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 #if IS_ENABLED(CONFIG_ZSMALLOC)
 722         "nr_zspages",
 723 #endif
 724 #ifdef CONFIG_NUMA
 725         "numa_hit",
 726         "numa_miss",
 727         "numa_foreign",
 728         "numa_interleave",
 729         "numa_local",
 730         "numa_other",
 731 #endif
 732         "workingset_refault",
 733         "workingset_activate",
 734         "workingset_nodereclaim",
 735         "nr_anon_transparent_hugepages",
 736         "nr_free_cma",
 737
 738         /* enum writeback_stat_item counters */
 739         "nr_dirty_threshold",
 740         "nr_dirty_background_threshold",
 741
 742 #ifdef CONFIG_VM_EVENT_COUNTERS
 743         /* enum vm_event_item counters */
 744         "pgpgin",
 745         "pgpgout",
 746         "pswpin",
 747         "pswpout",
 748
 749         TEXTS_FOR_ZONES("pgalloc")
 750
 751         "pgfree",
 752         "pgactivate",
 753         "pgdeactivate",
 754
 755         "pgfault",
 756         "pgmajfault",
 757         "pglazyfreed",
 758
 759         TEXTS_FOR_ZONES("pgrefill")
 760         TEXTS_FOR_ZONES("pgsteal_kswapd")
 761         TEXTS_FOR_ZONES("pgsteal_direct")
 762         TEXTS_FOR_ZONES("pgscan_kswapd")
 763         TEXTS_FOR_ZONES("pgscan_direct")
 764         "pgscan_direct_throttle",
 765
 766 #ifdef CONFIG_NUMA
 767         "zone_reclaim_failed",
 768 #endif
 769         "pginodesteal",
 770         "slabs_scanned",
 771         "kswapd_inodesteal",
 772         "kswapd_low_wmark_hit_quickly",
 773         "kswapd_high_wmark_hit_quickly",
 774         "pageoutrun",
 775         "allocstall",
 776
 777         "pgrotated",
 778
 779         "drop_pagecache",
 780         "drop_slab",
 781
 782 #ifdef CONFIG_NUMA_BALANCING
 783         "numa_pte_updates",
 784         "numa_huge_pte_updates",
 785         "numa_hint_faults",
 786         "numa_hint_faults_local",
 787         "numa_pages_migrated",
 788 #endif
 789 #ifdef CONFIG_MIGRATION
 790         "pgmigrate_success",
 791         "pgmigrate_fail",
 792 #endif
 793 #ifdef CONFIG_COMPACTION
 794         "compact_migrate_scanned",
 795         "compact_free_scanned",
 796         "compact_isolated",
 797         "compact_stall",
 798         "compact_fail",
 799         "compact_success",
 800         "compact_daemon_wake",
 801 #endif
 802
 803 #ifdef CONFIG_HUGETLB_PAGE
 804         "htlb_buddy_alloc_success",
 805         "htlb_buddy_alloc_fail",
 806 #endif
 807         "unevictable_pgs_culled",
 808         "unevictable_pgs_scanned",
 809         "unevictable_pgs_rescued",
 810         "unevictable_pgs_mlocked",
 811         "unevictable_pgs_munlocked",
 812         "unevictable_pgs_cleared",
 813         "unevictable_pgs_stranded",
 814
 815 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 816         "thp_fault_alloc",
 817         "thp_fault_fallback",
 818         "thp_collapse_alloc",
 819         "thp_collapse_alloc_failed",
 820         "thp_file_alloc",
 821         "thp_file_mapped",
 822         "thp_split_page",
 823         "thp_split_page_failed",
 824         "thp_deferred_split_page",
 825         "thp_split_pmd",
 826         "thp_zero_page_alloc",
 827         "thp_zero_page_alloc_failed",
 828 #endif
 829 #ifdef CONFIG_MEMORY_BALLOON
 830         "balloon_inflate",
 831         "balloon_deflate",
 832 #ifdef CONFIG_BALLOON_COMPACTION
 833         "balloon_migrate",
 834 #endif
 835 #endif /* CONFIG_MEMORY_BALLOON */
 836 #ifdef CONFIG_DEBUG_TLBFLUSH
 837 #ifdef CONFIG_SMP
 838         "nr_tlb_remote_flush",
 839         "nr_tlb_remote_flush_received",
 840 #endif /* CONFIG_SMP */
 841         "nr_tlb_local_flush_all",
 842         "nr_tlb_local_flush_one",
 843 #endif /* CONFIG_DEBUG_TLBFLUSH */
 844
 845 #ifdef CONFIG_DEBUG_VM_VMACACHE
 846         "vmacache_find_calls",
 847         "vmacache_find_hits",
 848         "vmacache_full_flushes",
 849 #endif
 850 #endif /* CONFIG_VM_EVENTS_COUNTERS */
 851 };
 852 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
 853
 854
 855 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
 856      defined(CONFIG_PROC_FS)
 857 static void *frag_start(struct seq_file *m, loff_t *pos)
 858 {
 859         pg_data_t *pgdat;
 860         loff_t node = *pos;
 861
 862         for (pgdat = first_online_pgdat();
 863              pgdat && node;
 864              pgdat = next_online_pgdat(pgdat))
 865                 --node;
 866
 867         return pgdat;
 868 }
 869
 870 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
 871 {
 872         pg_data_t *pgdat = (pg_data_t *)arg;
 873
 874         (*pos)++;
 875         return next_online_pgdat(pgdat);
 876 }
 877
 878 static void frag_stop(struct seq_file *m, void *arg)
 879 {
 880 }
 881
 882 /* Walk all the zones in a node and print using a callback */
 883 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
 884                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
 885 {
 886         struct zone *zone;
 887         struct zone *node_zones = pgdat->node_zones;
 888         unsigned long flags;
 889
 890         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 891                 if (!populated_zone(zone))
 892                         continue;
 893
 894                 spin_lock_irqsave(&zone->lock, flags);
 895                 print(m, pgdat, zone);
 896                 spin_unlock_irqrestore(&zone->lock, flags);
 897         }
 898 }
 899 #endif
 900
 901 #ifdef CONFIG_PROC_FS
 902 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
 903                                                 struct zone *zone)
 904 {
 905         int order;
 906
 907         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 908         for (order = 0; order < MAX_ORDER; ++order)
 909                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
 910         seq_putc(m, '\n');
 911 }
 912
 913 /*
 914  * This walks the free areas for each zone.
 915  */
 916 static int frag_show(struct seq_file *m, void *arg)
 917 {
 918         pg_data_t *pgdat = (pg_data_t *)arg;
 919         walk_zones_in_node(m, pgdat, frag_show_print);
 920         return 0;
 921 }
 922
 923 static void pagetypeinfo_showfree_print(struct seq_file *m,
 924                                         pg_data_t *pgdat, struct zone *zone)
 925 {
 926         int order, mtype;
 927
 928         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
 929                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
 930                                         pgdat->node_id,
 931                                         zone->name,
 932                                         migratetype_names[mtype]);
 933                 for (order = 0; order < MAX_ORDER; ++order) {
 934                         unsigned long freecount = 0;
 935                         struct free_area *area;
 936                         struct list_head *curr;
 937
 938                         area = &(zone->free_area[order]);
 939
 940                         list_for_each(curr, &area->free_list[mtype])
 941                                 freecount++;
 942                         seq_printf(m, "%6lu ", freecount);
 943                 }
 944                 seq_putc(m, '\n');
 945         }
 946 }
 947
 948 /* Print out the free pages at each order for each migatetype */
 949 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
 950 {
 951         int order;
 952         pg_data_t *pgdat = (pg_data_t *)arg;
 953
 954         /* Print header */
 955         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
 956         for (order = 0; order < MAX_ORDER; ++order)
 957                 seq_printf(m, "%6d ", order);
 958         seq_putc(m, '\n');
 959
 960         walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
 961
 962         return 0;
 963 }
 964
 965 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
 966                                         pg_data_t *pgdat, struct zone *zone)
 967 {
 968         int mtype;
 969         unsigned long pfn;
 970         unsigned long start_pfn = zone->zone_start_pfn;
 971         unsigned long end_pfn = zone_end_pfn(zone);
 972         unsigned long count[MIGRATE_TYPES] = { 0, };
 973
 974         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
 975                 struct page *page;
 976
 977                 if (!pfn_valid(pfn))
 978                         continue;
 979
 980                 page = pfn_to_page(pfn);
 981
 982                 /* Watch for unexpected holes punched in the memmap */
 983                 if (!memmap_valid_within(pfn, page, zone))
 984                         continue;
 985
 986                 if (page_zone(page) != zone)
 987                         continue;
 988
 989                 mtype = get_pageblock_migratetype(page);
 990
 991                 if (mtype < MIGRATE_TYPES)
 992                         count[mtype]++;
 993         }
 994
 995         /* Print counts */
 996         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 997         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
 998                 seq_printf(m, "%12lu ", count[mtype]);
 999         seq_putc(m, '\n');
1000 }
1001
1002 /* Print out the free pages at each order for each migratetype */
1003 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1004 {
1005         int mtype;
1006         pg_data_t *pgdat = (pg_data_t *)arg;
1007
1008         seq_printf(m, "\n%-23s", "Number of blocks type ");
1009         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1010                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1011         seq_putc(m, '\n');
1012         walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
1013
1014         return 0;
1015 }
1016
1017 #ifdef CONFIG_PAGE_OWNER
1018 static void pagetypeinfo_showmixedcount_print(struct seq_file *m,
1019                                                         pg_data_t *pgdat,
1020                                                         struct zone *zone)
1021 {
1022         struct page *page;
1023         struct page_ext *page_ext;
1024         unsigned long pfn = zone->zone_start_pfn, block_end_pfn;
1025         unsigned long end_pfn = pfn + zone->spanned_pages;
1026         unsigned long count[MIGRATE_TYPES] = { 0, };
1027         int pageblock_mt, page_mt;
1028         int i;
1029
1030         /* Scan block by block. First and last block may be incomplete */
1031         pfn = zone->zone_start_pfn;
1032
1033         /*
1034          * Walk the zone in pageblock_nr_pages steps. If a page block spans
1035          * a zone boundary, it will be double counted between zones. This does
1036          * not matter as the mixed block count will still be correct
1037          */
1038         for (; pfn < end_pfn; ) {
1039                 if (!pfn_valid(pfn)) {
1040                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
1041                         continue;
1042                 }
1043
1044                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
1045                 block_end_pfn = min(block_end_pfn, end_pfn);
1046
1047                 page = pfn_to_page(pfn);
1048                 pageblock_mt = get_pageblock_migratetype(page);
1049
1050                 for (; pfn < block_end_pfn; pfn++) {
1051                         if (!pfn_valid_within(pfn))
1052                                 continue;
1053
1054                         page = pfn_to_page(pfn);
1055
1056                         if (page_zone(page) != zone)
1057                                 continue;
1058
1059                         if (PageBuddy(page)) {
1060                                 pfn += (1UL << page_order(page)) - 1;
1061                                 continue;
1062                         }
1063
1064                         if (PageReserved(page))
1065                                 continue;
1066
1067                         page_ext = lookup_page_ext(page);
1068                         if (unlikely(!page_ext))
1069                                 continue;
1070
1071                         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
1072                                 continue;
1073
1074                         page_mt = gfpflags_to_migratetype(page_ext->gfp_mask);
1075                         if (pageblock_mt != page_mt) {
1076                                 if (is_migrate_cma(pageblock_mt))
1077                                         count[MIGRATE_MOVABLE]++;
1078                                 else
1079                                         count[pageblock_mt]++;
1080
1081                                 pfn = block_end_pfn;
1082                                 break;
1083                         }
1084                         pfn += (1UL << page_ext->order) - 1;
1085                 }
1086         }
1087
1088         /* Print counts */
1089         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1090         for (i = 0; i < MIGRATE_TYPES; i++)
1091                 seq_printf(m, "%12lu ", count[i]);
1092         seq_putc(m, '\n');
1093 }
1094 #endif /* CONFIG_PAGE_OWNER */
1095
1096 /*
1097  * Print out the number of pageblocks for each migratetype that contain pages
1098  * of other types. This gives an indication of how well fallbacks are being
1099  * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1100  * to determine what is going on
1101  */
1102 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1103 {
1104 #ifdef CONFIG_PAGE_OWNER
1105         int mtype;
1106
1107         if (!static_branch_unlikely(&page_owner_inited))
1108                 return;
1109
1110         drain_all_pages(NULL);
1111
1112         seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1113         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1114                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1115         seq_putc(m, '\n');
1116
1117         walk_zones_in_node(m, pgdat, pagetypeinfo_showmixedcount_print);
1118 #endif /* CONFIG_PAGE_OWNER */
1119 }
1120
1121 /*
1122  * This prints out statistics in relation to grouping pages by mobility.
1123  * It is expensive to collect so do not constantly read the file.
1124  */
1125 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1126 {
1127         pg_data_t *pgdat = (pg_data_t *)arg;
1128
1129         /* check memoryless node */
1130         if (!node_state(pgdat->node_id, N_MEMORY))
1131                 return 0;
1132
1133         seq_printf(m, "Page block order: %d\n", pageblock_order);
1134         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
1135         seq_putc(m, '\n');
1136         pagetypeinfo_showfree(m, pgdat);
1137         pagetypeinfo_showblockcount(m, pgdat);
1138         pagetypeinfo_showmixedcount(m, pgdat);
1139
1140         return 0;
1141 }
1142
1143 static const struct seq_operations fragmentation_op = {
1144         .start  = frag_start,
1145         .next   = frag_next,
1146         .stop   = frag_stop,
1147         .show   = frag_show,
1148 };
1149
1150 static int fragmentation_open(struct inode *inode, struct file *file)
1151 {
1152         return seq_open(file, &fragmentation_op);
1153 }
1154
1155 static const struct file_operations fragmentation_file_operations = {
1156         .open           = fragmentation_open,
1157         .read           = seq_read,
1158         .llseek         = seq_lseek,
1159         .release        = seq_release,
1160 };
1161
1162 static const struct seq_operations pagetypeinfo_op = {
1163         .start  = frag_start,
1164         .next   = frag_next,
1165         .stop   = frag_stop,
1166         .show   = pagetypeinfo_show,
1167 };
1168
1169 static int pagetypeinfo_open(struct inode *inode, struct file *file)
1170 {
1171         return seq_open(file, &pagetypeinfo_op);
1172 }
1173
1174 static const struct file_operations pagetypeinfo_file_ops = {
1175         .open           = pagetypeinfo_open,
1176         .read           = seq_read,
1177         .llseek         = seq_lseek,
1178         .release        = seq_release,
1179 };
1180
1181 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1182                                                         struct zone *zone)
1183 {
1184         int i;
1185         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1186         seq_printf(m,
1187                    "\n  pages free     %lu"
1188                    "\n        min      %lu"
1189                    "\n        low      %lu"
1190                    "\n        high     %lu"
1191                    "\n        scanned  %lu"
1192                    "\n        spanned  %lu"
1193                    "\n        present  %lu"
1194                    "\n        managed  %lu",
1195                    zone_page_state(zone, NR_FREE_PAGES),
1196                    min_wmark_pages(zone),
1197                    low_wmark_pages(zone),
1198                    high_wmark_pages(zone),
1199                    zone_page_state(zone, NR_PAGES_SCANNED),
1200                    zone->spanned_pages,
1201                    zone->present_pages,
1202                    zone->managed_pages);
1203
1204         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1205                 seq_printf(m, "\n    %-12s %lu", vmstat_text[i],
1206                                 zone_page_state(zone, i));
1207
1208         seq_printf(m,
1209                    "\n        protection: (%ld",
1210                    zone->lowmem_reserve[0]);
1211         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1212                 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1213         seq_printf(m,
1214                    ")"
1215                    "\n  pagesets");
1216         for_each_online_cpu(i) {
1217                 struct per_cpu_pageset *pageset;
1218
1219                 pageset = per_cpu_ptr(zone->pageset, i);
1220                 seq_printf(m,
1221                            "\n    cpu: %i"
1222                            "\n              count: %i"
1223                            "\n              high:  %i"
1224                            "\n              batch: %i",
1225                            i,
1226                            pageset->pcp.count,
1227                            pageset->pcp.high,
1228                            pageset->pcp.batch);
1229 #ifdef CONFIG_SMP
1230                 seq_printf(m, "\n  vm stats threshold: %d",
1231                                 pageset->stat_threshold);
1232 #endif
1233         }
1234         seq_printf(m,
1235                    "\n  all_unreclaimable: %u"
1236                    "\n  start_pfn:         %lu"
1237                    "\n  inactive_ratio:    %u",
1238                    !zone_reclaimable(zone),
1239                    zone->zone_start_pfn,
1240                    zone->inactive_ratio);
1241         seq_putc(m, '\n');
1242 }
1243
1244 /*
1245  * Output information about zones in @pgdat.
1246  */
1247 static int zoneinfo_show(struct seq_file *m, void *arg)
1248 {
1249         pg_data_t *pgdat = (pg_data_t *)arg;
1250         walk_zones_in_node(m, pgdat, zoneinfo_show_print);
1251         return 0;
1252 }
1253
1254 static const struct seq_operations zoneinfo_op = {
1255         .start  = frag_start, /* iterate over all zones. The same as in
1256                                * fragmentation. */
1257         .next   = frag_next,
1258         .stop   = frag_stop,
1259         .show   = zoneinfo_show,
1260 };
1261
1262 static int zoneinfo_open(struct inode *inode, struct file *file)
1263 {
1264         return seq_open(file, &zoneinfo_op);
1265 }
1266
1267 static const struct file_operations proc_zoneinfo_file_operations = {
1268         .open           = zoneinfo_open,
1269         .read           = seq_read,
1270         .llseek         = seq_lseek,
1271         .release        = seq_release,
1272 };
1273
1274 enum writeback_stat_item {
1275         NR_DIRTY_THRESHOLD,
1276         NR_DIRTY_BG_THRESHOLD,
1277         NR_VM_WRITEBACK_STAT_ITEMS,
1278 };
1279
1280 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1281 {
1282         unsigned long *v;
1283         int i, stat_items_size;
1284
1285         if (*pos >= ARRAY_SIZE(vmstat_text))
1286                 return NULL;
1287         stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1288                           NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1289
1290 #ifdef CONFIG_VM_EVENT_COUNTERS
1291         stat_items_size += sizeof(struct vm_event_state);
1292 #endif
1293
1294         v = kmalloc(stat_items_size, GFP_KERNEL);
1295         m->private = v;
1296         if (!v)
1297                 return ERR_PTR(-ENOMEM);
1298         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1299                 v[i] = global_page_state(i);
1300         v += NR_VM_ZONE_STAT_ITEMS;
1301
1302         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1303                             v + NR_DIRTY_THRESHOLD);
1304         v += NR_VM_WRITEBACK_STAT_ITEMS;
1305
1306 #ifdef CONFIG_VM_EVENT_COUNTERS
1307         all_vm_events(v);
1308         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1309         v[PGPGOUT] /= 2;
1310 #endif
1311         return (unsigned long *)m->private + *pos;
1312 }
1313
1314 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1315 {
1316         (*pos)++;
1317         if (*pos >= ARRAY_SIZE(vmstat_text))
1318                 return NULL;
1319         return (unsigned long *)m->private + *pos;
1320 }
1321
1322 static int vmstat_show(struct seq_file *m, void *arg)
1323 {
1324         unsigned long *l = arg;
1325         unsigned long off = l - (unsigned long *)m->private;
1326
1327         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1328         return 0;
1329 }
1330
1331 static void vmstat_stop(struct seq_file *m, void *arg)
1332 {
1333         kfree(m->private);
1334         m->private = NULL;
1335 }
1336
1337 static const struct seq_operations vmstat_op = {
1338         .start  = vmstat_start,
1339         .next   = vmstat_next,
1340         .stop   = vmstat_stop,
1341         .show   = vmstat_show,
1342 };
1343
1344 static int vmstat_open(struct inode *inode, struct file *file)
1345 {
1346         return seq_open(file, &vmstat_op);
1347 }
1348
1349 static const struct file_operations proc_vmstat_file_operations = {
1350         .open           = vmstat_open,
1351         .read           = seq_read,
1352         .llseek         = seq_lseek,
1353         .release        = seq_release,
1354 };
1355 #endif /* CONFIG_PROC_FS */
1356
1357 #ifdef CONFIG_SMP
1358 static struct workqueue_struct *vmstat_wq;
1359 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1360 int sysctl_stat_interval __read_mostly = HZ;
1361
1362 #ifdef CONFIG_PROC_FS
1363 static void refresh_vm_stats(struct work_struct *work)
1364 {
1365         refresh_cpu_vm_stats(true);
1366 }
1367
1368 int vmstat_refresh(struct ctl_table *table, int write,
1369                    void __user *buffer, size_t *lenp, loff_t *ppos)
1370 {
1371         long val;
1372         int err;
1373         int i;
1374
1375         /*
1376          * The regular update, every sysctl_stat_interval, may come later
1377          * than expected: leaving a significant amount in per_cpu buckets.
1378          * This is particularly misleading when checking a quantity of HUGE
1379          * pages, immediately after running a test.  /proc/sys/vm/stat_refresh,
1380          * which can equally be echo'ed to or cat'ted from (by root),
1381          * can be used to update the stats just before reading them.
1382          *
1383          * Oh, and since global_page_state() etc. are so careful to hide
1384          * transiently negative values, report an error here if any of
1385          * the stats is negative, so we know to go looking for imbalance.
1386          */
1387         err = schedule_on_each_cpu(refresh_vm_stats);
1388         if (err)
1389                 return err;
1390         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1391                 val = atomic_long_read(&vm_stat[i]);
1392                 if (val < 0) {
1393                         switch (i) {
1394                         case NR_ALLOC_BATCH:
1395                         case NR_PAGES_SCANNED:
1396                                 /*
1397                                  * These are often seen to go negative in
1398                                  * recent kernels, but not to go permanently
1399                                  * negative.  Whilst it would be nicer not to
1400                                  * have exceptions, rooting them out would be
1401                                  * another task, of rather low priority.
1402                                  */
1403                                 break;
1404                         default:
1405                                 pr_warn("%s: %s %ld\n",
1406                                         __func__, vmstat_text[i], val);
1407                                 err = -EINVAL;
1408                                 break;
1409                         }
1410                 }
1411         }
1412         if (err)
1413                 return err;
1414         if (write)
1415                 *ppos += *lenp;
1416         else
1417                 *lenp = 0;
1418         return 0;
1419 }
1420 #endif /* CONFIG_PROC_FS */
1421
1422 static void vmstat_update(struct work_struct *w)
1423 {
1424         if (refresh_cpu_vm_stats(true)) {
1425                 /*
1426                  * Counters were updated so we expect more updates
1427                  * to occur in the future. Keep on running the
1428                  * update worker thread.
1429                  */
1430                 queue_delayed_work_on(smp_processor_id(), vmstat_wq,
1431                                 this_cpu_ptr(&vmstat_work),
1432                                 round_jiffies_relative(sysctl_stat_interval));
1433         }
1434 }
1435
1436 /*
1437  * Switch off vmstat processing and then fold all the remaining differentials
1438  * until the diffs stay at zero. The function is used by NOHZ and can only be
1439  * invoked when tick processing is not active.
1440  */
1441 /*
1442  * Check if the diffs for a certain cpu indicate that
1443  * an update is needed.
1444  */
1445 static bool need_update(int cpu)
1446 {
1447         struct zone *zone;
1448
1449         for_each_populated_zone(zone) {
1450                 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1451
1452                 BUILD_BUG_ON(sizeof(p->vm_stat_diff[0]) != 1);
1453                 /*
1454                  * The fast way of checking if there are any vmstat diffs.
1455                  * This works because the diffs are byte sized items.
1456                  */
1457                 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS))
1458                         return true;
1459
1460         }
1461         return false;
1462 }
1463
1464 /*
1465  * Switch off vmstat processing and then fold all the remaining differentials
1466  * until the diffs stay at zero. The function is used by NOHZ and can only be
1467  * invoked when tick processing is not active.
1468  */
1469 void quiet_vmstat(void)
1470 {
1471         if (system_state != SYSTEM_RUNNING)
1472                 return;
1473
1474         if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
1475                 return;
1476
1477         if (!need_update(smp_processor_id()))
1478                 return;
1479
1480         /*
1481          * Just refresh counters and do not care about the pending delayed
1482          * vmstat_update. It doesn't fire that often to matter and canceling
1483          * it would be too expensive from this path.
1484          * vmstat_shepherd will take care about that for us.
1485          */
1486         refresh_cpu_vm_stats(false);
1487 }
1488
1489 /*
1490  * Shepherd worker thread that checks the
1491  * differentials of processors that have their worker
1492  * threads for vm statistics updates disabled because of
1493  * inactivity.
1494  */
1495 static void vmstat_shepherd(struct work_struct *w);
1496
1497 static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
1498
1499 static void vmstat_shepherd(struct work_struct *w)
1500 {
1501         int cpu;
1502
1503         get_online_cpus();
1504         /* Check processors whose vmstat worker threads have been disabled */
1505         for_each_online_cpu(cpu) {
1506                 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
1507
1508                 if (!delayed_work_pending(dw) && need_update(cpu))
1509                         queue_delayed_work_on(cpu, vmstat_wq, dw, 0);
1510         }
1511         put_online_cpus();
1512
1513         schedule_delayed_work(&shepherd,
1514                 round_jiffies_relative(sysctl_stat_interval));
1515 }
1516
1517 static void __init start_shepherd_timer(void)
1518 {
1519         int cpu;
1520
1521         for_each_possible_cpu(cpu)
1522                 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
1523                         vmstat_update);
1524
1525         vmstat_wq = alloc_workqueue("vmstat", WQ_FREEZABLE|WQ_MEM_RECLAIM, 0);
1526         schedule_delayed_work(&shepherd,
1527                 round_jiffies_relative(sysctl_stat_interval));
1528 }
1529
1530 static void vmstat_cpu_dead(int node)
1531 {
1532         int cpu;
1533
1534         get_online_cpus();
1535         for_each_online_cpu(cpu)
1536                 if (cpu_to_node(cpu) == node)
1537                         goto end;
1538
1539         node_clear_state(node, N_CPU);
1540 end:
1541         put_online_cpus();
1542 }
1543
1544 /*
1545  * Use the cpu notifier to insure that the thresholds are recalculated
1546  * when necessary.
1547  */
1548 static int vmstat_cpuup_callback(struct notifier_block *nfb,
1549                 unsigned long action,
1550                 void *hcpu)
1551 {
1552         long cpu = (long)hcpu;
1553
1554         switch (action) {
1555         case CPU_ONLINE:
1556         case CPU_ONLINE_FROZEN:
1557                 refresh_zone_stat_thresholds();
1558                 node_set_state(cpu_to_node(cpu), N_CPU);
1559                 break;
1560         case CPU_DOWN_PREPARE:
1561         case CPU_DOWN_PREPARE_FROZEN:
1562                 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1563                 break;
1564         case CPU_DOWN_FAILED:
1565         case CPU_DOWN_FAILED_FROZEN:
1566                 break;
1567         case CPU_DEAD:
1568         case CPU_DEAD_FROZEN:
1569                 refresh_zone_stat_thresholds();
1570                 vmstat_cpu_dead(cpu_to_node(cpu));
1571                 break;
1572         default:
1573                 break;
1574         }
1575         return NOTIFY_OK;
1576 }
1577
1578 static struct notifier_block vmstat_notifier =
1579         { &vmstat_cpuup_callback, NULL, 0 };
1580 #endif
1581
1582 static int __init setup_vmstat(void)
1583 {
1584 #ifdef CONFIG_SMP
1585         cpu_notifier_register_begin();
1586         __register_cpu_notifier(&vmstat_notifier);
1587
1588         start_shepherd_timer();
1589         cpu_notifier_register_done();
1590 #endif
1591 #ifdef CONFIG_PROC_FS
1592         proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1593         proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1594         proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1595         proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1596 #endif
1597         return 0;
1598 }
1599 module_init(setup_vmstat)
1600
1601 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1602
1603 /*
1604  * Return an index indicating how much of the available free memory is
1605  * unusable for an allocation of the requested size.
1606  */
1607 static int unusable_free_index(unsigned int order,
1608                                 struct contig_page_info *info)
1609 {
1610         /* No free memory is interpreted as all free memory is unusable */
1611         if (info->free_pages == 0)
1612                 return 1000;
1613
1614         /*
1615          * Index should be a value between 0 and 1. Return a value to 3
1616          * decimal places.
1617          *
1618          * 0 => no fragmentation
1619          * 1 => high fragmentation
1620          */
1621         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1622
1623 }
1624
1625 static void unusable_show_print(struct seq_file *m,
1626                                         pg_data_t *pgdat, struct zone *zone)
1627 {
1628         unsigned int order;
1629         int index;
1630         struct contig_page_info info;
1631
1632         seq_printf(m, "Node %d, zone %8s ",
1633                                 pgdat->node_id,
1634                                 zone->name);
1635         for (order = 0; order < MAX_ORDER; ++order) {
1636                 fill_contig_page_info(zone, order, &info);
1637                 index = unusable_free_index(order, &info);
1638                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1639         }
1640
1641         seq_putc(m, '\n');
1642 }
1643
1644 /*
1645  * Display unusable free space index
1646  *
1647  * The unusable free space index measures how much of the available free
1648  * memory cannot be used to satisfy an allocation of a given size and is a
1649  * value between 0 and 1. The higher the value, the more of free memory is
1650  * unusable and by implication, the worse the external fragmentation is. This
1651  * can be expressed as a percentage by multiplying by 100.
1652  */
1653 static int unusable_show(struct seq_file *m, void *arg)
1654 {
1655         pg_data_t *pgdat = (pg_data_t *)arg;
1656
1657         /* check memoryless node */
1658         if (!node_state(pgdat->node_id, N_MEMORY))
1659                 return 0;
1660
1661         walk_zones_in_node(m, pgdat, unusable_show_print);
1662
1663         return 0;
1664 }
1665
1666 static const struct seq_operations unusable_op = {
1667         .start  = frag_start,
1668         .next   = frag_next,
1669         .stop   = frag_stop,
1670         .show   = unusable_show,
1671 };
1672
1673 static int unusable_open(struct inode *inode, struct file *file)
1674 {
1675         return seq_open(file, &unusable_op);
1676 }
1677
1678 static const struct file_operations unusable_file_ops = {
1679         .open           = unusable_open,
1680         .read           = seq_read,
1681         .llseek         = seq_lseek,
1682         .release        = seq_release,
1683 };
1684
1685 static void extfrag_show_print(struct seq_file *m,
1686                                         pg_data_t *pgdat, struct zone *zone)
1687 {
1688         unsigned int order;
1689         int index;
1690
1691         /* Alloc on stack as interrupts are disabled for zone walk */
1692         struct contig_page_info info;
1693
1694         seq_printf(m, "Node %d, zone %8s ",
1695                                 pgdat->node_id,
1696                                 zone->name);
1697         for (order = 0; order < MAX_ORDER; ++order) {
1698                 fill_contig_page_info(zone, order, &info);
1699                 index = __fragmentation_index(order, &info);
1700                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1701         }
1702
1703         seq_putc(m, '\n');
1704 }
1705
1706 /*
1707  * Display fragmentation index for orders that allocations would fail for
1708  */
1709 static int extfrag_show(struct seq_file *m, void *arg)
1710 {
1711         pg_data_t *pgdat = (pg_data_t *)arg;
1712
1713         walk_zones_in_node(m, pgdat, extfrag_show_print);
1714
1715         return 0;
1716 }
1717
1718 static const struct seq_operations extfrag_op = {
1719         .start  = frag_start,
1720         .next   = frag_next,
1721         .stop   = frag_stop,
1722         .show   = extfrag_show,
1723 };
1724
1725 static int extfrag_open(struct inode *inode, struct file *file)
1726 {
1727         return seq_open(file, &extfrag_op);
1728 }
1729
1730 static const struct file_operations extfrag_file_ops = {
1731         .open           = extfrag_open,
1732         .read           = seq_read,
1733         .llseek         = seq_lseek,
1734         .release        = seq_release,
1735 };
1736
1737 static int __init extfrag_debug_init(void)
1738 {
1739         struct dentry *extfrag_debug_root;
1740
1741         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1742         if (!extfrag_debug_root)
1743                 return -ENOMEM;
1744
1745         if (!debugfs_create_file("unusable_index", 0444,
1746                         extfrag_debug_root, NULL, &unusable_file_ops))
1747                 goto fail;
1748
1749         if (!debugfs_create_file("extfrag_index", 0444,
1750                         extfrag_debug_root, NULL, &extfrag_file_ops))
1751                 goto fail;
1752
1753         return 0;
1754 fail:
1755         debugfs_remove_recursive(extfrag_debug_root);
1756         return -ENOMEM;
1757 }
1758
1759 module_init(extfrag_debug_init);
1760 #endif