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         x = delta + __this_cpu_read(*p);
 230
 231         t = __this_cpu_read(pcp->stat_threshold);
 232
 233         if (unlikely(x > t || x < -t)) {
 234                 zone_page_state_add(x, zone, item);
 235                 x = 0;
 236         }
 237         __this_cpu_write(*p, x);
 238 }
 239 EXPORT_SYMBOL(__mod_zone_page_state);
 240
 241 /*
 242  * Optimized increment and decrement functions.
 243  *
 244  * These are only for a single page and therefore can take a struct page *
 245  * argument instead of struct zone *. This allows the inclusion of the code
 246  * generated for page_zone(page) into the optimized functions.
 247  *
 248  * No overflow check is necessary and therefore the differential can be
 249  * incremented or decremented in place which may allow the compilers to
 250  * generate better code.
 251  * The increment or decrement is known and therefore one boundary check can
 252  * be omitted.
 253  *
 254  * NOTE: These functions are very performance sensitive. Change only
 255  * with care.
 256  *
 257  * Some processors have inc/dec instructions that are atomic vs an interrupt.
 258  * However, the code must first determine the differential location in a zone
 259  * based on the processor number and then inc/dec the counter. There is no
 260  * guarantee without disabling preemption that the processor will not change
 261  * in between and therefore the atomicity vs. interrupt cannot be exploited
 262  * in a useful way here.
 263  */
 264 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
 265 {
 266         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 267         s8 __percpu *p = pcp->vm_stat_diff + item;
 268         s8 v, t;
 269
 270         v = __this_cpu_inc_return(*p);
 271         t = __this_cpu_read(pcp->stat_threshold);
 272         if (unlikely(v > t)) {
 273                 s8 overstep = t >> 1;
 274
 275                 zone_page_state_add(v + overstep, zone, item);
 276                 __this_cpu_write(*p, -overstep);
 277         }
 278 }
 279
 280 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
 281 {
 282         __inc_zone_state(page_zone(page), item);
 283 }
 284 EXPORT_SYMBOL(__inc_zone_page_state);
 285
 286 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
 287 {
 288         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 289         s8 __percpu *p = pcp->vm_stat_diff + item;
 290         s8 v, t;
 291
 292         v = __this_cpu_dec_return(*p);
 293         t = __this_cpu_read(pcp->stat_threshold);
 294         if (unlikely(v < - t)) {
 295                 s8 overstep = t >> 1;
 296
 297                 zone_page_state_add(v - overstep, zone, item);
 298                 __this_cpu_write(*p, overstep);
 299         }
 300 }
 301
 302 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
 303 {
 304         __dec_zone_state(page_zone(page), item);
 305 }
 306 EXPORT_SYMBOL(__dec_zone_page_state);
 307
 308 #ifdef CONFIG_HAVE_CMPXCHG_LOCAL
 309 /*
 310  * If we have cmpxchg_local support then we do not need to incur the overhead
 311  * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
 312  *
 313  * mod_state() modifies the zone counter state through atomic per cpu
 314  * operations.
 315  *
 316  * Overstep mode specifies how overstep should handled:
 317  *     0       No overstepping
 318  *     1       Overstepping half of threshold
 319  *     -1      Overstepping minus half of threshold
 320 */
 321 static inline void mod_state(struct zone *zone,
 322        enum zone_stat_item item, int delta, int overstep_mode)
 323 {
 324         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 325         s8 __percpu *p = pcp->vm_stat_diff + item;
 326         long o, n, t, z;
 327
 328         do {
 329                 z = 0;  /* overflow to zone counters */
 330
 331                 /*
 332                  * The fetching of the stat_threshold is racy. We may apply
 333                  * a counter threshold to the wrong the cpu if we get
 334                  * rescheduled while executing here. However, the next
 335                  * counter update will apply the threshold again and
 336                  * therefore bring the counter under the threshold again.
 337                  *
 338                  * Most of the time the thresholds are the same anyways
 339                  * for all cpus in a zone.
 340                  */
 341                 t = this_cpu_read(pcp->stat_threshold);
 342
 343                 o = this_cpu_read(*p);
 344                 n = delta + o;
 345
 346                 if (n > t || n < -t) {
 347                         int os = overstep_mode * (t >> 1) ;
 348
 349                         /* Overflow must be added to zone counters */
 350                         z = n + os;
 351                         n = -os;
 352                 }
 353         } while (this_cpu_cmpxchg(*p, o, n) != o);
 354
 355         if (z)
 356                 zone_page_state_add(z, zone, item);
 357 }
 358
 359 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 360                                         int delta)
 361 {
 362         mod_state(zone, item, delta, 0);
 363 }
 364 EXPORT_SYMBOL(mod_zone_page_state);
 365
 366 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
 367 {
 368         mod_state(zone, item, 1, 1);
 369 }
 370
 371 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 372 {
 373         mod_state(page_zone(page), item, 1, 1);
 374 }
 375 EXPORT_SYMBOL(inc_zone_page_state);
 376
 377 void dec_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(dec_zone_page_state);
 382 #else
 383 /*
 384  * Use interrupt disable to serialize counter updates
 385  */
 386 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 387                                         int delta)
 388 {
 389         unsigned long flags;
 390
 391         local_irq_save(flags);
 392         __mod_zone_page_state(zone, item, delta);
 393         local_irq_restore(flags);
 394 }
 395 EXPORT_SYMBOL(mod_zone_page_state);
 396
 397 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
 398 {
 399         unsigned long flags;
 400
 401         local_irq_save(flags);
 402         __inc_zone_state(zone, item);
 403         local_irq_restore(flags);
 404 }
 405
 406 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 407 {
 408         unsigned long flags;
 409         struct zone *zone;
 410
 411         zone = page_zone(page);
 412         local_irq_save(flags);
 413         __inc_zone_state(zone, item);
 414         local_irq_restore(flags);
 415 }
 416 EXPORT_SYMBOL(inc_zone_page_state);
 417
 418 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 419 {
 420         unsigned long flags;
 421
 422         local_irq_save(flags);
 423         __dec_zone_page_state(page, item);
 424         local_irq_restore(flags);
 425 }
 426 EXPORT_SYMBOL(dec_zone_page_state);
 427 #endif
 428
 429
 430 /*
 431  * Fold a differential into the global counters.
 432  * Returns the number of counters updated.
 433  */
 434 static int fold_diff(int *diff)
 435 {
 436         int i;
 437         int changes = 0;
 438
 439         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 440                 if (diff[i]) {
 441                         atomic_long_add(diff[i], &vm_stat[i]);
 442                         changes++;
 443         }
 444         return changes;
 445 }
 446
 447 /*
 448  * Update the zone counters for the current cpu.
 449  *
 450  * Note that refresh_cpu_vm_stats strives to only access
 451  * node local memory. The per cpu pagesets on remote zones are placed
 452  * in the memory local to the processor using that pageset. So the
 453  * loop over all zones will access a series of cachelines local to
 454  * the processor.
 455  *
 456  * The call to zone_page_state_add updates the cachelines with the
 457  * statistics in the remote zone struct as well as the global cachelines
 458  * with the global counters. These could cause remote node cache line
 459  * bouncing and will have to be only done when necessary.
 460  *
 461  * The function returns the number of global counters updated.
 462  */
 463 static int refresh_cpu_vm_stats(void)
 464 {
 465         struct zone *zone;
 466         int i;
 467         int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 468         int changes = 0;
 469
 470         for_each_populated_zone(zone) {
 471                 struct per_cpu_pageset __percpu *p = zone->pageset;
 472
 473                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
 474                         int v;
 475
 476                         v = this_cpu_xchg(p->vm_stat_diff[i], 0);
 477                         if (v) {
 478
 479                                 atomic_long_add(v, &zone->vm_stat[i]);
 480                                 global_diff[i] += v;
 481 #ifdef CONFIG_NUMA
 482                                 /* 3 seconds idle till flush */
 483                                 __this_cpu_write(p->expire, 3);
 484 #endif
 485                         }
 486                 }
 487                 cond_resched();
 488 #ifdef CONFIG_NUMA
 489                 /*
 490                  * Deal with draining the remote pageset of this
 491                  * processor
 492                  *
 493                  * Check if there are pages remaining in this pageset
 494                  * if not then there is nothing to expire.
 495                  */
 496                 if (!__this_cpu_read(p->expire) ||
 497                                !__this_cpu_read(p->pcp.count))
 498                         continue;
 499
 500                 /*
 501                  * We never drain zones local to this processor.
 502                  */
 503                 if (zone_to_nid(zone) == numa_node_id()) {
 504                         __this_cpu_write(p->expire, 0);
 505                         continue;
 506                 }
 507
 508                 if (__this_cpu_dec_return(p->expire))
 509                         continue;
 510
 511                 if (__this_cpu_read(p->pcp.count)) {
 512                         drain_zone_pages(zone, this_cpu_ptr(&p->pcp));
 513                         changes++;
 514                 }
 515 #endif
 516         }
 517         changes += fold_diff(global_diff);
 518         return changes;
 519 }
 520
 521 /*
 522  * Fold the data for an offline cpu into the global array.
 523  * There cannot be any access by the offline cpu and therefore
 524  * synchronization is simplified.
 525  */
 526 void cpu_vm_stats_fold(int cpu)
 527 {
 528         struct zone *zone;
 529         int i;
 530         int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 531
 532         for_each_populated_zone(zone) {
 533                 struct per_cpu_pageset *p;
 534
 535                 p = per_cpu_ptr(zone->pageset, cpu);
 536
 537                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 538                         if (p->vm_stat_diff[i]) {
 539                                 int v;
 540
 541                                 v = p->vm_stat_diff[i];
 542                                 p->vm_stat_diff[i] = 0;
 543                                 atomic_long_add(v, &zone->vm_stat[i]);
 544                                 global_diff[i] += v;
 545                         }
 546         }
 547
 548         fold_diff(global_diff);
 549 }
 550
 551 /*
 552  * this is only called if !populated_zone(zone), which implies no other users of
 553  * pset->vm_stat_diff[] exsist.
 554  */
 555 void drain_zonestat(struct zone *zone, struct per_cpu_pageset *pset)
 556 {
 557         int i;
 558
 559         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 560                 if (pset->vm_stat_diff[i]) {
 561                         int v = pset->vm_stat_diff[i];
 562                         pset->vm_stat_diff[i] = 0;
 563                         atomic_long_add(v, &zone->vm_stat[i]);
 564                         atomic_long_add(v, &vm_stat[i]);
 565                 }
 566 }
 567 #endif
 568
 569 #ifdef CONFIG_NUMA
 570 /*
 571  * zonelist = the list of zones passed to the allocator
 572  * z        = the zone from which the allocation occurred.
 573  *
 574  * Must be called with interrupts disabled.
 575  *
 576  * When __GFP_OTHER_NODE is set assume the node of the preferred
 577  * zone is the local node. This is useful for daemons who allocate
 578  * memory on behalf of other processes.
 579  */
 580 void zone_statistics(struct zone *preferred_zone, struct zone *z, gfp_t flags)
 581 {
 582         if (z->zone_pgdat == preferred_zone->zone_pgdat) {
 583                 __inc_zone_state(z, NUMA_HIT);
 584         } else {
 585                 __inc_zone_state(z, NUMA_MISS);
 586                 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
 587         }
 588         if (z->node == ((flags & __GFP_OTHER_NODE) ?
 589                         preferred_zone->node : numa_node_id()))
 590                 __inc_zone_state(z, NUMA_LOCAL);
 591         else
 592                 __inc_zone_state(z, NUMA_OTHER);
 593 }
 594
 595 /*
 596  * Determine the per node value of a stat item.
 597  */
 598 unsigned long node_page_state(int node, enum zone_stat_item item)
 599 {
 600         struct zone *zones = NODE_DATA(node)->node_zones;
 601
 602         return
 603 #ifdef CONFIG_ZONE_DMA
 604                 zone_page_state(&zones[ZONE_DMA], item) +
 605 #endif
 606 #ifdef CONFIG_ZONE_DMA32
 607                 zone_page_state(&zones[ZONE_DMA32], item) +
 608 #endif
 609 #ifdef CONFIG_HIGHMEM
 610                 zone_page_state(&zones[ZONE_HIGHMEM], item) +
 611 #endif
 612                 zone_page_state(&zones[ZONE_NORMAL], item) +
 613                 zone_page_state(&zones[ZONE_MOVABLE], item);
 614 }
 615
 616 #endif
 617
 618 #ifdef CONFIG_COMPACTION
 619
 620 struct contig_page_info {
 621         unsigned long free_pages;
 622         unsigned long free_blocks_total;
 623         unsigned long free_blocks_suitable;
 624 };
 625
 626 /*
 627  * Calculate the number of free pages in a zone, how many contiguous
 628  * pages are free and how many are large enough to satisfy an allocation of
 629  * the target size. Note that this function makes no attempt to estimate
 630  * how many suitable free blocks there *might* be if MOVABLE pages were
 631  * migrated. Calculating that is possible, but expensive and can be
 632  * figured out from userspace
 633  */
 634 static void fill_contig_page_info(struct zone *zone,
 635                                 unsigned int suitable_order,
 636                                 struct contig_page_info *info)
 637 {
 638         unsigned int order;
 639
 640         info->free_pages = 0;
 641         info->free_blocks_total = 0;
 642         info->free_blocks_suitable = 0;
 643
 644         for (order = 0; order < MAX_ORDER; order++) {
 645                 unsigned long blocks;
 646
 647                 /* Count number of free blocks */
 648                 blocks = zone->free_area[order].nr_free;
 649                 info->free_blocks_total += blocks;
 650
 651                 /* Count free base pages */
 652                 info->free_pages += blocks << order;
 653
 654                 /* Count the suitable free blocks */
 655                 if (order >= suitable_order)
 656                         info->free_blocks_suitable += blocks <<
 657                                                 (order - suitable_order);
 658         }
 659 }
 660
 661 /*
 662  * A fragmentation index only makes sense if an allocation of a requested
 663  * size would fail. If that is true, the fragmentation index indicates
 664  * whether external fragmentation or a lack of memory was the problem.
 665  * The value can be used to determine if page reclaim or compaction
 666  * should be used
 667  */
 668 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
 669 {
 670         unsigned long requested = 1UL << order;
 671
 672         if (!info->free_blocks_total)
 673                 return 0;
 674
 675         /* Fragmentation index only makes sense when a request would fail */
 676         if (info->free_blocks_suitable)
 677                 return -1000;
 678
 679         /*
 680          * Index is between 0 and 1 so return within 3 decimal places
 681          *
 682          * 0 => allocation would fail due to lack of memory
 683          * 1 => allocation would fail due to fragmentation
 684          */
 685         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
 686 }
 687
 688 /* Same as __fragmentation index but allocs contig_page_info on stack */
 689 int fragmentation_index(struct zone *zone, unsigned int order)
 690 {
 691         struct contig_page_info info;
 692
 693         fill_contig_page_info(zone, order, &info);
 694         return __fragmentation_index(order, &info);
 695 }
 696 #endif
 697
 698 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || defined(CONFIG_NUMA)
 699 #ifdef CONFIG_ZONE_DMA
 700 #define TEXT_FOR_DMA(xx) xx "_dma",
 701 #else
 702 #define TEXT_FOR_DMA(xx)
 703 #endif
 704
 705 #ifdef CONFIG_ZONE_DMA32
 706 #define TEXT_FOR_DMA32(xx) xx "_dma32",
 707 #else
 708 #define TEXT_FOR_DMA32(xx)
 709 #endif
 710
 711 #ifdef CONFIG_HIGHMEM
 712 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
 713 #else
 714 #define TEXT_FOR_HIGHMEM(xx)
 715 #endif
 716
 717 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
 718                                         TEXT_FOR_HIGHMEM(xx) xx "_movable",
 719
 720 const char * const vmstat_text[] = {
 721         /* enum zone_stat_item countes */
 722         "nr_free_pages",
 723         "nr_alloc_batch",
 724         "nr_inactive_anon",
 725         "nr_active_anon",
 726         "nr_inactive_file",
 727         "nr_active_file",
 728         "nr_unevictable",
 729         "nr_mlock",
 730         "nr_anon_pages",
 731         "nr_mapped",
 732         "nr_file_pages",
 733         "nr_dirty",
 734         "nr_writeback",
 735         "nr_slab_reclaimable",
 736         "nr_slab_unreclaimable",
 737         "nr_page_table_pages",
 738         "nr_kernel_stack",
 739         "nr_unstable",
 740         "nr_bounce",
 741         "nr_vmscan_write",
 742         "nr_vmscan_immediate_reclaim",
 743         "nr_writeback_temp",
 744         "nr_isolated_anon",
 745         "nr_isolated_file",
 746         "nr_shmem",
 747         "nr_dirtied",
 748         "nr_written",
 749         "nr_pages_scanned",
 750
 751 #ifdef CONFIG_NUMA
 752         "numa_hit",
 753         "numa_miss",
 754         "numa_foreign",
 755         "numa_interleave",
 756         "numa_local",
 757         "numa_other",
 758 #endif
 759         "workingset_refault",
 760         "workingset_activate",
 761         "workingset_nodereclaim",
 762         "nr_anon_transparent_hugepages",
 763         "nr_free_cma",
 764
 765         /* enum writeback_stat_item counters */
 766         "nr_dirty_threshold",
 767         "nr_dirty_background_threshold",
 768
 769 #ifdef CONFIG_VM_EVENT_COUNTERS
 770         /* enum vm_event_item counters */
 771         "pgpgin",
 772         "pgpgout",
 773         "pswpin",
 774         "pswpout",
 775
 776         TEXTS_FOR_ZONES("pgalloc")
 777
 778         "pgfree",
 779         "pgactivate",
 780         "pgdeactivate",
 781
 782         "pgfault",
 783         "pgmajfault",
 784         "pglazyfreed",
 785
 786         TEXTS_FOR_ZONES("pgrefill")
 787         TEXTS_FOR_ZONES("pgsteal_kswapd")
 788         TEXTS_FOR_ZONES("pgsteal_direct")
 789         TEXTS_FOR_ZONES("pgscan_kswapd")
 790         TEXTS_FOR_ZONES("pgscan_direct")
 791         "pgscan_direct_throttle",
 792
 793 #ifdef CONFIG_NUMA
 794         "zone_reclaim_failed",
 795 #endif
 796         "pginodesteal",
 797         "slabs_scanned",
 798         "kswapd_inodesteal",
 799         "kswapd_low_wmark_hit_quickly",
 800         "kswapd_high_wmark_hit_quickly",
 801         "pageoutrun",
 802         "allocstall",
 803
 804         "pgrotated",
 805
 806         "drop_pagecache",
 807         "drop_slab",
 808
 809 #ifdef CONFIG_NUMA_BALANCING
 810         "numa_pte_updates",
 811         "numa_huge_pte_updates",
 812         "numa_hint_faults",
 813         "numa_hint_faults_local",
 814         "numa_pages_migrated",
 815 #endif
 816 #ifdef CONFIG_MIGRATION
 817         "pgmigrate_success",
 818         "pgmigrate_fail",
 819 #endif
 820 #ifdef CONFIG_COMPACTION
 821         "compact_migrate_scanned",
 822         "compact_free_scanned",
 823         "compact_isolated",
 824         "compact_stall",
 825         "compact_fail",
 826         "compact_success",
 827 #endif
 828
 829 #ifdef CONFIG_HUGETLB_PAGE
 830         "htlb_buddy_alloc_success",
 831         "htlb_buddy_alloc_fail",
 832 #endif
 833         "unevictable_pgs_culled",
 834         "unevictable_pgs_scanned",
 835         "unevictable_pgs_rescued",
 836         "unevictable_pgs_mlocked",
 837         "unevictable_pgs_munlocked",
 838         "unevictable_pgs_cleared",
 839         "unevictable_pgs_stranded",
 840
 841 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 842         "thp_fault_alloc",
 843         "thp_fault_fallback",
 844         "thp_collapse_alloc",
 845         "thp_collapse_alloc_failed",
 846         "thp_split_page",
 847         "thp_split_page_failed",
 848         "thp_split_pmd",
 849         "thp_zero_page_alloc",
 850         "thp_zero_page_alloc_failed",
 851 #endif
 852 #ifdef CONFIG_MEMORY_BALLOON
 853         "balloon_inflate",
 854         "balloon_deflate",
 855 #ifdef CONFIG_BALLOON_COMPACTION
 856         "balloon_migrate",
 857 #endif
 858 #endif /* CONFIG_MEMORY_BALLOON */
 859 #ifdef CONFIG_DEBUG_TLBFLUSH
 860 #ifdef CONFIG_SMP
 861         "nr_tlb_remote_flush",
 862         "nr_tlb_remote_flush_received",
 863 #endif /* CONFIG_SMP */
 864         "nr_tlb_local_flush_all",
 865         "nr_tlb_local_flush_one",
 866 #endif /* CONFIG_DEBUG_TLBFLUSH */
 867
 868 #ifdef CONFIG_DEBUG_VM_VMACACHE
 869         "vmacache_find_calls",
 870         "vmacache_find_hits",
 871         "vmacache_full_flushes",
 872 #endif
 873 #endif /* CONFIG_VM_EVENTS_COUNTERS */
 874 };
 875 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
 876
 877
 878 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
 879      defined(CONFIG_PROC_FS)
 880 static void *frag_start(struct seq_file *m, loff_t *pos)
 881 {
 882         pg_data_t *pgdat;
 883         loff_t node = *pos;
 884
 885         for (pgdat = first_online_pgdat();
 886              pgdat && node;
 887              pgdat = next_online_pgdat(pgdat))
 888                 --node;
 889
 890         return pgdat;
 891 }
 892
 893 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
 894 {
 895         pg_data_t *pgdat = (pg_data_t *)arg;
 896
 897         (*pos)++;
 898         return next_online_pgdat(pgdat);
 899 }
 900
 901 static void frag_stop(struct seq_file *m, void *arg)
 902 {
 903 }
 904
 905 /* Walk all the zones in a node and print using a callback */
 906 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
 907                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
 908 {
 909         struct zone *zone;
 910         struct zone *node_zones = pgdat->node_zones;
 911         unsigned long flags;
 912
 913         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 914                 if (!populated_zone(zone))
 915                         continue;
 916
 917                 spin_lock_irqsave(&zone->lock, flags);
 918                 print(m, pgdat, zone);
 919                 spin_unlock_irqrestore(&zone->lock, flags);
 920         }
 921 }
 922 #endif
 923
 924 #ifdef CONFIG_PROC_FS
 925 static char * const migratetype_names[MIGRATE_TYPES] = {
 926         "Unmovable",
 927         "Reclaimable",
 928         "Movable",
 929         "HighAtomic",
 930 #ifdef CONFIG_CMA
 931         "CMA",
 932 #endif
 933 #ifdef CONFIG_MEMORY_ISOLATION
 934         "Isolate",
 935 #endif
 936 };
 937
 938 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
 939                                                 struct zone *zone)
 940 {
 941         int order;
 942
 943         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 944         for (order = 0; order < MAX_ORDER; ++order)
 945                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
 946         seq_putc(m, '\n');
 947 }
 948
 949 /*
 950  * This walks the free areas for each zone.
 951  */
 952 static int frag_show(struct seq_file *m, void *arg)
 953 {
 954         pg_data_t *pgdat = (pg_data_t *)arg;
 955         walk_zones_in_node(m, pgdat, frag_show_print);
 956         return 0;
 957 }
 958
 959 static void pagetypeinfo_showfree_print(struct seq_file *m,
 960                                         pg_data_t *pgdat, struct zone *zone)
 961 {
 962         int order, mtype;
 963
 964         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
 965                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
 966                                         pgdat->node_id,
 967                                         zone->name,
 968                                         migratetype_names[mtype]);
 969                 for (order = 0; order < MAX_ORDER; ++order) {
 970                         unsigned long freecount = 0;
 971                         struct free_area *area;
 972                         struct list_head *curr;
 973
 974                         area = &(zone->free_area[order]);
 975
 976                         list_for_each(curr, &area->free_list[mtype])
 977                                 freecount++;
 978                         seq_printf(m, "%6lu ", freecount);
 979                 }
 980                 seq_putc(m, '\n');
 981         }
 982 }
 983
 984 /* Print out the free pages at each order for each migatetype */
 985 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
 986 {
 987         int order;
 988         pg_data_t *pgdat = (pg_data_t *)arg;
 989
 990         /* Print header */
 991         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
 992         for (order = 0; order < MAX_ORDER; ++order)
 993                 seq_printf(m, "%6d ", order);
 994         seq_putc(m, '\n');
 995
 996         walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
 997
 998         return 0;
 999 }
1000
1001 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1002                                         pg_data_t *pgdat, struct zone *zone)
1003 {
1004         int mtype;
1005         unsigned long pfn;
1006         unsigned long start_pfn = zone->zone_start_pfn;
1007         unsigned long end_pfn = zone_end_pfn(zone);
1008         unsigned long count[MIGRATE_TYPES] = { 0, };
1009
1010         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1011                 struct page *page;
1012
1013                 if (!pfn_valid(pfn))
1014                         continue;
1015
1016                 page = pfn_to_page(pfn);
1017
1018                 /* Watch for unexpected holes punched in the memmap */
1019                 if (!memmap_valid_within(pfn, page, zone))
1020                         continue;
1021
1022                 mtype = get_pageblock_migratetype(page);
1023
1024                 if (mtype < MIGRATE_TYPES)
1025                         count[mtype]++;
1026         }
1027
1028         /* Print counts */
1029         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1030         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1031                 seq_printf(m, "%12lu ", count[mtype]);
1032         seq_putc(m, '\n');
1033 }
1034
1035 /* Print out the free pages at each order for each migratetype */
1036 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1037 {
1038         int mtype;
1039         pg_data_t *pgdat = (pg_data_t *)arg;
1040
1041         seq_printf(m, "\n%-23s", "Number of blocks type ");
1042         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1043                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1044         seq_putc(m, '\n');
1045         walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
1046
1047         return 0;
1048 }
1049
1050 #ifdef CONFIG_PAGE_OWNER
1051 static void pagetypeinfo_showmixedcount_print(struct seq_file *m,
1052                                                         pg_data_t *pgdat,
1053                                                         struct zone *zone)
1054 {
1055         struct page *page;
1056         struct page_ext *page_ext;
1057         unsigned long pfn = zone->zone_start_pfn, block_end_pfn;
1058         unsigned long end_pfn = pfn + zone->spanned_pages;
1059         unsigned long count[MIGRATE_TYPES] = { 0, };
1060         int pageblock_mt, page_mt;
1061         int i;
1062
1063         /* Scan block by block. First and last block may be incomplete */
1064         pfn = zone->zone_start_pfn;
1065
1066         /*
1067          * Walk the zone in pageblock_nr_pages steps. If a page block spans
1068          * a zone boundary, it will be double counted between zones. This does
1069          * not matter as the mixed block count will still be correct
1070          */
1071         for (; pfn < end_pfn; ) {
1072                 if (!pfn_valid(pfn)) {
1073                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
1074                         continue;
1075                 }
1076
1077                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
1078                 block_end_pfn = min(block_end_pfn, end_pfn);
1079
1080                 page = pfn_to_page(pfn);
1081                 pageblock_mt = get_pfnblock_migratetype(page, pfn);
1082
1083                 for (; pfn < block_end_pfn; pfn++) {
1084                         if (!pfn_valid_within(pfn))
1085                                 continue;
1086
1087                         page = pfn_to_page(pfn);
1088                         if (PageBuddy(page)) {
1089                                 pfn += (1UL << page_order(page)) - 1;
1090                                 continue;
1091                         }
1092
1093                         if (PageReserved(page))
1094                                 continue;
1095
1096                         page_ext = lookup_page_ext(page);
1097
1098                         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
1099                                 continue;
1100
1101                         page_mt = gfpflags_to_migratetype(page_ext->gfp_mask);
1102                         if (pageblock_mt != page_mt) {
1103                                 if (is_migrate_cma(pageblock_mt))
1104                                         count[MIGRATE_MOVABLE]++;
1105                                 else
1106                                         count[pageblock_mt]++;
1107
1108                                 pfn = block_end_pfn;
1109                                 break;
1110                         }
1111                         pfn += (1UL << page_ext->order) - 1;
1112                 }
1113         }
1114
1115         /* Print counts */
1116         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1117         for (i = 0; i < MIGRATE_TYPES; i++)
1118                 seq_printf(m, "%12lu ", count[i]);
1119         seq_putc(m, '\n');
1120 }
1121 #endif /* CONFIG_PAGE_OWNER */
1122
1123 /*
1124  * Print out the number of pageblocks for each migratetype that contain pages
1125  * of other types. This gives an indication of how well fallbacks are being
1126  * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1127  * to determine what is going on
1128  */
1129 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1130 {
1131 #ifdef CONFIG_PAGE_OWNER
1132         int mtype;
1133
1134         if (!page_owner_inited)
1135                 return;
1136
1137         drain_all_pages(NULL);
1138
1139         seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1140         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1141                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1142         seq_putc(m, '\n');
1143
1144         walk_zones_in_node(m, pgdat, pagetypeinfo_showmixedcount_print);
1145 #endif /* CONFIG_PAGE_OWNER */
1146 }
1147
1148 /*
1149  * This prints out statistics in relation to grouping pages by mobility.
1150  * It is expensive to collect so do not constantly read the file.
1151  */
1152 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1153 {
1154         pg_data_t *pgdat = (pg_data_t *)arg;
1155
1156         /* check memoryless node */
1157         if (!node_state(pgdat->node_id, N_MEMORY))
1158                 return 0;
1159
1160         seq_printf(m, "Page block order: %d\n", pageblock_order);
1161         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
1162         seq_putc(m, '\n');
1163         pagetypeinfo_showfree(m, pgdat);
1164         pagetypeinfo_showblockcount(m, pgdat);
1165         pagetypeinfo_showmixedcount(m, pgdat);
1166
1167         return 0;
1168 }
1169
1170 static const struct seq_operations fragmentation_op = {
1171         .start  = frag_start,
1172         .next   = frag_next,
1173         .stop   = frag_stop,
1174         .show   = frag_show,
1175 };
1176
1177 static int fragmentation_open(struct inode *inode, struct file *file)
1178 {
1179         return seq_open(file, &fragmentation_op);
1180 }
1181
1182 static const struct file_operations fragmentation_file_operations = {
1183         .open           = fragmentation_open,
1184         .read           = seq_read,
1185         .llseek         = seq_lseek,
1186         .release        = seq_release,
1187 };
1188
1189 static const struct seq_operations pagetypeinfo_op = {
1190         .start  = frag_start,
1191         .next   = frag_next,
1192         .stop   = frag_stop,
1193         .show   = pagetypeinfo_show,
1194 };
1195
1196 static int pagetypeinfo_open(struct inode *inode, struct file *file)
1197 {
1198         return seq_open(file, &pagetypeinfo_op);
1199 }
1200
1201 static const struct file_operations pagetypeinfo_file_ops = {
1202         .open           = pagetypeinfo_open,
1203         .read           = seq_read,
1204         .llseek         = seq_lseek,
1205         .release        = seq_release,
1206 };
1207
1208 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1209                                                         struct zone *zone)
1210 {
1211         int i;
1212         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1213         seq_printf(m,
1214                    "\n  pages free     %lu"
1215                    "\n        min      %lu"
1216                    "\n        low      %lu"
1217                    "\n        high     %lu"
1218                    "\n        scanned  %lu"
1219                    "\n        spanned  %lu"
1220                    "\n        present  %lu"
1221                    "\n        managed  %lu",
1222                    zone_page_state(zone, NR_FREE_PAGES),
1223                    min_wmark_pages(zone),
1224                    low_wmark_pages(zone),
1225                    high_wmark_pages(zone),
1226                    zone_page_state(zone, NR_PAGES_SCANNED),
1227                    zone->spanned_pages,
1228                    zone->present_pages,
1229                    zone->managed_pages);
1230
1231         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1232                 seq_printf(m, "\n    %-12s %lu", vmstat_text[i],
1233                                 zone_page_state(zone, i));
1234
1235         seq_printf(m,
1236                    "\n        protection: (%ld",
1237                    zone->lowmem_reserve[0]);
1238         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1239                 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1240         seq_printf(m,
1241                    ")"
1242                    "\n  pagesets");
1243         for_each_online_cpu(i) {
1244                 struct per_cpu_pageset *pageset;
1245
1246                 pageset = per_cpu_ptr(zone->pageset, i);
1247                 seq_printf(m,
1248                            "\n    cpu: %i"
1249                            "\n              count: %i"
1250                            "\n              high:  %i"
1251                            "\n              batch: %i",
1252                            i,
1253                            pageset->pcp.count,
1254                            pageset->pcp.high,
1255                            pageset->pcp.batch);
1256 #ifdef CONFIG_SMP
1257                 seq_printf(m, "\n  vm stats threshold: %d",
1258                                 pageset->stat_threshold);
1259 #endif
1260         }
1261         seq_printf(m,
1262                    "\n  all_unreclaimable: %u"
1263                    "\n  start_pfn:         %lu"
1264                    "\n  inactive_ratio:    %u",
1265                    !zone_reclaimable(zone),
1266                    zone->zone_start_pfn,
1267                    zone->inactive_ratio);
1268         seq_putc(m, '\n');
1269 }
1270
1271 /*
1272  * Output information about zones in @pgdat.
1273  */
1274 static int zoneinfo_show(struct seq_file *m, void *arg)
1275 {
1276         pg_data_t *pgdat = (pg_data_t *)arg;
1277         walk_zones_in_node(m, pgdat, zoneinfo_show_print);
1278         return 0;
1279 }
1280
1281 static const struct seq_operations zoneinfo_op = {
1282         .start  = frag_start, /* iterate over all zones. The same as in
1283                                * fragmentation. */
1284         .next   = frag_next,
1285         .stop   = frag_stop,
1286         .show   = zoneinfo_show,
1287 };
1288
1289 static int zoneinfo_open(struct inode *inode, struct file *file)
1290 {
1291         return seq_open(file, &zoneinfo_op);
1292 }
1293
1294 static const struct file_operations proc_zoneinfo_file_operations = {
1295         .open           = zoneinfo_open,
1296         .read           = seq_read,
1297         .llseek         = seq_lseek,
1298         .release        = seq_release,
1299 };
1300
1301 enum writeback_stat_item {
1302         NR_DIRTY_THRESHOLD,
1303         NR_DIRTY_BG_THRESHOLD,
1304         NR_VM_WRITEBACK_STAT_ITEMS,
1305 };
1306
1307 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1308 {
1309         unsigned long *v;
1310         int i, stat_items_size;
1311
1312         if (*pos >= ARRAY_SIZE(vmstat_text))
1313                 return NULL;
1314         stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1315                           NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1316
1317 #ifdef CONFIG_VM_EVENT_COUNTERS
1318         stat_items_size += sizeof(struct vm_event_state);
1319 #endif
1320
1321         v = kmalloc(stat_items_size, GFP_KERNEL);
1322         m->private = v;
1323         if (!v)
1324                 return ERR_PTR(-ENOMEM);
1325         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1326                 v[i] = global_page_state(i);
1327         v += NR_VM_ZONE_STAT_ITEMS;
1328
1329         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1330                             v + NR_DIRTY_THRESHOLD);
1331         v += NR_VM_WRITEBACK_STAT_ITEMS;
1332
1333 #ifdef CONFIG_VM_EVENT_COUNTERS
1334         all_vm_events(v);
1335         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1336         v[PGPGOUT] /= 2;
1337 #endif
1338         return (unsigned long *)m->private + *pos;
1339 }
1340
1341 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1342 {
1343         (*pos)++;
1344         if (*pos >= ARRAY_SIZE(vmstat_text))
1345                 return NULL;
1346         return (unsigned long *)m->private + *pos;
1347 }
1348
1349 static int vmstat_show(struct seq_file *m, void *arg)
1350 {
1351         unsigned long *l = arg;
1352         unsigned long off = l - (unsigned long *)m->private;
1353
1354         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1355         return 0;
1356 }
1357
1358 static void vmstat_stop(struct seq_file *m, void *arg)
1359 {
1360         kfree(m->private);
1361         m->private = NULL;
1362 }
1363
1364 static const struct seq_operations vmstat_op = {
1365         .start  = vmstat_start,
1366         .next   = vmstat_next,
1367         .stop   = vmstat_stop,
1368         .show   = vmstat_show,
1369 };
1370
1371 static int vmstat_open(struct inode *inode, struct file *file)
1372 {
1373         return seq_open(file, &vmstat_op);
1374 }
1375
1376 static const struct file_operations proc_vmstat_file_operations = {
1377         .open           = vmstat_open,
1378         .read           = seq_read,
1379         .llseek         = seq_lseek,
1380         .release        = seq_release,
1381 };
1382 #endif /* CONFIG_PROC_FS */
1383
1384 #ifdef CONFIG_SMP
1385 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1386 int sysctl_stat_interval __read_mostly = HZ;
1387 static cpumask_var_t cpu_stat_off;
1388
1389 static void vmstat_update(struct work_struct *w)
1390 {
1391         if (refresh_cpu_vm_stats()) {
1392                 /*
1393                  * Counters were updated so we expect more updates
1394                  * to occur in the future. Keep on running the
1395                  * update worker thread.
1396                  */
1397                 schedule_delayed_work_on(smp_processor_id(),
1398                         this_cpu_ptr(&vmstat_work),
1399                         round_jiffies_relative(sysctl_stat_interval));
1400         } else {
1401                 /*
1402                  * We did not update any counters so the app may be in
1403                  * a mode where it does not cause counter updates.
1404                  * We may be uselessly running vmstat_update.
1405                  * Defer the checking for differentials to the
1406                  * shepherd thread on a different processor.
1407                  */
1408                 int r;
1409                 /*
1410                  * Shepherd work thread does not race since it never
1411                  * changes the bit if its zero but the cpu
1412                  * online / off line code may race if
1413                  * worker threads are still allowed during
1414                  * shutdown / startup.
1415                  */
1416                 r = cpumask_test_and_set_cpu(smp_processor_id(),
1417                         cpu_stat_off);
1418                 VM_BUG_ON(r);
1419         }
1420 }
1421
1422 /*
1423  * Check if the diffs for a certain cpu indicate that
1424  * an update is needed.
1425  */
1426 static bool need_update(int cpu)
1427 {
1428         struct zone *zone;
1429
1430         for_each_populated_zone(zone) {
1431                 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1432
1433                 BUILD_BUG_ON(sizeof(p->vm_stat_diff[0]) != 1);
1434                 /*
1435                  * The fast way of checking if there are any vmstat diffs.
1436                  * This works because the diffs are byte sized items.
1437                  */
1438                 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS))
1439                         return true;
1440
1441         }
1442         return false;
1443 }
1444
1445
1446 /*
1447  * Shepherd worker thread that checks the
1448  * differentials of processors that have their worker
1449  * threads for vm statistics updates disabled because of
1450  * inactivity.
1451  */
1452 static void vmstat_shepherd(struct work_struct *w);
1453
1454 static DECLARE_DELAYED_WORK(shepherd, vmstat_shepherd);
1455
1456 static void vmstat_shepherd(struct work_struct *w)
1457 {
1458         int cpu;
1459
1460         get_online_cpus();
1461         /* Check processors whose vmstat worker threads have been disabled */
1462         for_each_cpu(cpu, cpu_stat_off)
1463                 if (need_update(cpu) &&
1464                         cpumask_test_and_clear_cpu(cpu, cpu_stat_off))
1465
1466                         schedule_delayed_work_on(cpu,
1467                                 &per_cpu(vmstat_work, cpu), 0);
1468
1469         put_online_cpus();
1470
1471         schedule_delayed_work(&shepherd,
1472                 round_jiffies_relative(sysctl_stat_interval));
1473
1474 }
1475
1476 static void __init start_shepherd_timer(void)
1477 {
1478         int cpu;
1479
1480         for_each_possible_cpu(cpu)
1481                 INIT_DELAYED_WORK(per_cpu_ptr(&vmstat_work, cpu),
1482                         vmstat_update);
1483
1484         if (!alloc_cpumask_var(&cpu_stat_off, GFP_KERNEL))
1485                 BUG();
1486         cpumask_copy(cpu_stat_off, cpu_online_mask);
1487
1488         schedule_delayed_work(&shepherd,
1489                 round_jiffies_relative(sysctl_stat_interval));
1490 }
1491
1492 static void vmstat_cpu_dead(int node)
1493 {
1494         int cpu;
1495
1496         get_online_cpus();
1497         for_each_online_cpu(cpu)
1498                 if (cpu_to_node(cpu) == node)
1499                         goto end;
1500
1501         node_clear_state(node, N_CPU);
1502 end:
1503         put_online_cpus();
1504 }
1505
1506 /*
1507  * Use the cpu notifier to insure that the thresholds are recalculated
1508  * when necessary.
1509  */
1510 static int vmstat_cpuup_callback(struct notifier_block *nfb,
1511                 unsigned long action,
1512                 void *hcpu)
1513 {
1514         long cpu = (long)hcpu;
1515
1516         switch (action) {
1517         case CPU_ONLINE:
1518         case CPU_ONLINE_FROZEN:
1519                 refresh_zone_stat_thresholds();
1520                 node_set_state(cpu_to_node(cpu), N_CPU);
1521                 cpumask_set_cpu(cpu, cpu_stat_off);
1522                 break;
1523         case CPU_DOWN_PREPARE:
1524         case CPU_DOWN_PREPARE_FROZEN:
1525                 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1526                 cpumask_clear_cpu(cpu, cpu_stat_off);
1527                 break;
1528         case CPU_DOWN_FAILED:
1529         case CPU_DOWN_FAILED_FROZEN:
1530                 cpumask_set_cpu(cpu, cpu_stat_off);
1531                 break;
1532         case CPU_DEAD:
1533         case CPU_DEAD_FROZEN:
1534                 refresh_zone_stat_thresholds();
1535                 vmstat_cpu_dead(cpu_to_node(cpu));
1536                 break;
1537         default:
1538                 break;
1539         }
1540         return NOTIFY_OK;
1541 }
1542
1543 static struct notifier_block vmstat_notifier =
1544         { &vmstat_cpuup_callback, NULL, 0 };
1545 #endif
1546
1547 static int __init setup_vmstat(void)
1548 {
1549 #ifdef CONFIG_SMP
1550         cpu_notifier_register_begin();
1551         __register_cpu_notifier(&vmstat_notifier);
1552
1553         start_shepherd_timer();
1554         cpu_notifier_register_done();
1555 #endif
1556 #ifdef CONFIG_PROC_FS
1557         proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1558         proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1559         proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1560         proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1561 #endif
1562         return 0;
1563 }
1564 module_init(setup_vmstat)
1565
1566 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1567
1568 /*
1569  * Return an index indicating how much of the available free memory is
1570  * unusable for an allocation of the requested size.
1571  */
1572 static int unusable_free_index(unsigned int order,
1573                                 struct contig_page_info *info)
1574 {
1575         /* No free memory is interpreted as all free memory is unusable */
1576         if (info->free_pages == 0)
1577                 return 1000;
1578
1579         /*
1580          * Index should be a value between 0 and 1. Return a value to 3
1581          * decimal places.
1582          *
1583          * 0 => no fragmentation
1584          * 1 => high fragmentation
1585          */
1586         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1587
1588 }
1589
1590 static void unusable_show_print(struct seq_file *m,
1591                                         pg_data_t *pgdat, struct zone *zone)
1592 {
1593         unsigned int order;
1594         int index;
1595         struct contig_page_info info;
1596
1597         seq_printf(m, "Node %d, zone %8s ",
1598                                 pgdat->node_id,
1599                                 zone->name);
1600         for (order = 0; order < MAX_ORDER; ++order) {
1601                 fill_contig_page_info(zone, order, &info);
1602                 index = unusable_free_index(order, &info);
1603                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1604         }
1605
1606         seq_putc(m, '\n');
1607 }
1608
1609 /*
1610  * Display unusable free space index
1611  *
1612  * The unusable free space index measures how much of the available free
1613  * memory cannot be used to satisfy an allocation of a given size and is a
1614  * value between 0 and 1. The higher the value, the more of free memory is
1615  * unusable and by implication, the worse the external fragmentation is. This
1616  * can be expressed as a percentage by multiplying by 100.
1617  */
1618 static int unusable_show(struct seq_file *m, void *arg)
1619 {
1620         pg_data_t *pgdat = (pg_data_t *)arg;
1621
1622         /* check memoryless node */
1623         if (!node_state(pgdat->node_id, N_MEMORY))
1624                 return 0;
1625
1626         walk_zones_in_node(m, pgdat, unusable_show_print);
1627
1628         return 0;
1629 }
1630
1631 static const struct seq_operations unusable_op = {
1632         .start  = frag_start,
1633         .next   = frag_next,
1634         .stop   = frag_stop,
1635         .show   = unusable_show,
1636 };
1637
1638 static int unusable_open(struct inode *inode, struct file *file)
1639 {
1640         return seq_open(file, &unusable_op);
1641 }
1642
1643 static const struct file_operations unusable_file_ops = {
1644         .open           = unusable_open,
1645         .read           = seq_read,
1646         .llseek         = seq_lseek,
1647         .release        = seq_release,
1648 };
1649
1650 static void extfrag_show_print(struct seq_file *m,
1651                                         pg_data_t *pgdat, struct zone *zone)
1652 {
1653         unsigned int order;
1654         int index;
1655
1656         /* Alloc on stack as interrupts are disabled for zone walk */
1657         struct contig_page_info info;
1658
1659         seq_printf(m, "Node %d, zone %8s ",
1660                                 pgdat->node_id,
1661                                 zone->name);
1662         for (order = 0; order < MAX_ORDER; ++order) {
1663                 fill_contig_page_info(zone, order, &info);
1664                 index = __fragmentation_index(order, &info);
1665                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1666         }
1667
1668         seq_putc(m, '\n');
1669 }
1670
1671 /*
1672  * Display fragmentation index for orders that allocations would fail for
1673  */
1674 static int extfrag_show(struct seq_file *m, void *arg)
1675 {
1676         pg_data_t *pgdat = (pg_data_t *)arg;
1677
1678         walk_zones_in_node(m, pgdat, extfrag_show_print);
1679
1680         return 0;
1681 }
1682
1683 static const struct seq_operations extfrag_op = {
1684         .start  = frag_start,
1685         .next   = frag_next,
1686         .stop   = frag_stop,
1687         .show   = extfrag_show,
1688 };
1689
1690 static int extfrag_open(struct inode *inode, struct file *file)
1691 {
1692         return seq_open(file, &extfrag_op);
1693 }
1694
1695 static const struct file_operations extfrag_file_ops = {
1696         .open           = extfrag_open,
1697         .read           = seq_read,
1698         .llseek         = seq_lseek,
1699         .release        = seq_release,
1700 };
1701
1702 static int __init extfrag_debug_init(void)
1703 {
1704         struct dentry *extfrag_debug_root;
1705
1706         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1707         if (!extfrag_debug_root)
1708                 return -ENOMEM;
1709
1710         if (!debugfs_create_file("unusable_index", 0444,
1711                         extfrag_debug_root, NULL, &unusable_file_ops))
1712                 goto fail;
1713
1714         if (!debugfs_create_file("extfrag_index", 0444,
1715                         extfrag_debug_root, NULL, &extfrag_file_ops))
1716                 goto fail;
1717
1718         return 0;
1719 fail:
1720         debugfs_remove_recursive(extfrag_debug_root);
1721         return -ENOMEM;
1722 }
1723
1724 module_init(extfrag_debug_init);
1725 #endif