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 #include <linux/fs.h>
  12 #include <linux/mm.h>
  13 #include <linux/err.h>
  14 #include <linux/module.h>
  15 #include <linux/slab.h>
  16 #include <linux/cpu.h>
  17 #include <linux/vmstat.h>
  18 #include <linux/sched.h>
  19 #include <linux/math64.h>
  20 #include <linux/writeback.h>
  21 #include <linux/compaction.h>
  22
  23 #ifdef CONFIG_VM_EVENT_COUNTERS
  24 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
  25 EXPORT_PER_CPU_SYMBOL(vm_event_states);
  26
  27 static void sum_vm_events(unsigned long *ret)
  28 {
  29         int cpu;
  30         int i;
  31
  32         memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
  33
  34         for_each_online_cpu(cpu) {
  35                 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
  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         get_online_cpus();
  50         sum_vm_events(ret);
  51         put_online_cpus();
  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_populated_zone(zone) {
 143                 unsigned long max_drift, tolerate_drift;
 144
 145                 threshold = calculate_threshold(zone);
 146
 147                 for_each_online_cpu(cpu)
 148                         per_cpu_ptr(zone->pageset, cpu)->stat_threshold
 149                                                         = threshold;
 150
 151                 /*
 152                  * Only set percpu_drift_mark if there is a danger that
 153                  * NR_FREE_PAGES reports the low watermark is ok when in fact
 154                  * the min watermark could be breached by an allocation
 155                  */
 156                 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
 157                 max_drift = num_online_cpus() * threshold;
 158                 if (max_drift > tolerate_drift)
 159                         zone->percpu_drift_mark = high_wmark_pages(zone) +
 160                                         max_drift;
 161         }
 162 }
 163
 164 /*
 165  * For use when we know that interrupts are disabled.
 166  */
 167 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 168                                 int delta)
 169 {
 170         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 171         s8 __percpu *p = pcp->vm_stat_diff + item;
 172         long x;
 173         long t;
 174
 175         x = delta + __this_cpu_read(*p);
 176
 177         t = __this_cpu_read(pcp->stat_threshold);
 178
 179         if (unlikely(x > t || x < -t)) {
 180                 zone_page_state_add(x, zone, item);
 181                 x = 0;
 182         }
 183         __this_cpu_write(*p, x);
 184 }
 185 EXPORT_SYMBOL(__mod_zone_page_state);
 186
 187 /*
 188  * For an unknown interrupt state
 189  */
 190 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 191                                         int delta)
 192 {
 193         unsigned long flags;
 194
 195         local_irq_save(flags);
 196         __mod_zone_page_state(zone, item, delta);
 197         local_irq_restore(flags);
 198 }
 199 EXPORT_SYMBOL(mod_zone_page_state);
 200
 201 /*
 202  * Optimized increment and decrement functions.
 203  *
 204  * These are only for a single page and therefore can take a struct page *
 205  * argument instead of struct zone *. This allows the inclusion of the code
 206  * generated for page_zone(page) into the optimized functions.
 207  *
 208  * No overflow check is necessary and therefore the differential can be
 209  * incremented or decremented in place which may allow the compilers to
 210  * generate better code.
 211  * The increment or decrement is known and therefore one boundary check can
 212  * be omitted.
 213  *
 214  * NOTE: These functions are very performance sensitive. Change only
 215  * with care.
 216  *
 217  * Some processors have inc/dec instructions that are atomic vs an interrupt.
 218  * However, the code must first determine the differential location in a zone
 219  * based on the processor number and then inc/dec the counter. There is no
 220  * guarantee without disabling preemption that the processor will not change
 221  * in between and therefore the atomicity vs. interrupt cannot be exploited
 222  * in a useful way here.
 223  */
 224 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
 225 {
 226         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 227         s8 __percpu *p = pcp->vm_stat_diff + item;
 228         s8 v, t;
 229
 230         __this_cpu_inc(*p);
 231
 232         v = __this_cpu_read(*p);
 233         t = __this_cpu_read(pcp->stat_threshold);
 234         if (unlikely(v > t)) {
 235                 s8 overstep = t >> 1;
 236
 237                 zone_page_state_add(v + overstep, zone, item);
 238                 __this_cpu_write(*p, -overstep);
 239         }
 240 }
 241
 242 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
 243 {
 244         __inc_zone_state(page_zone(page), item);
 245 }
 246 EXPORT_SYMBOL(__inc_zone_page_state);
 247
 248 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
 249 {
 250         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 251         s8 __percpu *p = pcp->vm_stat_diff + item;
 252         s8 v, t;
 253
 254         __this_cpu_dec(*p);
 255
 256         v = __this_cpu_read(*p);
 257         t = __this_cpu_read(pcp->stat_threshold);
 258         if (unlikely(v < - t)) {
 259                 s8 overstep = t >> 1;
 260
 261                 zone_page_state_add(v - overstep, zone, item);
 262                 __this_cpu_write(*p, overstep);
 263         }
 264 }
 265
 266 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
 267 {
 268         __dec_zone_state(page_zone(page), item);
 269 }
 270 EXPORT_SYMBOL(__dec_zone_page_state);
 271
 272 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
 273 {
 274         unsigned long flags;
 275
 276         local_irq_save(flags);
 277         __inc_zone_state(zone, item);
 278         local_irq_restore(flags);
 279 }
 280
 281 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 282 {
 283         unsigned long flags;
 284         struct zone *zone;
 285
 286         zone = page_zone(page);
 287         local_irq_save(flags);
 288         __inc_zone_state(zone, item);
 289         local_irq_restore(flags);
 290 }
 291 EXPORT_SYMBOL(inc_zone_page_state);
 292
 293 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 294 {
 295         unsigned long flags;
 296
 297         local_irq_save(flags);
 298         __dec_zone_page_state(page, item);
 299         local_irq_restore(flags);
 300 }
 301 EXPORT_SYMBOL(dec_zone_page_state);
 302
 303 /*
 304  * Update the zone counters for one cpu.
 305  *
 306  * The cpu specified must be either the current cpu or a processor that
 307  * is not online. If it is the current cpu then the execution thread must
 308  * be pinned to the current cpu.
 309  *
 310  * Note that refresh_cpu_vm_stats strives to only access
 311  * node local memory. The per cpu pagesets on remote zones are placed
 312  * in the memory local to the processor using that pageset. So the
 313  * loop over all zones will access a series of cachelines local to
 314  * the processor.
 315  *
 316  * The call to zone_page_state_add updates the cachelines with the
 317  * statistics in the remote zone struct as well as the global cachelines
 318  * with the global counters. These could cause remote node cache line
 319  * bouncing and will have to be only done when necessary.
 320  */
 321 void refresh_cpu_vm_stats(int cpu)
 322 {
 323         struct zone *zone;
 324         int i;
 325         int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 326
 327         for_each_populated_zone(zone) {
 328                 struct per_cpu_pageset *p;
 329
 330                 p = per_cpu_ptr(zone->pageset, cpu);
 331
 332                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 333                         if (p->vm_stat_diff[i]) {
 334                                 unsigned long flags;
 335                                 int v;
 336
 337                                 local_irq_save(flags);
 338                                 v = p->vm_stat_diff[i];
 339                                 p->vm_stat_diff[i] = 0;
 340                                 local_irq_restore(flags);
 341                                 atomic_long_add(v, &zone->vm_stat[i]);
 342                                 global_diff[i] += v;
 343 #ifdef CONFIG_NUMA
 344                                 /* 3 seconds idle till flush */
 345                                 p->expire = 3;
 346 #endif
 347                         }
 348                 cond_resched();
 349 #ifdef CONFIG_NUMA
 350                 /*
 351                  * Deal with draining the remote pageset of this
 352                  * processor
 353                  *
 354                  * Check if there are pages remaining in this pageset
 355                  * if not then there is nothing to expire.
 356                  */
 357                 if (!p->expire || !p->pcp.count)
 358                         continue;
 359
 360                 /*
 361                  * We never drain zones local to this processor.
 362                  */
 363                 if (zone_to_nid(zone) == numa_node_id()) {
 364                         p->expire = 0;
 365                         continue;
 366                 }
 367
 368                 p->expire--;
 369                 if (p->expire)
 370                         continue;
 371
 372                 if (p->pcp.count)
 373                         drain_zone_pages(zone, &p->pcp);
 374 #endif
 375         }
 376
 377         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 378                 if (global_diff[i])
 379                         atomic_long_add(global_diff[i], &vm_stat[i]);
 380 }
 381
 382 #endif
 383
 384 #ifdef CONFIG_NUMA
 385 /*
 386  * zonelist = the list of zones passed to the allocator
 387  * z        = the zone from which the allocation occurred.
 388  *
 389  * Must be called with interrupts disabled.
 390  */
 391 void zone_statistics(struct zone *preferred_zone, struct zone *z)
 392 {
 393         if (z->zone_pgdat == preferred_zone->zone_pgdat) {
 394                 __inc_zone_state(z, NUMA_HIT);
 395         } else {
 396                 __inc_zone_state(z, NUMA_MISS);
 397                 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
 398         }
 399         if (z->node == numa_node_id())
 400                 __inc_zone_state(z, NUMA_LOCAL);
 401         else
 402                 __inc_zone_state(z, NUMA_OTHER);
 403 }
 404 #endif
 405
 406 #ifdef CONFIG_COMPACTION
 407
 408 struct contig_page_info {
 409         unsigned long free_pages;
 410         unsigned long free_blocks_total;
 411         unsigned long free_blocks_suitable;
 412 };
 413
 414 /*
 415  * Calculate the number of free pages in a zone, how many contiguous
 416  * pages are free and how many are large enough to satisfy an allocation of
 417  * the target size. Note that this function makes no attempt to estimate
 418  * how many suitable free blocks there *might* be if MOVABLE pages were
 419  * migrated. Calculating that is possible, but expensive and can be
 420  * figured out from userspace
 421  */
 422 static void fill_contig_page_info(struct zone *zone,
 423                                 unsigned int suitable_order,
 424                                 struct contig_page_info *info)
 425 {
 426         unsigned int order;
 427
 428         info->free_pages = 0;
 429         info->free_blocks_total = 0;
 430         info->free_blocks_suitable = 0;
 431
 432         for (order = 0; order < MAX_ORDER; order++) {
 433                 unsigned long blocks;
 434
 435                 /* Count number of free blocks */
 436                 blocks = zone->free_area[order].nr_free;
 437                 info->free_blocks_total += blocks;
 438
 439                 /* Count free base pages */
 440                 info->free_pages += blocks << order;
 441
 442                 /* Count the suitable free blocks */
 443                 if (order >= suitable_order)
 444                         info->free_blocks_suitable += blocks <<
 445                                                 (order - suitable_order);
 446         }
 447 }
 448
 449 /*
 450  * A fragmentation index only makes sense if an allocation of a requested
 451  * size would fail. If that is true, the fragmentation index indicates
 452  * whether external fragmentation or a lack of memory was the problem.
 453  * The value can be used to determine if page reclaim or compaction
 454  * should be used
 455  */
 456 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
 457 {
 458         unsigned long requested = 1UL << order;
 459
 460         if (!info->free_blocks_total)
 461                 return 0;
 462
 463         /* Fragmentation index only makes sense when a request would fail */
 464         if (info->free_blocks_suitable)
 465                 return -1000;
 466
 467         /*
 468          * Index is between 0 and 1 so return within 3 decimal places
 469          *
 470          * 0 => allocation would fail due to lack of memory
 471          * 1 => allocation would fail due to fragmentation
 472          */
 473         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
 474 }
 475
 476 /* Same as __fragmentation index but allocs contig_page_info on stack */
 477 int fragmentation_index(struct zone *zone, unsigned int order)
 478 {
 479         struct contig_page_info info;
 480
 481         fill_contig_page_info(zone, order, &info);
 482         return __fragmentation_index(order, &info);
 483 }
 484 #endif
 485
 486 #if defined(CONFIG_PROC_FS) || defined(CONFIG_COMPACTION)
 487 #include <linux/proc_fs.h>
 488 #include <linux/seq_file.h>
 489
 490 static char * const migratetype_names[MIGRATE_TYPES] = {
 491         "Unmovable",
 492         "Reclaimable",
 493         "Movable",
 494         "Reserve",
 495         "Isolate",
 496 };
 497
 498 static void *frag_start(struct seq_file *m, loff_t *pos)
 499 {
 500         pg_data_t *pgdat;
 501         loff_t node = *pos;
 502         for (pgdat = first_online_pgdat();
 503              pgdat && node;
 504              pgdat = next_online_pgdat(pgdat))
 505                 --node;
 506
 507         return pgdat;
 508 }
 509
 510 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
 511 {
 512         pg_data_t *pgdat = (pg_data_t *)arg;
 513
 514         (*pos)++;
 515         return next_online_pgdat(pgdat);
 516 }
 517
 518 static void frag_stop(struct seq_file *m, void *arg)
 519 {
 520 }
 521
 522 /* Walk all the zones in a node and print using a callback */
 523 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
 524                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
 525 {
 526         struct zone *zone;
 527         struct zone *node_zones = pgdat->node_zones;
 528         unsigned long flags;
 529
 530         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 531                 if (!populated_zone(zone))
 532                         continue;
 533
 534                 spin_lock_irqsave(&zone->lock, flags);
 535                 print(m, pgdat, zone);
 536                 spin_unlock_irqrestore(&zone->lock, flags);
 537         }
 538 }
 539 #endif
 540
 541 #ifdef CONFIG_PROC_FS
 542 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
 543                                                 struct zone *zone)
 544 {
 545         int order;
 546
 547         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 548         for (order = 0; order < MAX_ORDER; ++order)
 549                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
 550         seq_putc(m, '\n');
 551 }
 552
 553 /*
 554  * This walks the free areas for each zone.
 555  */
 556 static int frag_show(struct seq_file *m, void *arg)
 557 {
 558         pg_data_t *pgdat = (pg_data_t *)arg;
 559         walk_zones_in_node(m, pgdat, frag_show_print);
 560         return 0;
 561 }
 562
 563 static void pagetypeinfo_showfree_print(struct seq_file *m,
 564                                         pg_data_t *pgdat, struct zone *zone)
 565 {
 566         int order, mtype;
 567
 568         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
 569                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
 570                                         pgdat->node_id,
 571                                         zone->name,
 572                                         migratetype_names[mtype]);
 573                 for (order = 0; order < MAX_ORDER; ++order) {
 574                         unsigned long freecount = 0;
 575                         struct free_area *area;
 576                         struct list_head *curr;
 577
 578                         area = &(zone->free_area[order]);
 579
 580                         list_for_each(curr, &area->free_list[mtype])
 581                                 freecount++;
 582                         seq_printf(m, "%6lu ", freecount);
 583                 }
 584                 seq_putc(m, '\n');
 585         }
 586 }
 587
 588 /* Print out the free pages at each order for each migatetype */
 589 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
 590 {
 591         int order;
 592         pg_data_t *pgdat = (pg_data_t *)arg;
 593
 594         /* Print header */
 595         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
 596         for (order = 0; order < MAX_ORDER; ++order)
 597                 seq_printf(m, "%6d ", order);
 598         seq_putc(m, '\n');
 599
 600         walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
 601
 602         return 0;
 603 }
 604
 605 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
 606                                         pg_data_t *pgdat, struct zone *zone)
 607 {
 608         int mtype;
 609         unsigned long pfn;
 610         unsigned long start_pfn = zone->zone_start_pfn;
 611         unsigned long end_pfn = start_pfn + zone->spanned_pages;
 612         unsigned long count[MIGRATE_TYPES] = { 0, };
 613
 614         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
 615                 struct page *page;
 616
 617                 if (!pfn_valid(pfn))
 618                         continue;
 619
 620                 page = pfn_to_page(pfn);
 621
 622                 /* Watch for unexpected holes punched in the memmap */
 623                 if (!memmap_valid_within(pfn, page, zone))
 624                         continue;
 625
 626                 mtype = get_pageblock_migratetype(page);
 627
 628                 if (mtype < MIGRATE_TYPES)
 629                         count[mtype]++;
 630         }
 631
 632         /* Print counts */
 633         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 634         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
 635                 seq_printf(m, "%12lu ", count[mtype]);
 636         seq_putc(m, '\n');
 637 }
 638
 639 /* Print out the free pages at each order for each migratetype */
 640 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
 641 {
 642         int mtype;
 643         pg_data_t *pgdat = (pg_data_t *)arg;
 644
 645         seq_printf(m, "\n%-23s", "Number of blocks type ");
 646         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
 647                 seq_printf(m, "%12s ", migratetype_names[mtype]);
 648         seq_putc(m, '\n');
 649         walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
 650
 651         return 0;
 652 }
 653
 654 /*
 655  * This prints out statistics in relation to grouping pages by mobility.
 656  * It is expensive to collect so do not constantly read the file.
 657  */
 658 static int pagetypeinfo_show(struct seq_file *m, void *arg)
 659 {
 660         pg_data_t *pgdat = (pg_data_t *)arg;
 661
 662         /* check memoryless node */
 663         if (!node_state(pgdat->node_id, N_HIGH_MEMORY))
 664                 return 0;
 665
 666         seq_printf(m, "Page block order: %d\n", pageblock_order);
 667         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
 668         seq_putc(m, '\n');
 669         pagetypeinfo_showfree(m, pgdat);
 670         pagetypeinfo_showblockcount(m, pgdat);
 671
 672         return 0;
 673 }
 674
 675 static const struct seq_operations fragmentation_op = {
 676         .start  = frag_start,
 677         .next   = frag_next,
 678         .stop   = frag_stop,
 679         .show   = frag_show,
 680 };
 681
 682 static int fragmentation_open(struct inode *inode, struct file *file)
 683 {
 684         return seq_open(file, &fragmentation_op);
 685 }
 686
 687 static const struct file_operations fragmentation_file_operations = {
 688         .open           = fragmentation_open,
 689         .read           = seq_read,
 690         .llseek         = seq_lseek,
 691         .release        = seq_release,
 692 };
 693
 694 static const struct seq_operations pagetypeinfo_op = {
 695         .start  = frag_start,
 696         .next   = frag_next,
 697         .stop   = frag_stop,
 698         .show   = pagetypeinfo_show,
 699 };
 700
 701 static int pagetypeinfo_open(struct inode *inode, struct file *file)
 702 {
 703         return seq_open(file, &pagetypeinfo_op);
 704 }
 705
 706 static const struct file_operations pagetypeinfo_file_ops = {
 707         .open           = pagetypeinfo_open,
 708         .read           = seq_read,
 709         .llseek         = seq_lseek,
 710         .release        = seq_release,
 711 };
 712
 713 #ifdef CONFIG_ZONE_DMA
 714 #define TEXT_FOR_DMA(xx) xx "_dma",
 715 #else
 716 #define TEXT_FOR_DMA(xx)
 717 #endif
 718
 719 #ifdef CONFIG_ZONE_DMA32
 720 #define TEXT_FOR_DMA32(xx) xx "_dma32",
 721 #else
 722 #define TEXT_FOR_DMA32(xx)
 723 #endif
 724
 725 #ifdef CONFIG_HIGHMEM
 726 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
 727 #else
 728 #define TEXT_FOR_HIGHMEM(xx)
 729 #endif
 730
 731 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
 732                                         TEXT_FOR_HIGHMEM(xx) xx "_movable",
 733
 734 static const char * const vmstat_text[] = {
 735         /* Zoned VM counters */
 736         "nr_free_pages",
 737         "nr_inactive_anon",
 738         "nr_active_anon",
 739         "nr_inactive_file",
 740         "nr_active_file",
 741         "nr_unevictable",
 742         "nr_mlock",
 743         "nr_anon_pages",
 744         "nr_mapped",
 745         "nr_file_pages",
 746         "nr_dirty",
 747         "nr_writeback",
 748         "nr_slab_reclaimable",
 749         "nr_slab_unreclaimable",
 750         "nr_page_table_pages",
 751         "nr_kernel_stack",
 752         "nr_unstable",
 753         "nr_bounce",
 754         "nr_vmscan_write",
 755         "nr_writeback_temp",
 756         "nr_isolated_anon",
 757         "nr_isolated_file",
 758         "nr_shmem",
 759         "nr_dirtied",
 760         "nr_written",
 761
 762 #ifdef CONFIG_NUMA
 763         "numa_hit",
 764         "numa_miss",
 765         "numa_foreign",
 766         "numa_interleave",
 767         "numa_local",
 768         "numa_other",
 769 #endif
 770         "nr_dirty_threshold",
 771         "nr_dirty_background_threshold",
 772
 773 #ifdef CONFIG_VM_EVENT_COUNTERS
 774         "pgpgin",
 775         "pgpgout",
 776         "pswpin",
 777         "pswpout",
 778
 779         TEXTS_FOR_ZONES("pgalloc")
 780
 781         "pgfree",
 782         "pgactivate",
 783         "pgdeactivate",
 784
 785         "pgfault",
 786         "pgmajfault",
 787
 788         TEXTS_FOR_ZONES("pgrefill")
 789         TEXTS_FOR_ZONES("pgsteal")
 790         TEXTS_FOR_ZONES("pgscan_kswapd")
 791         TEXTS_FOR_ZONES("pgscan_direct")
 792
 793 #ifdef CONFIG_NUMA
 794         "zone_reclaim_failed",
 795 #endif
 796         "pginodesteal",
 797         "slabs_scanned",
 798         "kswapd_steal",
 799         "kswapd_inodesteal",
 800         "kswapd_low_wmark_hit_quickly",
 801         "kswapd_high_wmark_hit_quickly",
 802         "kswapd_skip_congestion_wait",
 803         "pageoutrun",
 804         "allocstall",
 805
 806         "pgrotated",
 807
 808 #ifdef CONFIG_COMPACTION
 809         "compact_blocks_moved",
 810         "compact_pages_moved",
 811         "compact_pagemigrate_failed",
 812         "compact_stall",
 813         "compact_fail",
 814         "compact_success",
 815 #endif
 816
 817 #ifdef CONFIG_HUGETLB_PAGE
 818         "htlb_buddy_alloc_success",
 819         "htlb_buddy_alloc_fail",
 820 #endif
 821         "unevictable_pgs_culled",
 822         "unevictable_pgs_scanned",
 823         "unevictable_pgs_rescued",
 824         "unevictable_pgs_mlocked",
 825         "unevictable_pgs_munlocked",
 826         "unevictable_pgs_cleared",
 827         "unevictable_pgs_stranded",
 828         "unevictable_pgs_mlockfreed",
 829 #endif
 830 };
 831
 832 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
 833                                                         struct zone *zone)
 834 {
 835         int i;
 836         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
 837         seq_printf(m,
 838                    "\n  pages free     %lu"
 839                    "\n        min      %lu"
 840                    "\n        low      %lu"
 841                    "\n        high     %lu"
 842                    "\n        scanned  %lu"
 843                    "\n        spanned  %lu"
 844                    "\n        present  %lu",
 845                    zone_nr_free_pages(zone),
 846                    min_wmark_pages(zone),
 847                    low_wmark_pages(zone),
 848                    high_wmark_pages(zone),
 849                    zone->pages_scanned,
 850                    zone->spanned_pages,
 851                    zone->present_pages);
 852
 853         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 854                 seq_printf(m, "\n    %-12s %lu", vmstat_text[i],
 855                                 zone_page_state(zone, i));
 856
 857         seq_printf(m,
 858                    "\n        protection: (%lu",
 859                    zone->lowmem_reserve[0]);
 860         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
 861                 seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
 862         seq_printf(m,
 863                    ")"
 864                    "\n  pagesets");
 865         for_each_online_cpu(i) {
 866                 struct per_cpu_pageset *pageset;
 867
 868                 pageset = per_cpu_ptr(zone->pageset, i);
 869                 seq_printf(m,
 870                            "\n    cpu: %i"
 871                            "\n              count: %i"
 872                            "\n              high:  %i"
 873                            "\n              batch: %i",
 874                            i,
 875                            pageset->pcp.count,
 876                            pageset->pcp.high,
 877                            pageset->pcp.batch);
 878 #ifdef CONFIG_SMP
 879                 seq_printf(m, "\n  vm stats threshold: %d",
 880                                 pageset->stat_threshold);
 881 #endif
 882         }
 883         seq_printf(m,
 884                    "\n  all_unreclaimable: %u"
 885                    "\n  start_pfn:         %lu"
 886                    "\n  inactive_ratio:    %u",
 887                    zone->all_unreclaimable,
 888                    zone->zone_start_pfn,
 889                    zone->inactive_ratio);
 890         seq_putc(m, '\n');
 891 }
 892
 893 /*
 894  * Output information about zones in @pgdat.
 895  */
 896 static int zoneinfo_show(struct seq_file *m, void *arg)
 897 {
 898         pg_data_t *pgdat = (pg_data_t *)arg;
 899         walk_zones_in_node(m, pgdat, zoneinfo_show_print);
 900         return 0;
 901 }
 902
 903 static const struct seq_operations zoneinfo_op = {
 904         .start  = frag_start, /* iterate over all zones. The same as in
 905                                * fragmentation. */
 906         .next   = frag_next,
 907         .stop   = frag_stop,
 908         .show   = zoneinfo_show,
 909 };
 910
 911 static int zoneinfo_open(struct inode *inode, struct file *file)
 912 {
 913         return seq_open(file, &zoneinfo_op);
 914 }
 915
 916 static const struct file_operations proc_zoneinfo_file_operations = {
 917         .open           = zoneinfo_open,
 918         .read           = seq_read,
 919         .llseek         = seq_lseek,
 920         .release        = seq_release,
 921 };
 922
 923 enum writeback_stat_item {
 924         NR_DIRTY_THRESHOLD,
 925         NR_DIRTY_BG_THRESHOLD,
 926         NR_VM_WRITEBACK_STAT_ITEMS,
 927 };
 928
 929 static void *vmstat_start(struct seq_file *m, loff_t *pos)
 930 {
 931         unsigned long *v;
 932         int i, stat_items_size;
 933
 934         if (*pos >= ARRAY_SIZE(vmstat_text))
 935                 return NULL;
 936         stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
 937                           NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
 938
 939 #ifdef CONFIG_VM_EVENT_COUNTERS
 940         stat_items_size += sizeof(struct vm_event_state);
 941 #endif
 942
 943         v = kmalloc(stat_items_size, GFP_KERNEL);
 944         m->private = v;
 945         if (!v)
 946                 return ERR_PTR(-ENOMEM);
 947         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 948                 v[i] = global_page_state(i);
 949         v += NR_VM_ZONE_STAT_ITEMS;
 950
 951         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
 952                             v + NR_DIRTY_THRESHOLD);
 953         v += NR_VM_WRITEBACK_STAT_ITEMS;
 954
 955 #ifdef CONFIG_VM_EVENT_COUNTERS
 956         all_vm_events(v);
 957         v[PGPGIN] /= 2;         /* sectors -> kbytes */
 958         v[PGPGOUT] /= 2;
 959 #endif
 960         return (unsigned long *)m->private + *pos;
 961 }
 962
 963 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
 964 {
 965         (*pos)++;
 966         if (*pos >= ARRAY_SIZE(vmstat_text))
 967                 return NULL;
 968         return (unsigned long *)m->private + *pos;
 969 }
 970
 971 static int vmstat_show(struct seq_file *m, void *arg)
 972 {
 973         unsigned long *l = arg;
 974         unsigned long off = l - (unsigned long *)m->private;
 975
 976         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
 977         return 0;
 978 }
 979
 980 static void vmstat_stop(struct seq_file *m, void *arg)
 981 {
 982         kfree(m->private);
 983         m->private = NULL;
 984 }
 985
 986 static const struct seq_operations vmstat_op = {
 987         .start  = vmstat_start,
 988         .next   = vmstat_next,
 989         .stop   = vmstat_stop,
 990         .show   = vmstat_show,
 991 };
 992
 993 static int vmstat_open(struct inode *inode, struct file *file)
 994 {
 995         return seq_open(file, &vmstat_op);
 996 }
 997
 998 static const struct file_operations proc_vmstat_file_operations = {
 999         .open           = vmstat_open,
1000         .read           = seq_read,
1001         .llseek         = seq_lseek,
1002         .release        = seq_release,
1003 };
1004 #endif /* CONFIG_PROC_FS */
1005
1006 #ifdef CONFIG_SMP
1007 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1008 int sysctl_stat_interval __read_mostly = HZ;
1009
1010 static void vmstat_update(struct work_struct *w)
1011 {
1012         refresh_cpu_vm_stats(smp_processor_id());
1013         schedule_delayed_work(&__get_cpu_var(vmstat_work),
1014                 round_jiffies_relative(sysctl_stat_interval));
1015 }
1016
1017 static void __cpuinit start_cpu_timer(int cpu)
1018 {
1019         struct delayed_work *work = &per_cpu(vmstat_work, cpu);
1020
1021         INIT_DELAYED_WORK_DEFERRABLE(work, vmstat_update);
1022         schedule_delayed_work_on(cpu, work, __round_jiffies_relative(HZ, cpu));
1023 }
1024
1025 /*
1026  * Use the cpu notifier to insure that the thresholds are recalculated
1027  * when necessary.
1028  */
1029 static int __cpuinit vmstat_cpuup_callback(struct notifier_block *nfb,
1030                 unsigned long action,
1031                 void *hcpu)
1032 {
1033         long cpu = (long)hcpu;
1034
1035         switch (action) {
1036         case CPU_ONLINE:
1037         case CPU_ONLINE_FROZEN:
1038                 refresh_zone_stat_thresholds();
1039                 start_cpu_timer(cpu);
1040                 node_set_state(cpu_to_node(cpu), N_CPU);
1041                 break;
1042         case CPU_DOWN_PREPARE:
1043         case CPU_DOWN_PREPARE_FROZEN:
1044                 cancel_rearming_delayed_work(&per_cpu(vmstat_work, cpu));
1045                 per_cpu(vmstat_work, cpu).work.func = NULL;
1046                 break;
1047         case CPU_DOWN_FAILED:
1048         case CPU_DOWN_FAILED_FROZEN:
1049                 start_cpu_timer(cpu);
1050                 break;
1051         case CPU_DEAD:
1052         case CPU_DEAD_FROZEN:
1053                 refresh_zone_stat_thresholds();
1054                 break;
1055         default:
1056                 break;
1057         }
1058         return NOTIFY_OK;
1059 }
1060
1061 static struct notifier_block __cpuinitdata vmstat_notifier =
1062         { &vmstat_cpuup_callback, NULL, 0 };
1063 #endif
1064
1065 static int __init setup_vmstat(void)
1066 {
1067 #ifdef CONFIG_SMP
1068         int cpu;
1069
1070         refresh_zone_stat_thresholds();
1071         register_cpu_notifier(&vmstat_notifier);
1072
1073         for_each_online_cpu(cpu)
1074                 start_cpu_timer(cpu);
1075 #endif
1076 #ifdef CONFIG_PROC_FS
1077         proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1078         proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1079         proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1080         proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1081 #endif
1082         return 0;
1083 }
1084 module_init(setup_vmstat)
1085
1086 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1087 #include <linux/debugfs.h>
1088
1089 static struct dentry *extfrag_debug_root;
1090
1091 /*
1092  * Return an index indicating how much of the available free memory is
1093  * unusable for an allocation of the requested size.
1094  */
1095 static int unusable_free_index(unsigned int order,
1096                                 struct contig_page_info *info)
1097 {
1098         /* No free memory is interpreted as all free memory is unusable */
1099         if (info->free_pages == 0)
1100                 return 1000;
1101
1102         /*
1103          * Index should be a value between 0 and 1. Return a value to 3
1104          * decimal places.
1105          *
1106          * 0 => no fragmentation
1107          * 1 => high fragmentation
1108          */
1109         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1110
1111 }
1112
1113 static void unusable_show_print(struct seq_file *m,
1114                                         pg_data_t *pgdat, struct zone *zone)
1115 {
1116         unsigned int order;
1117         int index;
1118         struct contig_page_info info;
1119
1120         seq_printf(m, "Node %d, zone %8s ",
1121                                 pgdat->node_id,
1122                                 zone->name);
1123         for (order = 0; order < MAX_ORDER; ++order) {
1124                 fill_contig_page_info(zone, order, &info);
1125                 index = unusable_free_index(order, &info);
1126                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1127         }
1128
1129         seq_putc(m, '\n');
1130 }
1131
1132 /*
1133  * Display unusable free space index
1134  *
1135  * The unusable free space index measures how much of the available free
1136  * memory cannot be used to satisfy an allocation of a given size and is a
1137  * value between 0 and 1. The higher the value, the more of free memory is
1138  * unusable and by implication, the worse the external fragmentation is. This
1139  * can be expressed as a percentage by multiplying by 100.
1140  */
1141 static int unusable_show(struct seq_file *m, void *arg)
1142 {
1143         pg_data_t *pgdat = (pg_data_t *)arg;
1144
1145         /* check memoryless node */
1146         if (!node_state(pgdat->node_id, N_HIGH_MEMORY))
1147                 return 0;
1148
1149         walk_zones_in_node(m, pgdat, unusable_show_print);
1150
1151         return 0;
1152 }
1153
1154 static const struct seq_operations unusable_op = {
1155         .start  = frag_start,
1156         .next   = frag_next,
1157         .stop   = frag_stop,
1158         .show   = unusable_show,
1159 };
1160
1161 static int unusable_open(struct inode *inode, struct file *file)
1162 {
1163         return seq_open(file, &unusable_op);
1164 }
1165
1166 static const struct file_operations unusable_file_ops = {
1167         .open           = unusable_open,
1168         .read           = seq_read,
1169         .llseek         = seq_lseek,
1170         .release        = seq_release,
1171 };
1172
1173 static void extfrag_show_print(struct seq_file *m,
1174                                         pg_data_t *pgdat, struct zone *zone)
1175 {
1176         unsigned int order;
1177         int index;
1178
1179         /* Alloc on stack as interrupts are disabled for zone walk */
1180         struct contig_page_info info;
1181
1182         seq_printf(m, "Node %d, zone %8s ",
1183                                 pgdat->node_id,
1184                                 zone->name);
1185         for (order = 0; order < MAX_ORDER; ++order) {
1186                 fill_contig_page_info(zone, order, &info);
1187                 index = __fragmentation_index(order, &info);
1188                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1189         }
1190
1191         seq_putc(m, '\n');
1192 }
1193
1194 /*
1195  * Display fragmentation index for orders that allocations would fail for
1196  */
1197 static int extfrag_show(struct seq_file *m, void *arg)
1198 {
1199         pg_data_t *pgdat = (pg_data_t *)arg;
1200
1201         walk_zones_in_node(m, pgdat, extfrag_show_print);
1202
1203         return 0;
1204 }
1205
1206 static const struct seq_operations extfrag_op = {
1207         .start  = frag_start,
1208         .next   = frag_next,
1209         .stop   = frag_stop,
1210         .show   = extfrag_show,
1211 };
1212
1213 static int extfrag_open(struct inode *inode, struct file *file)
1214 {
1215         return seq_open(file, &extfrag_op);
1216 }
1217
1218 static const struct file_operations extfrag_file_ops = {
1219         .open           = extfrag_open,
1220         .read           = seq_read,
1221         .llseek         = seq_lseek,
1222         .release        = seq_release,
1223 };
1224
1225 static int __init extfrag_debug_init(void)
1226 {
1227         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1228         if (!extfrag_debug_root)
1229                 return -ENOMEM;
1230
1231         if (!debugfs_create_file("unusable_index", 0444,
1232                         extfrag_debug_root, NULL, &unusable_file_ops))
1233                 return -ENOMEM;
1234
1235         if (!debugfs_create_file("extfrag_index", 0444,
1236                         extfrag_debug_root, NULL, &extfrag_file_ops))
1237                 return -ENOMEM;
1238
1239         return 0;
1240 }
1241
1242 module_init(extfrag_debug_init);
1243 #endif