kernel/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  *  Copyright (C) 2008-2014 Christoph Lameter
  11  */
  12 #include <linux/fs.h>
  13 #include <linux/mm.h>
  14 #include <linux/err.h>
  15 #include <linux/module.h>
  16 #include <linux/slab.h>
  17 #include <linux/cpu.h>
  18 #include <linux/cpumask.h>
  19 #include <linux/vmstat.h>
  20 #include <linux/proc_fs.h>
  21 #include <linux/seq_file.h>
  22 #include <linux/debugfs.h>
  23 #include <linux/sched.h>
  24 #include <linux/math64.h>
  25 #include <linux/writeback.h>
  26 #include <linux/compaction.h>
  27 #include <linux/mm_inline.h>
  28 #include <linux/page_ext.h>
  29 #include <linux/page_owner.h>
  30
  31 #include "internal.h"
  32
  33 #ifdef CONFIG_VM_EVENT_COUNTERS
  34 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
  35 EXPORT_PER_CPU_SYMBOL(vm_event_states);
  36
  37 static void sum_vm_events(unsigned long *ret)
  38 {
  39         int cpu;
  40         int i;
  41
  42         memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
  43
  44         for_each_online_cpu(cpu) {
  45                 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
  46
  47                 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
  48                         ret[i] += this->event[i];
  49         }
  50 }
  51
  52 /*
  53  * Accumulate the vm event counters across all CPUs.
  54  * The result is unavoidably approximate - it can change
  55  * during and after execution of this function.
  56 */
  57 void all_vm_events(unsigned long *ret)
  58 {
  59         get_online_cpus();
  60         sum_vm_events(ret);
  61         put_online_cpus();
  62 }
  63 EXPORT_SYMBOL_GPL(all_vm_events);
  64
  65 /*
  66  * Fold the foreign cpu events into our own.
  67  *
  68  * This is adding to the events on one processor
  69  * but keeps the global counts constant.
  70  */
  71 void vm_events_fold_cpu(int cpu)
  72 {
  73         struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
  74         int i;
  75
  76         for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
  77                 count_vm_events(i, fold_state->event[i]);
  78                 fold_state->event[i] = 0;
  79         }
  80 }
  81
  82 #endif /* CONFIG_VM_EVENT_COUNTERS */
  83
  84 /*
  85  * Manage combined zone based / global counters
  86  *
  87  * vm_stat contains the global counters
  88  */
  89 atomic_long_t vm_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
  90 EXPORT_SYMBOL(vm_stat);
  91
  92 #ifdef CONFIG_SMP
  93
  94 int calculate_pressure_threshold(struct zone *zone)
  95 {
  96         int threshold;
  97         int watermark_distance;
  98
  99         /*
 100          * As vmstats are not up to date, there is drift between the estimated
 101          * and real values. For high thresholds and a high number of CPUs, it
 102          * is possible for the min watermark to be breached while the estimated
 103          * value looks fine. The pressure threshold is a reduced value such
 104          * that even the maximum amount of drift will not accidentally breach
 105          * the min watermark
 106          */
 107         watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
 108         threshold = max(1, (int)(watermark_distance / num_online_cpus()));
 109
 110         /*
 111          * Maximum threshold is 125
 112          */
 113         threshold = min(125, threshold);
 114
 115         return threshold;
 116 }
 117
 118 int calculate_normal_threshold(struct zone *zone)
 119 {
 120         int threshold;
 121         int mem;        /* memory in 128 MB units */
 122
 123         /*
 124          * The threshold scales with the number of processors and the amount
 125          * of memory per zone. More memory means that we can defer updates for
 126          * longer, more processors could lead to more contention.
 127          * fls() is used to have a cheap way of logarithmic scaling.
 128          *
 129          * Some sample thresholds:
 130          *
 131          * Threshold    Processors      (fls)   Zonesize        fls(mem+1)
 132          * ------------------------------------------------------------------
 133          * 8            1               1       0.9-1 GB        4
 134          * 16           2               2       0.9-1 GB        4
 135          * 20           2               2       1-2 GB          5
 136          * 24           2               2       2-4 GB          6
 137          * 28           2               2       4-8 GB          7
 138          * 32           2               2       8-16 GB         8
 139          * 4            2               2       <128M           1
 140          * 30           4               3       2-4 GB          5
 141          * 48           4               3       8-16 GB         8
 142          * 32           8               4       1-2 GB          4
 143          * 32           8               4       0.9-1GB         4
 144          * 10           16              5       <128M           1
 145          * 40           16              5       900M            4
 146          * 70           64              7       2-4 GB          5
 147          * 84           64              7       4-8 GB          6
 148          * 108          512             9       4-8 GB          6
 149          * 125          1024            10      8-16 GB         8
 150          * 125          1024            10      16-32 GB        9
 151          */
 152
 153         mem = zone->managed_pages >> (27 - PAGE_SHIFT);
 154
 155         threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
 156
 157         /*
 158          * Maximum threshold is 125
 159          */
 160         threshold = min(125, threshold);
 161
 162         return threshold;
 163 }
 164
 165 /*
 166  * Refresh the thresholds for each zone.
 167  */
 168 void refresh_zone_stat_thresholds(void)
 169 {
 170         struct zone *zone;
 171         int cpu;
 172         int threshold;
 173
 174         for_each_populated_zone(zone) {
 175                 unsigned long max_drift, tolerate_drift;
 176
 177                 threshold = calculate_normal_threshold(zone);
 178
 179                 for_each_online_cpu(cpu)
 180                         per_cpu_ptr(zone->pageset, cpu)->stat_threshold
 181                                                         = threshold;
 182
 183                 /*
 184                  * Only set percpu_drift_mark if there is a danger that
 185                  * NR_FREE_PAGES reports the low watermark is ok when in fact
 186                  * the min watermark could be breached by an allocation
 187                  */
 188                 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
 189                 max_drift = num_online_cpus() * threshold;
 190                 if (max_drift > tolerate_drift)
 191                         zone->percpu_drift_mark = high_wmark_pages(zone) +
 192                                         max_drift;
 193         }
 194 }
 195
 196 void set_pgdat_percpu_threshold(pg_data_t *pgdat,
 197                                 int (*calculate_pressure)(struct zone *))
 198 {
 199         struct zone *zone;
 200         int cpu;
 201         int threshold;
 202         int i;
 203
 204         for (i = 0; i < pgdat->nr_zones; i++) {
 205                 zone = &pgdat->node_zones[i];
 206                 if (!zone->percpu_drift_mark)
 207                         continue;
 208
 209                 threshold = (*calculate_pressure)(zone);
 210                 for_each_online_cpu(cpu)
 211                         per_cpu_ptr(zone->pageset, cpu)->stat_threshold
 212                                                         = threshold;
 213         }
 214 }
 215
 216 /*
 217  * For use when we know that interrupts are disabled,
 218  * or when we know that preemption is disabled and that
 219  * particular counter cannot be updated from interrupt context.
 220  */
 221 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 222                                 int delta)
 223 {
 224         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 225         s8 __percpu *p = pcp->vm_stat_diff + item;
 226         long x;
 227         long t;
 228
 229         preempt_disable_rt();
 230         x = delta + __this_cpu_read(*p);
 231
 232         t = __this_cpu_read(pcp->stat_threshold);
 233
 234         if (unlikely(x > t || x < -t)) {
 235                 zone_page_state_add(x, zone, item);
 236                 x = 0;
 237         }
 238         __this_cpu_write(*p, x);
 239         preempt_enable_rt();
 240 }
 241 EXPORT_SYMBOL(__mod_zone_page_state);
 242
 243 /*
 244  * Optimized increment and decrement functions.
 245  *
 246  * These are only for a single page and therefore can take a struct page *
 247  * argument instead of struct zone *. This allows the inclusion of the code
 248  * generated for page_zone(page) into the optimized functions.
 249  *
 250  * No overflow check is necessary and therefore the differential can be
 251  * incremented or decremented in place which may allow the compilers to
 252  * generate better code.
 253  * The increment or decrement is known and therefore one boundary check can
 254  * be omitted.
 255  *
 256  * NOTE: These functions are very performance sensitive. Change only
 257  * with care.
 258  *
 259  * Some processors have inc/dec instructions that are atomic vs an interrupt.
 260  * However, the code must first determine the differential location in a zone
 261  * based on the processor number and then inc/dec the counter. There is no
 262  * guarantee without disabling preemption that the processor will not change
 263  * in between and therefore the atomicity vs. interrupt cannot be exploited
 264  * in a useful way here.
 265  */
 266 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
 267 {
 268         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 269         s8 __percpu *p = pcp->vm_stat_diff + item;
 270         s8 v, t;
 271
 272         preempt_disable_rt();
 273         v = __this_cpu_inc_return(*p);
 274         t = __this_cpu_read(pcp->stat_threshold);
 275         if (unlikely(v > t)) {
 276                 s8 overstep = t >> 1;
 277
 278                 zone_page_state_add(v + overstep, zone, item);
 279                 __this_cpu_write(*p, -overstep);
 280         }
 281         preempt_enable_rt();
 282 }
 283
 284 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
 285 {
 286         __inc_zone_state(page_zone(page), item);
 287 }
 288 EXPORT_SYMBOL(__inc_zone_page_state);
 289
 290 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
 291 {
 292         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 293         s8 __percpu *p = pcp->vm_stat_diff + item;
 294         s8 v, t;
 295
 296         preempt_disable_rt();
 297         v = __this_cpu_dec_return(*p);
 298         t = __this_cpu_read(pcp->stat_threshold);
 299         if (unlikely(v < - t)) {
 300                 s8 overstep = t >> 1;
 301
 302                 zone_page_state_add(v - overstep, zone, item);
 303                 __this_cpu_write(*p, overstep);
 304         }
 305         preempt_enable_rt();
 306 }
 307
 308 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
 309 {
 310         __dec_zone_state(page_zone(page), item);
 311 }
 312 EXPORT_SYMBOL(__dec_zone_page_state);
 313
 314 #ifdef CONFIG_HAVE_CMPXCHG_LOCAL
 315 /*
 316  * If we have cmpxchg_local support then we do not need to incur the overhead
 317  * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
 318  *
 319  * mod_state() modifies the zone counter state through atomic per cpu
 320  * operations.
 321  *
 322  * Overstep mode specifies how overstep should handled:
 323  *     0       No overstepping
 324  *     1       Overstepping half of threshold
 325  *     -1      Overstepping minus half of threshold
 326 */
 327 static inline void mod_state(struct zone *zone,
 328        enum zone_stat_item item, int delta, int overstep_mode)
 329 {
 330         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 331         s8 __percpu *p = pcp->vm_stat_diff + item;
 332         long o, n, t, z;
 333
 334         do {
 335                 z = 0;  /* overflow to zone counters */
 336
 337                 /*
 338                  * The fetching of the stat_threshold is racy. We may apply
 339                  * a counter threshold to the wrong the cpu if we get
 340                  * rescheduled while executing here. However, the next
 341                  * counter update will apply the threshold again and
 342                  * therefore bring the counter under the threshold again.
 343                  *
 344                  * Most of the time the thresholds are the same anyways
 345                  * for all cpus in a zone.
 346                  */
 347                 t = this_cpu_read(pcp->stat_threshold);
 348
 349                 o = this_cpu_read(*p);
 350                 n = delta + o;
 351
 352                 if (n > t || n < -t) {
 353                         int os = overstep_mode * (t >> 1) ;
 354
 355                         /* Overflow must be added to zone counters */
 356                         z = n + os;
 357                         n = -os;
 358                 }
 359         } while (this_cpu_cmpxchg(*p, o, n) != o);
 360
 361         if (z)
 362                 zone_page_state_add(z, zone, item);
 363 }
 364
 365 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 366                                         int delta)
 367 {
 368         mod_state(zone, item, delta, 0);
 369 }
 370 EXPORT_SYMBOL(mod_zone_page_state);
 371
 372 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
 373 {
 374         mod_state(zone, item, 1, 1);
 375 }
 376
 377 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 378 {
 379         mod_state(page_zone(page), item, 1, 1);
 380 }
 381 EXPORT_SYMBOL(inc_zone_page_state);
 382
 383 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 384 {
 385         mod_state(page_zone(page), item, -1, -1);
 386 }
 387 EXPORT_SYMBOL(dec_zone_page_state);
 388 #else
 389 /*
 390  * Use interrupt disable to serialize counter updates
 391  */
 392 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 393                                         int delta)
 394 {
 395         unsigned long flags;
 396
 397         local_irq_save(flags);
 398         __mod_zone_page_state(zone, item, delta);
 399         local_irq_restore(flags);
 400 }
 401 EXPORT_SYMBOL(mod_zone_page_state);
 402
 403 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
 404 {
 405         unsigned long flags;
 406
 407         local_irq_save(flags);
 408         __inc_zone_state(zone, item);
 409         local_irq_restore(flags);
 410 }
 411
 412 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 413 {
 414         unsigned long flags;
 415         struct zone *zone;
 416
 417         zone = page_zone(page);
 418         local_irq_save(flags);
 419         __inc_zone_state(zone, item);
 420         local_irq_restore(flags);
 421 }
 422 EXPORT_SYMBOL(inc_zone_page_state);
 423
 424 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 425 {
 426         unsigned long flags;
 427
 428         local_irq_save(flags);
 429         __dec_zone_page_state(page, item);
 430         local_irq_restore(flags);
 431 }
 432 EXPORT_SYMBOL(dec_zone_page_state);
 433 #endif
 434
 435
 436 /*
 437  * Fold a differential into the global counters.
 438  * Returns the number of counters updated.
 439  */
 440 static int fold_diff(int *diff)
 441 {
 442         int i;
 443         int changes = 0;
 444
 445         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 446                 if (diff[i]) {
 447                         atomic_long_add(diff[i], &vm_stat[i]);
 448                         changes++;
 449         }
 450         return changes;
 451 }
 452
 453 /*
 454  * Update the zone counters for the current cpu.
 455  *
 456  * Note that refresh_cpu_vm_stats strives to only access
 457  * node local memory. The per cpu pagesets on remote zones are placed
 458  * in the memory local to the processor using that pageset. So the
 459  * loop over all zones will access a series of cachelines local to
 460  * the processor.
 461  *
 462  * The call to zone_page_state_add updates the cachelines with the
 463  * statistics in the remote zone struct as well as the global cachelines
 464  * with the global counters. These could cause remote node cache line
 465  * bouncing and will have to be only done when necessary.
 466  *
 467  * The function returns the number of global counters updated.
 468  */
 469 static int refresh_cpu_vm_stats(void)
 470 {
 471         struct zone *zone;
 472         int i;
 473         int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 474         int changes = 0;
 475
 476         for_each_populated_zone(zone) {
 477                 struct per_cpu_pageset __percpu *p = zone->pageset;
 478
 479                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
 480                         int v;
 481
 482                         v = this_cpu_xchg(p->vm_stat_diff[i], 0);
 483                         if (v) {
 484
 485                                 atomic_long_add(v, &zone->vm_stat[i]);
 486                                 global_diff[i] += v;
 487 #ifdef CONFIG_NUMA
 488                                 /* 3 seconds idle till flush */
 489                                 __this_cpu_write(p->expire, 3);
 490 #endif
 491                         }
 492                 }
 493                 cond_resched();
 494 #ifdef CONFIG_NUMA
 495                 /*
 496                  * Deal with draining the remote pageset of this
 497                  * processor
 498                  *
 499                  * Check if there are pages remaining in this pageset
 500                  * if not then there is nothing to expire.
 501                  */
 502                 if (!__this_cpu_read(p->expire) ||
 503                                !__this_cpu_read(p->pcp.count))
 504                         continue;
 505
 506                 /*
 507                  * We never drain zones local to this processor.
 508                  */
 509                 if (zone_to_nid(zone) == numa_node_id()) {
 510                         __this_cpu_write(p->expire, 0);
 511                         continue;
 512                 }
 513
 514                 if (__this_cpu_dec_return(p->expire))
 515                         continue;
 516
 517                 if (__this_cpu_read(p->pcp.count)) {
 518                         drain_zone_pages(zone, this_cpu_ptr(&p->pcp));
 519                         changes++;
 520                 }
 521 #endif
 522         }
 523         changes += fold_diff(global_diff);
 524         return changes;
 525 }
 526
 527 /*
 528  * Fold the data for an offline cpu into the global array.
 529  * There cannot be any access by the offline cpu and therefore
 530  * synchronization is simplified.
 531  */
 532 void cpu_vm_stats_fold(int cpu)
 533 {
 534         struct zone *zone;
 535         int i;
 536         int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 537
 538         for_each_populated_zone(zone) {
 539                 struct per_cpu_pageset *p;
 540
 541                 p = per_cpu_ptr(zone->pageset, cpu);
 542
 543                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 544                         if (p->vm_stat_diff[i]) {
 545                                 int v;
 546
 547                                 v = p->vm_stat_diff[i];
 548                                 p->vm_stat_diff[i] = 0;
 549                                 atomic_long_add(v, &zone->vm_stat[i]);
 550                                 global_diff[i] += v;
 551                         }
 552         }
 553
 554         fold_diff(global_diff);
 555 }
 556
 557 /*
 558  * this is only called if !populated_zone(zone), which implies no other users of
 559  * pset->vm_stat_diff[] exsist.
 560  */
 561 void drain_zonestat(struct zone *zone, struct per_cpu_pageset *pset)
 562 {
 563         int i;
 564
 565         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 566                 if (pset->vm_stat_diff[i]) {
 567                         int v = pset->vm_stat_diff[i];
 568                         pset->vm_stat_diff[i] = 0;
 569                         atomic_long_add(v, &zone->vm_stat[i]);
 570                         atomic_long_add(v, &vm_stat[i]);
 571                 }
 572 }
 573 #endif
 574
 575 #ifdef CONFIG_NUMA
 576 /*
 577  * zonelist = the list of zones passed to the allocator
 578  * z        = the zone from which the allocation occurred.
 579  *
 580  * Must be called with interrupts disabled.
 581  *
 582  * When __GFP_OTHER_NODE is set assume the node of the preferred
 583  * zone is the local node. This is useful for daemons who allocate
 584  * memory on behalf of other processes.
 585  */
 586 void zone_statistics(struct zone *preferred_zone, struct zone *z, gfp_t flags)
 587 {
 588         if (z->zone_pgdat == preferred_zone->zone_pgdat) {
 589                 __inc_zone_state(z, NUMA_HIT);
 590         } else {
 591                 __inc_zone_state(z, NUMA_MISS);
 592                 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
 593         }
 594         if (z->node == ((flags & __GFP_OTHER_NODE) ?
 595                         preferred_zone->node : numa_node_id()))
 596                 __inc_zone_state(z, NUMA_LOCAL);
 597         else
 598                 __inc_zone_state(z, NUMA_OTHER);
 599 }
 600 #endif
 601
 602 #ifdef CONFIG_COMPACTION
 603
 604 struct contig_page_info {
 605         unsigned long free_pages;
 606         unsigned long free_blocks_total;
 607         unsigned long free_blocks_suitable;
 608 };
 609
 610 /*
 611  * Calculate the number of free pages in a zone, how many contiguous
 612  * pages are free and how many are large enough to satisfy an allocation of
 613  * the target size. Note that this function makes no attempt to estimate
 614  * how many suitable free blocks there *might* be if MOVABLE pages were
 615  * migrated. Calculating that is possible, but expensive and can be
 616  * figured out from userspace
 617  */
 618 static void fill_contig_page_info(struct zone *zone,
 619                                 unsigned int suitable_order,
 620                                 struct contig_page_info *info)
 621 {
 622         unsigned int order;
 623
 624         info->free_pages = 0;
 625         info->free_blocks_total = 0;
 626         info->free_blocks_suitable = 0;
 627
 628         for (order = 0; order < MAX_ORDER; order++) {
 629                 unsigned long blocks;
 630
 631                 /* Count number of free blocks */
 632                 blocks = zone->free_area[order].nr_free;
 633                 info->free_blocks_total += blocks;
 634
 635                 /* Count free base pages */
 636                 info->free_pages += blocks << order;
 637
 638                 /* Count the suitable free blocks */
 639                 if (order >= suitable_order)
 640                         info->free_blocks_suitable += blocks <<
 641                                                 (order - suitable_order);
 642         }
 643 }
 644
 645 /*
 646  * A fragmentation index only makes sense if an allocation of a requested
 647  * size would fail. If that is true, the fragmentation index indicates
 648  * whether external fragmentation or a lack of memory was the problem.
 649  * The value can be used to determine if page reclaim or compaction
 650  * should be used
 651  */
 652 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
 653 {
 654         unsigned long requested = 1UL << order;
 655
 656         if (!info->free_blocks_total)
 657                 return 0;
 658
 659         /* Fragmentation index only makes sense when a request would fail */
 660         if (info->free_blocks_suitable)
 661                 return -1000;
 662
 663         /*
 664          * Index is between 0 and 1 so return within 3 decimal places
 665          *
 666          * 0 => allocation would fail due to lack of memory
 667          * 1 => allocation would fail due to fragmentation
 668          */
 669         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
 670 }
 671
 672 /* Same as __fragmentation index but allocs contig_page_info on stack */
 673 int fragmentation_index(struct zone *zone, unsigned int order)
 674 {
 675         struct contig_page_info info;
 676
 677         fill_contig_page_info(zone, order, &info);
 678         return __fragmentation_index(order, &info);
 679 }
 680 #endif
 681
 682 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || defined(CONFIG_NUMA)
 683 #ifdef CONFIG_ZONE_DMA
 684 #define TEXT_FOR_DMA(xx) xx "_dma",
 685 #else
 686 #define TEXT_FOR_DMA(xx)
 687 #endif
 688
 689 #ifdef CONFIG_ZONE_DMA32
 690 #define TEXT_FOR_DMA32(xx) xx "_dma32",
 691 #else
 692 #define TEXT_FOR_DMA32(xx)
 693 #endif
 694
 695 #ifdef CONFIG_HIGHMEM
 696 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
 697 #else
 698 #define TEXT_FOR_HIGHMEM(xx)
 699 #endif
 700
 701 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
 702                                         TEXT_FOR_HIGHMEM(xx) xx "_movable",
 703
 704 const char * const vmstat_text[] = {
 705         /* enum zone_stat_item countes */
 706         "nr_free_pages",
 707         "nr_alloc_batch",
 708         "nr_inactive_anon",
 709         "nr_active_anon",
 710         "nr_inactive_file",
 711         "nr_active_file",
 712         "nr_unevictable",
 713         "nr_mlock",
 714         "nr_anon_pages",
 715         "nr_mapped",
 716         "nr_file_pages",
 717         "nr_dirty",
 718         "nr_writeback",
 719         "nr_slab_reclaimable",
 720         "nr_slab_unreclaimable",
 721         "nr_page_table_pages",
 722         "nr_kernel_stack",
 723         "nr_unstable",
 724         "nr_bounce",
 725         "nr_vmscan_write",
 726         "nr_vmscan_immediate_reclaim",
 727         "nr_writeback_temp",
 728         "nr_isolated_anon",
 729         "nr_isolated_file",
 730         "nr_shmem",
 731         "nr_dirtied",
 732         "nr_written",
 733         "nr_pages_scanned",
 734
 735 #ifdef CONFIG_NUMA
 736         "numa_hit",
 737         "numa_miss",
 738         "numa_foreign",
 739         "numa_interleave",
 740         "numa_local",
 741         "numa_other",
 742 #endif
 743         "workingset_refault",
 744         "workingset_activate",
 745         "workingset_nodereclaim",
 746         "nr_anon_transparent_hugepages",
 747         "nr_free_cma",
 748
 749         /* enum writeback_stat_item counters */
 750         "nr_dirty_threshold",
 751         "nr_dirty_background_threshold",
 752
 753 #ifdef CONFIG_VM_EVENT_COUNTERS
 754         /* enum vm_event_item counters */
 755         "pgpgin",
 756         "pgpgout",
 757         "pswpin",
 758         "pswpout",
 759
 760         TEXTS_FOR_ZONES("pgalloc")
 761
 762         "pgfree",
 763         "pgactivate",
 764         "pgdeactivate",
 765
 766         "pgfault",
 767         "pgmajfault",
 768
 769         TEXTS_FOR_ZONES("pgrefill")
 770         TEXTS_FOR_ZONES("pgsteal_kswapd")
 771         TEXTS_FOR_ZONES("pgsteal_direct")
 772         TEXTS_FOR_ZONES("pgscan_kswapd")
 773         TEXTS_FOR_ZONES("pgscan_direct")
 774         "pgscan_direct_throttle",
 775
 776 #ifdef CONFIG_NUMA
 777         "zone_reclaim_failed",
 778 #endif
 779         "pginodesteal",
 780         "slabs_scanned",
 781         "kswapd_inodesteal",
 782         "kswapd_low_wmark_hit_quickly",
 783         "kswapd_high_wmark_hit_quickly",
 784         "pageoutrun",
 785         "allocstall",
 786
 787         "pgrotated",
 788
 789         "drop_pagecache",
 790         "drop_slab",
 791
 792 #ifdef CONFIG_NUMA_BALANCING
 793         "numa_pte_updates",
 794         "numa_huge_pte_updates",
 795         "numa_hint_faults",
 796         "numa_hint_faults_local",
 797         "numa_pages_migrated",
 798 #endif
 799 #ifdef CONFIG_MIGRATION
 800         "pgmigrate_success",
 801         "pgmigrate_fail",
 802 #endif
 803 #ifdef CONFIG_COMPACTION
 804         "compact_migrate_scanned",
 805         "compact_free_scanned",
 806         "compact_isolated",
 807         "compact_stall",
 808         "compact_fail",
 809         "compact_success",
 810 #endif
 811
 812 #ifdef CONFIG_HUGETLB_PAGE
 813         "htlb_buddy_alloc_success",
 814         "htlb_buddy_alloc_fail",
 815 #endif
 816         "unevictable_pgs_culled",
 817         "unevictable_pgs_scanned",
 818         "unevictable_pgs_rescued",
 819         "unevictable_pgs_mlocked",
 820         "unevictable_pgs_munlocked",
 821         "unevictable_pgs_cleared",
 822         "unevictable_pgs_stranded",
 823
 824 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 825         "thp_fault_alloc",
 826         "thp_fault_fallback",
 827         "thp_collapse_alloc",
 828         "thp_collapse_alloc_failed",
 829         "thp_split",
 830         "thp_zero_page_alloc",
 831         "thp_zero_page_alloc_failed",
 832 #endif
 833 #ifdef CONFIG_MEMORY_BALLOON
 834         "balloon_inflate",
 835         "balloon_deflate",
 836 #ifdef CONFIG_BALLOON_COMPACTION
 837         "balloon_migrate",
 838 #endif
 839 #endif /* CONFIG_MEMORY_BALLOON */
 840 #ifdef CONFIG_DEBUG_TLBFLUSH
 841 #ifdef CONFIG_SMP
 842         "nr_tlb_remote_flush",
 843         "nr_tlb_remote_flush_received",
 844 #endif /* CONFIG_SMP */
 845         "nr_tlb_local_flush_all",
 846         "nr_tlb_local_flush_one",
 847 #endif /* CONFIG_DEBUG_TLBFLUSH */
 848
 849 #ifdef CONFIG_DEBUG_VM_VMACACHE
 850         "vmacache_find_calls",
 851         "vmacache_find_hits",
 852         "vmacache_full_flushes",
 853 #endif
 854 #endif /* CONFIG_VM_EVENTS_COUNTERS */
 855 };
 856 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
 857
 858
 859 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
 860      defined(CONFIG_PROC_FS)
 861 static void *frag_start(struct seq_file *m, loff_t *pos)
 862 {
 863         pg_data_t *pgdat;
 864         loff_t node = *pos;
 865
 866         for (pgdat = first_online_pgdat();
 867              pgdat && node;
 868              pgdat = next_online_pgdat(pgdat))
 869                 --node;
 870
 871         return pgdat;
 872 }
 873
 874 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
 875 {
 876         pg_data_t *pgdat = (pg_data_t *)arg;
 877
 878         (*pos)++;
 879         return next_online_pgdat(pgdat);
 880 }
 881
 882 static void frag_stop(struct seq_file *m, void *arg)
 883 {
 884 }
 885
 886 /* Walk all the zones in a node and print using a callback */
 887 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
 888                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
 889 {
 890         struct zone *zone;
 891         struct zone *node_zones = pgdat->node_zones;
 892         unsigned long flags;
 893
 894         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 895                 if (!populated_zone(zone))
 896                         continue;
 897
 898                 spin_lock_irqsave(&zone->lock, flags);
 899                 print(m, pgdat, zone);
 900                 spin_unlock_irqrestore(&zone->lock, flags);
 901         }
 902 }
 903 #endif
 904
 905 #ifdef CONFIG_PROC_FS
 906 static char * const migratetype_names[MIGRATE_TYPES] = {
 907         "Unmovable",
 908         "Reclaimable",
 909         "Movable",
 910         "Reserve",
 911 #ifdef CONFIG_CMA
 912         "CMA",
 913 #endif
 914 #ifdef CONFIG_MEMORY_ISOLATION
 915         "Isolate",
 916 #endif
 917 };
 918
 919 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
 920                                                 struct zone *zone)
 921 {
 922         int order;
 923
 924         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 925         for (order = 0; order < MAX_ORDER; ++order)
 926                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
 927         seq_putc(m, '\n');
 928 }
 929
 930 /*
 931  * This walks the free areas for each zone.
 932  */
 933 static int frag_show(struct seq_file *m, void *arg)
 934 {
 935         pg_data_t *pgdat = (pg_data_t *)arg;
 936         walk_zones_in_node(m, pgdat, frag_show_print);
 937         return 0;
 938 }
 939
 940 static void pagetypeinfo_showfree_print(struct seq_file *m,
 941                                         pg_data_t *pgdat, struct zone *zone)
 942 {
 943         int order, mtype;
 944
 945         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
 946                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
 947                                         pgdat->node_id,
 948                                         zone->name,
 949                                         migratetype_names[mtype]);
 950                 for (order = 0; order < MAX_ORDER; ++order) {
 951                         unsigned long freecount = 0;
 952                         struct free_area *area;
 953                         struct list_head *curr;
 954
 955                         area = &(zone->free_area[order]);
 956
 957                         list_for_each(curr, &area->free_list[mtype])
 958                                 freecount++;
 959                         seq_printf(m, "%6lu ", freecount);
 960                 }
 961                 seq_putc(m, '\n');
 962         }
 963 }
 964
 965 /* Print out the free pages at each order for each migatetype */
 966 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
 967 {
 968         int order;
 969         pg_data_t *pgdat = (pg_data_t *)arg;
 970
 971         /* Print header */
 972         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
 973         for (order = 0; order < MAX_ORDER; ++order)
 974                 seq_printf(m, "%6d ", order);
 975         seq_putc(m, '\n');
 976
 977         walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
 978
 979         return 0;
 980 }
 981
 982 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
 983                                         pg_data_t *pgdat, struct zone *zone)
 984 {
 985         int mtype;
 986         unsigned long pfn;
 987         unsigned long start_pfn = zone->zone_start_pfn;
 988         unsigned long end_pfn = zone_end_pfn(zone);
 989         unsigned long count[MIGRATE_TYPES] = { 0, };
 990
 991         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
 992                 struct page *page;
 993
 994                 if (!pfn_valid(pfn))
 995                         continue;
 996
 997                 page = pfn_to_page(pfn);
 998
 999                 /* Watch for unexpected holes punched in the memmap */
1000                 if (!memmap_valid_within(pfn, page, zone))
1001                         continue;
1002
1003                 mtype = get_pageblock_migratetype(page);
1004
1005                 if (mtype < MIGRATE_TYPES)
1006                         count[mtype]++;
1007         }
1008
1009         /* Print counts */
1010         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1011         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1012                 seq_printf(m, "%12lu ", count[mtype]);
1013         seq_putc(m, '\n');
1014 }
1015
1016 /* Print out the free pages at each order for each migratetype */
1017 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1018 {
1019         int mtype;
1020         pg_data_t *pgdat = (pg_data_t *)arg;
1021
1022         seq_printf(m, "\n%-23s", "Number of blocks type ");
1023         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1024                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1025         seq_putc(m, '\n');
1026         walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
1027
1028         return 0;
1029 }
1030
1031 #ifdef CONFIG_PAGE_OWNER
1032 static void pagetypeinfo_showmixedcount_print(struct seq_file *m,
1033                                                         pg_data_t *pgdat,
1034                                                         struct zone *zone)
1035 {
1036         struct page *page;
1037         struct page_ext *page_ext;
1038         unsigned long pfn = zone->zone_start_pfn, block_end_pfn;
1039         unsigned long end_pfn = pfn + zone->spanned_pages;
1040         unsigned long count[MIGRATE_TYPES] = { 0, };
1041         int pageblock_mt, page_mt;
1042         int i;
1043
1044         /* Scan block by block. First and last block may be incomplete */
1045         pfn = zone->zone_start_pfn;
1046
1047         /*
1048          * Walk the zone in pageblock_nr_pages steps. If a page block spans
1049          * a zone boundary, it will be double counted between zones. This does
1050          * not matter as the mixed block count will still be correct
1051          */
1052         for (; pfn < end_pfn; ) {
1053                 if (!pfn_valid(pfn)) {
1054                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
1055                         continue;
1056                 }
1057
1058                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
1059                 block_end_pfn = min(block_end_pfn, end_pfn);
1060
1061                 page = pfn_to_page(pfn);
1062                 pageblock_mt = get_pfnblock_migratetype(page, pfn);
1063
1064                 for (; pfn < block_end_pfn; pfn++) {
1065                         if (!pfn_valid_within(pfn))
1066                                 continue;
1067
1068                         page = pfn_to_page(pfn);
1069                         if (PageBuddy(page)) {
1070                                 pfn += (1UL << page_order(page)) - 1;
1071                                 continue;
1072                         }
1073
1074                         if (PageReserved(page))
1075                                 continue;
1076
1077                         page_ext = lookup_page_ext(page);
1078
1079                         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
1080                                 continue;
1081
1082                         page_mt = gfpflags_to_migratetype(page_ext->gfp_mask);
1083                         if (pageblock_mt != page_mt) {
1084                                 if (is_migrate_cma(pageblock_mt))
1085                                         count[MIGRATE_MOVABLE]++;
1086                                 else
1087                                         count[pageblock_mt]++;
1088
1089                                 pfn = block_end_pfn;
1090                                 break;
1091                         }
1092                         pfn += (1UL << page_ext->order) - 1;
1093                 }
1094         }
1095
1096         /* Print counts */
1097         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1098         for (i = 0; i < MIGRATE_TYPES; i++)
1099                 seq_printf(m, "%12lu ", count[i]);
1100         seq_putc(m, '\n');
1101 }
1102 #endif /* CONFIG_PAGE_OWNER */
1103
1104 /*
1105  * Print out the number of pageblocks for each migratetype that contain pages
1106  * of other types. This gives an indication of how well fallbacks are being
1107  * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1108  * to determine what is going on
1109  */
1110 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1111 {
1112 #ifdef CONFIG_PAGE_OWNER
1113         int mtype;
1114
1115         if (!page_owner_inited)
1116                 return;
1117
1118         drain_all_pages(NULL);
1119
1120         seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1121         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1122                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1123         seq_putc(m, '\n');
1124
1125         walk_zones_in_node(m, pgdat, pagetypeinfo_showmixedcount_print);
1126 #endif /* CONFIG_PAGE_OWNER */
1127 }
1128
1129 /*
1130  * This prints out statistics in relation to grouping pages by mobility.
1131  * It is expensive to collect so do not constantly read the file.
1132  */
1133 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1134 {
1135         pg_data_t *pgdat = (pg_data_t *)arg;
1136
1137         /* check memoryless node */
1138         if (!node_state(pgdat->node_id, N_MEMORY))
1139                 return 0;
1140
1141         seq_printf(m, "Page block order: %d\n", pageblock_order);
1142         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
1143         seq_putc(m, '\n');
1144         pagetypeinfo_showfree(m, pgdat);
1145         pagetypeinfo_showblockcount(m, pgdat);
1146         pagetypeinfo_showmixedcount(m, pgdat);
1147
1148         return 0;
1149 }
1150
1151 static const struct seq_operations fragmentation_op = {
1152         .start  = frag_start,
1153         .next   = frag_next,
1154         .stop   = frag_stop,
1155         .show   = frag_show,
1156 };
1157
1158 static int fragmentation_open(struct inode *inode, struct file *file)
1159 {
1160         return seq_open(file, &fragmentation_op);
1161 }
1162
1163 static const struct file_operations fragmentation_file_operations = {
1164         .open           = fragmentation_open,
1165         .read           = seq_read,
1166         .llseek         = seq_lseek,
1167         .release        = seq_release,
1168 };
1169
1170 static const struct seq_operations pagetypeinfo_op = {
1171         .start  = frag_start,
1172         .next   = frag_next,
1173         .stop   = frag_stop,
1174         .show   = pagetypeinfo_show,
1175 };
1176
1177 static int pagetypeinfo_open(struct inode *inode, struct file *file)
1178 {
1179         return seq_open(file, &pagetypeinfo_op);
1180 }
1181
1182 static const struct file_operations pagetypeinfo_file_ops = {
1183         .open           = pagetypeinfo_open,
1184         .read           = seq_read,
1185         .llseek         = seq_lseek,
1186         .release        = seq_release,
1187 };
1188
1189 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1190                                                         struct zone *zone)
1191 {
1192         int i;
1193         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1194         seq_printf(m,
1195                    "\n  pages free     %lu"
1196                    "\n        min      %lu"
1197                    "\n        low      %lu"
1198                    "\n        high     %lu"
1199                    "\n        scanned  %lu"
1200                    "\n        spanned  %lu"
1201                    "\n        present  %lu"
1202                    "\n        managed  %lu",
1203                    zone_page_state(zone, NR_FREE_PAGES),
1204                    min_wmark_pages(zone),
1205                    low_wmark_pages(zone),
1206                    high_wmark_pages(zone),
1207                    zone_page_state(zone, NR_PAGES_SCANNED),
1208                    zone->spanned_pages,
1209                    zone->present_pages,
1210                    zone->managed_pages);
1211
1212         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1213                 seq_printf(m, "\n    %-12s %lu", vmstat_text[i],
1214                                 zone_page_state(zone, i));
1215
1216         seq_printf(m,
1217                    "\n        protection: (%ld",
1218                    zone->lowmem_reserve[0]);
1219         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1220                 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1221         seq_printf(m,
1222                    ")"
1223                    "\n  pagesets");
1224         for_each_online_cpu(i) {
1225                 struct per_cpu_pageset *pageset;
1226
1227                 pageset = per_cpu_ptr(zone->pageset, i);
1228                 seq_printf(m,
1229                            "\n    cpu: %i"
1230                            "\n              count: %i"
1231                            "\n              high:  %i"
1232                            "\n              batch: %i",
1233                            i,
1234                            pageset->pcp.count,
1235                            pageset->pcp.high,
1236                            pageset->pcp.batch);
1237 #ifdef CONFIG_SMP
1238                 seq_printf(m, "\n  vm stats threshold: %d",
1239                                 pageset->stat_threshold);
1240 #endif
1241         }
1242         seq_printf(m,
1243                    "\n  all_unreclaimable: %u"
1244                    "\n  start_pfn:         %lu"
1245                    "\n  inactive_ratio:    %u",
1246                    !zone_reclaimable(zone),
1247                    zone->zone_start_pfn,
1248                    zone->inactive_ratio);
1249         seq_putc(m, '\n');
1250 }
1251
1252 /*
1253  * Output information about zones in @pgdat.
1254  */
1255 static int zoneinfo_show(struct seq_file *m, void *arg)
1256 {
1257         pg_data_t *pgdat = (pg_data_t *)arg;
1258         walk_zones_in_node(m, pgdat, zoneinfo_show_print);
1259         return 0;
1260 }
1261
1262 static const struct seq_operations zoneinfo_op = {
1263         .start  = frag_start, /* iterate over all zones. The same as in
1264                                * fragmentation. */
1265         .next   = frag_next,
1266         .stop   = frag_stop,
1267         .show   = zoneinfo_show,
1268 };
1269
1270 static int zoneinfo_open(struct inode *inode, struct file *file)
1271 {
1272         return seq_open(file, &zoneinfo_op);
1273 }
1274
1275 static const struct file_operations proc_zoneinfo_file_operations = {
1276         .open           = zoneinfo_open,
1277         .read           = seq_read,
1278         .llseek         = seq_lseek,
1279         .release        = seq_release,
1280 };
1281
1282 enum writeback_stat_item {
1283         NR_DIRTY_THRESHOLD,
1284         NR_DIRTY_BG_THRESHOLD,
1285         NR_VM_WRITEBACK_STAT_ITEMS,
1286 };
1287
1288 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1289 {
1290         unsigned long *v;
1291         int i, stat_items_size;
1292
1293         if (*pos >= ARRAY_SIZE(vmstat_text))
1294                 return NULL;
1295         stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1296                           NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1297
1298 #ifdef CONFIG_VM_EVENT_COUNTERS
1299         stat_items_size += sizeof(struct vm_event_state);
1300 #endif
1301
1302         v = kmalloc(stat_items_size, GFP_KERNEL);
1303         m->private = v;
1304         if (!v)
1305                 return ERR_PTR(-ENOMEM);
1306         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1307                 v[i] = global_page_state(i);
1308         v += NR_VM_ZONE_STAT_ITEMS;
1309
1310         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1311                             v + NR_DIRTY_THRESHOLD);
1312         v += NR_VM_WRITEBACK_STAT_ITEMS;
1313
1314 #ifdef CONFIG_VM_EVENT_COUNTERS
1315         all_vm_events(v);
1316         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1317         v[PGPGOUT] /= 2;
1318 #endif
1319         return (unsigned long *)m->private + *pos;
1320 }
1321
1322 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1323 {
1324         (*pos)++;
1325         if (*pos >= ARRAY_SIZE(vmstat_text))
1326                 return NULL;
1327         return (unsigned long *)m->private + *pos;
1328 }
1329
1330 static int vmstat_show(struct seq_file *m, void *arg)
1331 {
1332         unsigned long *l = arg;
1333         unsigned long off = l - (unsigned long *)m->private;
1334
1335         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1336         return 0;
1337 }
1338
1339 static void vmstat_stop(struct seq_file *m, void *arg)
1340 {
1341         kfree(m->private);
1342         m->private = NULL;
1343 }
1344
1345 static const struct seq_operations vmstat_op = {
1346         .start  = vmstat_start,
1347         .next   = vmstat_next,
1348         .stop   = vmstat_stop,
1349         .show   = vmstat_show,
1350 };
1351
1352 static int vmstat_open(struct inode *inode, struct file *file)
1353 {
1354         return seq_open(file, &vmstat_op);
1355 }
1356
1357 static const struct file_operations proc_vmstat_file_operations = {
1358         .open           = vmstat_open,
1359         .read           = seq_read,
1360         .llseek         = seq_lseek,
1361         .release        = seq_release,
1362 };
1363 #endif /* CONFIG_PROC_FS */
1364
1365 #ifdef CONFIG_SMP
1366 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1367 int sysctl_stat_interval __read_mostly = HZ;
1368 static cpumask_var_t cpu_stat_off;
1369
1370 static void vmstat_update(struct work_struct *w)
1371 {
1372         if (refresh_cpu_vm_stats())
1373                 /*
1374                  * Counters were updated so we expect more updates
1375                  * to occur in the future. Keep on running the
1376                  * update worker thread.
1377                  */
1378                 schedule_delayed_work(this_cpu_ptr(&vmstat_work),
1379                         round_jiffies_relative(sysctl_stat_interval));
1380         else {
1381                 /*
1382                  * We did not update any counters so the app may be in
1383                  * a mode where it does not cause counter updates.
1384                  * We may be uselessly running vmstat_update.
1385                  * Defer the checking for differentials to the
1386                  * shepherd thread on a different processor.
1387                  */
1388                 int r;
1389                 /*
1390                  * Shepherd work thread does not race since it never
1391                  * changes the bit if its zero but the cpu
1392                  * online / off line code may race if
1393                  * worker threads are still allowed during
1394                  * shutdown / startup.
1395                  */
1396                 r = cpumask_test_and_set_cpu(smp_processor_id(),
1397                         cpu_stat_off);
1398                 VM_BUG_ON(r);
1399         }
1400 }
1401
1402 /*
1403  * Check if the diffs for a certain cpu indicate that
1404  * an update is needed.
1405  */
1406 static bool need_update(int cpu)
1407 {
1408         struct zone *zone;
1409
1410         for_each_populated_zone(zone) {
1411                 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1412
1413                 BUILD_BUG_ON(sizeof(p->vm_stat_diff[0]) != 1);
1414                 /*
1415                  * The fast way of checking if there are any vmstat diffs.
1416                  * This works because the diffs are byte sized items.
1417                  */
1418                 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS))
1419                         return true;
1420
1421         }
1422         return false;
1423 }
1424
1425
1426 /*
1427  * Shepherd worker thread that checks the
1428  * differentials of processors that have their worker
1429  * threads for vm statistics updates disabled because of
1430  * inactivity.
1431  */
1432 static void vmstat_shepherd(struct work_struct *w);
1433
1434 static DECLARE_DELAYED_WORK(shepherd, vmstat_shepherd);
1435
1436 static void vmstat_shepherd(struct work_struct *w)
1437 {
1438         int cpu;
1439
1440         get_online_cpus();
1441         /* Check processors whose vmstat worker threads have been disabled */
1442         for_each_cpu(cpu, cpu_stat_off)
1443                 if (need_update(cpu) &&
1444                         cpumask_test_and_clear_cpu(cpu, cpu_stat_off))
1445
1446                         schedule_delayed_work_on(cpu,
1447                                 &per_cpu(vmstat_work, cpu), 0);
1448
1449         put_online_cpus();
1450
1451         schedule_delayed_work(&shepherd,
1452                 round_jiffies_relative(sysctl_stat_interval));
1453
1454 }
1455
1456 static void __init start_shepherd_timer(void)
1457 {
1458         int cpu;
1459
1460         for_each_possible_cpu(cpu)
1461                 INIT_DELAYED_WORK(per_cpu_ptr(&vmstat_work, cpu),
1462                         vmstat_update);
1463
1464         if (!alloc_cpumask_var(&cpu_stat_off, GFP_KERNEL))
1465                 BUG();
1466         cpumask_copy(cpu_stat_off, cpu_online_mask);
1467
1468         schedule_delayed_work(&shepherd,
1469                 round_jiffies_relative(sysctl_stat_interval));
1470 }
1471
1472 static void vmstat_cpu_dead(int node)
1473 {
1474         int cpu;
1475
1476         get_online_cpus();
1477         for_each_online_cpu(cpu)
1478                 if (cpu_to_node(cpu) == node)
1479                         goto end;
1480
1481         node_clear_state(node, N_CPU);
1482 end:
1483         put_online_cpus();
1484 }
1485
1486 /*
1487  * Use the cpu notifier to insure that the thresholds are recalculated
1488  * when necessary.
1489  */
1490 static int vmstat_cpuup_callback(struct notifier_block *nfb,
1491                 unsigned long action,
1492                 void *hcpu)
1493 {
1494         long cpu = (long)hcpu;
1495
1496         switch (action) {
1497         case CPU_ONLINE:
1498         case CPU_ONLINE_FROZEN:
1499                 refresh_zone_stat_thresholds();
1500                 node_set_state(cpu_to_node(cpu), N_CPU);
1501                 cpumask_set_cpu(cpu, cpu_stat_off);
1502                 break;
1503         case CPU_DOWN_PREPARE:
1504         case CPU_DOWN_PREPARE_FROZEN:
1505                 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1506                 cpumask_clear_cpu(cpu, cpu_stat_off);
1507                 break;
1508         case CPU_DOWN_FAILED:
1509         case CPU_DOWN_FAILED_FROZEN:
1510                 cpumask_set_cpu(cpu, cpu_stat_off);
1511                 break;
1512         case CPU_DEAD:
1513         case CPU_DEAD_FROZEN:
1514                 refresh_zone_stat_thresholds();
1515                 vmstat_cpu_dead(cpu_to_node(cpu));
1516                 break;
1517         default:
1518                 break;
1519         }
1520         return NOTIFY_OK;
1521 }
1522
1523 static struct notifier_block vmstat_notifier =
1524         { &vmstat_cpuup_callback, NULL, 0 };
1525 #endif
1526
1527 static int __init setup_vmstat(void)
1528 {
1529 #ifdef CONFIG_SMP
1530         cpu_notifier_register_begin();
1531         __register_cpu_notifier(&vmstat_notifier);
1532
1533         start_shepherd_timer();
1534         cpu_notifier_register_done();
1535 #endif
1536 #ifdef CONFIG_PROC_FS
1537         proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1538         proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1539         proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1540         proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1541 #endif
1542         return 0;
1543 }
1544 module_init(setup_vmstat)
1545
1546 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1547
1548 /*
1549  * Return an index indicating how much of the available free memory is
1550  * unusable for an allocation of the requested size.
1551  */
1552 static int unusable_free_index(unsigned int order,
1553                                 struct contig_page_info *info)
1554 {
1555         /* No free memory is interpreted as all free memory is unusable */
1556         if (info->free_pages == 0)
1557                 return 1000;
1558
1559         /*
1560          * Index should be a value between 0 and 1. Return a value to 3
1561          * decimal places.
1562          *
1563          * 0 => no fragmentation
1564          * 1 => high fragmentation
1565          */
1566         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1567
1568 }
1569
1570 static void unusable_show_print(struct seq_file *m,
1571                                         pg_data_t *pgdat, struct zone *zone)
1572 {
1573         unsigned int order;
1574         int index;
1575         struct contig_page_info info;
1576
1577         seq_printf(m, "Node %d, zone %8s ",
1578                                 pgdat->node_id,
1579                                 zone->name);
1580         for (order = 0; order < MAX_ORDER; ++order) {
1581                 fill_contig_page_info(zone, order, &info);
1582                 index = unusable_free_index(order, &info);
1583                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1584         }
1585
1586         seq_putc(m, '\n');
1587 }
1588
1589 /*
1590  * Display unusable free space index
1591  *
1592  * The unusable free space index measures how much of the available free
1593  * memory cannot be used to satisfy an allocation of a given size and is a
1594  * value between 0 and 1. The higher the value, the more of free memory is
1595  * unusable and by implication, the worse the external fragmentation is. This
1596  * can be expressed as a percentage by multiplying by 100.
1597  */
1598 static int unusable_show(struct seq_file *m, void *arg)
1599 {
1600         pg_data_t *pgdat = (pg_data_t *)arg;
1601
1602         /* check memoryless node */
1603         if (!node_state(pgdat->node_id, N_MEMORY))
1604                 return 0;
1605
1606         walk_zones_in_node(m, pgdat, unusable_show_print);
1607
1608         return 0;
1609 }
1610
1611 static const struct seq_operations unusable_op = {
1612         .start  = frag_start,
1613         .next   = frag_next,
1614         .stop   = frag_stop,
1615         .show   = unusable_show,
1616 };
1617
1618 static int unusable_open(struct inode *inode, struct file *file)
1619 {
1620         return seq_open(file, &unusable_op);
1621 }
1622
1623 static const struct file_operations unusable_file_ops = {
1624         .open           = unusable_open,
1625         .read           = seq_read,
1626         .llseek         = seq_lseek,
1627         .release        = seq_release,
1628 };
1629
1630 static void extfrag_show_print(struct seq_file *m,
1631                                         pg_data_t *pgdat, struct zone *zone)
1632 {
1633         unsigned int order;
1634         int index;
1635
1636         /* Alloc on stack as interrupts are disabled for zone walk */
1637         struct contig_page_info info;
1638
1639         seq_printf(m, "Node %d, zone %8s ",
1640                                 pgdat->node_id,
1641                                 zone->name);
1642         for (order = 0; order < MAX_ORDER; ++order) {
1643                 fill_contig_page_info(zone, order, &info);
1644                 index = __fragmentation_index(order, &info);
1645                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1646         }
1647
1648         seq_putc(m, '\n');
1649 }
1650
1651 /*
1652  * Display fragmentation index for orders that allocations would fail for
1653  */
1654 static int extfrag_show(struct seq_file *m, void *arg)
1655 {
1656         pg_data_t *pgdat = (pg_data_t *)arg;
1657
1658         walk_zones_in_node(m, pgdat, extfrag_show_print);
1659
1660         return 0;
1661 }
1662
1663 static const struct seq_operations extfrag_op = {
1664         .start  = frag_start,
1665         .next   = frag_next,
1666         .stop   = frag_stop,
1667         .show   = extfrag_show,
1668 };
1669
1670 static int extfrag_open(struct inode *inode, struct file *file)
1671 {
1672         return seq_open(file, &extfrag_op);
1673 }
1674
1675 static const struct file_operations extfrag_file_ops = {
1676         .open           = extfrag_open,
1677         .read           = seq_read,
1678         .llseek         = seq_lseek,
1679         .release        = seq_release,
1680 };
1681
1682 static int __init extfrag_debug_init(void)
1683 {
1684         struct dentry *extfrag_debug_root;
1685
1686         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1687         if (!extfrag_debug_root)
1688                 return -ENOMEM;
1689
1690         if (!debugfs_create_file("unusable_index", 0444,
1691                         extfrag_debug_root, NULL, &unusable_file_ops))
1692                 goto fail;
1693
1694         if (!debugfs_create_file("extfrag_index", 0444,
1695                         extfrag_debug_root, NULL, &extfrag_file_ops))
1696                 goto fail;
1697
1698         return 0;
1699 fail:
1700         debugfs_remove_recursive(extfrag_debug_root);
1701         return -ENOMEM;
1702 }
1703
1704 module_init(extfrag_debug_init);
1705 #endif