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>
37 #include <asm/tlbflush.h>
39 DECLARE_BITMAP(node_online_map, MAX_NUMNODES);
40 struct pglist_data *pgdat_list;
41 unsigned long totalram_pages;
42 unsigned long totalhigh_pages;
45 int sysctl_lower_zone_protection = 0;
47 EXPORT_SYMBOL(totalram_pages);
48 EXPORT_SYMBOL(nr_swap_pages);
51 * Used by page_zone() to look up the address of the struct zone whose
52 * id is encoded in the upper bits of page->flags
54 struct zone *zone_table[1 << (ZONES_SHIFT + NODES_SHIFT)];
55 EXPORT_SYMBOL(zone_table);
57 static char *zone_names[MAX_NR_ZONES] = { "DMA", "Normal", "HighMem" };
58 int min_free_kbytes = 1024;
61 * Temporary debugging check for pages not lying within a given zone.
63 static int bad_range(struct zone *zone, struct page *page)
65 if (page_to_pfn(page) >= zone->zone_start_pfn + zone->spanned_pages)
67 if (page_to_pfn(page) < zone->zone_start_pfn)
69 if (zone != page_zone(page))
74 static void bad_page(const char *function, struct page *page)
76 printk(KERN_EMERG "Bad page state at %s (in process '%s', page %p)\n",
77 function, current->comm, page);
78 printk(KERN_EMERG "flags:0x%08lx mapping:%p mapcount:%d count:%d\n",
79 (unsigned long)page->flags, page->mapping,
80 (int)page->mapcount, page_count(page));
81 printk(KERN_EMERG "Backtrace:\n");
83 printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n");
84 page->flags &= ~(1 << PG_private |
93 set_page_count(page, 0);
98 #ifndef CONFIG_HUGETLB_PAGE
99 #define prep_compound_page(page, order) do { } while (0)
100 #define destroy_compound_page(page, order) do { } while (0)
103 * Higher-order pages are called "compound pages". They are structured thusly:
105 * The first PAGE_SIZE page is called the "head page".
107 * The remaining PAGE_SIZE pages are called "tail pages".
109 * All pages have PG_compound set. All pages have their ->private pointing at
110 * the head page (even the head page has this).
112 * The first tail page's ->mapping, if non-zero, holds the address of the
113 * compound page's put_page() function.
115 * The order of the allocation is stored in the first tail page's ->index
116 * This is only for debug at present. This usage means that zero-order pages
117 * may not be compound.
119 static void prep_compound_page(struct page *page, unsigned long order)
122 int nr_pages = 1 << order;
125 page[1].index = order;
126 for (i = 0; i < nr_pages; i++) {
127 struct page *p = page + i;
130 p->private = (unsigned long)page;
134 static void destroy_compound_page(struct page *page, unsigned long order)
137 int nr_pages = 1 << order;
139 if (!PageCompound(page))
142 if (page[1].index != order)
143 bad_page(__FUNCTION__, page);
145 for (i = 0; i < nr_pages; i++) {
146 struct page *p = page + i;
148 if (!PageCompound(p))
149 bad_page(__FUNCTION__, page);
150 if (p->private != (unsigned long)page)
151 bad_page(__FUNCTION__, page);
152 ClearPageCompound(p);
155 #endif /* CONFIG_HUGETLB_PAGE */
158 * Freeing function for a buddy system allocator.
160 * The concept of a buddy system is to maintain direct-mapped table
161 * (containing bit values) for memory blocks of various "orders".
162 * The bottom level table contains the map for the smallest allocatable
163 * units of memory (here, pages), and each level above it describes
164 * pairs of units from the levels below, hence, "buddies".
165 * At a high level, all that happens here is marking the table entry
166 * at the bottom level available, and propagating the changes upward
167 * as necessary, plus some accounting needed to play nicely with other
168 * parts of the VM system.
169 * At each level, we keep one bit for each pair of blocks, which
170 * is set to 1 iff only one of the pair is allocated. So when we
171 * are allocating or freeing one, we can derive the state of the
172 * other. That is, if we allocate a small block, and both were
173 * free, the remainder of the region must be split into blocks.
174 * If a block is freed, and its buddy is also free, then this
175 * triggers coalescing into a block of larger size.
180 static inline void __free_pages_bulk (struct page *page, struct page *base,
181 struct zone *zone, struct free_area *area, unsigned long mask,
184 unsigned long page_idx, index;
187 destroy_compound_page(page, order);
188 page_idx = page - base;
189 if (page_idx & ~mask)
191 index = page_idx >> (1 + order);
193 zone->free_pages -= mask;
194 while (mask + (1 << (MAX_ORDER-1))) {
195 struct page *buddy1, *buddy2;
197 BUG_ON(area >= zone->free_area + MAX_ORDER);
198 if (!__test_and_change_bit(index, area->map))
200 * the buddy page is still allocated.
204 * Move the buddy up one level.
205 * This code is taking advantage of the identity:
208 buddy1 = base + (page_idx ^ -mask);
209 buddy2 = base + page_idx;
210 BUG_ON(bad_range(zone, buddy1));
211 BUG_ON(bad_range(zone, buddy2));
212 list_del(&buddy1->lru);
218 list_add(&(base + page_idx)->lru, &area->free_list);
221 static inline void free_pages_check(const char *function, struct page *page)
223 if ( page_mapped(page) ||
224 page->mapping != NULL ||
225 page_count(page) != 0 ||
236 1 << PG_writeback )))
237 bad_page(function, page);
239 ClearPageDirty(page);
243 * Frees a list of pages.
244 * Assumes all pages on list are in same zone, and of same order.
245 * count is the number of pages to free, or 0 for all on the list.
247 * If the zone was previously in an "all pages pinned" state then look to
248 * see if this freeing clears that state.
250 * And clear the zone's pages_scanned counter, to hold off the "all pages are
251 * pinned" detection logic.
254 free_pages_bulk(struct zone *zone, int count,
255 struct list_head *list, unsigned int order)
257 unsigned long mask, flags;
258 struct free_area *area;
259 struct page *base, *page = NULL;
262 mask = (~0UL) << order;
263 base = zone->zone_mem_map;
264 area = zone->free_area + order;
265 spin_lock_irqsave(&zone->lock, flags);
266 zone->all_unreclaimable = 0;
267 zone->pages_scanned = 0;
268 while (!list_empty(list) && count--) {
269 page = list_entry(list->prev, struct page, lru);
270 /* have to delete it as __free_pages_bulk list manipulates */
271 list_del(&page->lru);
272 __free_pages_bulk(page, base, zone, area, mask, order);
275 spin_unlock_irqrestore(&zone->lock, flags);
279 void __free_pages_ok(struct page *page, unsigned int order)
284 mod_page_state(pgfree, 1 << order);
285 for (i = 0 ; i < (1 << order) ; ++i)
286 free_pages_check(__FUNCTION__, page + i);
287 list_add(&page->lru, &list);
288 kernel_map_pages(page, 1<<order, 0);
289 free_pages_bulk(page_zone(page), 1, &list, order);
292 #define MARK_USED(index, order, area) \
293 __change_bit((index) >> (1+(order)), (area)->map)
295 static inline struct page *
296 expand(struct zone *zone, struct page *page,
297 unsigned long index, int low, int high, struct free_area *area)
299 unsigned long size = 1 << high;
302 BUG_ON(bad_range(zone, page));
306 list_add(&page->lru, &area->free_list);
307 MARK_USED(index, high, area);
314 static inline void set_page_refs(struct page *page, int order)
317 set_page_count(page, 1);
322 * We need to reference all the pages for this order, otherwise if
323 * anyone accesses one of the pages with (get/put) it will be freed.
325 for (i = 0; i < (1 << order); i++)
326 set_page_count(page+i, 1);
327 #endif /* CONFIG_MMU */
331 * This page is about to be returned from the page allocator
333 static void prep_new_page(struct page *page, int order)
335 if (page->mapping || page_mapped(page) ||
346 1 << PG_writeback )))
347 bad_page(__FUNCTION__, page);
349 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
350 1 << PG_referenced | 1 << PG_arch_1 |
351 1 << PG_checked | 1 << PG_mappedtodisk);
353 set_page_refs(page, order);
357 * Do the hard work of removing an element from the buddy allocator.
358 * Call me with the zone->lock already held.
360 static struct page *__rmqueue(struct zone *zone, unsigned int order)
362 struct free_area * area;
363 unsigned int current_order;
367 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
368 area = zone->free_area + current_order;
369 if (list_empty(&area->free_list))
372 page = list_entry(area->free_list.next, struct page, lru);
373 list_del(&page->lru);
374 index = page - zone->zone_mem_map;
375 if (current_order != MAX_ORDER-1)
376 MARK_USED(index, current_order, area);
377 zone->free_pages -= 1UL << order;
378 return expand(zone, page, index, order, current_order, area);
385 * Obtain a specified number of elements from the buddy allocator, all under
386 * a single hold of the lock, for efficiency. Add them to the supplied list.
387 * Returns the number of new pages which were placed at *list.
389 static int rmqueue_bulk(struct zone *zone, unsigned int order,
390 unsigned long count, struct list_head *list)
397 spin_lock_irqsave(&zone->lock, flags);
398 for (i = 0; i < count; ++i) {
399 page = __rmqueue(zone, order);
403 list_add_tail(&page->lru, list);
405 spin_unlock_irqrestore(&zone->lock, flags);
409 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
410 static void __drain_pages(unsigned int cpu)
415 for_each_zone(zone) {
416 struct per_cpu_pageset *pset;
418 pset = &zone->pageset[cpu];
419 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
420 struct per_cpu_pages *pcp;
423 pcp->count -= free_pages_bulk(zone, pcp->count,
428 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
431 int is_head_of_free_region(struct page *page)
433 struct zone *zone = page_zone(page);
436 struct list_head *curr;
439 * Should not matter as we need quiescent system for
440 * suspend anyway, but...
442 spin_lock_irqsave(&zone->lock, flags);
443 for (order = MAX_ORDER - 1; order >= 0; --order)
444 list_for_each(curr, &zone->free_area[order].free_list)
445 if (page == list_entry(curr, struct page, lru)) {
446 spin_unlock_irqrestore(&zone->lock, flags);
449 spin_unlock_irqrestore(&zone->lock, flags);
454 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
456 void drain_local_pages(void)
460 local_irq_save(flags);
461 __drain_pages(smp_processor_id());
462 local_irq_restore(flags);
464 #endif /* CONFIG_PM */
466 static void zone_statistics(struct zonelist *zonelist, struct zone *z)
471 pg_data_t *pg = z->zone_pgdat;
472 pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
473 struct per_cpu_pageset *p;
475 local_irq_save(flags);
476 cpu = smp_processor_id();
477 p = &z->pageset[cpu];
479 z->pageset[cpu].numa_hit++;
482 zonelist->zones[0]->pageset[cpu].numa_foreign++;
484 if (pg == NODE_DATA(numa_node_id()))
488 local_irq_restore(flags);
493 * Free a 0-order page
495 static void FASTCALL(free_hot_cold_page(struct page *page, int cold));
496 static void fastcall free_hot_cold_page(struct page *page, int cold)
498 struct zone *zone = page_zone(page);
499 struct per_cpu_pages *pcp;
502 kernel_map_pages(page, 1, 0);
503 inc_page_state(pgfree);
504 free_pages_check(__FUNCTION__, page);
505 pcp = &zone->pageset[get_cpu()].pcp[cold];
506 local_irq_save(flags);
507 if (pcp->count >= pcp->high)
508 pcp->count -= free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
509 list_add(&page->lru, &pcp->list);
511 local_irq_restore(flags);
515 void fastcall free_hot_page(struct page *page)
517 free_hot_cold_page(page, 0);
520 void fastcall free_cold_page(struct page *page)
522 free_hot_cold_page(page, 1);
526 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
527 * we cheat by calling it from here, in the order > 0 path. Saves a branch
532 buffered_rmqueue(struct zone *zone, int order, int gfp_flags)
535 struct page *page = NULL;
536 int cold = !!(gfp_flags & __GFP_COLD);
539 struct per_cpu_pages *pcp;
541 pcp = &zone->pageset[get_cpu()].pcp[cold];
542 local_irq_save(flags);
543 if (pcp->count <= pcp->low)
544 pcp->count += rmqueue_bulk(zone, 0,
545 pcp->batch, &pcp->list);
547 page = list_entry(pcp->list.next, struct page, lru);
548 list_del(&page->lru);
551 local_irq_restore(flags);
556 spin_lock_irqsave(&zone->lock, flags);
557 page = __rmqueue(zone, order);
558 spin_unlock_irqrestore(&zone->lock, flags);
562 BUG_ON(bad_range(zone, page));
563 mod_page_state_zone(zone, pgalloc, 1 << order);
564 prep_new_page(page, order);
565 if (order && (gfp_flags & __GFP_COMP))
566 prep_compound_page(page, order);
572 * This is the 'heart' of the zoned buddy allocator.
574 * Herein lies the mysterious "incremental min". That's the
576 * local_low = z->pages_low;
579 * thing. The intent here is to provide additional protection to low zones for
580 * allocation requests which _could_ use higher zones. So a GFP_HIGHMEM
581 * request is not allowed to dip as deeply into the normal zone as a GFP_KERNEL
582 * request. This preserves additional space in those lower zones for requests
583 * which really do need memory from those zones. It means that on a decent
584 * sized machine, GFP_HIGHMEM and GFP_KERNEL requests basically leave the DMA
587 struct page * fastcall
588 __alloc_pages(unsigned int gfp_mask, unsigned int order,
589 struct zonelist *zonelist)
591 const int wait = gfp_mask & __GFP_WAIT;
595 struct reclaim_state reclaim_state;
596 struct task_struct *p = current;
601 might_sleep_if(wait);
603 zones = zonelist->zones; /* the list of zones suitable for gfp_mask */
604 if (zones[0] == NULL) /* no zones in the zonelist */
607 alloc_type = zone_idx(zones[0]);
609 /* Go through the zonelist once, looking for a zone with enough free */
610 for (i = 0; zones[i] != NULL; i++) {
611 struct zone *z = zones[i];
613 min = (1<<order) + z->protection[alloc_type];
616 * We let real-time tasks dip their real-time paws a little
617 * deeper into reserves.
620 min -= z->pages_low >> 1;
622 if (z->free_pages >= min ||
623 (!wait && z->free_pages >= z->pages_high)) {
624 page = buffered_rmqueue(z, order, gfp_mask);
626 zone_statistics(zonelist, z);
632 /* we're somewhat low on memory, failed to find what we needed */
633 for (i = 0; zones[i] != NULL; i++)
634 wakeup_kswapd(zones[i]);
636 /* Go through the zonelist again, taking __GFP_HIGH into account */
637 for (i = 0; zones[i] != NULL; i++) {
638 struct zone *z = zones[i];
640 min = (1<<order) + z->protection[alloc_type];
642 if (gfp_mask & __GFP_HIGH)
643 min -= z->pages_low >> 2;
645 min -= z->pages_low >> 1;
647 if (z->free_pages >= min ||
648 (!wait && z->free_pages >= z->pages_high)) {
649 page = buffered_rmqueue(z, order, gfp_mask);
651 zone_statistics(zonelist, z);
657 /* here we're in the low on memory slow path */
660 if ((p->flags & (PF_MEMALLOC | PF_MEMDIE)) && !in_interrupt()) {
661 /* go through the zonelist yet again, ignoring mins */
662 for (i = 0; zones[i] != NULL; i++) {
663 struct zone *z = zones[i];
665 page = buffered_rmqueue(z, order, gfp_mask);
667 zone_statistics(zonelist, z);
674 /* Atomic allocations - we can't balance anything */
678 p->flags |= PF_MEMALLOC;
679 reclaim_state.reclaimed_slab = 0;
680 p->reclaim_state = &reclaim_state;
682 try_to_free_pages(zones, gfp_mask, order);
684 p->reclaim_state = NULL;
685 p->flags &= ~PF_MEMALLOC;
687 /* go through the zonelist yet one more time */
688 for (i = 0; zones[i] != NULL; i++) {
689 struct zone *z = zones[i];
691 min = (1UL << order) + z->protection[alloc_type];
693 if (z->free_pages >= min ||
694 (!wait && z->free_pages >= z->pages_high)) {
695 page = buffered_rmqueue(z, order, gfp_mask);
697 zone_statistics(zonelist, z);
704 * Don't let big-order allocations loop unless the caller explicitly
705 * requests that. Wait for some write requests to complete then retry.
707 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL, but that
708 * may not be true in other implementations.
711 if (!(gfp_mask & __GFP_NORETRY)) {
712 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
714 if (gfp_mask & __GFP_NOFAIL)
718 blk_congestion_wait(WRITE, HZ/50);
723 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
724 printk(KERN_WARNING "%s: page allocation failure."
725 " order:%d, mode:0x%x\n",
726 p->comm, order, gfp_mask);
731 kernel_map_pages(page, 1 << order, 1);
735 EXPORT_SYMBOL(__alloc_pages);
738 /* Early boot: Everything is done by one cpu, but the data structures will be
739 * used by all cpus - spread them on all nodes.
741 static __init unsigned long get_boot_pages(unsigned int gfp_mask, unsigned int order)
748 if (i > nodenr + numnodes)
750 if (node_present_pages(i%numnodes)) {
752 /* The node contains memory. Check that there is
753 * memory in the intended zonelist.
755 z = NODE_DATA(i%numnodes)->node_zonelists[gfp_mask & GFP_ZONEMASK].zones;
757 if ( (*z)->free_pages > (1UL<<order))
766 page = alloc_pages_node(i%numnodes, gfp_mask, order);
769 return (unsigned long) page_address(page);
774 * Common helper functions.
776 fastcall unsigned long __get_free_pages(unsigned int gfp_mask, unsigned int order)
781 if (unlikely(system_state == SYSTEM_BOOTING))
782 return get_boot_pages(gfp_mask, order);
784 page = alloc_pages(gfp_mask, order);
787 return (unsigned long) page_address(page);
790 EXPORT_SYMBOL(__get_free_pages);
792 fastcall unsigned long get_zeroed_page(unsigned int gfp_mask)
797 * get_zeroed_page() returns a 32-bit address, which cannot represent
800 BUG_ON(gfp_mask & __GFP_HIGHMEM);
802 page = alloc_pages(gfp_mask, 0);
804 void *address = page_address(page);
806 return (unsigned long) address;
811 EXPORT_SYMBOL(get_zeroed_page);
813 void __pagevec_free(struct pagevec *pvec)
815 int i = pagevec_count(pvec);
818 free_hot_cold_page(pvec->pages[i], pvec->cold);
821 fastcall void __free_pages(struct page *page, unsigned int order)
823 if (!PageReserved(page) && put_page_testzero(page)) {
827 __free_pages_ok(page, order);
831 EXPORT_SYMBOL(__free_pages);
833 fastcall void free_pages(unsigned long addr, unsigned int order)
836 BUG_ON(!virt_addr_valid(addr));
837 __free_pages(virt_to_page(addr), order);
841 EXPORT_SYMBOL(free_pages);
844 * Total amount of free (allocatable) RAM:
846 unsigned int nr_free_pages(void)
848 unsigned int sum = 0;
852 sum += zone->free_pages;
857 EXPORT_SYMBOL(nr_free_pages);
859 unsigned int nr_used_zone_pages(void)
861 unsigned int pages = 0;
865 pages += zone->nr_active + zone->nr_inactive;
871 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
873 unsigned int i, sum = 0;
875 for (i = 0; i < MAX_NR_ZONES; i++)
876 sum += pgdat->node_zones[i].free_pages;
882 static unsigned int nr_free_zone_pages(int offset)
885 unsigned int sum = 0;
887 for_each_pgdat(pgdat) {
888 struct zonelist *zonelist = pgdat->node_zonelists + offset;
889 struct zone **zonep = zonelist->zones;
892 for (zone = *zonep++; zone; zone = *zonep++) {
893 unsigned long size = zone->present_pages;
894 unsigned long high = zone->pages_high;
904 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
906 unsigned int nr_free_buffer_pages(void)
908 return nr_free_zone_pages(GFP_USER & GFP_ZONEMASK);
912 * Amount of free RAM allocatable within all zones
914 unsigned int nr_free_pagecache_pages(void)
916 return nr_free_zone_pages(GFP_HIGHUSER & GFP_ZONEMASK);
919 #ifdef CONFIG_HIGHMEM
920 unsigned int nr_free_highpages (void)
923 unsigned int pages = 0;
925 for_each_pgdat(pgdat)
926 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
933 static void show_node(struct zone *zone)
935 printk("Node %d ", zone->zone_pgdat->node_id);
938 #define show_node(zone) do { } while (0)
942 * Accumulate the page_state information across all CPUs.
943 * The result is unavoidably approximate - it can change
944 * during and after execution of this function.
946 DEFINE_PER_CPU(struct page_state, page_states) = {0};
947 EXPORT_PER_CPU_SYMBOL(page_states);
949 atomic_t nr_pagecache = ATOMIC_INIT(0);
950 EXPORT_SYMBOL(nr_pagecache);
952 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
955 void __get_page_state(struct page_state *ret, int nr)
959 memset(ret, 0, sizeof(*ret));
960 while (cpu < NR_CPUS) {
961 unsigned long *in, *out, off;
963 if (!cpu_possible(cpu)) {
968 in = (unsigned long *)&per_cpu(page_states, cpu);
970 if (cpu < NR_CPUS && cpu_possible(cpu))
971 prefetch(&per_cpu(page_states, cpu));
972 out = (unsigned long *)ret;
973 for (off = 0; off < nr; off++)
978 void get_page_state(struct page_state *ret)
982 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
983 nr /= sizeof(unsigned long);
985 __get_page_state(ret, nr + 1);
988 void get_full_page_state(struct page_state *ret)
990 __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long));
993 unsigned long __read_page_state(unsigned offset)
995 unsigned long ret = 0;
998 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1001 if (!cpu_possible(cpu))
1004 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
1005 ret += *((unsigned long *)in);
1010 void get_zone_counts(unsigned long *active,
1011 unsigned long *inactive, unsigned long *free)
1018 for_each_zone(zone) {
1019 *active += zone->nr_active;
1020 *inactive += zone->nr_inactive;
1021 *free += zone->free_pages;
1025 void si_meminfo(struct sysinfo *val)
1027 val->totalram = totalram_pages;
1029 val->freeram = nr_free_pages();
1030 val->bufferram = nr_blockdev_pages();
1031 #ifdef CONFIG_HIGHMEM
1032 val->totalhigh = totalhigh_pages;
1033 val->freehigh = nr_free_highpages();
1038 val->mem_unit = PAGE_SIZE;
1039 if (vx_flags(VXF_VIRT_MEM, 0))
1040 vx_vsi_meminfo(val);
1043 EXPORT_SYMBOL(si_meminfo);
1046 void si_meminfo_node(struct sysinfo *val, int nid)
1048 pg_data_t *pgdat = NODE_DATA(nid);
1050 val->totalram = pgdat->node_present_pages;
1051 val->freeram = nr_free_pages_pgdat(pgdat);
1052 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1053 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1054 val->mem_unit = PAGE_SIZE;
1058 #define K(x) ((x) << (PAGE_SHIFT-10))
1061 * Show free area list (used inside shift_scroll-lock stuff)
1062 * We also calculate the percentage fragmentation. We do this by counting the
1063 * memory on each free list with the exception of the first item on the list.
1065 void show_free_areas(void)
1067 struct page_state ps;
1068 int cpu, temperature;
1069 unsigned long active;
1070 unsigned long inactive;
1074 for_each_zone(zone) {
1076 printk("%s per-cpu:", zone->name);
1078 if (!zone->present_pages) {
1084 for (cpu = 0; cpu < NR_CPUS; ++cpu) {
1085 struct per_cpu_pageset *pageset;
1087 if (!cpu_possible(cpu))
1090 pageset = zone->pageset + cpu;
1092 for (temperature = 0; temperature < 2; temperature++)
1093 printk("cpu %d %s: low %d, high %d, batch %d\n",
1095 temperature ? "cold" : "hot",
1096 pageset->pcp[temperature].low,
1097 pageset->pcp[temperature].high,
1098 pageset->pcp[temperature].batch);
1102 get_page_state(&ps);
1103 get_zone_counts(&active, &inactive, &free);
1105 printk("\nFree pages: %11ukB (%ukB HighMem)\n",
1107 K(nr_free_highpages()));
1109 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1110 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1119 ps.nr_page_table_pages);
1121 for_each_zone(zone) {
1135 K(zone->free_pages),
1138 K(zone->pages_high),
1140 K(zone->nr_inactive),
1141 K(zone->present_pages)
1143 printk("protections[]:");
1144 for (i = 0; i < MAX_NR_ZONES; i++)
1145 printk(" %lu", zone->protection[i]);
1149 for_each_zone(zone) {
1150 struct list_head *elem;
1151 unsigned long nr, flags, order, total = 0;
1154 printk("%s: ", zone->name);
1155 if (!zone->present_pages) {
1160 spin_lock_irqsave(&zone->lock, flags);
1161 for (order = 0; order < MAX_ORDER; order++) {
1163 list_for_each(elem, &zone->free_area[order].free_list)
1165 total += nr << order;
1166 printk("%lu*%lukB ", nr, K(1UL) << order);
1168 spin_unlock_irqrestore(&zone->lock, flags);
1169 printk("= %lukB\n", K(total));
1172 show_swap_cache_info();
1176 * Builds allocation fallback zone lists.
1178 static int __init build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, int j, int k)
1185 zone = pgdat->node_zones + ZONE_HIGHMEM;
1186 if (zone->present_pages) {
1187 #ifndef CONFIG_HIGHMEM
1190 zonelist->zones[j++] = zone;
1193 zone = pgdat->node_zones + ZONE_NORMAL;
1194 if (zone->present_pages)
1195 zonelist->zones[j++] = zone;
1197 zone = pgdat->node_zones + ZONE_DMA;
1198 if (zone->present_pages)
1199 zonelist->zones[j++] = zone;
1206 #define MAX_NODE_LOAD (numnodes)
1207 static int __initdata node_load[MAX_NUMNODES];
1209 * find_next_best_node - find the next node that should appear in a given
1210 * node's fallback list
1211 * @node: node whose fallback list we're appending
1212 * @used_node_mask: pointer to the bitmap of already used nodes
1214 * We use a number of factors to determine which is the next node that should
1215 * appear on a given node's fallback list. The node should not have appeared
1216 * already in @node's fallback list, and it should be the next closest node
1217 * according to the distance array (which contains arbitrary distance values
1218 * from each node to each node in the system), and should also prefer nodes
1219 * with no CPUs, since presumably they'll have very little allocation pressure
1220 * on them otherwise.
1221 * It returns -1 if no node is found.
1223 static int __init find_next_best_node(int node, void *used_node_mask)
1226 int min_val = INT_MAX;
1229 for (i = 0; i < numnodes; i++) {
1232 /* Start from local node */
1233 n = (node+i)%numnodes;
1235 /* Don't want a node to appear more than once */
1236 if (test_bit(n, used_node_mask))
1239 /* Use the distance array to find the distance */
1240 val = node_distance(node, n);
1242 /* Give preference to headless and unused nodes */
1243 tmp = node_to_cpumask(n);
1244 if (!cpus_empty(tmp))
1245 val += PENALTY_FOR_NODE_WITH_CPUS;
1247 /* Slight preference for less loaded node */
1248 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1249 val += node_load[n];
1251 if (val < min_val) {
1258 set_bit(best_node, used_node_mask);
1263 static void __init build_zonelists(pg_data_t *pgdat)
1265 int i, j, k, node, local_node;
1266 int prev_node, load;
1267 struct zonelist *zonelist;
1268 DECLARE_BITMAP(used_mask, MAX_NUMNODES);
1270 /* initialize zonelists */
1271 for (i = 0; i < MAX_NR_ZONES; i++) {
1272 zonelist = pgdat->node_zonelists + i;
1273 memset(zonelist, 0, sizeof(*zonelist));
1274 zonelist->zones[0] = NULL;
1277 /* NUMA-aware ordering of nodes */
1278 local_node = pgdat->node_id;
1280 prev_node = local_node;
1281 bitmap_zero(used_mask, MAX_NUMNODES);
1282 while ((node = find_next_best_node(local_node, used_mask)) >= 0) {
1284 * We don't want to pressure a particular node.
1285 * So adding penalty to the first node in same
1286 * distance group to make it round-robin.
1288 if (node_distance(local_node, node) !=
1289 node_distance(local_node, prev_node))
1290 node_load[node] += load;
1293 for (i = 0; i < MAX_NR_ZONES; i++) {
1294 zonelist = pgdat->node_zonelists + i;
1295 for (j = 0; zonelist->zones[j] != NULL; j++);
1298 if (i & __GFP_HIGHMEM)
1303 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1304 zonelist->zones[j] = NULL;
1309 #else /* CONFIG_NUMA */
1311 static void __init build_zonelists(pg_data_t *pgdat)
1313 int i, j, k, node, local_node;
1315 local_node = pgdat->node_id;
1316 for (i = 0; i < MAX_NR_ZONES; i++) {
1317 struct zonelist *zonelist;
1319 zonelist = pgdat->node_zonelists + i;
1320 memset(zonelist, 0, sizeof(*zonelist));
1324 if (i & __GFP_HIGHMEM)
1329 j = build_zonelists_node(pgdat, zonelist, j, k);
1331 * Now we build the zonelist so that it contains the zones
1332 * of all the other nodes.
1333 * We don't want to pressure a particular node, so when
1334 * building the zones for node N, we make sure that the
1335 * zones coming right after the local ones are those from
1336 * node N+1 (modulo N)
1338 for (node = local_node + 1; node < numnodes; node++)
1339 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1340 for (node = 0; node < local_node; node++)
1341 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1343 zonelist->zones[j] = NULL;
1347 #endif /* CONFIG_NUMA */
1349 void __init build_all_zonelists(void)
1353 for(i = 0 ; i < numnodes ; i++)
1354 build_zonelists(NODE_DATA(i));
1355 printk("Built %i zonelists\n", numnodes);
1359 * Helper functions to size the waitqueue hash table.
1360 * Essentially these want to choose hash table sizes sufficiently
1361 * large so that collisions trying to wait on pages are rare.
1362 * But in fact, the number of active page waitqueues on typical
1363 * systems is ridiculously low, less than 200. So this is even
1364 * conservative, even though it seems large.
1366 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1367 * waitqueues, i.e. the size of the waitq table given the number of pages.
1369 #define PAGES_PER_WAITQUEUE 256
1371 static inline unsigned long wait_table_size(unsigned long pages)
1373 unsigned long size = 1;
1375 pages /= PAGES_PER_WAITQUEUE;
1377 while (size < pages)
1381 * Once we have dozens or even hundreds of threads sleeping
1382 * on IO we've got bigger problems than wait queue collision.
1383 * Limit the size of the wait table to a reasonable size.
1385 size = min(size, 4096UL);
1387 return max(size, 4UL);
1391 * This is an integer logarithm so that shifts can be used later
1392 * to extract the more random high bits from the multiplicative
1393 * hash function before the remainder is taken.
1395 static inline unsigned long wait_table_bits(unsigned long size)
1400 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1402 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1403 unsigned long *zones_size, unsigned long *zholes_size)
1405 unsigned long realtotalpages, totalpages = 0;
1408 for (i = 0; i < MAX_NR_ZONES; i++)
1409 totalpages += zones_size[i];
1410 pgdat->node_spanned_pages = totalpages;
1412 realtotalpages = totalpages;
1414 for (i = 0; i < MAX_NR_ZONES; i++)
1415 realtotalpages -= zholes_size[i];
1416 pgdat->node_present_pages = realtotalpages;
1417 printk("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1422 * Initially all pages are reserved - free ones are freed
1423 * up by free_all_bootmem() once the early boot process is
1424 * done. Non-atomic initialization, single-pass.
1426 void __init memmap_init_zone(struct page *start, unsigned long size, int nid,
1427 unsigned long zone, unsigned long start_pfn)
1431 for (page = start; page < (start + size); page++) {
1432 set_page_zone(page, NODEZONE(nid, zone));
1433 set_page_count(page, 0);
1434 SetPageReserved(page);
1435 INIT_LIST_HEAD(&page->lru);
1436 #ifdef WANT_PAGE_VIRTUAL
1437 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1438 if (zone != ZONE_HIGHMEM)
1439 set_page_address(page, __va(start_pfn << PAGE_SHIFT));
1445 #ifndef __HAVE_ARCH_MEMMAP_INIT
1446 #define memmap_init(start, size, nid, zone, start_pfn) \
1447 memmap_init_zone((start), (size), (nid), (zone), (start_pfn))
1451 * Set up the zone data structures:
1452 * - mark all pages reserved
1453 * - mark all memory queues empty
1454 * - clear the memory bitmaps
1456 static void __init free_area_init_core(struct pglist_data *pgdat,
1457 unsigned long *zones_size, unsigned long *zholes_size)
1460 const unsigned long zone_required_alignment = 1UL << (MAX_ORDER-1);
1461 int cpu, nid = pgdat->node_id;
1462 struct page *lmem_map = pgdat->node_mem_map;
1463 unsigned long zone_start_pfn = pgdat->node_start_pfn;
1465 pgdat->nr_zones = 0;
1466 init_waitqueue_head(&pgdat->kswapd_wait);
1468 for (j = 0; j < MAX_NR_ZONES; j++) {
1469 struct zone *zone = pgdat->node_zones + j;
1470 unsigned long size, realsize;
1471 unsigned long batch;
1473 zone_table[NODEZONE(nid, j)] = zone;
1474 realsize = size = zones_size[j];
1476 realsize -= zholes_size[j];
1478 zone->spanned_pages = size;
1479 zone->present_pages = realsize;
1480 zone->name = zone_names[j];
1481 spin_lock_init(&zone->lock);
1482 spin_lock_init(&zone->lru_lock);
1483 zone->zone_pgdat = pgdat;
1484 zone->free_pages = 0;
1486 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
1489 * The per-cpu-pages pools are set to around 1000th of the
1490 * size of the zone. But no more than 1/4 of a meg - there's
1491 * no point in going beyond the size of L2 cache.
1493 * OK, so we don't know how big the cache is. So guess.
1495 batch = zone->present_pages / 1024;
1496 if (batch * PAGE_SIZE > 256 * 1024)
1497 batch = (256 * 1024) / PAGE_SIZE;
1498 batch /= 4; /* We effectively *= 4 below */
1502 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1503 struct per_cpu_pages *pcp;
1505 pcp = &zone->pageset[cpu].pcp[0]; /* hot */
1507 pcp->low = 2 * batch;
1508 pcp->high = 6 * batch;
1509 pcp->batch = 1 * batch;
1510 INIT_LIST_HEAD(&pcp->list);
1512 pcp = &zone->pageset[cpu].pcp[1]; /* cold */
1515 pcp->high = 2 * batch;
1516 pcp->batch = 1 * batch;
1517 INIT_LIST_HEAD(&pcp->list);
1519 printk(" %s zone: %lu pages, LIFO batch:%lu\n",
1520 zone_names[j], realsize, batch);
1521 INIT_LIST_HEAD(&zone->active_list);
1522 INIT_LIST_HEAD(&zone->inactive_list);
1523 atomic_set(&zone->nr_scan_active, 0);
1524 atomic_set(&zone->nr_scan_inactive, 0);
1525 zone->nr_active = 0;
1526 zone->nr_inactive = 0;
1531 * The per-page waitqueue mechanism uses hashed waitqueues
1534 zone->wait_table_size = wait_table_size(size);
1535 zone->wait_table_bits =
1536 wait_table_bits(zone->wait_table_size);
1537 zone->wait_table = (wait_queue_head_t *)
1538 alloc_bootmem_node(pgdat, zone->wait_table_size
1539 * sizeof(wait_queue_head_t));
1541 for(i = 0; i < zone->wait_table_size; ++i)
1542 init_waitqueue_head(zone->wait_table + i);
1544 pgdat->nr_zones = j+1;
1546 zone->zone_mem_map = lmem_map;
1547 zone->zone_start_pfn = zone_start_pfn;
1549 if ((zone_start_pfn) & (zone_required_alignment-1))
1550 printk("BUG: wrong zone alignment, it will crash\n");
1552 memmap_init(lmem_map, size, nid, j, zone_start_pfn);
1554 zone_start_pfn += size;
1557 for (i = 0; ; i++) {
1558 unsigned long bitmap_size;
1560 INIT_LIST_HEAD(&zone->free_area[i].free_list);
1561 if (i == MAX_ORDER-1) {
1562 zone->free_area[i].map = NULL;
1567 * Page buddy system uses "index >> (i+1)",
1568 * where "index" is at most "size-1".
1570 * The extra "+3" is to round down to byte
1571 * size (8 bits per byte assumption). Thus
1572 * we get "(size-1) >> (i+4)" as the last byte
1575 * The "+1" is because we want to round the
1576 * byte allocation up rather than down. So
1577 * we should have had a "+7" before we shifted
1578 * down by three. Also, we have to add one as
1579 * we actually _use_ the last bit (it's [0,n]
1580 * inclusive, not [0,n[).
1582 * So we actually had +7+1 before we shift
1583 * down by 3. But (n+8) >> 3 == (n >> 3) + 1
1584 * (modulo overflows, which we do not have).
1586 * Finally, we LONG_ALIGN because all bitmap
1587 * operations are on longs.
1589 bitmap_size = (size-1) >> (i+4);
1590 bitmap_size = LONG_ALIGN(bitmap_size+1);
1591 zone->free_area[i].map =
1592 (unsigned long *) alloc_bootmem_node(pgdat, bitmap_size);
1597 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
1598 struct page *node_mem_map, unsigned long *zones_size,
1599 unsigned long node_start_pfn, unsigned long *zholes_size)
1603 pgdat->node_id = nid;
1604 pgdat->node_start_pfn = node_start_pfn;
1605 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
1606 if (!node_mem_map) {
1607 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
1608 node_mem_map = alloc_bootmem_node(pgdat, size);
1610 pgdat->node_mem_map = node_mem_map;
1612 free_area_init_core(pgdat, zones_size, zholes_size);
1615 #ifndef CONFIG_DISCONTIGMEM
1616 static bootmem_data_t contig_bootmem_data;
1617 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
1619 EXPORT_SYMBOL(contig_page_data);
1621 void __init free_area_init(unsigned long *zones_size)
1623 free_area_init_node(0, &contig_page_data, NULL, zones_size,
1624 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
1625 mem_map = contig_page_data.node_mem_map;
1629 #ifdef CONFIG_PROC_FS
1631 #include <linux/seq_file.h>
1633 static void *frag_start(struct seq_file *m, loff_t *pos)
1638 for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
1644 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1646 pg_data_t *pgdat = (pg_data_t *)arg;
1649 return pgdat->pgdat_next;
1652 static void frag_stop(struct seq_file *m, void *arg)
1657 * This walks the freelist for each zone. Whilst this is slow, I'd rather
1658 * be slow here than slow down the fast path by keeping stats - mjbligh
1660 static int frag_show(struct seq_file *m, void *arg)
1662 pg_data_t *pgdat = (pg_data_t *)arg;
1664 struct zone *node_zones = pgdat->node_zones;
1665 unsigned long flags;
1668 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1669 if (!zone->present_pages)
1672 spin_lock_irqsave(&zone->lock, flags);
1673 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1674 for (order = 0; order < MAX_ORDER; ++order) {
1675 unsigned long nr_bufs = 0;
1676 struct list_head *elem;
1678 list_for_each(elem, &(zone->free_area[order].free_list))
1680 seq_printf(m, "%6lu ", nr_bufs);
1682 spin_unlock_irqrestore(&zone->lock, flags);
1688 struct seq_operations fragmentation_op = {
1689 .start = frag_start,
1695 static char *vmstat_text[] = {
1699 "nr_page_table_pages",
1724 "pgscan_kswapd_high",
1725 "pgscan_kswapd_normal",
1727 "pgscan_kswapd_dma",
1728 "pgscan_direct_high",
1729 "pgscan_direct_normal",
1730 "pgscan_direct_dma",
1735 "kswapd_inodesteal",
1742 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1744 struct page_state *ps;
1746 if (*pos >= ARRAY_SIZE(vmstat_text))
1749 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
1752 return ERR_PTR(-ENOMEM);
1753 get_full_page_state(ps);
1754 ps->pgpgin /= 2; /* sectors -> kbytes */
1756 return (unsigned long *)ps + *pos;
1759 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1762 if (*pos >= ARRAY_SIZE(vmstat_text))
1764 return (unsigned long *)m->private + *pos;
1767 static int vmstat_show(struct seq_file *m, void *arg)
1769 unsigned long *l = arg;
1770 unsigned long off = l - (unsigned long *)m->private;
1772 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1776 static void vmstat_stop(struct seq_file *m, void *arg)
1782 struct seq_operations vmstat_op = {
1783 .start = vmstat_start,
1784 .next = vmstat_next,
1785 .stop = vmstat_stop,
1786 .show = vmstat_show,
1789 #endif /* CONFIG_PROC_FS */
1791 #ifdef CONFIG_HOTPLUG_CPU
1792 static int page_alloc_cpu_notify(struct notifier_block *self,
1793 unsigned long action, void *hcpu)
1795 int cpu = (unsigned long)hcpu;
1798 if (action == CPU_DEAD) {
1799 /* Drain local pagecache count. */
1800 count = &per_cpu(nr_pagecache_local, cpu);
1801 atomic_add(*count, &nr_pagecache);
1803 local_irq_disable();
1809 #endif /* CONFIG_HOTPLUG_CPU */
1811 void __init page_alloc_init(void)
1813 hotcpu_notifier(page_alloc_cpu_notify, 0);
1816 static unsigned long higherzone_val(struct zone *z, int max_zone,
1819 int z_idx = zone_idx(z);
1820 struct zone *higherzone;
1821 unsigned long pages;
1823 /* there is no higher zone to get a contribution from */
1824 if (z_idx == MAX_NR_ZONES-1)
1827 higherzone = &z->zone_pgdat->node_zones[z_idx+1];
1829 /* We always start with the higher zone's protection value */
1830 pages = higherzone->protection[alloc_type];
1833 * We get a lower-zone-protection contribution only if there are
1834 * pages in the higher zone and if we're not the highest zone
1835 * in the current zonelist. e.g., never happens for GFP_DMA. Happens
1836 * only for ZONE_DMA in a GFP_KERNEL allocation and happens for ZONE_DMA
1837 * and ZONE_NORMAL for a GFP_HIGHMEM allocation.
1839 if (higherzone->present_pages && z_idx < alloc_type)
1840 pages += higherzone->pages_low * sysctl_lower_zone_protection;
1846 * setup_per_zone_protection - called whenver min_free_kbytes or
1847 * sysctl_lower_zone_protection changes. Ensures that each zone
1848 * has a correct pages_protected value, so an adequate number of
1849 * pages are left in the zone after a successful __alloc_pages().
1851 * This algorithm is way confusing. I tries to keep the same behavior
1852 * as we had with the incremental min iterative algorithm.
1854 static void setup_per_zone_protection(void)
1856 struct pglist_data *pgdat;
1857 struct zone *zones, *zone;
1861 for_each_pgdat(pgdat) {
1862 zones = pgdat->node_zones;
1864 for (i = 0, max_zone = 0; i < MAX_NR_ZONES; i++)
1865 if (zones[i].present_pages)
1869 * For each of the different allocation types:
1870 * GFP_DMA -> GFP_KERNEL -> GFP_HIGHMEM
1872 for (i = 0; i < MAX_NR_ZONES; i++) {
1874 * For each of the zones:
1875 * ZONE_HIGHMEM -> ZONE_NORMAL -> ZONE_DMA
1877 for (j = MAX_NR_ZONES-1; j >= 0; j--) {
1881 * We never protect zones that don't have memory
1882 * in them (j>max_zone) or zones that aren't in
1883 * the zonelists for a certain type of
1884 * allocation (j>i). We have to assign these to
1885 * zero because the lower zones take
1886 * contributions from the higher zones.
1888 if (j > max_zone || j > i) {
1889 zone->protection[i] = 0;
1893 * The contribution of the next higher zone
1895 zone->protection[i] = higherzone_val(zone,
1897 zone->protection[i] += zone->pages_low;
1904 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
1905 * that the pages_{min,low,high} values for each zone are set correctly
1906 * with respect to min_free_kbytes.
1908 static void setup_per_zone_pages_min(void)
1910 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
1911 unsigned long lowmem_pages = 0;
1913 unsigned long flags;
1915 /* Calculate total number of !ZONE_HIGHMEM pages */
1916 for_each_zone(zone) {
1917 if (!is_highmem(zone))
1918 lowmem_pages += zone->present_pages;
1921 for_each_zone(zone) {
1922 spin_lock_irqsave(&zone->lru_lock, flags);
1923 if (is_highmem(zone)) {
1925 * Often, highmem doesn't need to reserve any pages.
1926 * But the pages_min/low/high values are also used for
1927 * batching up page reclaim activity so we need a
1928 * decent value here.
1932 min_pages = zone->present_pages / 1024;
1933 if (min_pages < SWAP_CLUSTER_MAX)
1934 min_pages = SWAP_CLUSTER_MAX;
1935 if (min_pages > 128)
1937 zone->pages_min = min_pages;
1939 /* if it's a lowmem zone, reserve a number of pages
1940 * proportionate to the zone's size.
1942 zone->pages_min = (pages_min * zone->present_pages) /
1946 zone->pages_low = zone->pages_min * 2;
1947 zone->pages_high = zone->pages_min * 3;
1948 spin_unlock_irqrestore(&zone->lru_lock, flags);
1953 * Initialise min_free_kbytes.
1955 * For small machines we want it small (128k min). For large machines
1956 * we want it large (16MB max). But it is not linear, because network
1957 * bandwidth does not increase linearly with machine size. We use
1959 * min_free_kbytes = sqrt(lowmem_kbytes)
1975 static int __init init_per_zone_pages_min(void)
1977 unsigned long lowmem_kbytes;
1979 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
1981 min_free_kbytes = int_sqrt(lowmem_kbytes);
1982 if (min_free_kbytes < 128)
1983 min_free_kbytes = 128;
1984 if (min_free_kbytes > 16384)
1985 min_free_kbytes = 16384;
1986 setup_per_zone_pages_min();
1987 setup_per_zone_protection();
1990 module_init(init_per_zone_pages_min)
1993 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
1994 * that we can call two helper functions whenever min_free_kbytes
1997 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
1998 struct file *file, void __user *buffer, size_t *length)
2000 proc_dointvec(table, write, file, buffer, length);
2001 setup_per_zone_pages_min();
2002 setup_per_zone_protection();
2007 * lower_zone_protection_sysctl_handler - just a wrapper around
2008 * proc_dointvec() so that we can call setup_per_zone_protection()
2009 * whenever sysctl_lower_zone_protection changes.
2011 int lower_zone_protection_sysctl_handler(ctl_table *table, int write,
2012 struct file *file, void __user *buffer, size_t *length)
2014 proc_dointvec_minmax(table, write, file, buffer, length);
2015 setup_per_zone_protection();