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 unsigned long __initdata nr_kernel_pages;
61 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%0*lx mapping:%p mapcount:%d count:%d\n",
82 (int)(2*sizeof(page_flags_t)), (unsigned long)page->flags,
83 page->mapping, page_mapcount(page), 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 |
94 set_page_count(page, 0);
95 reset_page_mapcount(page);
99 #ifndef CONFIG_HUGETLB_PAGE
100 #define prep_compound_page(page, order) do { } while (0)
101 #define destroy_compound_page(page, order) do { } while (0)
104 * Higher-order pages are called "compound pages". They are structured thusly:
106 * The first PAGE_SIZE page is called the "head page".
108 * The remaining PAGE_SIZE pages are called "tail pages".
110 * All pages have PG_compound set. All pages have their ->private pointing at
111 * the head page (even the head page has this).
113 * The first tail page's ->mapping, if non-zero, holds the address of the
114 * compound page's put_page() function.
116 * The order of the allocation is stored in the first tail page's ->index
117 * This is only for debug at present. This usage means that zero-order pages
118 * may not be compound.
120 static void prep_compound_page(struct page *page, unsigned long order)
123 int nr_pages = 1 << order;
125 page[1].mapping = NULL;
126 page[1].index = order;
127 for (i = 0; i < nr_pages; i++) {
128 struct page *p = page + i;
131 p->private = (unsigned long)page;
135 static void destroy_compound_page(struct page *page, unsigned long order)
138 int nr_pages = 1 << order;
140 if (!PageCompound(page))
143 if (page[1].index != order)
144 bad_page(__FUNCTION__, page);
146 for (i = 0; i < nr_pages; i++) {
147 struct page *p = page + i;
149 if (!PageCompound(p))
150 bad_page(__FUNCTION__, page);
151 if (p->private != (unsigned long)page)
152 bad_page(__FUNCTION__, page);
153 ClearPageCompound(p);
156 #endif /* CONFIG_HUGETLB_PAGE */
159 * Freeing function for a buddy system allocator.
161 * The concept of a buddy system is to maintain direct-mapped table
162 * (containing bit values) for memory blocks of various "orders".
163 * The bottom level table contains the map for the smallest allocatable
164 * units of memory (here, pages), and each level above it describes
165 * pairs of units from the levels below, hence, "buddies".
166 * At a high level, all that happens here is marking the table entry
167 * at the bottom level available, and propagating the changes upward
168 * as necessary, plus some accounting needed to play nicely with other
169 * parts of the VM system.
170 * At each level, we keep one bit for each pair of blocks, which
171 * is set to 1 iff only one of the pair is allocated. So when we
172 * are allocating or freeing one, we can derive the state of the
173 * other. That is, if we allocate a small block, and both were
174 * free, the remainder of the region must be split into blocks.
175 * If a block is freed, and its buddy is also free, then this
176 * triggers coalescing into a block of larger size.
181 static inline void __free_pages_bulk (struct page *page, struct page *base,
182 struct zone *zone, struct free_area *area, unsigned int order)
184 unsigned long page_idx, index, mask;
187 destroy_compound_page(page, order);
188 mask = (~0UL) << order;
189 page_idx = page - base;
190 if (page_idx & ~mask)
192 index = page_idx >> (1 + order);
194 zone->free_pages += 1 << order;
195 while (order < MAX_ORDER-1) {
196 struct page *buddy1, *buddy2;
198 BUG_ON(area >= zone->free_area + MAX_ORDER);
199 if (!__test_and_change_bit(index, area->map))
201 * the buddy page is still allocated.
205 /* Move the buddy up one level. */
206 buddy1 = base + (page_idx ^ (1 << order));
207 buddy2 = base + page_idx;
208 BUG_ON(bad_range(zone, buddy1));
209 BUG_ON(bad_range(zone, buddy2));
210 list_del(&buddy1->lru);
217 list_add(&(base + page_idx)->lru, &area->free_list);
220 static inline void free_pages_check(const char *function, struct page *page)
222 if ( page_mapped(page) ||
223 page->mapping != NULL ||
224 page_count(page) != 0 ||
233 1 << PG_writeback )))
234 bad_page(function, page);
236 ClearPageDirty(page);
240 * Frees a list of pages.
241 * Assumes all pages on list are in same zone, and of same order.
242 * count is the number of pages to free, or 0 for all on the list.
244 * If the zone was previously in an "all pages pinned" state then look to
245 * see if this freeing clears that state.
247 * And clear the zone's pages_scanned counter, to hold off the "all pages are
248 * pinned" detection logic.
251 free_pages_bulk(struct zone *zone, int count,
252 struct list_head *list, unsigned int order)
255 struct free_area *area;
256 struct page *base, *page = NULL;
259 base = zone->zone_mem_map;
260 area = zone->free_area + order;
261 spin_lock_irqsave(&zone->lock, flags);
262 zone->all_unreclaimable = 0;
263 zone->pages_scanned = 0;
264 while (!list_empty(list) && count--) {
265 page = list_entry(list->prev, struct page, lru);
266 /* have to delete it as __free_pages_bulk list manipulates */
267 list_del(&page->lru);
268 __free_pages_bulk(page, base, zone, area, order);
271 spin_unlock_irqrestore(&zone->lock, flags);
275 void __free_pages_ok(struct page *page, unsigned int order)
280 arch_free_page(page, order);
282 mod_page_state(pgfree, 1 << order);
283 for (i = 0 ; i < (1 << order) ; ++i)
284 free_pages_check(__FUNCTION__, page + i);
285 list_add(&page->lru, &list);
286 kernel_map_pages(page, 1<<order, 0);
287 free_pages_bulk(page_zone(page), 1, &list, order);
290 #define MARK_USED(index, order, area) \
291 __change_bit((index) >> (1+(order)), (area)->map)
294 * The order of subdivision here is critical for the IO subsystem.
295 * Please do not alter this order without good reasons and regression
296 * testing. Specifically, as large blocks of memory are subdivided,
297 * the order in which smaller blocks are delivered depends on the order
298 * they're subdivided in this function. This is the primary factor
299 * influencing the order in which pages are delivered to the IO
300 * subsystem according to empirical testing, and this is also justified
301 * by considering the behavior of a buddy system containing a single
302 * large block of memory acted on by a series of small allocations.
303 * This behavior is a critical factor in sglist merging's success.
307 static inline struct page *
308 expand(struct zone *zone, struct page *page,
309 unsigned long index, int low, int high, struct free_area *area)
311 unsigned long size = 1 << high;
317 BUG_ON(bad_range(zone, &page[size]));
318 list_add(&page[size].lru, &area->free_list);
319 MARK_USED(index + size, high, area);
324 static inline void set_page_refs(struct page *page, int order)
327 set_page_count(page, 1);
332 * We need to reference all the pages for this order, otherwise if
333 * anyone accesses one of the pages with (get/put) it will be freed.
335 for (i = 0; i < (1 << order); i++)
336 set_page_count(page+i, 1);
337 #endif /* CONFIG_MMU */
341 * This page is about to be returned from the page allocator
343 static void prep_new_page(struct page *page, int order)
345 if (page->mapping || page_mapped(page) ||
354 1 << PG_writeback )))
355 bad_page(__FUNCTION__, page);
357 page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
358 1 << PG_referenced | 1 << PG_arch_1 |
359 1 << PG_checked | 1 << PG_mappedtodisk);
361 set_page_refs(page, order);
365 * Do the hard work of removing an element from the buddy allocator.
366 * Call me with the zone->lock already held.
368 static struct page *__rmqueue(struct zone *zone, unsigned int order)
370 struct free_area * area;
371 unsigned int current_order;
375 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
376 area = zone->free_area + current_order;
377 if (list_empty(&area->free_list))
380 page = list_entry(area->free_list.next, struct page, lru);
381 list_del(&page->lru);
382 index = page - zone->zone_mem_map;
383 if (current_order != MAX_ORDER-1)
384 MARK_USED(index, current_order, area);
385 zone->free_pages -= 1UL << order;
386 return expand(zone, page, index, order, current_order, area);
393 * Obtain a specified number of elements from the buddy allocator, all under
394 * a single hold of the lock, for efficiency. Add them to the supplied list.
395 * Returns the number of new pages which were placed at *list.
397 static int rmqueue_bulk(struct zone *zone, unsigned int order,
398 unsigned long count, struct list_head *list)
405 spin_lock_irqsave(&zone->lock, flags);
406 for (i = 0; i < count; ++i) {
407 page = __rmqueue(zone, order);
411 list_add_tail(&page->lru, list);
413 spin_unlock_irqrestore(&zone->lock, flags);
417 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
418 static void __drain_pages(unsigned int cpu)
423 for_each_zone(zone) {
424 struct per_cpu_pageset *pset;
426 pset = &zone->pageset[cpu];
427 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
428 struct per_cpu_pages *pcp;
431 pcp->count -= free_pages_bulk(zone, pcp->count,
436 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
439 int is_head_of_free_region(struct page *page)
441 struct zone *zone = page_zone(page);
444 struct list_head *curr;
447 * Should not matter as we need quiescent system for
448 * suspend anyway, but...
450 spin_lock_irqsave(&zone->lock, flags);
451 for (order = MAX_ORDER - 1; order >= 0; --order)
452 list_for_each(curr, &zone->free_area[order].free_list)
453 if (page == list_entry(curr, struct page, lru)) {
454 spin_unlock_irqrestore(&zone->lock, flags);
457 spin_unlock_irqrestore(&zone->lock, flags);
462 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
464 void drain_local_pages(void)
468 local_irq_save(flags);
469 __drain_pages(smp_processor_id());
470 local_irq_restore(flags);
472 #endif /* CONFIG_PM */
474 static void zone_statistics(struct zonelist *zonelist, struct zone *z)
479 pg_data_t *pg = z->zone_pgdat;
480 pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
481 struct per_cpu_pageset *p;
483 local_irq_save(flags);
484 cpu = smp_processor_id();
485 p = &z->pageset[cpu];
487 z->pageset[cpu].numa_hit++;
490 zonelist->zones[0]->pageset[cpu].numa_foreign++;
492 if (pg == NODE_DATA(numa_node_id()))
496 local_irq_restore(flags);
501 * Free a 0-order page
503 static void FASTCALL(free_hot_cold_page(struct page *page, int cold));
504 static void fastcall free_hot_cold_page(struct page *page, int cold)
506 struct zone *zone = page_zone(page);
507 struct per_cpu_pages *pcp;
510 arch_free_page(page, 0);
512 kernel_map_pages(page, 1, 0);
513 inc_page_state(pgfree);
515 page->mapping = NULL;
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;
605 struct zone **zones, *z;
607 struct reclaim_state reclaim_state;
608 struct task_struct *p = current;
614 might_sleep_if(wait);
617 * The caller may dip into page reserves a bit more if the caller
618 * cannot run direct reclaim, or is the caller has realtime scheduling
621 can_try_harder = (unlikely(rt_task(p)) && !in_interrupt()) || !wait;
623 zones = zonelist->zones; /* the list of zones suitable for gfp_mask */
625 if (unlikely(zones[0] == NULL)) {
626 /* Should this ever happen?? */
630 alloc_type = zone_idx(zones[0]);
632 /* Go through the zonelist once, looking for a zone with enough free */
633 for (i = 0; (z = zones[i]) != NULL; i++) {
634 min = z->pages_low + (1<<order) + z->protection[alloc_type];
636 if (z->free_pages < min)
639 page = buffered_rmqueue(z, order, gfp_mask);
644 for (i = 0; (z = zones[i]) != NULL; i++)
648 * Go through the zonelist again. Let __GFP_HIGH and allocations
649 * coming from realtime tasks to go deeper into reserves
651 for (i = 0; (z = zones[i]) != NULL; i++) {
653 if (gfp_mask & __GFP_HIGH)
657 min += (1<<order) + z->protection[alloc_type];
659 if (z->free_pages < min)
662 page = buffered_rmqueue(z, order, gfp_mask);
667 /* This allocation should allow future memory freeing. */
668 if ((p->flags & (PF_MEMALLOC | PF_MEMDIE)) && !in_interrupt()) {
669 /* go through the zonelist yet again, ignoring mins */
670 for (i = 0; (z = zones[i]) != NULL; i++) {
671 page = buffered_rmqueue(z, order, gfp_mask);
678 /* Atomic allocations - we can't balance anything */
683 /* We now go into synchronous reclaim */
684 p->flags |= PF_MEMALLOC;
685 reclaim_state.reclaimed_slab = 0;
686 p->reclaim_state = &reclaim_state;
688 try_to_free_pages(zones, gfp_mask, order);
690 p->reclaim_state = NULL;
691 p->flags &= ~PF_MEMALLOC;
693 /* go through the zonelist yet one more time */
694 for (i = 0; (z = zones[i]) != NULL; i++) {
696 if (gfp_mask & __GFP_HIGH)
700 min += (1<<order) + z->protection[alloc_type];
702 if (z->free_pages < min)
705 page = buffered_rmqueue(z, order, gfp_mask);
711 * Don't let big-order allocations loop unless the caller explicitly
712 * requests that. Wait for some write requests to complete then retry.
714 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
715 * <= 3, but that may not be true in other implementations.
718 if (!(gfp_mask & __GFP_NORETRY)) {
719 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
721 if (gfp_mask & __GFP_NOFAIL)
725 blk_congestion_wait(WRITE, HZ/50);
730 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
731 printk(KERN_WARNING "%s: page allocation failure."
732 " order:%d, mode:0x%x\n",
733 p->comm, order, gfp_mask);
738 zone_statistics(zonelist, z);
739 kernel_map_pages(page, 1 << order, 1);
743 EXPORT_SYMBOL(__alloc_pages);
746 * Common helper functions.
748 fastcall unsigned long __get_free_pages(unsigned int gfp_mask, unsigned int order)
751 page = alloc_pages(gfp_mask, order);
754 return (unsigned long) page_address(page);
757 EXPORT_SYMBOL(__get_free_pages);
759 fastcall unsigned long get_zeroed_page(unsigned int gfp_mask)
764 * get_zeroed_page() returns a 32-bit address, which cannot represent
767 BUG_ON(gfp_mask & __GFP_HIGHMEM);
769 page = alloc_pages(gfp_mask, 0);
771 void *address = page_address(page);
773 return (unsigned long) address;
778 EXPORT_SYMBOL(get_zeroed_page);
780 void __pagevec_free(struct pagevec *pvec)
782 int i = pagevec_count(pvec);
785 free_hot_cold_page(pvec->pages[i], pvec->cold);
788 fastcall void __free_pages(struct page *page, unsigned int order)
790 if (!PageReserved(page) && put_page_testzero(page)) {
794 __free_pages_ok(page, order);
798 EXPORT_SYMBOL(__free_pages);
800 fastcall void free_pages(unsigned long addr, unsigned int order)
803 BUG_ON(!virt_addr_valid((void *)addr));
804 __free_pages(virt_to_page((void *)addr), order);
808 EXPORT_SYMBOL(free_pages);
811 * Total amount of free (allocatable) RAM:
813 unsigned int nr_free_pages(void)
815 unsigned int sum = 0;
819 sum += zone->free_pages;
824 EXPORT_SYMBOL(nr_free_pages);
827 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
829 unsigned int i, sum = 0;
831 for (i = 0; i < MAX_NR_ZONES; i++)
832 sum += pgdat->node_zones[i].free_pages;
838 static unsigned int nr_free_zone_pages(int offset)
841 unsigned int sum = 0;
843 for_each_pgdat(pgdat) {
844 struct zonelist *zonelist = pgdat->node_zonelists + offset;
845 struct zone **zonep = zonelist->zones;
848 for (zone = *zonep++; zone; zone = *zonep++) {
849 unsigned long size = zone->present_pages;
850 unsigned long high = zone->pages_high;
860 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
862 unsigned int nr_free_buffer_pages(void)
864 return nr_free_zone_pages(GFP_USER & GFP_ZONEMASK);
868 * Amount of free RAM allocatable within all zones
870 unsigned int nr_free_pagecache_pages(void)
872 return nr_free_zone_pages(GFP_HIGHUSER & GFP_ZONEMASK);
875 #ifdef CONFIG_HIGHMEM
876 unsigned int nr_free_highpages (void)
879 unsigned int pages = 0;
881 for_each_pgdat(pgdat)
882 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
889 static void show_node(struct zone *zone)
891 printk("Node %d ", zone->zone_pgdat->node_id);
894 #define show_node(zone) do { } while (0)
898 * Accumulate the page_state information across all CPUs.
899 * The result is unavoidably approximate - it can change
900 * during and after execution of this function.
902 DEFINE_PER_CPU(struct page_state, page_states) = {0};
903 EXPORT_PER_CPU_SYMBOL(page_states);
905 atomic_t nr_pagecache = ATOMIC_INIT(0);
906 EXPORT_SYMBOL(nr_pagecache);
908 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
911 void __get_page_state(struct page_state *ret, int nr)
915 memset(ret, 0, sizeof(*ret));
916 while (cpu < NR_CPUS) {
917 unsigned long *in, *out, off;
919 if (!cpu_possible(cpu)) {
924 in = (unsigned long *)&per_cpu(page_states, cpu);
926 if (cpu < NR_CPUS && cpu_possible(cpu))
927 prefetch(&per_cpu(page_states, cpu));
928 out = (unsigned long *)ret;
929 for (off = 0; off < nr; off++)
934 void get_page_state(struct page_state *ret)
938 nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
939 nr /= sizeof(unsigned long);
941 __get_page_state(ret, nr + 1);
944 void get_full_page_state(struct page_state *ret)
946 __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long));
949 unsigned long __read_page_state(unsigned offset)
951 unsigned long ret = 0;
954 for (cpu = 0; cpu < NR_CPUS; cpu++) {
957 if (!cpu_possible(cpu))
960 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
961 ret += *((unsigned long *)in);
966 void __get_zone_counts(unsigned long *active, unsigned long *inactive,
967 unsigned long *free, struct pglist_data *pgdat)
969 struct zone *zones = pgdat->node_zones;
975 for (i = 0; i < MAX_NR_ZONES; i++) {
976 *active += zones[i].nr_active;
977 *inactive += zones[i].nr_inactive;
978 *free += zones[i].free_pages;
982 void get_zone_counts(unsigned long *active,
983 unsigned long *inactive, unsigned long *free)
985 struct pglist_data *pgdat;
990 for_each_pgdat(pgdat) {
991 unsigned long l, m, n;
992 __get_zone_counts(&l, &m, &n, pgdat);
999 void si_meminfo(struct sysinfo *val)
1001 val->totalram = totalram_pages;
1003 val->freeram = nr_free_pages();
1004 val->bufferram = nr_blockdev_pages();
1005 #ifdef CONFIG_HIGHMEM
1006 val->totalhigh = totalhigh_pages;
1007 val->freehigh = nr_free_highpages();
1012 val->mem_unit = PAGE_SIZE;
1013 if (vx_flags(VXF_VIRT_MEM, 0))
1014 vx_vsi_meminfo(val);
1017 EXPORT_SYMBOL(si_meminfo);
1020 void si_meminfo_node(struct sysinfo *val, int nid)
1022 pg_data_t *pgdat = NODE_DATA(nid);
1024 val->totalram = pgdat->node_present_pages;
1025 val->freeram = nr_free_pages_pgdat(pgdat);
1026 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1027 val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1028 val->mem_unit = PAGE_SIZE;
1032 #define K(x) ((x) << (PAGE_SHIFT-10))
1035 * Show free area list (used inside shift_scroll-lock stuff)
1036 * We also calculate the percentage fragmentation. We do this by counting the
1037 * memory on each free list with the exception of the first item on the list.
1039 void show_free_areas(void)
1041 struct page_state ps;
1042 int cpu, temperature;
1043 unsigned long active;
1044 unsigned long inactive;
1048 for_each_zone(zone) {
1050 printk("%s per-cpu:", zone->name);
1052 if (!zone->present_pages) {
1058 for (cpu = 0; cpu < NR_CPUS; ++cpu) {
1059 struct per_cpu_pageset *pageset;
1061 if (!cpu_possible(cpu))
1064 pageset = zone->pageset + cpu;
1066 for (temperature = 0; temperature < 2; temperature++)
1067 printk("cpu %d %s: low %d, high %d, batch %d\n",
1069 temperature ? "cold" : "hot",
1070 pageset->pcp[temperature].low,
1071 pageset->pcp[temperature].high,
1072 pageset->pcp[temperature].batch);
1076 get_page_state(&ps);
1077 get_zone_counts(&active, &inactive, &free);
1079 printk("\nFree pages: %11ukB (%ukB HighMem)\n",
1081 K(nr_free_highpages()));
1083 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1084 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1093 ps.nr_page_table_pages);
1095 for_each_zone(zone) {
1109 K(zone->free_pages),
1112 K(zone->pages_high),
1114 K(zone->nr_inactive),
1115 K(zone->present_pages)
1117 printk("protections[]:");
1118 for (i = 0; i < MAX_NR_ZONES; i++)
1119 printk(" %lu", zone->protection[i]);
1123 for_each_zone(zone) {
1124 struct list_head *elem;
1125 unsigned long nr, flags, order, total = 0;
1128 printk("%s: ", zone->name);
1129 if (!zone->present_pages) {
1134 spin_lock_irqsave(&zone->lock, flags);
1135 for (order = 0; order < MAX_ORDER; order++) {
1137 list_for_each(elem, &zone->free_area[order].free_list)
1139 total += nr << order;
1140 printk("%lu*%lukB ", nr, K(1UL) << order);
1142 spin_unlock_irqrestore(&zone->lock, flags);
1143 printk("= %lukB\n", K(total));
1146 show_swap_cache_info();
1150 * Builds allocation fallback zone lists.
1152 static int __init build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, int j, int k)
1159 zone = pgdat->node_zones + ZONE_HIGHMEM;
1160 if (zone->present_pages) {
1161 #ifndef CONFIG_HIGHMEM
1164 zonelist->zones[j++] = zone;
1167 zone = pgdat->node_zones + ZONE_NORMAL;
1168 if (zone->present_pages)
1169 zonelist->zones[j++] = zone;
1171 zone = pgdat->node_zones + ZONE_DMA;
1172 if (zone->present_pages)
1173 zonelist->zones[j++] = zone;
1180 #define MAX_NODE_LOAD (numnodes)
1181 static int __initdata node_load[MAX_NUMNODES];
1183 * find_next_best_node - find the next node that should appear in a given
1184 * node's fallback list
1185 * @node: node whose fallback list we're appending
1186 * @used_node_mask: pointer to the bitmap of already used nodes
1188 * We use a number of factors to determine which is the next node that should
1189 * appear on a given node's fallback list. The node should not have appeared
1190 * already in @node's fallback list, and it should be the next closest node
1191 * according to the distance array (which contains arbitrary distance values
1192 * from each node to each node in the system), and should also prefer nodes
1193 * with no CPUs, since presumably they'll have very little allocation pressure
1194 * on them otherwise.
1195 * It returns -1 if no node is found.
1197 static int __init find_next_best_node(int node, void *used_node_mask)
1200 int min_val = INT_MAX;
1203 for (i = 0; i < numnodes; i++) {
1206 /* Start from local node */
1207 n = (node+i)%numnodes;
1209 /* Don't want a node to appear more than once */
1210 if (test_bit(n, used_node_mask))
1213 /* Use the distance array to find the distance */
1214 val = node_distance(node, n);
1216 /* Give preference to headless and unused nodes */
1217 tmp = node_to_cpumask(n);
1218 if (!cpus_empty(tmp))
1219 val += PENALTY_FOR_NODE_WITH_CPUS;
1221 /* Slight preference for less loaded node */
1222 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1223 val += node_load[n];
1225 if (val < min_val) {
1232 set_bit(best_node, used_node_mask);
1237 static void __init build_zonelists(pg_data_t *pgdat)
1239 int i, j, k, node, local_node;
1240 int prev_node, load;
1241 struct zonelist *zonelist;
1242 DECLARE_BITMAP(used_mask, MAX_NUMNODES);
1244 /* initialize zonelists */
1245 for (i = 0; i < GFP_ZONETYPES; i++) {
1246 zonelist = pgdat->node_zonelists + i;
1247 memset(zonelist, 0, sizeof(*zonelist));
1248 zonelist->zones[0] = NULL;
1251 /* NUMA-aware ordering of nodes */
1252 local_node = pgdat->node_id;
1254 prev_node = local_node;
1255 bitmap_zero(used_mask, MAX_NUMNODES);
1256 while ((node = find_next_best_node(local_node, used_mask)) >= 0) {
1258 * We don't want to pressure a particular node.
1259 * So adding penalty to the first node in same
1260 * distance group to make it round-robin.
1262 if (node_distance(local_node, node) !=
1263 node_distance(local_node, prev_node))
1264 node_load[node] += load;
1267 for (i = 0; i < GFP_ZONETYPES; i++) {
1268 zonelist = pgdat->node_zonelists + i;
1269 for (j = 0; zonelist->zones[j] != NULL; j++);
1272 if (i & __GFP_HIGHMEM)
1277 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1278 zonelist->zones[j] = NULL;
1283 #else /* CONFIG_NUMA */
1285 static void __init build_zonelists(pg_data_t *pgdat)
1287 int i, j, k, node, local_node;
1289 local_node = pgdat->node_id;
1290 for (i = 0; i < GFP_ZONETYPES; i++) {
1291 struct zonelist *zonelist;
1293 zonelist = pgdat->node_zonelists + i;
1294 memset(zonelist, 0, sizeof(*zonelist));
1298 if (i & __GFP_HIGHMEM)
1303 j = build_zonelists_node(pgdat, zonelist, j, k);
1305 * Now we build the zonelist so that it contains the zones
1306 * of all the other nodes.
1307 * We don't want to pressure a particular node, so when
1308 * building the zones for node N, we make sure that the
1309 * zones coming right after the local ones are those from
1310 * node N+1 (modulo N)
1312 for (node = local_node + 1; node < numnodes; node++)
1313 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1314 for (node = 0; node < local_node; node++)
1315 j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1317 zonelist->zones[j] = NULL;
1321 #endif /* CONFIG_NUMA */
1323 void __init build_all_zonelists(void)
1327 for(i = 0 ; i < numnodes ; i++)
1328 build_zonelists(NODE_DATA(i));
1329 printk("Built %i zonelists\n", numnodes);
1333 * Helper functions to size the waitqueue hash table.
1334 * Essentially these want to choose hash table sizes sufficiently
1335 * large so that collisions trying to wait on pages are rare.
1336 * But in fact, the number of active page waitqueues on typical
1337 * systems is ridiculously low, less than 200. So this is even
1338 * conservative, even though it seems large.
1340 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1341 * waitqueues, i.e. the size of the waitq table given the number of pages.
1343 #define PAGES_PER_WAITQUEUE 256
1345 static inline unsigned long wait_table_size(unsigned long pages)
1347 unsigned long size = 1;
1349 pages /= PAGES_PER_WAITQUEUE;
1351 while (size < pages)
1355 * Once we have dozens or even hundreds of threads sleeping
1356 * on IO we've got bigger problems than wait queue collision.
1357 * Limit the size of the wait table to a reasonable size.
1359 size = min(size, 4096UL);
1361 return max(size, 4UL);
1365 * This is an integer logarithm so that shifts can be used later
1366 * to extract the more random high bits from the multiplicative
1367 * hash function before the remainder is taken.
1369 static inline unsigned long wait_table_bits(unsigned long size)
1374 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1376 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1377 unsigned long *zones_size, unsigned long *zholes_size)
1379 unsigned long realtotalpages, totalpages = 0;
1382 for (i = 0; i < MAX_NR_ZONES; i++)
1383 totalpages += zones_size[i];
1384 pgdat->node_spanned_pages = totalpages;
1386 realtotalpages = totalpages;
1388 for (i = 0; i < MAX_NR_ZONES; i++)
1389 realtotalpages -= zholes_size[i];
1390 pgdat->node_present_pages = realtotalpages;
1391 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1396 * Initially all pages are reserved - free ones are freed
1397 * up by free_all_bootmem() once the early boot process is
1398 * done. Non-atomic initialization, single-pass.
1400 void __init memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1401 unsigned long start_pfn)
1403 struct page *start = pfn_to_page(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 reset_page_mapcount(page);
1410 SetPageReserved(page);
1411 INIT_LIST_HEAD(&page->lru);
1412 #ifdef WANT_PAGE_VIRTUAL
1413 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1414 if (!is_highmem_idx(zone))
1415 set_page_address(page, __va(start_pfn << PAGE_SHIFT));
1422 * Page buddy system uses "index >> (i+1)", where "index" is
1425 * The extra "+3" is to round down to byte size (8 bits per byte
1426 * assumption). Thus we get "(size-1) >> (i+4)" as the last byte
1429 * The "+1" is because we want to round the byte allocation up
1430 * rather than down. So we should have had a "+7" before we shifted
1431 * down by three. Also, we have to add one as we actually _use_ the
1432 * last bit (it's [0,n] inclusive, not [0,n[).
1434 * So we actually had +7+1 before we shift down by 3. But
1435 * (n+8) >> 3 == (n >> 3) + 1 (modulo overflows, which we do not have).
1437 * Finally, we LONG_ALIGN because all bitmap operations are on longs.
1439 unsigned long pages_to_bitmap_size(unsigned long order, unsigned long nr_pages)
1441 unsigned long bitmap_size;
1443 bitmap_size = (nr_pages-1) >> (order+4);
1444 bitmap_size = LONG_ALIGN(bitmap_size+1);
1449 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone, unsigned long size)
1452 for (order = 0; ; order++) {
1453 unsigned long bitmap_size;
1455 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1456 if (order == MAX_ORDER-1) {
1457 zone->free_area[order].map = NULL;
1461 bitmap_size = pages_to_bitmap_size(order, size);
1462 zone->free_area[order].map =
1463 (unsigned long *) alloc_bootmem_node(pgdat, bitmap_size);
1467 #ifndef __HAVE_ARCH_MEMMAP_INIT
1468 #define memmap_init(size, nid, zone, start_pfn) \
1469 memmap_init_zone((size), (nid), (zone), (start_pfn))
1473 * Set up the zone data structures:
1474 * - mark all pages reserved
1475 * - mark all memory queues empty
1476 * - clear the memory bitmaps
1478 static void __init free_area_init_core(struct pglist_data *pgdat,
1479 unsigned long *zones_size, unsigned long *zholes_size)
1482 const unsigned long zone_required_alignment = 1UL << (MAX_ORDER-1);
1483 int cpu, nid = pgdat->node_id;
1484 unsigned long zone_start_pfn = pgdat->node_start_pfn;
1486 pgdat->nr_zones = 0;
1487 init_waitqueue_head(&pgdat->kswapd_wait);
1489 for (j = 0; j < MAX_NR_ZONES; j++) {
1490 struct zone *zone = pgdat->node_zones + j;
1491 unsigned long size, realsize;
1492 unsigned long batch;
1494 zone_table[NODEZONE(nid, j)] = zone;
1495 realsize = size = zones_size[j];
1497 realsize -= zholes_size[j];
1499 if (j == ZONE_DMA || j == ZONE_NORMAL)
1500 nr_kernel_pages += realsize;
1501 nr_all_pages += realsize;
1503 zone->spanned_pages = size;
1504 zone->present_pages = realsize;
1505 zone->name = zone_names[j];
1506 spin_lock_init(&zone->lock);
1507 spin_lock_init(&zone->lru_lock);
1508 zone->zone_pgdat = pgdat;
1509 zone->free_pages = 0;
1511 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
1514 * The per-cpu-pages pools are set to around 1000th of the
1515 * size of the zone. But no more than 1/4 of a meg - there's
1516 * no point in going beyond the size of L2 cache.
1518 * OK, so we don't know how big the cache is. So guess.
1520 batch = zone->present_pages / 1024;
1521 if (batch * PAGE_SIZE > 256 * 1024)
1522 batch = (256 * 1024) / PAGE_SIZE;
1523 batch /= 4; /* We effectively *= 4 below */
1527 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1528 struct per_cpu_pages *pcp;
1530 pcp = &zone->pageset[cpu].pcp[0]; /* hot */
1532 pcp->low = 2 * batch;
1533 pcp->high = 6 * batch;
1534 pcp->batch = 1 * batch;
1535 INIT_LIST_HEAD(&pcp->list);
1537 pcp = &zone->pageset[cpu].pcp[1]; /* cold */
1540 pcp->high = 2 * batch;
1541 pcp->batch = 1 * batch;
1542 INIT_LIST_HEAD(&pcp->list);
1544 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n",
1545 zone_names[j], realsize, batch);
1546 INIT_LIST_HEAD(&zone->active_list);
1547 INIT_LIST_HEAD(&zone->inactive_list);
1548 zone->nr_scan_active = 0;
1549 zone->nr_scan_inactive = 0;
1550 zone->nr_active = 0;
1551 zone->nr_inactive = 0;
1556 * The per-page waitqueue mechanism uses hashed waitqueues
1559 zone->wait_table_size = wait_table_size(size);
1560 zone->wait_table_bits =
1561 wait_table_bits(zone->wait_table_size);
1562 zone->wait_table = (wait_queue_head_t *)
1563 alloc_bootmem_node(pgdat, zone->wait_table_size
1564 * sizeof(wait_queue_head_t));
1566 for(i = 0; i < zone->wait_table_size; ++i)
1567 init_waitqueue_head(zone->wait_table + i);
1569 pgdat->nr_zones = j+1;
1571 zone->zone_mem_map = pfn_to_page(zone_start_pfn);
1572 zone->zone_start_pfn = zone_start_pfn;
1574 if ((zone_start_pfn) & (zone_required_alignment-1))
1575 printk("BUG: wrong zone alignment, it will crash\n");
1577 memmap_init(size, nid, j, zone_start_pfn);
1579 zone_start_pfn += size;
1581 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1585 void __init node_alloc_mem_map(struct pglist_data *pgdat)
1589 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
1590 pgdat->node_mem_map = alloc_bootmem_node(pgdat, size);
1591 #ifndef CONFIG_DISCONTIGMEM
1592 mem_map = contig_page_data.node_mem_map;
1596 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
1597 unsigned long *zones_size, unsigned long node_start_pfn,
1598 unsigned long *zholes_size)
1600 pgdat->node_id = nid;
1601 pgdat->node_start_pfn = node_start_pfn;
1602 calculate_zone_totalpages(pgdat, zones_size, zholes_size);
1604 if (!pfn_to_page(node_start_pfn))
1605 node_alloc_mem_map(pgdat);
1607 free_area_init_core(pgdat, zones_size, zholes_size);
1610 #ifndef CONFIG_DISCONTIGMEM
1611 static bootmem_data_t contig_bootmem_data;
1612 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
1614 EXPORT_SYMBOL(contig_page_data);
1616 void __init free_area_init(unsigned long *zones_size)
1618 free_area_init_node(0, &contig_page_data, zones_size,
1619 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
1623 #ifdef CONFIG_PROC_FS
1625 #include <linux/seq_file.h>
1627 static void *frag_start(struct seq_file *m, loff_t *pos)
1632 for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
1638 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1640 pg_data_t *pgdat = (pg_data_t *)arg;
1643 return pgdat->pgdat_next;
1646 static void frag_stop(struct seq_file *m, void *arg)
1651 * This walks the freelist for each zone. Whilst this is slow, I'd rather
1652 * be slow here than slow down the fast path by keeping stats - mjbligh
1654 static int frag_show(struct seq_file *m, void *arg)
1656 pg_data_t *pgdat = (pg_data_t *)arg;
1658 struct zone *node_zones = pgdat->node_zones;
1659 unsigned long flags;
1662 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1663 if (!zone->present_pages)
1666 spin_lock_irqsave(&zone->lock, flags);
1667 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1668 for (order = 0; order < MAX_ORDER; ++order) {
1669 unsigned long nr_bufs = 0;
1670 struct list_head *elem;
1672 list_for_each(elem, &(zone->free_area[order].free_list))
1674 seq_printf(m, "%6lu ", nr_bufs);
1676 spin_unlock_irqrestore(&zone->lock, flags);
1682 struct seq_operations fragmentation_op = {
1683 .start = frag_start,
1689 static char *vmstat_text[] = {
1693 "nr_page_table_pages",
1718 "pgscan_kswapd_high",
1719 "pgscan_kswapd_normal",
1721 "pgscan_kswapd_dma",
1722 "pgscan_direct_high",
1723 "pgscan_direct_normal",
1724 "pgscan_direct_dma",
1729 "kswapd_inodesteal",
1736 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1738 struct page_state *ps;
1740 if (*pos >= ARRAY_SIZE(vmstat_text))
1743 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
1746 return ERR_PTR(-ENOMEM);
1747 get_full_page_state(ps);
1748 ps->pgpgin /= 2; /* sectors -> kbytes */
1750 return (unsigned long *)ps + *pos;
1753 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1756 if (*pos >= ARRAY_SIZE(vmstat_text))
1758 return (unsigned long *)m->private + *pos;
1761 static int vmstat_show(struct seq_file *m, void *arg)
1763 unsigned long *l = arg;
1764 unsigned long off = l - (unsigned long *)m->private;
1766 seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1770 static void vmstat_stop(struct seq_file *m, void *arg)
1776 struct seq_operations vmstat_op = {
1777 .start = vmstat_start,
1778 .next = vmstat_next,
1779 .stop = vmstat_stop,
1780 .show = vmstat_show,
1783 #endif /* CONFIG_PROC_FS */
1785 #ifdef CONFIG_HOTPLUG_CPU
1786 static int page_alloc_cpu_notify(struct notifier_block *self,
1787 unsigned long action, void *hcpu)
1789 int cpu = (unsigned long)hcpu;
1792 if (action == CPU_DEAD) {
1793 /* Drain local pagecache count. */
1794 count = &per_cpu(nr_pagecache_local, cpu);
1795 atomic_add(*count, &nr_pagecache);
1797 local_irq_disable();
1803 #endif /* CONFIG_HOTPLUG_CPU */
1805 void __init page_alloc_init(void)
1807 hotcpu_notifier(page_alloc_cpu_notify, 0);
1810 static unsigned long higherzone_val(struct zone *z, int max_zone,
1813 int z_idx = zone_idx(z);
1814 struct zone *higherzone;
1815 unsigned long pages;
1817 /* there is no higher zone to get a contribution from */
1818 if (z_idx == MAX_NR_ZONES-1)
1821 higherzone = &z->zone_pgdat->node_zones[z_idx+1];
1823 /* We always start with the higher zone's protection value */
1824 pages = higherzone->protection[alloc_type];
1827 * We get a lower-zone-protection contribution only if there are
1828 * pages in the higher zone and if we're not the highest zone
1829 * in the current zonelist. e.g., never happens for GFP_DMA. Happens
1830 * only for ZONE_DMA in a GFP_KERNEL allocation and happens for ZONE_DMA
1831 * and ZONE_NORMAL for a GFP_HIGHMEM allocation.
1833 if (higherzone->present_pages && z_idx < alloc_type)
1834 pages += higherzone->pages_low * sysctl_lower_zone_protection;
1840 * setup_per_zone_protection - called whenver min_free_kbytes or
1841 * sysctl_lower_zone_protection changes. Ensures that each zone
1842 * has a correct pages_protected value, so an adequate number of
1843 * pages are left in the zone after a successful __alloc_pages().
1845 * This algorithm is way confusing. I tries to keep the same behavior
1846 * as we had with the incremental min iterative algorithm.
1848 static void setup_per_zone_protection(void)
1850 struct pglist_data *pgdat;
1851 struct zone *zones, *zone;
1855 for_each_pgdat(pgdat) {
1856 zones = pgdat->node_zones;
1858 for (i = 0, max_zone = 0; i < MAX_NR_ZONES; i++)
1859 if (zones[i].present_pages)
1863 * For each of the different allocation types:
1864 * GFP_DMA -> GFP_KERNEL -> GFP_HIGHMEM
1866 for (i = 0; i < GFP_ZONETYPES; i++) {
1868 * For each of the zones:
1869 * ZONE_HIGHMEM -> ZONE_NORMAL -> ZONE_DMA
1871 for (j = MAX_NR_ZONES-1; j >= 0; j--) {
1875 * We never protect zones that don't have memory
1876 * in them (j>max_zone) or zones that aren't in
1877 * the zonelists for a certain type of
1878 * allocation (j>=i). We have to assign these
1879 * to zero because the lower zones take
1880 * contributions from the higher zones.
1882 if (j > max_zone || j >= i) {
1883 zone->protection[i] = 0;
1887 * The contribution of the next higher zone
1889 zone->protection[i] = higherzone_val(zone,
1897 * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures
1898 * that the pages_{min,low,high} values for each zone are set correctly
1899 * with respect to min_free_kbytes.
1901 static void setup_per_zone_pages_min(void)
1903 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
1904 unsigned long lowmem_pages = 0;
1906 unsigned long flags;
1908 /* Calculate total number of !ZONE_HIGHMEM pages */
1909 for_each_zone(zone) {
1910 if (!is_highmem(zone))
1911 lowmem_pages += zone->present_pages;
1914 for_each_zone(zone) {
1915 spin_lock_irqsave(&zone->lru_lock, flags);
1916 if (is_highmem(zone)) {
1918 * Often, highmem doesn't need to reserve any pages.
1919 * But the pages_min/low/high values are also used for
1920 * batching up page reclaim activity so we need a
1921 * decent value here.
1925 min_pages = zone->present_pages / 1024;
1926 if (min_pages < SWAP_CLUSTER_MAX)
1927 min_pages = SWAP_CLUSTER_MAX;
1928 if (min_pages > 128)
1930 zone->pages_min = min_pages;
1932 /* if it's a lowmem zone, reserve a number of pages
1933 * proportionate to the zone's size.
1935 zone->pages_min = (pages_min * zone->present_pages) /
1939 zone->pages_low = zone->pages_min * 2;
1940 zone->pages_high = zone->pages_min * 3;
1941 spin_unlock_irqrestore(&zone->lru_lock, flags);
1946 * Initialise min_free_kbytes.
1948 * For small machines we want it small (128k min). For large machines
1949 * we want it large (16MB max). But it is not linear, because network
1950 * bandwidth does not increase linearly with machine size. We use
1952 * min_free_kbytes = sqrt(lowmem_kbytes)
1968 static int __init init_per_zone_pages_min(void)
1970 unsigned long lowmem_kbytes;
1972 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
1974 min_free_kbytes = int_sqrt(lowmem_kbytes);
1975 if (min_free_kbytes < 128)
1976 min_free_kbytes = 128;
1977 if (min_free_kbytes > 16384)
1978 min_free_kbytes = 16384;
1979 setup_per_zone_pages_min();
1980 setup_per_zone_protection();
1983 module_init(init_per_zone_pages_min)
1986 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
1987 * that we can call two helper functions whenever min_free_kbytes
1990 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
1991 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
1993 proc_dointvec(table, write, file, buffer, length, ppos);
1994 setup_per_zone_pages_min();
1995 setup_per_zone_protection();
2000 * lower_zone_protection_sysctl_handler - just a wrapper around
2001 * proc_dointvec() so that we can call setup_per_zone_protection()
2002 * whenever sysctl_lower_zone_protection changes.
2004 int lower_zone_protection_sysctl_handler(ctl_table *table, int write,
2005 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2007 proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2008 setup_per_zone_protection();
2013 * allocate a large system hash table from bootmem
2014 * - it is assumed that the hash table must contain an exact power-of-2
2015 * quantity of entries
2017 void *__init alloc_large_system_hash(const char *tablename,
2018 unsigned long bucketsize,
2019 unsigned long numentries,
2021 int consider_highmem,
2022 unsigned int *_hash_shift,
2023 unsigned int *_hash_mask)
2025 unsigned long long max;
2026 unsigned long log2qty, size;
2029 /* allow the kernel cmdline to have a say */
2031 /* round applicable memory size up to nearest megabyte */
2032 numentries = consider_highmem ? nr_all_pages : nr_kernel_pages;
2033 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2034 numentries >>= 20 - PAGE_SHIFT;
2035 numentries <<= 20 - PAGE_SHIFT;
2037 /* limit to 1 bucket per 2^scale bytes of low memory */
2038 if (scale > PAGE_SHIFT)
2039 numentries >>= (scale - PAGE_SHIFT);
2041 numentries <<= (PAGE_SHIFT - scale);
2043 /* rounded up to nearest power of 2 in size */
2044 numentries = 1UL << (long_log2(numentries) + 1);
2046 /* limit allocation size to 1/16 total memory */
2047 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2048 do_div(max, bucketsize);
2050 if (numentries > max)
2053 log2qty = long_log2(numentries);
2056 size = bucketsize << log2qty;
2057 table = alloc_bootmem(size);
2058 } while (!table && size > PAGE_SIZE && --log2qty);
2061 panic("Failed to allocate %s hash table\n", tablename);
2063 printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2066 long_log2(size) - PAGE_SHIFT,
2070 *_hash_shift = log2qty;
2072 *_hash_mask = (1 << log2qty) - 1;