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