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