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