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/config.h>
  13 #include <linux/mm.h>
  14 #include <linux/module.h>
  15 #include <linux/cpu.h>
  16
  17 void __get_zone_counts(unsigned long *active, unsigned long *inactive,
  18                         unsigned long *free, struct pglist_data *pgdat)
  19 {
  20         struct zone *zones = pgdat->node_zones;
  21         int i;
  22
  23         *active = 0;
  24         *inactive = 0;
  25         *free = 0;
  26         for (i = 0; i < MAX_NR_ZONES; i++) {
  27                 *active += zones[i].nr_active;
  28                 *inactive += zones[i].nr_inactive;
  29                 *free += zones[i].free_pages;
  30         }
  31 }
  32
  33 void get_zone_counts(unsigned long *active,
  34                 unsigned long *inactive, unsigned long *free)
  35 {
  36         struct pglist_data *pgdat;
  37
  38         *active = 0;
  39         *inactive = 0;
  40         *free = 0;
  41         for_each_online_pgdat(pgdat) {
  42                 unsigned long l, m, n;
  43                 __get_zone_counts(&l, &m, &n, pgdat);
  44                 *active += l;
  45                 *inactive += m;
  46                 *free += n;
  47         }
  48 }
  49
  50 #ifdef CONFIG_VM_EVENT_COUNTERS
  51 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
  52 EXPORT_PER_CPU_SYMBOL(vm_event_states);
  53
  54 static void sum_vm_events(unsigned long *ret, cpumask_t *cpumask)
  55 {
  56         int cpu = 0;
  57         int i;
  58
  59         memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
  60
  61         cpu = first_cpu(*cpumask);
  62         while (cpu < NR_CPUS) {
  63                 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
  64
  65                 cpu = next_cpu(cpu, *cpumask);
  66
  67                 if (cpu < NR_CPUS)
  68                         prefetch(&per_cpu(vm_event_states, cpu));
  69
  70
  71                 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
  72                         ret[i] += this->event[i];
  73         }
  74 }
  75
  76 /*
  77  * Accumulate the vm event counters across all CPUs.
  78  * The result is unavoidably approximate - it can change
  79  * during and after execution of this function.
  80 */
  81 void all_vm_events(unsigned long *ret)
  82 {
  83         sum_vm_events(ret, &cpu_online_map);
  84 }
  85 EXPORT_SYMBOL_GPL(all_vm_events);
  86
  87 #ifdef CONFIG_HOTPLUG
  88 /*
  89  * Fold the foreign cpu events into our own.
  90  *
  91  * This is adding to the events on one processor
  92  * but keeps the global counts constant.
  93  */
  94 void vm_events_fold_cpu(int cpu)
  95 {
  96         struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
  97         int i;
  98
  99         for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
 100                 count_vm_events(i, fold_state->event[i]);
 101                 fold_state->event[i] = 0;
 102         }
 103 }
 104 #endif /* CONFIG_HOTPLUG */
 105
 106 #endif /* CONFIG_VM_EVENT_COUNTERS */
 107
 108 /*
 109  * Manage combined zone based / global counters
 110  *
 111  * vm_stat contains the global counters
 112  */
 113 atomic_long_t vm_stat[NR_VM_ZONE_STAT_ITEMS];
 114 EXPORT_SYMBOL(vm_stat);
 115
 116 #ifdef CONFIG_SMP
 117
 118 static int calculate_threshold(struct zone *zone)
 119 {
 120         int threshold;
 121         int mem;        /* memory in 128 MB units */
 122
 123         /*
 124          * The threshold scales with the number of processors and the amount
 125          * of memory per zone. More memory means that we can defer updates for
 126          * longer, more processors could lead to more contention.
 127          * fls() is used to have a cheap way of logarithmic scaling.
 128          *
 129          * Some sample thresholds:
 130          *
 131          * Threshold    Processors      (fls)   Zonesize        fls(mem+1)
 132          * ------------------------------------------------------------------
 133          * 8            1               1       0.9-1 GB        4
 134          * 16           2               2       0.9-1 GB        4
 135          * 20           2               2       1-2 GB          5
 136          * 24           2               2       2-4 GB          6
 137          * 28           2               2       4-8 GB          7
 138          * 32           2               2       8-16 GB         8
 139          * 4            2               2       <128M           1
 140          * 30           4               3       2-4 GB          5
 141          * 48           4               3       8-16 GB         8
 142          * 32           8               4       1-2 GB          4
 143          * 32           8               4       0.9-1GB         4
 144          * 10           16              5       <128M           1
 145          * 40           16              5       900M            4
 146          * 70           64              7       2-4 GB          5
 147          * 84           64              7       4-8 GB          6
 148          * 108          512             9       4-8 GB          6
 149          * 125          1024            10      8-16 GB         8
 150          * 125          1024            10      16-32 GB        9
 151          */
 152
 153         mem = zone->present_pages >> (27 - PAGE_SHIFT);
 154
 155         threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
 156
 157         /*
 158          * Maximum threshold is 125
 159          */
 160         threshold = min(125, threshold);
 161
 162         return threshold;
 163 }
 164
 165 /*
 166  * Refresh the thresholds for each zone.
 167  */
 168 static void refresh_zone_stat_thresholds(void)
 169 {
 170         struct zone *zone;
 171         int cpu;
 172         int threshold;
 173
 174         for_each_zone(zone) {
 175
 176                 if (!zone->present_pages)
 177                         continue;
 178
 179                 threshold = calculate_threshold(zone);
 180
 181                 for_each_online_cpu(cpu)
 182                         zone_pcp(zone, cpu)->stat_threshold = threshold;
 183         }
 184 }
 185
 186 /*
 187  * For use when we know that interrupts are disabled.
 188  */
 189 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 190                                 int delta)
 191 {
 192         struct per_cpu_pageset *pcp = zone_pcp(zone, smp_processor_id());
 193         s8 *p = pcp->vm_stat_diff + item;
 194         long x;
 195
 196         x = delta + *p;
 197
 198         if (unlikely(x > pcp->stat_threshold || x < -pcp->stat_threshold)) {
 199                 zone_page_state_add(x, zone, item);
 200                 x = 0;
 201         }
 202         *p = x;
 203 }
 204 EXPORT_SYMBOL(__mod_zone_page_state);
 205
 206 /*
 207  * For an unknown interrupt state
 208  */
 209 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 210                                         int delta)
 211 {
 212         unsigned long flags;
 213
 214         local_irq_save(flags);
 215         __mod_zone_page_state(zone, item, delta);
 216         local_irq_restore(flags);
 217 }
 218 EXPORT_SYMBOL(mod_zone_page_state);
 219
 220 /*
 221  * Optimized increment and decrement functions.
 222  *
 223  * These are only for a single page and therefore can take a struct page *
 224  * argument instead of struct zone *. This allows the inclusion of the code
 225  * generated for page_zone(page) into the optimized functions.
 226  *
 227  * No overflow check is necessary and therefore the differential can be
 228  * incremented or decremented in place which may allow the compilers to
 229  * generate better code.
 230  * The increment or decrement is known and therefore one boundary check can
 231  * be omitted.
 232  *
 233  * NOTE: These functions are very performance sensitive. Change only
 234  * with care.
 235  *
 236  * Some processors have inc/dec instructions that are atomic vs an interrupt.
 237  * However, the code must first determine the differential location in a zone
 238  * based on the processor number and then inc/dec the counter. There is no
 239  * guarantee without disabling preemption that the processor will not change
 240  * in between and therefore the atomicity vs. interrupt cannot be exploited
 241  * in a useful way here.
 242  */
 243 static void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
 244 {
 245         struct per_cpu_pageset *pcp = zone_pcp(zone, smp_processor_id());
 246         s8 *p = pcp->vm_stat_diff + item;
 247
 248         (*p)++;
 249
 250         if (unlikely(*p > pcp->stat_threshold)) {
 251                 int overstep = pcp->stat_threshold / 2;
 252
 253                 zone_page_state_add(*p + overstep, zone, item);
 254                 *p = -overstep;
 255         }
 256 }
 257
 258 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
 259 {
 260         __inc_zone_state(page_zone(page), item);
 261 }
 262 EXPORT_SYMBOL(__inc_zone_page_state);
 263
 264 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
 265 {
 266         struct zone *zone = page_zone(page);
 267         struct per_cpu_pageset *pcp = zone_pcp(zone, smp_processor_id());
 268         s8 *p = pcp->vm_stat_diff + item;
 269
 270         (*p)--;
 271
 272         if (unlikely(*p < - pcp->stat_threshold)) {
 273                 int overstep = pcp->stat_threshold / 2;
 274
 275                 zone_page_state_add(*p - overstep, zone, item);
 276                 *p = overstep;
 277         }
 278 }
 279 EXPORT_SYMBOL(__dec_zone_page_state);
 280
 281 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
 282 {
 283         unsigned long flags;
 284
 285         local_irq_save(flags);
 286         __inc_zone_state(zone, item);
 287         local_irq_restore(flags);
 288 }
 289
 290 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 291 {
 292         unsigned long flags;
 293         struct zone *zone;
 294
 295         zone = page_zone(page);
 296         local_irq_save(flags);
 297         __inc_zone_state(zone, item);
 298         local_irq_restore(flags);
 299 }
 300 EXPORT_SYMBOL(inc_zone_page_state);
 301
 302 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 303 {
 304         unsigned long flags;
 305
 306         local_irq_save(flags);
 307         __dec_zone_page_state(page, item);
 308         local_irq_restore(flags);
 309 }
 310 EXPORT_SYMBOL(dec_zone_page_state);
 311
 312 /*
 313  * Update the zone counters for one cpu.
 314  */
 315 void refresh_cpu_vm_stats(int cpu)
 316 {
 317         struct zone *zone;
 318         int i;
 319         unsigned long flags;
 320
 321         for_each_zone(zone) {
 322                 struct per_cpu_pageset *pcp;
 323
 324                 pcp = zone_pcp(zone, cpu);
 325
 326                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 327                         if (pcp->vm_stat_diff[i]) {
 328                                 local_irq_save(flags);
 329                                 zone_page_state_add(pcp->vm_stat_diff[i],
 330                                         zone, i);
 331                                 pcp->vm_stat_diff[i] = 0;
 332                                 local_irq_restore(flags);
 333                         }
 334         }
 335 }
 336
 337 static void __refresh_cpu_vm_stats(void *dummy)
 338 {
 339         refresh_cpu_vm_stats(smp_processor_id());
 340 }
 341
 342 /*
 343  * Consolidate all counters.
 344  *
 345  * Note that the result is less inaccurate but still inaccurate
 346  * if concurrent processes are allowed to run.
 347  */
 348 void refresh_vm_stats(void)
 349 {
 350         on_each_cpu(__refresh_cpu_vm_stats, NULL, 0, 1);
 351 }
 352 EXPORT_SYMBOL(refresh_vm_stats);
 353
 354 #endif
 355
 356 #ifdef CONFIG_NUMA
 357 /*
 358  * zonelist = the list of zones passed to the allocator
 359  * z        = the zone from which the allocation occurred.
 360  *
 361  * Must be called with interrupts disabled.
 362  */
 363 void zone_statistics(struct zonelist *zonelist, struct zone *z)
 364 {
 365         if (z->zone_pgdat == zonelist->zones[0]->zone_pgdat) {
 366                 __inc_zone_state(z, NUMA_HIT);
 367         } else {
 368                 __inc_zone_state(z, NUMA_MISS);
 369                 __inc_zone_state(zonelist->zones[0], NUMA_FOREIGN);
 370         }
 371         if (z->zone_pgdat == NODE_DATA(numa_node_id()))
 372                 __inc_zone_state(z, NUMA_LOCAL);
 373         else
 374                 __inc_zone_state(z, NUMA_OTHER);
 375 }
 376 #endif
 377
 378 #ifdef CONFIG_PROC_FS
 379
 380 #include <linux/seq_file.h>
 381
 382 static void *frag_start(struct seq_file *m, loff_t *pos)
 383 {
 384         pg_data_t *pgdat;
 385         loff_t node = *pos;
 386         for (pgdat = first_online_pgdat();
 387              pgdat && node;
 388              pgdat = next_online_pgdat(pgdat))
 389                 --node;
 390
 391         return pgdat;
 392 }
 393
 394 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
 395 {
 396         pg_data_t *pgdat = (pg_data_t *)arg;
 397
 398         (*pos)++;
 399         return next_online_pgdat(pgdat);
 400 }
 401
 402 static void frag_stop(struct seq_file *m, void *arg)
 403 {
 404 }
 405
 406 /*
 407  * This walks the free areas for each zone.
 408  */
 409 static int frag_show(struct seq_file *m, void *arg)
 410 {
 411         pg_data_t *pgdat = (pg_data_t *)arg;
 412         struct zone *zone;
 413         struct zone *node_zones = pgdat->node_zones;
 414         unsigned long flags;
 415         int order;
 416
 417         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 418                 if (!populated_zone(zone))
 419                         continue;
 420
 421                 spin_lock_irqsave(&zone->lock, flags);
 422                 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 423                 for (order = 0; order < MAX_ORDER; ++order)
 424                         seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
 425                 spin_unlock_irqrestore(&zone->lock, flags);
 426                 seq_putc(m, '\n');
 427         }
 428         return 0;
 429 }
 430
 431 struct seq_operations fragmentation_op = {
 432         .start  = frag_start,
 433         .next   = frag_next,
 434         .stop   = frag_stop,
 435         .show   = frag_show,
 436 };
 437
 438 static char *vmstat_text[] = {
 439         /* Zoned VM counters */
 440         "nr_anon_pages",
 441         "nr_mapped",
 442         "nr_file_pages",
 443         "nr_slab",
 444         "nr_page_table_pages",
 445         "nr_dirty",
 446         "nr_writeback",
 447         "nr_unstable",
 448         "nr_bounce",
 449
 450 #ifdef CONFIG_NUMA
 451         "numa_hit",
 452         "numa_miss",
 453         "numa_foreign",
 454         "numa_interleave",
 455         "numa_local",
 456         "numa_other",
 457 #endif
 458
 459 #ifdef CONFIG_VM_EVENT_COUNTERS
 460         "pgpgin",
 461         "pgpgout",
 462         "pswpin",
 463         "pswpout",
 464
 465         "pgalloc_dma",
 466         "pgalloc_dma32",
 467         "pgalloc_normal",
 468         "pgalloc_high",
 469
 470         "pgfree",
 471         "pgactivate",
 472         "pgdeactivate",
 473
 474         "pgfault",
 475         "pgmajfault",
 476
 477         "pgrefill_dma",
 478         "pgrefill_dma32",
 479         "pgrefill_normal",
 480         "pgrefill_high",
 481
 482         "pgsteal_dma",
 483         "pgsteal_dma32",
 484         "pgsteal_normal",
 485         "pgsteal_high",
 486
 487         "pgscan_kswapd_dma",
 488         "pgscan_kswapd_dma32",
 489         "pgscan_kswapd_normal",
 490         "pgscan_kswapd_high",
 491
 492         "pgscan_direct_dma",
 493         "pgscan_direct_dma32",
 494         "pgscan_direct_normal",
 495         "pgscan_direct_high",
 496
 497         "pginodesteal",
 498         "slabs_scanned",
 499         "kswapd_steal",
 500         "kswapd_inodesteal",
 501         "pageoutrun",
 502         "allocstall",
 503
 504         "pgrotated",
 505 #endif
 506 };
 507
 508 /*
 509  * Output information about zones in @pgdat.
 510  */
 511 static int zoneinfo_show(struct seq_file *m, void *arg)
 512 {
 513         pg_data_t *pgdat = arg;
 514         struct zone *zone;
 515         struct zone *node_zones = pgdat->node_zones;
 516         unsigned long flags;
 517
 518         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; zone++) {
 519                 int i;
 520
 521                 if (!populated_zone(zone))
 522                         continue;
 523
 524                 spin_lock_irqsave(&zone->lock, flags);
 525                 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
 526                 seq_printf(m,
 527                            "\n  pages free     %lu"
 528                            "\n        min      %lu"
 529                            "\n        low      %lu"
 530                            "\n        high     %lu"
 531                            "\n        active   %lu"
 532                            "\n        inactive %lu"
 533                            "\n        scanned  %lu (a: %lu i: %lu)"
 534                            "\n        spanned  %lu"
 535                            "\n        present  %lu",
 536                            zone->free_pages,
 537                            zone->pages_min,
 538                            zone->pages_low,
 539                            zone->pages_high,
 540                            zone->nr_active,
 541                            zone->nr_inactive,
 542                            zone->pages_scanned,
 543                            zone->nr_scan_active, zone->nr_scan_inactive,
 544                            zone->spanned_pages,
 545                            zone->present_pages);
 546
 547                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 548                         seq_printf(m, "\n    %-12s %lu", vmstat_text[i],
 549                                         zone_page_state(zone, i));
 550
 551                 seq_printf(m,
 552                            "\n        protection: (%lu",
 553                            zone->lowmem_reserve[0]);
 554                 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
 555                         seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
 556                 seq_printf(m,
 557                            ")"
 558                            "\n  pagesets");
 559                 for_each_online_cpu(i) {
 560                         struct per_cpu_pageset *pageset;
 561                         int j;
 562
 563                         pageset = zone_pcp(zone, i);
 564                         for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
 565                                 if (pageset->pcp[j].count)
 566                                         break;
 567                         }
 568                         if (j == ARRAY_SIZE(pageset->pcp))
 569                                 continue;
 570                         for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
 571                                 seq_printf(m,
 572                                            "\n    cpu: %i pcp: %i"
 573                                            "\n              count: %i"
 574                                            "\n              high:  %i"
 575                                            "\n              batch: %i",
 576                                            i, j,
 577                                            pageset->pcp[j].count,
 578                                            pageset->pcp[j].high,
 579                                            pageset->pcp[j].batch);
 580                         }
 581 #ifdef CONFIG_SMP
 582                         seq_printf(m, "\n  vm stats threshold: %d",
 583                                         pageset->stat_threshold);
 584 #endif
 585                 }
 586                 seq_printf(m,
 587                            "\n  all_unreclaimable: %u"
 588                            "\n  prev_priority:     %i"
 589                            "\n  temp_priority:     %i"
 590                            "\n  start_pfn:         %lu",
 591                            zone->all_unreclaimable,
 592                            zone->prev_priority,
 593                            zone->temp_priority,
 594                            zone->zone_start_pfn);
 595                 spin_unlock_irqrestore(&zone->lock, flags);
 596                 seq_putc(m, '\n');
 597         }
 598         return 0;
 599 }
 600
 601 struct seq_operations zoneinfo_op = {
 602         .start  = frag_start, /* iterate over all zones. The same as in
 603                                * fragmentation. */
 604         .next   = frag_next,
 605         .stop   = frag_stop,
 606         .show   = zoneinfo_show,
 607 };
 608
 609 static void *vmstat_start(struct seq_file *m, loff_t *pos)
 610 {
 611         unsigned long *v;
 612 #ifdef CONFIG_VM_EVENT_COUNTERS
 613         unsigned long *e;
 614 #endif
 615         int i;
 616
 617         if (*pos >= ARRAY_SIZE(vmstat_text))
 618                 return NULL;
 619
 620 #ifdef CONFIG_VM_EVENT_COUNTERS
 621         v = kmalloc(NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long)
 622                         + sizeof(struct vm_event_state), GFP_KERNEL);
 623 #else
 624         v = kmalloc(NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long),
 625                         GFP_KERNEL);
 626 #endif
 627         m->private = v;
 628         if (!v)
 629                 return ERR_PTR(-ENOMEM);
 630         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 631                 v[i] = global_page_state(i);
 632 #ifdef CONFIG_VM_EVENT_COUNTERS
 633         e = v + NR_VM_ZONE_STAT_ITEMS;
 634         all_vm_events(e);
 635         e[PGPGIN] /= 2;         /* sectors -> kbytes */
 636         e[PGPGOUT] /= 2;
 637 #endif
 638         return v + *pos;
 639 }
 640
 641 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
 642 {
 643         (*pos)++;
 644         if (*pos >= ARRAY_SIZE(vmstat_text))
 645                 return NULL;
 646         return (unsigned long *)m->private + *pos;
 647 }
 648
 649 static int vmstat_show(struct seq_file *m, void *arg)
 650 {
 651         unsigned long *l = arg;
 652         unsigned long off = l - (unsigned long *)m->private;
 653
 654         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
 655         return 0;
 656 }
 657
 658 static void vmstat_stop(struct seq_file *m, void *arg)
 659 {
 660         kfree(m->private);
 661         m->private = NULL;
 662 }
 663
 664 struct seq_operations vmstat_op = {
 665         .start  = vmstat_start,
 666         .next   = vmstat_next,
 667         .stop   = vmstat_stop,
 668         .show   = vmstat_show,
 669 };
 670
 671 #endif /* CONFIG_PROC_FS */
 672
 673 #ifdef CONFIG_SMP
 674 /*
 675  * Use the cpu notifier to insure that the thresholds are recalculated
 676  * when necessary.
 677  */
 678 static int __cpuinit vmstat_cpuup_callback(struct notifier_block *nfb,
 679                 unsigned long action,
 680                 void *hcpu)
 681 {
 682         switch (action) {
 683                 case CPU_UP_PREPARE:
 684                 case CPU_UP_CANCELED:
 685                 case CPU_DEAD:
 686                         refresh_zone_stat_thresholds();
 687                         break;
 688                 default:
 689                         break;
 690         }
 691         return NOTIFY_OK;
 692 }
 693
 694 static struct notifier_block __cpuinitdata vmstat_notifier =
 695         { &vmstat_cpuup_callback, NULL, 0 };
 696
 697 int __init setup_vmstat(void)
 698 {
 699         refresh_zone_stat_thresholds();
 700         register_cpu_notifier(&vmstat_notifier);
 701         return 0;
 702 }
 703 module_init(setup_vmstat)
 704 #endif