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
 604         return
 605 #ifdef CONFIG_ZONE_DMA
 606                 zone_page_state(&zones[ZONE_DMA], item) +
 607 #endif
 608 #ifdef CONFIG_ZONE_DMA32
 609                 zone_page_state(&zones[ZONE_DMA32], item) +
 610 #endif
 611 #ifdef CONFIG_HIGHMEM
 612                 zone_page_state(&zones[ZONE_HIGHMEM], item) +
 613 #endif
 614                 zone_page_state(&zones[ZONE_NORMAL], item) +
 615                 zone_page_state(&zones[ZONE_MOVABLE], item);
 616 }
 617
 618 #endif
 619
 620 #ifdef CONFIG_COMPACTION
 621
 622 struct contig_page_info {
 623         unsigned long free_pages;
 624         unsigned long free_blocks_total;
 625         unsigned long free_blocks_suitable;
 626 };
 627
 628 /*
 629  * Calculate the number of free pages in a zone, how many contiguous
 630  * pages are free and how many are large enough to satisfy an allocation of
 631  * the target size. Note that this function makes no attempt to estimate
 632  * how many suitable free blocks there *might* be if MOVABLE pages were
 633  * migrated. Calculating that is possible, but expensive and can be
 634  * figured out from userspace
 635  */
 636 static void fill_contig_page_info(struct zone *zone,
 637                                 unsigned int suitable_order,
 638                                 struct contig_page_info *info)
 639 {
 640         unsigned int order;
 641
 642         info->free_pages = 0;
 643         info->free_blocks_total = 0;
 644         info->free_blocks_suitable = 0;
 645
 646         for (order = 0; order < MAX_ORDER; order++) {
 647                 unsigned long blocks;
 648
 649                 /* Count number of free blocks */
 650                 blocks = zone->free_area[order].nr_free;
 651                 info->free_blocks_total += blocks;
 652
 653                 /* Count free base pages */
 654                 info->free_pages += blocks << order;
 655
 656                 /* Count the suitable free blocks */
 657                 if (order >= suitable_order)
 658                         info->free_blocks_suitable += blocks <<
 659                                                 (order - suitable_order);
 660         }
 661 }
 662
 663 /*
 664  * A fragmentation index only makes sense if an allocation of a requested
 665  * size would fail. If that is true, the fragmentation index indicates
 666  * whether external fragmentation or a lack of memory was the problem.
 667  * The value can be used to determine if page reclaim or compaction
 668  * should be used
 669  */
 670 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
 671 {
 672         unsigned long requested = 1UL << order;
 673
 674         if (!info->free_blocks_total)
 675                 return 0;
 676
 677         /* Fragmentation index only makes sense when a request would fail */
 678         if (info->free_blocks_suitable)
 679                 return -1000;
 680
 681         /*
 682          * Index is between 0 and 1 so return within 3 decimal places
 683          *
 684          * 0 => allocation would fail due to lack of memory
 685          * 1 => allocation would fail due to fragmentation
 686          */
 687         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
 688 }
 689
 690 /* Same as __fragmentation index but allocs contig_page_info on stack */
 691 int fragmentation_index(struct zone *zone, unsigned int order)
 692 {
 693         struct contig_page_info info;
 694
 695         fill_contig_page_info(zone, order, &info);
 696         return __fragmentation_index(order, &info);
 697 }
 698 #endif
 699
 700 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || defined(CONFIG_NUMA)
 701 #ifdef CONFIG_ZONE_DMA
 702 #define TEXT_FOR_DMA(xx) xx "_dma",
 703 #else
 704 #define TEXT_FOR_DMA(xx)
 705 #endif
 706
 707 #ifdef CONFIG_ZONE_DMA32
 708 #define TEXT_FOR_DMA32(xx) xx "_dma32",
 709 #else
 710 #define TEXT_FOR_DMA32(xx)
 711 #endif
 712
 713 #ifdef CONFIG_HIGHMEM
 714 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
 715 #else
 716 #define TEXT_FOR_HIGHMEM(xx)
 717 #endif
 718
 719 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
 720                                         TEXT_FOR_HIGHMEM(xx) xx "_movable",
 721
 722 const char * const vmstat_text[] = {
 723         /* enum zone_stat_item countes */
 724         "nr_free_pages",
 725         "nr_alloc_batch",
 726         "nr_inactive_anon",
 727         "nr_active_anon",
 728         "nr_inactive_file",
 729         "nr_active_file",
 730         "nr_unevictable",
 731         "nr_mlock",
 732         "nr_anon_pages",
 733         "nr_mapped",
 734         "nr_file_pages",
 735         "nr_dirty",
 736         "nr_writeback",
 737         "nr_slab_reclaimable",
 738         "nr_slab_unreclaimable",
 739         "nr_page_table_pages",
 740         "nr_kernel_stack",
 741         "nr_overhead",
 742         "nr_unstable",
 743         "nr_bounce",
 744         "nr_vmscan_write",
 745         "nr_vmscan_immediate_reclaim",
 746         "nr_writeback_temp",
 747         "nr_isolated_anon",
 748         "nr_isolated_file",
 749         "nr_shmem",
 750         "nr_dirtied",
 751         "nr_written",
 752         "nr_pages_scanned",
 753
 754 #ifdef CONFIG_NUMA
 755         "numa_hit",
 756         "numa_miss",
 757         "numa_foreign",
 758         "numa_interleave",
 759         "numa_local",
 760         "numa_other",
 761 #endif
 762         "workingset_refault",
 763         "workingset_activate",
 764         "workingset_nodereclaim",
 765         "nr_anon_transparent_hugepages",
 766         "nr_free_cma",
 767         "nr_swapcache",
 768         "nr_indirectly_reclaimable",
 769
 770         /* enum writeback_stat_item counters */
 771         "nr_dirty_threshold",
 772         "nr_dirty_background_threshold",
 773
 774 #ifdef CONFIG_VM_EVENT_COUNTERS
 775         /* enum vm_event_item counters */
 776         "pgpgin",
 777         "pgpgout",
 778         "pgpgoutclean",
 779         "pswpin",
 780         "pswpout",
 781
 782         TEXTS_FOR_ZONES("pgalloc")
 783
 784         "pgfree",
 785         "pgactivate",
 786         "pgdeactivate",
 787
 788         "pgfault",
 789         "pgmajfault",
 790
 791         TEXTS_FOR_ZONES("pgrefill")
 792         TEXTS_FOR_ZONES("pgsteal_kswapd")
 793         TEXTS_FOR_ZONES("pgsteal_direct")
 794         TEXTS_FOR_ZONES("pgscan_kswapd")
 795         TEXTS_FOR_ZONES("pgscan_direct")
 796         "pgscan_direct_throttle",
 797
 798 #ifdef CONFIG_NUMA
 799         "zone_reclaim_failed",
 800 #endif
 801         "pginodesteal",
 802         "slabs_scanned",
 803         "kswapd_inodesteal",
 804         "kswapd_low_wmark_hit_quickly",
 805         "kswapd_high_wmark_hit_quickly",
 806         "pageoutrun",
 807         "allocstall",
 808
 809         "pgrotated",
 810
 811         "drop_pagecache",
 812         "drop_slab",
 813
 814 #ifdef CONFIG_NUMA_BALANCING
 815         "numa_pte_updates",
 816         "numa_huge_pte_updates",
 817         "numa_hint_faults",
 818         "numa_hint_faults_local",
 819         "numa_pages_migrated",
 820 #endif
 821 #ifdef CONFIG_MIGRATION
 822         "pgmigrate_success",
 823         "pgmigrate_fail",
 824 #endif
 825 #ifdef CONFIG_COMPACTION
 826         "compact_migrate_scanned",
 827         "compact_free_scanned",
 828         "compact_isolated",
 829         "compact_stall",
 830         "compact_fail",
 831         "compact_success",
 832         "compact_daemon_wake",
 833 #endif
 834
 835 #ifdef CONFIG_HUGETLB_PAGE
 836         "htlb_buddy_alloc_success",
 837         "htlb_buddy_alloc_fail",
 838 #endif
 839         "unevictable_pgs_culled",
 840         "unevictable_pgs_scanned",
 841         "unevictable_pgs_rescued",
 842         "unevictable_pgs_mlocked",
 843         "unevictable_pgs_munlocked",
 844         "unevictable_pgs_cleared",
 845         "unevictable_pgs_stranded",
 846
 847 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 848         "thp_fault_alloc",
 849         "thp_fault_fallback",
 850         "thp_collapse_alloc",
 851         "thp_collapse_alloc_failed",
 852         "thp_split",
 853         "thp_zero_page_alloc",
 854         "thp_zero_page_alloc_failed",
 855 #endif
 856 #ifdef CONFIG_MEMORY_BALLOON
 857         "balloon_inflate",
 858         "balloon_deflate",
 859 #ifdef CONFIG_BALLOON_COMPACTION
 860         "balloon_migrate",
 861 #endif
 862 #endif /* CONFIG_MEMORY_BALLOON */
 863 #ifdef CONFIG_DEBUG_TLBFLUSH
 864         "nr_tlb_remote_flush",
 865         "nr_tlb_remote_flush_received",
 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 #endif
 874 #endif /* CONFIG_VM_EVENTS_COUNTERS */
 875 };
 876 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
 877
 878
 879 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
 880      defined(CONFIG_PROC_FS)
 881 static void *frag_start(struct seq_file *m, loff_t *pos)
 882 {
 883         pg_data_t *pgdat;
 884         loff_t node = *pos;
 885
 886         for (pgdat = first_online_pgdat();
 887              pgdat && node;
 888              pgdat = next_online_pgdat(pgdat))
 889                 --node;
 890
 891         return pgdat;
 892 }
 893
 894 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
 895 {
 896         pg_data_t *pgdat = (pg_data_t *)arg;
 897
 898         (*pos)++;
 899         return next_online_pgdat(pgdat);
 900 }
 901
 902 static void frag_stop(struct seq_file *m, void *arg)
 903 {
 904 }
 905
 906 /* Walk all the zones in a node and print using a callback */
 907 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
 908                 bool nolock,
 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                 if (!nolock)
 920                         spin_lock_irqsave(&zone->lock, flags);
 921                 print(m, pgdat, zone);
 922                 if (!nolock)
 923                         spin_unlock_irqrestore(&zone->lock, flags);
 924         }
 925 }
 926 #endif
 927
 928 #ifdef CONFIG_PROC_FS
 929 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
 930                                                 struct zone *zone)
 931 {
 932         int order;
 933
 934         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 935         for (order = 0; order < MAX_ORDER; ++order)
 936                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
 937         seq_putc(m, '\n');
 938 }
 939
 940 /*
 941  * This walks the free areas for each zone.
 942  */
 943 static int frag_show(struct seq_file *m, void *arg)
 944 {
 945         pg_data_t *pgdat = (pg_data_t *)arg;
 946         walk_zones_in_node(m, pgdat, false, frag_show_print);
 947         return 0;
 948 }
 949
 950 static void pagetypeinfo_showfree_print(struct seq_file *m,
 951                                         pg_data_t *pgdat, struct zone *zone)
 952 {
 953         int order, mtype;
 954
 955         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
 956                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
 957                                         pgdat->node_id,
 958                                         zone->name,
 959                                         migratetype_names[mtype]);
 960                 for (order = 0; order < MAX_ORDER; ++order) {
 961                         unsigned long freecount = 0;
 962                         struct free_area *area;
 963                         struct list_head *curr;
 964
 965                         area = &(zone->free_area[order]);
 966
 967                         list_for_each(curr, &area->free_list[mtype])
 968                                 freecount++;
 969                         seq_printf(m, "%6lu ", freecount);
 970                 }
 971                 seq_putc(m, '\n');
 972         }
 973 }
 974
 975 /* Print out the free pages at each order for each migatetype */
 976 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
 977 {
 978         int order;
 979         pg_data_t *pgdat = (pg_data_t *)arg;
 980
 981         /* Print header */
 982         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
 983         for (order = 0; order < MAX_ORDER; ++order)
 984                 seq_printf(m, "%6d ", order);
 985         seq_putc(m, '\n');
 986
 987         walk_zones_in_node(m, pgdat, false, pagetypeinfo_showfree_print);
 988
 989         return 0;
 990 }
 991
 992 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
 993                                         pg_data_t *pgdat, struct zone *zone)
 994 {
 995         int mtype;
 996         unsigned long pfn;
 997         unsigned long start_pfn = zone->zone_start_pfn;
 998         unsigned long end_pfn = zone_end_pfn(zone);
 999         unsigned long count[MIGRATE_TYPES] = { 0, };
1000
1001         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1002                 struct page *page;
1003
1004                 if (!pfn_valid(pfn))
1005                         continue;
1006
1007                 page = pfn_to_page(pfn);
1008
1009                 /* Watch for unexpected holes punched in the memmap */
1010                 if (!memmap_valid_within(pfn, page, zone))
1011                         continue;
1012
1013                 mtype = get_pageblock_migratetype(page);
1014
1015                 if (mtype < MIGRATE_TYPES)
1016                         count[mtype]++;
1017         }
1018
1019         /* Print counts */
1020         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1021         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1022                 seq_printf(m, "%12lu ", count[mtype]);
1023         seq_putc(m, '\n');
1024 }
1025
1026 /* Print out the free pages at each order for each migratetype */
1027 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1028 {
1029         int mtype;
1030         pg_data_t *pgdat = (pg_data_t *)arg;
1031
1032         seq_printf(m, "\n%-23s", "Number of blocks type ");
1033         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1034                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1035         seq_putc(m, '\n');
1036         walk_zones_in_node(m, pgdat, false, pagetypeinfo_showblockcount_print);
1037
1038         return 0;
1039 }
1040
1041 #ifdef CONFIG_PAGE_OWNER
1042 static void pagetypeinfo_showmixedcount_print(struct seq_file *m,
1043                                                         pg_data_t *pgdat,
1044                                                         struct zone *zone)
1045 {
1046         struct page *page;
1047         struct page_ext *page_ext;
1048         unsigned long pfn = zone->zone_start_pfn, block_end_pfn;
1049         unsigned long end_pfn = pfn + zone->spanned_pages;
1050         unsigned long count[MIGRATE_TYPES] = { 0, };
1051         int pageblock_mt, page_mt;
1052         int i;
1053
1054         /* Scan block by block. First and last block may be incomplete */
1055         pfn = zone->zone_start_pfn;
1056
1057         /*
1058          * Walk the zone in pageblock_nr_pages steps. If a page block spans
1059          * a zone boundary, it will be double counted between zones. This does
1060          * not matter as the mixed block count will still be correct
1061          */
1062         for (; pfn < end_pfn; ) {
1063                 if (!pfn_valid(pfn)) {
1064                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
1065                         continue;
1066                 }
1067
1068                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
1069                 block_end_pfn = min(block_end_pfn, end_pfn);
1070
1071                 page = pfn_to_page(pfn);
1072                 pageblock_mt = get_pfnblock_migratetype(page, pfn);
1073
1074                 for (; pfn < block_end_pfn; pfn++) {
1075                         if (!pfn_valid_within(pfn))
1076                                 continue;
1077
1078                         page = pfn_to_page(pfn);
1079                         if (PageBuddy(page)) {
1080                                 unsigned long freepage_order;
1081
1082                                 freepage_order = page_order_unsafe(page);
1083                                 if (freepage_order < MAX_ORDER)
1084                                         pfn += (1UL << freepage_order) - 1;
1085                                 continue;
1086                         }
1087
1088                         if (PageReserved(page))
1089                                 continue;
1090
1091                         page_ext = lookup_page_ext(page);
1092                         if (unlikely(!page_ext))
1093                                 continue;
1094
1095                         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
1096                                 continue;
1097
1098                         page_mt = gfpflags_to_migratetype(page_ext->gfp_mask);
1099                         if (pageblock_mt != page_mt) {
1100                                 if (is_migrate_cma(pageblock_mt))
1101                                         count[MIGRATE_MOVABLE]++;
1102                                 else
1103                                         count[pageblock_mt]++;
1104
1105                                 pfn = block_end_pfn;
1106                                 break;
1107                         }
1108                         pfn += (1UL << page_ext->order) - 1;
1109                 }
1110         }
1111
1112         /* Print counts */
1113         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1114         for (i = 0; i < MIGRATE_TYPES; i++)
1115                 seq_printf(m, "%12lu ", count[i]);
1116         seq_putc(m, '\n');
1117 }
1118 #endif /* CONFIG_PAGE_OWNER */
1119
1120 /*
1121  * Print out the number of pageblocks for each migratetype that contain pages
1122  * of other types. This gives an indication of how well fallbacks are being
1123  * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1124  * to determine what is going on
1125  */
1126 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1127 {
1128 #ifdef CONFIG_PAGE_OWNER
1129         int mtype;
1130
1131         if (!static_branch_unlikely(&page_owner_inited))
1132                 return;
1133
1134         drain_all_pages(NULL);
1135
1136         seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1137         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1138                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1139         seq_putc(m, '\n');
1140
1141         walk_zones_in_node(m, pgdat, true, pagetypeinfo_showmixedcount_print);
1142 #endif /* CONFIG_PAGE_OWNER */
1143 }
1144
1145 /*
1146  * This prints out statistics in relation to grouping pages by mobility.
1147  * It is expensive to collect so do not constantly read the file.
1148  */
1149 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1150 {
1151         pg_data_t *pgdat = (pg_data_t *)arg;
1152
1153         /* check memoryless node */
1154         if (!node_state(pgdat->node_id, N_MEMORY))
1155                 return 0;
1156
1157         seq_printf(m, "Page block order: %d\n", pageblock_order);
1158         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
1159         seq_putc(m, '\n');
1160         pagetypeinfo_showfree(m, pgdat);
1161         pagetypeinfo_showblockcount(m, pgdat);
1162         pagetypeinfo_showmixedcount(m, pgdat);
1163
1164         return 0;
1165 }
1166
1167 static const struct seq_operations fragmentation_op = {
1168         .start  = frag_start,
1169         .next   = frag_next,
1170         .stop   = frag_stop,
1171         .show   = frag_show,
1172 };
1173
1174 static int fragmentation_open(struct inode *inode, struct file *file)
1175 {
1176         return seq_open(file, &fragmentation_op);
1177 }
1178
1179 static const struct file_operations fragmentation_file_operations = {
1180         .open           = fragmentation_open,
1181         .read           = seq_read,
1182         .llseek         = seq_lseek,
1183         .release        = seq_release,
1184 };
1185
1186 static const struct seq_operations pagetypeinfo_op = {
1187         .start  = frag_start,
1188         .next   = frag_next,
1189         .stop   = frag_stop,
1190         .show   = pagetypeinfo_show,
1191 };
1192
1193 static int pagetypeinfo_open(struct inode *inode, struct file *file)
1194 {
1195         return seq_open(file, &pagetypeinfo_op);
1196 }
1197
1198 static const struct file_operations pagetypeinfo_file_ops = {
1199         .open           = pagetypeinfo_open,
1200         .read           = seq_read,
1201         .llseek         = seq_lseek,
1202         .release        = seq_release,
1203 };
1204
1205 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1206                                                         struct zone *zone)
1207 {
1208         int i;
1209         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1210         seq_printf(m,
1211                    "\n  pages free     %lu"
1212                    "\n        min      %lu"
1213                    "\n        low      %lu"
1214                    "\n        high     %lu"
1215                    "\n        scanned  %lu"
1216                    "\n        spanned  %lu"
1217                    "\n        present  %lu"
1218                    "\n        managed  %lu",
1219                    zone_page_state(zone, NR_FREE_PAGES),
1220                    min_wmark_pages(zone),
1221                    low_wmark_pages(zone),
1222                    high_wmark_pages(zone),
1223                    zone_page_state(zone, NR_PAGES_SCANNED),
1224                    zone->spanned_pages,
1225                    zone->present_pages,
1226                    zone->managed_pages);
1227
1228         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1229                 seq_printf(m, "\n    %-12s %lu", vmstat_text[i],
1230                                 zone_page_state(zone, i));
1231
1232         seq_printf(m,
1233                    "\n        protection: (%ld",
1234                    zone->lowmem_reserve[0]);
1235         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1236                 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1237         seq_printf(m,
1238                    ")"
1239                    "\n  pagesets");
1240         for_each_online_cpu(i) {
1241                 struct per_cpu_pageset *pageset;
1242
1243                 pageset = per_cpu_ptr(zone->pageset, i);
1244                 seq_printf(m,
1245                            "\n    cpu: %i"
1246                            "\n              count: %i"
1247                            "\n              high:  %i"
1248                            "\n              batch: %i",
1249                            i,
1250                            pageset->pcp.count,
1251                            pageset->pcp.high,
1252                            pageset->pcp.batch);
1253 #ifdef CONFIG_SMP
1254                 seq_printf(m, "\n  vm stats threshold: %d",
1255                                 pageset->stat_threshold);
1256 #endif
1257         }
1258         seq_printf(m,
1259                    "\n  all_unreclaimable: %u"
1260                    "\n  start_pfn:         %lu"
1261                    "\n  inactive_ratio:    %u",
1262                    !zone_reclaimable(zone),
1263                    zone->zone_start_pfn,
1264                    zone->inactive_ratio);
1265         seq_putc(m, '\n');
1266 }
1267
1268 /*
1269  * Output information about zones in @pgdat.
1270  */
1271 static int zoneinfo_show(struct seq_file *m, void *arg)
1272 {
1273         pg_data_t *pgdat = (pg_data_t *)arg;
1274         walk_zones_in_node(m, pgdat, false, zoneinfo_show_print);
1275         return 0;
1276 }
1277
1278 static const struct seq_operations zoneinfo_op = {
1279         .start  = frag_start, /* iterate over all zones. The same as in
1280                                * fragmentation. */
1281         .next   = frag_next,
1282         .stop   = frag_stop,
1283         .show   = zoneinfo_show,
1284 };
1285
1286 static int zoneinfo_open(struct inode *inode, struct file *file)
1287 {
1288         return seq_open(file, &zoneinfo_op);
1289 }
1290
1291 static const struct file_operations proc_zoneinfo_file_operations = {
1292         .open           = zoneinfo_open,
1293         .read           = seq_read,
1294         .llseek         = seq_lseek,
1295         .release        = seq_release,
1296 };
1297
1298 enum writeback_stat_item {
1299         NR_DIRTY_THRESHOLD,
1300         NR_DIRTY_BG_THRESHOLD,
1301         NR_VM_WRITEBACK_STAT_ITEMS,
1302 };
1303
1304 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1305 {
1306         unsigned long *v;
1307         int i, stat_items_size;
1308
1309         if (*pos >= ARRAY_SIZE(vmstat_text))
1310                 return NULL;
1311         stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1312                           NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1313
1314 #ifdef CONFIG_VM_EVENT_COUNTERS
1315         stat_items_size += sizeof(struct vm_event_state);
1316 #endif
1317
1318         v = kmalloc(stat_items_size, GFP_KERNEL);
1319         m->private = v;
1320         if (!v)
1321                 return ERR_PTR(-ENOMEM);
1322         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1323                 v[i] = global_page_state(i);
1324         v += NR_VM_ZONE_STAT_ITEMS;
1325
1326         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1327                             v + NR_DIRTY_THRESHOLD);
1328         v += NR_VM_WRITEBACK_STAT_ITEMS;
1329
1330 #ifdef CONFIG_VM_EVENT_COUNTERS
1331         all_vm_events(v);
1332         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1333         v[PGPGOUT] /= 2;
1334 #endif
1335         return (unsigned long *)m->private + *pos;
1336 }
1337
1338 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1339 {
1340         (*pos)++;
1341         if (*pos >= ARRAY_SIZE(vmstat_text))
1342                 return NULL;
1343         return (unsigned long *)m->private + *pos;
1344 }
1345
1346 static int vmstat_show(struct seq_file *m, void *arg)
1347 {
1348         unsigned long *l = arg;
1349         unsigned long off = l - (unsigned long *)m->private;
1350
1351         seq_puts(m, vmstat_text[off]);
1352         seq_put_decimal_ull(m, ' ', *l);
1353         seq_putc(m, '\n');
1354         return 0;
1355 }
1356
1357 static void vmstat_stop(struct seq_file *m, void *arg)
1358 {
1359         kfree(m->private);
1360         m->private = NULL;
1361 }
1362
1363 static const struct seq_operations vmstat_op = {
1364         .start  = vmstat_start,
1365         .next   = vmstat_next,
1366         .stop   = vmstat_stop,
1367         .show   = vmstat_show,
1368 };
1369
1370 static int vmstat_open(struct inode *inode, struct file *file)
1371 {
1372         return seq_open(file, &vmstat_op);
1373 }
1374
1375 static const struct file_operations proc_vmstat_file_operations = {
1376         .open           = vmstat_open,
1377         .read           = seq_read,
1378         .llseek         = seq_lseek,
1379         .release        = seq_release,
1380 };
1381 #endif /* CONFIG_PROC_FS */
1382
1383 #ifdef CONFIG_SMP
1384 static struct workqueue_struct *vmstat_wq;
1385 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1386 int sysctl_stat_interval __read_mostly = HZ;
1387 static cpumask_var_t cpu_stat_off;
1388
1389 static void vmstat_update(struct work_struct *w)
1390 {
1391         if (refresh_cpu_vm_stats(true) && !cpu_isolated(smp_processor_id())) {
1392                 /*
1393                  * Counters were updated so we expect more updates
1394                  * to occur in the future. Keep on running the
1395                  * update worker thread.
1396                  * If we were marked on cpu_stat_off clear the flag
1397                  * so that vmstat_shepherd doesn't schedule us again.
1398                  */
1399                 if (!cpumask_test_and_clear_cpu(smp_processor_id(),
1400                                                 cpu_stat_off)) {
1401                         queue_delayed_work_on(smp_processor_id(), vmstat_wq,
1402                                 this_cpu_ptr(&vmstat_work),
1403                                 round_jiffies_relative(sysctl_stat_interval));
1404                 }
1405         } else {
1406                 /*
1407                  * We did not update any counters so the app may be in
1408                  * a mode where it does not cause counter updates or the cpu
1409                  * was isolated.
1410                  * We may be uselessly running vmstat_update.
1411                  * Defer the checking for differentials to the
1412                  * shepherd thread on a different processor.
1413                  */
1414                 cpumask_set_cpu(smp_processor_id(), cpu_stat_off);
1415         }
1416 }
1417
1418 /*
1419  * Switch off vmstat processing and then fold all the remaining differentials
1420  * until the diffs stay at zero. The function is used by NOHZ and can only be
1421  * invoked when tick processing is not active.
1422  */
1423 /*
1424  * Check if the diffs for a certain cpu indicate that
1425  * an update is needed.
1426  */
1427 static bool need_update(int cpu)
1428 {
1429         struct zone *zone;
1430
1431         for_each_populated_zone(zone) {
1432                 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1433
1434                 BUILD_BUG_ON(sizeof(p->vm_stat_diff[0]) != 1);
1435                 /*
1436                  * The fast way of checking if there are any vmstat diffs.
1437                  * This works because the diffs are byte sized items.
1438                  */
1439                 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS))
1440                         return true;
1441
1442         }
1443         return false;
1444 }
1445
1446 void quiet_vmstat(void)
1447 {
1448         if (system_state != SYSTEM_RUNNING)
1449                 return;
1450
1451         /*
1452          * If we are already in hands of the shepherd then there
1453          * is nothing for us to do here.
1454          */
1455         if (cpumask_test_and_set_cpu(smp_processor_id(), cpu_stat_off))
1456                 return;
1457
1458         if (!need_update(smp_processor_id()))
1459                 return;
1460
1461         /*
1462          * Just refresh counters and do not care about the pending delayed
1463          * vmstat_update. It doesn't fire that often to matter and canceling
1464          * it would be too expensive from this path.
1465          * vmstat_shepherd will take care about that for us.
1466          */
1467         refresh_cpu_vm_stats(false);
1468 }
1469
1470
1471 /*
1472  * Shepherd worker thread that checks the
1473  * differentials of processors that have their worker
1474  * threads for vm statistics updates disabled because of
1475  * inactivity.
1476  */
1477 static void vmstat_shepherd(struct work_struct *w);
1478
1479 static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
1480
1481 static void vmstat_shepherd(struct work_struct *w)
1482 {
1483         int cpu;
1484
1485         get_online_cpus();
1486         /* Check processors whose vmstat worker threads have been disabled */
1487         for_each_cpu(cpu, cpu_stat_off) {
1488                 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
1489
1490                 if (!cpu_isolated(cpu) && need_update(cpu)) {
1491                         if (cpumask_test_and_clear_cpu(cpu, cpu_stat_off))
1492                                 queue_delayed_work_on(cpu, vmstat_wq, dw, 0);
1493                 } else {
1494                         /*
1495                          * Cancel the work if quiet_vmstat has put this
1496                          * cpu on cpu_stat_off because the work item might
1497                          * be still scheduled
1498                          */
1499                         cancel_delayed_work(dw);
1500                 }
1501         }
1502         put_online_cpus();
1503
1504         schedule_delayed_work(&shepherd,
1505                 round_jiffies_relative(sysctl_stat_interval));
1506 }
1507
1508 static void __init start_shepherd_timer(void)
1509 {
1510         int cpu;
1511
1512         for_each_possible_cpu(cpu)
1513                 INIT_DELAYED_WORK(per_cpu_ptr(&vmstat_work, cpu),
1514                         vmstat_update);
1515
1516         if (!alloc_cpumask_var(&cpu_stat_off, GFP_KERNEL))
1517                 BUG();
1518         cpumask_copy(cpu_stat_off, cpu_online_mask);
1519
1520         vmstat_wq = alloc_workqueue("vmstat", WQ_FREEZABLE|WQ_MEM_RECLAIM, 0);
1521         schedule_delayed_work(&shepherd,
1522                 round_jiffies_relative(sysctl_stat_interval));
1523 }
1524
1525 static void vmstat_cpu_dead(int node)
1526 {
1527         int cpu;
1528
1529         get_online_cpus();
1530         for_each_online_cpu(cpu)
1531                 if (cpu_to_node(cpu) == node)
1532                         goto end;
1533
1534         node_clear_state(node, N_CPU);
1535 end:
1536         put_online_cpus();
1537 }
1538
1539 /*
1540  * Use the cpu notifier to insure that the thresholds are recalculated
1541  * when necessary.
1542  */
1543 static int vmstat_cpuup_callback(struct notifier_block *nfb,
1544                 unsigned long action,
1545                 void *hcpu)
1546 {
1547         long cpu = (long)hcpu;
1548
1549         switch (action) {
1550         case CPU_ONLINE:
1551         case CPU_ONLINE_FROZEN:
1552                 refresh_zone_stat_thresholds();
1553                 node_set_state(cpu_to_node(cpu), N_CPU);
1554                 cpumask_set_cpu(cpu, cpu_stat_off);
1555                 break;
1556         case CPU_DOWN_PREPARE:
1557         case CPU_DOWN_PREPARE_FROZEN:
1558                 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1559                 cpumask_clear_cpu(cpu, cpu_stat_off);
1560                 break;
1561         case CPU_DOWN_FAILED:
1562         case CPU_DOWN_FAILED_FROZEN:
1563                 cpumask_set_cpu(cpu, cpu_stat_off);
1564                 break;
1565         case CPU_DEAD:
1566         case CPU_DEAD_FROZEN:
1567                 refresh_zone_stat_thresholds();
1568                 vmstat_cpu_dead(cpu_to_node(cpu));
1569                 break;
1570         default:
1571                 break;
1572         }
1573         return NOTIFY_OK;
1574 }
1575
1576 static struct notifier_block vmstat_notifier =
1577         { &vmstat_cpuup_callback, NULL, 0 };
1578 #endif
1579
1580 static int __init setup_vmstat(void)
1581 {
1582 #ifdef CONFIG_SMP
1583         cpu_notifier_register_begin();
1584         __register_cpu_notifier(&vmstat_notifier);
1585
1586         start_shepherd_timer();
1587         cpu_notifier_register_done();
1588 #endif
1589 #ifdef CONFIG_PROC_FS
1590         proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1591         proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1592         proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1593         proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1594 #endif
1595         return 0;
1596 }
1597 module_init(setup_vmstat)
1598
1599 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1600
1601 /*
1602  * Return an index indicating how much of the available free memory is
1603  * unusable for an allocation of the requested size.
1604  */
1605 static int unusable_free_index(unsigned int order,
1606                                 struct contig_page_info *info)
1607 {
1608         /* No free memory is interpreted as all free memory is unusable */
1609         if (info->free_pages == 0)
1610                 return 1000;
1611
1612         /*
1613          * Index should be a value between 0 and 1. Return a value to 3
1614          * decimal places.
1615          *
1616          * 0 => no fragmentation
1617          * 1 => high fragmentation
1618          */
1619         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1620
1621 }
1622
1623 static void unusable_show_print(struct seq_file *m,
1624                                         pg_data_t *pgdat, struct zone *zone)
1625 {
1626         unsigned int order;
1627         int index;
1628         struct contig_page_info info;
1629
1630         seq_printf(m, "Node %d, zone %8s ",
1631                                 pgdat->node_id,
1632                                 zone->name);
1633         for (order = 0; order < MAX_ORDER; ++order) {
1634                 fill_contig_page_info(zone, order, &info);
1635                 index = unusable_free_index(order, &info);
1636                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1637         }
1638
1639         seq_putc(m, '\n');
1640 }
1641
1642 /*
1643  * Display unusable free space index
1644  *
1645  * The unusable free space index measures how much of the available free
1646  * memory cannot be used to satisfy an allocation of a given size and is a
1647  * value between 0 and 1. The higher the value, the more of free memory is
1648  * unusable and by implication, the worse the external fragmentation is. This
1649  * can be expressed as a percentage by multiplying by 100.
1650  */
1651 static int unusable_show(struct seq_file *m, void *arg)
1652 {
1653         pg_data_t *pgdat = (pg_data_t *)arg;
1654
1655         /* check memoryless node */
1656         if (!node_state(pgdat->node_id, N_MEMORY))
1657                 return 0;
1658
1659         walk_zones_in_node(m, pgdat, false, unusable_show_print);
1660
1661         return 0;
1662 }
1663
1664 static const struct seq_operations unusable_op = {
1665         .start  = frag_start,
1666         .next   = frag_next,
1667         .stop   = frag_stop,
1668         .show   = unusable_show,
1669 };
1670
1671 static int unusable_open(struct inode *inode, struct file *file)
1672 {
1673         return seq_open(file, &unusable_op);
1674 }
1675
1676 static const struct file_operations unusable_file_ops = {
1677         .open           = unusable_open,
1678         .read           = seq_read,
1679         .llseek         = seq_lseek,
1680         .release        = seq_release,
1681 };
1682
1683 static void extfrag_show_print(struct seq_file *m,
1684                                         pg_data_t *pgdat, struct zone *zone)
1685 {
1686         unsigned int order;
1687         int index;
1688
1689         /* Alloc on stack as interrupts are disabled for zone walk */
1690         struct contig_page_info info;
1691
1692         seq_printf(m, "Node %d, zone %8s ",
1693                                 pgdat->node_id,
1694                                 zone->name);
1695         for (order = 0; order < MAX_ORDER; ++order) {
1696                 fill_contig_page_info(zone, order, &info);
1697                 index = __fragmentation_index(order, &info);
1698                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1699         }
1700
1701         seq_putc(m, '\n');
1702 }
1703
1704 /*
1705  * Display fragmentation index for orders that allocations would fail for
1706  */
1707 static int extfrag_show(struct seq_file *m, void *arg)
1708 {
1709         pg_data_t *pgdat = (pg_data_t *)arg;
1710
1711         walk_zones_in_node(m, pgdat, false, extfrag_show_print);
1712
1713         return 0;
1714 }
1715
1716 static const struct seq_operations extfrag_op = {
1717         .start  = frag_start,
1718         .next   = frag_next,
1719         .stop   = frag_stop,
1720         .show   = extfrag_show,
1721 };
1722
1723 static int extfrag_open(struct inode *inode, struct file *file)
1724 {
1725         return seq_open(file, &extfrag_op);
1726 }
1727
1728 static const struct file_operations extfrag_file_ops = {
1729         .open           = extfrag_open,
1730         .read           = seq_read,
1731         .llseek         = seq_lseek,
1732         .release        = seq_release,
1733 };
1734
1735 static int __init extfrag_debug_init(void)
1736 {
1737         struct dentry *extfrag_debug_root;
1738
1739         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1740         if (!extfrag_debug_root)
1741                 return -ENOMEM;
1742
1743         if (!debugfs_create_file("unusable_index", 0444,
1744                         extfrag_debug_root, NULL, &unusable_file_ops))
1745                 goto fail;
1746
1747         if (!debugfs_create_file("extfrag_index", 0444,
1748                         extfrag_debug_root, NULL, &extfrag_file_ops))
1749                 goto fail;
1750
1751         return 0;
1752 fail:
1753         debugfs_remove_recursive(extfrag_debug_root);
1754         return -ENOMEM;
1755 }
1756
1757 module_init(extfrag_debug_init);
1758 #endif