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
  22 #ifdef CONFIG_VM_EVENT_COUNTERS
  23 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
  24 EXPORT_PER_CPU_SYMBOL(vm_event_states);
  25
  26 static void sum_vm_events(unsigned long *ret)
  27 {
  28         int cpu;
  29         int i;
  30
  31         memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
  32
  33         for_each_online_cpu(cpu) {
  34                 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
  35
  36                 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
  37                         ret[i] += this->event[i];
  38         }
  39 }
  40
  41 /*
  42  * Accumulate the vm event counters across all CPUs.
  43  * The result is unavoidably approximate - it can change
  44  * during and after execution of this function.
  45 */
  46 void all_vm_events(unsigned long *ret)
  47 {
  48         get_online_cpus();
  49         sum_vm_events(ret);
  50         put_online_cpus();
  51 }
  52 EXPORT_SYMBOL_GPL(all_vm_events);
  53
  54 #ifdef CONFIG_HOTPLUG
  55 /*
  56  * Fold the foreign cpu events into our own.
  57  *
  58  * This is adding to the events on one processor
  59  * but keeps the global counts constant.
  60  */
  61 void vm_events_fold_cpu(int cpu)
  62 {
  63         struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
  64         int i;
  65
  66         for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
  67                 count_vm_events(i, fold_state->event[i]);
  68                 fold_state->event[i] = 0;
  69         }
  70 }
  71 #endif /* CONFIG_HOTPLUG */
  72
  73 #endif /* CONFIG_VM_EVENT_COUNTERS */
  74
  75 /*
  76  * Manage combined zone based / global counters
  77  *
  78  * vm_stat contains the global counters
  79  */
  80 atomic_long_t vm_stat[NR_VM_ZONE_STAT_ITEMS];
  81 EXPORT_SYMBOL(vm_stat);
  82
  83 #ifdef CONFIG_SMP
  84
  85 static int calculate_threshold(struct zone *zone)
  86 {
  87         int threshold;
  88         int mem;        /* memory in 128 MB units */
  89
  90         /*
  91          * The threshold scales with the number of processors and the amount
  92          * of memory per zone. More memory means that we can defer updates for
  93          * longer, more processors could lead to more contention.
  94          * fls() is used to have a cheap way of logarithmic scaling.
  95          *
  96          * Some sample thresholds:
  97          *
  98          * Threshold    Processors      (fls)   Zonesize        fls(mem+1)
  99          * ------------------------------------------------------------------
 100          * 8            1               1       0.9-1 GB        4
 101          * 16           2               2       0.9-1 GB        4
 102          * 20           2               2       1-2 GB          5
 103          * 24           2               2       2-4 GB          6
 104          * 28           2               2       4-8 GB          7
 105          * 32           2               2       8-16 GB         8
 106          * 4            2               2       <128M           1
 107          * 30           4               3       2-4 GB          5
 108          * 48           4               3       8-16 GB         8
 109          * 32           8               4       1-2 GB          4
 110          * 32           8               4       0.9-1GB         4
 111          * 10           16              5       <128M           1
 112          * 40           16              5       900M            4
 113          * 70           64              7       2-4 GB          5
 114          * 84           64              7       4-8 GB          6
 115          * 108          512             9       4-8 GB          6
 116          * 125          1024            10      8-16 GB         8
 117          * 125          1024            10      16-32 GB        9
 118          */
 119
 120         mem = zone->present_pages >> (27 - PAGE_SHIFT);
 121
 122         threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
 123
 124         /*
 125          * Maximum threshold is 125
 126          */
 127         threshold = min(125, threshold);
 128
 129         return threshold;
 130 }
 131
 132 /*
 133  * Refresh the thresholds for each zone.
 134  */
 135 static void refresh_zone_stat_thresholds(void)
 136 {
 137         struct zone *zone;
 138         int cpu;
 139         int threshold;
 140
 141         for_each_populated_zone(zone) {
 142                 unsigned long max_drift, tolerate_drift;
 143
 144                 threshold = calculate_threshold(zone);
 145
 146                 for_each_online_cpu(cpu)
 147                         per_cpu_ptr(zone->pageset, cpu)->stat_threshold
 148                                                         = threshold;
 149
 150                 /*
 151                  * Only set percpu_drift_mark if there is a danger that
 152                  * NR_FREE_PAGES reports the low watermark is ok when in fact
 153                  * the min watermark could be breached by an allocation
 154                  */
 155                 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
 156                 max_drift = num_online_cpus() * threshold;
 157                 if (max_drift > tolerate_drift)
 158                         zone->percpu_drift_mark = high_wmark_pages(zone) +
 159                                         max_drift;
 160         }
 161 }
 162
 163 /*
 164  * For use when we know that interrupts are disabled.
 165  */
 166 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 167                                 int delta)
 168 {
 169         struct per_cpu_pageset *pcp = this_cpu_ptr(zone->pageset);
 170
 171         s8 *p = pcp->vm_stat_diff + item;
 172         long x;
 173
 174         x = delta + *p;
 175
 176         if (unlikely(x > pcp->stat_threshold || x < -pcp->stat_threshold)) {
 177                 zone_page_state_add(x, zone, item);
 178                 x = 0;
 179         }
 180         *p = x;
 181 }
 182 EXPORT_SYMBOL(__mod_zone_page_state);
 183
 184 /*
 185  * For an unknown interrupt state
 186  */
 187 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 188                                         int delta)
 189 {
 190         unsigned long flags;
 191
 192         local_irq_save(flags);
 193         __mod_zone_page_state(zone, item, delta);
 194         local_irq_restore(flags);
 195 }
 196 EXPORT_SYMBOL(mod_zone_page_state);
 197
 198 /*
 199  * Optimized increment and decrement functions.
 200  *
 201  * These are only for a single page and therefore can take a struct page *
 202  * argument instead of struct zone *. This allows the inclusion of the code
 203  * generated for page_zone(page) into the optimized functions.
 204  *
 205  * No overflow check is necessary and therefore the differential can be
 206  * incremented or decremented in place which may allow the compilers to
 207  * generate better code.
 208  * The increment or decrement is known and therefore one boundary check can
 209  * be omitted.
 210  *
 211  * NOTE: These functions are very performance sensitive. Change only
 212  * with care.
 213  *
 214  * Some processors have inc/dec instructions that are atomic vs an interrupt.
 215  * However, the code must first determine the differential location in a zone
 216  * based on the processor number and then inc/dec the counter. There is no
 217  * guarantee without disabling preemption that the processor will not change
 218  * in between and therefore the atomicity vs. interrupt cannot be exploited
 219  * in a useful way here.
 220  */
 221 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
 222 {
 223         struct per_cpu_pageset *pcp = this_cpu_ptr(zone->pageset);
 224         s8 *p = pcp->vm_stat_diff + item;
 225
 226         (*p)++;
 227
 228         if (unlikely(*p > pcp->stat_threshold)) {
 229                 int overstep = pcp->stat_threshold / 2;
 230
 231                 zone_page_state_add(*p + overstep, zone, item);
 232                 *p = -overstep;
 233         }
 234 }
 235
 236 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
 237 {
 238         __inc_zone_state(page_zone(page), item);
 239 }
 240 EXPORT_SYMBOL(__inc_zone_page_state);
 241
 242 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
 243 {
 244         struct per_cpu_pageset *pcp = this_cpu_ptr(zone->pageset);
 245         s8 *p = pcp->vm_stat_diff + item;
 246
 247         (*p)--;
 248
 249         if (unlikely(*p < - pcp->stat_threshold)) {
 250                 int overstep = pcp->stat_threshold / 2;
 251
 252                 zone_page_state_add(*p - overstep, zone, item);
 253                 *p = overstep;
 254         }
 255 }
 256
 257 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
 258 {
 259         __dec_zone_state(page_zone(page), item);
 260 }
 261 EXPORT_SYMBOL(__dec_zone_page_state);
 262
 263 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
 264 {
 265         unsigned long flags;
 266
 267         local_irq_save(flags);
 268         __inc_zone_state(zone, item);
 269         local_irq_restore(flags);
 270 }
 271
 272 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 273 {
 274         unsigned long flags;
 275         struct zone *zone;
 276
 277         zone = page_zone(page);
 278         local_irq_save(flags);
 279         __inc_zone_state(zone, item);
 280         local_irq_restore(flags);
 281 }
 282 EXPORT_SYMBOL(inc_zone_page_state);
 283
 284 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 285 {
 286         unsigned long flags;
 287
 288         local_irq_save(flags);
 289         __dec_zone_page_state(page, item);
 290         local_irq_restore(flags);
 291 }
 292 EXPORT_SYMBOL(dec_zone_page_state);
 293
 294 /*
 295  * Update the zone counters for one cpu.
 296  *
 297  * The cpu specified must be either the current cpu or a processor that
 298  * is not online. If it is the current cpu then the execution thread must
 299  * be pinned to the current cpu.
 300  *
 301  * Note that refresh_cpu_vm_stats strives to only access
 302  * node local memory. The per cpu pagesets on remote zones are placed
 303  * in the memory local to the processor using that pageset. So the
 304  * loop over all zones will access a series of cachelines local to
 305  * the processor.
 306  *
 307  * The call to zone_page_state_add updates the cachelines with the
 308  * statistics in the remote zone struct as well as the global cachelines
 309  * with the global counters. These could cause remote node cache line
 310  * bouncing and will have to be only done when necessary.
 311  */
 312 void refresh_cpu_vm_stats(int cpu)
 313 {
 314         struct zone *zone;
 315         int i;
 316         int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 317
 318         for_each_populated_zone(zone) {
 319                 struct per_cpu_pageset *p;
 320
 321                 p = per_cpu_ptr(zone->pageset, cpu);
 322
 323                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 324                         if (p->vm_stat_diff[i]) {
 325                                 unsigned long flags;
 326                                 int v;
 327
 328                                 local_irq_save(flags);
 329                                 v = p->vm_stat_diff[i];
 330                                 p->vm_stat_diff[i] = 0;
 331                                 local_irq_restore(flags);
 332                                 atomic_long_add(v, &zone->vm_stat[i]);
 333                                 global_diff[i] += v;
 334 #ifdef CONFIG_NUMA
 335                                 /* 3 seconds idle till flush */
 336                                 p->expire = 3;
 337 #endif
 338                         }
 339                 cond_resched();
 340 #ifdef CONFIG_NUMA
 341                 /*
 342                  * Deal with draining the remote pageset of this
 343                  * processor
 344                  *
 345                  * Check if there are pages remaining in this pageset
 346                  * if not then there is nothing to expire.
 347                  */
 348                 if (!p->expire || !p->pcp.count)
 349                         continue;
 350
 351                 /*
 352                  * We never drain zones local to this processor.
 353                  */
 354                 if (zone_to_nid(zone) == numa_node_id()) {
 355                         p->expire = 0;
 356                         continue;
 357                 }
 358
 359                 p->expire--;
 360                 if (p->expire)
 361                         continue;
 362
 363                 if (p->pcp.count)
 364                         drain_zone_pages(zone, &p->pcp);
 365 #endif
 366         }
 367
 368         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 369                 if (global_diff[i])
 370                         atomic_long_add(global_diff[i], &vm_stat[i]);
 371 }
 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 zone *preferred_zone, struct zone *z)
 383 {
 384         if (z->zone_pgdat == preferred_zone->zone_pgdat) {
 385                 __inc_zone_state(z, NUMA_HIT);
 386         } else {
 387                 __inc_zone_state(z, NUMA_MISS);
 388                 __inc_zone_state(preferred_zone, 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_COMPACTION
 398 struct contig_page_info {
 399         unsigned long free_pages;
 400         unsigned long free_blocks_total;
 401         unsigned long free_blocks_suitable;
 402 };
 403
 404 /*
 405  * Calculate the number of free pages in a zone, how many contiguous
 406  * pages are free and how many are large enough to satisfy an allocation of
 407  * the target size. Note that this function makes no attempt to estimate
 408  * how many suitable free blocks there *might* be if MOVABLE pages were
 409  * migrated. Calculating that is possible, but expensive and can be
 410  * figured out from userspace
 411  */
 412 static void fill_contig_page_info(struct zone *zone,
 413                                 unsigned int suitable_order,
 414                                 struct contig_page_info *info)
 415 {
 416         unsigned int order;
 417
 418         info->free_pages = 0;
 419         info->free_blocks_total = 0;
 420         info->free_blocks_suitable = 0;
 421
 422         for (order = 0; order < MAX_ORDER; order++) {
 423                 unsigned long blocks;
 424
 425                 /* Count number of free blocks */
 426                 blocks = zone->free_area[order].nr_free;
 427                 info->free_blocks_total += blocks;
 428
 429                 /* Count free base pages */
 430                 info->free_pages += blocks << order;
 431
 432                 /* Count the suitable free blocks */
 433                 if (order >= suitable_order)
 434                         info->free_blocks_suitable += blocks <<
 435                                                 (order - suitable_order);
 436         }
 437 }
 438
 439 /*
 440  * A fragmentation index only makes sense if an allocation of a requested
 441  * size would fail. If that is true, the fragmentation index indicates
 442  * whether external fragmentation or a lack of memory was the problem.
 443  * The value can be used to determine if page reclaim or compaction
 444  * should be used
 445  */
 446 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
 447 {
 448         unsigned long requested = 1UL << order;
 449
 450         if (!info->free_blocks_total)
 451                 return 0;
 452
 453         /* Fragmentation index only makes sense when a request would fail */
 454         if (info->free_blocks_suitable)
 455                 return -1000;
 456
 457         /*
 458          * Index is between 0 and 1 so return within 3 decimal places
 459          *
 460          * 0 => allocation would fail due to lack of memory
 461          * 1 => allocation would fail due to fragmentation
 462          */
 463         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
 464 }
 465
 466 /* Same as __fragmentation index but allocs contig_page_info on stack */
 467 int fragmentation_index(struct zone *zone, unsigned int order)
 468 {
 469         struct contig_page_info info;
 470
 471         fill_contig_page_info(zone, order, &info);
 472         return __fragmentation_index(order, &info);
 473 }
 474 #endif
 475
 476 #if defined(CONFIG_PROC_FS) || defined(CONFIG_COMPACTION)
 477 #include <linux/proc_fs.h>
 478 #include <linux/seq_file.h>
 479
 480 static char * const migratetype_names[MIGRATE_TYPES] = {
 481         "Unmovable",
 482         "Reclaimable",
 483         "Movable",
 484         "Reserve",
 485         "Isolate",
 486 };
 487
 488 static void *frag_start(struct seq_file *m, loff_t *pos)
 489 {
 490         pg_data_t *pgdat;
 491         loff_t node = *pos;
 492         for (pgdat = first_online_pgdat();
 493              pgdat && node;
 494              pgdat = next_online_pgdat(pgdat))
 495                 --node;
 496
 497         return pgdat;
 498 }
 499
 500 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
 501 {
 502         pg_data_t *pgdat = (pg_data_t *)arg;
 503
 504         (*pos)++;
 505         return next_online_pgdat(pgdat);
 506 }
 507
 508 static void frag_stop(struct seq_file *m, void *arg)
 509 {
 510 }
 511
 512 /* Walk all the zones in a node and print using a callback */
 513 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
 514                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
 515 {
 516         struct zone *zone;
 517         struct zone *node_zones = pgdat->node_zones;
 518         unsigned long flags;
 519
 520         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 521                 if (!populated_zone(zone))
 522                         continue;
 523
 524                 spin_lock_irqsave(&zone->lock, flags);
 525                 print(m, pgdat, zone);
 526                 spin_unlock_irqrestore(&zone->lock, flags);
 527         }
 528 }
 529 #endif
 530
 531 #ifdef CONFIG_PROC_FS
 532 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
 533                                                 struct zone *zone)
 534 {
 535         int order;
 536
 537         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 538         for (order = 0; order < MAX_ORDER; ++order)
 539                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
 540         seq_putc(m, '\n');
 541 }
 542
 543 /*
 544  * This walks the free areas for each zone.
 545  */
 546 static int frag_show(struct seq_file *m, void *arg)
 547 {
 548         pg_data_t *pgdat = (pg_data_t *)arg;
 549         walk_zones_in_node(m, pgdat, frag_show_print);
 550         return 0;
 551 }
 552
 553 static void pagetypeinfo_showfree_print(struct seq_file *m,
 554                                         pg_data_t *pgdat, struct zone *zone)
 555 {
 556         int order, mtype;
 557
 558         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
 559                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
 560                                         pgdat->node_id,
 561                                         zone->name,
 562                                         migratetype_names[mtype]);
 563                 for (order = 0; order < MAX_ORDER; ++order) {
 564                         unsigned long freecount = 0;
 565                         struct free_area *area;
 566                         struct list_head *curr;
 567
 568                         area = &(zone->free_area[order]);
 569
 570                         list_for_each(curr, &area->free_list[mtype])
 571                                 freecount++;
 572                         seq_printf(m, "%6lu ", freecount);
 573                 }
 574                 seq_putc(m, '\n');
 575         }
 576 }
 577
 578 /* Print out the free pages at each order for each migatetype */
 579 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
 580 {
 581         int order;
 582         pg_data_t *pgdat = (pg_data_t *)arg;
 583
 584         /* Print header */
 585         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
 586         for (order = 0; order < MAX_ORDER; ++order)
 587                 seq_printf(m, "%6d ", order);
 588         seq_putc(m, '\n');
 589
 590         walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
 591
 592         return 0;
 593 }
 594
 595 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
 596                                         pg_data_t *pgdat, struct zone *zone)
 597 {
 598         int mtype;
 599         unsigned long pfn;
 600         unsigned long start_pfn = zone->zone_start_pfn;
 601         unsigned long end_pfn = start_pfn + zone->spanned_pages;
 602         unsigned long count[MIGRATE_TYPES] = { 0, };
 603
 604         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
 605                 struct page *page;
 606
 607                 if (!pfn_valid(pfn))
 608                         continue;
 609
 610                 page = pfn_to_page(pfn);
 611
 612                 /* Watch for unexpected holes punched in the memmap */
 613                 if (!memmap_valid_within(pfn, page, zone))
 614                         continue;
 615
 616                 mtype = get_pageblock_migratetype(page);
 617
 618                 if (mtype < MIGRATE_TYPES)
 619                         count[mtype]++;
 620         }
 621
 622         /* Print counts */
 623         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 624         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
 625                 seq_printf(m, "%12lu ", count[mtype]);
 626         seq_putc(m, '\n');
 627 }
 628
 629 /* Print out the free pages at each order for each migratetype */
 630 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
 631 {
 632         int mtype;
 633         pg_data_t *pgdat = (pg_data_t *)arg;
 634
 635         seq_printf(m, "\n%-23s", "Number of blocks type ");
 636         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
 637                 seq_printf(m, "%12s ", migratetype_names[mtype]);
 638         seq_putc(m, '\n');
 639         walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
 640
 641         return 0;
 642 }
 643
 644 /*
 645  * This prints out statistics in relation to grouping pages by mobility.
 646  * It is expensive to collect so do not constantly read the file.
 647  */
 648 static int pagetypeinfo_show(struct seq_file *m, void *arg)
 649 {
 650         pg_data_t *pgdat = (pg_data_t *)arg;
 651
 652         /* check memoryless node */
 653         if (!node_state(pgdat->node_id, N_HIGH_MEMORY))
 654                 return 0;
 655
 656         seq_printf(m, "Page block order: %d\n", pageblock_order);
 657         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
 658         seq_putc(m, '\n');
 659         pagetypeinfo_showfree(m, pgdat);
 660         pagetypeinfo_showblockcount(m, pgdat);
 661
 662         return 0;
 663 }
 664
 665 static const struct seq_operations fragmentation_op = {
 666         .start  = frag_start,
 667         .next   = frag_next,
 668         .stop   = frag_stop,
 669         .show   = frag_show,
 670 };
 671
 672 static int fragmentation_open(struct inode *inode, struct file *file)
 673 {
 674         return seq_open(file, &fragmentation_op);
 675 }
 676
 677 static const struct file_operations fragmentation_file_operations = {
 678         .open           = fragmentation_open,
 679         .read           = seq_read,
 680         .llseek         = seq_lseek,
 681         .release        = seq_release,
 682 };
 683
 684 static const struct seq_operations pagetypeinfo_op = {
 685         .start  = frag_start,
 686         .next   = frag_next,
 687         .stop   = frag_stop,
 688         .show   = pagetypeinfo_show,
 689 };
 690
 691 static int pagetypeinfo_open(struct inode *inode, struct file *file)
 692 {
 693         return seq_open(file, &pagetypeinfo_op);
 694 }
 695
 696 static const struct file_operations pagetypeinfo_file_ops = {
 697         .open           = pagetypeinfo_open,
 698         .read           = seq_read,
 699         .llseek         = seq_lseek,
 700         .release        = seq_release,
 701 };
 702
 703 #ifdef CONFIG_ZONE_DMA
 704 #define TEXT_FOR_DMA(xx) xx "_dma",
 705 #else
 706 #define TEXT_FOR_DMA(xx)
 707 #endif
 708
 709 #ifdef CONFIG_ZONE_DMA32
 710 #define TEXT_FOR_DMA32(xx) xx "_dma32",
 711 #else
 712 #define TEXT_FOR_DMA32(xx)
 713 #endif
 714
 715 #ifdef CONFIG_HIGHMEM
 716 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
 717 #else
 718 #define TEXT_FOR_HIGHMEM(xx)
 719 #endif
 720
 721 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
 722                                         TEXT_FOR_HIGHMEM(xx) xx "_movable",
 723
 724 static const char * const vmstat_text[] = {
 725         /* Zoned VM counters */
 726         "nr_free_pages",
 727         "nr_inactive_anon",
 728         "nr_active_anon",
 729         "nr_inactive_file",
 730         "nr_active_file",
 731         "nr_unevictable",
 732         "nr_mlock",
 733         "nr_anon_pages",
 734         "nr_mapped",
 735         "nr_file_pages",
 736         "nr_dirty",
 737         "nr_writeback",
 738         "nr_slab_reclaimable",
 739         "nr_slab_unreclaimable",
 740         "nr_page_table_pages",
 741         "nr_kernel_stack",
 742         "nr_unstable",
 743         "nr_bounce",
 744         "nr_vmscan_write",
 745         "nr_writeback_temp",
 746         "nr_isolated_anon",
 747         "nr_isolated_file",
 748         "nr_shmem",
 749         "nr_dirtied",
 750         "nr_written",
 751         "nr_dirty_threshold",
 752         "nr_dirty_background_threshold",
 753
 754 #ifdef CONFIG_NUMA
 755         "numa_hit",
 756         "numa_miss",
 757         "numa_foreign",
 758         "numa_interleave",
 759         "numa_local",
 760         "numa_other",
 761 #endif
 762
 763 #ifdef CONFIG_VM_EVENT_COUNTERS
 764         "pgpgin",
 765         "pgpgout",
 766         "pswpin",
 767         "pswpout",
 768
 769         TEXTS_FOR_ZONES("pgalloc")
 770
 771         "pgfree",
 772         "pgactivate",
 773         "pgdeactivate",
 774
 775         "pgfault",
 776         "pgmajfault",
 777
 778         TEXTS_FOR_ZONES("pgrefill")
 779         TEXTS_FOR_ZONES("pgsteal")
 780         TEXTS_FOR_ZONES("pgscan_kswapd")
 781         TEXTS_FOR_ZONES("pgscan_direct")
 782
 783 #ifdef CONFIG_NUMA
 784         "zone_reclaim_failed",
 785 #endif
 786         "pginodesteal",
 787         "slabs_scanned",
 788         "kswapd_steal",
 789         "kswapd_inodesteal",
 790         "kswapd_low_wmark_hit_quickly",
 791         "kswapd_high_wmark_hit_quickly",
 792         "kswapd_skip_congestion_wait",
 793         "pageoutrun",
 794         "allocstall",
 795
 796         "pgrotated",
 797
 798 #ifdef CONFIG_COMPACTION
 799         "compact_blocks_moved",
 800         "compact_pages_moved",
 801         "compact_pagemigrate_failed",
 802         "compact_stall",
 803         "compact_fail",
 804         "compact_success",
 805 #endif
 806
 807 #ifdef CONFIG_HUGETLB_PAGE
 808         "htlb_buddy_alloc_success",
 809         "htlb_buddy_alloc_fail",
 810 #endif
 811         "unevictable_pgs_culled",
 812         "unevictable_pgs_scanned",
 813         "unevictable_pgs_rescued",
 814         "unevictable_pgs_mlocked",
 815         "unevictable_pgs_munlocked",
 816         "unevictable_pgs_cleared",
 817         "unevictable_pgs_stranded",
 818         "unevictable_pgs_mlockfreed",
 819 #endif
 820 };
 821
 822 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
 823                                                         struct zone *zone)
 824 {
 825         int i;
 826         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
 827         seq_printf(m,
 828                    "\n  pages free     %lu"
 829                    "\n        min      %lu"
 830                    "\n        low      %lu"
 831                    "\n        high     %lu"
 832                    "\n        scanned  %lu"
 833                    "\n        spanned  %lu"
 834                    "\n        present  %lu",
 835                    zone_nr_free_pages(zone),
 836                    min_wmark_pages(zone),
 837                    low_wmark_pages(zone),
 838                    high_wmark_pages(zone),
 839                    zone->pages_scanned,
 840                    zone->spanned_pages,
 841                    zone->present_pages);
 842
 843         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 844                 seq_printf(m, "\n    %-12s %lu", vmstat_text[i],
 845                                 zone_page_state(zone, i));
 846
 847         seq_printf(m,
 848                    "\n        protection: (%lu",
 849                    zone->lowmem_reserve[0]);
 850         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
 851                 seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
 852         seq_printf(m,
 853                    ")"
 854                    "\n  pagesets");
 855         for_each_online_cpu(i) {
 856                 struct per_cpu_pageset *pageset;
 857
 858                 pageset = per_cpu_ptr(zone->pageset, i);
 859                 seq_printf(m,
 860                            "\n    cpu: %i"
 861                            "\n              count: %i"
 862                            "\n              high:  %i"
 863                            "\n              batch: %i",
 864                            i,
 865                            pageset->pcp.count,
 866                            pageset->pcp.high,
 867                            pageset->pcp.batch);
 868 #ifdef CONFIG_SMP
 869                 seq_printf(m, "\n  vm stats threshold: %d",
 870                                 pageset->stat_threshold);
 871 #endif
 872         }
 873         seq_printf(m,
 874                    "\n  all_unreclaimable: %u"
 875                    "\n  start_pfn:         %lu"
 876                    "\n  inactive_ratio:    %u",
 877                    zone->all_unreclaimable,
 878                    zone->zone_start_pfn,
 879                    zone->inactive_ratio);
 880         seq_putc(m, '\n');
 881 }
 882
 883 /*
 884  * Output information about zones in @pgdat.
 885  */
 886 static int zoneinfo_show(struct seq_file *m, void *arg)
 887 {
 888         pg_data_t *pgdat = (pg_data_t *)arg;
 889         walk_zones_in_node(m, pgdat, zoneinfo_show_print);
 890         return 0;
 891 }
 892
 893 static const struct seq_operations zoneinfo_op = {
 894         .start  = frag_start, /* iterate over all zones. The same as in
 895                                * fragmentation. */
 896         .next   = frag_next,
 897         .stop   = frag_stop,
 898         .show   = zoneinfo_show,
 899 };
 900
 901 static int zoneinfo_open(struct inode *inode, struct file *file)
 902 {
 903         return seq_open(file, &zoneinfo_op);
 904 }
 905
 906 static const struct file_operations proc_zoneinfo_file_operations = {
 907         .open           = zoneinfo_open,
 908         .read           = seq_read,
 909         .llseek         = seq_lseek,
 910         .release        = seq_release,
 911 };
 912
 913 enum writeback_stat_item {
 914         NR_DIRTY_THRESHOLD,
 915         NR_DIRTY_BG_THRESHOLD,
 916         NR_VM_WRITEBACK_STAT_ITEMS,
 917 };
 918
 919 static void *vmstat_start(struct seq_file *m, loff_t *pos)
 920 {
 921         unsigned long *v;
 922         int i, stat_items_size;
 923
 924         if (*pos >= ARRAY_SIZE(vmstat_text))
 925                 return NULL;
 926         stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
 927                           NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
 928
 929 #ifdef CONFIG_VM_EVENT_COUNTERS
 930         stat_items_size += sizeof(struct vm_event_state);
 931 #endif
 932
 933         v = kmalloc(stat_items_size, GFP_KERNEL);
 934         m->private = v;
 935         if (!v)
 936                 return ERR_PTR(-ENOMEM);
 937         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 938                 v[i] = global_page_state(i);
 939         v += NR_VM_ZONE_STAT_ITEMS;
 940
 941         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
 942                             v + NR_DIRTY_THRESHOLD);
 943         v += NR_VM_WRITEBACK_STAT_ITEMS;
 944
 945 #ifdef CONFIG_VM_EVENT_COUNTERS
 946         all_vm_events(v);
 947         v[PGPGIN] /= 2;         /* sectors -> kbytes */
 948         v[PGPGOUT] /= 2;
 949 #endif
 950         return m->private + *pos;
 951 }
 952
 953 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
 954 {
 955         (*pos)++;
 956         if (*pos >= ARRAY_SIZE(vmstat_text))
 957                 return NULL;
 958         return (unsigned long *)m->private + *pos;
 959 }
 960
 961 static int vmstat_show(struct seq_file *m, void *arg)
 962 {
 963         unsigned long *l = arg;
 964         unsigned long off = l - (unsigned long *)m->private;
 965
 966         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
 967         return 0;
 968 }
 969
 970 static void vmstat_stop(struct seq_file *m, void *arg)
 971 {
 972         kfree(m->private);
 973         m->private = NULL;
 974 }
 975
 976 static const struct seq_operations vmstat_op = {
 977         .start  = vmstat_start,
 978         .next   = vmstat_next,
 979         .stop   = vmstat_stop,
 980         .show   = vmstat_show,
 981 };
 982
 983 static int vmstat_open(struct inode *inode, struct file *file)
 984 {
 985         return seq_open(file, &vmstat_op);
 986 }
 987
 988 static const struct file_operations proc_vmstat_file_operations = {
 989         .open           = vmstat_open,
 990         .read           = seq_read,
 991         .llseek         = seq_lseek,
 992         .release        = seq_release,
 993 };
 994 #endif /* CONFIG_PROC_FS */
 995
 996 #ifdef CONFIG_SMP
 997 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
 998 int sysctl_stat_interval __read_mostly = HZ;
 999
1000 static void vmstat_update(struct work_struct *w)
1001 {
1002         refresh_cpu_vm_stats(smp_processor_id());
1003         schedule_delayed_work(&__get_cpu_var(vmstat_work),
1004                 round_jiffies_relative(sysctl_stat_interval));
1005 }
1006
1007 static void __cpuinit start_cpu_timer(int cpu)
1008 {
1009         struct delayed_work *work = &per_cpu(vmstat_work, cpu);
1010
1011         INIT_DELAYED_WORK_DEFERRABLE(work, vmstat_update);
1012         schedule_delayed_work_on(cpu, work, __round_jiffies_relative(HZ, cpu));
1013 }
1014
1015 /*
1016  * Use the cpu notifier to insure that the thresholds are recalculated
1017  * when necessary.
1018  */
1019 static int __cpuinit vmstat_cpuup_callback(struct notifier_block *nfb,
1020                 unsigned long action,
1021                 void *hcpu)
1022 {
1023         long cpu = (long)hcpu;
1024
1025         switch (action) {
1026         case CPU_ONLINE:
1027         case CPU_ONLINE_FROZEN:
1028                 refresh_zone_stat_thresholds();
1029                 start_cpu_timer(cpu);
1030                 node_set_state(cpu_to_node(cpu), N_CPU);
1031                 break;
1032         case CPU_DOWN_PREPARE:
1033         case CPU_DOWN_PREPARE_FROZEN:
1034                 cancel_rearming_delayed_work(&per_cpu(vmstat_work, cpu));
1035                 per_cpu(vmstat_work, cpu).work.func = NULL;
1036                 break;
1037         case CPU_DOWN_FAILED:
1038         case CPU_DOWN_FAILED_FROZEN:
1039                 start_cpu_timer(cpu);
1040                 break;
1041         case CPU_DEAD:
1042         case CPU_DEAD_FROZEN:
1043                 refresh_zone_stat_thresholds();
1044                 break;
1045         default:
1046                 break;
1047         }
1048         return NOTIFY_OK;
1049 }
1050
1051 static struct notifier_block __cpuinitdata vmstat_notifier =
1052         { &vmstat_cpuup_callback, NULL, 0 };
1053 #endif
1054
1055 static int __init setup_vmstat(void)
1056 {
1057 #ifdef CONFIG_SMP
1058         int cpu;
1059
1060         refresh_zone_stat_thresholds();
1061         register_cpu_notifier(&vmstat_notifier);
1062
1063         for_each_online_cpu(cpu)
1064                 start_cpu_timer(cpu);
1065 #endif
1066 #ifdef CONFIG_PROC_FS
1067         proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1068         proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1069         proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1070         proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1071 #endif
1072         return 0;
1073 }
1074 module_init(setup_vmstat)
1075
1076 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1077 #include <linux/debugfs.h>
1078
1079 static struct dentry *extfrag_debug_root;
1080
1081 /*
1082  * Return an index indicating how much of the available free memory is
1083  * unusable for an allocation of the requested size.
1084  */
1085 static int unusable_free_index(unsigned int order,
1086                                 struct contig_page_info *info)
1087 {
1088         /* No free memory is interpreted as all free memory is unusable */
1089         if (info->free_pages == 0)
1090                 return 1000;
1091
1092         /*
1093          * Index should be a value between 0 and 1. Return a value to 3
1094          * decimal places.
1095          *
1096          * 0 => no fragmentation
1097          * 1 => high fragmentation
1098          */
1099         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1100
1101 }
1102
1103 static void unusable_show_print(struct seq_file *m,
1104                                         pg_data_t *pgdat, struct zone *zone)
1105 {
1106         unsigned int order;
1107         int index;
1108         struct contig_page_info info;
1109
1110         seq_printf(m, "Node %d, zone %8s ",
1111                                 pgdat->node_id,
1112                                 zone->name);
1113         for (order = 0; order < MAX_ORDER; ++order) {
1114                 fill_contig_page_info(zone, order, &info);
1115                 index = unusable_free_index(order, &info);
1116                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1117         }
1118
1119         seq_putc(m, '\n');
1120 }
1121
1122 /*
1123  * Display unusable free space index
1124  *
1125  * The unusable free space index measures how much of the available free
1126  * memory cannot be used to satisfy an allocation of a given size and is a
1127  * value between 0 and 1. The higher the value, the more of free memory is
1128  * unusable and by implication, the worse the external fragmentation is. This
1129  * can be expressed as a percentage by multiplying by 100.
1130  */
1131 static int unusable_show(struct seq_file *m, void *arg)
1132 {
1133         pg_data_t *pgdat = (pg_data_t *)arg;
1134
1135         /* check memoryless node */
1136         if (!node_state(pgdat->node_id, N_HIGH_MEMORY))
1137                 return 0;
1138
1139         walk_zones_in_node(m, pgdat, unusable_show_print);
1140
1141         return 0;
1142 }
1143
1144 static const struct seq_operations unusable_op = {
1145         .start  = frag_start,
1146         .next   = frag_next,
1147         .stop   = frag_stop,
1148         .show   = unusable_show,
1149 };
1150
1151 static int unusable_open(struct inode *inode, struct file *file)
1152 {
1153         return seq_open(file, &unusable_op);
1154 }
1155
1156 static const struct file_operations unusable_file_ops = {
1157         .open           = unusable_open,
1158         .read           = seq_read,
1159         .llseek         = seq_lseek,
1160         .release        = seq_release,
1161 };
1162
1163 static void extfrag_show_print(struct seq_file *m,
1164                                         pg_data_t *pgdat, struct zone *zone)
1165 {
1166         unsigned int order;
1167         int index;
1168
1169         /* Alloc on stack as interrupts are disabled for zone walk */
1170         struct contig_page_info info;
1171
1172         seq_printf(m, "Node %d, zone %8s ",
1173                                 pgdat->node_id,
1174                                 zone->name);
1175         for (order = 0; order < MAX_ORDER; ++order) {
1176                 fill_contig_page_info(zone, order, &info);
1177                 index = __fragmentation_index(order, &info);
1178                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1179         }
1180
1181         seq_putc(m, '\n');
1182 }
1183
1184 /*
1185  * Display fragmentation index for orders that allocations would fail for
1186  */
1187 static int extfrag_show(struct seq_file *m, void *arg)
1188 {
1189         pg_data_t *pgdat = (pg_data_t *)arg;
1190
1191         walk_zones_in_node(m, pgdat, extfrag_show_print);
1192
1193         return 0;
1194 }
1195
1196 static const struct seq_operations extfrag_op = {
1197         .start  = frag_start,
1198         .next   = frag_next,
1199         .stop   = frag_stop,
1200         .show   = extfrag_show,
1201 };
1202
1203 static int extfrag_open(struct inode *inode, struct file *file)
1204 {
1205         return seq_open(file, &extfrag_op);
1206 }
1207
1208 static const struct file_operations extfrag_file_ops = {
1209         .open           = extfrag_open,
1210         .read           = seq_read,
1211         .llseek         = seq_lseek,
1212         .release        = seq_release,
1213 };
1214
1215 static int __init extfrag_debug_init(void)
1216 {
1217         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1218         if (!extfrag_debug_root)
1219                 return -ENOMEM;
1220
1221         if (!debugfs_create_file("unusable_index", 0444,
1222                         extfrag_debug_root, NULL, &unusable_file_ops))
1223                 return -ENOMEM;
1224
1225         if (!debugfs_create_file("extfrag_index", 0444,
1226                         extfrag_debug_root, NULL, &extfrag_file_ops))
1227                 return -ENOMEM;
1228
1229         return 0;
1230 }
1231
1232 module_init(extfrag_debug_init);
1233 #endif