Make vm statistics update interval configurable
[linux-2.6.git] / mm / vmstat.c
1 /*
2  *  linux/mm/vmstat.c
3  *
4  *  Manages VM statistics
5  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
6  *
7  *  zoned VM statistics
8  *  Copyright (C) 2006 Silicon Graphics, Inc.,
9  *              Christoph Lameter <christoph@lameter.com>
10  */
11
12 #include <linux/mm.h>
13 #include <linux/module.h>
14 #include <linux/cpu.h>
15
16 #ifdef CONFIG_VM_EVENT_COUNTERS
17 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
18 EXPORT_PER_CPU_SYMBOL(vm_event_states);
19
20 static void sum_vm_events(unsigned long *ret, cpumask_t *cpumask)
21 {
22         int cpu = 0;
23         int i;
24
25         memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
26
27         cpu = first_cpu(*cpumask);
28         while (cpu < NR_CPUS) {
29                 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
30
31                 cpu = next_cpu(cpu, *cpumask);
32
33                 if (cpu < NR_CPUS)
34                         prefetch(&per_cpu(vm_event_states, cpu));
35
36
37                 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
38                         ret[i] += this->event[i];
39         }
40 }
41
42 /*
43  * Accumulate the vm event counters across all CPUs.
44  * The result is unavoidably approximate - it can change
45  * during and after execution of this function.
46 */
47 void all_vm_events(unsigned long *ret)
48 {
49         sum_vm_events(ret, &cpu_online_map);
50 }
51 EXPORT_SYMBOL_GPL(all_vm_events);
52
53 #ifdef CONFIG_HOTPLUG
54 /*
55  * Fold the foreign cpu events into our own.
56  *
57  * This is adding to the events on one processor
58  * but keeps the global counts constant.
59  */
60 void vm_events_fold_cpu(int cpu)
61 {
62         struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
63         int i;
64
65         for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
66                 count_vm_events(i, fold_state->event[i]);
67                 fold_state->event[i] = 0;
68         }
69 }
70 #endif /* CONFIG_HOTPLUG */
71
72 #endif /* CONFIG_VM_EVENT_COUNTERS */
73
74 /*
75  * Manage combined zone based / global counters
76  *
77  * vm_stat contains the global counters
78  */
79 atomic_long_t vm_stat[NR_VM_ZONE_STAT_ITEMS];
80 EXPORT_SYMBOL(vm_stat);
81
82 #ifdef CONFIG_SMP
83
84 static int calculate_threshold(struct zone *zone)
85 {
86         int threshold;
87         int mem;        /* memory in 128 MB units */
88
89         /*
90          * The threshold scales with the number of processors and the amount
91          * of memory per zone. More memory means that we can defer updates for
92          * longer, more processors could lead to more contention.
93          * fls() is used to have a cheap way of logarithmic scaling.
94          *
95          * Some sample thresholds:
96          *
97          * Threshold    Processors      (fls)   Zonesize        fls(mem+1)
98          * ------------------------------------------------------------------
99          * 8            1               1       0.9-1 GB        4
100          * 16           2               2       0.9-1 GB        4
101          * 20           2               2       1-2 GB          5
102          * 24           2               2       2-4 GB          6
103          * 28           2               2       4-8 GB          7
104          * 32           2               2       8-16 GB         8
105          * 4            2               2       <128M           1
106          * 30           4               3       2-4 GB          5
107          * 48           4               3       8-16 GB         8
108          * 32           8               4       1-2 GB          4
109          * 32           8               4       0.9-1GB         4
110          * 10           16              5       <128M           1
111          * 40           16              5       900M            4
112          * 70           64              7       2-4 GB          5
113          * 84           64              7       4-8 GB          6
114          * 108          512             9       4-8 GB          6
115          * 125          1024            10      8-16 GB         8
116          * 125          1024            10      16-32 GB        9
117          */
118
119         mem = zone->present_pages >> (27 - PAGE_SHIFT);
120
121         threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
122
123         /*
124          * Maximum threshold is 125
125          */
126         threshold = min(125, threshold);
127
128         return threshold;
129 }
130
131 /*
132  * Refresh the thresholds for each zone.
133  */
134 static void refresh_zone_stat_thresholds(void)
135 {
136         struct zone *zone;
137         int cpu;
138         int threshold;
139
140         for_each_zone(zone) {
141
142                 if (!zone->present_pages)
143                         continue;
144
145                 threshold = calculate_threshold(zone);
146
147                 for_each_online_cpu(cpu)
148                         zone_pcp(zone, cpu)->stat_threshold = threshold;
149         }
150 }
151
152 /*
153  * For use when we know that interrupts are disabled.
154  */
155 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
156                                 int delta)
157 {
158         struct per_cpu_pageset *pcp = zone_pcp(zone, smp_processor_id());
159         s8 *p = pcp->vm_stat_diff + item;
160         long x;
161
162         x = delta + *p;
163
164         if (unlikely(x > pcp->stat_threshold || x < -pcp->stat_threshold)) {
165                 zone_page_state_add(x, zone, item);
166                 x = 0;
167         }
168         *p = x;
169 }
170 EXPORT_SYMBOL(__mod_zone_page_state);
171
172 /*
173  * For an unknown interrupt state
174  */
175 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
176                                         int delta)
177 {
178         unsigned long flags;
179
180         local_irq_save(flags);
181         __mod_zone_page_state(zone, item, delta);
182         local_irq_restore(flags);
183 }
184 EXPORT_SYMBOL(mod_zone_page_state);
185
186 /*
187  * Optimized increment and decrement functions.
188  *
189  * These are only for a single page and therefore can take a struct page *
190  * argument instead of struct zone *. This allows the inclusion of the code
191  * generated for page_zone(page) into the optimized functions.
192  *
193  * No overflow check is necessary and therefore the differential can be
194  * incremented or decremented in place which may allow the compilers to
195  * generate better code.
196  * The increment or decrement is known and therefore one boundary check can
197  * be omitted.
198  *
199  * NOTE: These functions are very performance sensitive. Change only
200  * with care.
201  *
202  * Some processors have inc/dec instructions that are atomic vs an interrupt.
203  * However, the code must first determine the differential location in a zone
204  * based on the processor number and then inc/dec the counter. There is no
205  * guarantee without disabling preemption that the processor will not change
206  * in between and therefore the atomicity vs. interrupt cannot be exploited
207  * in a useful way here.
208  */
209 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
210 {
211         struct per_cpu_pageset *pcp = zone_pcp(zone, smp_processor_id());
212         s8 *p = pcp->vm_stat_diff + item;
213
214         (*p)++;
215
216         if (unlikely(*p > pcp->stat_threshold)) {
217                 int overstep = pcp->stat_threshold / 2;
218
219                 zone_page_state_add(*p + overstep, zone, item);
220                 *p = -overstep;
221         }
222 }
223
224 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
225 {
226         __inc_zone_state(page_zone(page), item);
227 }
228 EXPORT_SYMBOL(__inc_zone_page_state);
229
230 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
231 {
232         struct per_cpu_pageset *pcp = zone_pcp(zone, smp_processor_id());
233         s8 *p = pcp->vm_stat_diff + item;
234
235         (*p)--;
236
237         if (unlikely(*p < - pcp->stat_threshold)) {
238                 int overstep = pcp->stat_threshold / 2;
239
240                 zone_page_state_add(*p - overstep, zone, item);
241                 *p = overstep;
242         }
243 }
244
245 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
246 {
247         __dec_zone_state(page_zone(page), item);
248 }
249 EXPORT_SYMBOL(__dec_zone_page_state);
250
251 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
252 {
253         unsigned long flags;
254
255         local_irq_save(flags);
256         __inc_zone_state(zone, item);
257         local_irq_restore(flags);
258 }
259
260 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
261 {
262         unsigned long flags;
263         struct zone *zone;
264
265         zone = page_zone(page);
266         local_irq_save(flags);
267         __inc_zone_state(zone, item);
268         local_irq_restore(flags);
269 }
270 EXPORT_SYMBOL(inc_zone_page_state);
271
272 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
273 {
274         unsigned long flags;
275
276         local_irq_save(flags);
277         __dec_zone_page_state(page, item);
278         local_irq_restore(flags);
279 }
280 EXPORT_SYMBOL(dec_zone_page_state);
281
282 /*
283  * Update the zone counters for one cpu.
284  */
285 void refresh_cpu_vm_stats(int cpu)
286 {
287         struct zone *zone;
288         int i;
289         unsigned long flags;
290
291         for_each_zone(zone) {
292                 struct per_cpu_pageset *pcp;
293
294                 if (!populated_zone(zone))
295                         continue;
296
297                 pcp = zone_pcp(zone, cpu);
298
299                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
300                         if (pcp->vm_stat_diff[i]) {
301                                 local_irq_save(flags);
302                                 zone_page_state_add(pcp->vm_stat_diff[i],
303                                         zone, i);
304                                 pcp->vm_stat_diff[i] = 0;
305                                 local_irq_restore(flags);
306                         }
307         }
308 }
309
310 static void __refresh_cpu_vm_stats(void *dummy)
311 {
312         refresh_cpu_vm_stats(smp_processor_id());
313 }
314
315 /*
316  * Consolidate all counters.
317  *
318  * Note that the result is less inaccurate but still inaccurate
319  * if concurrent processes are allowed to run.
320  */
321 void refresh_vm_stats(void)
322 {
323         on_each_cpu(__refresh_cpu_vm_stats, NULL, 0, 1);
324 }
325 EXPORT_SYMBOL(refresh_vm_stats);
326
327 #endif
328
329 #ifdef CONFIG_NUMA
330 /*
331  * zonelist = the list of zones passed to the allocator
332  * z        = the zone from which the allocation occurred.
333  *
334  * Must be called with interrupts disabled.
335  */
336 void zone_statistics(struct zonelist *zonelist, struct zone *z)
337 {
338         if (z->zone_pgdat == zonelist->zones[0]->zone_pgdat) {
339                 __inc_zone_state(z, NUMA_HIT);
340         } else {
341                 __inc_zone_state(z, NUMA_MISS);
342                 __inc_zone_state(zonelist->zones[0], NUMA_FOREIGN);
343         }
344         if (z->node == numa_node_id())
345                 __inc_zone_state(z, NUMA_LOCAL);
346         else
347                 __inc_zone_state(z, NUMA_OTHER);
348 }
349 #endif
350
351 #ifdef CONFIG_PROC_FS
352
353 #include <linux/seq_file.h>
354
355 static void *frag_start(struct seq_file *m, loff_t *pos)
356 {
357         pg_data_t *pgdat;
358         loff_t node = *pos;
359         for (pgdat = first_online_pgdat();
360              pgdat && node;
361              pgdat = next_online_pgdat(pgdat))
362                 --node;
363
364         return pgdat;
365 }
366
367 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
368 {
369         pg_data_t *pgdat = (pg_data_t *)arg;
370
371         (*pos)++;
372         return next_online_pgdat(pgdat);
373 }
374
375 static void frag_stop(struct seq_file *m, void *arg)
376 {
377 }
378
379 /*
380  * This walks the free areas for each zone.
381  */
382 static int frag_show(struct seq_file *m, void *arg)
383 {
384         pg_data_t *pgdat = (pg_data_t *)arg;
385         struct zone *zone;
386         struct zone *node_zones = pgdat->node_zones;
387         unsigned long flags;
388         int order;
389
390         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
391                 if (!populated_zone(zone))
392                         continue;
393
394                 spin_lock_irqsave(&zone->lock, flags);
395                 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
396                 for (order = 0; order < MAX_ORDER; ++order)
397                         seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
398                 spin_unlock_irqrestore(&zone->lock, flags);
399                 seq_putc(m, '\n');
400         }
401         return 0;
402 }
403
404 const struct seq_operations fragmentation_op = {
405         .start  = frag_start,
406         .next   = frag_next,
407         .stop   = frag_stop,
408         .show   = frag_show,
409 };
410
411 #ifdef CONFIG_ZONE_DMA
412 #define TEXT_FOR_DMA(xx) xx "_dma",
413 #else
414 #define TEXT_FOR_DMA(xx)
415 #endif
416
417 #ifdef CONFIG_ZONE_DMA32
418 #define TEXT_FOR_DMA32(xx) xx "_dma32",
419 #else
420 #define TEXT_FOR_DMA32(xx)
421 #endif
422
423 #ifdef CONFIG_HIGHMEM
424 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
425 #else
426 #define TEXT_FOR_HIGHMEM(xx)
427 #endif
428
429 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
430                                         TEXT_FOR_HIGHMEM(xx)
431
432 static const char * const vmstat_text[] = {
433         /* Zoned VM counters */
434         "nr_free_pages",
435         "nr_active",
436         "nr_inactive",
437         "nr_anon_pages",
438         "nr_mapped",
439         "nr_file_pages",
440         "nr_dirty",
441         "nr_writeback",
442         "nr_slab_reclaimable",
443         "nr_slab_unreclaimable",
444         "nr_page_table_pages",
445         "nr_unstable",
446         "nr_bounce",
447         "nr_vmscan_write",
448
449 #ifdef CONFIG_NUMA
450         "numa_hit",
451         "numa_miss",
452         "numa_foreign",
453         "numa_interleave",
454         "numa_local",
455         "numa_other",
456 #endif
457
458 #ifdef CONFIG_VM_EVENT_COUNTERS
459         "pgpgin",
460         "pgpgout",
461         "pswpin",
462         "pswpout",
463
464         TEXTS_FOR_ZONES("pgalloc")
465
466         "pgfree",
467         "pgactivate",
468         "pgdeactivate",
469
470         "pgfault",
471         "pgmajfault",
472
473         TEXTS_FOR_ZONES("pgrefill")
474         TEXTS_FOR_ZONES("pgsteal")
475         TEXTS_FOR_ZONES("pgscan_kswapd")
476         TEXTS_FOR_ZONES("pgscan_direct")
477
478         "pginodesteal",
479         "slabs_scanned",
480         "kswapd_steal",
481         "kswapd_inodesteal",
482         "pageoutrun",
483         "allocstall",
484
485         "pgrotated",
486 #endif
487 };
488
489 /*
490  * Output information about zones in @pgdat.
491  */
492 static int zoneinfo_show(struct seq_file *m, void *arg)
493 {
494         pg_data_t *pgdat = arg;
495         struct zone *zone;
496         struct zone *node_zones = pgdat->node_zones;
497         unsigned long flags;
498
499         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; zone++) {
500                 int i;
501
502                 if (!populated_zone(zone))
503                         continue;
504
505                 spin_lock_irqsave(&zone->lock, flags);
506                 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
507                 seq_printf(m,
508                            "\n  pages free     %lu"
509                            "\n        min      %lu"
510                            "\n        low      %lu"
511                            "\n        high     %lu"
512                            "\n        scanned  %lu (a: %lu i: %lu)"
513                            "\n        spanned  %lu"
514                            "\n        present  %lu",
515                            zone_page_state(zone, NR_FREE_PAGES),
516                            zone->pages_min,
517                            zone->pages_low,
518                            zone->pages_high,
519                            zone->pages_scanned,
520                            zone->nr_scan_active, zone->nr_scan_inactive,
521                            zone->spanned_pages,
522                            zone->present_pages);
523
524                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
525                         seq_printf(m, "\n    %-12s %lu", vmstat_text[i],
526                                         zone_page_state(zone, i));
527
528                 seq_printf(m,
529                            "\n        protection: (%lu",
530                            zone->lowmem_reserve[0]);
531                 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
532                         seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
533                 seq_printf(m,
534                            ")"
535                            "\n  pagesets");
536                 for_each_online_cpu(i) {
537                         struct per_cpu_pageset *pageset;
538                         int j;
539
540                         pageset = zone_pcp(zone, i);
541                         for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
542                                 seq_printf(m,
543                                            "\n    cpu: %i pcp: %i"
544                                            "\n              count: %i"
545                                            "\n              high:  %i"
546                                            "\n              batch: %i",
547                                            i, j,
548                                            pageset->pcp[j].count,
549                                            pageset->pcp[j].high,
550                                            pageset->pcp[j].batch);
551                         }
552 #ifdef CONFIG_SMP
553                         seq_printf(m, "\n  vm stats threshold: %d",
554                                         pageset->stat_threshold);
555 #endif
556                 }
557                 seq_printf(m,
558                            "\n  all_unreclaimable: %u"
559                            "\n  prev_priority:     %i"
560                            "\n  start_pfn:         %lu",
561                            zone->all_unreclaimable,
562                            zone->prev_priority,
563                            zone->zone_start_pfn);
564                 spin_unlock_irqrestore(&zone->lock, flags);
565                 seq_putc(m, '\n');
566         }
567         return 0;
568 }
569
570 const struct seq_operations zoneinfo_op = {
571         .start  = frag_start, /* iterate over all zones. The same as in
572                                * fragmentation. */
573         .next   = frag_next,
574         .stop   = frag_stop,
575         .show   = zoneinfo_show,
576 };
577
578 static void *vmstat_start(struct seq_file *m, loff_t *pos)
579 {
580         unsigned long *v;
581 #ifdef CONFIG_VM_EVENT_COUNTERS
582         unsigned long *e;
583 #endif
584         int i;
585
586         if (*pos >= ARRAY_SIZE(vmstat_text))
587                 return NULL;
588
589 #ifdef CONFIG_VM_EVENT_COUNTERS
590         v = kmalloc(NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long)
591                         + sizeof(struct vm_event_state), GFP_KERNEL);
592 #else
593         v = kmalloc(NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long),
594                         GFP_KERNEL);
595 #endif
596         m->private = v;
597         if (!v)
598                 return ERR_PTR(-ENOMEM);
599         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
600                 v[i] = global_page_state(i);
601 #ifdef CONFIG_VM_EVENT_COUNTERS
602         e = v + NR_VM_ZONE_STAT_ITEMS;
603         all_vm_events(e);
604         e[PGPGIN] /= 2;         /* sectors -> kbytes */
605         e[PGPGOUT] /= 2;
606 #endif
607         return v + *pos;
608 }
609
610 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
611 {
612         (*pos)++;
613         if (*pos >= ARRAY_SIZE(vmstat_text))
614                 return NULL;
615         return (unsigned long *)m->private + *pos;
616 }
617
618 static int vmstat_show(struct seq_file *m, void *arg)
619 {
620         unsigned long *l = arg;
621         unsigned long off = l - (unsigned long *)m->private;
622
623         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
624         return 0;
625 }
626
627 static void vmstat_stop(struct seq_file *m, void *arg)
628 {
629         kfree(m->private);
630         m->private = NULL;
631 }
632
633 const struct seq_operations vmstat_op = {
634         .start  = vmstat_start,
635         .next   = vmstat_next,
636         .stop   = vmstat_stop,
637         .show   = vmstat_show,
638 };
639
640 #endif /* CONFIG_PROC_FS */
641
642 #ifdef CONFIG_SMP
643 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
644 int sysctl_stat_interval __read_mostly = HZ;
645
646 static void vmstat_update(struct work_struct *w)
647 {
648         refresh_cpu_vm_stats(smp_processor_id());
649         schedule_delayed_work(&__get_cpu_var(vmstat_work),
650                 sysctl_stat_interval);
651 }
652
653 static void __devinit start_cpu_timer(int cpu)
654 {
655         struct delayed_work *vmstat_work = &per_cpu(vmstat_work, cpu);
656
657         INIT_DELAYED_WORK(vmstat_work, vmstat_update);
658         schedule_delayed_work_on(cpu, vmstat_work, HZ + cpu);
659 }
660
661 /*
662  * Use the cpu notifier to insure that the thresholds are recalculated
663  * when necessary.
664  */
665 static int __cpuinit vmstat_cpuup_callback(struct notifier_block *nfb,
666                 unsigned long action,
667                 void *hcpu)
668 {
669         long cpu = (long)hcpu;
670
671         switch (action) {
672         case CPU_ONLINE:
673         case CPU_ONLINE_FROZEN:
674                 start_cpu_timer(cpu);
675                 break;
676         case CPU_DOWN_PREPARE:
677         case CPU_DOWN_PREPARE_FROZEN:
678                 cancel_rearming_delayed_work(&per_cpu(vmstat_work, cpu));
679                 per_cpu(vmstat_work, cpu).work.func = NULL;
680                 break;
681         case CPU_DOWN_FAILED:
682         case CPU_DOWN_FAILED_FROZEN:
683                 start_cpu_timer(cpu);
684                 break;
685         case CPU_DEAD:
686         case CPU_DEAD_FROZEN:
687                 refresh_zone_stat_thresholds();
688                 break;
689         default:
690                 break;
691         }
692         return NOTIFY_OK;
693 }
694
695 static struct notifier_block __cpuinitdata vmstat_notifier =
696         { &vmstat_cpuup_callback, NULL, 0 };
697
698 int __init setup_vmstat(void)
699 {
700         int cpu;
701
702         refresh_zone_stat_thresholds();
703         register_cpu_notifier(&vmstat_notifier);
704
705         for_each_online_cpu(cpu)
706                 start_cpu_timer(cpu);
707         return 0;
708 }
709 module_init(setup_vmstat)
710 #endif