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;
60 static unsigned long __initdata nr_kernel_pages;
61 static unsigned long __initdata nr_all_pages;
64 * Temporary debugging check for pages not lying within a given zone.
66 static int bad_range(struct zone *zone, struct page *page)
68 if (page_to_pfn(page) >= zone->zone_start_pfn + zone->spanned_pages)
70 if (page_to_pfn(page) < zone->zone_start_pfn)
72 if (zone != page_zone(page))
77 static void bad_page(const char *function, struct page *page)
79 printk(KERN_EMERG "Bad page state at %s (in process '%s', page %p)\n",
80 function, current->comm, page);
81 printk(KERN_EMERG "flags:0x%08lx mapping:%p mapcount:%d count:%d\n",
82 (unsigned long)page->flags, page->mapping,
83 (int)page->mapcount, page_count(page));
84 printk(KERN_EMERG "Backtrace:\n");
86 printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n");
87 page->flags &= ~(1 << PG_private |
96 set_page_count(page, 0);
101 #ifndef CONFIG_HUGETLB_PAGE
102 #define prep_compound_page(page, order) do { } while (0)
103 #define destroy_compound_page(page, order) do { } while (0)
106 * Higher-order pages are called "compound pages". They are structured thusly:
108 * The first PAGE_SIZE page is called the "head page".
110 * The remaining PAGE_SIZE pages are called "tail pages".
112 * All pages have PG_compound set. All pages have their ->private pointing at
113 * the head page (even the head page has this).
115 * The first tail page's ->mapping, if non-zero, holds the address of the
116 * compound page's put_page() function.
118 * The order of the allocation is stored in the first tail page's ->index
119 * This is only for debug at present. This usage means that zero-order pages
120 * may not be compound.
122 static void prep_compound_page(struct page *page, unsigned long order)
125 int nr_pages = 1 << order;
127 page[1].mapping = NULL;
128 page[1].index = order;
129 for (i = 0; i < nr_pages; i++) {
130 struct page *p = page + i;
133 p->private = (unsigned long)page;
137 static void destroy_compound_page(struct page *page, unsigned long order)
140 int nr_pages = 1 << order;
142 if (!PageCompound(page))
145 if (page[1].index != order)
146 bad_page(__FUNCTION__, page);
148 for (i = 0; i < nr_pages; i++) {
149 struct page *p = page + i;
151 if (!PageCompound(p))
152 bad_page(__FUNCTION__, page);
153 if (p->private != (unsigned long)page)
154 bad_page(__FUNCTION__, page);
155 ClearPageCompound(p);
158 #endif /* CONFIG_HUGETLB_PAGE */
161 * Freeing function for a buddy system allocator.
163 * The concept of a buddy system is to maintain direct-mapped table
164 * (containing bit values) for memory blocks of various "orders".
165 * The bottom level table contains the map for the smallest allocatable
166 * units of memory (here, pages), and each level above it describes
167 * pairs of units from the levels below, hence, "buddies".
168 * At a high level, all that happens here is marking the table entry
169 * at the bottom level available, and propagating the changes upward
170 * as necessary, plus some accounting needed to play nicely with other
171 * parts of the VM system.
172 * At each level, we keep one bit for each pair of blocks, which
173 * is set to 1 iff only one of the pair is allocated. So when we
174 * are allocating or freeing one, we can derive the state of the
175 * other. That is, if we allocate a small block, and both were
176 * free, the remainder of the region must be split into blocks.
177 * If a block is freed, and its buddy is also free, then this
178 * triggers coalescing into a block of larger size.
183 static inline void __free_pages_bulk (struct page *page, struct page *base,
184 struct zone *zone, struct free_area *area, unsigned int order)
186 unsigned long page_idx, index, mask;
189 destroy_compound_page(page, order);
190 mask = (~0UL) << order;
191 page_idx = page - base;
192 if (page_idx & ~mask)
194 index = page_idx >> (1 + order);
196 zone->free_pages += 1 << order;
197 while (order < MAX_ORDER-1) {
198 struct page *buddy1, *buddy2;
200 BUG_ON(area >= zone->free_area + MAX_ORDER);
201 if (!__test_and_change_bit(index, area->map))
203 * the buddy page is still allocated.
207 /* Move the buddy up one level. */
208 buddy1 = base + (page_idx ^ (1 << order));
209 buddy2 = base + page_idx;
210 BUG_ON(bad_range(zone, buddy1));
211 BUG_ON(bad_range(zone, buddy2));
212 list_del(&buddy1->lru);
219 list_add(&(base + page_idx)->lru, &area->free_list);
222 static inline void free_pages_check(const char *function, struct page *page)
224 if ( page_mapped(page) ||
225 page->mapping != NULL ||
226 page_count(page) != 0 ||
237 1 << PG_writeback )))
238 bad_page(function, page);
240 ClearPageDirty(page);
244 * Frees a list of pages.
245 * Assumes all pages on list are in same zone, and of same order.
246 * count is the number of pages to free, or 0 for all on the list.
248 * If the zone was previously in an "all pages pinned" state then look to
249 * see if this freeing clears that state.
251 * And clear the zone's pages_scanned counter, to hold off the "all pages are
252 * pinned" detection logic.
255 free_pages_bulk(struct zone *zone, int count,
256 struct list_head *list, unsigned int order)
259 struct free_area *area;
260 struct page *base, *page = NULL;
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, order);
275 spin_unlock_irqrestore(&zone->lock, flags);
279 void __free_pages_ok(struct page *page, unsigned int order)
284 arch_free_page(page, order);
286 mod_page_state(pgfree, 1 << order);
287 for (i = 0 ; i < (1 << order) ; ++i)
288 free_pages_check(__FUNCTION__, page + i);
289 list_add(&page->lru, &list);
290 kernel_map_pages(page, 1<<order, 0);
291 free_pages_bulk(page_zone(page), 1, &list, order);
294 #define MARK_USED(index, order, area) \
295 __change_bit((index) >> (1+(order)), (area)->map)
298 * The order of subdivision here is critical for the IO subsystem.
299 * Please do not alter this order without good reasons and regression
300 * testing. Specifically, as large blocks of memory are subdivided,
301 * the order in which smaller blocks are delivered depends on the order
302 * they're subdivided in this function. This is the primary factor
303 * influencing the order in which pages are delivered to the IO
304 * subsystem according to empirical testing, and this is also justified
305 * by considering the behavior of a buddy system containing a single
306 * large block of memory acted on by a series of small allocations.
307 * This behavior is a critical factor in sglist merging's success.
311 static inline struct page *
312 expand(struct zone *zone, struct page *page,
313 unsigned long index, int low, int high, struct free_area *area)
315 unsigned long size = 1 << high;
321 BUG_ON(bad_range(zone, &page[size]));
322 list_add(&page[size].lru, &area->free_list);
323 MARK_USED(index + size, high, area);
328 static inline void set_page_refs(struct page *page, int order)
331 set_page_count(page, 1);
336 * We need to reference all the pages for this order, otherwise if
337 * anyone accesses one of the pages with (get/put) it will be freed.
339 for (i = 0; i < (1 << order); i++)
340 set_page_count(page+i, 1);
341 #endif /* CONFIG_MMU */
345 * This page is about to be returned from the page allocator
347 static void prep_new_page(struct page *page, int order)
349 if (page->mapping || page_mapped(page) ||
360 1 << PG_writeback )))
361 bad_page(__FUNCTION__, page);
363 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
364 1 << PG_referenced | 1 << PG_arch_1 |
365 1 << PG_checked | 1 << PG_mappedtodisk);
367 set_page_refs(page, order);
371 * Do the hard work of removing an element from the buddy allocator.
372 * Call me with the zone->lock already held.
374 static struct page *__rmqueue(struct zone *zone, unsigned int order)
376 struct free_area * area;
377 unsigned int current_order;
381 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
382 area = zone->free_area + current_order;
383 if (list_empty(&area->free_list))
386 page = list_entry(area->free_list.next, struct page, lru);
387 list_del(&page->lru);
388 index = page - zone->zone_mem_map;
389 if (current_order != MAX_ORDER-1)
390 MARK_USED(index, current_order, area);
391 zone->free_pages -= 1UL << order;
392 return expand(zone, page, index, order, current_order, area);
399 * Obtain a specified number of elements from the buddy allocator, all under
400 * a single hold of the lock, for efficiency. Add them to the supplied list.
401 * Returns the number of new pages which were placed at *list.
403 static int rmqueue_bulk(struct zone *zone, unsigned int order,
404 unsigned long count, struct list_head *list)
411 spin_lock_irqsave(&zone->lock, flags);
412 for (i = 0; i < count; ++i) {
413 page = __rmqueue(zone, order);
417 list_add_tail(&page->lru, list);
419 spin_unlock_irqrestore(&zone->lock, flags);
423 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
424 static void __drain_pages(unsigned int cpu)
429 for_each_zone(zone) {
430 struct per_cpu_pageset *pset;
432 pset = &zone->pageset[cpu];
433 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
434 struct per_cpu_pages *pcp;
437 pcp->count -= free_pages_bulk(zone, pcp->count,
442 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
445 int is_head_of_free_region(struct page *page)
447 struct zone *zone = page_zone(page);
450 struct list_head *curr;
453 * Should not matter as we need quiescent system for
454 * suspend anyway, but...
456 spin_lock_irqsave(&zone->lock, flags);
457 for (order = MAX_ORDER - 1; order >= 0; --order)
458 list_for_each(curr, &zone->free_area[order].free_list)
459 if (page == list_entry(curr, struct page, lru)) {
460 spin_unlock_irqrestore(&zone->lock, flags);
463 spin_unlock_irqrestore(&zone->lock, flags);
468 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
470 void drain_local_pages(void)
474 local_irq_save(flags);
475 __drain_pages(smp_processor_id());
476 local_irq_restore(flags);
478 #endif /* CONFIG_PM */
480 static void zone_statistics(struct zonelist *zonelist, struct zone *z)
485 pg_data_t *pg = z->zone_pgdat;
486 pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
487 struct per_cpu_pageset *p;
489 local_irq_save(flags);
490 cpu = smp_processor_id();
491 p = &z->pageset[cpu];
493 z->pageset[cpu].numa_hit++;
496 zonelist->zones[0]->pageset[cpu].numa_foreign++;
498 if (pg == NODE_DATA(numa_node_id()))
502 local_irq_restore(flags);
507 * Free a 0-order page
509 static void FASTCALL(free_hot_cold_page(struct page *page, int cold));
510 static void fastcall free_hot_cold_page(struct page *page, int cold)
512 struct zone *zone = page_zone(page);
513 struct per_cpu_pages *pcp;
516 arch_free_page(page, 0);
518 kernel_map_pages(page, 1, 0);
519 inc_page_state(pgfree);
520 free_pages_check(__FUNCTION__, page);
521 pcp = &zone->pageset[get_cpu()].pcp[cold];
522 local_irq_save(flags);
523 if (pcp->count >= pcp->high)
524 pcp->count -= free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
525 list_add(&page->lru, &pcp->list);
527 local_irq_restore(flags);
531 void fastcall free_hot_page(struct page *page)
533 free_hot_cold_page(page, 0);
536 void fastcall free_cold_page(struct page *page)
538 free_hot_cold_page(page, 1);
542 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
543 * we cheat by calling it from here, in the order > 0 path. Saves a branch
548 buffered_rmqueue(struct zone *zone, int order, int gfp_flags)
551 struct page *page = NULL;
552 int cold = !!(gfp_flags & __GFP_COLD);
555 struct per_cpu_pages *pcp;
557 pcp = &zone->pageset[get_cpu()].pcp[cold];
558 local_irq_save(flags);
559 if (pcp->count <= pcp->low)
560 pcp->count += rmqueue_bulk(zone, 0,
561 pcp->batch, &pcp->list);
563 page = list_entry(pcp->list.next, struct page, lru);
564 list_del(&page->lru);
567 local_irq_restore(flags);
572 spin_lock_irqsave(&zone->lock, flags);
573 page = __rmqueue(zone, order);
574 spin_unlock_irqrestore(&zone->lock, flags);
578 BUG_ON(bad_range(zone, page));
579 mod_page_state_zone(zone, pgalloc, 1 << order);
580 prep_new_page(page, order);
581 if (order && (gfp_flags & __GFP_COMP))
582 prep_compound_page(page, order);
588 * This is the 'heart' of the zoned buddy allocator.
590 * Herein lies the mysterious "incremental min". That's the
592 * local_low = z->pages_low;
595 * thing. The intent here is to provide additional protection to low zones for
596 * allocation requests which _could_ use higher zones. So a GFP_HIGHMEM
597 * request is not allowed to dip as deeply into the normal zone as a GFP_KERNEL
598 * request. This preserves additional space in those lower zones for requests
599 * which really do need memory from those zones. It means that on a decent
600 * sized machine, GFP_HIGHMEM and GFP_KERNEL requests basically leave the DMA
603 struct page * fastcall
604 __alloc_pages(unsigned int gfp_mask, unsigned int order,
605 struct zonelist *zonelist)
607 const int wait = gfp_mask & __GFP_WAIT;
611 struct reclaim_state reclaim_state;
612 struct task_struct *p = current;
617 might_sleep_if(wait);
619 zones = zonelist->zones; /* the list of zones suitable for gfp_mask */
620 if (zones[0] == NULL) /* no zones in the zonelist */
623 alloc_type = zone_idx(zones[0]);
625 /* Go through the zonelist once, looking for a zone with enough free */
626 for (i = 0; zones[i] != NULL; i++) {
627 struct zone *z = zones[i];
629 min = (1<<order) + z->protection[alloc_type];
632 * We let real-time tasks dip their real-time paws a little
633 * deeper into reserves.
636 min -= z->pages_low >> 1;
638 if (z->free_pages >= min ||
639 (!wait && z->free_pages >= z->pages_high)) {
640 page = buffered_rmqueue(z, order, gfp_mask);
642 zone_statistics(zonelist, z);
648 /* we're somewhat low on memory, failed to find what we needed */
649 for (i = 0; zones[i] != NULL; i++)
650 wakeup_kswapd(zones[i]);
652 /* Go through the zonelist again, taking __GFP_HIGH into account */
653 for (i = 0; zones[i] != NULL; i++) {
654 struct zone *z = zones[i];
656 min = (1<<order) + z->protection[alloc_type];
658 if (gfp_mask & __GFP_HIGH)
659 min -= z->pages_low >> 2;
661 min -= z->pages_low >> 1;
663 if (z->free_pages >= min ||
664 (!wait && z->free_pages >= z->pages_high)) {
665 page = buffered_rmqueue(z, order, gfp_mask);
667 zone_statistics(zonelist, z);
673 /* here we're in the low on memory slow path */
676 if ((p->flags & (PF_MEMALLOC | PF_MEMDIE)) && !in_interrupt()) {
677 /* go through the zonelist yet again, ignoring mins */
678 for (i = 0; zones[i] != NULL; i++) {
679 struct zone *z = zones[i];
681 page = buffered_rmqueue(z, order, gfp_mask);
683 zone_statistics(zonelist, z);
690 /* Atomic allocations - we can't balance anything */
694 p->flags |= PF_MEMALLOC;
695 reclaim_state.reclaimed_slab = 0;
696 p->reclaim_state = &reclaim_state;
698 try_to_free_pages(zones, gfp_mask, order);
700 p->reclaim_state = NULL;
701 p->flags &= ~PF_MEMALLOC;
703 /* go through the zonelist yet one more time */
704 for (i = 0; zones[i] != NULL; i++) {
705 struct zone *z = zones[i];
707 min = (1UL << order) + z->protection[alloc_type];
709 if (z->free_pages >= min ||
710 (!wait && z->free_pages >= z->pages_high)) {
711 page = buffered_rmqueue(z, order, gfp_mask);
713 zone_statistics(zonelist, z);
720 * Don't let big-order allocations loop unless the caller explicitly
721 * requests that. Wait for some write requests to complete then retry.
723 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL, but that
724 * may not be true in other implementations.
727 if (!(gfp_mask & __GFP_NORETRY)) {
728 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
730 if (gfp_mask & __GFP_NOFAIL)
734 blk_congestion_wait(WRITE, HZ/50);
739 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
740 printk(KERN_WARNING "%s: page allocation failure."
741 " order:%d, mode:0x%x\n",
742 p->comm, order, gfp_mask);
747 kernel_map_pages(page, 1 << order, 1);
751 EXPORT_SYMBOL(__alloc_pages);
754 * Common helper functions.
756 fastcall unsigned long __get_free_pages(unsigned int gfp_mask, unsigned int order)
759 page = alloc_pages(gfp_mask, order);
762 return (unsigned long) page_address(page);
765 EXPORT_SYMBOL(__get_free_pages);
767 fastcall unsigned long get_zeroed_page(unsigned int gfp_mask)
772 * get_zeroed_page() returns a 32-bit address, which cannot represent
775 BUG_ON(gfp_mask & __GFP_HIGHMEM);
777 page = alloc_pages(gfp_mask, 0);
779 void *address = page_address(page);
781 return (unsigned long) address;
786 EXPORT_SYMBOL(get_zeroed_page);
788 void __pagevec_free(struct pagevec *pvec)
790 int i = pagevec_count(pvec);
793 free_hot_cold_page(pvec->pages[i], pvec->cold);
796 fastcall void __free_pages(struct page *page, unsigned int order)
798 if (!PageReserved(page) && put_page_testzero(page)) {
802 __free_pages_ok(page, order);
806 EXPORT_SYMBOL(__free_pages);
808 fastcall void free_pages(unsigned long addr, unsigned int order)
811 BUG_ON(!virt_addr_valid(addr));
812 __free_pages(virt_to_page(addr), order);
816 EXPORT_SYMBOL(free_pages);
819 * Total amount of free (allocatable) RAM:
821 unsigned int nr_free_pages(void)
823 unsigned int sum = 0;
827 sum += zone->free_pages;
832 EXPORT_SYMBOL(nr_free_pages);
834 unsigned int nr_used_zone_pages(void)
836 unsigned int pages = 0;
840 pages += zone->nr_active + zone->nr_inactive;
846 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
848 unsigned int i, sum = 0;
850 for (i = 0; i < MAX_NR_ZONES; i++)
851 sum += pgdat->node_zones[i].free_pages;
857 static unsigned int nr_free_zone_pages(int offset)
860 unsigned int sum = 0;
862 for_each_pgdat(pgdat) {
863 struct zonelist *zonelist = pgdat->node_zonelists + offset;
864 struct zone **zonep = zonelist->zones;
867 for (zone = *zonep++; zone; zone = *zonep++) {
868 unsigned long size = zone->present_pages;
869 unsigned long high = zone->pages_high;
879 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
881 unsigned int nr_free_buffer_pages(void)
883 return nr_free_zone_pages(GFP_USER & GFP_ZONEMASK);
887 * Amount of free RAM allocatable within all zones
889 unsigned int nr_free_pagecache_pages(void)
891 return nr_free_zone_pages(GFP_HIGHUSER & GFP_ZONEMASK);
894 #ifdef CONFIG_HIGHMEM
895 unsigned int nr_free_highpages (void)
898 unsigned int pages = 0;
900 for_each_pgdat(pgdat)
901 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
908 static void show_node(struct zone *zone)
910 printk("Node %d ", zone->zone_pgdat->node_id);
913 #define show_node(zone) do { } while (0)
917 * Accumulate the page_state information across all CPUs.
918 * The result is unavoidably approximate - it can change
919 * during and after execution of this function.
921 DEFINE_PER_CPU(struct page_state, page_states) = {0};
922 EXPORT_PER_CPU_SYMBOL(page_states);
924 atomic_t nr_pagecache = ATOMIC_INIT(0);
925 EXPORT_SYMBOL(nr_pagecache);
927 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
930 void __get_page_state(struct page_state *ret, int nr)
934 memset(ret, 0, sizeof(*ret));
935 while (cpu < NR_CPUS) {
936 unsigned long *in, *out, off;
938 if (!cpu_possible(cpu)) {
943 in = (unsigned long *)&per_cpu(page_states, cpu);
945 if (cpu < NR_CPUS && cpu_possible(cpu))
946 prefetch(&per_cpu(page_states, cpu));
947 out = (unsigned long *)ret;
948 for (off = 0; off < nr; off++)
953 void get_page_state(struct page_state *ret)
957 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
958 nr /= sizeof(unsigned long);
960 __get_page_state(ret, nr + 1);
963 void get_full_page_state(struct page_state *ret)
965 __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long));
968 unsigned long __read_page_state(unsigned offset)
970 unsigned long ret = 0;
973 for (cpu = 0; cpu < NR_CPUS; cpu++) {
976 if (!cpu_possible(cpu))
979 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
980 ret += *((unsigned long *)in);
985 void get_zone_counts(unsigned long *active,
986 unsigned long *inactive, unsigned long *free)
993 for_each_zone(zone) {
994 *active += zone->nr_active;
995 *inactive += zone->nr_inactive;
996 *free += zone->free_pages;
1000 void si_meminfo(struct sysinfo *val)
1002 val->totalram = totalram_pages;
1004 val->freeram = nr_free_pages();
1005 val->bufferram = nr_blockdev_pages();
1006 #ifdef CONFIG_HIGHMEM
1007 val->totalhigh = totalhigh_pages;
1008 val->freehigh = nr_free_highpages();
1013 val->mem_unit = PAGE_SIZE;
1014 if (vx_flags(VXF_VIRT_MEM, 0))
1015 vx_vsi_meminfo(val);
1018 EXPORT_SYMBOL(si_meminfo);
1021 void si_meminfo_node(struct sysinfo *val, int nid)
1023 pg_data_t *pgdat = NODE_DATA(nid);
1025 val->totalram = pgdat->node_present_pages;
1026 val->freeram = nr_free_pages_pgdat(pgdat);
1027 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1028 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1029 val->mem_unit = PAGE_SIZE;
1033 #define K(x) ((x) << (PAGE_SHIFT-10))
1036 * Show free area list (used inside shift_scroll-lock stuff)
1037 * We also calculate the percentage fragmentation. We do this by counting the
1038 * memory on each free list with the exception of the first item on the list.
1040 void show_free_areas(void)
1042 struct page_state ps;
1043 int cpu, temperature;
1044 unsigned long active;
1045 unsigned long inactive;
1049 for_each_zone(zone) {
1051 printk("%s per-cpu:", zone->name);
1053 if (!zone->present_pages) {
1059 for (cpu = 0; cpu < NR_CPUS; ++cpu) {
1060 struct per_cpu_pageset *pageset;
1062 if (!cpu_possible(cpu))
1065 pageset = zone->pageset + cpu;
1067 for (temperature = 0; temperature < 2; temperature++)
1068 printk("cpu %d %s: low %d, high %d, batch %d\n",
1070 temperature ? "cold" : "hot",
1071 pageset->pcp[temperature].low,
1072 pageset->pcp[temperature].high,
1073 pageset->pcp[temperature].batch);
1077 get_page_state(&ps);
1078 get_zone_counts(&active, &inactive, &free);
1080 printk("\nFree pages: %11ukB (%ukB HighMem)\n",
1082 K(nr_free_highpages()));
1084 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1085 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1094 ps.nr_page_table_pages);
1096 for_each_zone(zone) {
1110 K(zone->free_pages),
1113 K(zone->pages_high),
1115 K(zone->nr_inactive),
1116 K(zone->present_pages)
1118 printk("protections[]:");
1119 for (i = 0; i < MAX_NR_ZONES; i++)
1120 printk(" %lu", zone->protection[i]);
1124 for_each_zone(zone) {
1125 struct list_head *elem;
1126 unsigned long nr, flags, order, total = 0;
1129 printk("%s: ", zone->name);
1130 if (!zone->present_pages) {
1135 spin_lock_irqsave(&zone->lock, flags);
1136 for (order = 0; order < MAX_ORDER; order++) {
1138 list_for_each(elem, &zone->free_area[order].free_list)
1140 total += nr << order;
1141 printk("%lu*%lukB ", nr, K(1UL) << order);
1143 spin_unlock_irqrestore(&zone->lock, flags);
1144 printk("= %lukB\n", K(total));
1147 show_swap_cache_info();
1151 * Builds allocation fallback zone lists.
1153 static int __init build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, int j, int k)
1160 zone = pgdat->node_zones + ZONE_HIGHMEM;
1161 if (zone->present_pages) {
1162 #ifndef CONFIG_HIGHMEM
1165 zonelist->zones[j++] = zone;
1168 zone = pgdat->node_zones + ZONE_NORMAL;
1169 if (zone->present_pages)
1170 zonelist->zones[j++] = zone;
1172 zone = pgdat->node_zones + ZONE_DMA;
1173 if (zone->present_pages)
1174 zonelist->zones[j++] = zone;
1181 #define MAX_NODE_LOAD (numnodes)
1182 static int __initdata node_load[MAX_NUMNODES];
1184 * find_next_best_node - find the next node that should appear in a given
1185 * node's fallback list
1186 * @node: node whose fallback list we're appending
1187 * @used_node_mask: pointer to the bitmap of already used nodes
1189 * We use a number of factors to determine which is the next node that should
1190 * appear on a given node's fallback list. The node should not have appeared
1191 * already in @node's fallback list, and it should be the next closest node
1192 * according to the distance array (which contains arbitrary distance values
1193 * from each node to each node in the system), and should also prefer nodes
1194 * with no CPUs, since presumably they'll have very little allocation pressure
1195 * on them otherwise.
1196 * It returns -1 if no node is found.
1198 static int __init find_next_best_node(int node, void *used_node_mask)
1201 int min_val = INT_MAX;
1204 for (i = 0; i < numnodes; i++) {
1207 /* Start from local node */
1208 n = (node+i)%numnodes;
1210 /* Don't want a node to appear more than once */
1211 if (test_bit(n, used_node_mask))
1214 /* Use the distance array to find the distance */
1215 val = node_distance(node, n);
1217 /* Give preference to headless and unused nodes */
1218 tmp = node_to_cpumask(n);
1219 if (!cpus_empty(tmp))
1220 val += PENALTY_FOR_NODE_WITH_CPUS;
1222 /* Slight preference for less loaded node */
1223 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1224 val += node_load[n];
1226 if (val < min_val) {
1233 set_bit(best_node, used_node_mask);
1238 static void __init build_zonelists(pg_data_t *pgdat)
1240 int i, j, k, node, local_node;
1241 int prev_node, load;
1242 struct zonelist *zonelist;
1243 DECLARE_BITMAP(used_mask, MAX_NUMNODES);
1245 /* initialize zonelists */
1246 for (i = 0; i < GFP_ZONETYPES; i++) {
1247 zonelist = pgdat->node_zonelists + i;
1248 memset(zonelist, 0, sizeof(*zonelist));
1249 zonelist->zones[0] = NULL;
1252 /* NUMA-aware ordering of nodes */
1253 local_node = pgdat->node_id;
1255 prev_node = local_node;
1256 bitmap_zero(used_mask, MAX_NUMNODES);
1257 while ((node = find_next_best_node(local_node, used_mask)) >= 0) {
1259 * We don't want to pressure a particular node.
1260 * So adding penalty to the first node in same
1261 * distance group to make it round-robin.
1263 if (node_distance(local_node, node) !=
1264 node_distance(local_node, prev_node))
1265 node_load[node] += load;
1268 for (i = 0; i < GFP_ZONETYPES; i++) {
1269 zonelist = pgdat->node_zonelists + i;
1270 for (j = 0; zonelist->zones[j] != NULL; j++);
1273 if (i & __GFP_HIGHMEM)
1278 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1279 zonelist->zones[j] = NULL;
1284 #else /* CONFIG_NUMA */
1286 static void __init build_zonelists(pg_data_t *pgdat)
1288 int i, j, k, node, local_node;
1290 local_node = pgdat->node_id;
1291 for (i = 0; i < GFP_ZONETYPES; i++) {
1292 struct zonelist *zonelist;
1294 zonelist = pgdat->node_zonelists + i;
1295 memset(zonelist, 0, sizeof(*zonelist));
1299 if (i & __GFP_HIGHMEM)
1304 j = build_zonelists_node(pgdat, zonelist, j, k);
1306 * Now we build the zonelist so that it contains the zones
1307 * of all the other nodes.
1308 * We don't want to pressure a particular node, so when
1309 * building the zones for node N, we make sure that the
1310 * zones coming right after the local ones are those from
1311 * node N+1 (modulo N)
1313 for (node = local_node + 1; node < numnodes; node++)
1314 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1315 for (node = 0; node < local_node; node++)
1316 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1318 zonelist->zones[j] = NULL;
1322 #endif /* CONFIG_NUMA */
1324 void __init build_all_zonelists(void)
1328 for(i = 0 ; i < numnodes ; i++)
1329 build_zonelists(NODE_DATA(i));
1330 printk("Built %i zonelists\n", numnodes);
1334 * Helper functions to size the waitqueue hash table.
1335 * Essentially these want to choose hash table sizes sufficiently
1336 * large so that collisions trying to wait on pages are rare.
1337 * But in fact, the number of active page waitqueues on typical
1338 * systems is ridiculously low, less than 200. So this is even
1339 * conservative, even though it seems large.
1341 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1342 * waitqueues, i.e. the size of the waitq table given the number of pages.
1344 #define PAGES_PER_WAITQUEUE 256
1346 static inline unsigned long wait_table_size(unsigned long pages)
1348 unsigned long size = 1;
1350 pages /= PAGES_PER_WAITQUEUE;
1352 while (size < pages)
1356 * Once we have dozens or even hundreds of threads sleeping
1357 * on IO we've got bigger problems than wait queue collision.
1358 * Limit the size of the wait table to a reasonable size.
1360 size = min(size, 4096UL);
1362 return max(size, 4UL);
1366 * This is an integer logarithm so that shifts can be used later
1367 * to extract the more random high bits from the multiplicative
1368 * hash function before the remainder is taken.
1370 static inline unsigned long wait_table_bits(unsigned long size)
1375 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1377 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1378 unsigned long *zones_size, unsigned long *zholes_size)
1380 unsigned long realtotalpages, totalpages = 0;
1383 for (i = 0; i < MAX_NR_ZONES; i++)
1384 totalpages += zones_size[i];
1385 pgdat->node_spanned_pages = totalpages;
1387 realtotalpages = totalpages;
1389 for (i = 0; i < MAX_NR_ZONES; i++)
1390 realtotalpages -= zholes_size[i];
1391 pgdat->node_present_pages = realtotalpages;
1392 printk("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1397 * Initially all pages are reserved - free ones are freed
1398 * up by free_all_bootmem() once the early boot process is
1399 * done. Non-atomic initialization, single-pass.
1401 void __init memmap_init_zone(struct page *start, unsigned long size, int nid,
1402 unsigned long zone, unsigned long start_pfn)
1406 for (page = start; page < (start + size); page++) {
1407 set_page_zone(page, NODEZONE(nid, zone));
1408 set_page_count(page, 0);
1409 SetPageReserved(page);
1410 INIT_LIST_HEAD(&page->lru);
1411 #ifdef WANT_PAGE_VIRTUAL
1412 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1413 if (!is_highmem(zone))
1414 set_page_address(page, __va(start_pfn << PAGE_SHIFT));
1420 #ifndef __HAVE_ARCH_MEMMAP_INIT
1421 #define memmap_init(start, size, nid, zone, start_pfn) \
1422 memmap_init_zone((start), (size), (nid), (zone), (start_pfn))
1426 * Set up the zone data structures:
1427 * - mark all pages reserved
1428 * - mark all memory queues empty
1429 * - clear the memory bitmaps
1431 static void __init free_area_init_core(struct pglist_data *pgdat,
1432 unsigned long *zones_size, unsigned long *zholes_size)
1435 const unsigned long zone_required_alignment = 1UL << (MAX_ORDER-1);
1436 int cpu, nid = pgdat->node_id;
1437 struct page *lmem_map = pgdat->node_mem_map;
1438 unsigned long zone_start_pfn = pgdat->node_start_pfn;
1440 pgdat->nr_zones = 0;
1441 init_waitqueue_head(&pgdat->kswapd_wait);
1443 for (j = 0; j < MAX_NR_ZONES; j++) {
1444 struct zone *zone = pgdat->node_zones + j;
1445 unsigned long size, realsize;
1446 unsigned long batch;
1448 zone_table[NODEZONE(nid, j)] = zone;
1449 realsize = size = zones_size[j];
1451 realsize -= zholes_size[j];
1453 if (j == ZONE_DMA || j == ZONE_NORMAL)
1454 nr_kernel_pages += realsize;
1455 nr_all_pages += realsize;
1457 zone->spanned_pages = size;
1458 zone->present_pages = realsize;
1459 zone->name = zone_names[j];
1460 spin_lock_init(&zone->lock);
1461 spin_lock_init(&zone->lru_lock);
1462 zone->zone_pgdat = pgdat;
1463 zone->free_pages = 0;
1465 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
1468 * The per-cpu-pages pools are set to around 1000th of the
1469 * size of the zone. But no more than 1/4 of a meg - there's
1470 * no point in going beyond the size of L2 cache.
1472 * OK, so we don't know how big the cache is. So guess.
1474 batch = zone->present_pages / 1024;
1475 if (batch * PAGE_SIZE > 256 * 1024)
1476 batch = (256 * 1024) / PAGE_SIZE;
1477 batch /= 4; /* We effectively *= 4 below */
1481 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1482 struct per_cpu_pages *pcp;
1484 pcp = &zone->pageset[cpu].pcp[0]; /* hot */
1486 pcp->low = 2 * batch;
1487 pcp->high = 6 * batch;
1488 pcp->batch = 1 * batch;
1489 INIT_LIST_HEAD(&pcp->list);
1491 pcp = &zone->pageset[cpu].pcp[1]; /* cold */
1494 pcp->high = 2 * batch;
1495 pcp->batch = 1 * batch;
1496 INIT_LIST_HEAD(&pcp->list);
1498 printk(" %s zone: %lu pages, LIFO batch:%lu\n",
1499 zone_names[j], realsize, batch);
1500 INIT_LIST_HEAD(&zone->active_list);
1501 INIT_LIST_HEAD(&zone->inactive_list);
1502 zone->nr_scan_active = 0;
1503 zone->nr_scan_inactive = 0;
1504 zone->nr_active = 0;
1505 zone->nr_inactive = 0;
1510 * The per-page waitqueue mechanism uses hashed waitqueues
1513 zone->wait_table_size = wait_table_size(size);
1514 zone->wait_table_bits =
1515 wait_table_bits(zone->wait_table_size);
1516 zone->wait_table = (wait_queue_head_t *)
1517 alloc_bootmem_node(pgdat, zone->wait_table_size
1518 * sizeof(wait_queue_head_t));
1520 for(i = 0; i < zone->wait_table_size; ++i)
1521 init_waitqueue_head(zone->wait_table + i);
1523 pgdat->nr_zones = j+1;
1525 zone->zone_mem_map = lmem_map;
1526 zone->zone_start_pfn = zone_start_pfn;
1528 if ((zone_start_pfn) & (zone_required_alignment-1))
1529 printk("BUG: wrong zone alignment, it will crash\n");
1531 memmap_init(lmem_map, size, nid, j, zone_start_pfn);
1533 zone_start_pfn += size;
1536 for (i = 0; ; i++) {
1537 unsigned long bitmap_size;
1539 INIT_LIST_HEAD(&zone->free_area[i].free_list);
1540 if (i == MAX_ORDER-1) {
1541 zone->free_area[i].map = NULL;
1546 * Page buddy system uses "index >> (i+1)",
1547 * where "index" is at most "size-1".
1549 * The extra "+3" is to round down to byte
1550 * size (8 bits per byte assumption). Thus
1551 * we get "(size-1) >> (i+4)" as the last byte
1554 * The "+1" is because we want to round the
1555 * byte allocation up rather than down. So
1556 * we should have had a "+7" before we shifted
1557 * down by three. Also, we have to add one as
1558 * we actually _use_ the last bit (it's [0,n]
1559 * inclusive, not [0,n[).
1561 * So we actually had +7+1 before we shift
1562 * down by 3. But (n+8) >> 3 == (n >> 3) + 1
1563 * (modulo overflows, which we do not have).
1565 * Finally, we LONG_ALIGN because all bitmap
1566 * operations are on longs.
1568 bitmap_size = (size-1) >> (i+4);
1569 bitmap_size = LONG_ALIGN(bitmap_size+1);
1570 zone->free_area[i].map =
1571 (unsigned long *) alloc_bootmem_node(pgdat, bitmap_size);
1576 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
1577 struct page *node_mem_map, unsigned long *zones_size,
1578 unsigned long node_start_pfn, unsigned long *zholes_size)
1582 pgdat->node_id = nid;
1583 pgdat->node_start_pfn = node_start_pfn;
1584 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
1585 if (!node_mem_map) {
1586 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
1587 node_mem_map = alloc_bootmem_node(pgdat, size);
1589 pgdat->node_mem_map = node_mem_map;
1591 free_area_init_core(pgdat, zones_size, zholes_size);
1594 #ifndef CONFIG_DISCONTIGMEM
1595 static bootmem_data_t contig_bootmem_data;
1596 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
1598 EXPORT_SYMBOL(contig_page_data);
1600 void __init free_area_init(unsigned long *zones_size)
1602 free_area_init_node(0, &contig_page_data, NULL, zones_size,
1603 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
1604 mem_map = contig_page_data.node_mem_map;
1608 #ifdef CONFIG_PROC_FS
1610 #include <linux/seq_file.h>
1612 static void *frag_start(struct seq_file *m, loff_t *pos)
1617 for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
1623 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1625 pg_data_t *pgdat = (pg_data_t *)arg;
1628 return pgdat->pgdat_next;
1631 static void frag_stop(struct seq_file *m, void *arg)
1636 * This walks the freelist for each zone. Whilst this is slow, I'd rather
1637 * be slow here than slow down the fast path by keeping stats - mjbligh
1639 static int frag_show(struct seq_file *m, void *arg)
1641 pg_data_t *pgdat = (pg_data_t *)arg;
1643 struct zone *node_zones = pgdat->node_zones;
1644 unsigned long flags;
1647 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1648 if (!zone->present_pages)
1651 spin_lock_irqsave(&zone->lock, flags);
1652 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1653 for (order = 0; order < MAX_ORDER; ++order) {
1654 unsigned long nr_bufs = 0;
1655 struct list_head *elem;
1657 list_for_each(elem, &(zone->free_area[order].free_list))
1659 seq_printf(m, "%6lu ", nr_bufs);
1661 spin_unlock_irqrestore(&zone->lock, flags);
1667 struct seq_operations fragmentation_op = {
1668 .start = frag_start,
1674 static char *vmstat_text[] = {
1678 "nr_page_table_pages",
1703 "pgscan_kswapd_high",
1704 "pgscan_kswapd_normal",
1706 "pgscan_kswapd_dma",
1707 "pgscan_direct_high",
1708 "pgscan_direct_normal",
1709 "pgscan_direct_dma",
1714 "kswapd_inodesteal",
1721 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1723 struct page_state *ps;
1725 if (*pos >= ARRAY_SIZE(vmstat_text))
1728 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
1731 return ERR_PTR(-ENOMEM);
1732 get_full_page_state(ps);
1733 ps->pgpgin /= 2; /* sectors -> kbytes */
1735 return (unsigned long *)ps + *pos;
1738 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1741 if (*pos >= ARRAY_SIZE(vmstat_text))
1743 return (unsigned long *)m->private + *pos;
1746 static int vmstat_show(struct seq_file *m, void *arg)
1748 unsigned long *l = arg;
1749 unsigned long off = l - (unsigned long *)m->private;
1751 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1755 static void vmstat_stop(struct seq_file *m, void *arg)
1761 struct seq_operations vmstat_op = {
1762 .start = vmstat_start,
1763 .next = vmstat_next,
1764 .stop = vmstat_stop,
1765 .show = vmstat_show,
1768 #endif /* CONFIG_PROC_FS */
1770 #ifdef CONFIG_HOTPLUG_CPU
1771 static int page_alloc_cpu_notify(struct notifier_block *self,
1772 unsigned long action, void *hcpu)
1774 int cpu = (unsigned long)hcpu;
1777 if (action == CPU_DEAD) {
1778 /* Drain local pagecache count. */
1779 count = &per_cpu(nr_pagecache_local, cpu);
1780 atomic_add(*count, &nr_pagecache);
1782 local_irq_disable();
1788 #endif /* CONFIG_HOTPLUG_CPU */
1790 void __init page_alloc_init(void)
1792 hotcpu_notifier(page_alloc_cpu_notify, 0);
1795 static unsigned long higherzone_val(struct zone *z, int max_zone,
1798 int z_idx = zone_idx(z);
1799 struct zone *higherzone;
1800 unsigned long pages;
1802 /* there is no higher zone to get a contribution from */
1803 if (z_idx == MAX_NR_ZONES-1)
1806 higherzone = &z->zone_pgdat->node_zones[z_idx+1];
1808 /* We always start with the higher zone's protection value */
1809 pages = higherzone->protection[alloc_type];
1812 * We get a lower-zone-protection contribution only if there are
1813 * pages in the higher zone and if we're not the highest zone
1814 * in the current zonelist. e.g., never happens for GFP_DMA. Happens
1815 * only for ZONE_DMA in a GFP_KERNEL allocation and happens for ZONE_DMA
1816 * and ZONE_NORMAL for a GFP_HIGHMEM allocation.
1818 if (higherzone->present_pages && z_idx < alloc_type)
1819 pages += higherzone->pages_low * sysctl_lower_zone_protection;
1825 * setup_per_zone_protection - called whenver min_free_kbytes or
1826 * sysctl_lower_zone_protection changes. Ensures that each zone
1827 * has a correct pages_protected value, so an adequate number of
1828 * pages are left in the zone after a successful __alloc_pages().
1830 * This algorithm is way confusing. I tries to keep the same behavior
1831 * as we had with the incremental min iterative algorithm.
1833 static void setup_per_zone_protection(void)
1835 struct pglist_data *pgdat;
1836 struct zone *zones, *zone;
1840 for_each_pgdat(pgdat) {
1841 zones = pgdat->node_zones;
1843 for (i = 0, max_zone = 0; i < MAX_NR_ZONES; i++)
1844 if (zones[i].present_pages)
1848 * For each of the different allocation types:
1849 * GFP_DMA -> GFP_KERNEL -> GFP_HIGHMEM
1851 for (i = 0; i < GFP_ZONETYPES; i++) {
1853 * For each of the zones:
1854 * ZONE_HIGHMEM -> ZONE_NORMAL -> ZONE_DMA
1856 for (j = MAX_NR_ZONES-1; j >= 0; j--) {
1860 * We never protect zones that don't have memory
1861 * in them (j>max_zone) or zones that aren't in
1862 * the zonelists for a certain type of
1863 * allocation (j>i). We have to assign these to
1864 * zero because the lower zones take
1865 * contributions from the higher zones.
1867 if (j > max_zone || j > i) {
1868 zone->protection[i] = 0;
1872 * The contribution of the next higher zone
1874 zone->protection[i] = higherzone_val(zone,
1876 zone->protection[i] += zone->pages_low;
1883 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
1884 * that the pages_{min,low,high} values for each zone are set correctly
1885 * with respect to min_free_kbytes.
1887 static void setup_per_zone_pages_min(void)
1889 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
1890 unsigned long lowmem_pages = 0;
1892 unsigned long flags;
1894 /* Calculate total number of !ZONE_HIGHMEM pages */
1895 for_each_zone(zone) {
1896 if (!is_highmem(zone))
1897 lowmem_pages += zone->present_pages;
1900 for_each_zone(zone) {
1901 spin_lock_irqsave(&zone->lru_lock, flags);
1902 if (is_highmem(zone)) {
1904 * Often, highmem doesn't need to reserve any pages.
1905 * But the pages_min/low/high values are also used for
1906 * batching up page reclaim activity so we need a
1907 * decent value here.
1911 min_pages = zone->present_pages / 1024;
1912 if (min_pages < SWAP_CLUSTER_MAX)
1913 min_pages = SWAP_CLUSTER_MAX;
1914 if (min_pages > 128)
1916 zone->pages_min = min_pages;
1918 /* if it's a lowmem zone, reserve a number of pages
1919 * proportionate to the zone's size.
1921 zone->pages_min = (pages_min * zone->present_pages) /
1925 zone->pages_low = zone->pages_min * 2;
1926 zone->pages_high = zone->pages_min * 3;
1927 spin_unlock_irqrestore(&zone->lru_lock, flags);
1932 * Initialise min_free_kbytes.
1934 * For small machines we want it small (128k min). For large machines
1935 * we want it large (16MB max). But it is not linear, because network
1936 * bandwidth does not increase linearly with machine size. We use
1938 * min_free_kbytes = sqrt(lowmem_kbytes)
1954 static int __init init_per_zone_pages_min(void)
1956 unsigned long lowmem_kbytes;
1958 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
1960 min_free_kbytes = int_sqrt(lowmem_kbytes);
1961 if (min_free_kbytes < 128)
1962 min_free_kbytes = 128;
1963 if (min_free_kbytes > 16384)
1964 min_free_kbytes = 16384;
1965 setup_per_zone_pages_min();
1966 setup_per_zone_protection();
1969 module_init(init_per_zone_pages_min)
1972 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
1973 * that we can call two helper functions whenever min_free_kbytes
1976 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
1977 struct file *file, void __user *buffer, size_t *length)
1979 proc_dointvec(table, write, file, buffer, length);
1980 setup_per_zone_pages_min();
1981 setup_per_zone_protection();
1986 * lower_zone_protection_sysctl_handler - just a wrapper around
1987 * proc_dointvec() so that we can call setup_per_zone_protection()
1988 * whenever sysctl_lower_zone_protection changes.
1990 int lower_zone_protection_sysctl_handler(ctl_table *table, int write,
1991 struct file *file, void __user *buffer, size_t *length)
1993 proc_dointvec_minmax(table, write, file, buffer, length);
1994 setup_per_zone_protection();
1999 * allocate a large system hash table from bootmem
2000 * - it is assumed that the hash table must contain an exact power-of-2
2001 * quantity of entries
2003 void *__init alloc_large_system_hash(const char *tablename,
2004 unsigned long bucketsize,
2005 unsigned long numentries,
2007 int consider_highmem,
2008 unsigned int *_hash_shift,
2009 unsigned int *_hash_mask)
2011 unsigned long mem, max, log2qty, size;
2014 /* round applicable memory size up to nearest megabyte */
2015 mem = consider_highmem ? nr_all_pages : nr_kernel_pages;
2016 mem += (1UL << (20 - PAGE_SHIFT)) - 1;
2017 mem >>= 20 - PAGE_SHIFT;
2018 mem <<= 20 - PAGE_SHIFT;
2020 /* limit to 1 bucket per 2^scale bytes of low memory (rounded up to
2021 * nearest power of 2 in size) */
2022 if (scale > PAGE_SHIFT)
2023 mem >>= (scale - PAGE_SHIFT);
2025 mem <<= (PAGE_SHIFT - scale);
2027 mem = 1UL << (long_log2(mem) + 1);
2029 /* limit allocation size */
2030 max = (1UL << (PAGE_SHIFT + MAX_SYS_HASH_TABLE_ORDER)) / bucketsize;
2034 /* allow the kernel cmdline to have a say */
2035 if (!numentries || numentries > max)
2038 log2qty = long_log2(numentries);
2041 size = bucketsize << log2qty;
2043 table = (void *) alloc_bootmem(size);
2045 } while (!table && size > PAGE_SIZE);
2048 panic("Failed to allocate %s hash table\n", tablename);
2050 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2053 long_log2(size) - PAGE_SHIFT,
2057 *_hash_shift = log2qty;
2059 *_hash_mask = (1 << log2qty) - 1;