mm/vmstat.c

   1 /*
   2  *  linux/mm/vmstat.c
   3  *
   4  *  Manages VM statistics
   5  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   6  *
   7  *  zoned VM statistics
   8  *  Copyright (C) 2006 Silicon Graphics, Inc.,
   9  *              Christoph Lameter <christoph@lameter.com>
  10  */
  11 #include <linux/fs.h>
  12 #include <linux/mm.h>
  13 #include <linux/err.h>
  14 #include <linux/module.h>
  15 #include <linux/slab.h>
  16 #include <linux/cpu.h>
  17 #include <linux/vmstat.h>
  18 #include <linux/sched.h>
  19 #include <linux/math64.h>
  20
  21 #ifdef CONFIG_VM_EVENT_COUNTERS
  22 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
  23 EXPORT_PER_CPU_SYMBOL(vm_event_states);
  24
  25 static void sum_vm_events(unsigned long *ret, const struct cpumask *cpumask)
  26 {
  27         int cpu;
  28         int i;
  29
  30         memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
  31
  32         for_each_cpu(cpu, cpumask) {
  33                 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
  34
  35                 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
  36                         ret[i] += this->event[i];
  37         }
  38 }
  39
  40 /*
  41  * Accumulate the vm event counters across all CPUs.
  42  * The result is unavoidably approximate - it can change
  43  * during and after execution of this function.
  44 */
  45 void all_vm_events(unsigned long *ret)
  46 {
  47         get_online_cpus();
  48         sum_vm_events(ret, cpu_online_mask);
  49         put_online_cpus();
  50 }
  51 EXPORT_SYMBOL_GPL(all_vm_events);
  52
  53 #ifdef CONFIG_HOTPLUG
  54 /*
  55  * Fold the foreign cpu events into our own.
  56  *
  57  * This is adding to the events on one processor
  58  * but keeps the global counts constant.
  59  */
  60 void vm_events_fold_cpu(int cpu)
  61 {
  62         struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
  63         int i;
  64
  65         for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
  66                 count_vm_events(i, fold_state->event[i]);
  67                 fold_state->event[i] = 0;
  68         }
  69 }
  70 #endif /* CONFIG_HOTPLUG */
  71
  72 #endif /* CONFIG_VM_EVENT_COUNTERS */
  73
  74 /*
  75  * Manage combined zone based / global counters
  76  *
  77  * vm_stat contains the global counters
  78  */
  79 atomic_long_t vm_stat[NR_VM_ZONE_STAT_ITEMS];
  80 EXPORT_SYMBOL(vm_stat);
  81
  82 #ifdef CONFIG_SMP
  83
  84 static int calculate_threshold(struct zone *zone)
  85 {
  86         int threshold;
  87         int mem;        /* memory in 128 MB units */
  88
  89         /*
  90          * The threshold scales with the number of processors and the amount
  91          * of memory per zone. More memory means that we can defer updates for
  92          * longer, more processors could lead to more contention.
  93          * fls() is used to have a cheap way of logarithmic scaling.
  94          *
  95          * Some sample thresholds:
  96          *
  97          * Threshold    Processors      (fls)   Zonesize        fls(mem+1)
  98          * ------------------------------------------------------------------
  99          * 8            1               1       0.9-1 GB        4
 100          * 16           2               2       0.9-1 GB        4
 101          * 20           2               2       1-2 GB          5
 102          * 24           2               2       2-4 GB          6
 103          * 28           2               2       4-8 GB          7
 104          * 32           2               2       8-16 GB         8
 105          * 4            2               2       <128M           1
 106          * 30           4               3       2-4 GB          5
 107          * 48           4               3       8-16 GB         8
 108          * 32           8               4       1-2 GB          4
 109          * 32           8               4       0.9-1GB         4
 110          * 10           16              5       <128M           1
 111          * 40           16              5       900M            4
 112          * 70           64              7       2-4 GB          5
 113          * 84           64              7       4-8 GB          6
 114          * 108          512             9       4-8 GB          6
 115          * 125          1024            10      8-16 GB         8
 116          * 125          1024            10      16-32 GB        9
 117          */
 118
 119         mem = zone->present_pages >> (27 - PAGE_SHIFT);
 120
 121         threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
 122
 123         /*
 124          * Maximum threshold is 125
 125          */
 126         threshold = min(125, threshold);
 127
 128         return threshold;
 129 }
 130
 131 /*
 132  * Refresh the thresholds for each zone.
 133  */
 134 static void refresh_zone_stat_thresholds(void)
 135 {
 136         struct zone *zone;
 137         int cpu;
 138         int threshold;
 139
 140         for_each_populated_zone(zone) {
 141                 threshold = calculate_threshold(zone);
 142
 143                 for_each_online_cpu(cpu)
 144                         per_cpu_ptr(zone->pageset, cpu)->stat_threshold
 145                                                         = threshold;
 146         }
 147 }
 148
 149 /*
 150  * For use when we know that interrupts are disabled.
 151  */
 152 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 153                                 int delta)
 154 {
 155         struct per_cpu_pageset *pcp = this_cpu_ptr(zone->pageset);
 156
 157         s8 *p = pcp->vm_stat_diff + item;
 158         long x;
 159
 160         x = delta + *p;
 161
 162         if (unlikely(x > pcp->stat_threshold || x < -pcp->stat_threshold)) {
 163                 zone_page_state_add(x, zone, item);
 164                 x = 0;
 165         }
 166         *p = x;
 167 }
 168 EXPORT_SYMBOL(__mod_zone_page_state);
 169
 170 /*
 171  * For an unknown interrupt state
 172  */
 173 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 174                                         int delta)
 175 {
 176         unsigned long flags;
 177
 178         local_irq_save(flags);
 179         __mod_zone_page_state(zone, item, delta);
 180         local_irq_restore(flags);
 181 }
 182 EXPORT_SYMBOL(mod_zone_page_state);
 183
 184 /*
 185  * Optimized increment and decrement functions.
 186  *
 187  * These are only for a single page and therefore can take a struct page *
 188  * argument instead of struct zone *. This allows the inclusion of the code
 189  * generated for page_zone(page) into the optimized functions.
 190  *
 191  * No overflow check is necessary and therefore the differential can be
 192  * incremented or decremented in place which may allow the compilers to
 193  * generate better code.
 194  * The increment or decrement is known and therefore one boundary check can
 195  * be omitted.
 196  *
 197  * NOTE: These functions are very performance sensitive. Change only
 198  * with care.
 199  *
 200  * Some processors have inc/dec instructions that are atomic vs an interrupt.
 201  * However, the code must first determine the differential location in a zone
 202  * based on the processor number and then inc/dec the counter. There is no
 203  * guarantee without disabling preemption that the processor will not change
 204  * in between and therefore the atomicity vs. interrupt cannot be exploited
 205  * in a useful way here.
 206  */
 207 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
 208 {
 209         struct per_cpu_pageset *pcp = this_cpu_ptr(zone->pageset);
 210         s8 *p = pcp->vm_stat_diff + item;
 211
 212         (*p)++;
 213
 214         if (unlikely(*p > pcp->stat_threshold)) {
 215                 int overstep = pcp->stat_threshold / 2;
 216
 217                 zone_page_state_add(*p + overstep, zone, item);
 218                 *p = -overstep;
 219         }
 220 }
 221
 222 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
 223 {
 224         __inc_zone_state(page_zone(page), item);
 225 }
 226 EXPORT_SYMBOL(__inc_zone_page_state);
 227
 228 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
 229 {
 230         struct per_cpu_pageset *pcp = this_cpu_ptr(zone->pageset);
 231         s8 *p = pcp->vm_stat_diff + item;
 232
 233         (*p)--;
 234
 235         if (unlikely(*p < - pcp->stat_threshold)) {
 236                 int overstep = pcp->stat_threshold / 2;
 237
 238                 zone_page_state_add(*p - overstep, zone, item);
 239                 *p = overstep;
 240         }
 241 }
 242
 243 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
 244 {
 245         __dec_zone_state(page_zone(page), item);
 246 }
 247 EXPORT_SYMBOL(__dec_zone_page_state);
 248
 249 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
 250 {
 251         unsigned long flags;
 252
 253         local_irq_save(flags);
 254         __inc_zone_state(zone, item);
 255         local_irq_restore(flags);
 256 }
 257
 258 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 259 {
 260         unsigned long flags;
 261         struct zone *zone;
 262
 263         zone = page_zone(page);
 264         local_irq_save(flags);
 265         __inc_zone_state(zone, item);
 266         local_irq_restore(flags);
 267 }
 268 EXPORT_SYMBOL(inc_zone_page_state);
 269
 270 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 271 {
 272         unsigned long flags;
 273
 274         local_irq_save(flags);
 275         __dec_zone_page_state(page, item);
 276         local_irq_restore(flags);
 277 }
 278 EXPORT_SYMBOL(dec_zone_page_state);
 279
 280 /*
 281  * Update the zone counters for one cpu.
 282  *
 283  * The cpu specified must be either the current cpu or a processor that
 284  * is not online. If it is the current cpu then the execution thread must
 285  * be pinned to the current cpu.
 286  *
 287  * Note that refresh_cpu_vm_stats strives to only access
 288  * node local memory. The per cpu pagesets on remote zones are placed
 289  * in the memory local to the processor using that pageset. So the
 290  * loop over all zones will access a series of cachelines local to
 291  * the processor.
 292  *
 293  * The call to zone_page_state_add updates the cachelines with the
 294  * statistics in the remote zone struct as well as the global cachelines
 295  * with the global counters. These could cause remote node cache line
 296  * bouncing and will have to be only done when necessary.
 297  */
 298 void refresh_cpu_vm_stats(int cpu)
 299 {
 300         struct zone *zone;
 301         int i;
 302         int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 303
 304         for_each_populated_zone(zone) {
 305                 struct per_cpu_pageset *p;
 306
 307                 p = per_cpu_ptr(zone->pageset, cpu);
 308
 309                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 310                         if (p->vm_stat_diff[i]) {
 311                                 unsigned long flags;
 312                                 int v;
 313
 314                                 local_irq_save(flags);
 315                                 v = p->vm_stat_diff[i];
 316                                 p->vm_stat_diff[i] = 0;
 317                                 local_irq_restore(flags);
 318                                 atomic_long_add(v, &zone->vm_stat[i]);
 319                                 global_diff[i] += v;
 320 #ifdef CONFIG_NUMA
 321                                 /* 3 seconds idle till flush */
 322                                 p->expire = 3;
 323 #endif
 324                         }
 325                 cond_resched();
 326 #ifdef CONFIG_NUMA
 327                 /*
 328                  * Deal with draining the remote pageset of this
 329                  * processor
 330                  *
 331                  * Check if there are pages remaining in this pageset
 332                  * if not then there is nothing to expire.
 333                  */
 334                 if (!p->expire || !p->pcp.count)
 335                         continue;
 336
 337                 /*
 338                  * We never drain zones local to this processor.
 339                  */
 340                 if (zone_to_nid(zone) == numa_node_id()) {
 341                         p->expire = 0;
 342                         continue;
 343                 }
 344
 345                 p->expire--;
 346                 if (p->expire)
 347                         continue;
 348
 349                 if (p->pcp.count)
 350                         drain_zone_pages(zone, &p->pcp);
 351 #endif
 352         }
 353
 354         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 355                 if (global_diff[i])
 356                         atomic_long_add(global_diff[i], &vm_stat[i]);
 357 }
 358
 359 #endif
 360
 361 #ifdef CONFIG_NUMA
 362 /*
 363  * zonelist = the list of zones passed to the allocator
 364  * z        = the zone from which the allocation occurred.
 365  *
 366  * Must be called with interrupts disabled.
 367  */
 368 void zone_statistics(struct zone *preferred_zone, struct zone *z)
 369 {
 370         if (z->zone_pgdat == preferred_zone->zone_pgdat) {
 371                 __inc_zone_state(z, NUMA_HIT);
 372         } else {
 373                 __inc_zone_state(z, NUMA_MISS);
 374                 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
 375         }
 376         if (z->node == numa_node_id())
 377                 __inc_zone_state(z, NUMA_LOCAL);
 378         else
 379                 __inc_zone_state(z, NUMA_OTHER);
 380 }
 381 #endif
 382
 383 #ifdef CONFIG_COMPACTION
 384 struct contig_page_info {
 385         unsigned long free_pages;
 386         unsigned long free_blocks_total;
 387         unsigned long free_blocks_suitable;
 388 };
 389
 390 /*
 391  * Calculate the number of free pages in a zone, how many contiguous
 392  * pages are free and how many are large enough to satisfy an allocation of
 393  * the target size. Note that this function makes no attempt to estimate
 394  * how many suitable free blocks there *might* be if MOVABLE pages were
 395  * migrated. Calculating that is possible, but expensive and can be
 396  * figured out from userspace
 397  */
 398 static void fill_contig_page_info(struct zone *zone,
 399                                 unsigned int suitable_order,
 400                                 struct contig_page_info *info)
 401 {
 402         unsigned int order;
 403
 404         info->free_pages = 0;
 405         info->free_blocks_total = 0;
 406         info->free_blocks_suitable = 0;
 407
 408         for (order = 0; order < MAX_ORDER; order++) {
 409                 unsigned long blocks;
 410
 411                 /* Count number of free blocks */
 412                 blocks = zone->free_area[order].nr_free;
 413                 info->free_blocks_total += blocks;
 414
 415                 /* Count free base pages */
 416                 info->free_pages += blocks << order;
 417
 418                 /* Count the suitable free blocks */
 419                 if (order >= suitable_order)
 420                         info->free_blocks_suitable += blocks <<
 421                                                 (order - suitable_order);
 422         }
 423 }
 424
 425 /*
 426  * A fragmentation index only makes sense if an allocation of a requested
 427  * size would fail. If that is true, the fragmentation index indicates
 428  * whether external fragmentation or a lack of memory was the problem.
 429  * The value can be used to determine if page reclaim or compaction
 430  * should be used
 431  */
 432 int fragmentation_index(unsigned int order, struct contig_page_info *info)
 433 {
 434         unsigned long requested = 1UL << order;
 435
 436         if (!info->free_blocks_total)
 437                 return 0;
 438
 439         /* Fragmentation index only makes sense when a request would fail */
 440         if (info->free_blocks_suitable)
 441                 return -1000;
 442
 443         /*
 444          * Index is between 0 and 1 so return within 3 decimal places
 445          *
 446          * 0 => allocation would fail due to lack of memory
 447          * 1 => allocation would fail due to fragmentation
 448          */
 449         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
 450 }
 451 #endif
 452
 453 #if defined(CONFIG_PROC_FS) || defined(CONFIG_COMPACTION)
 454 #include <linux/proc_fs.h>
 455 #include <linux/seq_file.h>
 456
 457 static char * const migratetype_names[MIGRATE_TYPES] = {
 458         "Unmovable",
 459         "Reclaimable",
 460         "Movable",
 461         "Reserve",
 462         "Isolate",
 463 };
 464
 465 static void *frag_start(struct seq_file *m, loff_t *pos)
 466 {
 467         pg_data_t *pgdat;
 468         loff_t node = *pos;
 469         for (pgdat = first_online_pgdat();
 470              pgdat && node;
 471              pgdat = next_online_pgdat(pgdat))
 472                 --node;
 473
 474         return pgdat;
 475 }
 476
 477 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
 478 {
 479         pg_data_t *pgdat = (pg_data_t *)arg;
 480
 481         (*pos)++;
 482         return next_online_pgdat(pgdat);
 483 }
 484
 485 static void frag_stop(struct seq_file *m, void *arg)
 486 {
 487 }
 488
 489 /* Walk all the zones in a node and print using a callback */
 490 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
 491                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
 492 {
 493         struct zone *zone;
 494         struct zone *node_zones = pgdat->node_zones;
 495         unsigned long flags;
 496
 497         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 498                 if (!populated_zone(zone))
 499                         continue;
 500
 501                 spin_lock_irqsave(&zone->lock, flags);
 502                 print(m, pgdat, zone);
 503                 spin_unlock_irqrestore(&zone->lock, flags);
 504         }
 505 }
 506 #endif
 507
 508 #ifdef CONFIG_PROC_FS
 509 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
 510                                                 struct zone *zone)
 511 {
 512         int order;
 513
 514         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 515         for (order = 0; order < MAX_ORDER; ++order)
 516                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
 517         seq_putc(m, '\n');
 518 }
 519
 520 /*
 521  * This walks the free areas for each zone.
 522  */
 523 static int frag_show(struct seq_file *m, void *arg)
 524 {
 525         pg_data_t *pgdat = (pg_data_t *)arg;
 526         walk_zones_in_node(m, pgdat, frag_show_print);
 527         return 0;
 528 }
 529
 530 static void pagetypeinfo_showfree_print(struct seq_file *m,
 531                                         pg_data_t *pgdat, struct zone *zone)
 532 {
 533         int order, mtype;
 534
 535         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
 536                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
 537                                         pgdat->node_id,
 538                                         zone->name,
 539                                         migratetype_names[mtype]);
 540                 for (order = 0; order < MAX_ORDER; ++order) {
 541                         unsigned long freecount = 0;
 542                         struct free_area *area;
 543                         struct list_head *curr;
 544
 545                         area = &(zone->free_area[order]);
 546
 547                         list_for_each(curr, &area->free_list[mtype])
 548                                 freecount++;
 549                         seq_printf(m, "%6lu ", freecount);
 550                 }
 551                 seq_putc(m, '\n');
 552         }
 553 }
 554
 555 /* Print out the free pages at each order for each migatetype */
 556 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
 557 {
 558         int order;
 559         pg_data_t *pgdat = (pg_data_t *)arg;
 560
 561         /* Print header */
 562         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
 563         for (order = 0; order < MAX_ORDER; ++order)
 564                 seq_printf(m, "%6d ", order);
 565         seq_putc(m, '\n');
 566
 567         walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
 568
 569         return 0;
 570 }
 571
 572 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
 573                                         pg_data_t *pgdat, struct zone *zone)
 574 {
 575         int mtype;
 576         unsigned long pfn;
 577         unsigned long start_pfn = zone->zone_start_pfn;
 578         unsigned long end_pfn = start_pfn + zone->spanned_pages;
 579         unsigned long count[MIGRATE_TYPES] = { 0, };
 580
 581         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
 582                 struct page *page;
 583
 584                 if (!pfn_valid(pfn))
 585                         continue;
 586
 587                 page = pfn_to_page(pfn);
 588
 589                 /* Watch for unexpected holes punched in the memmap */
 590                 if (!memmap_valid_within(pfn, page, zone))
 591                         continue;
 592
 593                 mtype = get_pageblock_migratetype(page);
 594
 595                 if (mtype < MIGRATE_TYPES)
 596                         count[mtype]++;
 597         }
 598
 599         /* Print counts */
 600         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 601         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
 602                 seq_printf(m, "%12lu ", count[mtype]);
 603         seq_putc(m, '\n');
 604 }
 605
 606 /* Print out the free pages at each order for each migratetype */
 607 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
 608 {
 609         int mtype;
 610         pg_data_t *pgdat = (pg_data_t *)arg;
 611
 612         seq_printf(m, "\n%-23s", "Number of blocks type ");
 613         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
 614                 seq_printf(m, "%12s ", migratetype_names[mtype]);
 615         seq_putc(m, '\n');
 616         walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
 617
 618         return 0;
 619 }
 620
 621 /*
 622  * This prints out statistics in relation to grouping pages by mobility.
 623  * It is expensive to collect so do not constantly read the file.
 624  */
 625 static int pagetypeinfo_show(struct seq_file *m, void *arg)
 626 {
 627         pg_data_t *pgdat = (pg_data_t *)arg;
 628
 629         /* check memoryless node */
 630         if (!node_state(pgdat->node_id, N_HIGH_MEMORY))
 631                 return 0;
 632
 633         seq_printf(m, "Page block order: %d\n", pageblock_order);
 634         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
 635         seq_putc(m, '\n');
 636         pagetypeinfo_showfree(m, pgdat);
 637         pagetypeinfo_showblockcount(m, pgdat);
 638
 639         return 0;
 640 }
 641
 642 static const struct seq_operations fragmentation_op = {
 643         .start  = frag_start,
 644         .next   = frag_next,
 645         .stop   = frag_stop,
 646         .show   = frag_show,
 647 };
 648
 649 static int fragmentation_open(struct inode *inode, struct file *file)
 650 {
 651         return seq_open(file, &fragmentation_op);
 652 }
 653
 654 static const struct file_operations fragmentation_file_operations = {
 655         .open           = fragmentation_open,
 656         .read           = seq_read,
 657         .llseek         = seq_lseek,
 658         .release        = seq_release,
 659 };
 660
 661 static const struct seq_operations pagetypeinfo_op = {
 662         .start  = frag_start,
 663         .next   = frag_next,
 664         .stop   = frag_stop,
 665         .show   = pagetypeinfo_show,
 666 };
 667
 668 static int pagetypeinfo_open(struct inode *inode, struct file *file)
 669 {
 670         return seq_open(file, &pagetypeinfo_op);
 671 }
 672
 673 static const struct file_operations pagetypeinfo_file_ops = {
 674         .open           = pagetypeinfo_open,
 675         .read           = seq_read,
 676         .llseek         = seq_lseek,
 677         .release        = seq_release,
 678 };
 679
 680 #ifdef CONFIG_ZONE_DMA
 681 #define TEXT_FOR_DMA(xx) xx "_dma",
 682 #else
 683 #define TEXT_FOR_DMA(xx)
 684 #endif
 685
 686 #ifdef CONFIG_ZONE_DMA32
 687 #define TEXT_FOR_DMA32(xx) xx "_dma32",
 688 #else
 689 #define TEXT_FOR_DMA32(xx)
 690 #endif
 691
 692 #ifdef CONFIG_HIGHMEM
 693 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
 694 #else
 695 #define TEXT_FOR_HIGHMEM(xx)
 696 #endif
 697
 698 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
 699                                         TEXT_FOR_HIGHMEM(xx) xx "_movable",
 700
 701 static const char * const vmstat_text[] = {
 702         /* Zoned VM counters */
 703         "nr_free_pages",
 704         "nr_inactive_anon",
 705         "nr_active_anon",
 706         "nr_inactive_file",
 707         "nr_active_file",
 708         "nr_unevictable",
 709         "nr_mlock",
 710         "nr_anon_pages",
 711         "nr_mapped",
 712         "nr_file_pages",
 713         "nr_dirty",
 714         "nr_writeback",
 715         "nr_slab_reclaimable",
 716         "nr_slab_unreclaimable",
 717         "nr_page_table_pages",
 718         "nr_kernel_stack",
 719         "nr_unstable",
 720         "nr_bounce",
 721         "nr_vmscan_write",
 722         "nr_writeback_temp",
 723         "nr_isolated_anon",
 724         "nr_isolated_file",
 725         "nr_shmem",
 726 #ifdef CONFIG_NUMA
 727         "numa_hit",
 728         "numa_miss",
 729         "numa_foreign",
 730         "numa_interleave",
 731         "numa_local",
 732         "numa_other",
 733 #endif
 734
 735 #ifdef CONFIG_VM_EVENT_COUNTERS
 736         "pgpgin",
 737         "pgpgout",
 738         "pswpin",
 739         "pswpout",
 740
 741         TEXTS_FOR_ZONES("pgalloc")
 742
 743         "pgfree",
 744         "pgactivate",
 745         "pgdeactivate",
 746
 747         "pgfault",
 748         "pgmajfault",
 749
 750         TEXTS_FOR_ZONES("pgrefill")
 751         TEXTS_FOR_ZONES("pgsteal")
 752         TEXTS_FOR_ZONES("pgscan_kswapd")
 753         TEXTS_FOR_ZONES("pgscan_direct")
 754
 755 #ifdef CONFIG_NUMA
 756         "zone_reclaim_failed",
 757 #endif
 758         "pginodesteal",
 759         "slabs_scanned",
 760         "kswapd_steal",
 761         "kswapd_inodesteal",
 762         "kswapd_low_wmark_hit_quickly",
 763         "kswapd_high_wmark_hit_quickly",
 764         "kswapd_skip_congestion_wait",
 765         "pageoutrun",
 766         "allocstall",
 767
 768         "pgrotated",
 769
 770 #ifdef CONFIG_COMPACTION
 771         "compact_blocks_moved",
 772         "compact_pages_moved",
 773         "compact_pagemigrate_failed",
 774 #endif
 775
 776 #ifdef CONFIG_HUGETLB_PAGE
 777         "htlb_buddy_alloc_success",
 778         "htlb_buddy_alloc_fail",
 779 #endif
 780         "unevictable_pgs_culled",
 781         "unevictable_pgs_scanned",
 782         "unevictable_pgs_rescued",
 783         "unevictable_pgs_mlocked",
 784         "unevictable_pgs_munlocked",
 785         "unevictable_pgs_cleared",
 786         "unevictable_pgs_stranded",
 787         "unevictable_pgs_mlockfreed",
 788 #endif
 789 };
 790
 791 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
 792                                                         struct zone *zone)
 793 {
 794         int i;
 795         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
 796         seq_printf(m,
 797                    "\n  pages free     %lu"
 798                    "\n        min      %lu"
 799                    "\n        low      %lu"
 800                    "\n        high     %lu"
 801                    "\n        scanned  %lu"
 802                    "\n        spanned  %lu"
 803                    "\n        present  %lu",
 804                    zone_page_state(zone, NR_FREE_PAGES),
 805                    min_wmark_pages(zone),
 806                    low_wmark_pages(zone),
 807                    high_wmark_pages(zone),
 808                    zone->pages_scanned,
 809                    zone->spanned_pages,
 810                    zone->present_pages);
 811
 812         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 813                 seq_printf(m, "\n    %-12s %lu", vmstat_text[i],
 814                                 zone_page_state(zone, i));
 815
 816         seq_printf(m,
 817                    "\n        protection: (%lu",
 818                    zone->lowmem_reserve[0]);
 819         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
 820                 seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
 821         seq_printf(m,
 822                    ")"
 823                    "\n  pagesets");
 824         for_each_online_cpu(i) {
 825                 struct per_cpu_pageset *pageset;
 826
 827                 pageset = per_cpu_ptr(zone->pageset, i);
 828                 seq_printf(m,
 829                            "\n    cpu: %i"
 830                            "\n              count: %i"
 831                            "\n              high:  %i"
 832                            "\n              batch: %i",
 833                            i,
 834                            pageset->pcp.count,
 835                            pageset->pcp.high,
 836                            pageset->pcp.batch);
 837 #ifdef CONFIG_SMP
 838                 seq_printf(m, "\n  vm stats threshold: %d",
 839                                 pageset->stat_threshold);
 840 #endif
 841         }
 842         seq_printf(m,
 843                    "\n  all_unreclaimable: %u"
 844                    "\n  prev_priority:     %i"
 845                    "\n  start_pfn:         %lu"
 846                    "\n  inactive_ratio:    %u",
 847                    zone->all_unreclaimable,
 848                    zone->prev_priority,
 849                    zone->zone_start_pfn,
 850                    zone->inactive_ratio);
 851         seq_putc(m, '\n');
 852 }
 853
 854 /*
 855  * Output information about zones in @pgdat.
 856  */
 857 static int zoneinfo_show(struct seq_file *m, void *arg)
 858 {
 859         pg_data_t *pgdat = (pg_data_t *)arg;
 860         walk_zones_in_node(m, pgdat, zoneinfo_show_print);
 861         return 0;
 862 }
 863
 864 static const struct seq_operations zoneinfo_op = {
 865         .start  = frag_start, /* iterate over all zones. The same as in
 866                                * fragmentation. */
 867         .next   = frag_next,
 868         .stop   = frag_stop,
 869         .show   = zoneinfo_show,
 870 };
 871
 872 static int zoneinfo_open(struct inode *inode, struct file *file)
 873 {
 874         return seq_open(file, &zoneinfo_op);
 875 }
 876
 877 static const struct file_operations proc_zoneinfo_file_operations = {
 878         .open           = zoneinfo_open,
 879         .read           = seq_read,
 880         .llseek         = seq_lseek,
 881         .release        = seq_release,
 882 };
 883
 884 static void *vmstat_start(struct seq_file *m, loff_t *pos)
 885 {
 886         unsigned long *v;
 887 #ifdef CONFIG_VM_EVENT_COUNTERS
 888         unsigned long *e;
 889 #endif
 890         int i;
 891
 892         if (*pos >= ARRAY_SIZE(vmstat_text))
 893                 return NULL;
 894
 895 #ifdef CONFIG_VM_EVENT_COUNTERS
 896         v = kmalloc(NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long)
 897                         + sizeof(struct vm_event_state), GFP_KERNEL);
 898 #else
 899         v = kmalloc(NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long),
 900                         GFP_KERNEL);
 901 #endif
 902         m->private = v;
 903         if (!v)
 904                 return ERR_PTR(-ENOMEM);
 905         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 906                 v[i] = global_page_state(i);
 907 #ifdef CONFIG_VM_EVENT_COUNTERS
 908         e = v + NR_VM_ZONE_STAT_ITEMS;
 909         all_vm_events(e);
 910         e[PGPGIN] /= 2;         /* sectors -> kbytes */
 911         e[PGPGOUT] /= 2;
 912 #endif
 913         return v + *pos;
 914 }
 915
 916 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
 917 {
 918         (*pos)++;
 919         if (*pos >= ARRAY_SIZE(vmstat_text))
 920                 return NULL;
 921         return (unsigned long *)m->private + *pos;
 922 }
 923
 924 static int vmstat_show(struct seq_file *m, void *arg)
 925 {
 926         unsigned long *l = arg;
 927         unsigned long off = l - (unsigned long *)m->private;
 928
 929         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
 930         return 0;
 931 }
 932
 933 static void vmstat_stop(struct seq_file *m, void *arg)
 934 {
 935         kfree(m->private);
 936         m->private = NULL;
 937 }
 938
 939 static const struct seq_operations vmstat_op = {
 940         .start  = vmstat_start,
 941         .next   = vmstat_next,
 942         .stop   = vmstat_stop,
 943         .show   = vmstat_show,
 944 };
 945
 946 static int vmstat_open(struct inode *inode, struct file *file)
 947 {
 948         return seq_open(file, &vmstat_op);
 949 }
 950
 951 static const struct file_operations proc_vmstat_file_operations = {
 952         .open           = vmstat_open,
 953         .read           = seq_read,
 954         .llseek         = seq_lseek,
 955         .release        = seq_release,
 956 };
 957 #endif /* CONFIG_PROC_FS */
 958
 959 #ifdef CONFIG_SMP
 960 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
 961 int sysctl_stat_interval __read_mostly = HZ;
 962
 963 static void vmstat_update(struct work_struct *w)
 964 {
 965         refresh_cpu_vm_stats(smp_processor_id());
 966         schedule_delayed_work(&__get_cpu_var(vmstat_work),
 967                 round_jiffies_relative(sysctl_stat_interval));
 968 }
 969
 970 static void __cpuinit start_cpu_timer(int cpu)
 971 {
 972         struct delayed_work *work = &per_cpu(vmstat_work, cpu);
 973
 974         INIT_DELAYED_WORK_DEFERRABLE(work, vmstat_update);
 975         schedule_delayed_work_on(cpu, work, __round_jiffies_relative(HZ, cpu));
 976 }
 977
 978 /*
 979  * Use the cpu notifier to insure that the thresholds are recalculated
 980  * when necessary.
 981  */
 982 static int __cpuinit vmstat_cpuup_callback(struct notifier_block *nfb,
 983                 unsigned long action,
 984                 void *hcpu)
 985 {
 986         long cpu = (long)hcpu;
 987
 988         switch (action) {
 989         case CPU_ONLINE:
 990         case CPU_ONLINE_FROZEN:
 991                 start_cpu_timer(cpu);
 992                 node_set_state(cpu_to_node(cpu), N_CPU);
 993                 break;
 994         case CPU_DOWN_PREPARE:
 995         case CPU_DOWN_PREPARE_FROZEN:
 996                 cancel_rearming_delayed_work(&per_cpu(vmstat_work, cpu));
 997                 per_cpu(vmstat_work, cpu).work.func = NULL;
 998                 break;
 999         case CPU_DOWN_FAILED:
1000         case CPU_DOWN_FAILED_FROZEN:
1001                 start_cpu_timer(cpu);
1002                 break;
1003         case CPU_DEAD:
1004         case CPU_DEAD_FROZEN:
1005                 refresh_zone_stat_thresholds();
1006                 break;
1007         default:
1008                 break;
1009         }
1010         return NOTIFY_OK;
1011 }
1012
1013 static struct notifier_block __cpuinitdata vmstat_notifier =
1014         { &vmstat_cpuup_callback, NULL, 0 };
1015 #endif
1016
1017 static int __init setup_vmstat(void)
1018 {
1019 #ifdef CONFIG_SMP
1020         int cpu;
1021
1022         refresh_zone_stat_thresholds();
1023         register_cpu_notifier(&vmstat_notifier);
1024
1025         for_each_online_cpu(cpu)
1026                 start_cpu_timer(cpu);
1027 #endif
1028 #ifdef CONFIG_PROC_FS
1029         proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1030         proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1031         proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1032         proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1033 #endif
1034         return 0;
1035 }
1036 module_init(setup_vmstat)
1037
1038 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1039 #include <linux/debugfs.h>
1040
1041 static struct dentry *extfrag_debug_root;
1042
1043 /*
1044  * Return an index indicating how much of the available free memory is
1045  * unusable for an allocation of the requested size.
1046  */
1047 static int unusable_free_index(unsigned int order,
1048                                 struct contig_page_info *info)
1049 {
1050         /* No free memory is interpreted as all free memory is unusable */
1051         if (info->free_pages == 0)
1052                 return 1000;
1053
1054         /*
1055          * Index should be a value between 0 and 1. Return a value to 3
1056          * decimal places.
1057          *
1058          * 0 => no fragmentation
1059          * 1 => high fragmentation
1060          */
1061         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1062
1063 }
1064
1065 static void unusable_show_print(struct seq_file *m,
1066                                         pg_data_t *pgdat, struct zone *zone)
1067 {
1068         unsigned int order;
1069         int index;
1070         struct contig_page_info info;
1071
1072         seq_printf(m, "Node %d, zone %8s ",
1073                                 pgdat->node_id,
1074                                 zone->name);
1075         for (order = 0; order < MAX_ORDER; ++order) {
1076                 fill_contig_page_info(zone, order, &info);
1077                 index = unusable_free_index(order, &info);
1078                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1079         }
1080
1081         seq_putc(m, '\n');
1082 }
1083
1084 /*
1085  * Display unusable free space index
1086  *
1087  * The unusable free space index measures how much of the available free
1088  * memory cannot be used to satisfy an allocation of a given size and is a
1089  * value between 0 and 1. The higher the value, the more of free memory is
1090  * unusable and by implication, the worse the external fragmentation is. This
1091  * can be expressed as a percentage by multiplying by 100.
1092  */
1093 static int unusable_show(struct seq_file *m, void *arg)
1094 {
1095         pg_data_t *pgdat = (pg_data_t *)arg;
1096
1097         /* check memoryless node */
1098         if (!node_state(pgdat->node_id, N_HIGH_MEMORY))
1099                 return 0;
1100
1101         walk_zones_in_node(m, pgdat, unusable_show_print);
1102
1103         return 0;
1104 }
1105
1106 static const struct seq_operations unusable_op = {
1107         .start  = frag_start,
1108         .next   = frag_next,
1109         .stop   = frag_stop,
1110         .show   = unusable_show,
1111 };
1112
1113 static int unusable_open(struct inode *inode, struct file *file)
1114 {
1115         return seq_open(file, &unusable_op);
1116 }
1117
1118 static const struct file_operations unusable_file_ops = {
1119         .open           = unusable_open,
1120         .read           = seq_read,
1121         .llseek         = seq_lseek,
1122         .release        = seq_release,
1123 };
1124
1125 static void extfrag_show_print(struct seq_file *m,
1126                                         pg_data_t *pgdat, struct zone *zone)
1127 {
1128         unsigned int order;
1129         int index;
1130
1131         /* Alloc on stack as interrupts are disabled for zone walk */
1132         struct contig_page_info info;
1133
1134         seq_printf(m, "Node %d, zone %8s ",
1135                                 pgdat->node_id,
1136                                 zone->name);
1137         for (order = 0; order < MAX_ORDER; ++order) {
1138                 fill_contig_page_info(zone, order, &info);
1139                 index = fragmentation_index(order, &info);
1140                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1141         }
1142
1143         seq_putc(m, '\n');
1144 }
1145
1146 /*
1147  * Display fragmentation index for orders that allocations would fail for
1148  */
1149 static int extfrag_show(struct seq_file *m, void *arg)
1150 {
1151         pg_data_t *pgdat = (pg_data_t *)arg;
1152
1153         walk_zones_in_node(m, pgdat, extfrag_show_print);
1154
1155         return 0;
1156 }
1157
1158 static const struct seq_operations extfrag_op = {
1159         .start  = frag_start,
1160         .next   = frag_next,
1161         .stop   = frag_stop,
1162         .show   = extfrag_show,
1163 };
1164
1165 static int extfrag_open(struct inode *inode, struct file *file)
1166 {
1167         return seq_open(file, &extfrag_op);
1168 }
1169
1170 static const struct file_operations extfrag_file_ops = {
1171         .open           = extfrag_open,
1172         .read           = seq_read,
1173         .llseek         = seq_lseek,
1174         .release        = seq_release,
1175 };
1176
1177 static int __init extfrag_debug_init(void)
1178 {
1179         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1180         if (!extfrag_debug_root)
1181                 return -ENOMEM;
1182
1183         if (!debugfs_create_file("unusable_index", 0444,
1184                         extfrag_debug_root, NULL, &unusable_file_ops))
1185                 return -ENOMEM;
1186
1187         if (!debugfs_create_file("extfrag_index", 0444,
1188                         extfrag_debug_root, NULL, &extfrag_file_ops))
1189                 return -ENOMEM;
1190
1191         return 0;
1192 }
1193
1194 module_init(extfrag_debug_init);
1195 #endif