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 int calculate_pressure_threshold(struct zone *zone)
  87 {
  88         int threshold;
  89         int watermark_distance;
  90
  91         /*
  92          * As vmstats are not up to date, there is drift between the estimated
  93          * and real values. For high thresholds and a high number of CPUs, it
  94          * is possible for the min watermark to be breached while the estimated
  95          * value looks fine. The pressure threshold is a reduced value such
  96          * that even the maximum amount of drift will not accidentally breach
  97          * the min watermark
  98          */
  99         watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
 100         threshold = max(1, (int)(watermark_distance / num_online_cpus()));
 101
 102         /*
 103          * Maximum threshold is 125
 104          */
 105         threshold = min(125, threshold);
 106
 107         return threshold;
 108 }
 109
 110 int calculate_normal_threshold(struct zone *zone)
 111 {
 112         int threshold;
 113         int mem;        /* memory in 128 MB units */
 114
 115         /*
 116          * The threshold scales with the number of processors and the amount
 117          * of memory per zone. More memory means that we can defer updates for
 118          * longer, more processors could lead to more contention.
 119          * fls() is used to have a cheap way of logarithmic scaling.
 120          *
 121          * Some sample thresholds:
 122          *
 123          * Threshold    Processors      (fls)   Zonesize        fls(mem+1)
 124          * ------------------------------------------------------------------
 125          * 8            1               1       0.9-1 GB        4
 126          * 16           2               2       0.9-1 GB        4
 127          * 20           2               2       1-2 GB          5
 128          * 24           2               2       2-4 GB          6
 129          * 28           2               2       4-8 GB          7
 130          * 32           2               2       8-16 GB         8
 131          * 4            2               2       <128M           1
 132          * 30           4               3       2-4 GB          5
 133          * 48           4               3       8-16 GB         8
 134          * 32           8               4       1-2 GB          4
 135          * 32           8               4       0.9-1GB         4
 136          * 10           16              5       <128M           1
 137          * 40           16              5       900M            4
 138          * 70           64              7       2-4 GB          5
 139          * 84           64              7       4-8 GB          6
 140          * 108          512             9       4-8 GB          6
 141          * 125          1024            10      8-16 GB         8
 142          * 125          1024            10      16-32 GB        9
 143          */
 144
 145         mem = zone->present_pages >> (27 - PAGE_SHIFT);
 146
 147         threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
 148
 149         /*
 150          * Maximum threshold is 125
 151          */
 152         threshold = min(125, threshold);
 153
 154         return threshold;
 155 }
 156
 157 /*
 158  * Refresh the thresholds for each zone.
 159  */
 160 static void refresh_zone_stat_thresholds(void)
 161 {
 162         struct zone *zone;
 163         int cpu;
 164         int threshold;
 165
 166         for_each_populated_zone(zone) {
 167                 unsigned long max_drift, tolerate_drift;
 168
 169                 threshold = calculate_normal_threshold(zone);
 170
 171                 for_each_online_cpu(cpu)
 172                         per_cpu_ptr(zone->pageset, cpu)->stat_threshold
 173                                                         = threshold;
 174
 175                 /*
 176                  * Only set percpu_drift_mark if there is a danger that
 177                  * NR_FREE_PAGES reports the low watermark is ok when in fact
 178                  * the min watermark could be breached by an allocation
 179                  */
 180                 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
 181                 max_drift = num_online_cpus() * threshold;
 182                 if (max_drift > tolerate_drift)
 183                         zone->percpu_drift_mark = high_wmark_pages(zone) +
 184                                         max_drift;
 185         }
 186 }
 187
 188 void set_pgdat_percpu_threshold(pg_data_t *pgdat,
 189                                 int (*calculate_pressure)(struct zone *))
 190 {
 191         struct zone *zone;
 192         int cpu;
 193         int threshold;
 194         int i;
 195
 196         for (i = 0; i < pgdat->nr_zones; i++) {
 197                 zone = &pgdat->node_zones[i];
 198                 if (!zone->percpu_drift_mark)
 199                         continue;
 200
 201                 threshold = (*calculate_pressure)(zone);
 202                 for_each_possible_cpu(cpu)
 203                         per_cpu_ptr(zone->pageset, cpu)->stat_threshold
 204                                                         = threshold;
 205         }
 206 }
 207
 208 /*
 209  * For use when we know that interrupts are disabled.
 210  */
 211 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 212                                 int delta)
 213 {
 214         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 215         s8 __percpu *p = pcp->vm_stat_diff + item;
 216         long x;
 217         long t;
 218
 219         x = delta + __this_cpu_read(*p);
 220
 221         t = __this_cpu_read(pcp->stat_threshold);
 222
 223         if (unlikely(x > t || x < -t)) {
 224                 zone_page_state_add(x, zone, item);
 225                 x = 0;
 226         }
 227         __this_cpu_write(*p, x);
 228 }
 229 EXPORT_SYMBOL(__mod_zone_page_state);
 230
 231 /*
 232  * Optimized increment and decrement functions.
 233  *
 234  * These are only for a single page and therefore can take a struct page *
 235  * argument instead of struct zone *. This allows the inclusion of the code
 236  * generated for page_zone(page) into the optimized functions.
 237  *
 238  * No overflow check is necessary and therefore the differential can be
 239  * incremented or decremented in place which may allow the compilers to
 240  * generate better code.
 241  * The increment or decrement is known and therefore one boundary check can
 242  * be omitted.
 243  *
 244  * NOTE: These functions are very performance sensitive. Change only
 245  * with care.
 246  *
 247  * Some processors have inc/dec instructions that are atomic vs an interrupt.
 248  * However, the code must first determine the differential location in a zone
 249  * based on the processor number and then inc/dec the counter. There is no
 250  * guarantee without disabling preemption that the processor will not change
 251  * in between and therefore the atomicity vs. interrupt cannot be exploited
 252  * in a useful way here.
 253  */
 254 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
 255 {
 256         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 257         s8 __percpu *p = pcp->vm_stat_diff + item;
 258         s8 v, t;
 259
 260         v = __this_cpu_inc_return(*p);
 261         t = __this_cpu_read(pcp->stat_threshold);
 262         if (unlikely(v > t)) {
 263                 s8 overstep = t >> 1;
 264
 265                 zone_page_state_add(v + overstep, zone, item);
 266                 __this_cpu_write(*p, -overstep);
 267         }
 268 }
 269
 270 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
 271 {
 272         __inc_zone_state(page_zone(page), item);
 273 }
 274 EXPORT_SYMBOL(__inc_zone_page_state);
 275
 276 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
 277 {
 278         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 279         s8 __percpu *p = pcp->vm_stat_diff + item;
 280         s8 v, t;
 281
 282         v = __this_cpu_dec_return(*p);
 283         t = __this_cpu_read(pcp->stat_threshold);
 284         if (unlikely(v < - t)) {
 285                 s8 overstep = t >> 1;
 286
 287                 zone_page_state_add(v - overstep, zone, item);
 288                 __this_cpu_write(*p, overstep);
 289         }
 290 }
 291
 292 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
 293 {
 294         __dec_zone_state(page_zone(page), item);
 295 }
 296 EXPORT_SYMBOL(__dec_zone_page_state);
 297
 298 #ifdef CONFIG_CMPXCHG_LOCAL
 299 /*
 300  * If we have cmpxchg_local support then we do not need to incur the overhead
 301  * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
 302  *
 303  * mod_state() modifies the zone counter state through atomic per cpu
 304  * operations.
 305  *
 306  * Overstep mode specifies how overstep should handled:
 307  *     0       No overstepping
 308  *     1       Overstepping half of threshold
 309  *     -1      Overstepping minus half of threshold
 310 */
 311 static inline void mod_state(struct zone *zone,
 312        enum zone_stat_item item, int delta, int overstep_mode)
 313 {
 314         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 315         s8 __percpu *p = pcp->vm_stat_diff + item;
 316         long o, n, t, z;
 317
 318         do {
 319                 z = 0;  /* overflow to zone counters */
 320
 321                 /*
 322                  * The fetching of the stat_threshold is racy. We may apply
 323                  * a counter threshold to the wrong the cpu if we get
 324                  * rescheduled while executing here. However, the following
 325                  * will apply the threshold again and therefore bring the
 326                  * counter under the threshold.
 327                  */
 328                 t = this_cpu_read(pcp->stat_threshold);
 329
 330                 o = this_cpu_read(*p);
 331                 n = delta + o;
 332
 333                 if (n > t || n < -t) {
 334                         int os = overstep_mode * (t >> 1) ;
 335
 336                         /* Overflow must be added to zone counters */
 337                         z = n + os;
 338                         n = -os;
 339                 }
 340         } while (this_cpu_cmpxchg(*p, o, n) != o);
 341
 342         if (z)
 343                 zone_page_state_add(z, zone, item);
 344 }
 345
 346 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 347                                         int delta)
 348 {
 349         mod_state(zone, item, delta, 0);
 350 }
 351 EXPORT_SYMBOL(mod_zone_page_state);
 352
 353 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
 354 {
 355         mod_state(zone, item, 1, 1);
 356 }
 357
 358 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 359 {
 360         mod_state(page_zone(page), item, 1, 1);
 361 }
 362 EXPORT_SYMBOL(inc_zone_page_state);
 363
 364 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 365 {
 366         mod_state(page_zone(page), item, -1, -1);
 367 }
 368 EXPORT_SYMBOL(dec_zone_page_state);
 369 #else
 370 /*
 371  * Use interrupt disable to serialize counter updates
 372  */
 373 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 374                                         int delta)
 375 {
 376         unsigned long flags;
 377
 378         local_irq_save(flags);
 379         __mod_zone_page_state(zone, item, delta);
 380         local_irq_restore(flags);
 381 }
 382 EXPORT_SYMBOL(mod_zone_page_state);
 383
 384 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
 385 {
 386         unsigned long flags;
 387
 388         local_irq_save(flags);
 389         __inc_zone_state(zone, item);
 390         local_irq_restore(flags);
 391 }
 392
 393 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 394 {
 395         unsigned long flags;
 396         struct zone *zone;
 397
 398         zone = page_zone(page);
 399         local_irq_save(flags);
 400         __inc_zone_state(zone, item);
 401         local_irq_restore(flags);
 402 }
 403 EXPORT_SYMBOL(inc_zone_page_state);
 404
 405 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 406 {
 407         unsigned long flags;
 408
 409         local_irq_save(flags);
 410         __dec_zone_page_state(page, item);
 411         local_irq_restore(flags);
 412 }
 413 EXPORT_SYMBOL(dec_zone_page_state);
 414 #endif
 415
 416 /*
 417  * Update the zone counters for one cpu.
 418  *
 419  * The cpu specified must be either the current cpu or a processor that
 420  * is not online. If it is the current cpu then the execution thread must
 421  * be pinned to the current cpu.
 422  *
 423  * Note that refresh_cpu_vm_stats strives to only access
 424  * node local memory. The per cpu pagesets on remote zones are placed
 425  * in the memory local to the processor using that pageset. So the
 426  * loop over all zones will access a series of cachelines local to
 427  * the processor.
 428  *
 429  * The call to zone_page_state_add updates the cachelines with the
 430  * statistics in the remote zone struct as well as the global cachelines
 431  * with the global counters. These could cause remote node cache line
 432  * bouncing and will have to be only done when necessary.
 433  */
 434 void refresh_cpu_vm_stats(int cpu)
 435 {
 436         struct zone *zone;
 437         int i;
 438         int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 439
 440         for_each_populated_zone(zone) {
 441                 struct per_cpu_pageset *p;
 442
 443                 p = per_cpu_ptr(zone->pageset, cpu);
 444
 445                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 446                         if (p->vm_stat_diff[i]) {
 447                                 unsigned long flags;
 448                                 int v;
 449
 450                                 local_irq_save(flags);
 451                                 v = p->vm_stat_diff[i];
 452                                 p->vm_stat_diff[i] = 0;
 453                                 local_irq_restore(flags);
 454                                 atomic_long_add(v, &zone->vm_stat[i]);
 455                                 global_diff[i] += v;
 456 #ifdef CONFIG_NUMA
 457                                 /* 3 seconds idle till flush */
 458                                 p->expire = 3;
 459 #endif
 460                         }
 461                 cond_resched();
 462 #ifdef CONFIG_NUMA
 463                 /*
 464                  * Deal with draining the remote pageset of this
 465                  * processor
 466                  *
 467                  * Check if there are pages remaining in this pageset
 468                  * if not then there is nothing to expire.
 469                  */
 470                 if (!p->expire || !p->pcp.count)
 471                         continue;
 472
 473                 /*
 474                  * We never drain zones local to this processor.
 475                  */
 476                 if (zone_to_nid(zone) == numa_node_id()) {
 477                         p->expire = 0;
 478                         continue;
 479                 }
 480
 481                 p->expire--;
 482                 if (p->expire)
 483                         continue;
 484
 485                 if (p->pcp.count)
 486                         drain_zone_pages(zone, &p->pcp);
 487 #endif
 488         }
 489
 490         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 491                 if (global_diff[i])
 492                         atomic_long_add(global_diff[i], &vm_stat[i]);
 493 }
 494
 495 #endif
 496
 497 #ifdef CONFIG_NUMA
 498 /*
 499  * zonelist = the list of zones passed to the allocator
 500  * z        = the zone from which the allocation occurred.
 501  *
 502  * Must be called with interrupts disabled.
 503  */
 504 void zone_statistics(struct zone *preferred_zone, struct zone *z)
 505 {
 506         if (z->zone_pgdat == preferred_zone->zone_pgdat) {
 507                 __inc_zone_state(z, NUMA_HIT);
 508         } else {
 509                 __inc_zone_state(z, NUMA_MISS);
 510                 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
 511         }
 512         if (z->node == numa_node_id())
 513                 __inc_zone_state(z, NUMA_LOCAL);
 514         else
 515                 __inc_zone_state(z, NUMA_OTHER);
 516 }
 517 #endif
 518
 519 #ifdef CONFIG_COMPACTION
 520
 521 struct contig_page_info {
 522         unsigned long free_pages;
 523         unsigned long free_blocks_total;
 524         unsigned long free_blocks_suitable;
 525 };
 526
 527 /*
 528  * Calculate the number of free pages in a zone, how many contiguous
 529  * pages are free and how many are large enough to satisfy an allocation of
 530  * the target size. Note that this function makes no attempt to estimate
 531  * how many suitable free blocks there *might* be if MOVABLE pages were
 532  * migrated. Calculating that is possible, but expensive and can be
 533  * figured out from userspace
 534  */
 535 static void fill_contig_page_info(struct zone *zone,
 536                                 unsigned int suitable_order,
 537                                 struct contig_page_info *info)
 538 {
 539         unsigned int order;
 540
 541         info->free_pages = 0;
 542         info->free_blocks_total = 0;
 543         info->free_blocks_suitable = 0;
 544
 545         for (order = 0; order < MAX_ORDER; order++) {
 546                 unsigned long blocks;
 547
 548                 /* Count number of free blocks */
 549                 blocks = zone->free_area[order].nr_free;
 550                 info->free_blocks_total += blocks;
 551
 552                 /* Count free base pages */
 553                 info->free_pages += blocks << order;
 554
 555                 /* Count the suitable free blocks */
 556                 if (order >= suitable_order)
 557                         info->free_blocks_suitable += blocks <<
 558                                                 (order - suitable_order);
 559         }
 560 }
 561
 562 /*
 563  * A fragmentation index only makes sense if an allocation of a requested
 564  * size would fail. If that is true, the fragmentation index indicates
 565  * whether external fragmentation or a lack of memory was the problem.
 566  * The value can be used to determine if page reclaim or compaction
 567  * should be used
 568  */
 569 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
 570 {
 571         unsigned long requested = 1UL << order;
 572
 573         if (!info->free_blocks_total)
 574                 return 0;
 575
 576         /* Fragmentation index only makes sense when a request would fail */
 577         if (info->free_blocks_suitable)
 578                 return -1000;
 579
 580         /*
 581          * Index is between 0 and 1 so return within 3 decimal places
 582          *
 583          * 0 => allocation would fail due to lack of memory
 584          * 1 => allocation would fail due to fragmentation
 585          */
 586         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
 587 }
 588
 589 /* Same as __fragmentation index but allocs contig_page_info on stack */
 590 int fragmentation_index(struct zone *zone, unsigned int order)
 591 {
 592         struct contig_page_info info;
 593
 594         fill_contig_page_info(zone, order, &info);
 595         return __fragmentation_index(order, &info);
 596 }
 597 #endif
 598
 599 #if defined(CONFIG_PROC_FS) || defined(CONFIG_COMPACTION)
 600 #include <linux/proc_fs.h>
 601 #include <linux/seq_file.h>
 602
 603 static char * const migratetype_names[MIGRATE_TYPES] = {
 604         "Unmovable",
 605         "Reclaimable",
 606         "Movable",
 607         "Reserve",
 608         "Isolate",
 609 };
 610
 611 static void *frag_start(struct seq_file *m, loff_t *pos)
 612 {
 613         pg_data_t *pgdat;
 614         loff_t node = *pos;
 615         for (pgdat = first_online_pgdat();
 616              pgdat && node;
 617              pgdat = next_online_pgdat(pgdat))
 618                 --node;
 619
 620         return pgdat;
 621 }
 622
 623 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
 624 {
 625         pg_data_t *pgdat = (pg_data_t *)arg;
 626
 627         (*pos)++;
 628         return next_online_pgdat(pgdat);
 629 }
 630
 631 static void frag_stop(struct seq_file *m, void *arg)
 632 {
 633 }
 634
 635 /* Walk all the zones in a node and print using a callback */
 636 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
 637                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
 638 {
 639         struct zone *zone;
 640         struct zone *node_zones = pgdat->node_zones;
 641         unsigned long flags;
 642
 643         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 644                 if (!populated_zone(zone))
 645                         continue;
 646
 647                 spin_lock_irqsave(&zone->lock, flags);
 648                 print(m, pgdat, zone);
 649                 spin_unlock_irqrestore(&zone->lock, flags);
 650         }
 651 }
 652 #endif
 653
 654 #ifdef CONFIG_PROC_FS
 655 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
 656                                                 struct zone *zone)
 657 {
 658         int order;
 659
 660         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 661         for (order = 0; order < MAX_ORDER; ++order)
 662                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
 663         seq_putc(m, '\n');
 664 }
 665
 666 /*
 667  * This walks the free areas for each zone.
 668  */
 669 static int frag_show(struct seq_file *m, void *arg)
 670 {
 671         pg_data_t *pgdat = (pg_data_t *)arg;
 672         walk_zones_in_node(m, pgdat, frag_show_print);
 673         return 0;
 674 }
 675
 676 static void pagetypeinfo_showfree_print(struct seq_file *m,
 677                                         pg_data_t *pgdat, struct zone *zone)
 678 {
 679         int order, mtype;
 680
 681         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
 682                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
 683                                         pgdat->node_id,
 684                                         zone->name,
 685                                         migratetype_names[mtype]);
 686                 for (order = 0; order < MAX_ORDER; ++order) {
 687                         unsigned long freecount = 0;
 688                         struct free_area *area;
 689                         struct list_head *curr;
 690
 691                         area = &(zone->free_area[order]);
 692
 693                         list_for_each(curr, &area->free_list[mtype])
 694                                 freecount++;
 695                         seq_printf(m, "%6lu ", freecount);
 696                 }
 697                 seq_putc(m, '\n');
 698         }
 699 }
 700
 701 /* Print out the free pages at each order for each migatetype */
 702 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
 703 {
 704         int order;
 705         pg_data_t *pgdat = (pg_data_t *)arg;
 706
 707         /* Print header */
 708         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
 709         for (order = 0; order < MAX_ORDER; ++order)
 710                 seq_printf(m, "%6d ", order);
 711         seq_putc(m, '\n');
 712
 713         walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
 714
 715         return 0;
 716 }
 717
 718 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
 719                                         pg_data_t *pgdat, struct zone *zone)
 720 {
 721         int mtype;
 722         unsigned long pfn;
 723         unsigned long start_pfn = zone->zone_start_pfn;
 724         unsigned long end_pfn = start_pfn + zone->spanned_pages;
 725         unsigned long count[MIGRATE_TYPES] = { 0, };
 726
 727         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
 728                 struct page *page;
 729
 730                 if (!pfn_valid(pfn))
 731                         continue;
 732
 733                 page = pfn_to_page(pfn);
 734
 735                 /* Watch for unexpected holes punched in the memmap */
 736                 if (!memmap_valid_within(pfn, page, zone))
 737                         continue;
 738
 739                 mtype = get_pageblock_migratetype(page);
 740
 741                 if (mtype < MIGRATE_TYPES)
 742                         count[mtype]++;
 743         }
 744
 745         /* Print counts */
 746         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 747         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
 748                 seq_printf(m, "%12lu ", count[mtype]);
 749         seq_putc(m, '\n');
 750 }
 751
 752 /* Print out the free pages at each order for each migratetype */
 753 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
 754 {
 755         int mtype;
 756         pg_data_t *pgdat = (pg_data_t *)arg;
 757
 758         seq_printf(m, "\n%-23s", "Number of blocks type ");
 759         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
 760                 seq_printf(m, "%12s ", migratetype_names[mtype]);
 761         seq_putc(m, '\n');
 762         walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
 763
 764         return 0;
 765 }
 766
 767 /*
 768  * This prints out statistics in relation to grouping pages by mobility.
 769  * It is expensive to collect so do not constantly read the file.
 770  */
 771 static int pagetypeinfo_show(struct seq_file *m, void *arg)
 772 {
 773         pg_data_t *pgdat = (pg_data_t *)arg;
 774
 775         /* check memoryless node */
 776         if (!node_state(pgdat->node_id, N_HIGH_MEMORY))
 777                 return 0;
 778
 779         seq_printf(m, "Page block order: %d\n", pageblock_order);
 780         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
 781         seq_putc(m, '\n');
 782         pagetypeinfo_showfree(m, pgdat);
 783         pagetypeinfo_showblockcount(m, pgdat);
 784
 785         return 0;
 786 }
 787
 788 static const struct seq_operations fragmentation_op = {
 789         .start  = frag_start,
 790         .next   = frag_next,
 791         .stop   = frag_stop,
 792         .show   = frag_show,
 793 };
 794
 795 static int fragmentation_open(struct inode *inode, struct file *file)
 796 {
 797         return seq_open(file, &fragmentation_op);
 798 }
 799
 800 static const struct file_operations fragmentation_file_operations = {
 801         .open           = fragmentation_open,
 802         .read           = seq_read,
 803         .llseek         = seq_lseek,
 804         .release        = seq_release,
 805 };
 806
 807 static const struct seq_operations pagetypeinfo_op = {
 808         .start  = frag_start,
 809         .next   = frag_next,
 810         .stop   = frag_stop,
 811         .show   = pagetypeinfo_show,
 812 };
 813
 814 static int pagetypeinfo_open(struct inode *inode, struct file *file)
 815 {
 816         return seq_open(file, &pagetypeinfo_op);
 817 }
 818
 819 static const struct file_operations pagetypeinfo_file_ops = {
 820         .open           = pagetypeinfo_open,
 821         .read           = seq_read,
 822         .llseek         = seq_lseek,
 823         .release        = seq_release,
 824 };
 825
 826 #ifdef CONFIG_ZONE_DMA
 827 #define TEXT_FOR_DMA(xx) xx "_dma",
 828 #else
 829 #define TEXT_FOR_DMA(xx)
 830 #endif
 831
 832 #ifdef CONFIG_ZONE_DMA32
 833 #define TEXT_FOR_DMA32(xx) xx "_dma32",
 834 #else
 835 #define TEXT_FOR_DMA32(xx)
 836 #endif
 837
 838 #ifdef CONFIG_HIGHMEM
 839 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
 840 #else
 841 #define TEXT_FOR_HIGHMEM(xx)
 842 #endif
 843
 844 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
 845                                         TEXT_FOR_HIGHMEM(xx) xx "_movable",
 846
 847 static const char * const vmstat_text[] = {
 848         /* Zoned VM counters */
 849         "nr_free_pages",
 850         "nr_inactive_anon",
 851         "nr_active_anon",
 852         "nr_inactive_file",
 853         "nr_active_file",
 854         "nr_unevictable",
 855         "nr_mlock",
 856         "nr_anon_pages",
 857         "nr_mapped",
 858         "nr_file_pages",
 859         "nr_dirty",
 860         "nr_writeback",
 861         "nr_slab_reclaimable",
 862         "nr_slab_unreclaimable",
 863         "nr_page_table_pages",
 864         "nr_kernel_stack",
 865         "nr_unstable",
 866         "nr_bounce",
 867         "nr_vmscan_write",
 868         "nr_writeback_temp",
 869         "nr_isolated_anon",
 870         "nr_isolated_file",
 871         "nr_shmem",
 872         "nr_dirtied",
 873         "nr_written",
 874
 875 #ifdef CONFIG_NUMA
 876         "numa_hit",
 877         "numa_miss",
 878         "numa_foreign",
 879         "numa_interleave",
 880         "numa_local",
 881         "numa_other",
 882 #endif
 883         "nr_dirty_threshold",
 884         "nr_dirty_background_threshold",
 885
 886 #ifdef CONFIG_VM_EVENT_COUNTERS
 887         "pgpgin",
 888         "pgpgout",
 889         "pswpin",
 890         "pswpout",
 891
 892         TEXTS_FOR_ZONES("pgalloc")
 893
 894         "pgfree",
 895         "pgactivate",
 896         "pgdeactivate",
 897
 898         "pgfault",
 899         "pgmajfault",
 900
 901         TEXTS_FOR_ZONES("pgrefill")
 902         TEXTS_FOR_ZONES("pgsteal")
 903         TEXTS_FOR_ZONES("pgscan_kswapd")
 904         TEXTS_FOR_ZONES("pgscan_direct")
 905
 906 #ifdef CONFIG_NUMA
 907         "zone_reclaim_failed",
 908 #endif
 909         "pginodesteal",
 910         "slabs_scanned",
 911         "kswapd_steal",
 912         "kswapd_inodesteal",
 913         "kswapd_low_wmark_hit_quickly",
 914         "kswapd_high_wmark_hit_quickly",
 915         "kswapd_skip_congestion_wait",
 916         "pageoutrun",
 917         "allocstall",
 918
 919         "pgrotated",
 920
 921 #ifdef CONFIG_COMPACTION
 922         "compact_blocks_moved",
 923         "compact_pages_moved",
 924         "compact_pagemigrate_failed",
 925         "compact_stall",
 926         "compact_fail",
 927         "compact_success",
 928 #endif
 929
 930 #ifdef CONFIG_HUGETLB_PAGE
 931         "htlb_buddy_alloc_success",
 932         "htlb_buddy_alloc_fail",
 933 #endif
 934         "unevictable_pgs_culled",
 935         "unevictable_pgs_scanned",
 936         "unevictable_pgs_rescued",
 937         "unevictable_pgs_mlocked",
 938         "unevictable_pgs_munlocked",
 939         "unevictable_pgs_cleared",
 940         "unevictable_pgs_stranded",
 941         "unevictable_pgs_mlockfreed",
 942 #endif
 943 };
 944
 945 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
 946                                                         struct zone *zone)
 947 {
 948         int i;
 949         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
 950         seq_printf(m,
 951                    "\n  pages free     %lu"
 952                    "\n        min      %lu"
 953                    "\n        low      %lu"
 954                    "\n        high     %lu"
 955                    "\n        scanned  %lu"
 956                    "\n        spanned  %lu"
 957                    "\n        present  %lu",
 958                    zone_page_state(zone, NR_FREE_PAGES),
 959                    min_wmark_pages(zone),
 960                    low_wmark_pages(zone),
 961                    high_wmark_pages(zone),
 962                    zone->pages_scanned,
 963                    zone->spanned_pages,
 964                    zone->present_pages);
 965
 966         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 967                 seq_printf(m, "\n    %-12s %lu", vmstat_text[i],
 968                                 zone_page_state(zone, i));
 969
 970         seq_printf(m,
 971                    "\n        protection: (%lu",
 972                    zone->lowmem_reserve[0]);
 973         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
 974                 seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
 975         seq_printf(m,
 976                    ")"
 977                    "\n  pagesets");
 978         for_each_online_cpu(i) {
 979                 struct per_cpu_pageset *pageset;
 980
 981                 pageset = per_cpu_ptr(zone->pageset, i);
 982                 seq_printf(m,
 983                            "\n    cpu: %i"
 984                            "\n              count: %i"
 985                            "\n              high:  %i"
 986                            "\n              batch: %i",
 987                            i,
 988                            pageset->pcp.count,
 989                            pageset->pcp.high,
 990                            pageset->pcp.batch);
 991 #ifdef CONFIG_SMP
 992                 seq_printf(m, "\n  vm stats threshold: %d",
 993                                 pageset->stat_threshold);
 994 #endif
 995         }
 996         seq_printf(m,
 997                    "\n  all_unreclaimable: %u"
 998                    "\n  start_pfn:         %lu"
 999                    "\n  inactive_ratio:    %u",
1000                    zone->all_unreclaimable,
1001                    zone->zone_start_pfn,
1002                    zone->inactive_ratio);
1003         seq_putc(m, '\n');
1004 }
1005
1006 /*
1007  * Output information about zones in @pgdat.
1008  */
1009 static int zoneinfo_show(struct seq_file *m, void *arg)
1010 {
1011         pg_data_t *pgdat = (pg_data_t *)arg;
1012         walk_zones_in_node(m, pgdat, zoneinfo_show_print);
1013         return 0;
1014 }
1015
1016 static const struct seq_operations zoneinfo_op = {
1017         .start  = frag_start, /* iterate over all zones. The same as in
1018                                * fragmentation. */
1019         .next   = frag_next,
1020         .stop   = frag_stop,
1021         .show   = zoneinfo_show,
1022 };
1023
1024 static int zoneinfo_open(struct inode *inode, struct file *file)
1025 {
1026         return seq_open(file, &zoneinfo_op);
1027 }
1028
1029 static const struct file_operations proc_zoneinfo_file_operations = {
1030         .open           = zoneinfo_open,
1031         .read           = seq_read,
1032         .llseek         = seq_lseek,
1033         .release        = seq_release,
1034 };
1035
1036 enum writeback_stat_item {
1037         NR_DIRTY_THRESHOLD,
1038         NR_DIRTY_BG_THRESHOLD,
1039         NR_VM_WRITEBACK_STAT_ITEMS,
1040 };
1041
1042 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1043 {
1044         unsigned long *v;
1045         int i, stat_items_size;
1046
1047         if (*pos >= ARRAY_SIZE(vmstat_text))
1048                 return NULL;
1049         stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1050                           NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1051
1052 #ifdef CONFIG_VM_EVENT_COUNTERS
1053         stat_items_size += sizeof(struct vm_event_state);
1054 #endif
1055
1056         v = kmalloc(stat_items_size, GFP_KERNEL);
1057         m->private = v;
1058         if (!v)
1059                 return ERR_PTR(-ENOMEM);
1060         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1061                 v[i] = global_page_state(i);
1062         v += NR_VM_ZONE_STAT_ITEMS;
1063
1064         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1065                             v + NR_DIRTY_THRESHOLD);
1066         v += NR_VM_WRITEBACK_STAT_ITEMS;
1067
1068 #ifdef CONFIG_VM_EVENT_COUNTERS
1069         all_vm_events(v);
1070         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1071         v[PGPGOUT] /= 2;
1072 #endif
1073         return (unsigned long *)m->private + *pos;
1074 }
1075
1076 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1077 {
1078         (*pos)++;
1079         if (*pos >= ARRAY_SIZE(vmstat_text))
1080                 return NULL;
1081         return (unsigned long *)m->private + *pos;
1082 }
1083
1084 static int vmstat_show(struct seq_file *m, void *arg)
1085 {
1086         unsigned long *l = arg;
1087         unsigned long off = l - (unsigned long *)m->private;
1088
1089         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1090         return 0;
1091 }
1092
1093 static void vmstat_stop(struct seq_file *m, void *arg)
1094 {
1095         kfree(m->private);
1096         m->private = NULL;
1097 }
1098
1099 static const struct seq_operations vmstat_op = {
1100         .start  = vmstat_start,
1101         .next   = vmstat_next,
1102         .stop   = vmstat_stop,
1103         .show   = vmstat_show,
1104 };
1105
1106 static int vmstat_open(struct inode *inode, struct file *file)
1107 {
1108         return seq_open(file, &vmstat_op);
1109 }
1110
1111 static const struct file_operations proc_vmstat_file_operations = {
1112         .open           = vmstat_open,
1113         .read           = seq_read,
1114         .llseek         = seq_lseek,
1115         .release        = seq_release,
1116 };
1117 #endif /* CONFIG_PROC_FS */
1118
1119 #ifdef CONFIG_SMP
1120 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1121 int sysctl_stat_interval __read_mostly = HZ;
1122
1123 static void vmstat_update(struct work_struct *w)
1124 {
1125         refresh_cpu_vm_stats(smp_processor_id());
1126         schedule_delayed_work(&__get_cpu_var(vmstat_work),
1127                 round_jiffies_relative(sysctl_stat_interval));
1128 }
1129
1130 static void __cpuinit start_cpu_timer(int cpu)
1131 {
1132         struct delayed_work *work = &per_cpu(vmstat_work, cpu);
1133
1134         INIT_DELAYED_WORK_DEFERRABLE(work, vmstat_update);
1135         schedule_delayed_work_on(cpu, work, __round_jiffies_relative(HZ, cpu));
1136 }
1137
1138 /*
1139  * Use the cpu notifier to insure that the thresholds are recalculated
1140  * when necessary.
1141  */
1142 static int __cpuinit vmstat_cpuup_callback(struct notifier_block *nfb,
1143                 unsigned long action,
1144                 void *hcpu)
1145 {
1146         long cpu = (long)hcpu;
1147
1148         switch (action) {
1149         case CPU_ONLINE:
1150         case CPU_ONLINE_FROZEN:
1151                 refresh_zone_stat_thresholds();
1152                 start_cpu_timer(cpu);
1153                 node_set_state(cpu_to_node(cpu), N_CPU);
1154                 break;
1155         case CPU_DOWN_PREPARE:
1156         case CPU_DOWN_PREPARE_FROZEN:
1157                 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1158                 per_cpu(vmstat_work, cpu).work.func = NULL;
1159                 break;
1160         case CPU_DOWN_FAILED:
1161         case CPU_DOWN_FAILED_FROZEN:
1162                 start_cpu_timer(cpu);
1163                 break;
1164         case CPU_DEAD:
1165         case CPU_DEAD_FROZEN:
1166                 refresh_zone_stat_thresholds();
1167                 break;
1168         default:
1169                 break;
1170         }
1171         return NOTIFY_OK;
1172 }
1173
1174 static struct notifier_block __cpuinitdata vmstat_notifier =
1175         { &vmstat_cpuup_callback, NULL, 0 };
1176 #endif
1177
1178 static int __init setup_vmstat(void)
1179 {
1180 #ifdef CONFIG_SMP
1181         int cpu;
1182
1183         refresh_zone_stat_thresholds();
1184         register_cpu_notifier(&vmstat_notifier);
1185
1186         for_each_online_cpu(cpu)
1187                 start_cpu_timer(cpu);
1188 #endif
1189 #ifdef CONFIG_PROC_FS
1190         proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1191         proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1192         proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1193         proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1194 #endif
1195         return 0;
1196 }
1197 module_init(setup_vmstat)
1198
1199 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1200 #include <linux/debugfs.h>
1201
1202 static struct dentry *extfrag_debug_root;
1203
1204 /*
1205  * Return an index indicating how much of the available free memory is
1206  * unusable for an allocation of the requested size.
1207  */
1208 static int unusable_free_index(unsigned int order,
1209                                 struct contig_page_info *info)
1210 {
1211         /* No free memory is interpreted as all free memory is unusable */
1212         if (info->free_pages == 0)
1213                 return 1000;
1214
1215         /*
1216          * Index should be a value between 0 and 1. Return a value to 3
1217          * decimal places.
1218          *
1219          * 0 => no fragmentation
1220          * 1 => high fragmentation
1221          */
1222         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1223
1224 }
1225
1226 static void unusable_show_print(struct seq_file *m,
1227                                         pg_data_t *pgdat, struct zone *zone)
1228 {
1229         unsigned int order;
1230         int index;
1231         struct contig_page_info info;
1232
1233         seq_printf(m, "Node %d, zone %8s ",
1234                                 pgdat->node_id,
1235                                 zone->name);
1236         for (order = 0; order < MAX_ORDER; ++order) {
1237                 fill_contig_page_info(zone, order, &info);
1238                 index = unusable_free_index(order, &info);
1239                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1240         }
1241
1242         seq_putc(m, '\n');
1243 }
1244
1245 /*
1246  * Display unusable free space index
1247  *
1248  * The unusable free space index measures how much of the available free
1249  * memory cannot be used to satisfy an allocation of a given size and is a
1250  * value between 0 and 1. The higher the value, the more of free memory is
1251  * unusable and by implication, the worse the external fragmentation is. This
1252  * can be expressed as a percentage by multiplying by 100.
1253  */
1254 static int unusable_show(struct seq_file *m, void *arg)
1255 {
1256         pg_data_t *pgdat = (pg_data_t *)arg;
1257
1258         /* check memoryless node */
1259         if (!node_state(pgdat->node_id, N_HIGH_MEMORY))
1260                 return 0;
1261
1262         walk_zones_in_node(m, pgdat, unusable_show_print);
1263
1264         return 0;
1265 }
1266
1267 static const struct seq_operations unusable_op = {
1268         .start  = frag_start,
1269         .next   = frag_next,
1270         .stop   = frag_stop,
1271         .show   = unusable_show,
1272 };
1273
1274 static int unusable_open(struct inode *inode, struct file *file)
1275 {
1276         return seq_open(file, &unusable_op);
1277 }
1278
1279 static const struct file_operations unusable_file_ops = {
1280         .open           = unusable_open,
1281         .read           = seq_read,
1282         .llseek         = seq_lseek,
1283         .release        = seq_release,
1284 };
1285
1286 static void extfrag_show_print(struct seq_file *m,
1287                                         pg_data_t *pgdat, struct zone *zone)
1288 {
1289         unsigned int order;
1290         int index;
1291
1292         /* Alloc on stack as interrupts are disabled for zone walk */
1293         struct contig_page_info info;
1294
1295         seq_printf(m, "Node %d, zone %8s ",
1296                                 pgdat->node_id,
1297                                 zone->name);
1298         for (order = 0; order < MAX_ORDER; ++order) {
1299                 fill_contig_page_info(zone, order, &info);
1300                 index = __fragmentation_index(order, &info);
1301                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1302         }
1303
1304         seq_putc(m, '\n');
1305 }
1306
1307 /*
1308  * Display fragmentation index for orders that allocations would fail for
1309  */
1310 static int extfrag_show(struct seq_file *m, void *arg)
1311 {
1312         pg_data_t *pgdat = (pg_data_t *)arg;
1313
1314         walk_zones_in_node(m, pgdat, extfrag_show_print);
1315
1316         return 0;
1317 }
1318
1319 static const struct seq_operations extfrag_op = {
1320         .start  = frag_start,
1321         .next   = frag_next,
1322         .stop   = frag_stop,
1323         .show   = extfrag_show,
1324 };
1325
1326 static int extfrag_open(struct inode *inode, struct file *file)
1327 {
1328         return seq_open(file, &extfrag_op);
1329 }
1330
1331 static const struct file_operations extfrag_file_ops = {
1332         .open           = extfrag_open,
1333         .read           = seq_read,
1334         .llseek         = seq_lseek,
1335         .release        = seq_release,
1336 };
1337
1338 static int __init extfrag_debug_init(void)
1339 {
1340         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1341         if (!extfrag_debug_root)
1342                 return -ENOMEM;
1343
1344         if (!debugfs_create_file("unusable_index", 0444,
1345                         extfrag_debug_root, NULL, &unusable_file_ops))
1346                 return -ENOMEM;
1347
1348         if (!debugfs_create_file("extfrag_index", 0444,
1349                         extfrag_debug_root, NULL, &extfrag_file_ops))
1350                 return -ENOMEM;
1351
1352         return 0;
1353 }
1354
1355 module_init(extfrag_debug_init);
1356 #endif