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 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)
296 * The order of subdivision here is critical for the IO subsystem.
297 * Please do not alter this order without good reasons and regression
298 * testing. Specifically, as large blocks of memory are subdivided,
299 * the order in which smaller blocks are delivered depends on the order
300 * they're subdivided in this function. This is the primary factor
301 * influencing the order in which pages are delivered to the IO
302 * subsystem according to empirical testing, and this is also justified
303 * by considering the behavior of a buddy system containing a single
304 * large block of memory acted on by a series of small allocations.
305 * This behavior is a critical factor in sglist merging's success.
309 static inline struct page *
310 expand(struct zone *zone, struct page *page,
311 unsigned long index, int low, int high, struct free_area *area)
313 unsigned long size = 1 << high;
319 BUG_ON(bad_range(zone, &page[size]));
320 list_add(&page[size].lru, &area->free_list);
321 MARK_USED(index + size, high, area);
326 static inline void set_page_refs(struct page *page, int order)
329 set_page_count(page, 1);
334 * We need to reference all the pages for this order, otherwise if
335 * anyone accesses one of the pages with (get/put) it will be freed.
337 for (i = 0; i < (1 << order); i++)
338 set_page_count(page+i, 1);
339 #endif /* CONFIG_MMU */
343 * This page is about to be returned from the page allocator
345 static void prep_new_page(struct page *page, int order)
347 if (page->mapping || page_mapped(page) ||
358 1 << PG_writeback )))
359 bad_page(__FUNCTION__, page);
361 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
362 1 << PG_referenced | 1 << PG_arch_1 |
363 1 << PG_checked | 1 << PG_mappedtodisk);
365 set_page_refs(page, order);
369 * Do the hard work of removing an element from the buddy allocator.
370 * Call me with the zone->lock already held.
372 static struct page *__rmqueue(struct zone *zone, unsigned int order)
374 struct free_area * area;
375 unsigned int current_order;
379 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
380 area = zone->free_area + current_order;
381 if (list_empty(&area->free_list))
384 page = list_entry(area->free_list.next, struct page, lru);
385 list_del(&page->lru);
386 index = page - zone->zone_mem_map;
387 if (current_order != MAX_ORDER-1)
388 MARK_USED(index, current_order, area);
389 zone->free_pages -= 1UL << order;
390 return expand(zone, page, index, order, current_order, area);
397 * Obtain a specified number of elements from the buddy allocator, all under
398 * a single hold of the lock, for efficiency. Add them to the supplied list.
399 * Returns the number of new pages which were placed at *list.
401 static int rmqueue_bulk(struct zone *zone, unsigned int order,
402 unsigned long count, struct list_head *list)
409 spin_lock_irqsave(&zone->lock, flags);
410 for (i = 0; i < count; ++i) {
411 page = __rmqueue(zone, order);
415 list_add_tail(&page->lru, list);
417 spin_unlock_irqrestore(&zone->lock, flags);
421 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
422 static void __drain_pages(unsigned int cpu)
427 for_each_zone(zone) {
428 struct per_cpu_pageset *pset;
430 pset = &zone->pageset[cpu];
431 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
432 struct per_cpu_pages *pcp;
435 pcp->count -= free_pages_bulk(zone, pcp->count,
440 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
443 int is_head_of_free_region(struct page *page)
445 struct zone *zone = page_zone(page);
448 struct list_head *curr;
451 * Should not matter as we need quiescent system for
452 * suspend anyway, but...
454 spin_lock_irqsave(&zone->lock, flags);
455 for (order = MAX_ORDER - 1; order >= 0; --order)
456 list_for_each(curr, &zone->free_area[order].free_list)
457 if (page == list_entry(curr, struct page, lru)) {
458 spin_unlock_irqrestore(&zone->lock, flags);
461 spin_unlock_irqrestore(&zone->lock, flags);
466 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
468 void drain_local_pages(void)
472 local_irq_save(flags);
473 __drain_pages(smp_processor_id());
474 local_irq_restore(flags);
476 #endif /* CONFIG_PM */
478 static void zone_statistics(struct zonelist *zonelist, struct zone *z)
483 pg_data_t *pg = z->zone_pgdat;
484 pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
485 struct per_cpu_pageset *p;
487 local_irq_save(flags);
488 cpu = smp_processor_id();
489 p = &z->pageset[cpu];
491 z->pageset[cpu].numa_hit++;
494 zonelist->zones[0]->pageset[cpu].numa_foreign++;
496 if (pg == NODE_DATA(numa_node_id()))
500 local_irq_restore(flags);
505 * Free a 0-order page
507 static void FASTCALL(free_hot_cold_page(struct page *page, int cold));
508 static void fastcall free_hot_cold_page(struct page *page, int cold)
510 struct zone *zone = page_zone(page);
511 struct per_cpu_pages *pcp;
514 kernel_map_pages(page, 1, 0);
515 inc_page_state(pgfree);
516 free_pages_check(__FUNCTION__, page);
517 pcp = &zone->pageset[get_cpu()].pcp[cold];
518 local_irq_save(flags);
519 if (pcp->count >= pcp->high)
520 pcp->count -= free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
521 list_add(&page->lru, &pcp->list);
523 local_irq_restore(flags);
527 void fastcall free_hot_page(struct page *page)
529 free_hot_cold_page(page, 0);
532 void fastcall free_cold_page(struct page *page)
534 free_hot_cold_page(page, 1);
538 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
539 * we cheat by calling it from here, in the order > 0 path. Saves a branch
544 buffered_rmqueue(struct zone *zone, int order, int gfp_flags)
547 struct page *page = NULL;
548 int cold = !!(gfp_flags & __GFP_COLD);
551 struct per_cpu_pages *pcp;
553 pcp = &zone->pageset[get_cpu()].pcp[cold];
554 local_irq_save(flags);
555 if (pcp->count <= pcp->low)
556 pcp->count += rmqueue_bulk(zone, 0,
557 pcp->batch, &pcp->list);
559 page = list_entry(pcp->list.next, struct page, lru);
560 list_del(&page->lru);
563 local_irq_restore(flags);
568 spin_lock_irqsave(&zone->lock, flags);
569 page = __rmqueue(zone, order);
570 spin_unlock_irqrestore(&zone->lock, flags);
574 BUG_ON(bad_range(zone, page));
575 mod_page_state_zone(zone, pgalloc, 1 << order);
576 prep_new_page(page, order);
577 if (order && (gfp_flags & __GFP_COMP))
578 prep_compound_page(page, order);
584 * This is the 'heart' of the zoned buddy allocator.
586 * Herein lies the mysterious "incremental min". That's the
588 * local_low = z->pages_low;
591 * thing. The intent here is to provide additional protection to low zones for
592 * allocation requests which _could_ use higher zones. So a GFP_HIGHMEM
593 * request is not allowed to dip as deeply into the normal zone as a GFP_KERNEL
594 * request. This preserves additional space in those lower zones for requests
595 * which really do need memory from those zones. It means that on a decent
596 * sized machine, GFP_HIGHMEM and GFP_KERNEL requests basically leave the DMA
599 struct page * fastcall
600 __alloc_pages(unsigned int gfp_mask, unsigned int order,
601 struct zonelist *zonelist)
603 const int wait = gfp_mask & __GFP_WAIT;
607 struct reclaim_state reclaim_state;
608 struct task_struct *p = current;
613 might_sleep_if(wait);
615 zones = zonelist->zones; /* the list of zones suitable for gfp_mask */
616 if (zones[0] == NULL) /* no zones in the zonelist */
619 alloc_type = zone_idx(zones[0]);
621 /* Go through the zonelist once, looking for a zone with enough free */
622 for (i = 0; zones[i] != NULL; i++) {
623 struct zone *z = zones[i];
625 min = (1<<order) + z->protection[alloc_type];
628 * We let real-time tasks dip their real-time paws a little
629 * deeper into reserves.
632 min -= z->pages_low >> 1;
634 if (z->free_pages >= min ||
635 (!wait && z->free_pages >= z->pages_high)) {
636 page = buffered_rmqueue(z, order, gfp_mask);
638 zone_statistics(zonelist, z);
644 /* we're somewhat low on memory, failed to find what we needed */
645 for (i = 0; zones[i] != NULL; i++)
646 wakeup_kswapd(zones[i]);
648 /* Go through the zonelist again, taking __GFP_HIGH into account */
649 for (i = 0; zones[i] != NULL; i++) {
650 struct zone *z = zones[i];
652 min = (1<<order) + z->protection[alloc_type];
654 if (gfp_mask & __GFP_HIGH)
655 min -= z->pages_low >> 2;
657 min -= z->pages_low >> 1;
659 if (z->free_pages >= min ||
660 (!wait && z->free_pages >= z->pages_high)) {
661 page = buffered_rmqueue(z, order, gfp_mask);
663 zone_statistics(zonelist, z);
669 /* here we're in the low on memory slow path */
672 if ((p->flags & (PF_MEMALLOC | PF_MEMDIE)) && !in_interrupt()) {
673 /* go through the zonelist yet again, ignoring mins */
674 for (i = 0; zones[i] != NULL; i++) {
675 struct zone *z = zones[i];
677 page = buffered_rmqueue(z, order, gfp_mask);
679 zone_statistics(zonelist, z);
686 /* Atomic allocations - we can't balance anything */
690 p->flags |= PF_MEMALLOC;
691 reclaim_state.reclaimed_slab = 0;
692 p->reclaim_state = &reclaim_state;
694 try_to_free_pages(zones, gfp_mask, order);
696 p->reclaim_state = NULL;
697 p->flags &= ~PF_MEMALLOC;
699 /* go through the zonelist yet one more time */
700 for (i = 0; zones[i] != NULL; i++) {
701 struct zone *z = zones[i];
703 min = (1UL << order) + z->protection[alloc_type];
705 if (z->free_pages >= min ||
706 (!wait && z->free_pages >= z->pages_high)) {
707 page = buffered_rmqueue(z, order, gfp_mask);
709 zone_statistics(zonelist, z);
716 * Don't let big-order allocations loop unless the caller explicitly
717 * requests that. Wait for some write requests to complete then retry.
719 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL, but that
720 * may not be true in other implementations.
723 if (!(gfp_mask & __GFP_NORETRY)) {
724 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
726 if (gfp_mask & __GFP_NOFAIL)
730 blk_congestion_wait(WRITE, HZ/50);
735 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
736 printk(KERN_WARNING "%s: page allocation failure."
737 " order:%d, mode:0x%x\n",
738 p->comm, order, gfp_mask);
743 kernel_map_pages(page, 1 << order, 1);
747 EXPORT_SYMBOL(__alloc_pages);
750 * Common helper functions.
752 fastcall unsigned long __get_free_pages(unsigned int gfp_mask, unsigned int order)
755 page = alloc_pages(gfp_mask, order);
758 return (unsigned long) page_address(page);
761 EXPORT_SYMBOL(__get_free_pages);
763 fastcall unsigned long get_zeroed_page(unsigned int gfp_mask)
768 * get_zeroed_page() returns a 32-bit address, which cannot represent
771 BUG_ON(gfp_mask & __GFP_HIGHMEM);
773 page = alloc_pages(gfp_mask, 0);
775 void *address = page_address(page);
777 return (unsigned long) address;
782 EXPORT_SYMBOL(get_zeroed_page);
784 void __pagevec_free(struct pagevec *pvec)
786 int i = pagevec_count(pvec);
789 free_hot_cold_page(pvec->pages[i], pvec->cold);
792 fastcall void __free_pages(struct page *page, unsigned int order)
794 if (!PageReserved(page) && put_page_testzero(page)) {
798 __free_pages_ok(page, order);
802 EXPORT_SYMBOL(__free_pages);
804 fastcall void free_pages(unsigned long addr, unsigned int order)
807 BUG_ON(!virt_addr_valid(addr));
808 __free_pages(virt_to_page(addr), order);
812 EXPORT_SYMBOL(free_pages);
815 * Total amount of free (allocatable) RAM:
817 unsigned int nr_free_pages(void)
819 unsigned int sum = 0;
823 sum += zone->free_pages;
828 EXPORT_SYMBOL(nr_free_pages);
831 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
833 unsigned int i, sum = 0;
835 for (i = 0; i < MAX_NR_ZONES; i++)
836 sum += pgdat->node_zones[i].free_pages;
842 static unsigned int nr_free_zone_pages(int offset)
845 unsigned int sum = 0;
847 for_each_pgdat(pgdat) {
848 struct zonelist *zonelist = pgdat->node_zonelists + offset;
849 struct zone **zonep = zonelist->zones;
852 for (zone = *zonep++; zone; zone = *zonep++) {
853 unsigned long size = zone->present_pages;
854 unsigned long high = zone->pages_high;
864 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
866 unsigned int nr_free_buffer_pages(void)
868 return nr_free_zone_pages(GFP_USER & GFP_ZONEMASK);
872 * Amount of free RAM allocatable within all zones
874 unsigned int nr_free_pagecache_pages(void)
876 return nr_free_zone_pages(GFP_HIGHUSER & GFP_ZONEMASK);
879 #ifdef CONFIG_HIGHMEM
880 unsigned int nr_free_highpages (void)
883 unsigned int pages = 0;
885 for_each_pgdat(pgdat)
886 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
893 static void show_node(struct zone *zone)
895 printk("Node %d ", zone->zone_pgdat->node_id);
898 #define show_node(zone) do { } while (0)
902 * Accumulate the page_state information across all CPUs.
903 * The result is unavoidably approximate - it can change
904 * during and after execution of this function.
906 DEFINE_PER_CPU(struct page_state, page_states) = {0};
907 EXPORT_PER_CPU_SYMBOL(page_states);
909 atomic_t nr_pagecache = ATOMIC_INIT(0);
910 EXPORT_SYMBOL(nr_pagecache);
912 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
915 void __get_page_state(struct page_state *ret, int nr)
919 memset(ret, 0, sizeof(*ret));
920 while (cpu < NR_CPUS) {
921 unsigned long *in, *out, off;
923 if (!cpu_possible(cpu)) {
928 in = (unsigned long *)&per_cpu(page_states, cpu);
930 if (cpu < NR_CPUS && cpu_possible(cpu))
931 prefetch(&per_cpu(page_states, cpu));
932 out = (unsigned long *)ret;
933 for (off = 0; off < nr; off++)
938 void get_page_state(struct page_state *ret)
942 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
943 nr /= sizeof(unsigned long);
945 __get_page_state(ret, nr + 1);
948 void get_full_page_state(struct page_state *ret)
950 __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long));
953 unsigned long __read_page_state(unsigned offset)
955 unsigned long ret = 0;
958 for (cpu = 0; cpu < NR_CPUS; cpu++) {
961 if (!cpu_possible(cpu))
964 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
965 ret += *((unsigned long *)in);
970 void get_zone_counts(unsigned long *active,
971 unsigned long *inactive, unsigned long *free)
978 for_each_zone(zone) {
979 *active += zone->nr_active;
980 *inactive += zone->nr_inactive;
981 *free += zone->free_pages;
985 void si_meminfo(struct sysinfo *val)
987 val->totalram = totalram_pages;
989 val->freeram = nr_free_pages();
990 val->bufferram = nr_blockdev_pages();
991 #ifdef CONFIG_HIGHMEM
992 val->totalhigh = totalhigh_pages;
993 val->freehigh = nr_free_highpages();
998 val->mem_unit = PAGE_SIZE;
999 if (vx_flags(VXF_VIRT_MEM, 0))
1000 vx_vsi_meminfo(val);
1003 EXPORT_SYMBOL(si_meminfo);
1006 void si_meminfo_node(struct sysinfo *val, int nid)
1008 pg_data_t *pgdat = NODE_DATA(nid);
1010 val->totalram = pgdat->node_present_pages;
1011 val->freeram = nr_free_pages_pgdat(pgdat);
1012 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1013 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1014 val->mem_unit = PAGE_SIZE;
1018 #define K(x) ((x) << (PAGE_SHIFT-10))
1021 * Show free area list (used inside shift_scroll-lock stuff)
1022 * We also calculate the percentage fragmentation. We do this by counting the
1023 * memory on each free list with the exception of the first item on the list.
1025 void show_free_areas(void)
1027 struct page_state ps;
1028 int cpu, temperature;
1029 unsigned long active;
1030 unsigned long inactive;
1034 for_each_zone(zone) {
1036 printk("%s per-cpu:", zone->name);
1038 if (!zone->present_pages) {
1044 for (cpu = 0; cpu < NR_CPUS; ++cpu) {
1045 struct per_cpu_pageset *pageset;
1047 if (!cpu_possible(cpu))
1050 pageset = zone->pageset + cpu;
1052 for (temperature = 0; temperature < 2; temperature++)
1053 printk("cpu %d %s: low %d, high %d, batch %d\n",
1055 temperature ? "cold" : "hot",
1056 pageset->pcp[temperature].low,
1057 pageset->pcp[temperature].high,
1058 pageset->pcp[temperature].batch);
1062 get_page_state(&ps);
1063 get_zone_counts(&active, &inactive, &free);
1065 printk("\nFree pages: %11ukB (%ukB HighMem)\n",
1067 K(nr_free_highpages()));
1069 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1070 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1079 ps.nr_page_table_pages);
1081 for_each_zone(zone) {
1095 K(zone->free_pages),
1098 K(zone->pages_high),
1100 K(zone->nr_inactive),
1101 K(zone->present_pages)
1103 printk("protections[]:");
1104 for (i = 0; i < MAX_NR_ZONES; i++)
1105 printk(" %lu", zone->protection[i]);
1109 for_each_zone(zone) {
1110 struct list_head *elem;
1111 unsigned long nr, flags, order, total = 0;
1114 printk("%s: ", zone->name);
1115 if (!zone->present_pages) {
1120 spin_lock_irqsave(&zone->lock, flags);
1121 for (order = 0; order < MAX_ORDER; order++) {
1123 list_for_each(elem, &zone->free_area[order].free_list)
1125 total += nr << order;
1126 printk("%lu*%lukB ", nr, K(1UL) << order);
1128 spin_unlock_irqrestore(&zone->lock, flags);
1129 printk("= %lukB\n", K(total));
1132 show_swap_cache_info();
1136 * Builds allocation fallback zone lists.
1138 static int __init build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, int j, int k)
1145 zone = pgdat->node_zones + ZONE_HIGHMEM;
1146 if (zone->present_pages) {
1147 #ifndef CONFIG_HIGHMEM
1150 zonelist->zones[j++] = zone;
1153 zone = pgdat->node_zones + ZONE_NORMAL;
1154 if (zone->present_pages)
1155 zonelist->zones[j++] = zone;
1157 zone = pgdat->node_zones + ZONE_DMA;
1158 if (zone->present_pages)
1159 zonelist->zones[j++] = zone;
1166 #define MAX_NODE_LOAD (numnodes)
1167 static int __initdata node_load[MAX_NUMNODES];
1169 * find_next_best_node - find the next node that should appear in a given
1170 * node's fallback list
1171 * @node: node whose fallback list we're appending
1172 * @used_node_mask: pointer to the bitmap of already used nodes
1174 * We use a number of factors to determine which is the next node that should
1175 * appear on a given node's fallback list. The node should not have appeared
1176 * already in @node's fallback list, and it should be the next closest node
1177 * according to the distance array (which contains arbitrary distance values
1178 * from each node to each node in the system), and should also prefer nodes
1179 * with no CPUs, since presumably they'll have very little allocation pressure
1180 * on them otherwise.
1181 * It returns -1 if no node is found.
1183 static int __init find_next_best_node(int node, void *used_node_mask)
1186 int min_val = INT_MAX;
1189 for (i = 0; i < numnodes; i++) {
1192 /* Start from local node */
1193 n = (node+i)%numnodes;
1195 /* Don't want a node to appear more than once */
1196 if (test_bit(n, used_node_mask))
1199 /* Use the distance array to find the distance */
1200 val = node_distance(node, n);
1202 /* Give preference to headless and unused nodes */
1203 tmp = node_to_cpumask(n);
1204 if (!cpus_empty(tmp))
1205 val += PENALTY_FOR_NODE_WITH_CPUS;
1207 /* Slight preference for less loaded node */
1208 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1209 val += node_load[n];
1211 if (val < min_val) {
1218 set_bit(best_node, used_node_mask);
1223 static void __init build_zonelists(pg_data_t *pgdat)
1225 int i, j, k, node, local_node;
1226 int prev_node, load;
1227 struct zonelist *zonelist;
1228 DECLARE_BITMAP(used_mask, MAX_NUMNODES);
1230 /* initialize zonelists */
1231 for (i = 0; i < GFP_ZONETYPES; i++) {
1232 zonelist = pgdat->node_zonelists + i;
1233 memset(zonelist, 0, sizeof(*zonelist));
1234 zonelist->zones[0] = NULL;
1237 /* NUMA-aware ordering of nodes */
1238 local_node = pgdat->node_id;
1240 prev_node = local_node;
1241 bitmap_zero(used_mask, MAX_NUMNODES);
1242 while ((node = find_next_best_node(local_node, used_mask)) >= 0) {
1244 * We don't want to pressure a particular node.
1245 * So adding penalty to the first node in same
1246 * distance group to make it round-robin.
1248 if (node_distance(local_node, node) !=
1249 node_distance(local_node, prev_node))
1250 node_load[node] += load;
1253 for (i = 0; i < GFP_ZONETYPES; i++) {
1254 zonelist = pgdat->node_zonelists + i;
1255 for (j = 0; zonelist->zones[j] != NULL; j++);
1258 if (i & __GFP_HIGHMEM)
1263 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1264 zonelist->zones[j] = NULL;
1269 #else /* CONFIG_NUMA */
1271 static void __init build_zonelists(pg_data_t *pgdat)
1273 int i, j, k, node, local_node;
1275 local_node = pgdat->node_id;
1276 for (i = 0; i < GFP_ZONETYPES; i++) {
1277 struct zonelist *zonelist;
1279 zonelist = pgdat->node_zonelists + i;
1280 memset(zonelist, 0, sizeof(*zonelist));
1284 if (i & __GFP_HIGHMEM)
1289 j = build_zonelists_node(pgdat, zonelist, j, k);
1291 * Now we build the zonelist so that it contains the zones
1292 * of all the other nodes.
1293 * We don't want to pressure a particular node, so when
1294 * building the zones for node N, we make sure that the
1295 * zones coming right after the local ones are those from
1296 * node N+1 (modulo N)
1298 for (node = local_node + 1; node < numnodes; node++)
1299 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1300 for (node = 0; node < local_node; node++)
1301 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1303 zonelist->zones[j] = NULL;
1307 #endif /* CONFIG_NUMA */
1309 void __init build_all_zonelists(void)
1313 for(i = 0 ; i < numnodes ; i++)
1314 build_zonelists(NODE_DATA(i));
1315 printk("Built %i zonelists\n", numnodes);
1319 * Helper functions to size the waitqueue hash table.
1320 * Essentially these want to choose hash table sizes sufficiently
1321 * large so that collisions trying to wait on pages are rare.
1322 * But in fact, the number of active page waitqueues on typical
1323 * systems is ridiculously low, less than 200. So this is even
1324 * conservative, even though it seems large.
1326 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1327 * waitqueues, i.e. the size of the waitq table given the number of pages.
1329 #define PAGES_PER_WAITQUEUE 256
1331 static inline unsigned long wait_table_size(unsigned long pages)
1333 unsigned long size = 1;
1335 pages /= PAGES_PER_WAITQUEUE;
1337 while (size < pages)
1341 * Once we have dozens or even hundreds of threads sleeping
1342 * on IO we've got bigger problems than wait queue collision.
1343 * Limit the size of the wait table to a reasonable size.
1345 size = min(size, 4096UL);
1347 return max(size, 4UL);
1351 * This is an integer logarithm so that shifts can be used later
1352 * to extract the more random high bits from the multiplicative
1353 * hash function before the remainder is taken.
1355 static inline unsigned long wait_table_bits(unsigned long size)
1360 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1362 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1363 unsigned long *zones_size, unsigned long *zholes_size)
1365 unsigned long realtotalpages, totalpages = 0;
1368 for (i = 0; i < MAX_NR_ZONES; i++)
1369 totalpages += zones_size[i];
1370 pgdat->node_spanned_pages = totalpages;
1372 realtotalpages = totalpages;
1374 for (i = 0; i < MAX_NR_ZONES; i++)
1375 realtotalpages -= zholes_size[i];
1376 pgdat->node_present_pages = realtotalpages;
1377 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1382 * Initially all pages are reserved - free ones are freed
1383 * up by free_all_bootmem() once the early boot process is
1384 * done. Non-atomic initialization, single-pass.
1386 void __init memmap_init_zone(struct page *start, unsigned long size, int nid,
1387 unsigned long zone, unsigned long start_pfn)
1391 for (page = start; page < (start + size); page++) {
1392 set_page_zone(page, NODEZONE(nid, zone));
1393 set_page_count(page, 0);
1394 SetPageReserved(page);
1395 INIT_LIST_HEAD(&page->lru);
1396 #ifdef WANT_PAGE_VIRTUAL
1397 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1398 if (!is_highmem_idx(zone))
1399 set_page_address(page, __va(start_pfn << PAGE_SHIFT));
1405 #ifndef __HAVE_ARCH_MEMMAP_INIT
1406 #define memmap_init(start, size, nid, zone, start_pfn) \
1407 memmap_init_zone((start), (size), (nid), (zone), (start_pfn))
1411 * Set up the zone data structures:
1412 * - mark all pages reserved
1413 * - mark all memory queues empty
1414 * - clear the memory bitmaps
1416 static void __init free_area_init_core(struct pglist_data *pgdat,
1417 unsigned long *zones_size, unsigned long *zholes_size)
1420 const unsigned long zone_required_alignment = 1UL << (MAX_ORDER-1);
1421 int cpu, nid = pgdat->node_id;
1422 struct page *lmem_map = pgdat->node_mem_map;
1423 unsigned long zone_start_pfn = pgdat->node_start_pfn;
1425 pgdat->nr_zones = 0;
1426 init_waitqueue_head(&pgdat->kswapd_wait);
1428 for (j = 0; j < MAX_NR_ZONES; j++) {
1429 struct zone *zone = pgdat->node_zones + j;
1430 unsigned long size, realsize;
1431 unsigned long batch;
1433 zone_table[NODEZONE(nid, j)] = zone;
1434 realsize = size = zones_size[j];
1436 realsize -= zholes_size[j];
1438 if (j == ZONE_DMA || j == ZONE_NORMAL)
1439 nr_kernel_pages += realsize;
1440 nr_all_pages += realsize;
1442 zone->spanned_pages = size;
1443 zone->present_pages = realsize;
1444 zone->name = zone_names[j];
1445 spin_lock_init(&zone->lock);
1446 spin_lock_init(&zone->lru_lock);
1447 zone->zone_pgdat = pgdat;
1448 zone->free_pages = 0;
1450 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
1453 * The per-cpu-pages pools are set to around 1000th of the
1454 * size of the zone. But no more than 1/4 of a meg - there's
1455 * no point in going beyond the size of L2 cache.
1457 * OK, so we don't know how big the cache is. So guess.
1459 batch = zone->present_pages / 1024;
1460 if (batch * PAGE_SIZE > 256 * 1024)
1461 batch = (256 * 1024) / PAGE_SIZE;
1462 batch /= 4; /* We effectively *= 4 below */
1466 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1467 struct per_cpu_pages *pcp;
1469 pcp = &zone->pageset[cpu].pcp[0]; /* hot */
1471 pcp->low = 2 * batch;
1472 pcp->high = 6 * batch;
1473 pcp->batch = 1 * batch;
1474 INIT_LIST_HEAD(&pcp->list);
1476 pcp = &zone->pageset[cpu].pcp[1]; /* cold */
1479 pcp->high = 2 * batch;
1480 pcp->batch = 1 * batch;
1481 INIT_LIST_HEAD(&pcp->list);
1483 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
1484 zone_names[j], realsize, batch);
1485 INIT_LIST_HEAD(&zone->active_list);
1486 INIT_LIST_HEAD(&zone->inactive_list);
1487 zone->nr_scan_active = 0;
1488 zone->nr_scan_inactive = 0;
1489 zone->nr_active = 0;
1490 zone->nr_inactive = 0;
1495 * The per-page waitqueue mechanism uses hashed waitqueues
1498 zone->wait_table_size = wait_table_size(size);
1499 zone->wait_table_bits =
1500 wait_table_bits(zone->wait_table_size);
1501 zone->wait_table = (wait_queue_head_t *)
1502 alloc_bootmem_node(pgdat, zone->wait_table_size
1503 * sizeof(wait_queue_head_t));
1505 for(i = 0; i < zone->wait_table_size; ++i)
1506 init_waitqueue_head(zone->wait_table + i);
1508 pgdat->nr_zones = j+1;
1510 zone->zone_mem_map = lmem_map;
1511 zone->zone_start_pfn = zone_start_pfn;
1513 if ((zone_start_pfn) & (zone_required_alignment-1))
1514 printk("BUG: wrong zone alignment, it will crash\n");
1516 memmap_init(lmem_map, size, nid, j, zone_start_pfn);
1518 zone_start_pfn += size;
1521 for (i = 0; ; i++) {
1522 unsigned long bitmap_size;
1524 INIT_LIST_HEAD(&zone->free_area[i].free_list);
1525 if (i == MAX_ORDER-1) {
1526 zone->free_area[i].map = NULL;
1531 * Page buddy system uses "index >> (i+1)",
1532 * where "index" is at most "size-1".
1534 * The extra "+3" is to round down to byte
1535 * size (8 bits per byte assumption). Thus
1536 * we get "(size-1) >> (i+4)" as the last byte
1539 * The "+1" is because we want to round the
1540 * byte allocation up rather than down. So
1541 * we should have had a "+7" before we shifted
1542 * down by three. Also, we have to add one as
1543 * we actually _use_ the last bit (it's [0,n]
1544 * inclusive, not [0,n[).
1546 * So we actually had +7+1 before we shift
1547 * down by 3. But (n+8) >> 3 == (n >> 3) + 1
1548 * (modulo overflows, which we do not have).
1550 * Finally, we LONG_ALIGN because all bitmap
1551 * operations are on longs.
1553 bitmap_size = (size-1) >> (i+4);
1554 bitmap_size = LONG_ALIGN(bitmap_size+1);
1555 zone->free_area[i].map =
1556 (unsigned long *) alloc_bootmem_node(pgdat, bitmap_size);
1561 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
1562 struct page *node_mem_map, unsigned long *zones_size,
1563 unsigned long node_start_pfn, unsigned long *zholes_size)
1567 pgdat->node_id = nid;
1568 pgdat->node_start_pfn = node_start_pfn;
1569 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
1570 if (!node_mem_map) {
1571 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
1572 node_mem_map = alloc_bootmem_node(pgdat, size);
1574 pgdat->node_mem_map = node_mem_map;
1576 free_area_init_core(pgdat, zones_size, zholes_size);
1579 #ifndef CONFIG_DISCONTIGMEM
1580 static bootmem_data_t contig_bootmem_data;
1581 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
1583 EXPORT_SYMBOL(contig_page_data);
1585 void __init free_area_init(unsigned long *zones_size)
1587 free_area_init_node(0, &contig_page_data, NULL, zones_size,
1588 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
1589 mem_map = contig_page_data.node_mem_map;
1593 #ifdef CONFIG_PROC_FS
1595 #include <linux/seq_file.h>
1597 static void *frag_start(struct seq_file *m, loff_t *pos)
1602 for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
1608 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1610 pg_data_t *pgdat = (pg_data_t *)arg;
1613 return pgdat->pgdat_next;
1616 static void frag_stop(struct seq_file *m, void *arg)
1621 * This walks the freelist for each zone. Whilst this is slow, I'd rather
1622 * be slow here than slow down the fast path by keeping stats - mjbligh
1624 static int frag_show(struct seq_file *m, void *arg)
1626 pg_data_t *pgdat = (pg_data_t *)arg;
1628 struct zone *node_zones = pgdat->node_zones;
1629 unsigned long flags;
1632 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1633 if (!zone->present_pages)
1636 spin_lock_irqsave(&zone->lock, flags);
1637 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1638 for (order = 0; order < MAX_ORDER; ++order) {
1639 unsigned long nr_bufs = 0;
1640 struct list_head *elem;
1642 list_for_each(elem, &(zone->free_area[order].free_list))
1644 seq_printf(m, "%6lu ", nr_bufs);
1646 spin_unlock_irqrestore(&zone->lock, flags);
1652 struct seq_operations fragmentation_op = {
1653 .start = frag_start,
1659 static char *vmstat_text[] = {
1663 "nr_page_table_pages",
1688 "pgscan_kswapd_high",
1689 "pgscan_kswapd_normal",
1691 "pgscan_kswapd_dma",
1692 "pgscan_direct_high",
1693 "pgscan_direct_normal",
1694 "pgscan_direct_dma",
1699 "kswapd_inodesteal",
1706 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1708 struct page_state *ps;
1710 if (*pos >= ARRAY_SIZE(vmstat_text))
1713 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
1716 return ERR_PTR(-ENOMEM);
1717 get_full_page_state(ps);
1718 ps->pgpgin /= 2; /* sectors -> kbytes */
1720 return (unsigned long *)ps + *pos;
1723 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1726 if (*pos >= ARRAY_SIZE(vmstat_text))
1728 return (unsigned long *)m->private + *pos;
1731 static int vmstat_show(struct seq_file *m, void *arg)
1733 unsigned long *l = arg;
1734 unsigned long off = l - (unsigned long *)m->private;
1736 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1740 static void vmstat_stop(struct seq_file *m, void *arg)
1746 struct seq_operations vmstat_op = {
1747 .start = vmstat_start,
1748 .next = vmstat_next,
1749 .stop = vmstat_stop,
1750 .show = vmstat_show,
1753 #endif /* CONFIG_PROC_FS */
1755 #ifdef CONFIG_HOTPLUG_CPU
1756 static int page_alloc_cpu_notify(struct notifier_block *self,
1757 unsigned long action, void *hcpu)
1759 int cpu = (unsigned long)hcpu;
1762 if (action == CPU_DEAD) {
1763 /* Drain local pagecache count. */
1764 count = &per_cpu(nr_pagecache_local, cpu);
1765 atomic_add(*count, &nr_pagecache);
1767 local_irq_disable();
1773 #endif /* CONFIG_HOTPLUG_CPU */
1775 void __init page_alloc_init(void)
1777 hotcpu_notifier(page_alloc_cpu_notify, 0);
1780 static unsigned long higherzone_val(struct zone *z, int max_zone,
1783 int z_idx = zone_idx(z);
1784 struct zone *higherzone;
1785 unsigned long pages;
1787 /* there is no higher zone to get a contribution from */
1788 if (z_idx == MAX_NR_ZONES-1)
1791 higherzone = &z->zone_pgdat->node_zones[z_idx+1];
1793 /* We always start with the higher zone's protection value */
1794 pages = higherzone->protection[alloc_type];
1797 * We get a lower-zone-protection contribution only if there are
1798 * pages in the higher zone and if we're not the highest zone
1799 * in the current zonelist. e.g., never happens for GFP_DMA. Happens
1800 * only for ZONE_DMA in a GFP_KERNEL allocation and happens for ZONE_DMA
1801 * and ZONE_NORMAL for a GFP_HIGHMEM allocation.
1803 if (higherzone->present_pages && z_idx < alloc_type)
1804 pages += higherzone->pages_low * sysctl_lower_zone_protection;
1810 * setup_per_zone_protection - called whenver min_free_kbytes or
1811 * sysctl_lower_zone_protection changes. Ensures that each zone
1812 * has a correct pages_protected value, so an adequate number of
1813 * pages are left in the zone after a successful __alloc_pages().
1815 * This algorithm is way confusing. I tries to keep the same behavior
1816 * as we had with the incremental min iterative algorithm.
1818 static void setup_per_zone_protection(void)
1820 struct pglist_data *pgdat;
1821 struct zone *zones, *zone;
1825 for_each_pgdat(pgdat) {
1826 zones = pgdat->node_zones;
1828 for (i = 0, max_zone = 0; i < MAX_NR_ZONES; i++)
1829 if (zones[i].present_pages)
1833 * For each of the different allocation types:
1834 * GFP_DMA -> GFP_KERNEL -> GFP_HIGHMEM
1836 for (i = 0; i < GFP_ZONETYPES; i++) {
1838 * For each of the zones:
1839 * ZONE_HIGHMEM -> ZONE_NORMAL -> ZONE_DMA
1841 for (j = MAX_NR_ZONES-1; j >= 0; j--) {
1845 * We never protect zones that don't have memory
1846 * in them (j>max_zone) or zones that aren't in
1847 * the zonelists for a certain type of
1848 * allocation (j>i). We have to assign these to
1849 * zero because the lower zones take
1850 * contributions from the higher zones.
1852 if (j > max_zone || j > i) {
1853 zone->protection[i] = 0;
1857 * The contribution of the next higher zone
1859 zone->protection[i] = higherzone_val(zone,
1861 zone->protection[i] += zone->pages_low;
1868 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
1869 * that the pages_{min,low,high} values for each zone are set correctly
1870 * with respect to min_free_kbytes.
1872 static void setup_per_zone_pages_min(void)
1874 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
1875 unsigned long lowmem_pages = 0;
1877 unsigned long flags;
1879 /* Calculate total number of !ZONE_HIGHMEM pages */
1880 for_each_zone(zone) {
1881 if (!is_highmem(zone))
1882 lowmem_pages += zone->present_pages;
1885 for_each_zone(zone) {
1886 spin_lock_irqsave(&zone->lru_lock, flags);
1887 if (is_highmem(zone)) {
1889 * Often, highmem doesn't need to reserve any pages.
1890 * But the pages_min/low/high values are also used for
1891 * batching up page reclaim activity so we need a
1892 * decent value here.
1896 min_pages = zone->present_pages / 1024;
1897 if (min_pages < SWAP_CLUSTER_MAX)
1898 min_pages = SWAP_CLUSTER_MAX;
1899 if (min_pages > 128)
1901 zone->pages_min = min_pages;
1903 /* if it's a lowmem zone, reserve a number of pages
1904 * proportionate to the zone's size.
1906 zone->pages_min = (pages_min * zone->present_pages) /
1910 zone->pages_low = zone->pages_min * 2;
1911 zone->pages_high = zone->pages_min * 3;
1912 spin_unlock_irqrestore(&zone->lru_lock, flags);
1917 * Initialise min_free_kbytes.
1919 * For small machines we want it small (128k min). For large machines
1920 * we want it large (16MB max). But it is not linear, because network
1921 * bandwidth does not increase linearly with machine size. We use
1923 * min_free_kbytes = sqrt(lowmem_kbytes)
1939 static int __init init_per_zone_pages_min(void)
1941 unsigned long lowmem_kbytes;
1943 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
1945 min_free_kbytes = int_sqrt(lowmem_kbytes);
1946 if (min_free_kbytes < 128)
1947 min_free_kbytes = 128;
1948 if (min_free_kbytes > 16384)
1949 min_free_kbytes = 16384;
1950 setup_per_zone_pages_min();
1951 setup_per_zone_protection();
1954 module_init(init_per_zone_pages_min)
1957 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
1958 * that we can call two helper functions whenever min_free_kbytes
1961 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
1962 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
1964 proc_dointvec(table, write, file, buffer, length, ppos);
1965 setup_per_zone_pages_min();
1966 setup_per_zone_protection();
1971 * lower_zone_protection_sysctl_handler - just a wrapper around
1972 * proc_dointvec() so that we can call setup_per_zone_protection()
1973 * whenever sysctl_lower_zone_protection changes.
1975 int lower_zone_protection_sysctl_handler(ctl_table *table, int write,
1976 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
1978 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
1979 setup_per_zone_protection();
1984 * allocate a large system hash table from bootmem
1985 * - it is assumed that the hash table must contain an exact power-of-2
1986 * quantity of entries
1988 void *__init alloc_large_system_hash(const char *tablename,
1989 unsigned long bucketsize,
1990 unsigned long numentries,
1992 int consider_highmem,
1993 unsigned int *_hash_shift,
1994 unsigned int *_hash_mask)
1996 unsigned long mem, max, log2qty, size;
1999 /* round applicable memory size up to nearest megabyte */
2000 mem = consider_highmem ? nr_all_pages : nr_kernel_pages;
2001 mem += (1UL << (20 - PAGE_SHIFT)) - 1;
2002 mem >>= 20 - PAGE_SHIFT;
2003 mem <<= 20 - PAGE_SHIFT;
2005 /* limit to 1 bucket per 2^scale bytes of low memory (rounded up to
2006 * nearest power of 2 in size) */
2007 if (scale > PAGE_SHIFT)
2008 mem >>= (scale - PAGE_SHIFT);
2010 mem <<= (PAGE_SHIFT - scale);
2012 mem = 1UL << (long_log2(mem) + 1);
2014 /* limit allocation size */
2015 max = (1UL << (PAGE_SHIFT + MAX_SYS_HASH_TABLE_ORDER)) / bucketsize;
2019 /* allow the kernel cmdline to have a say */
2020 if (!numentries || numentries > max)
2023 log2qty = long_log2(numentries);
2026 size = bucketsize << log2qty;
2028 table = (void *) alloc_bootmem(size);
2030 } while (!table && size > PAGE_SIZE);
2033 panic("Failed to allocate %s hash table\n", tablename);
2035 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2038 long_log2(size) - PAGE_SHIFT,
2042 *_hash_shift = log2qty;
2044 *_hash_mask = (1 << log2qty) - 1;