 |
373 1 << PG_checked | 1 << PG_mappedtodisk);
375 set_page_refs(page, order);
379 * Do the hard work of removing an element from the buddy allocator.
380 * Call me with the zone->lock already held.
382 static struct page *__rmqueue(struct zone *zone, unsigned int order)
384 struct free_area * area;
385 unsigned int current_order;
389 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
390 area = zone->free_area + current_order;
391 if (list_empty(&area->free_list))
394 page = list_entry(area->free_list.next, struct page, lru);
395 list_del(&page->lru);
396 index = page - zone->zone_mem_map;
397 if (current_order != MAX_ORDER-1)
398 MARK_USED(index, current_order, area);
399 zone->free_pages -= 1UL << order;
400 return expand(zone, page, index, order, current_order, area);
407 * Obtain a specified number of elements from the buddy allocator, all under
408 * a single hold of the lock, for efficiency. Add them to the supplied list.
409 * Returns the number of new pages which were placed at *list.
411 static int rmqueue_bulk(struct zone *zone, unsigned int order,
412 unsigned long count, struct list_head *list)
419 spin_lock_irqsave(&zone->lock, flags);
420 for (i = 0; i < count; ++i) {
421 page = __rmqueue(zone, order);
425 list_add_tail(&page->lru, list);
427 spin_unlock_irqrestore(&zone->lock, flags);
431 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
432 static void __drain_pages(unsigned int cpu)
437 for_each_zone(zone) {
438 struct per_cpu_pageset *pset;
440 pset = &zone->pageset[cpu];
441 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
442 struct per_cpu_pages *pcp;
445 pcp->count -= free_pages_bulk(zone, pcp->count,
450 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
453 int is_head_of_free_region(struct page *page)
455 struct zone *zone = page_zone(page);
458 struct list_head *curr;
461 * Should not matter as we need quiescent system for
462 * suspend anyway, but...
464 spin_lock_irqsave(&zone->lock, flags);
465 for (order = MAX_ORDER - 1; order >= 0; --order)
466 list_for_each(curr, &zone->free_area[order].free_list)
467 if (page == list_entry(curr, struct page, lru)) {
468 spin_unlock_irqrestore(&zone->lock, flags);
471 spin_unlock_irqrestore(&zone->lock, flags);
476 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
478 void drain_local_pages(void)
482 local_irq_save(flags);
483 __drain_pages(smp_processor_id());
484 local_irq_restore(flags);
486 #endif /* CONFIG_PM */
488 static void zone_statistics(struct zonelist *zonelist, struct zone *z)
493 pg_data_t *pg = z->zone_pgdat;
494 pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
495 struct per_cpu_pageset *p;
497 local_irq_save(flags);
498 cpu = smp_processor_id();
499 p = &z->pageset[cpu];
501 z->pageset[cpu].numa_hit++;
504 zonelist->zones[0]->pageset[cpu].numa_foreign++;
506 if (pg == NODE_DATA(numa_node_id()))
510 local_irq_restore(flags);
515 * Free a 0-order page
517 static void FASTCALL(free_hot_cold_page(struct page *page, int cold));
518 static void fastcall free_hot_cold_page(struct page *page, int cold)
520 struct zone *zone = page_zone(page);
521 struct per_cpu_pages *pcp;
524 arch_free_page(page, 0);
526 kernel_map_pages(page, 1, 0);
527 inc_page_state(pgfree);
528 free_pages_check(__FUNCTION__, page);
529 pcp = &zone->pageset[get_cpu()].pcp[cold];
530 local_irq_save(flags);
531 if (pcp->count >= pcp->high)
532 pcp->count -= free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
533 list_add(&page->lru, &pcp->list);
535 local_irq_restore(flags);
539 void fastcall free_hot_page(struct page *page)
541 free_hot_cold_page(page, 0);
544 void fastcall free_cold_page(struct page *page)
546 free_hot_cold_page(page, 1);
550 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
551 * we cheat by calling it from here, in the order > 0 path. Saves a branch
556 buffered_rmqueue(struct zone *zone, int order, int gfp_flags)
559 struct page *page = NULL;
560 int cold = !!(gfp_flags & __GFP_COLD);
563 struct per_cpu_pages *pcp;
565 pcp = &zone->pageset[get_cpu()].pcp[cold];
566 local_irq_save(flags);
567 if (pcp->count <= pcp->low)
568 pcp->count += rmqueue_bulk(zone, 0,
569 pcp->batch, &pcp->list);
571 page = list_entry(pcp->list.next, struct page, lru);
572 list_del(&page->lru);
575 local_irq_restore(flags);
580 spin_lock_irqsave(&zone->lock, flags);
581 page = __rmqueue(zone, order);
582 spin_unlock_irqrestore(&zone->lock, flags);
586 BUG_ON(bad_range(zone, page));
587 mod_page_state_zone(zone, pgalloc, 1 << order);
588 prep_new_page(page, order);
589 if (order && (gfp_flags & __GFP_COMP))
590 prep_compound_page(page, order);
596 * This is the 'heart' of the zoned buddy allocator.
598 * Herein lies the mysterious "incremental min". That's the
600 * local_low = z->pages_low;
603 * thing. The intent here is to provide additional protection to low zones for
604 * allocation requests which _could_ use higher zones. So a GFP_HIGHMEM
605 * request is not allowed to dip as deeply into the normal zone as a GFP_KERNEL
606 * request. This preserves additional space in those lower zones for requests
607 * which really do need memory from those zones. It means that on a decent
608 * sized machine, GFP_HIGHMEM and GFP_KERNEL requests basically leave the DMA
611 struct page * fastcall
612 __alloc_pages(unsigned int gfp_mask, unsigned int order,
613 struct zonelist *zonelist)
615 const int wait = gfp_mask & __GFP_WAIT;
619 struct reclaim_state reclaim_state;
620 struct task_struct *p = current;
625 might_sleep_if(wait);
627 if (!ckrm_class_limit_ok((GET_MEM_CLASS(current)))) {
631 zones = zonelist->zones; /* the list of zones suitable for gfp_mask */
632 if (zones[0] == NULL) /* no zones in the zonelist */
635 alloc_type = zone_idx(zones[0]);
637 /* Go through the zonelist once, looking for a zone with enough free */
638 for (i = 0; zones[i] != NULL; i++) {
639 struct zone *z = zones[i];
641 min = (1<<order) + z->protection[alloc_type];
644 * We let real-time tasks dip their real-time paws a little
645 * deeper into reserves.
648 min -= z->pages_low >> 1;
650 if (z->free_pages >= min ||
651 (!wait && z->free_pages >= z->pages_high)) {
652 page = buffered_rmqueue(z, order, gfp_mask);
654 zone_statistics(zonelist, z);
660 /* we're somewhat low on memory, failed to find what we needed */
661 for (i = 0; zones[i] != NULL; i++)
662 wakeup_kswapd(zones[i]);
664 /* Go through the zonelist again, taking __GFP_HIGH into account */
665 for (i = 0; zones[i] != NULL; i++) {
666 struct zone *z = zones[i];
668 min = (1<<order) + z->protection[alloc_type];
670 if (gfp_mask & __GFP_HIGH)
671 min -= z->pages_low >> 2;
673 min -= z->pages_low >> 1;
675 if (z->free_pages >= min ||
676 (!wait && z->free_pages >= z->pages_high)) {
677 page = buffered_rmqueue(z, order, gfp_mask);
679 zone_statistics(zonelist, z);
685 /* here we're in the low on memory slow path */
688 if ((p->flags & (PF_MEMALLOC | PF_MEMDIE)) && !in_interrupt()) {
689 /* go through the zonelist yet again, ignoring mins */
690 for (i = 0; zones[i] != NULL; i++) {
691 struct zone *z = zones[i];
693 page = buffered_rmqueue(z, order, gfp_mask);
695 zone_statistics(zonelist, z);
702 /* Atomic allocations - we can't balance anything */
706 p->flags |= PF_MEMALLOC;
707 reclaim_state.reclaimed_slab = 0;
708 p->reclaim_state = &reclaim_state;
710 try_to_free_pages(zones, gfp_mask, order);
712 p->reclaim_state = NULL;
713 p->flags &= ~PF_MEMALLOC;
715 /* go through the zonelist yet one more time */
716 for (i = 0; zones[i] != NULL; i++) {
717 struct zone *z = zones[i];
719 min = (1UL << order) + z->protection[alloc_type];
721 if (z->free_pages >= min ||
722 (!wait && z->free_pages >= z->pages_high)) {
723 page = buffered_rmqueue(z, order, gfp_mask);
725 zone_statistics(zonelist, z);
732 * Don't let big-order allocations loop unless the caller explicitly
733 * requests that. Wait for some write requests to complete then retry.
735 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL, but that
736 * may not be true in other implementations.
739 if (!(gfp_mask & __GFP_NORETRY)) {
740 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
742 if (gfp_mask & __GFP_NOFAIL)
746 blk_congestion_wait(WRITE, HZ/50);
751 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
752 printk(KERN_WARNING "%s: page allocation failure."
753 " order:%d, mode:0x%x\n",
754 p->comm, order, gfp_mask);
759 kernel_map_pages(page, 1 << order, 1);
760 ckrm_set_pages_class(page, 1 << order, GET_MEM_CLASS(current));
764 EXPORT_SYMBOL(__alloc_pages);
767 * Common helper functions.
769 fastcall unsigned long __get_free_pages(unsigned int gfp_mask, unsigned int order)
772 page = alloc_pages(gfp_mask, order);
775 return (unsigned long) page_address(page);
778 EXPORT_SYMBOL(__get_free_pages);
780 fastcall unsigned long get_zeroed_page(unsigned int gfp_mask)
785 * get_zeroed_page() returns a 32-bit address, which cannot represent
788 BUG_ON(gfp_mask & __GFP_HIGHMEM);
790 page = alloc_pages(gfp_mask, 0);
792 void *address = page_address(page);
794 return (unsigned long) address;
799 EXPORT_SYMBOL(get_zeroed_page);
801 void __pagevec_free(struct pagevec *pvec)
803 int i = pagevec_count(pvec);
806 free_hot_cold_page(pvec->pages[i], pvec->cold);
809 fastcall void __free_pages(struct page *page, unsigned int order)
811 if (!PageReserved(page) && put_page_testzero(page)) {
815 __free_pages_ok(page, order);
819 EXPORT_SYMBOL(__free_pages);
821 fastcall void free_pages(unsigned long addr, unsigned int order)
824 BUG_ON(!virt_addr_valid(addr));
825 __free_pages(virt_to_page(addr), order);
829 EXPORT_SYMBOL(free_pages);
832 * Total amount of free (allocatable) RAM:
834 unsigned int nr_free_pages(void)
836 unsigned int sum = 0;
840 sum += zone->free_pages;
845 EXPORT_SYMBOL(nr_free_pages);
848 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
850 unsigned int i, sum = 0;
852 for (i = 0; i < MAX_NR_ZONES; i++)
853 sum += pgdat->node_zones[i].free_pages;
859 static unsigned int nr_free_zone_pages(int offset)
862 unsigned int sum = 0;
864 for_each_pgdat(pgdat) {
865 struct zonelist *zonelist = pgdat->node_zonelists + offset;
866 struct zone **zonep = zonelist->zones;
869 for (zone = *zonep++; zone; zone = *zonep++) {
870 unsigned long size = zone->present_pages;
871 unsigned long high = zone->pages_high;
881 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
883 unsigned int nr_free_buffer_pages(void)
885 return nr_free_zone_pages(GFP_USER & GFP_ZONEMASK);
889 * Amount of free RAM allocatable within all zones
891 unsigned int nr_free_pagecache_pages(void)
893 return nr_free_zone_pages(GFP_HIGHUSER & GFP_ZONEMASK);
896 #ifdef CONFIG_HIGHMEM
897 unsigned int nr_free_highpages (void)
900 unsigned int pages = 0;
902 for_each_pgdat(pgdat)
903 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
910 static void show_node(struct zone *zone)
912 printk("Node %d ", zone->zone_pgdat->node_id);
915 #define show_node(zone) do { } while (0)
919 * Accumulate the page_state information across all CPUs.
920 * The result is unavoidably approximate - it can change
921 * during and after execution of this function.
923 DEFINE_PER_CPU(struct page_state, page_states) = {0};
924 EXPORT_PER_CPU_SYMBOL(page_states);
926 atomic_t nr_pagecache = ATOMIC_INIT(0);
927 EXPORT_SYMBOL(nr_pagecache);
929 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
932 void __get_page_state(struct page_state *ret, int nr)
936 memset(ret, 0, sizeof(*ret));
937 while (cpu < NR_CPUS) {
938 unsigned long *in, *out, off;
940 if (!cpu_possible(cpu)) {
945 in = (unsigned long *)&per_cpu(page_states, cpu);
947 if (cpu < NR_CPUS && cpu_possible(cpu))
948 prefetch(&per_cpu(page_states, cpu));
949 out = (unsigned long *)ret;
950 for (off = 0; off < nr; off++)
955 void get_page_state(struct page_state *ret)
959 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
960 nr /= sizeof(unsigned long);
962 __get_page_state(ret, nr + 1);
965 void get_full_page_state(struct page_state *ret)
967 __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long));
970 unsigned long __read_page_state(unsigned offset)
972 unsigned long ret = 0;
975 for (cpu = 0; cpu < NR_CPUS; cpu++) {
978 if (!cpu_possible(cpu))
981 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
982 ret += *((unsigned long *)in);
987 void get_zone_counts(unsigned long *active,
988 unsigned long *inactive, unsigned long *free)
995 for_each_zone(zone) {
996 *active += zone->nr_active;
997 *inactive += zone->nr_inactive;
998 *free += zone->free_pages;
1002 void si_meminfo(struct sysinfo *val)
1004 val->totalram = totalram_pages;
1006 val->freeram = nr_free_pages();
1007 val->bufferram = nr_blockdev_pages();
1008 #ifdef CONFIG_HIGHMEM
1009 val->totalhigh = totalhigh_pages;
1010 val->freehigh = nr_free_highpages();
1015 val->mem_unit = PAGE_SIZE;
1016 if (vx_flags(VXF_VIRT_MEM, 0))
1017 vx_vsi_meminfo(val);
1020 EXPORT_SYMBOL(si_meminfo);
1023 void si_meminfo_node(struct sysinfo *val, int nid)
1025 pg_data_t *pgdat = NODE_DATA(nid);
1027 val->totalram = pgdat->node_present_pages;
1028 val->freeram = nr_free_pages_pgdat(pgdat);
1029 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1030 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1031 val->mem_unit = PAGE_SIZE;
1035 #define K(x) ((x) << (PAGE_SHIFT-10))
1038 * Show free area list (used inside shift_scroll-lock stuff)
1039 * We also calculate the percentage fragmentation. We do this by counting the
1040 * memory on each free list with the exception of the first item on the list.
1042 void show_free_areas(void)
1044 struct page_state ps;
1045 int cpu, temperature;
1046 unsigned long active;
1047 unsigned long inactive;
1051 for_each_zone(zone) {
1053 printk("%s per-cpu:", zone->name);
1055 if (!zone->present_pages) {
1061 for (cpu = 0; cpu < NR_CPUS; ++cpu) {
1062 struct per_cpu_pageset *pageset;
1064 if (!cpu_possible(cpu))
1067 pageset = zone->pageset + cpu;
1069 for (temperature = 0; temperature < 2; temperature++)
1070 printk("cpu %d %s: low %d, high %d, batch %d\n",
1072 temperature ? "cold" : "hot",
1073 pageset->pcp[temperature].low,
1074 pageset->pcp[temperature].high,
1075 pageset->pcp[temperature].batch);
1079 get_page_state(&ps);
1080 get_zone_counts(&active, &inactive, &free);
1082 printk("\nFree pages: %11ukB (%ukB HighMem)\n",
1084 K(nr_free_highpages()));
1086 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1087 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1096 ps.nr_page_table_pages);
1098 for_each_zone(zone) {
1112 K(zone->free_pages),
1115 K(zone->pages_high),
1117 K(zone->nr_inactive),
1118 K(zone->present_pages)
1120 printk("protections[]:");
1121 for (i = 0; i < MAX_NR_ZONES; i++)
1122 printk(" %lu", zone->protection[i]);
1126 for_each_zone(zone) {
1127 struct list_head *elem;
1128 unsigned long nr, flags, order, total = 0;
1131 printk("%s: ", zone->name);
1132 if (!zone->present_pages) {
1137 spin_lock_irqsave(&zone->lock, flags);
1138 for (order = 0; order < MAX_ORDER; order++) {
1140 list_for_each(elem, &zone->free_area[order].free_list)
1142 total += nr << order;
1143 printk("%lu*%lukB ", nr, K(1UL) << order);
1145 spin_unlock_irqrestore(&zone->lock, flags);
1146 printk("= %lukB\n", K(total));
1149 show_swap_cache_info();
1153 * Builds allocation fallback zone lists.
1155 static int __init build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, int j, int k)
1162 zone = pgdat->node_zones + ZONE_HIGHMEM;
1163 if (zone->present_pages) {
1164 #ifndef CONFIG_HIGHMEM
1167 zonelist->zones[j++] = zone;
1170 zone = pgdat->node_zones + ZONE_NORMAL;
1171 if (zone->present_pages)
1172 zonelist->zones[j++] = zone;
1174 zone = pgdat->node_zones + ZONE_DMA;
1175 if (zone->present_pages)
1176 zonelist->zones[j++] = zone;
1183 #define MAX_NODE_LOAD (numnodes)
1184 static int __initdata node_load[MAX_NUMNODES];
1186 * find_next_best_node - find the next node that should appear in a given
1187 * node's fallback list
1188 * @node: node whose fallback list we're appending
1189 * @used_node_mask: pointer to the bitmap of already used nodes
1191 * We use a number of factors to determine which is the next node that should
1192 * appear on a given node's fallback list. The node should not have appeared
1193 * already in @node's fallback list, and it should be the next closest node
1194 * according to the distance array (which contains arbitrary distance values
1195 * from each node to each node in the system), and should also prefer nodes
1196 * with no CPUs, since presumably they'll have very little allocation pressure
1197 * on them otherwise.
1198 * It returns -1 if no node is found.
1200 static int __init find_next_best_node(int node, void *used_node_mask)
1203 int min_val = INT_MAX;
1206 for (i = 0; i < numnodes; i++) {
1209 /* Start from local node */
1210 n = (node+i)%numnodes;
1212 /* Don't want a node to appear more than once */
1213 if (test_bit(n, used_node_mask))
1216 /* Use the distance array to find the distance */
1217 val = node_distance(node, n);
1219 /* Give preference to headless and unused nodes */
1220 tmp = node_to_cpumask(n);
1221 if (!cpus_empty(tmp))
1222 val += PENALTY_FOR_NODE_WITH_CPUS;
1224 /* Slight preference for less loaded node */
1225 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1226 val += node_load[n];
1228 if (val < min_val) {
1235 set_bit(best_node, used_node_mask);
1240 static void __init build_zonelists(pg_data_t *pgdat)
1242 int i, j, k, node, local_node;
1243 int prev_node, load;
1244 struct zonelist *zonelist;
1245 DECLARE_BITMAP(used_mask, MAX_NUMNODES);
1247 /* initialize zonelists */
1248 for (i = 0; i < GFP_ZONETYPES; i++) {
1249 zonelist = pgdat->node_zonelists + i;
1250 memset(zonelist, 0, sizeof(*zonelist));
1251 zonelist->zones[0] = NULL;
1254 /* NUMA-aware ordering of nodes */
1255 local_node = pgdat->node_id;
1257 prev_node = local_node;
1258 bitmap_zero(used_mask, MAX_NUMNODES);
1259 while ((node = find_next_best_node(local_node, used_mask)) >= 0) {
1261 * We don't want to pressure a particular node.
1262 * So adding penalty to the first node in same
1263 * distance group to make it round-robin.
1265 if (node_distance(local_node, node) !=
1266 node_distance(local_node, prev_node))
1267 node_load[node] += load;
1270 for (i = 0; i < GFP_ZONETYPES; i++) {
1271 zonelist = pgdat->node_zonelists + i;
1272 for (j = 0; zonelist->zones[j] != NULL; j++);
1275 if (i & __GFP_HIGHMEM)
1280 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1281 zonelist->zones[j] = NULL;
1286 #else /* CONFIG_NUMA */
1288 static void __init build_zonelists(pg_data_t *pgdat)
1290 int i, j, k, node, local_node;
1292 local_node = pgdat->node_id;
1293 for (i = 0; i < GFP_ZONETYPES; i++) {
1294 struct zonelist *zonelist;
1296 zonelist = pgdat->node_zonelists + i;
1297 memset(zonelist, 0, sizeof(*zonelist));
1301 if (i & __GFP_HIGHMEM)
1306 j = build_zonelists_node(pgdat, zonelist, j, k);
1308 * Now we build the zonelist so that it contains the zones
1309 * of all the other nodes.
1310 * We don't want to pressure a particular node, so when
1311 * building the zones for node N, we make sure that the
1312 * zones coming right after the local ones are those from
1313 * node N+1 (modulo N)
1315 for (node = local_node + 1; node < numnodes; node++)
1316 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1317 for (node = 0; node < local_node; node++)
1318 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1320 zonelist->zones[j] = NULL;
1324 #endif /* CONFIG_NUMA */
1326 void __init build_all_zonelists(void)
1330 for(i = 0 ; i < numnodes ; i++)
1331 build_zonelists(NODE_DATA(i));
1332 printk("Built %i zonelists\n", numnodes);
1336 * Helper functions to size the waitqueue hash table.
1337 * Essentially these want to choose hash table sizes sufficiently
1338 * large so that collisions trying to wait on pages are rare.
1339 * But in fact, the number of active page waitqueues on typical
1340 * systems is ridiculously low, less than 200. So this is even
1341 * conservative, even though it seems large.
1343 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1344 * waitqueues, i.e. the size of the waitq table given the number of pages.
1346 #define PAGES_PER_WAITQUEUE 256
1348 static inline unsigned long wait_table_size(unsigned long pages)
1350 unsigned long size = 1;
1352 pages /= PAGES_PER_WAITQUEUE;
1354 while (size < pages)
1358 * Once we have dozens or even hundreds of threads sleeping
1359 * on IO we've got bigger problems than wait queue collision.
1360 * Limit the size of the wait table to a reasonable size.
1362 size = min(size, 4096UL);
1364 return max(size, 4UL);
1368 * This is an integer logarithm so that shifts can be used later
1369 * to extract the more random high bits from the multiplicative
1370 * hash function before the remainder is taken.
1372 static inline unsigned long wait_table_bits(unsigned long size)
1377 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1379 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1380 unsigned long *zones_size, unsigned long *zholes_size)
1382 unsigned long realtotalpages, totalpages = 0;
1385 for (i = 0; i < MAX_NR_ZONES; i++)
1386 totalpages += zones_size[i];
1387 pgdat->node_spanned_pages = totalpages;
1389 realtotalpages = totalpages;
1391 for (i = 0; i < MAX_NR_ZONES; i++)
1392 realtotalpages -= zholes_size[i];
1393 pgdat->node_present_pages = realtotalpages;
1394 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1399 * Initially all pages are reserved - free ones are freed
1400 * up by free_all_bootmem() once the early boot process is
1401 * done. Non-atomic initialization, single-pass.
1403 void __init memmap_init_zone(struct page *start, unsigned long size, int nid,
1404 unsigned long zone, unsigned long start_pfn)
1408 for (page = start; page < (start + size); page++) {
1409 set_page_zone(page, NODEZONE(nid, zone));
1410 set_page_count(page, 0);
1411 SetPageReserved(page);
1412 INIT_LIST_HEAD(&page->lru);
1413 #ifdef WANT_PAGE_VIRTUAL
1414 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1415 if (!is_highmem_idx(zone))
1416 set_page_address(page, __va(start_pfn << PAGE_SHIFT));
1422 #ifndef __HAVE_ARCH_MEMMAP_INIT
1423 #define memmap_init(start, size, nid, zone, start_pfn) \
1424 memmap_init_zone((start), (size), (nid), (zone), (start_pfn))
1428 * Set up the zone data structures:
1429 * - mark all pages reserved
1430 * - mark all memory queues empty
1431 * - clear the memory bitmaps
1433 static void __init free_area_init_core(struct pglist_data *pgdat,
1434 unsigned long *zones_size, unsigned long *zholes_size)
1437 const unsigned long zone_required_alignment = 1UL << (MAX_ORDER-1);
1438 int cpu, nid = pgdat->node_id;
1439 struct page *lmem_map = pgdat->node_mem_map;
1440 unsigned long zone_start_pfn = pgdat->node_start_pfn;
1442 pgdat->nr_zones = 0;
1443 init_waitqueue_head(&pgdat->kswapd_wait);
1445 for (j = 0; j < MAX_NR_ZONES; j++) {
1446 struct zone *zone = pgdat->node_zones + j;
1447 unsigned long size, realsize;
1448 unsigned long batch;
1450 zone_table[NODEZONE(nid, j)] = zone;
1451 realsize = size = zones_size[j];
1453 realsize -= zholes_size[j];
1455 if (j == ZONE_DMA || j == ZONE_NORMAL)
1456 nr_kernel_pages += realsize;
1457 nr_all_pages += realsize;
1459 zone->spanned_pages = size;
1460 zone->present_pages = realsize;
1461 zone->name = zone_names[j];
1462 spin_lock_init(&zone->lock);
1463 spin_lock_init(&zone->lru_lock);
1464 zone->zone_pgdat = pgdat;
1465 zone->free_pages = 0;
1467 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
1470 * The per-cpu-pages pools are set to around 1000th of the
1471 * size of the zone. But no more than 1/4 of a meg - there's
1472 * no point in going beyond the size of L2 cache.
1474 * OK, so we don't know how big the cache is. So guess.
1476 batch = zone->present_pages / 1024;
1477 if (batch * PAGE_SIZE > 256 * 1024)
1478 batch = (256 * 1024) / PAGE_SIZE;
1479 batch /= 4; /* We effectively *= 4 below */
1483 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1484 struct per_cpu_pages *pcp;
1486 pcp = &zone->pageset[cpu].pcp[0]; /* hot */
1488 pcp->low = 2 * batch;
1489 pcp->high = 6 * batch;
1490 pcp->batch = 1 * batch;
1491 INIT_LIST_HEAD(&pcp->list);
1493 pcp = &zone->pageset[cpu].pcp[1]; /* cold */
1496 pcp->high = 2 * batch;
1497 pcp->batch = 1 * batch;
1498 INIT_LIST_HEAD(&pcp->list);
1500 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
1501 zone_names[j], realsize, batch);
1502 INIT_LIST_HEAD(&zone->active_list);
1503 INIT_LIST_HEAD(&zone->inactive_list);
1504 zone->nr_scan_active = 0;
1505 zone->nr_scan_inactive = 0;
1506 zone->nr_active = 0;
1507 zone->nr_inactive = 0;
1512 * The per-page waitqueue mechanism uses hashed waitqueues
1515 zone->wait_table_size = wait_table_size(size);
1516 zone->wait_table_bits =
1517 wait_table_bits(zone->wait_table_size);
1518 zone->wait_table = (wait_queue_head_t *)
1519 alloc_bootmem_node(pgdat, zone->wait_table_size
1520 * sizeof(wait_queue_head_t));
1522 for(i = 0; i < zone->wait_table_size; ++i)
1523 init_waitqueue_head(zone->wait_table + i);
1525 pgdat->nr_zones = j+1;
1527 zone->zone_mem_map = lmem_map;
1528 zone->zone_start_pfn = zone_start_pfn;
1530 if ((zone_start_pfn) & (zone_required_alignment-1))
1531 printk("BUG: wrong zone alignment, it will crash\n");
1533 memmap_init(lmem_map, size, nid, j, zone_start_pfn);
1535 zone_start_pfn += size;
1538 for (i = 0; ; i++) {
1539 unsigned long bitmap_size;
1541 INIT_LIST_HEAD(&zone->free_area[i].free_list);
1542 if (i == MAX_ORDER-1) {
1543 zone->free_area[i].map = NULL;
1548 * Page buddy system uses "index >> (i+1)",
1549 * where "index" is at most "size-1".
1551 * The extra "+3" is to round down to byte
1552 * size (8 bits per byte assumption). Thus
1553 * we get "(size-1) >> (i+4)" as the last byte
1556 * The "+1" is because we want to round the
1557 * byte allocation up rather than down. So
1558 * we should have had a "+7" before we shifted
1559 * down by three. Also, we have to add one as
1560 * we actually _use_ the last bit (it's [0,n]
1561 * inclusive, not [0,n[).
1563 * So we actually had +7+1 before we shift
1564 * down by 3. But (n+8) >> 3 == (n >> 3) + 1
1565 * (modulo overflows, which we do not have).
1567 * Finally, we LONG_ALIGN because all bitmap
1568 * operations are on longs.
1570 bitmap_size = (size-1) >> (i+4);
1571 bitmap_size = LONG_ALIGN(bitmap_size+1);
1572 zone->free_area[i].map =
1573 (unsigned long *) alloc_bootmem_node(pgdat, bitmap_size);
1578 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
1579 struct page *node_mem_map, unsigned long *zones_size,
1580 unsigned long node_start_pfn, unsigned long *zholes_size)
1584 pgdat->node_id = nid;
1585 pgdat->node_start_pfn = node_start_pfn;
1586 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
1587 if (!node_mem_map) {
1588 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
1589 node_mem_map = alloc_bootmem_node(pgdat, size);
1591 pgdat->node_mem_map = node_mem_map;
1593 free_area_init_core(pgdat, zones_size, zholes_size);
1596 #ifndef CONFIG_DISCONTIGMEM
1597 static bootmem_data_t contig_bootmem_data;
1598 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
1600 EXPORT_SYMBOL(contig_page_data);
1602 void __init free_area_init(unsigned long *zones_size)
1604 free_area_init_node(0, &contig_page_data, NULL, zones_size,
1605 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
1606 mem_map = contig_page_data.node_mem_map;
1610 #ifdef CONFIG_PROC_FS
1612 #include <linux/seq_file.h>
1614 static void *frag_start(struct seq_file *m, loff_t *pos)
1619 for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
1625 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1627 pg_data_t *pgdat = (pg_data_t *)arg;
1630 return pgdat->pgdat_next;
1633 static void frag_stop(struct seq_file *m, void *arg)
1638 * This walks the freelist for each zone. Whilst this is slow, I'd rather
1639 * be slow here than slow down the fast path by keeping stats - mjbligh
1641 static int frag_show(struct seq_file *m, void *arg)
1643 pg_data_t *pgdat = (pg_data_t *)arg;
1645 struct zone *node_zones = pgdat->node_zones;
1646 unsigned long flags;
1649 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1650 if (!zone->present_pages)
1653 spin_lock_irqsave(&zone->lock, flags);
1654 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1655 for (order = 0; order < MAX_ORDER; ++order) {
1656 unsigned long nr_bufs = 0;
1657 struct list_head *elem;
1659 list_for_each(elem, &(zone->free_area[order].free_list))
1661 seq_printf(m, "%6lu ", nr_bufs);
1663 spin_unlock_irqrestore(&zone->lock, flags);
1669 struct seq_operations fragmentation_op = {
1670 .start = frag_start,
1676 static char *vmstat_text[] = {
1680 "nr_page_table_pages",
1705 "pgscan_kswapd_high",
1706 "pgscan_kswapd_normal",
1708 "pgscan_kswapd_dma",
1709 "pgscan_direct_high",
1710 "pgscan_direct_normal",
1711 "pgscan_direct_dma",
1716 "kswapd_inodesteal",
1723 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1725 struct page_state *ps;
1727 if (*pos >= ARRAY_SIZE(vmstat_text))
1730 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
1733 return ERR_PTR(-ENOMEM);
1734 get_full_page_state(ps);
1735 ps->pgpgin /= 2; /* sectors -> kbytes */
1737 return (unsigned long *)ps + *pos;
1740 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1743 if (*pos >= ARRAY_SIZE(vmstat_text))
1745 return (unsigned long *)m->private + *pos;
1748 static int vmstat_show(struct seq_file *m, void *arg)
1750 unsigned long *l = arg;
1751 unsigned long off = l - (unsigned long *)m->private;
1753 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1757 static void vmstat_stop(struct seq_file *m, void *arg)
1763 struct seq_operations vmstat_op = {
1764 .start = vmstat_start,
1765 .next = vmstat_next,
1766 .stop = vmstat_stop,
1767 .show = vmstat_show,
1770 #endif /* CONFIG_PROC_FS */
1772 #ifdef CONFIG_HOTPLUG_CPU
1773 static int page_alloc_cpu_notify(struct notifier_block *self,
1774 unsigned long action, void *hcpu)
1776 int cpu = (unsigned long)hcpu;
1779 if (action == CPU_DEAD) {
1780 /* Drain local pagecache count. */
1781 count = &per_cpu(nr_pagecache_local, cpu);
1782 atomic_add(*count, &nr_pagecache);
1784 local_irq_disable();
1790 #endif /* CONFIG_HOTPLUG_CPU */
1792 void __init page_alloc_init(void)
1794 hotcpu_notifier(page_alloc_cpu_notify, 0);
1797 static unsigned long higherzone_val(struct zone *z, int max_zone,
1800 int z_idx = zone_idx(z);
1801 struct zone *higherzone;
1802 unsigned long pages;
1804 /* there is no higher zone to get a contribution from */
1805 if (z_idx == MAX_NR_ZONES-1)
1808 higherzone = &z->zone_pgdat->node_zones[z_idx+1];
1810 /* We always start with the higher zone's protection value */
1811 pages = higherzone->protection[alloc_type];
1814 * We get a lower-zone-protection contribution only if there are
1815 * pages in the higher zone and if we're not the highest zone
1816 * in the current zonelist. e.g., never happens for GFP_DMA. Happens
1817 * only for ZONE_DMA in a GFP_KERNEL allocation and happens for ZONE_DMA
1818 * and ZONE_NORMAL for a GFP_HIGHMEM allocation.
1820 if (higherzone->present_pages && z_idx < alloc_type)
1821 pages += higherzone->pages_low * sysctl_lower_zone_protection;
1827 * setup_per_zone_protection - called whenver min_free_kbytes or
1828 * sysctl_lower_zone_protection changes. Ensures that each zone
1829 * has a correct pages_protected value, so an adequate number of
1830 * pages are left in the zone after a successful __alloc_pages().
1832 * This algorithm is way confusing. I tries to keep the same behavior
1833 * as we had with the incremental min iterative algorithm.
1835 static void setup_per_zone_protection(void)
1837 struct pglist_data *pgdat;
1838 struct zone *zones, *zone;
1842 for_each_pgdat(pgdat) {
1843 zones = pgdat->node_zones;
1845 for (i = 0, max_zone = 0; i < MAX_NR_ZONES; i++)
1846 if (zones[i].present_pages)
1850 * For each of the different allocation types:
1851 * GFP_DMA -> GFP_KERNEL -> GFP_HIGHMEM
1853 for (i = 0; i < GFP_ZONETYPES; i++) {
1855 * For each of the zones:
1856 * ZONE_HIGHMEM -> ZONE_NORMAL -> ZONE_DMA
1858 for (j = MAX_NR_ZONES-1; j >= 0; j--) {
1862 * We never protect zones that don't have memory
1863 * in them (j>max_zone) or zones that aren't in
1864 * the zonelists for a certain type of
1865 * allocation (j>i). We have to assign these to
1866 * zero because the lower zones take
1867 * contributions from the higher zones.
1869 if (j > max_zone || j > i) {
1870 zone->protection[i] = 0;
1874 * The contribution of the next higher zone
1876 zone->protection[i] = higherzone_val(zone,
1878 zone->protection[i] += zone->pages_low;
1885 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
1886 * that the pages_{min,low,high} values for each zone are set correctly
1887 * with respect to min_free_kbytes.
1889 static void setup_per_zone_pages_min(void)
1891 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
1892 unsigned long lowmem_pages = 0;
1894 unsigned long flags;
1896 /* Calculate total number of !ZONE_HIGHMEM pages */
1897 for_each_zone(zone) {
1898 if (!is_highmem(zone))
1899 lowmem_pages += zone->present_pages;
1902 for_each_zone(zone) {
1903 spin_lock_irqsave(&zone->lru_lock, flags);
1904 if (is_highmem(zone)) {
1906 * Often, highmem doesn't need to reserve any pages.
1907 * But the pages_min/low/high values are also used for
1908 * batching up page reclaim activity so we need a
1909 * decent value here.
1913 min_pages = zone->present_pages / 1024;
1914 if (min_pages < SWAP_CLUSTER_MAX)
1915 min_pages = SWAP_CLUSTER_MAX;
1916 if (min_pages > 128)
1918 zone->pages_min = min_pages;
1920 /* if it's a lowmem zone, reserve a number of pages
1921 * proportionate to the zone's size.
1923 zone->pages_min = (pages_min * zone->present_pages) /
1927 zone->pages_low = zone->pages_min * 2;
1928 zone->pages_high = zone->pages_min * 3;
1929 spin_unlock_irqrestore(&zone->lru_lock, flags);
1934 * Initialise min_free_kbytes.
1936 * For small machines we want it small (128k min). For large machines
1937 * we want it large (16MB max). But it is not linear, because network
1938 * bandwidth does not increase linearly with machine size. We use
1940 * min_free_kbytes = sqrt(lowmem_kbytes)
1956 static int __init init_per_zone_pages_min(void)
1958 unsigned long lowmem_kbytes;
1960 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
1962 min_free_kbytes = int_sqrt(lowmem_kbytes);
1963 if (min_free_kbytes < 128)
1964 min_free_kbytes = 128;
1965 if (min_free_kbytes > 16384)
1966 min_free_kbytes = 16384;
1967 setup_per_zone_pages_min();
1968 setup_per_zone_protection();
1971 module_init(init_per_zone_pages_min)
1974 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
1975 * that we can call two helper functions whenever min_free_kbytes
1978 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
1979 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
1981 proc_dointvec(table, write, file, buffer, length, ppos);
1982 setup_per_zone_pages_min();
1983 setup_per_zone_protection();
1988 * lower_zone_protection_sysctl_handler - just a wrapper around
1989 * proc_dointvec() so that we can call setup_per_zone_protection()
1990 * whenever sysctl_lower_zone_protection changes.
1992 int lower_zone_protection_sysctl_handler(ctl_table *table, int write,
1993 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
1995 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
1996 setup_per_zone_protection();
2001 * allocate a large system hash table from bootmem
2002 * - it is assumed that the hash table must contain an exact power-of-2
2003 * quantity of entries
2005 void *__init alloc_large_system_hash(const char *tablename,
2006 unsigned long bucketsize,
2007 unsigned long numentries,
2009 int consider_highmem,
2010 unsigned int *_hash_shift,
2011 unsigned int *_hash_mask)
2013 unsigned long mem, max, log2qty, size;
2016 /* round applicable memory size up to nearest megabyte */
2017 mem = consider_highmem ? nr_all_pages : nr_kernel_pages;
2018 mem += (1UL << (20 - PAGE_SHIFT)) - 1;
2019 mem >>= 20 - PAGE_SHIFT;
2020 mem <<= 20 - PAGE_SHIFT;
2022 /* limit to 1 bucket per 2^scale bytes of low memory (rounded up to
2023 * nearest power of 2 in size) */
2024 if (scale > PAGE_SHIFT)
2025 mem >>= (scale - PAGE_SHIFT);
2027 mem <<= (PAGE_SHIFT - scale);
2029 mem = 1UL << (long_log2(mem) + 1);
2031 /* limit allocation size */
2032 max = (1UL << (PAGE_SHIFT + MAX_SYS_HASH_TABLE_ORDER)) / bucketsize;
2036 /* allow the kernel cmdline to have a say */
2037 if (!numentries || numentries > max)
2040 log2qty = long_log2(numentries);
2043 size = bucketsize << log2qty;
2045 table = (void *) alloc_bootmem(size);
2047 } while (!table && size > PAGE_SIZE);
2050 panic("Failed to allocate %s hash table\n", tablename);
2052 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2055 long_log2(size) - PAGE_SHIFT,
2059 *_hash_shift = log2qty;
2061 *_hash_mask = (1 << log2qty) - 1;