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