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                            long 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, enum zone_stat_item item,
 328                              long 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                          long 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                          long 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
 601 /*
 602  * Determine the per node value of a stat item.
 603  */
 604 unsigned long node_page_state(int node, enum zone_stat_item item)
 605 {
 606         struct zone *zones = NODE_DATA(node)->node_zones;
 607
 608         return
 609 #ifdef CONFIG_ZONE_DMA
 610                 zone_page_state(&zones[ZONE_DMA], item) +
 611 #endif
 612 #ifdef CONFIG_ZONE_DMA32
 613                 zone_page_state(&zones[ZONE_DMA32], item) +
 614 #endif
 615 #ifdef CONFIG_HIGHMEM
 616                 zone_page_state(&zones[ZONE_HIGHMEM], item) +
 617 #endif
 618                 zone_page_state(&zones[ZONE_NORMAL], item) +
 619                 zone_page_state(&zones[ZONE_MOVABLE], item);
 620 }
 621
 622 #endif
 623
 624 #ifdef CONFIG_COMPACTION
 625
 626 struct contig_page_info {
 627         unsigned long free_pages;
 628         unsigned long free_blocks_total;
 629         unsigned long free_blocks_suitable;
 630 };
 631
 632 /*
 633  * Calculate the number of free pages in a zone, how many contiguous
 634  * pages are free and how many are large enough to satisfy an allocation of
 635  * the target size. Note that this function makes no attempt to estimate
 636  * how many suitable free blocks there *might* be if MOVABLE pages were
 637  * migrated. Calculating that is possible, but expensive and can be
 638  * figured out from userspace
 639  */
 640 static void fill_contig_page_info(struct zone *zone,
 641                                 unsigned int suitable_order,
 642                                 struct contig_page_info *info)
 643 {
 644         unsigned int order;
 645
 646         info->free_pages = 0;
 647         info->free_blocks_total = 0;
 648         info->free_blocks_suitable = 0;
 649
 650         for (order = 0; order < MAX_ORDER; order++) {
 651                 unsigned long blocks;
 652
 653                 /* Count number of free blocks */
 654                 blocks = zone->free_area[order].nr_free;
 655                 info->free_blocks_total += blocks;
 656
 657                 /* Count free base pages */
 658                 info->free_pages += blocks << order;
 659
 660                 /* Count the suitable free blocks */
 661                 if (order >= suitable_order)
 662                         info->free_blocks_suitable += blocks <<
 663                                                 (order - suitable_order);
 664         }
 665 }
 666
 667 /*
 668  * A fragmentation index only makes sense if an allocation of a requested
 669  * size would fail. If that is true, the fragmentation index indicates
 670  * whether external fragmentation or a lack of memory was the problem.
 671  * The value can be used to determine if page reclaim or compaction
 672  * should be used
 673  */
 674 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
 675 {
 676         unsigned long requested = 1UL << order;
 677
 678         if (!info->free_blocks_total)
 679                 return 0;
 680
 681         /* Fragmentation index only makes sense when a request would fail */
 682         if (info->free_blocks_suitable)
 683                 return -1000;
 684
 685         /*
 686          * Index is between 0 and 1 so return within 3 decimal places
 687          *
 688          * 0 => allocation would fail due to lack of memory
 689          * 1 => allocation would fail due to fragmentation
 690          */
 691         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
 692 }
 693
 694 /* Same as __fragmentation index but allocs contig_page_info on stack */
 695 int fragmentation_index(struct zone *zone, unsigned int order)
 696 {
 697         struct contig_page_info info;
 698
 699         fill_contig_page_info(zone, order, &info);
 700         return __fragmentation_index(order, &info);
 701 }
 702 #endif
 703
 704 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || defined(CONFIG_NUMA)
 705 #ifdef CONFIG_ZONE_DMA
 706 #define TEXT_FOR_DMA(xx) xx "_dma",
 707 #else
 708 #define TEXT_FOR_DMA(xx)
 709 #endif
 710
 711 #ifdef CONFIG_ZONE_DMA32
 712 #define TEXT_FOR_DMA32(xx) xx "_dma32",
 713 #else
 714 #define TEXT_FOR_DMA32(xx)
 715 #endif
 716
 717 #ifdef CONFIG_HIGHMEM
 718 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
 719 #else
 720 #define TEXT_FOR_HIGHMEM(xx)
 721 #endif
 722
 723 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
 724                                         TEXT_FOR_HIGHMEM(xx) xx "_movable",
 725
 726 const char * const vmstat_text[] = {
 727         /* enum zone_stat_item countes */
 728         "nr_free_pages",
 729         "nr_alloc_batch",
 730         "nr_inactive_anon",
 731         "nr_active_anon",
 732         "nr_inactive_file",
 733         "nr_active_file",
 734         "nr_unevictable",
 735         "nr_mlock",
 736         "nr_anon_pages",
 737         "nr_mapped",
 738         "nr_file_pages",
 739         "nr_dirty",
 740         "nr_writeback",
 741         "nr_slab_reclaimable",
 742         "nr_slab_unreclaimable",
 743         "nr_page_table_pages",
 744         "nr_kernel_stack",
 745         "nr_unstable",
 746         "nr_bounce",
 747         "nr_vmscan_write",
 748         "nr_vmscan_immediate_reclaim",
 749         "nr_writeback_temp",
 750         "nr_isolated_anon",
 751         "nr_isolated_file",
 752         "nr_shmem",
 753         "nr_dirtied",
 754         "nr_written",
 755         "nr_pages_scanned",
 756
 757 #ifdef CONFIG_NUMA
 758         "numa_hit",
 759         "numa_miss",
 760         "numa_foreign",
 761         "numa_interleave",
 762         "numa_local",
 763         "numa_other",
 764 #endif
 765         "workingset_refault",
 766         "workingset_activate",
 767         "workingset_nodereclaim",
 768         "nr_anon_transparent_hugepages",
 769         "nr_free_cma",
 770
 771         /* enum writeback_stat_item counters */
 772         "nr_dirty_threshold",
 773         "nr_dirty_background_threshold",
 774
 775 #ifdef CONFIG_VM_EVENT_COUNTERS
 776         /* enum vm_event_item counters */
 777         "pgpgin",
 778         "pgpgout",
 779         "pswpin",
 780         "pswpout",
 781
 782         TEXTS_FOR_ZONES("pgalloc")
 783
 784         "pgfree",
 785         "pgactivate",
 786         "pgdeactivate",
 787
 788         "pgfault",
 789         "pgmajfault",
 790
 791         TEXTS_FOR_ZONES("pgrefill")
 792         TEXTS_FOR_ZONES("pgsteal_kswapd")
 793         TEXTS_FOR_ZONES("pgsteal_direct")
 794         TEXTS_FOR_ZONES("pgscan_kswapd")
 795         TEXTS_FOR_ZONES("pgscan_direct")
 796         "pgscan_direct_throttle",
 797
 798 #ifdef CONFIG_NUMA
 799         "zone_reclaim_failed",
 800 #endif
 801         "pginodesteal",
 802         "slabs_scanned",
 803         "kswapd_inodesteal",
 804         "kswapd_low_wmark_hit_quickly",
 805         "kswapd_high_wmark_hit_quickly",
 806         "pageoutrun",
 807         "allocstall",
 808
 809         "pgrotated",
 810
 811         "drop_pagecache",
 812         "drop_slab",
 813
 814 #ifdef CONFIG_NUMA_BALANCING
 815         "numa_pte_updates",
 816         "numa_huge_pte_updates",
 817         "numa_hint_faults",
 818         "numa_hint_faults_local",
 819         "numa_pages_migrated",
 820 #endif
 821 #ifdef CONFIG_MIGRATION
 822         "pgmigrate_success",
 823         "pgmigrate_fail",
 824 #endif
 825 #ifdef CONFIG_COMPACTION
 826         "compact_migrate_scanned",
 827         "compact_free_scanned",
 828         "compact_isolated",
 829         "compact_stall",
 830         "compact_fail",
 831         "compact_success",
 832 #endif
 833
 834 #ifdef CONFIG_HUGETLB_PAGE
 835         "htlb_buddy_alloc_success",
 836         "htlb_buddy_alloc_fail",
 837 #endif
 838         "unevictable_pgs_culled",
 839         "unevictable_pgs_scanned",
 840         "unevictable_pgs_rescued",
 841         "unevictable_pgs_mlocked",
 842         "unevictable_pgs_munlocked",
 843         "unevictable_pgs_cleared",
 844         "unevictable_pgs_stranded",
 845
 846 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 847         "thp_fault_alloc",
 848         "thp_fault_fallback",
 849         "thp_collapse_alloc",
 850         "thp_collapse_alloc_failed",
 851         "thp_split",
 852         "thp_zero_page_alloc",
 853         "thp_zero_page_alloc_failed",
 854 #endif
 855 #ifdef CONFIG_MEMORY_BALLOON
 856         "balloon_inflate",
 857         "balloon_deflate",
 858 #ifdef CONFIG_BALLOON_COMPACTION
 859         "balloon_migrate",
 860 #endif
 861 #endif /* CONFIG_MEMORY_BALLOON */
 862 #ifdef CONFIG_DEBUG_TLBFLUSH
 863 #ifdef CONFIG_SMP
 864         "nr_tlb_remote_flush",
 865         "nr_tlb_remote_flush_received",
 866 #endif /* CONFIG_SMP */
 867         "nr_tlb_local_flush_all",
 868         "nr_tlb_local_flush_one",
 869 #endif /* CONFIG_DEBUG_TLBFLUSH */
 870
 871 #ifdef CONFIG_DEBUG_VM_VMACACHE
 872         "vmacache_find_calls",
 873         "vmacache_find_hits",
 874         "vmacache_full_flushes",
 875 #endif
 876 #endif /* CONFIG_VM_EVENTS_COUNTERS */
 877 };
 878 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
 879
 880
 881 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
 882      defined(CONFIG_PROC_FS)
 883 static void *frag_start(struct seq_file *m, loff_t *pos)
 884 {
 885         pg_data_t *pgdat;
 886         loff_t node = *pos;
 887
 888         for (pgdat = first_online_pgdat();
 889              pgdat && node;
 890              pgdat = next_online_pgdat(pgdat))
 891                 --node;
 892
 893         return pgdat;
 894 }
 895
 896 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
 897 {
 898         pg_data_t *pgdat = (pg_data_t *)arg;
 899
 900         (*pos)++;
 901         return next_online_pgdat(pgdat);
 902 }
 903
 904 static void frag_stop(struct seq_file *m, void *arg)
 905 {
 906 }
 907
 908 /* Walk all the zones in a node and print using a callback */
 909 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
 910                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
 911 {
 912         struct zone *zone;
 913         struct zone *node_zones = pgdat->node_zones;
 914         unsigned long flags;
 915
 916         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 917                 if (!populated_zone(zone))
 918                         continue;
 919
 920                 spin_lock_irqsave(&zone->lock, flags);
 921                 print(m, pgdat, zone);
 922                 spin_unlock_irqrestore(&zone->lock, flags);
 923         }
 924 }
 925 #endif
 926
 927 #ifdef CONFIG_PROC_FS
 928 static char * const migratetype_names[MIGRATE_TYPES] = {
 929         "Unmovable",
 930         "Movable",
 931         "Reclaimable",
 932         "HighAtomic",
 933 #ifdef CONFIG_CMA
 934         "CMA",
 935 #endif
 936 #ifdef CONFIG_MEMORY_ISOLATION
 937         "Isolate",
 938 #endif
 939 };
 940
 941 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
 942                                                 struct zone *zone)
 943 {
 944         int order;
 945
 946         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 947         for (order = 0; order < MAX_ORDER; ++order)
 948                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
 949         seq_putc(m, '\n');
 950 }
 951
 952 /*
 953  * This walks the free areas for each zone.
 954  */
 955 static int frag_show(struct seq_file *m, void *arg)
 956 {
 957         pg_data_t *pgdat = (pg_data_t *)arg;
 958         walk_zones_in_node(m, pgdat, frag_show_print);
 959         return 0;
 960 }
 961
 962 static void pagetypeinfo_showfree_print(struct seq_file *m,
 963                                         pg_data_t *pgdat, struct zone *zone)
 964 {
 965         int order, mtype;
 966
 967         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
 968                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
 969                                         pgdat->node_id,
 970                                         zone->name,
 971                                         migratetype_names[mtype]);
 972                 for (order = 0; order < MAX_ORDER; ++order) {
 973                         unsigned long freecount = 0;
 974                         struct free_area *area;
 975                         struct list_head *curr;
 976
 977                         area = &(zone->free_area[order]);
 978
 979                         list_for_each(curr, &area->free_list[mtype])
 980                                 freecount++;
 981                         seq_printf(m, "%6lu ", freecount);
 982                 }
 983                 seq_putc(m, '\n');
 984         }
 985 }
 986
 987 /* Print out the free pages at each order for each migatetype */
 988 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
 989 {
 990         int order;
 991         pg_data_t *pgdat = (pg_data_t *)arg;
 992
 993         /* Print header */
 994         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
 995         for (order = 0; order < MAX_ORDER; ++order)
 996                 seq_printf(m, "%6d ", order);
 997         seq_putc(m, '\n');
 998
 999         walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
1000
1001         return 0;
1002 }
1003
1004 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1005                                         pg_data_t *pgdat, struct zone *zone)
1006 {
1007         int mtype;
1008         unsigned long pfn;
1009         unsigned long start_pfn = zone->zone_start_pfn;
1010         unsigned long end_pfn = zone_end_pfn(zone);
1011         unsigned long count[MIGRATE_TYPES] = { 0, };
1012
1013         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1014                 struct page *page;
1015
1016                 if (!pfn_valid(pfn))
1017                         continue;
1018
1019                 page = pfn_to_page(pfn);
1020
1021                 /* Watch for unexpected holes punched in the memmap */
1022                 if (!memmap_valid_within(pfn, page, zone))
1023                         continue;
1024
1025                 mtype = get_pageblock_migratetype(page);
1026
1027                 if (mtype < MIGRATE_TYPES)
1028                         count[mtype]++;
1029         }
1030
1031         /* Print counts */
1032         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1033         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1034                 seq_printf(m, "%12lu ", count[mtype]);
1035         seq_putc(m, '\n');
1036 }
1037
1038 /* Print out the free pages at each order for each migratetype */
1039 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1040 {
1041         int mtype;
1042         pg_data_t *pgdat = (pg_data_t *)arg;
1043
1044         seq_printf(m, "\n%-23s", "Number of blocks type ");
1045         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1046                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1047         seq_putc(m, '\n');
1048         walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
1049
1050         return 0;
1051 }
1052
1053 #ifdef CONFIG_PAGE_OWNER
1054 static void pagetypeinfo_showmixedcount_print(struct seq_file *m,
1055                                                         pg_data_t *pgdat,
1056                                                         struct zone *zone)
1057 {
1058         struct page *page;
1059         struct page_ext *page_ext;
1060         unsigned long pfn = zone->zone_start_pfn, block_end_pfn;
1061         unsigned long end_pfn = pfn + zone->spanned_pages;
1062         unsigned long count[MIGRATE_TYPES] = { 0, };
1063         int pageblock_mt, page_mt;
1064         int i;
1065
1066         /* Scan block by block. First and last block may be incomplete */
1067         pfn = zone->zone_start_pfn;
1068
1069         /*
1070          * Walk the zone in pageblock_nr_pages steps. If a page block spans
1071          * a zone boundary, it will be double counted between zones. This does
1072          * not matter as the mixed block count will still be correct
1073          */
1074         for (; pfn < end_pfn; ) {
1075                 if (!pfn_valid(pfn)) {
1076                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
1077                         continue;
1078                 }
1079
1080                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
1081                 block_end_pfn = min(block_end_pfn, end_pfn);
1082
1083                 page = pfn_to_page(pfn);
1084                 pageblock_mt = get_pfnblock_migratetype(page, pfn);
1085
1086                 for (; pfn < block_end_pfn; pfn++) {
1087                         if (!pfn_valid_within(pfn))
1088                                 continue;
1089
1090                         page = pfn_to_page(pfn);
1091                         if (PageBuddy(page)) {
1092                                 pfn += (1UL << page_order(page)) - 1;
1093                                 continue;
1094                         }
1095
1096                         if (PageReserved(page))
1097                                 continue;
1098
1099                         page_ext = lookup_page_ext(page);
1100
1101                         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
1102                                 continue;
1103
1104                         page_mt = gfpflags_to_migratetype(page_ext->gfp_mask);
1105                         if (pageblock_mt != page_mt) {
1106                                 if (is_migrate_cma(pageblock_mt))
1107                                         count[MIGRATE_MOVABLE]++;
1108                                 else
1109                                         count[pageblock_mt]++;
1110
1111                                 pfn = block_end_pfn;
1112                                 break;
1113                         }
1114                         pfn += (1UL << page_ext->order) - 1;
1115                 }
1116         }
1117
1118         /* Print counts */
1119         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1120         for (i = 0; i < MIGRATE_TYPES; i++)
1121                 seq_printf(m, "%12lu ", count[i]);
1122         seq_putc(m, '\n');
1123 }
1124 #endif /* CONFIG_PAGE_OWNER */
1125
1126 /*
1127  * Print out the number of pageblocks for each migratetype that contain pages
1128  * of other types. This gives an indication of how well fallbacks are being
1129  * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1130  * to determine what is going on
1131  */
1132 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1133 {
1134 #ifdef CONFIG_PAGE_OWNER
1135         int mtype;
1136
1137         if (!page_owner_inited)
1138                 return;
1139
1140         drain_all_pages(NULL);
1141
1142         seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1143         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1144                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1145         seq_putc(m, '\n');
1146
1147         walk_zones_in_node(m, pgdat, pagetypeinfo_showmixedcount_print);
1148 #endif /* CONFIG_PAGE_OWNER */
1149 }
1150
1151 /*
1152  * This prints out statistics in relation to grouping pages by mobility.
1153  * It is expensive to collect so do not constantly read the file.
1154  */
1155 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1156 {
1157         pg_data_t *pgdat = (pg_data_t *)arg;
1158
1159         /* check memoryless node */
1160         if (!node_state(pgdat->node_id, N_MEMORY))
1161                 return 0;
1162
1163         seq_printf(m, "Page block order: %d\n", pageblock_order);
1164         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
1165         seq_putc(m, '\n');
1166         pagetypeinfo_showfree(m, pgdat);
1167         pagetypeinfo_showblockcount(m, pgdat);
1168         pagetypeinfo_showmixedcount(m, pgdat);
1169
1170         return 0;
1171 }
1172
1173 static const struct seq_operations fragmentation_op = {
1174         .start  = frag_start,
1175         .next   = frag_next,
1176         .stop   = frag_stop,
1177         .show   = frag_show,
1178 };
1179
1180 static int fragmentation_open(struct inode *inode, struct file *file)
1181 {
1182         return seq_open(file, &fragmentation_op);
1183 }
1184
1185 static const struct file_operations fragmentation_file_operations = {
1186         .open           = fragmentation_open,
1187         .read           = seq_read,
1188         .llseek         = seq_lseek,
1189         .release        = seq_release,
1190 };
1191
1192 static const struct seq_operations pagetypeinfo_op = {
1193         .start  = frag_start,
1194         .next   = frag_next,
1195         .stop   = frag_stop,
1196         .show   = pagetypeinfo_show,
1197 };
1198
1199 static int pagetypeinfo_open(struct inode *inode, struct file *file)
1200 {
1201         return seq_open(file, &pagetypeinfo_op);
1202 }
1203
1204 static const struct file_operations pagetypeinfo_file_ops = {
1205         .open           = pagetypeinfo_open,
1206         .read           = seq_read,
1207         .llseek         = seq_lseek,
1208         .release        = seq_release,
1209 };
1210
1211 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1212                                                         struct zone *zone)
1213 {
1214         int i;
1215         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1216         seq_printf(m,
1217                    "\n  pages free     %lu"
1218                    "\n        min      %lu"
1219                    "\n        low      %lu"
1220                    "\n        high     %lu"
1221                    "\n        scanned  %lu"
1222                    "\n        spanned  %lu"
1223                    "\n        present  %lu"
1224                    "\n        managed  %lu",
1225                    zone_page_state(zone, NR_FREE_PAGES),
1226                    min_wmark_pages(zone),
1227                    low_wmark_pages(zone),
1228                    high_wmark_pages(zone),
1229                    zone_page_state(zone, NR_PAGES_SCANNED),
1230                    zone->spanned_pages,
1231                    zone->present_pages,
1232                    zone->managed_pages);
1233
1234         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1235                 seq_printf(m, "\n    %-12s %lu", vmstat_text[i],
1236                                 zone_page_state(zone, i));
1237
1238         seq_printf(m,
1239                    "\n        protection: (%ld",
1240                    zone->lowmem_reserve[0]);
1241         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1242                 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1243         seq_printf(m,
1244                    ")"
1245                    "\n  pagesets");
1246         for_each_online_cpu(i) {
1247                 struct per_cpu_pageset *pageset;
1248
1249                 pageset = per_cpu_ptr(zone->pageset, i);
1250                 seq_printf(m,
1251                            "\n    cpu: %i"
1252                            "\n              count: %i"
1253                            "\n              high:  %i"
1254                            "\n              batch: %i",
1255                            i,
1256                            pageset->pcp.count,
1257                            pageset->pcp.high,
1258                            pageset->pcp.batch);
1259 #ifdef CONFIG_SMP
1260                 seq_printf(m, "\n  vm stats threshold: %d",
1261                                 pageset->stat_threshold);
1262 #endif
1263         }
1264         seq_printf(m,
1265                    "\n  all_unreclaimable: %u"
1266                    "\n  start_pfn:         %lu"
1267                    "\n  inactive_ratio:    %u",
1268                    !zone_reclaimable(zone),
1269                    zone->zone_start_pfn,
1270                    zone->inactive_ratio);
1271         seq_putc(m, '\n');
1272 }
1273
1274 /*
1275  * Output information about zones in @pgdat.
1276  */
1277 static int zoneinfo_show(struct seq_file *m, void *arg)
1278 {
1279         pg_data_t *pgdat = (pg_data_t *)arg;
1280         walk_zones_in_node(m, pgdat, zoneinfo_show_print);
1281         return 0;
1282 }
1283
1284 static const struct seq_operations zoneinfo_op = {
1285         .start  = frag_start, /* iterate over all zones. The same as in
1286                                * fragmentation. */
1287         .next   = frag_next,
1288         .stop   = frag_stop,
1289         .show   = zoneinfo_show,
1290 };
1291
1292 static int zoneinfo_open(struct inode *inode, struct file *file)
1293 {
1294         return seq_open(file, &zoneinfo_op);
1295 }
1296
1297 static const struct file_operations proc_zoneinfo_file_operations = {
1298         .open           = zoneinfo_open,
1299         .read           = seq_read,
1300         .llseek         = seq_lseek,
1301         .release        = seq_release,
1302 };
1303
1304 enum writeback_stat_item {
1305         NR_DIRTY_THRESHOLD,
1306         NR_DIRTY_BG_THRESHOLD,
1307         NR_VM_WRITEBACK_STAT_ITEMS,
1308 };
1309
1310 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1311 {
1312         unsigned long *v;
1313         int i, stat_items_size;
1314
1315         if (*pos >= ARRAY_SIZE(vmstat_text))
1316                 return NULL;
1317         stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1318                           NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1319
1320 #ifdef CONFIG_VM_EVENT_COUNTERS
1321         stat_items_size += sizeof(struct vm_event_state);
1322 #endif
1323
1324         v = kmalloc(stat_items_size, GFP_KERNEL);
1325         m->private = v;
1326         if (!v)
1327                 return ERR_PTR(-ENOMEM);
1328         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1329                 v[i] = global_page_state(i);
1330         v += NR_VM_ZONE_STAT_ITEMS;
1331
1332         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1333                             v + NR_DIRTY_THRESHOLD);
1334         v += NR_VM_WRITEBACK_STAT_ITEMS;
1335
1336 #ifdef CONFIG_VM_EVENT_COUNTERS
1337         all_vm_events(v);
1338         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1339         v[PGPGOUT] /= 2;
1340 #endif
1341         return (unsigned long *)m->private + *pos;
1342 }
1343
1344 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1345 {
1346         (*pos)++;
1347         if (*pos >= ARRAY_SIZE(vmstat_text))
1348                 return NULL;
1349         return (unsigned long *)m->private + *pos;
1350 }
1351
1352 static int vmstat_show(struct seq_file *m, void *arg)
1353 {
1354         unsigned long *l = arg;
1355         unsigned long off = l - (unsigned long *)m->private;
1356
1357         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1358         return 0;
1359 }
1360
1361 static void vmstat_stop(struct seq_file *m, void *arg)
1362 {
1363         kfree(m->private);
1364         m->private = NULL;
1365 }
1366
1367 static const struct seq_operations vmstat_op = {
1368         .start  = vmstat_start,
1369         .next   = vmstat_next,
1370         .stop   = vmstat_stop,
1371         .show   = vmstat_show,
1372 };
1373
1374 static int vmstat_open(struct inode *inode, struct file *file)
1375 {
1376         return seq_open(file, &vmstat_op);
1377 }
1378
1379 static const struct file_operations proc_vmstat_file_operations = {
1380         .open           = vmstat_open,
1381         .read           = seq_read,
1382         .llseek         = seq_lseek,
1383         .release        = seq_release,
1384 };
1385 #endif /* CONFIG_PROC_FS */
1386
1387 #ifdef CONFIG_SMP
1388 static struct workqueue_struct *vmstat_wq;
1389 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1390 int sysctl_stat_interval __read_mostly = HZ;
1391 static cpumask_var_t cpu_stat_off;
1392
1393 static void vmstat_update(struct work_struct *w)
1394 {
1395         if (refresh_cpu_vm_stats()) {
1396                 /*
1397                  * Counters were updated so we expect more updates
1398                  * to occur in the future. Keep on running the
1399                  * update worker thread.
1400                  */
1401                 queue_delayed_work_on(smp_processor_id(), vmstat_wq,
1402                         this_cpu_ptr(&vmstat_work),
1403                         round_jiffies_relative(sysctl_stat_interval));
1404         } else {
1405                 /*
1406                  * We did not update any counters so the app may be in
1407                  * a mode where it does not cause counter updates.
1408                  * We may be uselessly running vmstat_update.
1409                  * Defer the checking for differentials to the
1410                  * shepherd thread on a different processor.
1411                  */
1412                 int r;
1413                 /*
1414                  * Shepherd work thread does not race since it never
1415                  * changes the bit if its zero but the cpu
1416                  * online / off line code may race if
1417                  * worker threads are still allowed during
1418                  * shutdown / startup.
1419                  */
1420                 r = cpumask_test_and_set_cpu(smp_processor_id(),
1421                         cpu_stat_off);
1422                 VM_BUG_ON(r);
1423         }
1424 }
1425
1426 /*
1427  * Check if the diffs for a certain cpu indicate that
1428  * an update is needed.
1429  */
1430 static bool need_update(int cpu)
1431 {
1432         struct zone *zone;
1433
1434         for_each_populated_zone(zone) {
1435                 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1436
1437                 BUILD_BUG_ON(sizeof(p->vm_stat_diff[0]) != 1);
1438                 /*
1439                  * The fast way of checking if there are any vmstat diffs.
1440                  * This works because the diffs are byte sized items.
1441                  */
1442                 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS))
1443                         return true;
1444
1445         }
1446         return false;
1447 }
1448
1449
1450 /*
1451  * Shepherd worker thread that checks the
1452  * differentials of processors that have their worker
1453  * threads for vm statistics updates disabled because of
1454  * inactivity.
1455  */
1456 static void vmstat_shepherd(struct work_struct *w);
1457
1458 static DECLARE_DELAYED_WORK(shepherd, vmstat_shepherd);
1459
1460 static void vmstat_shepherd(struct work_struct *w)
1461 {
1462         int cpu;
1463
1464         get_online_cpus();
1465         /* Check processors whose vmstat worker threads have been disabled */
1466         for_each_cpu(cpu, cpu_stat_off)
1467                 if (need_update(cpu) &&
1468                         cpumask_test_and_clear_cpu(cpu, cpu_stat_off))
1469
1470                         queue_delayed_work_on(cpu, vmstat_wq,
1471                                 &per_cpu(vmstat_work, cpu), 0);
1472
1473         put_online_cpus();
1474
1475         schedule_delayed_work(&shepherd,
1476                 round_jiffies_relative(sysctl_stat_interval));
1477
1478 }
1479
1480 static void __init start_shepherd_timer(void)
1481 {
1482         int cpu;
1483
1484         for_each_possible_cpu(cpu)
1485                 INIT_DELAYED_WORK(per_cpu_ptr(&vmstat_work, cpu),
1486                         vmstat_update);
1487
1488         if (!alloc_cpumask_var(&cpu_stat_off, GFP_KERNEL))
1489                 BUG();
1490         cpumask_copy(cpu_stat_off, cpu_online_mask);
1491
1492         vmstat_wq = alloc_workqueue("vmstat", WQ_FREEZABLE|WQ_MEM_RECLAIM, 0);
1493         schedule_delayed_work(&shepherd,
1494                 round_jiffies_relative(sysctl_stat_interval));
1495 }
1496
1497 static void vmstat_cpu_dead(int node)
1498 {
1499         int cpu;
1500
1501         get_online_cpus();
1502         for_each_online_cpu(cpu)
1503                 if (cpu_to_node(cpu) == node)
1504                         goto end;
1505
1506         node_clear_state(node, N_CPU);
1507 end:
1508         put_online_cpus();
1509 }
1510
1511 /*
1512  * Use the cpu notifier to insure that the thresholds are recalculated
1513  * when necessary.
1514  */
1515 static int vmstat_cpuup_callback(struct notifier_block *nfb,
1516                 unsigned long action,
1517                 void *hcpu)
1518 {
1519         long cpu = (long)hcpu;
1520
1521         switch (action) {
1522         case CPU_ONLINE:
1523         case CPU_ONLINE_FROZEN:
1524                 refresh_zone_stat_thresholds();
1525                 node_set_state(cpu_to_node(cpu), N_CPU);
1526                 cpumask_set_cpu(cpu, cpu_stat_off);
1527                 break;
1528         case CPU_DOWN_PREPARE:
1529         case CPU_DOWN_PREPARE_FROZEN:
1530                 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1531                 cpumask_clear_cpu(cpu, cpu_stat_off);
1532                 break;
1533         case CPU_DOWN_FAILED:
1534         case CPU_DOWN_FAILED_FROZEN:
1535                 cpumask_set_cpu(cpu, cpu_stat_off);
1536                 break;
1537         case CPU_DEAD:
1538         case CPU_DEAD_FROZEN:
1539                 refresh_zone_stat_thresholds();
1540                 vmstat_cpu_dead(cpu_to_node(cpu));
1541                 break;
1542         default:
1543                 break;
1544         }
1545         return NOTIFY_OK;
1546 }
1547
1548 static struct notifier_block vmstat_notifier =
1549         { &vmstat_cpuup_callback, NULL, 0 };
1550 #endif
1551
1552 static int __init setup_vmstat(void)
1553 {
1554 #ifdef CONFIG_SMP
1555         cpu_notifier_register_begin();
1556         __register_cpu_notifier(&vmstat_notifier);
1557
1558         start_shepherd_timer();
1559         cpu_notifier_register_done();
1560 #endif
1561 #ifdef CONFIG_PROC_FS
1562         proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1563         proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1564         proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1565         proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1566 #endif
1567         return 0;
1568 }
1569 module_init(setup_vmstat)
1570
1571 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1572
1573 /*
1574  * Return an index indicating how much of the available free memory is
1575  * unusable for an allocation of the requested size.
1576  */
1577 static int unusable_free_index(unsigned int order,
1578                                 struct contig_page_info *info)
1579 {
1580         /* No free memory is interpreted as all free memory is unusable */
1581         if (info->free_pages == 0)
1582                 return 1000;
1583
1584         /*
1585          * Index should be a value between 0 and 1. Return a value to 3
1586          * decimal places.
1587          *
1588          * 0 => no fragmentation
1589          * 1 => high fragmentation
1590          */
1591         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1592
1593 }
1594
1595 static void unusable_show_print(struct seq_file *m,
1596                                         pg_data_t *pgdat, struct zone *zone)
1597 {
1598         unsigned int order;
1599         int index;
1600         struct contig_page_info info;
1601
1602         seq_printf(m, "Node %d, zone %8s ",
1603                                 pgdat->node_id,
1604                                 zone->name);
1605         for (order = 0; order < MAX_ORDER; ++order) {
1606                 fill_contig_page_info(zone, order, &info);
1607                 index = unusable_free_index(order, &info);
1608                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1609         }
1610
1611         seq_putc(m, '\n');
1612 }
1613
1614 /*
1615  * Display unusable free space index
1616  *
1617  * The unusable free space index measures how much of the available free
1618  * memory cannot be used to satisfy an allocation of a given size and is a
1619  * value between 0 and 1. The higher the value, the more of free memory is
1620  * unusable and by implication, the worse the external fragmentation is. This
1621  * can be expressed as a percentage by multiplying by 100.
1622  */
1623 static int unusable_show(struct seq_file *m, void *arg)
1624 {
1625         pg_data_t *pgdat = (pg_data_t *)arg;
1626
1627         /* check memoryless node */
1628         if (!node_state(pgdat->node_id, N_MEMORY))
1629                 return 0;
1630
1631         walk_zones_in_node(m, pgdat, unusable_show_print);
1632
1633         return 0;
1634 }
1635
1636 static const struct seq_operations unusable_op = {
1637         .start  = frag_start,
1638         .next   = frag_next,
1639         .stop   = frag_stop,
1640         .show   = unusable_show,
1641 };
1642
1643 static int unusable_open(struct inode *inode, struct file *file)
1644 {
1645         return seq_open(file, &unusable_op);
1646 }
1647
1648 static const struct file_operations unusable_file_ops = {
1649         .open           = unusable_open,
1650         .read           = seq_read,
1651         .llseek         = seq_lseek,
1652         .release        = seq_release,
1653 };
1654
1655 static void extfrag_show_print(struct seq_file *m,
1656                                         pg_data_t *pgdat, struct zone *zone)
1657 {
1658         unsigned int order;
1659         int index;
1660
1661         /* Alloc on stack as interrupts are disabled for zone walk */
1662         struct contig_page_info info;
1663
1664         seq_printf(m, "Node %d, zone %8s ",
1665                                 pgdat->node_id,
1666                                 zone->name);
1667         for (order = 0; order < MAX_ORDER; ++order) {
1668                 fill_contig_page_info(zone, order, &info);
1669                 index = __fragmentation_index(order, &info);
1670                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1671         }
1672
1673         seq_putc(m, '\n');
1674 }
1675
1676 /*
1677  * Display fragmentation index for orders that allocations would fail for
1678  */
1679 static int extfrag_show(struct seq_file *m, void *arg)
1680 {
1681         pg_data_t *pgdat = (pg_data_t *)arg;
1682
1683         walk_zones_in_node(m, pgdat, extfrag_show_print);
1684
1685         return 0;
1686 }
1687
1688 static const struct seq_operations extfrag_op = {
1689         .start  = frag_start,
1690         .next   = frag_next,
1691         .stop   = frag_stop,
1692         .show   = extfrag_show,
1693 };
1694
1695 static int extfrag_open(struct inode *inode, struct file *file)
1696 {
1697         return seq_open(file, &extfrag_op);
1698 }
1699
1700 static const struct file_operations extfrag_file_ops = {
1701         .open           = extfrag_open,
1702         .read           = seq_read,
1703         .llseek         = seq_lseek,
1704         .release        = seq_release,
1705 };
1706
1707 static int __init extfrag_debug_init(void)
1708 {
1709         struct dentry *extfrag_debug_root;
1710
1711         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1712         if (!extfrag_debug_root)
1713                 return -ENOMEM;
1714
1715         if (!debugfs_create_file("unusable_index", 0444,
1716                         extfrag_debug_root, NULL, &unusable_file_ops))
1717                 goto fail;
1718
1719         if (!debugfs_create_file("extfrag_index", 0444,
1720                         extfrag_debug_root, NULL, &extfrag_file_ops))
1721                 goto fail;
1722
1723         return 0;
1724 fail:
1725         debugfs_remove_recursive(extfrag_debug_root);
1726         return -ENOMEM;
1727 }
1728
1729 module_init(extfrag_debug_init);
1730 #endif