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         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, enum zone_stat_item item,
 322                              long 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                          long 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                          long 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(bool do_pagesets)
 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 #ifdef CONFIG_NUMA
 488                 if (do_pagesets) {
 489                         cond_resched();
 490                         /*
 491                          * Deal with draining the remote pageset of this
 492                          * processor
 493                          *
 494                          * Check if there are pages remaining in this pageset
 495                          * if not then there is nothing to expire.
 496                          */
 497                         if (!__this_cpu_read(p->expire) ||
 498                                !__this_cpu_read(p->pcp.count))
 499                                 continue;
 500
 501                         /*
 502                          * We never drain zones local to this processor.
 503                          */
 504                         if (zone_to_nid(zone) == numa_node_id()) {
 505                                 __this_cpu_write(p->expire, 0);
 506                                 continue;
 507                         }
 508
 509                         if (__this_cpu_dec_return(p->expire))
 510                                 continue;
 511
 512                         if (__this_cpu_read(p->pcp.count)) {
 513                                 drain_zone_pages(zone, this_cpu_ptr(&p->pcp));
 514                                 changes++;
 515                         }
 516                 }
 517 #endif
 518         }
 519         changes += fold_diff(global_diff);
 520         return changes;
 521 }
 522
 523 /*
 524  * Fold the data for an offline cpu into the global array.
 525  * There cannot be any access by the offline cpu and therefore
 526  * synchronization is simplified.
 527  */
 528 void cpu_vm_stats_fold(int cpu)
 529 {
 530         struct zone *zone;
 531         int i;
 532         int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 533
 534         for_each_populated_zone(zone) {
 535                 struct per_cpu_pageset *p;
 536
 537                 p = per_cpu_ptr(zone->pageset, cpu);
 538
 539                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 540                         if (p->vm_stat_diff[i]) {
 541                                 int v;
 542
 543                                 v = p->vm_stat_diff[i];
 544                                 p->vm_stat_diff[i] = 0;
 545                                 atomic_long_add(v, &zone->vm_stat[i]);
 546                                 global_diff[i] += v;
 547                         }
 548         }
 549
 550         fold_diff(global_diff);
 551 }
 552
 553 /*
 554  * this is only called if !populated_zone(zone), which implies no other users of
 555  * pset->vm_stat_diff[] exsist.
 556  */
 557 void drain_zonestat(struct zone *zone, struct per_cpu_pageset *pset)
 558 {
 559         int i;
 560
 561         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 562                 if (pset->vm_stat_diff[i]) {
 563                         int v = pset->vm_stat_diff[i];
 564                         pset->vm_stat_diff[i] = 0;
 565                         atomic_long_add(v, &zone->vm_stat[i]);
 566                         atomic_long_add(v, &vm_stat[i]);
 567                 }
 568 }
 569 #endif
 570
 571 #ifdef CONFIG_NUMA
 572 /*
 573  * zonelist = the list of zones passed to the allocator
 574  * z        = the zone from which the allocation occurred.
 575  *
 576  * Must be called with interrupts disabled.
 577  *
 578  * When __GFP_OTHER_NODE is set assume the node of the preferred
 579  * zone is the local node. This is useful for daemons who allocate
 580  * memory on behalf of other processes.
 581  */
 582 void zone_statistics(struct zone *preferred_zone, struct zone *z, gfp_t flags)
 583 {
 584         if (z->zone_pgdat == preferred_zone->zone_pgdat) {
 585                 __inc_zone_state(z, NUMA_HIT);
 586         } else {
 587                 __inc_zone_state(z, NUMA_MISS);
 588                 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
 589         }
 590         if (z->node == ((flags & __GFP_OTHER_NODE) ?
 591                         preferred_zone->node : numa_node_id()))
 592                 __inc_zone_state(z, NUMA_LOCAL);
 593         else
 594                 __inc_zone_state(z, NUMA_OTHER);
 595 }
 596
 597 /*
 598  * Determine the per node value of a stat item.
 599  */
 600 unsigned long node_page_state(int node, enum zone_stat_item item)
 601 {
 602         struct zone *zones = NODE_DATA(node)->node_zones;
 603         int i;
 604         unsigned long count = 0;
 605
 606         for (i = 0; i < MAX_NR_ZONES; i++)
 607                 count += zone_page_state(zones + i, item);
 608
 609         return count;
 610 }
 611
 612 #endif
 613
 614 #ifdef CONFIG_COMPACTION
 615
 616 struct contig_page_info {
 617         unsigned long free_pages;
 618         unsigned long free_blocks_total;
 619         unsigned long free_blocks_suitable;
 620 };
 621
 622 /*
 623  * Calculate the number of free pages in a zone, how many contiguous
 624  * pages are free and how many are large enough to satisfy an allocation of
 625  * the target size. Note that this function makes no attempt to estimate
 626  * how many suitable free blocks there *might* be if MOVABLE pages were
 627  * migrated. Calculating that is possible, but expensive and can be
 628  * figured out from userspace
 629  */
 630 static void fill_contig_page_info(struct zone *zone,
 631                                 unsigned int suitable_order,
 632                                 struct contig_page_info *info)
 633 {
 634         unsigned int order;
 635
 636         info->free_pages = 0;
 637         info->free_blocks_total = 0;
 638         info->free_blocks_suitable = 0;
 639
 640         for (order = 0; order < MAX_ORDER; order++) {
 641                 unsigned long blocks;
 642
 643                 /* Count number of free blocks */
 644                 blocks = zone->free_area[order].nr_free;
 645                 info->free_blocks_total += blocks;
 646
 647                 /* Count free base pages */
 648                 info->free_pages += blocks << order;
 649
 650                 /* Count the suitable free blocks */
 651                 if (order >= suitable_order)
 652                         info->free_blocks_suitable += blocks <<
 653                                                 (order - suitable_order);
 654         }
 655 }
 656
 657 /*
 658  * A fragmentation index only makes sense if an allocation of a requested
 659  * size would fail. If that is true, the fragmentation index indicates
 660  * whether external fragmentation or a lack of memory was the problem.
 661  * The value can be used to determine if page reclaim or compaction
 662  * should be used
 663  */
 664 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
 665 {
 666         unsigned long requested = 1UL << order;
 667
 668         if (!info->free_blocks_total)
 669                 return 0;
 670
 671         /* Fragmentation index only makes sense when a request would fail */
 672         if (info->free_blocks_suitable)
 673                 return -1000;
 674
 675         /*
 676          * Index is between 0 and 1 so return within 3 decimal places
 677          *
 678          * 0 => allocation would fail due to lack of memory
 679          * 1 => allocation would fail due to fragmentation
 680          */
 681         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
 682 }
 683
 684 /* Same as __fragmentation index but allocs contig_page_info on stack */
 685 int fragmentation_index(struct zone *zone, unsigned int order)
 686 {
 687         struct contig_page_info info;
 688
 689         fill_contig_page_info(zone, order, &info);
 690         return __fragmentation_index(order, &info);
 691 }
 692 #endif
 693
 694 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || defined(CONFIG_NUMA)
 695 #ifdef CONFIG_ZONE_DMA
 696 #define TEXT_FOR_DMA(xx) xx "_dma",
 697 #else
 698 #define TEXT_FOR_DMA(xx)
 699 #endif
 700
 701 #ifdef CONFIG_ZONE_DMA32
 702 #define TEXT_FOR_DMA32(xx) xx "_dma32",
 703 #else
 704 #define TEXT_FOR_DMA32(xx)
 705 #endif
 706
 707 #ifdef CONFIG_HIGHMEM
 708 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
 709 #else
 710 #define TEXT_FOR_HIGHMEM(xx)
 711 #endif
 712
 713 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
 714                                         TEXT_FOR_HIGHMEM(xx) xx "_movable",
 715
 716 const char * const vmstat_text[] = {
 717         /* enum zone_stat_item countes */
 718         "nr_free_pages",
 719         "nr_alloc_batch",
 720         "nr_inactive_anon",
 721         "nr_active_anon",
 722         "nr_inactive_file",
 723         "nr_active_file",
 724         "nr_unevictable",
 725         "nr_mlock",
 726         "nr_anon_pages",
 727         "nr_mapped",
 728         "nr_file_pages",
 729         "nr_dirty",
 730         "nr_writeback",
 731         "nr_slab_reclaimable",
 732         "nr_slab_unreclaimable",
 733         "nr_page_table_pages",
 734         "nr_kernel_stack",
 735         "nr_unstable",
 736         "nr_bounce",
 737         "nr_vmscan_write",
 738         "nr_vmscan_immediate_reclaim",
 739         "nr_writeback_temp",
 740         "nr_isolated_anon",
 741         "nr_isolated_file",
 742         "nr_shmem",
 743         "nr_dirtied",
 744         "nr_written",
 745         "nr_pages_scanned",
 746
 747 #ifdef CONFIG_NUMA
 748         "numa_hit",
 749         "numa_miss",
 750         "numa_foreign",
 751         "numa_interleave",
 752         "numa_local",
 753         "numa_other",
 754 #endif
 755         "workingset_refault",
 756         "workingset_activate",
 757         "workingset_nodereclaim",
 758         "nr_anon_transparent_hugepages",
 759         "nr_free_cma",
 760
 761         /* enum writeback_stat_item counters */
 762         "nr_dirty_threshold",
 763         "nr_dirty_background_threshold",
 764
 765 #ifdef CONFIG_VM_EVENT_COUNTERS
 766         /* enum vm_event_item counters */
 767         "pgpgin",
 768         "pgpgout",
 769         "pswpin",
 770         "pswpout",
 771
 772         TEXTS_FOR_ZONES("pgalloc")
 773
 774         "pgfree",
 775         "pgactivate",
 776         "pgdeactivate",
 777
 778         "pgfault",
 779         "pgmajfault",
 780         "pglazyfreed",
 781
 782         TEXTS_FOR_ZONES("pgrefill")
 783         TEXTS_FOR_ZONES("pgsteal_kswapd")
 784         TEXTS_FOR_ZONES("pgsteal_direct")
 785         TEXTS_FOR_ZONES("pgscan_kswapd")
 786         TEXTS_FOR_ZONES("pgscan_direct")
 787         "pgscan_direct_throttle",
 788
 789 #ifdef CONFIG_NUMA
 790         "zone_reclaim_failed",
 791 #endif
 792         "pginodesteal",
 793         "slabs_scanned",
 794         "kswapd_inodesteal",
 795         "kswapd_low_wmark_hit_quickly",
 796         "kswapd_high_wmark_hit_quickly",
 797         "pageoutrun",
 798         "allocstall",
 799
 800         "pgrotated",
 801
 802         "drop_pagecache",
 803         "drop_slab",
 804
 805 #ifdef CONFIG_NUMA_BALANCING
 806         "numa_pte_updates",
 807         "numa_huge_pte_updates",
 808         "numa_hint_faults",
 809         "numa_hint_faults_local",
 810         "numa_pages_migrated",
 811 #endif
 812 #ifdef CONFIG_MIGRATION
 813         "pgmigrate_success",
 814         "pgmigrate_fail",
 815 #endif
 816 #ifdef CONFIG_COMPACTION
 817         "compact_migrate_scanned",
 818         "compact_free_scanned",
 819         "compact_isolated",
 820         "compact_stall",
 821         "compact_fail",
 822         "compact_success",
 823         "compact_daemon_wake",
 824 #endif
 825
 826 #ifdef CONFIG_HUGETLB_PAGE
 827         "htlb_buddy_alloc_success",
 828         "htlb_buddy_alloc_fail",
 829 #endif
 830         "unevictable_pgs_culled",
 831         "unevictable_pgs_scanned",
 832         "unevictable_pgs_rescued",
 833         "unevictable_pgs_mlocked",
 834         "unevictable_pgs_munlocked",
 835         "unevictable_pgs_cleared",
 836         "unevictable_pgs_stranded",
 837
 838 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 839         "thp_fault_alloc",
 840         "thp_fault_fallback",
 841         "thp_collapse_alloc",
 842         "thp_collapse_alloc_failed",
 843         "thp_split_page",
 844         "thp_split_page_failed",
 845         "thp_deferred_split_page",
 846         "thp_split_pmd",
 847         "thp_zero_page_alloc",
 848         "thp_zero_page_alloc_failed",
 849 #endif
 850 #ifdef CONFIG_MEMORY_BALLOON
 851         "balloon_inflate",
 852         "balloon_deflate",
 853 #ifdef CONFIG_BALLOON_COMPACTION
 854         "balloon_migrate",
 855 #endif
 856 #endif /* CONFIG_MEMORY_BALLOON */
 857 #ifdef CONFIG_DEBUG_TLBFLUSH
 858 #ifdef CONFIG_SMP
 859         "nr_tlb_remote_flush",
 860         "nr_tlb_remote_flush_received",
 861 #endif /* CONFIG_SMP */
 862         "nr_tlb_local_flush_all",
 863         "nr_tlb_local_flush_one",
 864 #endif /* CONFIG_DEBUG_TLBFLUSH */
 865
 866 #ifdef CONFIG_DEBUG_VM_VMACACHE
 867         "vmacache_find_calls",
 868         "vmacache_find_hits",
 869         "vmacache_full_flushes",
 870 #endif
 871 #endif /* CONFIG_VM_EVENTS_COUNTERS */
 872 };
 873 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
 874
 875
 876 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
 877      defined(CONFIG_PROC_FS)
 878 static void *frag_start(struct seq_file *m, loff_t *pos)
 879 {
 880         pg_data_t *pgdat;
 881         loff_t node = *pos;
 882
 883         for (pgdat = first_online_pgdat();
 884              pgdat && node;
 885              pgdat = next_online_pgdat(pgdat))
 886                 --node;
 887
 888         return pgdat;
 889 }
 890
 891 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
 892 {
 893         pg_data_t *pgdat = (pg_data_t *)arg;
 894
 895         (*pos)++;
 896         return next_online_pgdat(pgdat);
 897 }
 898
 899 static void frag_stop(struct seq_file *m, void *arg)
 900 {
 901 }
 902
 903 /* Walk all the zones in a node and print using a callback */
 904 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
 905                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
 906 {
 907         struct zone *zone;
 908         struct zone *node_zones = pgdat->node_zones;
 909         unsigned long flags;
 910
 911         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 912                 if (!populated_zone(zone))
 913                         continue;
 914
 915                 spin_lock_irqsave(&zone->lock, flags);
 916                 print(m, pgdat, zone);
 917                 spin_unlock_irqrestore(&zone->lock, flags);
 918         }
 919 }
 920 #endif
 921
 922 #ifdef CONFIG_PROC_FS
 923 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
 924                                                 struct zone *zone)
 925 {
 926         int order;
 927
 928         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 929         for (order = 0; order < MAX_ORDER; ++order)
 930                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
 931         seq_putc(m, '\n');
 932 }
 933
 934 /*
 935  * This walks the free areas for each zone.
 936  */
 937 static int frag_show(struct seq_file *m, void *arg)
 938 {
 939         pg_data_t *pgdat = (pg_data_t *)arg;
 940         walk_zones_in_node(m, pgdat, frag_show_print);
 941         return 0;
 942 }
 943
 944 static void pagetypeinfo_showfree_print(struct seq_file *m,
 945                                         pg_data_t *pgdat, struct zone *zone)
 946 {
 947         int order, mtype;
 948
 949         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
 950                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
 951                                         pgdat->node_id,
 952                                         zone->name,
 953                                         migratetype_names[mtype]);
 954                 for (order = 0; order < MAX_ORDER; ++order) {
 955                         unsigned long freecount = 0;
 956                         struct free_area *area;
 957                         struct list_head *curr;
 958
 959                         area = &(zone->free_area[order]);
 960
 961                         list_for_each(curr, &area->free_list[mtype])
 962                                 freecount++;
 963                         seq_printf(m, "%6lu ", freecount);
 964                 }
 965                 seq_putc(m, '\n');
 966         }
 967 }
 968
 969 /* Print out the free pages at each order for each migatetype */
 970 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
 971 {
 972         int order;
 973         pg_data_t *pgdat = (pg_data_t *)arg;
 974
 975         /* Print header */
 976         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
 977         for (order = 0; order < MAX_ORDER; ++order)
 978                 seq_printf(m, "%6d ", order);
 979         seq_putc(m, '\n');
 980
 981         walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
 982
 983         return 0;
 984 }
 985
 986 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
 987                                         pg_data_t *pgdat, struct zone *zone)
 988 {
 989         int mtype;
 990         unsigned long pfn;
 991         unsigned long start_pfn = zone->zone_start_pfn;
 992         unsigned long end_pfn = zone_end_pfn(zone);
 993         unsigned long count[MIGRATE_TYPES] = { 0, };
 994
 995         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
 996                 struct page *page;
 997
 998                 if (!pfn_valid(pfn))
 999                         continue;
1000
1001                 page = pfn_to_page(pfn);
1002
1003                 /* Watch for unexpected holes punched in the memmap */
1004                 if (!memmap_valid_within(pfn, page, zone))
1005                         continue;
1006
1007                 if (page_zone(page) != zone)
1008                         continue;
1009
1010                 mtype = get_pageblock_migratetype(page);
1011
1012                 if (mtype < MIGRATE_TYPES)
1013                         count[mtype]++;
1014         }
1015
1016         /* Print counts */
1017         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1018         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1019                 seq_printf(m, "%12lu ", count[mtype]);
1020         seq_putc(m, '\n');
1021 }
1022
1023 /* Print out the free pages at each order for each migratetype */
1024 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1025 {
1026         int mtype;
1027         pg_data_t *pgdat = (pg_data_t *)arg;
1028
1029         seq_printf(m, "\n%-23s", "Number of blocks type ");
1030         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1031                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1032         seq_putc(m, '\n');
1033         walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
1034
1035         return 0;
1036 }
1037
1038 #ifdef CONFIG_PAGE_OWNER
1039 static void pagetypeinfo_showmixedcount_print(struct seq_file *m,
1040                                                         pg_data_t *pgdat,
1041                                                         struct zone *zone)
1042 {
1043         struct page *page;
1044         struct page_ext *page_ext;
1045         unsigned long pfn = zone->zone_start_pfn, block_end_pfn;
1046         unsigned long end_pfn = pfn + zone->spanned_pages;
1047         unsigned long count[MIGRATE_TYPES] = { 0, };
1048         int pageblock_mt, page_mt;
1049         int i;
1050
1051         /* Scan block by block. First and last block may be incomplete */
1052         pfn = zone->zone_start_pfn;
1053
1054         /*
1055          * Walk the zone in pageblock_nr_pages steps. If a page block spans
1056          * a zone boundary, it will be double counted between zones. This does
1057          * not matter as the mixed block count will still be correct
1058          */
1059         for (; pfn < end_pfn; ) {
1060                 if (!pfn_valid(pfn)) {
1061                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
1062                         continue;
1063                 }
1064
1065                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
1066                 block_end_pfn = min(block_end_pfn, end_pfn);
1067
1068                 page = pfn_to_page(pfn);
1069                 pageblock_mt = get_pfnblock_migratetype(page, pfn);
1070
1071                 for (; pfn < block_end_pfn; pfn++) {
1072                         if (!pfn_valid_within(pfn))
1073                                 continue;
1074
1075                         page = pfn_to_page(pfn);
1076
1077                         if (page_zone(page) != zone)
1078                                 continue;
1079
1080                         if (PageBuddy(page)) {
1081                                 pfn += (1UL << page_order(page)) - 1;
1082                                 continue;
1083                         }
1084
1085                         if (PageReserved(page))
1086                                 continue;
1087
1088                         page_ext = lookup_page_ext(page);
1089
1090                         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
1091                                 continue;
1092
1093                         page_mt = gfpflags_to_migratetype(page_ext->gfp_mask);
1094                         if (pageblock_mt != page_mt) {
1095                                 if (is_migrate_cma(pageblock_mt))
1096                                         count[MIGRATE_MOVABLE]++;
1097                                 else
1098                                         count[pageblock_mt]++;
1099
1100                                 pfn = block_end_pfn;
1101                                 break;
1102                         }
1103                         pfn += (1UL << page_ext->order) - 1;
1104                 }
1105         }
1106
1107         /* Print counts */
1108         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1109         for (i = 0; i < MIGRATE_TYPES; i++)
1110                 seq_printf(m, "%12lu ", count[i]);
1111         seq_putc(m, '\n');
1112 }
1113 #endif /* CONFIG_PAGE_OWNER */
1114
1115 /*
1116  * Print out the number of pageblocks for each migratetype that contain pages
1117  * of other types. This gives an indication of how well fallbacks are being
1118  * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1119  * to determine what is going on
1120  */
1121 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1122 {
1123 #ifdef CONFIG_PAGE_OWNER
1124         int mtype;
1125
1126         if (!static_branch_unlikely(&page_owner_inited))
1127                 return;
1128
1129         drain_all_pages(NULL);
1130
1131         seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1132         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1133                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1134         seq_putc(m, '\n');
1135
1136         walk_zones_in_node(m, pgdat, pagetypeinfo_showmixedcount_print);
1137 #endif /* CONFIG_PAGE_OWNER */
1138 }
1139
1140 /*
1141  * This prints out statistics in relation to grouping pages by mobility.
1142  * It is expensive to collect so do not constantly read the file.
1143  */
1144 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1145 {
1146         pg_data_t *pgdat = (pg_data_t *)arg;
1147
1148         /* check memoryless node */
1149         if (!node_state(pgdat->node_id, N_MEMORY))
1150                 return 0;
1151
1152         seq_printf(m, "Page block order: %d\n", pageblock_order);
1153         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
1154         seq_putc(m, '\n');
1155         pagetypeinfo_showfree(m, pgdat);
1156         pagetypeinfo_showblockcount(m, pgdat);
1157         pagetypeinfo_showmixedcount(m, pgdat);
1158
1159         return 0;
1160 }
1161
1162 static const struct seq_operations fragmentation_op = {
1163         .start  = frag_start,
1164         .next   = frag_next,
1165         .stop   = frag_stop,
1166         .show   = frag_show,
1167 };
1168
1169 static int fragmentation_open(struct inode *inode, struct file *file)
1170 {
1171         return seq_open(file, &fragmentation_op);
1172 }
1173
1174 static const struct file_operations fragmentation_file_operations = {
1175         .open           = fragmentation_open,
1176         .read           = seq_read,
1177         .llseek         = seq_lseek,
1178         .release        = seq_release,
1179 };
1180
1181 static const struct seq_operations pagetypeinfo_op = {
1182         .start  = frag_start,
1183         .next   = frag_next,
1184         .stop   = frag_stop,
1185         .show   = pagetypeinfo_show,
1186 };
1187
1188 static int pagetypeinfo_open(struct inode *inode, struct file *file)
1189 {
1190         return seq_open(file, &pagetypeinfo_op);
1191 }
1192
1193 static const struct file_operations pagetypeinfo_file_ops = {
1194         .open           = pagetypeinfo_open,
1195         .read           = seq_read,
1196         .llseek         = seq_lseek,
1197         .release        = seq_release,
1198 };
1199
1200 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1201                                                         struct zone *zone)
1202 {
1203         int i;
1204         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1205         seq_printf(m,
1206                    "\n  pages free     %lu"
1207                    "\n        min      %lu"
1208                    "\n        low      %lu"
1209                    "\n        high     %lu"
1210                    "\n        scanned  %lu"
1211                    "\n        spanned  %lu"
1212                    "\n        present  %lu"
1213                    "\n        managed  %lu",
1214                    zone_page_state(zone, NR_FREE_PAGES),
1215                    min_wmark_pages(zone),
1216                    low_wmark_pages(zone),
1217                    high_wmark_pages(zone),
1218                    zone_page_state(zone, NR_PAGES_SCANNED),
1219                    zone->spanned_pages,
1220                    zone->present_pages,
1221                    zone->managed_pages);
1222
1223         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1224                 seq_printf(m, "\n    %-12s %lu", vmstat_text[i],
1225                                 zone_page_state(zone, i));
1226
1227         seq_printf(m,
1228                    "\n        protection: (%ld",
1229                    zone->lowmem_reserve[0]);
1230         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1231                 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1232         seq_printf(m,
1233                    ")"
1234                    "\n  pagesets");
1235         for_each_online_cpu(i) {
1236                 struct per_cpu_pageset *pageset;
1237
1238                 pageset = per_cpu_ptr(zone->pageset, i);
1239                 seq_printf(m,
1240                            "\n    cpu: %i"
1241                            "\n              count: %i"
1242                            "\n              high:  %i"
1243                            "\n              batch: %i",
1244                            i,
1245                            pageset->pcp.count,
1246                            pageset->pcp.high,
1247                            pageset->pcp.batch);
1248 #ifdef CONFIG_SMP
1249                 seq_printf(m, "\n  vm stats threshold: %d",
1250                                 pageset->stat_threshold);
1251 #endif
1252         }
1253         seq_printf(m,
1254                    "\n  all_unreclaimable: %u"
1255                    "\n  start_pfn:         %lu"
1256                    "\n  inactive_ratio:    %u",
1257                    !zone_reclaimable(zone),
1258                    zone->zone_start_pfn,
1259                    zone->inactive_ratio);
1260         seq_putc(m, '\n');
1261 }
1262
1263 /*
1264  * Output information about zones in @pgdat.
1265  */
1266 static int zoneinfo_show(struct seq_file *m, void *arg)
1267 {
1268         pg_data_t *pgdat = (pg_data_t *)arg;
1269         walk_zones_in_node(m, pgdat, zoneinfo_show_print);
1270         return 0;
1271 }
1272
1273 static const struct seq_operations zoneinfo_op = {
1274         .start  = frag_start, /* iterate over all zones. The same as in
1275                                * fragmentation. */
1276         .next   = frag_next,
1277         .stop   = frag_stop,
1278         .show   = zoneinfo_show,
1279 };
1280
1281 static int zoneinfo_open(struct inode *inode, struct file *file)
1282 {
1283         return seq_open(file, &zoneinfo_op);
1284 }
1285
1286 static const struct file_operations proc_zoneinfo_file_operations = {
1287         .open           = zoneinfo_open,
1288         .read           = seq_read,
1289         .llseek         = seq_lseek,
1290         .release        = seq_release,
1291 };
1292
1293 enum writeback_stat_item {
1294         NR_DIRTY_THRESHOLD,
1295         NR_DIRTY_BG_THRESHOLD,
1296         NR_VM_WRITEBACK_STAT_ITEMS,
1297 };
1298
1299 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1300 {
1301         unsigned long *v;
1302         int i, stat_items_size;
1303
1304         if (*pos >= ARRAY_SIZE(vmstat_text))
1305                 return NULL;
1306         stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1307                           NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1308
1309 #ifdef CONFIG_VM_EVENT_COUNTERS
1310         stat_items_size += sizeof(struct vm_event_state);
1311 #endif
1312
1313         v = kmalloc(stat_items_size, GFP_KERNEL);
1314         m->private = v;
1315         if (!v)
1316                 return ERR_PTR(-ENOMEM);
1317         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1318                 v[i] = global_page_state(i);
1319         v += NR_VM_ZONE_STAT_ITEMS;
1320
1321         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1322                             v + NR_DIRTY_THRESHOLD);
1323         v += NR_VM_WRITEBACK_STAT_ITEMS;
1324
1325 #ifdef CONFIG_VM_EVENT_COUNTERS
1326         all_vm_events(v);
1327         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1328         v[PGPGOUT] /= 2;
1329 #endif
1330         return (unsigned long *)m->private + *pos;
1331 }
1332
1333 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1334 {
1335         (*pos)++;
1336         if (*pos >= ARRAY_SIZE(vmstat_text))
1337                 return NULL;
1338         return (unsigned long *)m->private + *pos;
1339 }
1340
1341 static int vmstat_show(struct seq_file *m, void *arg)
1342 {
1343         unsigned long *l = arg;
1344         unsigned long off = l - (unsigned long *)m->private;
1345
1346         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1347         return 0;
1348 }
1349
1350 static void vmstat_stop(struct seq_file *m, void *arg)
1351 {
1352         kfree(m->private);
1353         m->private = NULL;
1354 }
1355
1356 static const struct seq_operations vmstat_op = {
1357         .start  = vmstat_start,
1358         .next   = vmstat_next,
1359         .stop   = vmstat_stop,
1360         .show   = vmstat_show,
1361 };
1362
1363 static int vmstat_open(struct inode *inode, struct file *file)
1364 {
1365         return seq_open(file, &vmstat_op);
1366 }
1367
1368 static const struct file_operations proc_vmstat_file_operations = {
1369         .open           = vmstat_open,
1370         .read           = seq_read,
1371         .llseek         = seq_lseek,
1372         .release        = seq_release,
1373 };
1374 #endif /* CONFIG_PROC_FS */
1375
1376 #ifdef CONFIG_SMP
1377 static struct workqueue_struct *vmstat_wq;
1378 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1379 int sysctl_stat_interval __read_mostly = HZ;
1380 static cpumask_var_t cpu_stat_off;
1381
1382 #ifdef CONFIG_PROC_FS
1383 static void refresh_vm_stats(struct work_struct *work)
1384 {
1385         refresh_cpu_vm_stats(true);
1386 }
1387
1388 int vmstat_refresh(struct ctl_table *table, int write,
1389                    void __user *buffer, size_t *lenp, loff_t *ppos)
1390 {
1391         long val;
1392         int err;
1393         int i;
1394
1395         /*
1396          * The regular update, every sysctl_stat_interval, may come later
1397          * than expected: leaving a significant amount in per_cpu buckets.
1398          * This is particularly misleading when checking a quantity of HUGE
1399          * pages, immediately after running a test.  /proc/sys/vm/stat_refresh,
1400          * which can equally be echo'ed to or cat'ted from (by root),
1401          * can be used to update the stats just before reading them.
1402          *
1403          * Oh, and since global_page_state() etc. are so careful to hide
1404          * transiently negative values, report an error here if any of
1405          * the stats is negative, so we know to go looking for imbalance.
1406          */
1407         err = schedule_on_each_cpu(refresh_vm_stats);
1408         if (err)
1409                 return err;
1410         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1411                 val = atomic_long_read(&vm_stat[i]);
1412                 if (val < 0) {
1413                         switch (i) {
1414                         case NR_ALLOC_BATCH:
1415                         case NR_PAGES_SCANNED:
1416                                 /*
1417                                  * These are often seen to go negative in
1418                                  * recent kernels, but not to go permanently
1419                                  * negative.  Whilst it would be nicer not to
1420                                  * have exceptions, rooting them out would be
1421                                  * another task, of rather low priority.
1422                                  */
1423                                 break;
1424                         default:
1425                                 pr_warn("%s: %s %ld\n",
1426                                         __func__, vmstat_text[i], val);
1427                                 err = -EINVAL;
1428                                 break;
1429                         }
1430                 }
1431         }
1432         if (err)
1433                 return err;
1434         if (write)
1435                 *ppos += *lenp;
1436         else
1437                 *lenp = 0;
1438         return 0;
1439 }
1440 #endif /* CONFIG_PROC_FS */
1441
1442 static void vmstat_update(struct work_struct *w)
1443 {
1444         if (refresh_cpu_vm_stats(true)) {
1445                 /*
1446                  * Counters were updated so we expect more updates
1447                  * to occur in the future. Keep on running the
1448                  * update worker thread.
1449                  * If we were marked on cpu_stat_off clear the flag
1450                  * so that vmstat_shepherd doesn't schedule us again.
1451                  */
1452                 if (!cpumask_test_and_clear_cpu(smp_processor_id(),
1453                                                 cpu_stat_off)) {
1454                         queue_delayed_work_on(smp_processor_id(), vmstat_wq,
1455                                 this_cpu_ptr(&vmstat_work),
1456                                 round_jiffies_relative(sysctl_stat_interval));
1457                 }
1458         } else {
1459                 /*
1460                  * We did not update any counters so the app may be in
1461                  * a mode where it does not cause counter updates.
1462                  * We may be uselessly running vmstat_update.
1463                  * Defer the checking for differentials to the
1464                  * shepherd thread on a different processor.
1465                  */
1466                 cpumask_set_cpu(smp_processor_id(), cpu_stat_off);
1467         }
1468 }
1469
1470 /*
1471  * Switch off vmstat processing and then fold all the remaining differentials
1472  * until the diffs stay at zero. The function is used by NOHZ and can only be
1473  * invoked when tick processing is not active.
1474  */
1475 /*
1476  * Check if the diffs for a certain cpu indicate that
1477  * an update is needed.
1478  */
1479 static bool need_update(int cpu)
1480 {
1481         struct zone *zone;
1482
1483         for_each_populated_zone(zone) {
1484                 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1485
1486                 BUILD_BUG_ON(sizeof(p->vm_stat_diff[0]) != 1);
1487                 /*
1488                  * The fast way of checking if there are any vmstat diffs.
1489                  * This works because the diffs are byte sized items.
1490                  */
1491                 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS))
1492                         return true;
1493
1494         }
1495         return false;
1496 }
1497
1498 void quiet_vmstat(void)
1499 {
1500         if (system_state != SYSTEM_RUNNING)
1501                 return;
1502
1503         /*
1504          * If we are already in hands of the shepherd then there
1505          * is nothing for us to do here.
1506          */
1507         if (cpumask_test_and_set_cpu(smp_processor_id(), cpu_stat_off))
1508                 return;
1509
1510         if (!need_update(smp_processor_id()))
1511                 return;
1512
1513         /*
1514          * Just refresh counters and do not care about the pending delayed
1515          * vmstat_update. It doesn't fire that often to matter and canceling
1516          * it would be too expensive from this path.
1517          * vmstat_shepherd will take care about that for us.
1518          */
1519         refresh_cpu_vm_stats(false);
1520 }
1521
1522
1523 /*
1524  * Shepherd worker thread that checks the
1525  * differentials of processors that have their worker
1526  * threads for vm statistics updates disabled because of
1527  * inactivity.
1528  */
1529 static void vmstat_shepherd(struct work_struct *w);
1530
1531 static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
1532
1533 static void vmstat_shepherd(struct work_struct *w)
1534 {
1535         int cpu;
1536
1537         get_online_cpus();
1538         /* Check processors whose vmstat worker threads have been disabled */
1539         for_each_cpu(cpu, cpu_stat_off) {
1540                 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
1541
1542                 if (need_update(cpu)) {
1543                         if (cpumask_test_and_clear_cpu(cpu, cpu_stat_off))
1544                                 queue_delayed_work_on(cpu, vmstat_wq, dw, 0);
1545                 } else {
1546                         /*
1547                          * Cancel the work if quiet_vmstat has put this
1548                          * cpu on cpu_stat_off because the work item might
1549                          * be still scheduled
1550                          */
1551                         cancel_delayed_work(dw);
1552                 }
1553         }
1554         put_online_cpus();
1555
1556         schedule_delayed_work(&shepherd,
1557                 round_jiffies_relative(sysctl_stat_interval));
1558 }
1559
1560 static void __init start_shepherd_timer(void)
1561 {
1562         int cpu;
1563
1564         for_each_possible_cpu(cpu)
1565                 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
1566                         vmstat_update);
1567
1568         if (!alloc_cpumask_var(&cpu_stat_off, GFP_KERNEL))
1569                 BUG();
1570         cpumask_copy(cpu_stat_off, cpu_online_mask);
1571
1572         vmstat_wq = alloc_workqueue("vmstat", WQ_FREEZABLE|WQ_MEM_RECLAIM, 0);
1573         schedule_delayed_work(&shepherd,
1574                 round_jiffies_relative(sysctl_stat_interval));
1575 }
1576
1577 static void vmstat_cpu_dead(int node)
1578 {
1579         int cpu;
1580
1581         get_online_cpus();
1582         for_each_online_cpu(cpu)
1583                 if (cpu_to_node(cpu) == node)
1584                         goto end;
1585
1586         node_clear_state(node, N_CPU);
1587 end:
1588         put_online_cpus();
1589 }
1590
1591 /*
1592  * Use the cpu notifier to insure that the thresholds are recalculated
1593  * when necessary.
1594  */
1595 static int vmstat_cpuup_callback(struct notifier_block *nfb,
1596                 unsigned long action,
1597                 void *hcpu)
1598 {
1599         long cpu = (long)hcpu;
1600
1601         switch (action) {
1602         case CPU_ONLINE:
1603         case CPU_ONLINE_FROZEN:
1604                 refresh_zone_stat_thresholds();
1605                 node_set_state(cpu_to_node(cpu), N_CPU);
1606                 cpumask_set_cpu(cpu, cpu_stat_off);
1607                 break;
1608         case CPU_DOWN_PREPARE:
1609         case CPU_DOWN_PREPARE_FROZEN:
1610                 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1611                 cpumask_clear_cpu(cpu, cpu_stat_off);
1612                 break;
1613         case CPU_DOWN_FAILED:
1614         case CPU_DOWN_FAILED_FROZEN:
1615                 cpumask_set_cpu(cpu, cpu_stat_off);
1616                 break;
1617         case CPU_DEAD:
1618         case CPU_DEAD_FROZEN:
1619                 refresh_zone_stat_thresholds();
1620                 vmstat_cpu_dead(cpu_to_node(cpu));
1621                 break;
1622         default:
1623                 break;
1624         }
1625         return NOTIFY_OK;
1626 }
1627
1628 static struct notifier_block vmstat_notifier =
1629         { &vmstat_cpuup_callback, NULL, 0 };
1630 #endif
1631
1632 static int __init setup_vmstat(void)
1633 {
1634 #ifdef CONFIG_SMP
1635         cpu_notifier_register_begin();
1636         __register_cpu_notifier(&vmstat_notifier);
1637
1638         start_shepherd_timer();
1639         cpu_notifier_register_done();
1640 #endif
1641 #ifdef CONFIG_PROC_FS
1642         proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1643         proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1644         proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1645         proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1646 #endif
1647         return 0;
1648 }
1649 module_init(setup_vmstat)
1650
1651 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1652
1653 /*
1654  * Return an index indicating how much of the available free memory is
1655  * unusable for an allocation of the requested size.
1656  */
1657 static int unusable_free_index(unsigned int order,
1658                                 struct contig_page_info *info)
1659 {
1660         /* No free memory is interpreted as all free memory is unusable */
1661         if (info->free_pages == 0)
1662                 return 1000;
1663
1664         /*
1665          * Index should be a value between 0 and 1. Return a value to 3
1666          * decimal places.
1667          *
1668          * 0 => no fragmentation
1669          * 1 => high fragmentation
1670          */
1671         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1672
1673 }
1674
1675 static void unusable_show_print(struct seq_file *m,
1676                                         pg_data_t *pgdat, struct zone *zone)
1677 {
1678         unsigned int order;
1679         int index;
1680         struct contig_page_info info;
1681
1682         seq_printf(m, "Node %d, zone %8s ",
1683                                 pgdat->node_id,
1684                                 zone->name);
1685         for (order = 0; order < MAX_ORDER; ++order) {
1686                 fill_contig_page_info(zone, order, &info);
1687                 index = unusable_free_index(order, &info);
1688                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1689         }
1690
1691         seq_putc(m, '\n');
1692 }
1693
1694 /*
1695  * Display unusable free space index
1696  *
1697  * The unusable free space index measures how much of the available free
1698  * memory cannot be used to satisfy an allocation of a given size and is a
1699  * value between 0 and 1. The higher the value, the more of free memory is
1700  * unusable and by implication, the worse the external fragmentation is. This
1701  * can be expressed as a percentage by multiplying by 100.
1702  */
1703 static int unusable_show(struct seq_file *m, void *arg)
1704 {
1705         pg_data_t *pgdat = (pg_data_t *)arg;
1706
1707         /* check memoryless node */
1708         if (!node_state(pgdat->node_id, N_MEMORY))
1709                 return 0;
1710
1711         walk_zones_in_node(m, pgdat, unusable_show_print);
1712
1713         return 0;
1714 }
1715
1716 static const struct seq_operations unusable_op = {
1717         .start  = frag_start,
1718         .next   = frag_next,
1719         .stop   = frag_stop,
1720         .show   = unusable_show,
1721 };
1722
1723 static int unusable_open(struct inode *inode, struct file *file)
1724 {
1725         return seq_open(file, &unusable_op);
1726 }
1727
1728 static const struct file_operations unusable_file_ops = {
1729         .open           = unusable_open,
1730         .read           = seq_read,
1731         .llseek         = seq_lseek,
1732         .release        = seq_release,
1733 };
1734
1735 static void extfrag_show_print(struct seq_file *m,
1736                                         pg_data_t *pgdat, struct zone *zone)
1737 {
1738         unsigned int order;
1739         int index;
1740
1741         /* Alloc on stack as interrupts are disabled for zone walk */
1742         struct contig_page_info info;
1743
1744         seq_printf(m, "Node %d, zone %8s ",
1745                                 pgdat->node_id,
1746                                 zone->name);
1747         for (order = 0; order < MAX_ORDER; ++order) {
1748                 fill_contig_page_info(zone, order, &info);
1749                 index = __fragmentation_index(order, &info);
1750                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1751         }
1752
1753         seq_putc(m, '\n');
1754 }
1755
1756 /*
1757  * Display fragmentation index for orders that allocations would fail for
1758  */
1759 static int extfrag_show(struct seq_file *m, void *arg)
1760 {
1761         pg_data_t *pgdat = (pg_data_t *)arg;
1762
1763         walk_zones_in_node(m, pgdat, extfrag_show_print);
1764
1765         return 0;
1766 }
1767
1768 static const struct seq_operations extfrag_op = {
1769         .start  = frag_start,
1770         .next   = frag_next,
1771         .stop   = frag_stop,
1772         .show   = extfrag_show,
1773 };
1774
1775 static int extfrag_open(struct inode *inode, struct file *file)
1776 {
1777         return seq_open(file, &extfrag_op);
1778 }
1779
1780 static const struct file_operations extfrag_file_ops = {
1781         .open           = extfrag_open,
1782         .read           = seq_read,
1783         .llseek         = seq_lseek,
1784         .release        = seq_release,
1785 };
1786
1787 static int __init extfrag_debug_init(void)
1788 {
1789         struct dentry *extfrag_debug_root;
1790
1791         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1792         if (!extfrag_debug_root)
1793                 return -ENOMEM;
1794
1795         if (!debugfs_create_file("unusable_index", 0444,
1796                         extfrag_debug_root, NULL, &unusable_file_ops))
1797                 goto fail;
1798
1799         if (!debugfs_create_file("extfrag_index", 0444,
1800                         extfrag_debug_root, NULL, &extfrag_file_ops))
1801                 goto fail;
1802
1803         return 0;
1804 fail:
1805         debugfs_remove_recursive(extfrag_debug_root);
1806         return -ENOMEM;
1807 }
1808
1809 module_init(extfrag_debug_init);
1810 #endif