1 /* 2 * linux/mm/vmstat.c 3 * 4 * Manages VM statistics 5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 6 * 7 * zoned VM statistics 8 * Copyright (C) 2006 Silicon Graphics, Inc., 9 * Christoph Lameter <christoph@lameter.com> 10 */ 11 12 #include <linux/mm.h> 13 #include <linux/module.h> 14 #include <linux/cpu.h> 15 16 #ifdef CONFIG_VM_EVENT_COUNTERS 17 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}}; 18 EXPORT_PER_CPU_SYMBOL(vm_event_states); 19 20 static void sum_vm_events(unsigned long *ret, cpumask_t *cpumask) 21 { 22 int cpu = 0; 23 int i; 24 25 memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long)); 26 27 cpu = first_cpu(*cpumask); 28 while (cpu < NR_CPUS) { 29 struct vm_event_state *this = &per_cpu(vm_event_states, cpu); 30 31 cpu = next_cpu(cpu, *cpumask); 32 33 if (cpu < NR_CPUS) 34 prefetch(&per_cpu(vm_event_states, cpu)); 35 36 37 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) 38 ret[i] += this->event[i]; 39 } 40 } 41 42 /* 43 * Accumulate the vm event counters across all CPUs. 44 * The result is unavoidably approximate - it can change 45 * during and after execution of this function. 46 */ 47 void all_vm_events(unsigned long *ret) 48 { 49 sum_vm_events(ret, &cpu_online_map); 50 } 51 EXPORT_SYMBOL_GPL(all_vm_events); 52 53 #ifdef CONFIG_HOTPLUG 54 /* 55 * Fold the foreign cpu events into our own. 56 * 57 * This is adding to the events on one processor 58 * but keeps the global counts constant. 59 */ 60 void vm_events_fold_cpu(int cpu) 61 { 62 struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu); 63 int i; 64 65 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) { 66 count_vm_events(i, fold_state->event[i]); 67 fold_state->event[i] = 0; 68 } 69 } 70 #endif /* CONFIG_HOTPLUG */ 71 72 #endif /* CONFIG_VM_EVENT_COUNTERS */ 73 74 /* 75 * Manage combined zone based / global counters 76 * 77 * vm_stat contains the global counters 78 */ 79 atomic_long_t vm_stat[NR_VM_ZONE_STAT_ITEMS]; 80 EXPORT_SYMBOL(vm_stat); 81 82 #ifdef CONFIG_SMP 83 84 static int calculate_threshold(struct zone *zone) 85 { 86 int threshold; 87 int mem; /* memory in 128 MB units */ 88 89 /* 90 * The threshold scales with the number of processors and the amount 91 * of memory per zone. More memory means that we can defer updates for 92 * longer, more processors could lead to more contention. 93 * fls() is used to have a cheap way of logarithmic scaling. 94 * 95 * Some sample thresholds: 96 * 97 * Threshold Processors (fls) Zonesize fls(mem+1) 98 * ------------------------------------------------------------------ 99 * 8 1 1 0.9-1 GB 4 100 * 16 2 2 0.9-1 GB 4 101 * 20 2 2 1-2 GB 5 102 * 24 2 2 2-4 GB 6 103 * 28 2 2 4-8 GB 7 104 * 32 2 2 8-16 GB 8 105 * 4 2 2 <128M 1 106 * 30 4 3 2-4 GB 5 107 * 48 4 3 8-16 GB 8 108 * 32 8 4 1-2 GB 4 109 * 32 8 4 0.9-1GB 4 110 * 10 16 5 <128M 1 111 * 40 16 5 900M 4 112 * 70 64 7 2-4 GB 5 113 * 84 64 7 4-8 GB 6 114 * 108 512 9 4-8 GB 6 115 * 125 1024 10 8-16 GB 8 116 * 125 1024 10 16-32 GB 9 117 */ 118 119 mem = zone->present_pages >> (27 - PAGE_SHIFT); 120 121 threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem)); 122 123 /* 124 * Maximum threshold is 125 125 */ 126 threshold = min(125, threshold); 127 128 return threshold; 129 } 130 131 /* 132 * Refresh the thresholds for each zone. 133 */ 134 static void refresh_zone_stat_thresholds(void) 135 { 136 struct zone *zone; 137 int cpu; 138 int threshold; 139 140 for_each_zone(zone) { 141 142 if (!zone->present_pages) 143 continue; 144 145 threshold = calculate_threshold(zone); 146 147 for_each_online_cpu(cpu) 148 zone_pcp(zone, cpu)->stat_threshold = threshold; 149 } 150 } 151 152 /* 153 * For use when we know that interrupts are disabled. 154 */ 155 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item, 156 int delta) 157 { 158 struct per_cpu_pageset *pcp = zone_pcp(zone, smp_processor_id()); 159 s8 *p = pcp->vm_stat_diff + item; 160 long x; 161 162 x = delta + *p; 163 164 if (unlikely(x > pcp->stat_threshold || x < -pcp->stat_threshold)) { 165 zone_page_state_add(x, zone, item); 166 x = 0; 167 } 168 *p = x; 169 } 170 EXPORT_SYMBOL(__mod_zone_page_state); 171 172 /* 173 * For an unknown interrupt state 174 */ 175 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item, 176 int delta) 177 { 178 unsigned long flags; 179 180 local_irq_save(flags); 181 __mod_zone_page_state(zone, item, delta); 182 local_irq_restore(flags); 183 } 184 EXPORT_SYMBOL(mod_zone_page_state); 185 186 /* 187 * Optimized increment and decrement functions. 188 * 189 * These are only for a single page and therefore can take a struct page * 190 * argument instead of struct zone *. This allows the inclusion of the code 191 * generated for page_zone(page) into the optimized functions. 192 * 193 * No overflow check is necessary and therefore the differential can be 194 * incremented or decremented in place which may allow the compilers to 195 * generate better code. 196 * The increment or decrement is known and therefore one boundary check can 197 * be omitted. 198 * 199 * NOTE: These functions are very performance sensitive. Change only 200 * with care. 201 * 202 * Some processors have inc/dec instructions that are atomic vs an interrupt. 203 * However, the code must first determine the differential location in a zone 204 * based on the processor number and then inc/dec the counter. There is no 205 * guarantee without disabling preemption that the processor will not change 206 * in between and therefore the atomicity vs. interrupt cannot be exploited 207 * in a useful way here. 208 */ 209 void __inc_zone_state(struct zone *zone, enum zone_stat_item item) 210 { 211 struct per_cpu_pageset *pcp = zone_pcp(zone, smp_processor_id()); 212 s8 *p = pcp->vm_stat_diff + item; 213 214 (*p)++; 215 216 if (unlikely(*p > pcp->stat_threshold)) { 217 int overstep = pcp->stat_threshold / 2; 218 219 zone_page_state_add(*p + overstep, zone, item); 220 *p = -overstep; 221 } 222 } 223 224 void __inc_zone_page_state(struct page *page, enum zone_stat_item item) 225 { 226 __inc_zone_state(page_zone(page), item); 227 } 228 EXPORT_SYMBOL(__inc_zone_page_state); 229 230 void __dec_zone_state(struct zone *zone, enum zone_stat_item item) 231 { 232 struct per_cpu_pageset *pcp = zone_pcp(zone, smp_processor_id()); 233 s8 *p = pcp->vm_stat_diff + item; 234 235 (*p)--; 236 237 if (unlikely(*p < - pcp->stat_threshold)) { 238 int overstep = pcp->stat_threshold / 2; 239 240 zone_page_state_add(*p - overstep, zone, item); 241 *p = overstep; 242 } 243 } 244 245 void __dec_zone_page_state(struct page *page, enum zone_stat_item item) 246 { 247 __dec_zone_state(page_zone(page), item); 248 } 249 EXPORT_SYMBOL(__dec_zone_page_state); 250 251 void inc_zone_state(struct zone *zone, enum zone_stat_item item) 252 { 253 unsigned long flags; 254 255 local_irq_save(flags); 256 __inc_zone_state(zone, item); 257 local_irq_restore(flags); 258 } 259 260 void inc_zone_page_state(struct page *page, enum zone_stat_item item) 261 { 262 unsigned long flags; 263 struct zone *zone; 264 265 zone = page_zone(page); 266 local_irq_save(flags); 267 __inc_zone_state(zone, item); 268 local_irq_restore(flags); 269 } 270 EXPORT_SYMBOL(inc_zone_page_state); 271 272 void dec_zone_page_state(struct page *page, enum zone_stat_item item) 273 { 274 unsigned long flags; 275 276 local_irq_save(flags); 277 __dec_zone_page_state(page, item); 278 local_irq_restore(flags); 279 } 280 EXPORT_SYMBOL(dec_zone_page_state); 281 282 /* 283 * Update the zone counters for one cpu. 284 * 285 * Note that refresh_cpu_vm_stats strives to only access 286 * node local memory. The per cpu pagesets on remote zones are placed 287 * in the memory local to the processor using that pageset. So the 288 * loop over all zones will access a series of cachelines local to 289 * the processor. 290 * 291 * The call to zone_page_state_add updates the cachelines with the 292 * statistics in the remote zone struct as well as the global cachelines 293 * with the global counters. These could cause remote node cache line 294 * bouncing and will have to be only done when necessary. 295 */ 296 void refresh_cpu_vm_stats(int cpu) 297 { 298 struct zone *zone; 299 int i; 300 unsigned long flags; 301 302 for_each_zone(zone) { 303 struct per_cpu_pageset *p; 304 305 if (!populated_zone(zone)) 306 continue; 307 308 p = zone_pcp(zone, cpu); 309 310 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) 311 if (p->vm_stat_diff[i]) { 312 local_irq_save(flags); 313 zone_page_state_add(p->vm_stat_diff[i], 314 zone, i); 315 p->vm_stat_diff[i] = 0; 316 #ifdef CONFIG_NUMA 317 /* 3 seconds idle till flush */ 318 p->expire = 3; 319 #endif 320 local_irq_restore(flags); 321 } 322 #ifdef CONFIG_NUMA 323 /* 324 * Deal with draining the remote pageset of this 325 * processor 326 * 327 * Check if there are pages remaining in this pageset 328 * if not then there is nothing to expire. 329 */ 330 if (!p->expire || (!p->pcp[0].count && !p->pcp[1].count)) 331 continue; 332 333 /* 334 * We never drain zones local to this processor. 335 */ 336 if (zone_to_nid(zone) == numa_node_id()) { 337 p->expire = 0; 338 continue; 339 } 340 341 p->expire--; 342 if (p->expire) 343 continue; 344 345 if (p->pcp[0].count) 346 drain_zone_pages(zone, p->pcp + 0); 347 348 if (p->pcp[1].count) 349 drain_zone_pages(zone, p->pcp + 1); 350 #endif 351 } 352 } 353 354 static void __refresh_cpu_vm_stats(void *dummy) 355 { 356 refresh_cpu_vm_stats(smp_processor_id()); 357 } 358 359 /* 360 * Consolidate all counters. 361 * 362 * Note that the result is less inaccurate but still inaccurate 363 * if concurrent processes are allowed to run. 364 */ 365 void refresh_vm_stats(void) 366 { 367 on_each_cpu(__refresh_cpu_vm_stats, NULL, 0, 1); 368 } 369 EXPORT_SYMBOL(refresh_vm_stats); 370 371 #endif 372 373 #ifdef CONFIG_NUMA 374 /* 375 * zonelist = the list of zones passed to the allocator 376 * z = the zone from which the allocation occurred. 377 * 378 * Must be called with interrupts disabled. 379 */ 380 void zone_statistics(struct zonelist *zonelist, struct zone *z) 381 { 382 if (z->zone_pgdat == zonelist->zones[0]->zone_pgdat) { 383 __inc_zone_state(z, NUMA_HIT); 384 } else { 385 __inc_zone_state(z, NUMA_MISS); 386 __inc_zone_state(zonelist->zones[0], NUMA_FOREIGN); 387 } 388 if (z->node == numa_node_id()) 389 __inc_zone_state(z, NUMA_LOCAL); 390 else 391 __inc_zone_state(z, NUMA_OTHER); 392 } 393 #endif 394 395 #ifdef CONFIG_PROC_FS 396 397 #include <linux/seq_file.h> 398 399 static void *frag_start(struct seq_file *m, loff_t *pos) 400 { 401 pg_data_t *pgdat; 402 loff_t node = *pos; 403 for (pgdat = first_online_pgdat(); 404 pgdat && node; 405 pgdat = next_online_pgdat(pgdat)) 406 --node; 407 408 return pgdat; 409 } 410 411 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos) 412 { 413 pg_data_t *pgdat = (pg_data_t *)arg; 414 415 (*pos)++; 416 return next_online_pgdat(pgdat); 417 } 418 419 static void frag_stop(struct seq_file *m, void *arg) 420 { 421 } 422 423 /* 424 * This walks the free areas for each zone. 425 */ 426 static int frag_show(struct seq_file *m, void *arg) 427 { 428 pg_data_t *pgdat = (pg_data_t *)arg; 429 struct zone *zone; 430 struct zone *node_zones = pgdat->node_zones; 431 unsigned long flags; 432 int order; 433 434 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) { 435 if (!populated_zone(zone)) 436 continue; 437 438 spin_lock_irqsave(&zone->lock, flags); 439 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name); 440 for (order = 0; order < MAX_ORDER; ++order) 441 seq_printf(m, "%6lu ", zone->free_area[order].nr_free); 442 spin_unlock_irqrestore(&zone->lock, flags); 443 seq_putc(m, '\n'); 444 } 445 return 0; 446 } 447 448 const struct seq_operations fragmentation_op = { 449 .start = frag_start, 450 .next = frag_next, 451 .stop = frag_stop, 452 .show = frag_show, 453 }; 454 455 #ifdef CONFIG_ZONE_DMA 456 #define TEXT_FOR_DMA(xx) xx "_dma", 457 #else 458 #define TEXT_FOR_DMA(xx) 459 #endif 460 461 #ifdef CONFIG_ZONE_DMA32 462 #define TEXT_FOR_DMA32(xx) xx "_dma32", 463 #else 464 #define TEXT_FOR_DMA32(xx) 465 #endif 466 467 #ifdef CONFIG_HIGHMEM 468 #define TEXT_FOR_HIGHMEM(xx) xx "_high", 469 #else 470 #define TEXT_FOR_HIGHMEM(xx) 471 #endif 472 473 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \ 474 TEXT_FOR_HIGHMEM(xx) 475 476 static const char * const vmstat_text[] = { 477 /* Zoned VM counters */ 478 "nr_free_pages", 479 "nr_active", 480 "nr_inactive", 481 "nr_anon_pages", 482 "nr_mapped", 483 "nr_file_pages", 484 "nr_dirty", 485 "nr_writeback", 486 "nr_slab_reclaimable", 487 "nr_slab_unreclaimable", 488 "nr_page_table_pages", 489 "nr_unstable", 490 "nr_bounce", 491 "nr_vmscan_write", 492 493 #ifdef CONFIG_NUMA 494 "numa_hit", 495 "numa_miss", 496 "numa_foreign", 497 "numa_interleave", 498 "numa_local", 499 "numa_other", 500 #endif 501 502 #ifdef CONFIG_VM_EVENT_COUNTERS 503 "pgpgin", 504 "pgpgout", 505 "pswpin", 506 "pswpout", 507 508 TEXTS_FOR_ZONES("pgalloc") 509 510 "pgfree", 511 "pgactivate", 512 "pgdeactivate", 513 514 "pgfault", 515 "pgmajfault", 516 517 TEXTS_FOR_ZONES("pgrefill") 518 TEXTS_FOR_ZONES("pgsteal") 519 TEXTS_FOR_ZONES("pgscan_kswapd") 520 TEXTS_FOR_ZONES("pgscan_direct") 521 522 "pginodesteal", 523 "slabs_scanned", 524 "kswapd_steal", 525 "kswapd_inodesteal", 526 "pageoutrun", 527 "allocstall", 528 529 "pgrotated", 530 #endif 531 }; 532 533 /* 534 * Output information about zones in @pgdat. 535 */ 536 static int zoneinfo_show(struct seq_file *m, void *arg) 537 { 538 pg_data_t *pgdat = arg; 539 struct zone *zone; 540 struct zone *node_zones = pgdat->node_zones; 541 unsigned long flags; 542 543 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; zone++) { 544 int i; 545 546 if (!populated_zone(zone)) 547 continue; 548 549 spin_lock_irqsave(&zone->lock, flags); 550 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name); 551 seq_printf(m, 552 "\n pages free %lu" 553 "\n min %lu" 554 "\n low %lu" 555 "\n high %lu" 556 "\n scanned %lu (a: %lu i: %lu)" 557 "\n spanned %lu" 558 "\n present %lu", 559 zone_page_state(zone, NR_FREE_PAGES), 560 zone->pages_min, 561 zone->pages_low, 562 zone->pages_high, 563 zone->pages_scanned, 564 zone->nr_scan_active, zone->nr_scan_inactive, 565 zone->spanned_pages, 566 zone->present_pages); 567 568 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) 569 seq_printf(m, "\n %-12s %lu", vmstat_text[i], 570 zone_page_state(zone, i)); 571 572 seq_printf(m, 573 "\n protection: (%lu", 574 zone->lowmem_reserve[0]); 575 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++) 576 seq_printf(m, ", %lu", zone->lowmem_reserve[i]); 577 seq_printf(m, 578 ")" 579 "\n pagesets"); 580 for_each_online_cpu(i) { 581 struct per_cpu_pageset *pageset; 582 int j; 583 584 pageset = zone_pcp(zone, i); 585 for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) { 586 seq_printf(m, 587 "\n cpu: %i pcp: %i" 588 "\n count: %i" 589 "\n high: %i" 590 "\n batch: %i", 591 i, j, 592 pageset->pcp[j].count, 593 pageset->pcp[j].high, 594 pageset->pcp[j].batch); 595 } 596 #ifdef CONFIG_SMP 597 seq_printf(m, "\n vm stats threshold: %d", 598 pageset->stat_threshold); 599 #endif 600 } 601 seq_printf(m, 602 "\n all_unreclaimable: %u" 603 "\n prev_priority: %i" 604 "\n start_pfn: %lu", 605 zone->all_unreclaimable, 606 zone->prev_priority, 607 zone->zone_start_pfn); 608 spin_unlock_irqrestore(&zone->lock, flags); 609 seq_putc(m, '\n'); 610 } 611 return 0; 612 } 613 614 const struct seq_operations zoneinfo_op = { 615 .start = frag_start, /* iterate over all zones. The same as in 616 * fragmentation. */ 617 .next = frag_next, 618 .stop = frag_stop, 619 .show = zoneinfo_show, 620 }; 621 622 static void *vmstat_start(struct seq_file *m, loff_t *pos) 623 { 624 unsigned long *v; 625 #ifdef CONFIG_VM_EVENT_COUNTERS 626 unsigned long *e; 627 #endif 628 int i; 629 630 if (*pos >= ARRAY_SIZE(vmstat_text)) 631 return NULL; 632 633 #ifdef CONFIG_VM_EVENT_COUNTERS 634 v = kmalloc(NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) 635 + sizeof(struct vm_event_state), GFP_KERNEL); 636 #else 637 v = kmalloc(NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long), 638 GFP_KERNEL); 639 #endif 640 m->private = v; 641 if (!v) 642 return ERR_PTR(-ENOMEM); 643 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) 644 v[i] = global_page_state(i); 645 #ifdef CONFIG_VM_EVENT_COUNTERS 646 e = v + NR_VM_ZONE_STAT_ITEMS; 647 all_vm_events(e); 648 e[PGPGIN] /= 2; /* sectors -> kbytes */ 649 e[PGPGOUT] /= 2; 650 #endif 651 return v + *pos; 652 } 653 654 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos) 655 { 656 (*pos)++; 657 if (*pos >= ARRAY_SIZE(vmstat_text)) 658 return NULL; 659 return (unsigned long *)m->private + *pos; 660 } 661 662 static int vmstat_show(struct seq_file *m, void *arg) 663 { 664 unsigned long *l = arg; 665 unsigned long off = l - (unsigned long *)m->private; 666 667 seq_printf(m, "%s %lu\n", vmstat_text[off], *l); 668 return 0; 669 } 670 671 static void vmstat_stop(struct seq_file *m, void *arg) 672 { 673 kfree(m->private); 674 m->private = NULL; 675 } 676 677 const struct seq_operations vmstat_op = { 678 .start = vmstat_start, 679 .next = vmstat_next, 680 .stop = vmstat_stop, 681 .show = vmstat_show, 682 }; 683 684 #endif /* CONFIG_PROC_FS */ 685 686 #ifdef CONFIG_SMP 687 static DEFINE_PER_CPU(struct delayed_work, vmstat_work); 688 int sysctl_stat_interval __read_mostly = HZ; 689 690 static void vmstat_update(struct work_struct *w) 691 { 692 refresh_cpu_vm_stats(smp_processor_id()); 693 schedule_delayed_work(&__get_cpu_var(vmstat_work), 694 sysctl_stat_interval); 695 } 696 697 static void __devinit start_cpu_timer(int cpu) 698 { 699 struct delayed_work *vmstat_work = &per_cpu(vmstat_work, cpu); 700 701 INIT_DELAYED_WORK_DEFERRABLE(vmstat_work, vmstat_update); 702 schedule_delayed_work_on(cpu, vmstat_work, HZ + cpu); 703 } 704 705 /* 706 * Use the cpu notifier to insure that the thresholds are recalculated 707 * when necessary. 708 */ 709 static int __cpuinit vmstat_cpuup_callback(struct notifier_block *nfb, 710 unsigned long action, 711 void *hcpu) 712 { 713 long cpu = (long)hcpu; 714 715 switch (action) { 716 case CPU_ONLINE: 717 case CPU_ONLINE_FROZEN: 718 start_cpu_timer(cpu); 719 break; 720 case CPU_DOWN_PREPARE: 721 case CPU_DOWN_PREPARE_FROZEN: 722 cancel_rearming_delayed_work(&per_cpu(vmstat_work, cpu)); 723 per_cpu(vmstat_work, cpu).work.func = NULL; 724 break; 725 case CPU_DOWN_FAILED: 726 case CPU_DOWN_FAILED_FROZEN: 727 start_cpu_timer(cpu); 728 break; 729 case CPU_DEAD: 730 case CPU_DEAD_FROZEN: 731 refresh_zone_stat_thresholds(); 732 break; 733 default: 734 break; 735 } 736 return NOTIFY_OK; 737 } 738 739 static struct notifier_block __cpuinitdata vmstat_notifier = 740 { &vmstat_cpuup_callback, NULL, 0 }; 741 742 int __init setup_vmstat(void) 743 { 744 int cpu; 745 746 refresh_zone_stat_thresholds(); 747 register_cpu_notifier(&vmstat_notifier); 748 749 for_each_online_cpu(cpu) 750 start_cpu_timer(cpu); 751 return 0; 752 } 753 module_init(setup_vmstat) 754 #endif 755