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