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
  21 #ifdef CONFIG_VM_EVENT_COUNTERS
  22 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
  23 EXPORT_PER_CPU_SYMBOL(vm_event_states);
  24
  25 static void sum_vm_events(unsigned long *ret, const struct cpumask *cpumask)
  26 {
  27         int cpu;
  28         int i;
  29
  30         memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
  31
  32         for_each_cpu(cpu, cpumask) {
  33                 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
  34
  35                 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
  36                         ret[i] += this->event[i];
  37         }
  38 }
  39
  40 /*
  41  * Accumulate the vm event counters across all CPUs.
  42  * The result is unavoidably approximate - it can change
  43  * during and after execution of this function.
  44 */
  45 void all_vm_events(unsigned long *ret)
  46 {
  47         get_online_cpus();
  48         sum_vm_events(ret, cpu_online_mask);
  49         put_online_cpus();
  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_populated_zone(zone) {
 141                 unsigned long max_drift, tolerate_drift;
 142
 143                 threshold = calculate_threshold(zone);
 144
 145                 for_each_online_cpu(cpu)
 146                         per_cpu_ptr(zone->pageset, cpu)->stat_threshold
 147                                                         = threshold;
 148
 149                 /*
 150                  * Only set percpu_drift_mark if there is a danger that
 151                  * NR_FREE_PAGES reports the low watermark is ok when in fact
 152                  * the min watermark could be breached by an allocation
 153                  */
 154                 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
 155                 max_drift = num_online_cpus() * threshold;
 156                 if (max_drift > tolerate_drift)
 157                         zone->percpu_drift_mark = high_wmark_pages(zone) +
 158                                         max_drift;
 159         }
 160 }
 161
 162 /*
 163  * For use when we know that interrupts are disabled.
 164  */
 165 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 166                                 int delta)
 167 {
 168         struct per_cpu_pageset *pcp = this_cpu_ptr(zone->pageset);
 169
 170         s8 *p = pcp->vm_stat_diff + item;
 171         long x;
 172
 173         x = delta + *p;
 174
 175         if (unlikely(x > pcp->stat_threshold || x < -pcp->stat_threshold)) {
 176                 zone_page_state_add(x, zone, item);
 177                 x = 0;
 178         }
 179         *p = x;
 180 }
 181 EXPORT_SYMBOL(__mod_zone_page_state);
 182
 183 /*
 184  * For an unknown interrupt state
 185  */
 186 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 187                                         int delta)
 188 {
 189         unsigned long flags;
 190
 191         local_irq_save(flags);
 192         __mod_zone_page_state(zone, item, delta);
 193         local_irq_restore(flags);
 194 }
 195 EXPORT_SYMBOL(mod_zone_page_state);
 196
 197 /*
 198  * Optimized increment and decrement functions.
 199  *
 200  * These are only for a single page and therefore can take a struct page *
 201  * argument instead of struct zone *. This allows the inclusion of the code
 202  * generated for page_zone(page) into the optimized functions.
 203  *
 204  * No overflow check is necessary and therefore the differential can be
 205  * incremented or decremented in place which may allow the compilers to
 206  * generate better code.
 207  * The increment or decrement is known and therefore one boundary check can
 208  * be omitted.
 209  *
 210  * NOTE: These functions are very performance sensitive. Change only
 211  * with care.
 212  *
 213  * Some processors have inc/dec instructions that are atomic vs an interrupt.
 214  * However, the code must first determine the differential location in a zone
 215  * based on the processor number and then inc/dec the counter. There is no
 216  * guarantee without disabling preemption that the processor will not change
 217  * in between and therefore the atomicity vs. interrupt cannot be exploited
 218  * in a useful way here.
 219  */
 220 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
 221 {
 222         struct per_cpu_pageset *pcp = this_cpu_ptr(zone->pageset);
 223         s8 *p = pcp->vm_stat_diff + item;
 224
 225         (*p)++;
 226
 227         if (unlikely(*p > pcp->stat_threshold)) {
 228                 int overstep = pcp->stat_threshold / 2;
 229
 230                 zone_page_state_add(*p + overstep, zone, item);
 231                 *p = -overstep;
 232         }
 233 }
 234
 235 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
 236 {
 237         __inc_zone_state(page_zone(page), item);
 238 }
 239 EXPORT_SYMBOL(__inc_zone_page_state);
 240
 241 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
 242 {
 243         struct per_cpu_pageset *pcp = this_cpu_ptr(zone->pageset);
 244         s8 *p = pcp->vm_stat_diff + item;
 245
 246         (*p)--;
 247
 248         if (unlikely(*p < - pcp->stat_threshold)) {
 249                 int overstep = pcp->stat_threshold / 2;
 250
 251                 zone_page_state_add(*p - overstep, zone, item);
 252                 *p = overstep;
 253         }
 254 }
 255
 256 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
 257 {
 258         __dec_zone_state(page_zone(page), item);
 259 }
 260 EXPORT_SYMBOL(__dec_zone_page_state);
 261
 262 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
 263 {
 264         unsigned long flags;
 265
 266         local_irq_save(flags);
 267         __inc_zone_state(zone, item);
 268         local_irq_restore(flags);
 269 }
 270
 271 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 272 {
 273         unsigned long flags;
 274         struct zone *zone;
 275
 276         zone = page_zone(page);
 277         local_irq_save(flags);
 278         __inc_zone_state(zone, item);
 279         local_irq_restore(flags);
 280 }
 281 EXPORT_SYMBOL(inc_zone_page_state);
 282
 283 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 284 {
 285         unsigned long flags;
 286
 287         local_irq_save(flags);
 288         __dec_zone_page_state(page, item);
 289         local_irq_restore(flags);
 290 }
 291 EXPORT_SYMBOL(dec_zone_page_state);
 292
 293 /*
 294  * Update the zone counters for one cpu.
 295  *
 296  * The cpu specified must be either the current cpu or a processor that
 297  * is not online. If it is the current cpu then the execution thread must
 298  * be pinned to the current cpu.
 299  *
 300  * Note that refresh_cpu_vm_stats strives to only access
 301  * node local memory. The per cpu pagesets on remote zones are placed
 302  * in the memory local to the processor using that pageset. So the
 303  * loop over all zones will access a series of cachelines local to
 304  * the processor.
 305  *
 306  * The call to zone_page_state_add updates the cachelines with the
 307  * statistics in the remote zone struct as well as the global cachelines
 308  * with the global counters. These could cause remote node cache line
 309  * bouncing and will have to be only done when necessary.
 310  */
 311 void refresh_cpu_vm_stats(int cpu)
 312 {
 313         struct zone *zone;
 314         int i;
 315         int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 316
 317         for_each_populated_zone(zone) {
 318                 struct per_cpu_pageset *p;
 319
 320                 p = per_cpu_ptr(zone->pageset, cpu);
 321
 322                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 323                         if (p->vm_stat_diff[i]) {
 324                                 unsigned long flags;
 325                                 int v;
 326
 327                                 local_irq_save(flags);
 328                                 v = p->vm_stat_diff[i];
 329                                 p->vm_stat_diff[i] = 0;
 330                                 local_irq_restore(flags);
 331                                 atomic_long_add(v, &zone->vm_stat[i]);
 332                                 global_diff[i] += v;
 333 #ifdef CONFIG_NUMA
 334                                 /* 3 seconds idle till flush */
 335                                 p->expire = 3;
 336 #endif
 337                         }
 338                 cond_resched();
 339 #ifdef CONFIG_NUMA
 340                 /*
 341                  * Deal with draining the remote pageset of this
 342                  * processor
 343                  *
 344                  * Check if there are pages remaining in this pageset
 345                  * if not then there is nothing to expire.
 346                  */
 347                 if (!p->expire || !p->pcp.count)
 348                         continue;
 349
 350                 /*
 351                  * We never drain zones local to this processor.
 352                  */
 353                 if (zone_to_nid(zone) == numa_node_id()) {
 354                         p->expire = 0;
 355                         continue;
 356                 }
 357
 358                 p->expire--;
 359                 if (p->expire)
 360                         continue;
 361
 362                 if (p->pcp.count)
 363                         drain_zone_pages(zone, &p->pcp);
 364 #endif
 365         }
 366
 367         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 368                 if (global_diff[i])
 369                         atomic_long_add(global_diff[i], &vm_stat[i]);
 370 }
 371
 372 #endif
 373
 374 #ifdef CONFIG_NUMA
 375 /*
 376  * zonelist = the list of zones passed to the allocator
 377  * z        = the zone from which the allocation occurred.
 378  *
 379  * Must be called with interrupts disabled.
 380  */
 381 void zone_statistics(struct zone *preferred_zone, struct zone *z)
 382 {
 383         if (z->zone_pgdat == preferred_zone->zone_pgdat) {
 384                 __inc_zone_state(z, NUMA_HIT);
 385         } else {
 386                 __inc_zone_state(z, NUMA_MISS);
 387                 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
 388         }
 389         if (z->node == numa_node_id())
 390                 __inc_zone_state(z, NUMA_LOCAL);
 391         else
 392                 __inc_zone_state(z, NUMA_OTHER);
 393 }
 394 #endif
 395
 396 #ifdef CONFIG_COMPACTION
 397 struct contig_page_info {
 398         unsigned long free_pages;
 399         unsigned long free_blocks_total;
 400         unsigned long free_blocks_suitable;
 401 };
 402
 403 /*
 404  * Calculate the number of free pages in a zone, how many contiguous
 405  * pages are free and how many are large enough to satisfy an allocation of
 406  * the target size. Note that this function makes no attempt to estimate
 407  * how many suitable free blocks there *might* be if MOVABLE pages were
 408  * migrated. Calculating that is possible, but expensive and can be
 409  * figured out from userspace
 410  */
 411 static void fill_contig_page_info(struct zone *zone,
 412                                 unsigned int suitable_order,
 413                                 struct contig_page_info *info)
 414 {
 415         unsigned int order;
 416
 417         info->free_pages = 0;
 418         info->free_blocks_total = 0;
 419         info->free_blocks_suitable = 0;
 420
 421         for (order = 0; order < MAX_ORDER; order++) {
 422                 unsigned long blocks;
 423
 424                 /* Count number of free blocks */
 425                 blocks = zone->free_area[order].nr_free;
 426                 info->free_blocks_total += blocks;
 427
 428                 /* Count free base pages */
 429                 info->free_pages += blocks << order;
 430
 431                 /* Count the suitable free blocks */
 432                 if (order >= suitable_order)
 433                         info->free_blocks_suitable += blocks <<
 434                                                 (order - suitable_order);
 435         }
 436 }
 437
 438 /*
 439  * A fragmentation index only makes sense if an allocation of a requested
 440  * size would fail. If that is true, the fragmentation index indicates
 441  * whether external fragmentation or a lack of memory was the problem.
 442  * The value can be used to determine if page reclaim or compaction
 443  * should be used
 444  */
 445 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
 446 {
 447         unsigned long requested = 1UL << order;
 448
 449         if (!info->free_blocks_total)
 450                 return 0;
 451
 452         /* Fragmentation index only makes sense when a request would fail */
 453         if (info->free_blocks_suitable)
 454                 return -1000;
 455
 456         /*
 457          * Index is between 0 and 1 so return within 3 decimal places
 458          *
 459          * 0 => allocation would fail due to lack of memory
 460          * 1 => allocation would fail due to fragmentation
 461          */
 462         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
 463 }
 464
 465 /* Same as __fragmentation index but allocs contig_page_info on stack */
 466 int fragmentation_index(struct zone *zone, unsigned int order)
 467 {
 468         struct contig_page_info info;
 469
 470         fill_contig_page_info(zone, order, &info);
 471         return __fragmentation_index(order, &info);
 472 }
 473 #endif
 474
 475 #if defined(CONFIG_PROC_FS) || defined(CONFIG_COMPACTION)
 476 #include <linux/proc_fs.h>
 477 #include <linux/seq_file.h>
 478
 479 static char * const migratetype_names[MIGRATE_TYPES] = {
 480         "Unmovable",
 481         "Reclaimable",
 482         "Movable",
 483         "Reserve",
 484         "Isolate",
 485 };
 486
 487 static void *frag_start(struct seq_file *m, loff_t *pos)
 488 {
 489         pg_data_t *pgdat;
 490         loff_t node = *pos;
 491         for (pgdat = first_online_pgdat();
 492              pgdat && node;
 493              pgdat = next_online_pgdat(pgdat))
 494                 --node;
 495
 496         return pgdat;
 497 }
 498
 499 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
 500 {
 501         pg_data_t *pgdat = (pg_data_t *)arg;
 502
 503         (*pos)++;
 504         return next_online_pgdat(pgdat);
 505 }
 506
 507 static void frag_stop(struct seq_file *m, void *arg)
 508 {
 509 }
 510
 511 /* Walk all the zones in a node and print using a callback */
 512 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
 513                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
 514 {
 515         struct zone *zone;
 516         struct zone *node_zones = pgdat->node_zones;
 517         unsigned long flags;
 518
 519         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 520                 if (!populated_zone(zone))
 521                         continue;
 522
 523                 spin_lock_irqsave(&zone->lock, flags);
 524                 print(m, pgdat, zone);
 525                 spin_unlock_irqrestore(&zone->lock, flags);
 526         }
 527 }
 528 #endif
 529
 530 #ifdef CONFIG_PROC_FS
 531 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
 532                                                 struct zone *zone)
 533 {
 534         int order;
 535
 536         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 537         for (order = 0; order < MAX_ORDER; ++order)
 538                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
 539         seq_putc(m, '\n');
 540 }
 541
 542 /*
 543  * This walks the free areas for each zone.
 544  */
 545 static int frag_show(struct seq_file *m, void *arg)
 546 {
 547         pg_data_t *pgdat = (pg_data_t *)arg;
 548         walk_zones_in_node(m, pgdat, frag_show_print);
 549         return 0;
 550 }
 551
 552 static void pagetypeinfo_showfree_print(struct seq_file *m,
 553                                         pg_data_t *pgdat, struct zone *zone)
 554 {
 555         int order, mtype;
 556
 557         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
 558                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
 559                                         pgdat->node_id,
 560                                         zone->name,
 561                                         migratetype_names[mtype]);
 562                 for (order = 0; order < MAX_ORDER; ++order) {
 563                         unsigned long freecount = 0;
 564                         struct free_area *area;
 565                         struct list_head *curr;
 566
 567                         area = &(zone->free_area[order]);
 568
 569                         list_for_each(curr, &area->free_list[mtype])
 570                                 freecount++;
 571                         seq_printf(m, "%6lu ", freecount);
 572                 }
 573                 seq_putc(m, '\n');
 574         }
 575 }
 576
 577 /* Print out the free pages at each order for each migatetype */
 578 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
 579 {
 580         int order;
 581         pg_data_t *pgdat = (pg_data_t *)arg;
 582
 583         /* Print header */
 584         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
 585         for (order = 0; order < MAX_ORDER; ++order)
 586                 seq_printf(m, "%6d ", order);
 587         seq_putc(m, '\n');
 588
 589         walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
 590
 591         return 0;
 592 }
 593
 594 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
 595                                         pg_data_t *pgdat, struct zone *zone)
 596 {
 597         int mtype;
 598         unsigned long pfn;
 599         unsigned long start_pfn = zone->zone_start_pfn;
 600         unsigned long end_pfn = start_pfn + zone->spanned_pages;
 601         unsigned long count[MIGRATE_TYPES] = { 0, };
 602
 603         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
 604                 struct page *page;
 605
 606                 if (!pfn_valid(pfn))
 607                         continue;
 608
 609                 page = pfn_to_page(pfn);
 610
 611                 /* Watch for unexpected holes punched in the memmap */
 612                 if (!memmap_valid_within(pfn, page, zone))
 613                         continue;
 614
 615                 mtype = get_pageblock_migratetype(page);
 616
 617                 if (mtype < MIGRATE_TYPES)
 618                         count[mtype]++;
 619         }
 620
 621         /* Print counts */
 622         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 623         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
 624                 seq_printf(m, "%12lu ", count[mtype]);
 625         seq_putc(m, '\n');
 626 }
 627
 628 /* Print out the free pages at each order for each migratetype */
 629 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
 630 {
 631         int mtype;
 632         pg_data_t *pgdat = (pg_data_t *)arg;
 633
 634         seq_printf(m, "\n%-23s", "Number of blocks type ");
 635         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
 636                 seq_printf(m, "%12s ", migratetype_names[mtype]);
 637         seq_putc(m, '\n');
 638         walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
 639
 640         return 0;
 641 }
 642
 643 /*
 644  * This prints out statistics in relation to grouping pages by mobility.
 645  * It is expensive to collect so do not constantly read the file.
 646  */
 647 static int pagetypeinfo_show(struct seq_file *m, void *arg)
 648 {
 649         pg_data_t *pgdat = (pg_data_t *)arg;
 650
 651         /* check memoryless node */
 652         if (!node_state(pgdat->node_id, N_HIGH_MEMORY))
 653                 return 0;
 654
 655         seq_printf(m, "Page block order: %d\n", pageblock_order);
 656         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
 657         seq_putc(m, '\n');
 658         pagetypeinfo_showfree(m, pgdat);
 659         pagetypeinfo_showblockcount(m, pgdat);
 660
 661         return 0;
 662 }
 663
 664 static const struct seq_operations fragmentation_op = {
 665         .start  = frag_start,
 666         .next   = frag_next,
 667         .stop   = frag_stop,
 668         .show   = frag_show,
 669 };
 670
 671 static int fragmentation_open(struct inode *inode, struct file *file)
 672 {
 673         return seq_open(file, &fragmentation_op);
 674 }
 675
 676 static const struct file_operations fragmentation_file_operations = {
 677         .open           = fragmentation_open,
 678         .read           = seq_read,
 679         .llseek         = seq_lseek,
 680         .release        = seq_release,
 681 };
 682
 683 static const struct seq_operations pagetypeinfo_op = {
 684         .start  = frag_start,
 685         .next   = frag_next,
 686         .stop   = frag_stop,
 687         .show   = pagetypeinfo_show,
 688 };
 689
 690 static int pagetypeinfo_open(struct inode *inode, struct file *file)
 691 {
 692         return seq_open(file, &pagetypeinfo_op);
 693 }
 694
 695 static const struct file_operations pagetypeinfo_file_ops = {
 696         .open           = pagetypeinfo_open,
 697         .read           = seq_read,
 698         .llseek         = seq_lseek,
 699         .release        = seq_release,
 700 };
 701
 702 #ifdef CONFIG_ZONE_DMA
 703 #define TEXT_FOR_DMA(xx) xx "_dma",
 704 #else
 705 #define TEXT_FOR_DMA(xx)
 706 #endif
 707
 708 #ifdef CONFIG_ZONE_DMA32
 709 #define TEXT_FOR_DMA32(xx) xx "_dma32",
 710 #else
 711 #define TEXT_FOR_DMA32(xx)
 712 #endif
 713
 714 #ifdef CONFIG_HIGHMEM
 715 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
 716 #else
 717 #define TEXT_FOR_HIGHMEM(xx)
 718 #endif
 719
 720 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
 721                                         TEXT_FOR_HIGHMEM(xx) xx "_movable",
 722
 723 static const char * const vmstat_text[] = {
 724         /* Zoned VM counters */
 725         "nr_free_pages",
 726         "nr_inactive_anon",
 727         "nr_active_anon",
 728         "nr_inactive_file",
 729         "nr_active_file",
 730         "nr_unevictable",
 731         "nr_mlock",
 732         "nr_anon_pages",
 733         "nr_mapped",
 734         "nr_file_pages",
 735         "nr_dirty",
 736         "nr_writeback",
 737         "nr_slab_reclaimable",
 738         "nr_slab_unreclaimable",
 739         "nr_page_table_pages",
 740         "nr_kernel_stack",
 741         "nr_unstable",
 742         "nr_bounce",
 743         "nr_vmscan_write",
 744         "nr_writeback_temp",
 745         "nr_isolated_anon",
 746         "nr_isolated_file",
 747         "nr_shmem",
 748 #ifdef CONFIG_NUMA
 749         "numa_hit",
 750         "numa_miss",
 751         "numa_foreign",
 752         "numa_interleave",
 753         "numa_local",
 754         "numa_other",
 755 #endif
 756
 757 #ifdef CONFIG_VM_EVENT_COUNTERS
 758         "pgpgin",
 759         "pgpgout",
 760         "pswpin",
 761         "pswpout",
 762
 763         TEXTS_FOR_ZONES("pgalloc")
 764
 765         "pgfree",
 766         "pgactivate",
 767         "pgdeactivate",
 768
 769         "pgfault",
 770         "pgmajfault",
 771
 772         TEXTS_FOR_ZONES("pgrefill")
 773         TEXTS_FOR_ZONES("pgsteal")
 774         TEXTS_FOR_ZONES("pgscan_kswapd")
 775         TEXTS_FOR_ZONES("pgscan_direct")
 776
 777 #ifdef CONFIG_NUMA
 778         "zone_reclaim_failed",
 779 #endif
 780         "pginodesteal",
 781         "slabs_scanned",
 782         "kswapd_steal",
 783         "kswapd_inodesteal",
 784         "kswapd_low_wmark_hit_quickly",
 785         "kswapd_high_wmark_hit_quickly",
 786         "kswapd_skip_congestion_wait",
 787         "pageoutrun",
 788         "allocstall",
 789
 790         "pgrotated",
 791
 792 #ifdef CONFIG_COMPACTION
 793         "compact_blocks_moved",
 794         "compact_pages_moved",
 795         "compact_pagemigrate_failed",
 796         "compact_stall",
 797         "compact_fail",
 798         "compact_success",
 799 #endif
 800
 801 #ifdef CONFIG_HUGETLB_PAGE
 802         "htlb_buddy_alloc_success",
 803         "htlb_buddy_alloc_fail",
 804 #endif
 805         "unevictable_pgs_culled",
 806         "unevictable_pgs_scanned",
 807         "unevictable_pgs_rescued",
 808         "unevictable_pgs_mlocked",
 809         "unevictable_pgs_munlocked",
 810         "unevictable_pgs_cleared",
 811         "unevictable_pgs_stranded",
 812         "unevictable_pgs_mlockfreed",
 813 #endif
 814 };
 815
 816 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
 817                                                         struct zone *zone)
 818 {
 819         int i;
 820         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
 821         seq_printf(m,
 822                    "\n  pages free     %lu"
 823                    "\n        min      %lu"
 824                    "\n        low      %lu"
 825                    "\n        high     %lu"
 826                    "\n        scanned  %lu"
 827                    "\n        spanned  %lu"
 828                    "\n        present  %lu",
 829                    zone_nr_free_pages(zone),
 830                    min_wmark_pages(zone),
 831                    low_wmark_pages(zone),
 832                    high_wmark_pages(zone),
 833                    zone->pages_scanned,
 834                    zone->spanned_pages,
 835                    zone->present_pages);
 836
 837         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 838                 seq_printf(m, "\n    %-12s %lu", vmstat_text[i],
 839                                 zone_page_state(zone, i));
 840
 841         seq_printf(m,
 842                    "\n        protection: (%lu",
 843                    zone->lowmem_reserve[0]);
 844         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
 845                 seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
 846         seq_printf(m,
 847                    ")"
 848                    "\n  pagesets");
 849         for_each_online_cpu(i) {
 850                 struct per_cpu_pageset *pageset;
 851
 852                 pageset = per_cpu_ptr(zone->pageset, i);
 853                 seq_printf(m,
 854                            "\n    cpu: %i"
 855                            "\n              count: %i"
 856                            "\n              high:  %i"
 857                            "\n              batch: %i",
 858                            i,
 859                            pageset->pcp.count,
 860                            pageset->pcp.high,
 861                            pageset->pcp.batch);
 862 #ifdef CONFIG_SMP
 863                 seq_printf(m, "\n  vm stats threshold: %d",
 864                                 pageset->stat_threshold);
 865 #endif
 866         }
 867         seq_printf(m,
 868                    "\n  all_unreclaimable: %u"
 869                    "\n  prev_priority:     %i"
 870                    "\n  start_pfn:         %lu"
 871                    "\n  inactive_ratio:    %u",
 872                    zone->all_unreclaimable,
 873                    zone->prev_priority,
 874                    zone->zone_start_pfn,
 875                    zone->inactive_ratio);
 876         seq_putc(m, '\n');
 877 }
 878
 879 /*
 880  * Output information about zones in @pgdat.
 881  */
 882 static int zoneinfo_show(struct seq_file *m, void *arg)
 883 {
 884         pg_data_t *pgdat = (pg_data_t *)arg;
 885         walk_zones_in_node(m, pgdat, zoneinfo_show_print);
 886         return 0;
 887 }
 888
 889 static const struct seq_operations zoneinfo_op = {
 890         .start  = frag_start, /* iterate over all zones. The same as in
 891                                * fragmentation. */
 892         .next   = frag_next,
 893         .stop   = frag_stop,
 894         .show   = zoneinfo_show,
 895 };
 896
 897 static int zoneinfo_open(struct inode *inode, struct file *file)
 898 {
 899         return seq_open(file, &zoneinfo_op);
 900 }
 901
 902 static const struct file_operations proc_zoneinfo_file_operations = {
 903         .open           = zoneinfo_open,
 904         .read           = seq_read,
 905         .llseek         = seq_lseek,
 906         .release        = seq_release,
 907 };
 908
 909 static void *vmstat_start(struct seq_file *m, loff_t *pos)
 910 {
 911         unsigned long *v;
 912 #ifdef CONFIG_VM_EVENT_COUNTERS
 913         unsigned long *e;
 914 #endif
 915         int i;
 916
 917         if (*pos >= ARRAY_SIZE(vmstat_text))
 918                 return NULL;
 919
 920 #ifdef CONFIG_VM_EVENT_COUNTERS
 921         v = kmalloc(NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long)
 922                         + sizeof(struct vm_event_state), GFP_KERNEL);
 923 #else
 924         v = kmalloc(NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long),
 925                         GFP_KERNEL);
 926 #endif
 927         m->private = v;
 928         if (!v)
 929                 return ERR_PTR(-ENOMEM);
 930         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 931                 v[i] = global_page_state(i);
 932 #ifdef CONFIG_VM_EVENT_COUNTERS
 933         e = v + NR_VM_ZONE_STAT_ITEMS;
 934         all_vm_events(e);
 935         e[PGPGIN] /= 2;         /* sectors -> kbytes */
 936         e[PGPGOUT] /= 2;
 937 #endif
 938         return v + *pos;
 939 }
 940
 941 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
 942 {
 943         (*pos)++;
 944         if (*pos >= ARRAY_SIZE(vmstat_text))
 945                 return NULL;
 946         return (unsigned long *)m->private + *pos;
 947 }
 948
 949 static int vmstat_show(struct seq_file *m, void *arg)
 950 {
 951         unsigned long *l = arg;
 952         unsigned long off = l - (unsigned long *)m->private;
 953
 954         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
 955         return 0;
 956 }
 957
 958 static void vmstat_stop(struct seq_file *m, void *arg)
 959 {
 960         kfree(m->private);
 961         m->private = NULL;
 962 }
 963
 964 static const struct seq_operations vmstat_op = {
 965         .start  = vmstat_start,
 966         .next   = vmstat_next,
 967         .stop   = vmstat_stop,
 968         .show   = vmstat_show,
 969 };
 970
 971 static int vmstat_open(struct inode *inode, struct file *file)
 972 {
 973         return seq_open(file, &vmstat_op);
 974 }
 975
 976 static const struct file_operations proc_vmstat_file_operations = {
 977         .open           = vmstat_open,
 978         .read           = seq_read,
 979         .llseek         = seq_lseek,
 980         .release        = seq_release,
 981 };
 982 #endif /* CONFIG_PROC_FS */
 983
 984 #ifdef CONFIG_SMP
 985 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
 986 int sysctl_stat_interval __read_mostly = HZ;
 987
 988 static void vmstat_update(struct work_struct *w)
 989 {
 990         refresh_cpu_vm_stats(smp_processor_id());
 991         schedule_delayed_work(&__get_cpu_var(vmstat_work),
 992                 round_jiffies_relative(sysctl_stat_interval));
 993 }
 994
 995 static void __cpuinit start_cpu_timer(int cpu)
 996 {
 997         struct delayed_work *work = &per_cpu(vmstat_work, cpu);
 998
 999         INIT_DELAYED_WORK_DEFERRABLE(work, vmstat_update);
1000         schedule_delayed_work_on(cpu, work, __round_jiffies_relative(HZ, cpu));
1001 }
1002
1003 /*
1004  * Use the cpu notifier to insure that the thresholds are recalculated
1005  * when necessary.
1006  */
1007 static int __cpuinit vmstat_cpuup_callback(struct notifier_block *nfb,
1008                 unsigned long action,
1009                 void *hcpu)
1010 {
1011         long cpu = (long)hcpu;
1012
1013         switch (action) {
1014         case CPU_ONLINE:
1015         case CPU_ONLINE_FROZEN:
1016                 start_cpu_timer(cpu);
1017                 node_set_state(cpu_to_node(cpu), N_CPU);
1018                 break;
1019         case CPU_DOWN_PREPARE:
1020         case CPU_DOWN_PREPARE_FROZEN:
1021                 cancel_rearming_delayed_work(&per_cpu(vmstat_work, cpu));
1022                 per_cpu(vmstat_work, cpu).work.func = NULL;
1023                 break;
1024         case CPU_DOWN_FAILED:
1025         case CPU_DOWN_FAILED_FROZEN:
1026                 start_cpu_timer(cpu);
1027                 break;
1028         case CPU_DEAD:
1029         case CPU_DEAD_FROZEN:
1030                 refresh_zone_stat_thresholds();
1031                 break;
1032         default:
1033                 break;
1034         }
1035         return NOTIFY_OK;
1036 }
1037
1038 static struct notifier_block __cpuinitdata vmstat_notifier =
1039         { &vmstat_cpuup_callback, NULL, 0 };
1040 #endif
1041
1042 static int __init setup_vmstat(void)
1043 {
1044 #ifdef CONFIG_SMP
1045         int cpu;
1046
1047         refresh_zone_stat_thresholds();
1048         register_cpu_notifier(&vmstat_notifier);
1049
1050         for_each_online_cpu(cpu)
1051                 start_cpu_timer(cpu);
1052 #endif
1053 #ifdef CONFIG_PROC_FS
1054         proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1055         proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1056         proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1057         proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1058 #endif
1059         return 0;
1060 }
1061 module_init(setup_vmstat)
1062
1063 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1064 #include <linux/debugfs.h>
1065
1066 static struct dentry *extfrag_debug_root;
1067
1068 /*
1069  * Return an index indicating how much of the available free memory is
1070  * unusable for an allocation of the requested size.
1071  */
1072 static int unusable_free_index(unsigned int order,
1073                                 struct contig_page_info *info)
1074 {
1075         /* No free memory is interpreted as all free memory is unusable */
1076         if (info->free_pages == 0)
1077                 return 1000;
1078
1079         /*
1080          * Index should be a value between 0 and 1. Return a value to 3
1081          * decimal places.
1082          *
1083          * 0 => no fragmentation
1084          * 1 => high fragmentation
1085          */
1086         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1087
1088 }
1089
1090 static void unusable_show_print(struct seq_file *m,
1091                                         pg_data_t *pgdat, struct zone *zone)
1092 {
1093         unsigned int order;
1094         int index;
1095         struct contig_page_info info;
1096
1097         seq_printf(m, "Node %d, zone %8s ",
1098                                 pgdat->node_id,
1099                                 zone->name);
1100         for (order = 0; order < MAX_ORDER; ++order) {
1101                 fill_contig_page_info(zone, order, &info);
1102                 index = unusable_free_index(order, &info);
1103                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1104         }
1105
1106         seq_putc(m, '\n');
1107 }
1108
1109 /*
1110  * Display unusable free space index
1111  *
1112  * The unusable free space index measures how much of the available free
1113  * memory cannot be used to satisfy an allocation of a given size and is a
1114  * value between 0 and 1. The higher the value, the more of free memory is
1115  * unusable and by implication, the worse the external fragmentation is. This
1116  * can be expressed as a percentage by multiplying by 100.
1117  */
1118 static int unusable_show(struct seq_file *m, void *arg)
1119 {
1120         pg_data_t *pgdat = (pg_data_t *)arg;
1121
1122         /* check memoryless node */
1123         if (!node_state(pgdat->node_id, N_HIGH_MEMORY))
1124                 return 0;
1125
1126         walk_zones_in_node(m, pgdat, unusable_show_print);
1127
1128         return 0;
1129 }
1130
1131 static const struct seq_operations unusable_op = {
1132         .start  = frag_start,
1133         .next   = frag_next,
1134         .stop   = frag_stop,
1135         .show   = unusable_show,
1136 };
1137
1138 static int unusable_open(struct inode *inode, struct file *file)
1139 {
1140         return seq_open(file, &unusable_op);
1141 }
1142
1143 static const struct file_operations unusable_file_ops = {
1144         .open           = unusable_open,
1145         .read           = seq_read,
1146         .llseek         = seq_lseek,
1147         .release        = seq_release,
1148 };
1149
1150 static void extfrag_show_print(struct seq_file *m,
1151                                         pg_data_t *pgdat, struct zone *zone)
1152 {
1153         unsigned int order;
1154         int index;
1155
1156         /* Alloc on stack as interrupts are disabled for zone walk */
1157         struct contig_page_info info;
1158
1159         seq_printf(m, "Node %d, zone %8s ",
1160                                 pgdat->node_id,
1161                                 zone->name);
1162         for (order = 0; order < MAX_ORDER; ++order) {
1163                 fill_contig_page_info(zone, order, &info);
1164                 index = __fragmentation_index(order, &info);
1165                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1166         }
1167
1168         seq_putc(m, '\n');
1169 }
1170
1171 /*
1172  * Display fragmentation index for orders that allocations would fail for
1173  */
1174 static int extfrag_show(struct seq_file *m, void *arg)
1175 {
1176         pg_data_t *pgdat = (pg_data_t *)arg;
1177
1178         walk_zones_in_node(m, pgdat, extfrag_show_print);
1179
1180         return 0;
1181 }
1182
1183 static const struct seq_operations extfrag_op = {
1184         .start  = frag_start,
1185         .next   = frag_next,
1186         .stop   = frag_stop,
1187         .show   = extfrag_show,
1188 };
1189
1190 static int extfrag_open(struct inode *inode, struct file *file)
1191 {
1192         return seq_open(file, &extfrag_op);
1193 }
1194
1195 static const struct file_operations extfrag_file_ops = {
1196         .open           = extfrag_open,
1197         .read           = seq_read,
1198         .llseek         = seq_lseek,
1199         .release        = seq_release,
1200 };
1201
1202 static int __init extfrag_debug_init(void)
1203 {
1204         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1205         if (!extfrag_debug_root)
1206                 return -ENOMEM;
1207
1208         if (!debugfs_create_file("unusable_index", 0444,
1209                         extfrag_debug_root, NULL, &unusable_file_ops))
1210                 return -ENOMEM;
1211
1212         if (!debugfs_create_file("extfrag_index", 0444,
1213                         extfrag_debug_root, NULL, &extfrag_file_ops))
1214                 return -ENOMEM;
1215
1216         return 0;
1217 }
1218
1219 module_init(extfrag_debug_init);
1220 #endif