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         int local_nid = numa_node_id();
 585         enum zone_stat_item local_stat = NUMA_LOCAL;
 586
 587         if (unlikely(flags & __GFP_OTHER_NODE)) {
 588                 local_stat = NUMA_OTHER;
 589                 local_nid = preferred_zone->node;
 590         }
 591
 592         if (z->node == local_nid) {
 593                 __inc_zone_state(z, NUMA_HIT);
 594                 __inc_zone_state(z, local_stat);
 595         } else {
 596                 __inc_zone_state(z, NUMA_MISS);
 597                 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
 598         }
 599 }
 600
 601 /*
 602  * Determine the per node value of a stat item.
 603  */
 604 unsigned long node_page_state(int node, enum zone_stat_item item)
 605 {
 606         struct zone *zones = NODE_DATA(node)->node_zones;
 607         int i;
 608         unsigned long count = 0;
 609
 610         for (i = 0; i < MAX_NR_ZONES; i++)
 611                 count += zone_page_state(zones + i, item);
 612
 613         return count;
 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         "compact_daemon_wake",
 828 #endif
 829
 830 #ifdef CONFIG_HUGETLB_PAGE
 831         "htlb_buddy_alloc_success",
 832         "htlb_buddy_alloc_fail",
 833 #endif
 834         "unevictable_pgs_culled",
 835         "unevictable_pgs_scanned",
 836         "unevictable_pgs_rescued",
 837         "unevictable_pgs_mlocked",
 838         "unevictable_pgs_munlocked",
 839         "unevictable_pgs_cleared",
 840         "unevictable_pgs_stranded",
 841
 842 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 843         "thp_fault_alloc",
 844         "thp_fault_fallback",
 845         "thp_collapse_alloc",
 846         "thp_collapse_alloc_failed",
 847         "thp_split_page",
 848         "thp_split_page_failed",
 849         "thp_deferred_split_page",
 850         "thp_split_pmd",
 851         "thp_zero_page_alloc",
 852         "thp_zero_page_alloc_failed",
 853 #endif
 854 #ifdef CONFIG_MEMORY_BALLOON
 855         "balloon_inflate",
 856         "balloon_deflate",
 857 #ifdef CONFIG_BALLOON_COMPACTION
 858         "balloon_migrate",
 859 #endif
 860 #endif /* CONFIG_MEMORY_BALLOON */
 861 #ifdef CONFIG_DEBUG_TLBFLUSH
 862 #ifdef CONFIG_SMP
 863         "nr_tlb_remote_flush",
 864         "nr_tlb_remote_flush_received",
 865 #endif /* CONFIG_SMP */
 866         "nr_tlb_local_flush_all",
 867         "nr_tlb_local_flush_one",
 868 #endif /* CONFIG_DEBUG_TLBFLUSH */
 869
 870 #ifdef CONFIG_DEBUG_VM_VMACACHE
 871         "vmacache_find_calls",
 872         "vmacache_find_hits",
 873         "vmacache_full_flushes",
 874 #endif
 875 #endif /* CONFIG_VM_EVENTS_COUNTERS */
 876 };
 877 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
 878
 879
 880 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
 881      defined(CONFIG_PROC_FS)
 882 static void *frag_start(struct seq_file *m, loff_t *pos)
 883 {
 884         pg_data_t *pgdat;
 885         loff_t node = *pos;
 886
 887         for (pgdat = first_online_pgdat();
 888              pgdat && node;
 889              pgdat = next_online_pgdat(pgdat))
 890                 --node;
 891
 892         return pgdat;
 893 }
 894
 895 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
 896 {
 897         pg_data_t *pgdat = (pg_data_t *)arg;
 898
 899         (*pos)++;
 900         return next_online_pgdat(pgdat);
 901 }
 902
 903 static void frag_stop(struct seq_file *m, void *arg)
 904 {
 905 }
 906
 907 /* Walk all the zones in a node and print using a callback */
 908 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
 909                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
 910 {
 911         struct zone *zone;
 912         struct zone *node_zones = pgdat->node_zones;
 913         unsigned long flags;
 914
 915         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 916                 if (!populated_zone(zone))
 917                         continue;
 918
 919                 spin_lock_irqsave(&zone->lock, flags);
 920                 print(m, pgdat, zone);
 921                 spin_unlock_irqrestore(&zone->lock, flags);
 922         }
 923 }
 924 #endif
 925
 926 #ifdef CONFIG_PROC_FS
 927 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
 928                                                 struct zone *zone)
 929 {
 930         int order;
 931
 932         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 933         for (order = 0; order < MAX_ORDER; ++order)
 934                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
 935         seq_putc(m, '\n');
 936 }
 937
 938 /*
 939  * This walks the free areas for each zone.
 940  */
 941 static int frag_show(struct seq_file *m, void *arg)
 942 {
 943         pg_data_t *pgdat = (pg_data_t *)arg;
 944         walk_zones_in_node(m, pgdat, frag_show_print);
 945         return 0;
 946 }
 947
 948 static void pagetypeinfo_showfree_print(struct seq_file *m,
 949                                         pg_data_t *pgdat, struct zone *zone)
 950 {
 951         int order, mtype;
 952
 953         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
 954                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
 955                                         pgdat->node_id,
 956                                         zone->name,
 957                                         migratetype_names[mtype]);
 958                 for (order = 0; order < MAX_ORDER; ++order) {
 959                         unsigned long freecount = 0;
 960                         struct free_area *area;
 961                         struct list_head *curr;
 962
 963                         area = &(zone->free_area[order]);
 964
 965                         list_for_each(curr, &area->free_list[mtype])
 966                                 freecount++;
 967                         seq_printf(m, "%6lu ", freecount);
 968                 }
 969                 seq_putc(m, '\n');
 970         }
 971 }
 972
 973 /* Print out the free pages at each order for each migatetype */
 974 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
 975 {
 976         int order;
 977         pg_data_t *pgdat = (pg_data_t *)arg;
 978
 979         /* Print header */
 980         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
 981         for (order = 0; order < MAX_ORDER; ++order)
 982                 seq_printf(m, "%6d ", order);
 983         seq_putc(m, '\n');
 984
 985         walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
 986
 987         return 0;
 988 }
 989
 990 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
 991                                         pg_data_t *pgdat, struct zone *zone)
 992 {
 993         int mtype;
 994         unsigned long pfn;
 995         unsigned long start_pfn = zone->zone_start_pfn;
 996         unsigned long end_pfn = zone_end_pfn(zone);
 997         unsigned long count[MIGRATE_TYPES] = { 0, };
 998
 999         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1000                 struct page *page;
1001
1002                 if (!pfn_valid(pfn))
1003                         continue;
1004
1005                 page = pfn_to_page(pfn);
1006
1007                 /* Watch for unexpected holes punched in the memmap */
1008                 if (!memmap_valid_within(pfn, page, zone))
1009                         continue;
1010
1011                 if (page_zone(page) != zone)
1012                         continue;
1013
1014                 mtype = get_pageblock_migratetype(page);
1015
1016                 if (mtype < MIGRATE_TYPES)
1017                         count[mtype]++;
1018         }
1019
1020         /* Print counts */
1021         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1022         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1023                 seq_printf(m, "%12lu ", count[mtype]);
1024         seq_putc(m, '\n');
1025 }
1026
1027 /* Print out the free pages at each order for each migratetype */
1028 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1029 {
1030         int mtype;
1031         pg_data_t *pgdat = (pg_data_t *)arg;
1032
1033         seq_printf(m, "\n%-23s", "Number of blocks type ");
1034         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1035                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1036         seq_putc(m, '\n');
1037         walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
1038
1039         return 0;
1040 }
1041
1042 #ifdef CONFIG_PAGE_OWNER
1043 static void pagetypeinfo_showmixedcount_print(struct seq_file *m,
1044                                                         pg_data_t *pgdat,
1045                                                         struct zone *zone)
1046 {
1047         struct page *page;
1048         struct page_ext *page_ext;
1049         unsigned long pfn = zone->zone_start_pfn, block_end_pfn;
1050         unsigned long end_pfn = pfn + zone->spanned_pages;
1051         unsigned long count[MIGRATE_TYPES] = { 0, };
1052         int pageblock_mt, page_mt;
1053         int i;
1054
1055         /* Scan block by block. First and last block may be incomplete */
1056         pfn = zone->zone_start_pfn;
1057
1058         /*
1059          * Walk the zone in pageblock_nr_pages steps. If a page block spans
1060          * a zone boundary, it will be double counted between zones. This does
1061          * not matter as the mixed block count will still be correct
1062          */
1063         for (; pfn < end_pfn; ) {
1064                 if (!pfn_valid(pfn)) {
1065                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
1066                         continue;
1067                 }
1068
1069                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
1070                 block_end_pfn = min(block_end_pfn, end_pfn);
1071
1072                 page = pfn_to_page(pfn);
1073                 pageblock_mt = get_pfnblock_migratetype(page, pfn);
1074
1075                 for (; pfn < block_end_pfn; pfn++) {
1076                         if (!pfn_valid_within(pfn))
1077                                 continue;
1078
1079                         page = pfn_to_page(pfn);
1080
1081                         if (page_zone(page) != zone)
1082                                 continue;
1083
1084                         if (PageBuddy(page)) {
1085                                 pfn += (1UL << page_order(page)) - 1;
1086                                 continue;
1087                         }
1088
1089                         if (PageReserved(page))
1090                                 continue;
1091
1092                         page_ext = lookup_page_ext(page);
1093
1094                         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
1095                                 continue;
1096
1097                         page_mt = gfpflags_to_migratetype(page_ext->gfp_mask);
1098                         if (pageblock_mt != page_mt) {
1099                                 if (is_migrate_cma(pageblock_mt))
1100                                         count[MIGRATE_MOVABLE]++;
1101                                 else
1102                                         count[pageblock_mt]++;
1103
1104                                 pfn = block_end_pfn;
1105                                 break;
1106                         }
1107                         pfn += (1UL << page_ext->order) - 1;
1108                 }
1109         }
1110
1111         /* Print counts */
1112         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1113         for (i = 0; i < MIGRATE_TYPES; i++)
1114                 seq_printf(m, "%12lu ", count[i]);
1115         seq_putc(m, '\n');
1116 }
1117 #endif /* CONFIG_PAGE_OWNER */
1118
1119 /*
1120  * Print out the number of pageblocks for each migratetype that contain pages
1121  * of other types. This gives an indication of how well fallbacks are being
1122  * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1123  * to determine what is going on
1124  */
1125 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1126 {
1127 #ifdef CONFIG_PAGE_OWNER
1128         int mtype;
1129
1130         if (!static_branch_unlikely(&page_owner_inited))
1131                 return;
1132
1133         drain_all_pages(NULL);
1134
1135         seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1136         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1137                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1138         seq_putc(m, '\n');
1139
1140         walk_zones_in_node(m, pgdat, pagetypeinfo_showmixedcount_print);
1141 #endif /* CONFIG_PAGE_OWNER */
1142 }
1143
1144 /*
1145  * This prints out statistics in relation to grouping pages by mobility.
1146  * It is expensive to collect so do not constantly read the file.
1147  */
1148 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1149 {
1150         pg_data_t *pgdat = (pg_data_t *)arg;
1151
1152         /* check memoryless node */
1153         if (!node_state(pgdat->node_id, N_MEMORY))
1154                 return 0;
1155
1156         seq_printf(m, "Page block order: %d\n", pageblock_order);
1157         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
1158         seq_putc(m, '\n');
1159         pagetypeinfo_showfree(m, pgdat);
1160         pagetypeinfo_showblockcount(m, pgdat);
1161         pagetypeinfo_showmixedcount(m, pgdat);
1162
1163         return 0;
1164 }
1165
1166 static const struct seq_operations fragmentation_op = {
1167         .start  = frag_start,
1168         .next   = frag_next,
1169         .stop   = frag_stop,
1170         .show   = frag_show,
1171 };
1172
1173 static int fragmentation_open(struct inode *inode, struct file *file)
1174 {
1175         return seq_open(file, &fragmentation_op);
1176 }
1177
1178 static const struct file_operations fragmentation_file_operations = {
1179         .open           = fragmentation_open,
1180         .read           = seq_read,
1181         .llseek         = seq_lseek,
1182         .release        = seq_release,
1183 };
1184
1185 static const struct seq_operations pagetypeinfo_op = {
1186         .start  = frag_start,
1187         .next   = frag_next,
1188         .stop   = frag_stop,
1189         .show   = pagetypeinfo_show,
1190 };
1191
1192 static int pagetypeinfo_open(struct inode *inode, struct file *file)
1193 {
1194         return seq_open(file, &pagetypeinfo_op);
1195 }
1196
1197 static const struct file_operations pagetypeinfo_file_ops = {
1198         .open           = pagetypeinfo_open,
1199         .read           = seq_read,
1200         .llseek         = seq_lseek,
1201         .release        = seq_release,
1202 };
1203
1204 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1205                                                         struct zone *zone)
1206 {
1207         int i;
1208         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1209         seq_printf(m,
1210                    "\n  pages free     %lu"
1211                    "\n        min      %lu"
1212                    "\n        low      %lu"
1213                    "\n        high     %lu"
1214                    "\n        scanned  %lu"
1215                    "\n        spanned  %lu"
1216                    "\n        present  %lu"
1217                    "\n        managed  %lu",
1218                    zone_page_state(zone, NR_FREE_PAGES),
1219                    min_wmark_pages(zone),
1220                    low_wmark_pages(zone),
1221                    high_wmark_pages(zone),
1222                    zone_page_state(zone, NR_PAGES_SCANNED),
1223                    zone->spanned_pages,
1224                    zone->present_pages,
1225                    zone->managed_pages);
1226
1227         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1228                 seq_printf(m, "\n    %-12s %lu", vmstat_text[i],
1229                                 zone_page_state(zone, i));
1230
1231         seq_printf(m,
1232                    "\n        protection: (%ld",
1233                    zone->lowmem_reserve[0]);
1234         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1235                 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1236         seq_printf(m,
1237                    ")"
1238                    "\n  pagesets");
1239         for_each_online_cpu(i) {
1240                 struct per_cpu_pageset *pageset;
1241
1242                 pageset = per_cpu_ptr(zone->pageset, i);
1243                 seq_printf(m,
1244                            "\n    cpu: %i"
1245                            "\n              count: %i"
1246                            "\n              high:  %i"
1247                            "\n              batch: %i",
1248                            i,
1249                            pageset->pcp.count,
1250                            pageset->pcp.high,
1251                            pageset->pcp.batch);
1252 #ifdef CONFIG_SMP
1253                 seq_printf(m, "\n  vm stats threshold: %d",
1254                                 pageset->stat_threshold);
1255 #endif
1256         }
1257         seq_printf(m,
1258                    "\n  all_unreclaimable: %u"
1259                    "\n  start_pfn:         %lu"
1260                    "\n  inactive_ratio:    %u",
1261                    !zone_reclaimable(zone),
1262                    zone->zone_start_pfn,
1263                    zone->inactive_ratio);
1264         seq_putc(m, '\n');
1265 }
1266
1267 /*
1268  * Output information about zones in @pgdat.
1269  */
1270 static int zoneinfo_show(struct seq_file *m, void *arg)
1271 {
1272         pg_data_t *pgdat = (pg_data_t *)arg;
1273         walk_zones_in_node(m, pgdat, zoneinfo_show_print);
1274         return 0;
1275 }
1276
1277 static const struct seq_operations zoneinfo_op = {
1278         .start  = frag_start, /* iterate over all zones. The same as in
1279                                * fragmentation. */
1280         .next   = frag_next,
1281         .stop   = frag_stop,
1282         .show   = zoneinfo_show,
1283 };
1284
1285 static int zoneinfo_open(struct inode *inode, struct file *file)
1286 {
1287         return seq_open(file, &zoneinfo_op);
1288 }
1289
1290 static const struct file_operations proc_zoneinfo_file_operations = {
1291         .open           = zoneinfo_open,
1292         .read           = seq_read,
1293         .llseek         = seq_lseek,
1294         .release        = seq_release,
1295 };
1296
1297 enum writeback_stat_item {
1298         NR_DIRTY_THRESHOLD,
1299         NR_DIRTY_BG_THRESHOLD,
1300         NR_VM_WRITEBACK_STAT_ITEMS,
1301 };
1302
1303 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1304 {
1305         unsigned long *v;
1306         int i, stat_items_size;
1307
1308         if (*pos >= ARRAY_SIZE(vmstat_text))
1309                 return NULL;
1310         stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1311                           NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1312
1313 #ifdef CONFIG_VM_EVENT_COUNTERS
1314         stat_items_size += sizeof(struct vm_event_state);
1315 #endif
1316
1317         v = kmalloc(stat_items_size, GFP_KERNEL);
1318         m->private = v;
1319         if (!v)
1320                 return ERR_PTR(-ENOMEM);
1321         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1322                 v[i] = global_page_state(i);
1323         v += NR_VM_ZONE_STAT_ITEMS;
1324
1325         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1326                             v + NR_DIRTY_THRESHOLD);
1327         v += NR_VM_WRITEBACK_STAT_ITEMS;
1328
1329 #ifdef CONFIG_VM_EVENT_COUNTERS
1330         all_vm_events(v);
1331         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1332         v[PGPGOUT] /= 2;
1333 #endif
1334         return (unsigned long *)m->private + *pos;
1335 }
1336
1337 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1338 {
1339         (*pos)++;
1340         if (*pos >= ARRAY_SIZE(vmstat_text))
1341                 return NULL;
1342         return (unsigned long *)m->private + *pos;
1343 }
1344
1345 static int vmstat_show(struct seq_file *m, void *arg)
1346 {
1347         unsigned long *l = arg;
1348         unsigned long off = l - (unsigned long *)m->private;
1349
1350         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1351         return 0;
1352 }
1353
1354 static void vmstat_stop(struct seq_file *m, void *arg)
1355 {
1356         kfree(m->private);
1357         m->private = NULL;
1358 }
1359
1360 static const struct seq_operations vmstat_op = {
1361         .start  = vmstat_start,
1362         .next   = vmstat_next,
1363         .stop   = vmstat_stop,
1364         .show   = vmstat_show,
1365 };
1366
1367 static int vmstat_open(struct inode *inode, struct file *file)
1368 {
1369         return seq_open(file, &vmstat_op);
1370 }
1371
1372 static const struct file_operations proc_vmstat_file_operations = {
1373         .open           = vmstat_open,
1374         .read           = seq_read,
1375         .llseek         = seq_lseek,
1376         .release        = seq_release,
1377 };
1378 #endif /* CONFIG_PROC_FS */
1379
1380 #ifdef CONFIG_SMP
1381 static struct workqueue_struct *vmstat_wq;
1382 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1383 int sysctl_stat_interval __read_mostly = HZ;
1384 static cpumask_var_t cpu_stat_off;
1385
1386 #ifdef CONFIG_PROC_FS
1387 static void refresh_vm_stats(struct work_struct *work)
1388 {
1389         refresh_cpu_vm_stats(true);
1390 }
1391
1392 int vmstat_refresh(struct ctl_table *table, int write,
1393                    void __user *buffer, size_t *lenp, loff_t *ppos)
1394 {
1395         long val;
1396         int err;
1397         int i;
1398
1399         /*
1400          * The regular update, every sysctl_stat_interval, may come later
1401          * than expected: leaving a significant amount in per_cpu buckets.
1402          * This is particularly misleading when checking a quantity of HUGE
1403          * pages, immediately after running a test.  /proc/sys/vm/stat_refresh,
1404          * which can equally be echo'ed to or cat'ted from (by root),
1405          * can be used to update the stats just before reading them.
1406          *
1407          * Oh, and since global_page_state() etc. are so careful to hide
1408          * transiently negative values, report an error here if any of
1409          * the stats is negative, so we know to go looking for imbalance.
1410          */
1411         err = schedule_on_each_cpu(refresh_vm_stats);
1412         if (err)
1413                 return err;
1414         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1415                 val = atomic_long_read(&vm_stat[i]);
1416                 if (val < 0) {
1417                         switch (i) {
1418                         case NR_ALLOC_BATCH:
1419                         case NR_PAGES_SCANNED:
1420                                 /*
1421                                  * These are often seen to go negative in
1422                                  * recent kernels, but not to go permanently
1423                                  * negative.  Whilst it would be nicer not to
1424                                  * have exceptions, rooting them out would be
1425                                  * another task, of rather low priority.
1426                                  */
1427                                 break;
1428                         default:
1429                                 pr_warn("%s: %s %ld\n",
1430                                         __func__, vmstat_text[i], val);
1431                                 err = -EINVAL;
1432                                 break;
1433                         }
1434                 }
1435         }
1436         if (err)
1437                 return err;
1438         if (write)
1439                 *ppos += *lenp;
1440         else
1441                 *lenp = 0;
1442         return 0;
1443 }
1444 #endif /* CONFIG_PROC_FS */
1445
1446 static void vmstat_update(struct work_struct *w)
1447 {
1448         if (refresh_cpu_vm_stats(true)) {
1449                 /*
1450                  * Counters were updated so we expect more updates
1451                  * to occur in the future. Keep on running the
1452                  * update worker thread.
1453                  * If we were marked on cpu_stat_off clear the flag
1454                  * so that vmstat_shepherd doesn't schedule us again.
1455                  */
1456                 if (!cpumask_test_and_clear_cpu(smp_processor_id(),
1457                                                 cpu_stat_off)) {
1458                         queue_delayed_work_on(smp_processor_id(), vmstat_wq,
1459                                 this_cpu_ptr(&vmstat_work),
1460                                 round_jiffies_relative(sysctl_stat_interval));
1461                 }
1462         } else {
1463                 /*
1464                  * We did not update any counters so the app may be in
1465                  * a mode where it does not cause counter updates.
1466                  * We may be uselessly running vmstat_update.
1467                  * Defer the checking for differentials to the
1468                  * shepherd thread on a different processor.
1469                  */
1470                 cpumask_set_cpu(smp_processor_id(), cpu_stat_off);
1471         }
1472 }
1473
1474 /*
1475  * Switch off vmstat processing and then fold all the remaining differentials
1476  * until the diffs stay at zero. The function is used by NOHZ and can only be
1477  * invoked when tick processing is not active.
1478  */
1479 /*
1480  * Check if the diffs for a certain cpu indicate that
1481  * an update is needed.
1482  */
1483 static bool need_update(int cpu)
1484 {
1485         struct zone *zone;
1486
1487         for_each_populated_zone(zone) {
1488                 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1489
1490                 BUILD_BUG_ON(sizeof(p->vm_stat_diff[0]) != 1);
1491                 /*
1492                  * The fast way of checking if there are any vmstat diffs.
1493                  * This works because the diffs are byte sized items.
1494                  */
1495                 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS))
1496                         return true;
1497
1498         }
1499         return false;
1500 }
1501
1502 void quiet_vmstat(void)
1503 {
1504         if (system_state != SYSTEM_RUNNING)
1505                 return;
1506
1507         /*
1508          * If we are already in hands of the shepherd then there
1509          * is nothing for us to do here.
1510          */
1511         if (cpumask_test_and_set_cpu(smp_processor_id(), cpu_stat_off))
1512                 return;
1513
1514         if (!need_update(smp_processor_id()))
1515                 return;
1516
1517         /*
1518          * Just refresh counters and do not care about the pending delayed
1519          * vmstat_update. It doesn't fire that often to matter and canceling
1520          * it would be too expensive from this path.
1521          * vmstat_shepherd will take care about that for us.
1522          */
1523         refresh_cpu_vm_stats(false);
1524 }
1525
1526
1527 /*
1528  * Shepherd worker thread that checks the
1529  * differentials of processors that have their worker
1530  * threads for vm statistics updates disabled because of
1531  * inactivity.
1532  */
1533 static void vmstat_shepherd(struct work_struct *w);
1534
1535 static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
1536
1537 static void vmstat_shepherd(struct work_struct *w)
1538 {
1539         int cpu;
1540
1541         get_online_cpus();
1542         /* Check processors whose vmstat worker threads have been disabled */
1543         for_each_cpu(cpu, cpu_stat_off) {
1544                 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
1545
1546                 if (need_update(cpu)) {
1547                         if (cpumask_test_and_clear_cpu(cpu, cpu_stat_off))
1548                                 queue_delayed_work_on(cpu, vmstat_wq, dw, 0);
1549                 } else {
1550                         /*
1551                          * Cancel the work if quiet_vmstat has put this
1552                          * cpu on cpu_stat_off because the work item might
1553                          * be still scheduled
1554                          */
1555                         cancel_delayed_work(dw);
1556                 }
1557         }
1558         put_online_cpus();
1559
1560         schedule_delayed_work(&shepherd,
1561                 round_jiffies_relative(sysctl_stat_interval));
1562 }
1563
1564 static void __init start_shepherd_timer(void)
1565 {
1566         int cpu;
1567
1568         for_each_possible_cpu(cpu)
1569                 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
1570                         vmstat_update);
1571
1572         if (!alloc_cpumask_var(&cpu_stat_off, GFP_KERNEL))
1573                 BUG();
1574         cpumask_copy(cpu_stat_off, cpu_online_mask);
1575
1576         vmstat_wq = alloc_workqueue("vmstat", WQ_FREEZABLE|WQ_MEM_RECLAIM, 0);
1577         schedule_delayed_work(&shepherd,
1578                 round_jiffies_relative(sysctl_stat_interval));
1579 }
1580
1581 static void vmstat_cpu_dead(int node)
1582 {
1583         int cpu;
1584
1585         get_online_cpus();
1586         for_each_online_cpu(cpu)
1587                 if (cpu_to_node(cpu) == node)
1588                         goto end;
1589
1590         node_clear_state(node, N_CPU);
1591 end:
1592         put_online_cpus();
1593 }
1594
1595 /*
1596  * Use the cpu notifier to insure that the thresholds are recalculated
1597  * when necessary.
1598  */
1599 static int vmstat_cpuup_callback(struct notifier_block *nfb,
1600                 unsigned long action,
1601                 void *hcpu)
1602 {
1603         long cpu = (long)hcpu;
1604
1605         switch (action) {
1606         case CPU_ONLINE:
1607         case CPU_ONLINE_FROZEN:
1608                 refresh_zone_stat_thresholds();
1609                 node_set_state(cpu_to_node(cpu), N_CPU);
1610                 cpumask_set_cpu(cpu, cpu_stat_off);
1611                 break;
1612         case CPU_DOWN_PREPARE:
1613         case CPU_DOWN_PREPARE_FROZEN:
1614                 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1615                 cpumask_clear_cpu(cpu, cpu_stat_off);
1616                 break;
1617         case CPU_DOWN_FAILED:
1618         case CPU_DOWN_FAILED_FROZEN:
1619                 cpumask_set_cpu(cpu, cpu_stat_off);
1620                 break;
1621         case CPU_DEAD:
1622         case CPU_DEAD_FROZEN:
1623                 refresh_zone_stat_thresholds();
1624                 vmstat_cpu_dead(cpu_to_node(cpu));
1625                 break;
1626         default:
1627                 break;
1628         }
1629         return NOTIFY_OK;
1630 }
1631
1632 static struct notifier_block vmstat_notifier =
1633         { &vmstat_cpuup_callback, NULL, 0 };
1634 #endif
1635
1636 static int __init setup_vmstat(void)
1637 {
1638 #ifdef CONFIG_SMP
1639         cpu_notifier_register_begin();
1640         __register_cpu_notifier(&vmstat_notifier);
1641
1642         start_shepherd_timer();
1643         cpu_notifier_register_done();
1644 #endif
1645 #ifdef CONFIG_PROC_FS
1646         proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1647         proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1648         proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1649         proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1650 #endif
1651         return 0;
1652 }
1653 module_init(setup_vmstat)
1654
1655 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1656
1657 /*
1658  * Return an index indicating how much of the available free memory is
1659  * unusable for an allocation of the requested size.
1660  */
1661 static int unusable_free_index(unsigned int order,
1662                                 struct contig_page_info *info)
1663 {
1664         /* No free memory is interpreted as all free memory is unusable */
1665         if (info->free_pages == 0)
1666                 return 1000;
1667
1668         /*
1669          * Index should be a value between 0 and 1. Return a value to 3
1670          * decimal places.
1671          *
1672          * 0 => no fragmentation
1673          * 1 => high fragmentation
1674          */
1675         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1676
1677 }
1678
1679 static void unusable_show_print(struct seq_file *m,
1680                                         pg_data_t *pgdat, struct zone *zone)
1681 {
1682         unsigned int order;
1683         int index;
1684         struct contig_page_info info;
1685
1686         seq_printf(m, "Node %d, zone %8s ",
1687                                 pgdat->node_id,
1688                                 zone->name);
1689         for (order = 0; order < MAX_ORDER; ++order) {
1690                 fill_contig_page_info(zone, order, &info);
1691                 index = unusable_free_index(order, &info);
1692                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1693         }
1694
1695         seq_putc(m, '\n');
1696 }
1697
1698 /*
1699  * Display unusable free space index
1700  *
1701  * The unusable free space index measures how much of the available free
1702  * memory cannot be used to satisfy an allocation of a given size and is a
1703  * value between 0 and 1. The higher the value, the more of free memory is
1704  * unusable and by implication, the worse the external fragmentation is. This
1705  * can be expressed as a percentage by multiplying by 100.
1706  */
1707 static int unusable_show(struct seq_file *m, void *arg)
1708 {
1709         pg_data_t *pgdat = (pg_data_t *)arg;
1710
1711         /* check memoryless node */
1712         if (!node_state(pgdat->node_id, N_MEMORY))
1713                 return 0;
1714
1715         walk_zones_in_node(m, pgdat, unusable_show_print);
1716
1717         return 0;
1718 }
1719
1720 static const struct seq_operations unusable_op = {
1721         .start  = frag_start,
1722         .next   = frag_next,
1723         .stop   = frag_stop,
1724         .show   = unusable_show,
1725 };
1726
1727 static int unusable_open(struct inode *inode, struct file *file)
1728 {
1729         return seq_open(file, &unusable_op);
1730 }
1731
1732 static const struct file_operations unusable_file_ops = {
1733         .open           = unusable_open,
1734         .read           = seq_read,
1735         .llseek         = seq_lseek,
1736         .release        = seq_release,
1737 };
1738
1739 static void extfrag_show_print(struct seq_file *m,
1740                                         pg_data_t *pgdat, struct zone *zone)
1741 {
1742         unsigned int order;
1743         int index;
1744
1745         /* Alloc on stack as interrupts are disabled for zone walk */
1746         struct contig_page_info info;
1747
1748         seq_printf(m, "Node %d, zone %8s ",
1749                                 pgdat->node_id,
1750                                 zone->name);
1751         for (order = 0; order < MAX_ORDER; ++order) {
1752                 fill_contig_page_info(zone, order, &info);
1753                 index = __fragmentation_index(order, &info);
1754                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1755         }
1756
1757         seq_putc(m, '\n');
1758 }
1759
1760 /*
1761  * Display fragmentation index for orders that allocations would fail for
1762  */
1763 static int extfrag_show(struct seq_file *m, void *arg)
1764 {
1765         pg_data_t *pgdat = (pg_data_t *)arg;
1766
1767         walk_zones_in_node(m, pgdat, extfrag_show_print);
1768
1769         return 0;
1770 }
1771
1772 static const struct seq_operations extfrag_op = {
1773         .start  = frag_start,
1774         .next   = frag_next,
1775         .stop   = frag_stop,
1776         .show   = extfrag_show,
1777 };
1778
1779 static int extfrag_open(struct inode *inode, struct file *file)
1780 {
1781         return seq_open(file, &extfrag_op);
1782 }
1783
1784 static const struct file_operations extfrag_file_ops = {
1785         .open           = extfrag_open,
1786         .read           = seq_read,
1787         .llseek         = seq_lseek,
1788         .release        = seq_release,
1789 };
1790
1791 static int __init extfrag_debug_init(void)
1792 {
1793         struct dentry *extfrag_debug_root;
1794
1795         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1796         if (!extfrag_debug_root)
1797                 return -ENOMEM;
1798
1799         if (!debugfs_create_file("unusable_index", 0444,
1800                         extfrag_debug_root, NULL, &unusable_file_ops))
1801                 goto fail;
1802
1803         if (!debugfs_create_file("extfrag_index", 0444,
1804                         extfrag_debug_root, NULL, &extfrag_file_ops))
1805                 goto fail;
1806
1807         return 0;
1808 fail:
1809         debugfs_remove_recursive(extfrag_debug_root);
1810         return -ENOMEM;
1811 }
1812
1813 module_init(extfrag_debug_init);
1814 #endif