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/err.h>
  14 #include <linux/module.h>
  15 #include <linux/cpu.h>
  16 #include <linux/sched.h>
  17
  18 #ifdef CONFIG_VM_EVENT_COUNTERS
  19 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
  20 EXPORT_PER_CPU_SYMBOL(vm_event_states);
  21
  22 static void sum_vm_events(unsigned long *ret, cpumask_t *cpumask)
  23 {
  24         int cpu = 0;
  25         int i;
  26
  27         memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
  28
  29         cpu = first_cpu(*cpumask);
  30         while (cpu < NR_CPUS) {
  31                 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
  32
  33                 cpu = next_cpu(cpu, *cpumask);
  34
  35                 if (cpu < NR_CPUS)
  36                         prefetch(&per_cpu(vm_event_states, cpu));
  37
  38
  39                 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
  40                         ret[i] += this->event[i];
  41         }
  42 }
  43
  44 /*
  45  * Accumulate the vm event counters across all CPUs.
  46  * The result is unavoidably approximate - it can change
  47  * during and after execution of this function.
  48 */
  49 void all_vm_events(unsigned long *ret)
  50 {
  51         sum_vm_events(ret, &cpu_online_map);
  52 }
  53 EXPORT_SYMBOL_GPL(all_vm_events);
  54
  55 #ifdef CONFIG_HOTPLUG
  56 /*
  57  * Fold the foreign cpu events into our own.
  58  *
  59  * This is adding to the events on one processor
  60  * but keeps the global counts constant.
  61  */
  62 void vm_events_fold_cpu(int cpu)
  63 {
  64         struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
  65         int i;
  66
  67         for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
  68                 count_vm_events(i, fold_state->event[i]);
  69                 fold_state->event[i] = 0;
  70         }
  71 }
  72 #endif /* CONFIG_HOTPLUG */
  73
  74 #endif /* CONFIG_VM_EVENT_COUNTERS */
  75
  76 /*
  77  * Manage combined zone based / global counters
  78  *
  79  * vm_stat contains the global counters
  80  */
  81 atomic_long_t vm_stat[NR_VM_ZONE_STAT_ITEMS];
  82 EXPORT_SYMBOL(vm_stat);
  83
  84 #ifdef CONFIG_SMP
  85
  86 static int calculate_threshold(struct zone *zone)
  87 {
  88         int threshold;
  89         int mem;        /* memory in 128 MB units */
  90
  91         /*
  92          * The threshold scales with the number of processors and the amount
  93          * of memory per zone. More memory means that we can defer updates for
  94          * longer, more processors could lead to more contention.
  95          * fls() is used to have a cheap way of logarithmic scaling.
  96          *
  97          * Some sample thresholds:
  98          *
  99          * Threshold    Processors      (fls)   Zonesize        fls(mem+1)
 100          * ------------------------------------------------------------------
 101          * 8            1               1       0.9-1 GB        4
 102          * 16           2               2       0.9-1 GB        4
 103          * 20           2               2       1-2 GB          5
 104          * 24           2               2       2-4 GB          6
 105          * 28           2               2       4-8 GB          7
 106          * 32           2               2       8-16 GB         8
 107          * 4            2               2       <128M           1
 108          * 30           4               3       2-4 GB          5
 109          * 48           4               3       8-16 GB         8
 110          * 32           8               4       1-2 GB          4
 111          * 32           8               4       0.9-1GB         4
 112          * 10           16              5       <128M           1
 113          * 40           16              5       900M            4
 114          * 70           64              7       2-4 GB          5
 115          * 84           64              7       4-8 GB          6
 116          * 108          512             9       4-8 GB          6
 117          * 125          1024            10      8-16 GB         8
 118          * 125          1024            10      16-32 GB        9
 119          */
 120
 121         mem = zone->present_pages >> (27 - PAGE_SHIFT);
 122
 123         threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
 124
 125         /*
 126          * Maximum threshold is 125
 127          */
 128         threshold = min(125, threshold);
 129
 130         return threshold;
 131 }
 132
 133 /*
 134  * Refresh the thresholds for each zone.
 135  */
 136 static void refresh_zone_stat_thresholds(void)
 137 {
 138         struct zone *zone;
 139         int cpu;
 140         int threshold;
 141
 142         for_each_zone(zone) {
 143
 144                 if (!zone->present_pages)
 145                         continue;
 146
 147                 threshold = calculate_threshold(zone);
 148
 149                 for_each_online_cpu(cpu)
 150                         zone_pcp(zone, cpu)->stat_threshold = threshold;
 151         }
 152 }
 153
 154 /*
 155  * For use when we know that interrupts are disabled.
 156  */
 157 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 158                                 int delta)
 159 {
 160         struct per_cpu_pageset *pcp = zone_pcp(zone, smp_processor_id());
 161         s8 *p = pcp->vm_stat_diff + item;
 162         long x;
 163
 164         x = delta + *p;
 165
 166         if (unlikely(x > pcp->stat_threshold || x < -pcp->stat_threshold)) {
 167                 zone_page_state_add(x, zone, item);
 168                 x = 0;
 169         }
 170         *p = x;
 171 }
 172 EXPORT_SYMBOL(__mod_zone_page_state);
 173
 174 /*
 175  * For an unknown interrupt state
 176  */
 177 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 178                                         int delta)
 179 {
 180         unsigned long flags;
 181
 182         local_irq_save(flags);
 183         __mod_zone_page_state(zone, item, delta);
 184         local_irq_restore(flags);
 185 }
 186 EXPORT_SYMBOL(mod_zone_page_state);
 187
 188 /*
 189  * Optimized increment and decrement functions.
 190  *
 191  * These are only for a single page and therefore can take a struct page *
 192  * argument instead of struct zone *. This allows the inclusion of the code
 193  * generated for page_zone(page) into the optimized functions.
 194  *
 195  * No overflow check is necessary and therefore the differential can be
 196  * incremented or decremented in place which may allow the compilers to
 197  * generate better code.
 198  * The increment or decrement is known and therefore one boundary check can
 199  * be omitted.
 200  *
 201  * NOTE: These functions are very performance sensitive. Change only
 202  * with care.
 203  *
 204  * Some processors have inc/dec instructions that are atomic vs an interrupt.
 205  * However, the code must first determine the differential location in a zone
 206  * based on the processor number and then inc/dec the counter. There is no
 207  * guarantee without disabling preemption that the processor will not change
 208  * in between and therefore the atomicity vs. interrupt cannot be exploited
 209  * in a useful way here.
 210  */
 211 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
 212 {
 213         struct per_cpu_pageset *pcp = zone_pcp(zone, smp_processor_id());
 214         s8 *p = pcp->vm_stat_diff + item;
 215
 216         (*p)++;
 217
 218         if (unlikely(*p > pcp->stat_threshold)) {
 219                 int overstep = pcp->stat_threshold / 2;
 220
 221                 zone_page_state_add(*p + overstep, zone, item);
 222                 *p = -overstep;
 223         }
 224 }
 225
 226 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
 227 {
 228         __inc_zone_state(page_zone(page), item);
 229 }
 230 EXPORT_SYMBOL(__inc_zone_page_state);
 231
 232 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
 233 {
 234         struct per_cpu_pageset *pcp = zone_pcp(zone, smp_processor_id());
 235         s8 *p = pcp->vm_stat_diff + item;
 236
 237         (*p)--;
 238
 239         if (unlikely(*p < - pcp->stat_threshold)) {
 240                 int overstep = pcp->stat_threshold / 2;
 241
 242                 zone_page_state_add(*p - overstep, zone, item);
 243                 *p = overstep;
 244         }
 245 }
 246
 247 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
 248 {
 249         __dec_zone_state(page_zone(page), item);
 250 }
 251 EXPORT_SYMBOL(__dec_zone_page_state);
 252
 253 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
 254 {
 255         unsigned long flags;
 256
 257         local_irq_save(flags);
 258         __inc_zone_state(zone, item);
 259         local_irq_restore(flags);
 260 }
 261
 262 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 263 {
 264         unsigned long flags;
 265         struct zone *zone;
 266
 267         zone = page_zone(page);
 268         local_irq_save(flags);
 269         __inc_zone_state(zone, item);
 270         local_irq_restore(flags);
 271 }
 272 EXPORT_SYMBOL(inc_zone_page_state);
 273
 274 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 275 {
 276         unsigned long flags;
 277
 278         local_irq_save(flags);
 279         __dec_zone_page_state(page, item);
 280         local_irq_restore(flags);
 281 }
 282 EXPORT_SYMBOL(dec_zone_page_state);
 283
 284 /*
 285  * Update the zone counters for one cpu.
 286  *
 287  * Note that refresh_cpu_vm_stats strives to only access
 288  * node local memory. The per cpu pagesets on remote zones are placed
 289  * in the memory local to the processor using that pageset. So the
 290  * loop over all zones will access a series of cachelines local to
 291  * the processor.
 292  *
 293  * The call to zone_page_state_add updates the cachelines with the
 294  * statistics in the remote zone struct as well as the global cachelines
 295  * with the global counters. These could cause remote node cache line
 296  * bouncing and will have to be only done when necessary.
 297  */
 298 void refresh_cpu_vm_stats(int cpu)
 299 {
 300         struct zone *zone;
 301         int i;
 302         unsigned long flags;
 303
 304         for_each_zone(zone) {
 305                 struct per_cpu_pageset *p;
 306
 307                 if (!populated_zone(zone))
 308                         continue;
 309
 310                 p = zone_pcp(zone, cpu);
 311
 312                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 313                         if (p->vm_stat_diff[i]) {
 314                                 local_irq_save(flags);
 315                                 zone_page_state_add(p->vm_stat_diff[i],
 316                                         zone, i);
 317                                 p->vm_stat_diff[i] = 0;
 318 #ifdef CONFIG_NUMA
 319                                 /* 3 seconds idle till flush */
 320                                 p->expire = 3;
 321 #endif
 322                                 local_irq_restore(flags);
 323                         }
 324 #ifdef CONFIG_NUMA
 325                 /*
 326                  * Deal with draining the remote pageset of this
 327                  * processor
 328                  *
 329                  * Check if there are pages remaining in this pageset
 330                  * if not then there is nothing to expire.
 331                  */
 332                 if (!p->expire || (!p->pcp[0].count && !p->pcp[1].count))
 333                         continue;
 334
 335                 /*
 336                  * We never drain zones local to this processor.
 337                  */
 338                 if (zone_to_nid(zone) == numa_node_id()) {
 339                         p->expire = 0;
 340                         continue;
 341                 }
 342
 343                 p->expire--;
 344                 if (p->expire)
 345                         continue;
 346
 347                 if (p->pcp[0].count)
 348                         drain_zone_pages(zone, p->pcp + 0);
 349
 350                 if (p->pcp[1].count)
 351                         drain_zone_pages(zone, p->pcp + 1);
 352 #endif
 353         }
 354 }
 355
 356 static void __refresh_cpu_vm_stats(void *dummy)
 357 {
 358         refresh_cpu_vm_stats(smp_processor_id());
 359 }
 360
 361 /*
 362  * Consolidate all counters.
 363  *
 364  * Note that the result is less inaccurate but still inaccurate
 365  * if concurrent processes are allowed to run.
 366  */
 367 void refresh_vm_stats(void)
 368 {
 369         on_each_cpu(__refresh_cpu_vm_stats, NULL, 0, 1);
 370 }
 371 EXPORT_SYMBOL(refresh_vm_stats);
 372
 373 #endif
 374
 375 #ifdef CONFIG_NUMA
 376 /*
 377  * zonelist = the list of zones passed to the allocator
 378  * z        = the zone from which the allocation occurred.
 379  *
 380  * Must be called with interrupts disabled.
 381  */
 382 void zone_statistics(struct zonelist *zonelist, struct zone *z)
 383 {
 384         if (z->zone_pgdat == zonelist->zones[0]->zone_pgdat) {
 385                 __inc_zone_state(z, NUMA_HIT);
 386         } else {
 387                 __inc_zone_state(z, NUMA_MISS);
 388                 __inc_zone_state(zonelist->zones[0], NUMA_FOREIGN);
 389         }
 390         if (z->node == numa_node_id())
 391                 __inc_zone_state(z, NUMA_LOCAL);
 392         else
 393                 __inc_zone_state(z, NUMA_OTHER);
 394 }
 395 #endif
 396
 397 #ifdef CONFIG_PROC_FS
 398
 399 #include <linux/seq_file.h>
 400
 401 static void *frag_start(struct seq_file *m, loff_t *pos)
 402 {
 403         pg_data_t *pgdat;
 404         loff_t node = *pos;
 405         for (pgdat = first_online_pgdat();
 406              pgdat && node;
 407              pgdat = next_online_pgdat(pgdat))
 408                 --node;
 409
 410         return pgdat;
 411 }
 412
 413 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
 414 {
 415         pg_data_t *pgdat = (pg_data_t *)arg;
 416
 417         (*pos)++;
 418         return next_online_pgdat(pgdat);
 419 }
 420
 421 static void frag_stop(struct seq_file *m, void *arg)
 422 {
 423 }
 424
 425 /*
 426  * This walks the free areas for each zone.
 427  */
 428 static int frag_show(struct seq_file *m, void *arg)
 429 {
 430         pg_data_t *pgdat = (pg_data_t *)arg;
 431         struct zone *zone;
 432         struct zone *node_zones = pgdat->node_zones;
 433         unsigned long flags;
 434         int order;
 435
 436         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 437                 if (!populated_zone(zone))
 438                         continue;
 439
 440                 spin_lock_irqsave(&zone->lock, flags);
 441                 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 442                 for (order = 0; order < MAX_ORDER; ++order)
 443                         seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
 444                 spin_unlock_irqrestore(&zone->lock, flags);
 445                 seq_putc(m, '\n');
 446         }
 447         return 0;
 448 }
 449
 450 const struct seq_operations fragmentation_op = {
 451         .start  = frag_start,
 452         .next   = frag_next,
 453         .stop   = frag_stop,
 454         .show   = frag_show,
 455 };
 456
 457 #ifdef CONFIG_ZONE_DMA
 458 #define TEXT_FOR_DMA(xx) xx "_dma",
 459 #else
 460 #define TEXT_FOR_DMA(xx)
 461 #endif
 462
 463 #ifdef CONFIG_ZONE_DMA32
 464 #define TEXT_FOR_DMA32(xx) xx "_dma32",
 465 #else
 466 #define TEXT_FOR_DMA32(xx)
 467 #endif
 468
 469 #ifdef CONFIG_HIGHMEM
 470 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
 471 #else
 472 #define TEXT_FOR_HIGHMEM(xx)
 473 #endif
 474
 475 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
 476                                         TEXT_FOR_HIGHMEM(xx) xx "_movable",
 477
 478 static const char * const vmstat_text[] = {
 479         /* Zoned VM counters */
 480         "nr_free_pages",
 481         "nr_inactive",
 482         "nr_active",
 483         "nr_anon_pages",
 484         "nr_mapped",
 485         "nr_file_pages",
 486         "nr_dirty",
 487         "nr_writeback",
 488         "nr_slab_reclaimable",
 489         "nr_slab_unreclaimable",
 490         "nr_page_table_pages",
 491         "nr_unstable",
 492         "nr_bounce",
 493         "nr_vmscan_write",
 494
 495 #ifdef CONFIG_NUMA
 496         "numa_hit",
 497         "numa_miss",
 498         "numa_foreign",
 499         "numa_interleave",
 500         "numa_local",
 501         "numa_other",
 502 #endif
 503
 504 #ifdef CONFIG_VM_EVENT_COUNTERS
 505         "pgpgin",
 506         "pgpgout",
 507         "pswpin",
 508         "pswpout",
 509
 510         TEXTS_FOR_ZONES("pgalloc")
 511
 512         "pgfree",
 513         "pgactivate",
 514         "pgdeactivate",
 515
 516         "pgfault",
 517         "pgmajfault",
 518
 519         TEXTS_FOR_ZONES("pgrefill")
 520         TEXTS_FOR_ZONES("pgsteal")
 521         TEXTS_FOR_ZONES("pgscan_kswapd")
 522         TEXTS_FOR_ZONES("pgscan_direct")
 523
 524         "pginodesteal",
 525         "slabs_scanned",
 526         "kswapd_steal",
 527         "kswapd_inodesteal",
 528         "pageoutrun",
 529         "allocstall",
 530
 531         "pgrotated",
 532 #endif
 533 };
 534
 535 /*
 536  * Output information about zones in @pgdat.
 537  */
 538 static int zoneinfo_show(struct seq_file *m, void *arg)
 539 {
 540         pg_data_t *pgdat = arg;
 541         struct zone *zone;
 542         struct zone *node_zones = pgdat->node_zones;
 543         unsigned long flags;
 544
 545         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; zone++) {
 546                 int i;
 547
 548                 if (!populated_zone(zone))
 549                         continue;
 550
 551                 spin_lock_irqsave(&zone->lock, flags);
 552                 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
 553                 seq_printf(m,
 554                            "\n  pages free     %lu"
 555                            "\n        min      %lu"
 556                            "\n        low      %lu"
 557                            "\n        high     %lu"
 558                            "\n        scanned  %lu (a: %lu i: %lu)"
 559                            "\n        spanned  %lu"
 560                            "\n        present  %lu",
 561                            zone_page_state(zone, NR_FREE_PAGES),
 562                            zone->pages_min,
 563                            zone->pages_low,
 564                            zone->pages_high,
 565                            zone->pages_scanned,
 566                            zone->nr_scan_active, zone->nr_scan_inactive,
 567                            zone->spanned_pages,
 568                            zone->present_pages);
 569
 570                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 571                         seq_printf(m, "\n    %-12s %lu", vmstat_text[i],
 572                                         zone_page_state(zone, i));
 573
 574                 seq_printf(m,
 575                            "\n        protection: (%lu",
 576                            zone->lowmem_reserve[0]);
 577                 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
 578                         seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
 579                 seq_printf(m,
 580                            ")"
 581                            "\n  pagesets");
 582                 for_each_online_cpu(i) {
 583                         struct per_cpu_pageset *pageset;
 584                         int j;
 585
 586                         pageset = zone_pcp(zone, i);
 587                         for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
 588                                 seq_printf(m,
 589                                            "\n    cpu: %i pcp: %i"
 590                                            "\n              count: %i"
 591                                            "\n              high:  %i"
 592                                            "\n              batch: %i",
 593                                            i, j,
 594                                            pageset->pcp[j].count,
 595                                            pageset->pcp[j].high,
 596                                            pageset->pcp[j].batch);
 597                         }
 598 #ifdef CONFIG_SMP
 599                         seq_printf(m, "\n  vm stats threshold: %d",
 600                                         pageset->stat_threshold);
 601 #endif
 602                 }
 603                 seq_printf(m,
 604                            "\n  all_unreclaimable: %u"
 605                            "\n  prev_priority:     %i"
 606                            "\n  start_pfn:         %lu",
 607                            zone->all_unreclaimable,
 608                            zone->prev_priority,
 609                            zone->zone_start_pfn);
 610                 spin_unlock_irqrestore(&zone->lock, flags);
 611                 seq_putc(m, '\n');
 612         }
 613         return 0;
 614 }
 615
 616 const struct seq_operations zoneinfo_op = {
 617         .start  = frag_start, /* iterate over all zones. The same as in
 618                                * fragmentation. */
 619         .next   = frag_next,
 620         .stop   = frag_stop,
 621         .show   = zoneinfo_show,
 622 };
 623
 624 static void *vmstat_start(struct seq_file *m, loff_t *pos)
 625 {
 626         unsigned long *v;
 627 #ifdef CONFIG_VM_EVENT_COUNTERS
 628         unsigned long *e;
 629 #endif
 630         int i;
 631
 632         if (*pos >= ARRAY_SIZE(vmstat_text))
 633                 return NULL;
 634
 635 #ifdef CONFIG_VM_EVENT_COUNTERS
 636         v = kmalloc(NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long)
 637                         + sizeof(struct vm_event_state), GFP_KERNEL);
 638 #else
 639         v = kmalloc(NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long),
 640                         GFP_KERNEL);
 641 #endif
 642         m->private = v;
 643         if (!v)
 644                 return ERR_PTR(-ENOMEM);
 645         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 646                 v[i] = global_page_state(i);
 647 #ifdef CONFIG_VM_EVENT_COUNTERS
 648         e = v + NR_VM_ZONE_STAT_ITEMS;
 649         all_vm_events(e);
 650         e[PGPGIN] /= 2;         /* sectors -> kbytes */
 651         e[PGPGOUT] /= 2;
 652 #endif
 653         return v + *pos;
 654 }
 655
 656 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
 657 {
 658         (*pos)++;
 659         if (*pos >= ARRAY_SIZE(vmstat_text))
 660                 return NULL;
 661         return (unsigned long *)m->private + *pos;
 662 }
 663
 664 static int vmstat_show(struct seq_file *m, void *arg)
 665 {
 666         unsigned long *l = arg;
 667         unsigned long off = l - (unsigned long *)m->private;
 668
 669         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
 670         return 0;
 671 }
 672
 673 static void vmstat_stop(struct seq_file *m, void *arg)
 674 {
 675         kfree(m->private);
 676         m->private = NULL;
 677 }
 678
 679 const struct seq_operations vmstat_op = {
 680         .start  = vmstat_start,
 681         .next   = vmstat_next,
 682         .stop   = vmstat_stop,
 683         .show   = vmstat_show,
 684 };
 685
 686 #endif /* CONFIG_PROC_FS */
 687
 688 #ifdef CONFIG_SMP
 689 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
 690 int sysctl_stat_interval __read_mostly = HZ;
 691
 692 static void vmstat_update(struct work_struct *w)
 693 {
 694         refresh_cpu_vm_stats(smp_processor_id());
 695         schedule_delayed_work(&__get_cpu_var(vmstat_work),
 696                 sysctl_stat_interval);
 697 }
 698
 699 static void __devinit start_cpu_timer(int cpu)
 700 {
 701         struct delayed_work *vmstat_work = &per_cpu(vmstat_work, cpu);
 702
 703         INIT_DELAYED_WORK_DEFERRABLE(vmstat_work, vmstat_update);
 704         schedule_delayed_work_on(cpu, vmstat_work, HZ + cpu);
 705 }
 706
 707 /*
 708  * Use the cpu notifier to insure that the thresholds are recalculated
 709  * when necessary.
 710  */
 711 static int __cpuinit vmstat_cpuup_callback(struct notifier_block *nfb,
 712                 unsigned long action,
 713                 void *hcpu)
 714 {
 715         long cpu = (long)hcpu;
 716
 717         switch (action) {
 718         case CPU_ONLINE:
 719         case CPU_ONLINE_FROZEN:
 720                 start_cpu_timer(cpu);
 721                 break;
 722         case CPU_DOWN_PREPARE:
 723         case CPU_DOWN_PREPARE_FROZEN:
 724                 cancel_rearming_delayed_work(&per_cpu(vmstat_work, cpu));
 725                 per_cpu(vmstat_work, cpu).work.func = NULL;
 726                 break;
 727         case CPU_DOWN_FAILED:
 728         case CPU_DOWN_FAILED_FROZEN:
 729                 start_cpu_timer(cpu);
 730                 break;
 731         case CPU_DEAD:
 732         case CPU_DEAD_FROZEN:
 733                 refresh_zone_stat_thresholds();
 734                 break;
 735         default:
 736                 break;
 737         }
 738         return NOTIFY_OK;
 739 }
 740
 741 static struct notifier_block __cpuinitdata vmstat_notifier =
 742         { &vmstat_cpuup_callback, NULL, 0 };
 743
 744 int __init setup_vmstat(void)
 745 {
 746         int cpu;
 747
 748         refresh_zone_stat_thresholds();
 749         register_cpu_notifier(&vmstat_notifier);
 750
 751         for_each_online_cpu(cpu)
 752                 start_cpu_timer(cpu);
 753         return 0;
 754 }
 755 module_init(setup_vmstat)
 756 #endif