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 #define NUMA_STATS_THRESHOLD (U16_MAX - 2)
  34
  35 #ifdef CONFIG_VM_EVENT_COUNTERS
  36 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
  37 EXPORT_PER_CPU_SYMBOL(vm_event_states);
  38
  39 static void sum_vm_events(unsigned long *ret)
  40 {
  41         int cpu;
  42         int i;
  43
  44         memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
  45
  46         for_each_online_cpu(cpu) {
  47                 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
  48
  49                 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
  50                         ret[i] += this->event[i];
  51         }
  52 }
  53
  54 /*
  55  * Accumulate the vm event counters across all CPUs.
  56  * The result is unavoidably approximate - it can change
  57  * during and after execution of this function.
  58 */
  59 void all_vm_events(unsigned long *ret)
  60 {
  61         get_online_cpus();
  62         sum_vm_events(ret);
  63         put_online_cpus();
  64 }
  65 EXPORT_SYMBOL_GPL(all_vm_events);
  66
  67 /*
  68  * Fold the foreign cpu events into our own.
  69  *
  70  * This is adding to the events on one processor
  71  * but keeps the global counts constant.
  72  */
  73 void vm_events_fold_cpu(int cpu)
  74 {
  75         struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
  76         int i;
  77
  78         for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
  79                 count_vm_events(i, fold_state->event[i]);
  80                 fold_state->event[i] = 0;
  81         }
  82 }
  83
  84 #endif /* CONFIG_VM_EVENT_COUNTERS */
  85
  86 /*
  87  * Manage combined zone based / global counters
  88  *
  89  * vm_stat contains the global counters
  90  */
  91 atomic_long_t vm_zone_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
  92 atomic_long_t vm_numa_stat[NR_VM_NUMA_STAT_ITEMS] __cacheline_aligned_in_smp;
  93 atomic_long_t vm_node_stat[NR_VM_NODE_STAT_ITEMS] __cacheline_aligned_in_smp;
  94 EXPORT_SYMBOL(vm_zone_stat);
  95 EXPORT_SYMBOL(vm_numa_stat);
  96 EXPORT_SYMBOL(vm_node_stat);
  97
  98 #ifdef CONFIG_SMP
  99
 100 int calculate_pressure_threshold(struct zone *zone)
 101 {
 102         int threshold;
 103         int watermark_distance;
 104
 105         /*
 106          * As vmstats are not up to date, there is drift between the estimated
 107          * and real values. For high thresholds and a high number of CPUs, it
 108          * is possible for the min watermark to be breached while the estimated
 109          * value looks fine. The pressure threshold is a reduced value such
 110          * that even the maximum amount of drift will not accidentally breach
 111          * the min watermark
 112          */
 113         watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
 114         threshold = max(1, (int)(watermark_distance / num_online_cpus()));
 115
 116         /*
 117          * Maximum threshold is 125
 118          */
 119         threshold = min(125, threshold);
 120
 121         return threshold;
 122 }
 123
 124 int calculate_normal_threshold(struct zone *zone)
 125 {
 126         int threshold;
 127         int mem;        /* memory in 128 MB units */
 128
 129         /*
 130          * The threshold scales with the number of processors and the amount
 131          * of memory per zone. More memory means that we can defer updates for
 132          * longer, more processors could lead to more contention.
 133          * fls() is used to have a cheap way of logarithmic scaling.
 134          *
 135          * Some sample thresholds:
 136          *
 137          * Threshold    Processors      (fls)   Zonesize        fls(mem+1)
 138          * ------------------------------------------------------------------
 139          * 8            1               1       0.9-1 GB        4
 140          * 16           2               2       0.9-1 GB        4
 141          * 20           2               2       1-2 GB          5
 142          * 24           2               2       2-4 GB          6
 143          * 28           2               2       4-8 GB          7
 144          * 32           2               2       8-16 GB         8
 145          * 4            2               2       <128M           1
 146          * 30           4               3       2-4 GB          5
 147          * 48           4               3       8-16 GB         8
 148          * 32           8               4       1-2 GB          4
 149          * 32           8               4       0.9-1GB         4
 150          * 10           16              5       <128M           1
 151          * 40           16              5       900M            4
 152          * 70           64              7       2-4 GB          5
 153          * 84           64              7       4-8 GB          6
 154          * 108          512             9       4-8 GB          6
 155          * 125          1024            10      8-16 GB         8
 156          * 125          1024            10      16-32 GB        9
 157          */
 158
 159         mem = zone->managed_pages >> (27 - PAGE_SHIFT);
 160
 161         threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
 162
 163         /*
 164          * Maximum threshold is 125
 165          */
 166         threshold = min(125, threshold);
 167
 168         return threshold;
 169 }
 170
 171 /*
 172  * Refresh the thresholds for each zone.
 173  */
 174 void refresh_zone_stat_thresholds(void)
 175 {
 176         struct pglist_data *pgdat;
 177         struct zone *zone;
 178         int cpu;
 179         int threshold;
 180
 181         /* Zero current pgdat thresholds */
 182         for_each_online_pgdat(pgdat) {
 183                 for_each_online_cpu(cpu) {
 184                         per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold = 0;
 185                 }
 186         }
 187
 188         for_each_populated_zone(zone) {
 189                 struct pglist_data *pgdat = zone->zone_pgdat;
 190                 unsigned long max_drift, tolerate_drift;
 191
 192                 threshold = calculate_normal_threshold(zone);
 193
 194                 for_each_online_cpu(cpu) {
 195                         int pgdat_threshold;
 196
 197                         per_cpu_ptr(zone->pageset, cpu)->stat_threshold
 198                                                         = threshold;
 199
 200                         /* Base nodestat threshold on the largest populated zone. */
 201                         pgdat_threshold = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold;
 202                         per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold
 203                                 = max(threshold, pgdat_threshold);
 204                 }
 205
 206                 /*
 207                  * Only set percpu_drift_mark if there is a danger that
 208                  * NR_FREE_PAGES reports the low watermark is ok when in fact
 209                  * the min watermark could be breached by an allocation
 210                  */
 211                 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
 212                 max_drift = num_online_cpus() * threshold;
 213                 if (max_drift > tolerate_drift)
 214                         zone->percpu_drift_mark = high_wmark_pages(zone) +
 215                                         max_drift;
 216         }
 217 }
 218
 219 void set_pgdat_percpu_threshold(pg_data_t *pgdat,
 220                                 int (*calculate_pressure)(struct zone *))
 221 {
 222         struct zone *zone;
 223         int cpu;
 224         int threshold;
 225         int i;
 226
 227         for (i = 0; i < pgdat->nr_zones; i++) {
 228                 zone = &pgdat->node_zones[i];
 229                 if (!zone->percpu_drift_mark)
 230                         continue;
 231
 232                 threshold = (*calculate_pressure)(zone);
 233                 for_each_online_cpu(cpu)
 234                         per_cpu_ptr(zone->pageset, cpu)->stat_threshold
 235                                                         = threshold;
 236         }
 237 }
 238
 239 /*
 240  * For use when we know that interrupts are disabled,
 241  * or when we know that preemption is disabled and that
 242  * particular counter cannot be updated from interrupt context.
 243  */
 244 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 245                            long delta)
 246 {
 247         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 248         s8 __percpu *p = pcp->vm_stat_diff + item;
 249         long x;
 250         long t;
 251
 252         x = delta + __this_cpu_read(*p);
 253
 254         t = __this_cpu_read(pcp->stat_threshold);
 255
 256         if (unlikely(x > t || x < -t)) {
 257                 zone_page_state_add(x, zone, item);
 258                 x = 0;
 259         }
 260         __this_cpu_write(*p, x);
 261 }
 262 EXPORT_SYMBOL(__mod_zone_page_state);
 263
 264 void __mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
 265                                 long delta)
 266 {
 267         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 268         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 269         long x;
 270         long t;
 271
 272         x = delta + __this_cpu_read(*p);
 273
 274         t = __this_cpu_read(pcp->stat_threshold);
 275
 276         if (unlikely(x > t || x < -t)) {
 277                 node_page_state_add(x, pgdat, item);
 278                 x = 0;
 279         }
 280         __this_cpu_write(*p, x);
 281 }
 282 EXPORT_SYMBOL(__mod_node_page_state);
 283
 284 /*
 285  * Optimized increment and decrement functions.
 286  *
 287  * These are only for a single page and therefore can take a struct page *
 288  * argument instead of struct zone *. This allows the inclusion of the code
 289  * generated for page_zone(page) into the optimized functions.
 290  *
 291  * No overflow check is necessary and therefore the differential can be
 292  * incremented or decremented in place which may allow the compilers to
 293  * generate better code.
 294  * The increment or decrement is known and therefore one boundary check can
 295  * be omitted.
 296  *
 297  * NOTE: These functions are very performance sensitive. Change only
 298  * with care.
 299  *
 300  * Some processors have inc/dec instructions that are atomic vs an interrupt.
 301  * However, the code must first determine the differential location in a zone
 302  * based on the processor number and then inc/dec the counter. There is no
 303  * guarantee without disabling preemption that the processor will not change
 304  * in between and therefore the atomicity vs. interrupt cannot be exploited
 305  * in a useful way here.
 306  */
 307 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
 308 {
 309         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 310         s8 __percpu *p = pcp->vm_stat_diff + item;
 311         s8 v, t;
 312
 313         v = __this_cpu_inc_return(*p);
 314         t = __this_cpu_read(pcp->stat_threshold);
 315         if (unlikely(v > t)) {
 316                 s8 overstep = t >> 1;
 317
 318                 zone_page_state_add(v + overstep, zone, item);
 319                 __this_cpu_write(*p, -overstep);
 320         }
 321 }
 322
 323 void __inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 324 {
 325         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 326         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 327         s8 v, t;
 328
 329         v = __this_cpu_inc_return(*p);
 330         t = __this_cpu_read(pcp->stat_threshold);
 331         if (unlikely(v > t)) {
 332                 s8 overstep = t >> 1;
 333
 334                 node_page_state_add(v + overstep, pgdat, item);
 335                 __this_cpu_write(*p, -overstep);
 336         }
 337 }
 338
 339 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
 340 {
 341         __inc_zone_state(page_zone(page), item);
 342 }
 343 EXPORT_SYMBOL(__inc_zone_page_state);
 344
 345 void __inc_node_page_state(struct page *page, enum node_stat_item item)
 346 {
 347         __inc_node_state(page_pgdat(page), item);
 348 }
 349 EXPORT_SYMBOL(__inc_node_page_state);
 350
 351 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
 352 {
 353         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 354         s8 __percpu *p = pcp->vm_stat_diff + item;
 355         s8 v, t;
 356
 357         v = __this_cpu_dec_return(*p);
 358         t = __this_cpu_read(pcp->stat_threshold);
 359         if (unlikely(v < - t)) {
 360                 s8 overstep = t >> 1;
 361
 362                 zone_page_state_add(v - overstep, zone, item);
 363                 __this_cpu_write(*p, overstep);
 364         }
 365 }
 366
 367 void __dec_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 368 {
 369         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 370         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 371         s8 v, t;
 372
 373         v = __this_cpu_dec_return(*p);
 374         t = __this_cpu_read(pcp->stat_threshold);
 375         if (unlikely(v < - t)) {
 376                 s8 overstep = t >> 1;
 377
 378                 node_page_state_add(v - overstep, pgdat, item);
 379                 __this_cpu_write(*p, overstep);
 380         }
 381 }
 382
 383 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
 384 {
 385         __dec_zone_state(page_zone(page), item);
 386 }
 387 EXPORT_SYMBOL(__dec_zone_page_state);
 388
 389 void __dec_node_page_state(struct page *page, enum node_stat_item item)
 390 {
 391         __dec_node_state(page_pgdat(page), item);
 392 }
 393 EXPORT_SYMBOL(__dec_node_page_state);
 394
 395 #ifdef CONFIG_HAVE_CMPXCHG_LOCAL
 396 /*
 397  * If we have cmpxchg_local support then we do not need to incur the overhead
 398  * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
 399  *
 400  * mod_state() modifies the zone counter state through atomic per cpu
 401  * operations.
 402  *
 403  * Overstep mode specifies how overstep should handled:
 404  *     0       No overstepping
 405  *     1       Overstepping half of threshold
 406  *     -1      Overstepping minus half of threshold
 407 */
 408 static inline void mod_zone_state(struct zone *zone,
 409        enum zone_stat_item item, long delta, int overstep_mode)
 410 {
 411         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 412         s8 __percpu *p = pcp->vm_stat_diff + item;
 413         long o, n, t, z;
 414
 415         do {
 416                 z = 0;  /* overflow to zone counters */
 417
 418                 /*
 419                  * The fetching of the stat_threshold is racy. We may apply
 420                  * a counter threshold to the wrong the cpu if we get
 421                  * rescheduled while executing here. However, the next
 422                  * counter update will apply the threshold again and
 423                  * therefore bring the counter under the threshold again.
 424                  *
 425                  * Most of the time the thresholds are the same anyways
 426                  * for all cpus in a zone.
 427                  */
 428                 t = this_cpu_read(pcp->stat_threshold);
 429
 430                 o = this_cpu_read(*p);
 431                 n = delta + o;
 432
 433                 if (n > t || n < -t) {
 434                         int os = overstep_mode * (t >> 1) ;
 435
 436                         /* Overflow must be added to zone counters */
 437                         z = n + os;
 438                         n = -os;
 439                 }
 440         } while (this_cpu_cmpxchg(*p, o, n) != o);
 441
 442         if (z)
 443                 zone_page_state_add(z, zone, item);
 444 }
 445
 446 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 447                          long delta)
 448 {
 449         mod_zone_state(zone, item, delta, 0);
 450 }
 451 EXPORT_SYMBOL(mod_zone_page_state);
 452
 453 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 454 {
 455         mod_zone_state(page_zone(page), item, 1, 1);
 456 }
 457 EXPORT_SYMBOL(inc_zone_page_state);
 458
 459 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 460 {
 461         mod_zone_state(page_zone(page), item, -1, -1);
 462 }
 463 EXPORT_SYMBOL(dec_zone_page_state);
 464
 465 static inline void mod_node_state(struct pglist_data *pgdat,
 466        enum node_stat_item item, int delta, int overstep_mode)
 467 {
 468         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 469         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 470         long o, n, t, z;
 471
 472         do {
 473                 z = 0;  /* overflow to node counters */
 474
 475                 /*
 476                  * The fetching of the stat_threshold is racy. We may apply
 477                  * a counter threshold to the wrong the cpu if we get
 478                  * rescheduled while executing here. However, the next
 479                  * counter update will apply the threshold again and
 480                  * therefore bring the counter under the threshold again.
 481                  *
 482                  * Most of the time the thresholds are the same anyways
 483                  * for all cpus in a node.
 484                  */
 485                 t = this_cpu_read(pcp->stat_threshold);
 486
 487                 o = this_cpu_read(*p);
 488                 n = delta + o;
 489
 490                 if (n > t || n < -t) {
 491                         int os = overstep_mode * (t >> 1) ;
 492
 493                         /* Overflow must be added to node counters */
 494                         z = n + os;
 495                         n = -os;
 496                 }
 497         } while (this_cpu_cmpxchg(*p, o, n) != o);
 498
 499         if (z)
 500                 node_page_state_add(z, pgdat, item);
 501 }
 502
 503 void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
 504                                         long delta)
 505 {
 506         mod_node_state(pgdat, item, delta, 0);
 507 }
 508 EXPORT_SYMBOL(mod_node_page_state);
 509
 510 void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 511 {
 512         mod_node_state(pgdat, item, 1, 1);
 513 }
 514
 515 void inc_node_page_state(struct page *page, enum node_stat_item item)
 516 {
 517         mod_node_state(page_pgdat(page), item, 1, 1);
 518 }
 519 EXPORT_SYMBOL(inc_node_page_state);
 520
 521 void dec_node_page_state(struct page *page, enum node_stat_item item)
 522 {
 523         mod_node_state(page_pgdat(page), item, -1, -1);
 524 }
 525 EXPORT_SYMBOL(dec_node_page_state);
 526 #else
 527 /*
 528  * Use interrupt disable to serialize counter updates
 529  */
 530 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 531                          long delta)
 532 {
 533         unsigned long flags;
 534
 535         local_irq_save(flags);
 536         __mod_zone_page_state(zone, item, delta);
 537         local_irq_restore(flags);
 538 }
 539 EXPORT_SYMBOL(mod_zone_page_state);
 540
 541 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 542 {
 543         unsigned long flags;
 544         struct zone *zone;
 545
 546         zone = page_zone(page);
 547         local_irq_save(flags);
 548         __inc_zone_state(zone, item);
 549         local_irq_restore(flags);
 550 }
 551 EXPORT_SYMBOL(inc_zone_page_state);
 552
 553 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 554 {
 555         unsigned long flags;
 556
 557         local_irq_save(flags);
 558         __dec_zone_page_state(page, item);
 559         local_irq_restore(flags);
 560 }
 561 EXPORT_SYMBOL(dec_zone_page_state);
 562
 563 void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 564 {
 565         unsigned long flags;
 566
 567         local_irq_save(flags);
 568         __inc_node_state(pgdat, item);
 569         local_irq_restore(flags);
 570 }
 571 EXPORT_SYMBOL(inc_node_state);
 572
 573 void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
 574                                         long delta)
 575 {
 576         unsigned long flags;
 577
 578         local_irq_save(flags);
 579         __mod_node_page_state(pgdat, item, delta);
 580         local_irq_restore(flags);
 581 }
 582 EXPORT_SYMBOL(mod_node_page_state);
 583
 584 void inc_node_page_state(struct page *page, enum node_stat_item item)
 585 {
 586         unsigned long flags;
 587         struct pglist_data *pgdat;
 588
 589         pgdat = page_pgdat(page);
 590         local_irq_save(flags);
 591         __inc_node_state(pgdat, item);
 592         local_irq_restore(flags);
 593 }
 594 EXPORT_SYMBOL(inc_node_page_state);
 595
 596 void dec_node_page_state(struct page *page, enum node_stat_item item)
 597 {
 598         unsigned long flags;
 599
 600         local_irq_save(flags);
 601         __dec_node_page_state(page, item);
 602         local_irq_restore(flags);
 603 }
 604 EXPORT_SYMBOL(dec_node_page_state);
 605 #endif
 606
 607 /*
 608  * Fold a differential into the global counters.
 609  * Returns the number of counters updated.
 610  */
 611 #ifdef CONFIG_NUMA
 612 static int fold_diff(int *zone_diff, int *numa_diff, int *node_diff)
 613 {
 614         int i;
 615         int changes = 0;
 616
 617         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 618                 if (zone_diff[i]) {
 619                         atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
 620                         changes++;
 621         }
 622
 623         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
 624                 if (numa_diff[i]) {
 625                         atomic_long_add(numa_diff[i], &vm_numa_stat[i]);
 626                         changes++;
 627         }
 628
 629         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 630                 if (node_diff[i]) {
 631                         atomic_long_add(node_diff[i], &vm_node_stat[i]);
 632                         changes++;
 633         }
 634         return changes;
 635 }
 636 #else
 637 static int fold_diff(int *zone_diff, int *node_diff)
 638 {
 639         int i;
 640         int changes = 0;
 641
 642         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 643                 if (zone_diff[i]) {
 644                         atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
 645                         changes++;
 646         }
 647
 648         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 649                 if (node_diff[i]) {
 650                         atomic_long_add(node_diff[i], &vm_node_stat[i]);
 651                         changes++;
 652         }
 653         return changes;
 654 }
 655 #endif /* CONFIG_NUMA */
 656
 657 /*
 658  * Update the zone counters for the current cpu.
 659  *
 660  * Note that refresh_cpu_vm_stats strives to only access
 661  * node local memory. The per cpu pagesets on remote zones are placed
 662  * in the memory local to the processor using that pageset. So the
 663  * loop over all zones will access a series of cachelines local to
 664  * the processor.
 665  *
 666  * The call to zone_page_state_add updates the cachelines with the
 667  * statistics in the remote zone struct as well as the global cachelines
 668  * with the global counters. These could cause remote node cache line
 669  * bouncing and will have to be only done when necessary.
 670  *
 671  * The function returns the number of global counters updated.
 672  */
 673 static int refresh_cpu_vm_stats(bool do_pagesets)
 674 {
 675         struct pglist_data *pgdat;
 676         struct zone *zone;
 677         int i;
 678         int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 679 #ifdef CONFIG_NUMA
 680         int global_numa_diff[NR_VM_NUMA_STAT_ITEMS] = { 0, };
 681 #endif
 682         int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
 683         int changes = 0;
 684
 685         for_each_populated_zone(zone) {
 686                 struct per_cpu_pageset __percpu *p = zone->pageset;
 687
 688                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
 689                         int v;
 690
 691                         v = this_cpu_xchg(p->vm_stat_diff[i], 0);
 692                         if (v) {
 693
 694                                 atomic_long_add(v, &zone->vm_stat[i]);
 695                                 global_zone_diff[i] += v;
 696 #ifdef CONFIG_NUMA
 697                                 /* 3 seconds idle till flush */
 698                                 __this_cpu_write(p->expire, 3);
 699 #endif
 700                         }
 701                 }
 702 #ifdef CONFIG_NUMA
 703                 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) {
 704                         int v;
 705
 706                         v = this_cpu_xchg(p->vm_numa_stat_diff[i], 0);
 707                         if (v) {
 708
 709                                 atomic_long_add(v, &zone->vm_numa_stat[i]);
 710                                 global_numa_diff[i] += v;
 711                                 __this_cpu_write(p->expire, 3);
 712                         }
 713                 }
 714
 715                 if (do_pagesets) {
 716                         cond_resched();
 717                         /*
 718                          * Deal with draining the remote pageset of this
 719                          * processor
 720                          *
 721                          * Check if there are pages remaining in this pageset
 722                          * if not then there is nothing to expire.
 723                          */
 724                         if (!__this_cpu_read(p->expire) ||
 725                                !__this_cpu_read(p->pcp.count))
 726                                 continue;
 727
 728                         /*
 729                          * We never drain zones local to this processor.
 730                          */
 731                         if (zone_to_nid(zone) == numa_node_id()) {
 732                                 __this_cpu_write(p->expire, 0);
 733                                 continue;
 734                         }
 735
 736                         if (__this_cpu_dec_return(p->expire))
 737                                 continue;
 738
 739                         if (__this_cpu_read(p->pcp.count)) {
 740                                 drain_zone_pages(zone, this_cpu_ptr(&p->pcp));
 741                                 changes++;
 742                         }
 743                 }
 744 #endif
 745         }
 746
 747         for_each_online_pgdat(pgdat) {
 748                 struct per_cpu_nodestat __percpu *p = pgdat->per_cpu_nodestats;
 749
 750                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
 751                         int v;
 752
 753                         v = this_cpu_xchg(p->vm_node_stat_diff[i], 0);
 754                         if (v) {
 755                                 atomic_long_add(v, &pgdat->vm_stat[i]);
 756                                 global_node_diff[i] += v;
 757                         }
 758                 }
 759         }
 760
 761 #ifdef CONFIG_NUMA
 762         changes += fold_diff(global_zone_diff, global_numa_diff,
 763                              global_node_diff);
 764 #else
 765         changes += fold_diff(global_zone_diff, global_node_diff);
 766 #endif
 767         return changes;
 768 }
 769
 770 /*
 771  * Fold the data for an offline cpu into the global array.
 772  * There cannot be any access by the offline cpu and therefore
 773  * synchronization is simplified.
 774  */
 775 void cpu_vm_stats_fold(int cpu)
 776 {
 777         struct pglist_data *pgdat;
 778         struct zone *zone;
 779         int i;
 780         int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 781 #ifdef CONFIG_NUMA
 782         int global_numa_diff[NR_VM_NUMA_STAT_ITEMS] = { 0, };
 783 #endif
 784         int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
 785
 786         for_each_populated_zone(zone) {
 787                 struct per_cpu_pageset *p;
 788
 789                 p = per_cpu_ptr(zone->pageset, cpu);
 790
 791                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 792                         if (p->vm_stat_diff[i]) {
 793                                 int v;
 794
 795                                 v = p->vm_stat_diff[i];
 796                                 p->vm_stat_diff[i] = 0;
 797                                 atomic_long_add(v, &zone->vm_stat[i]);
 798                                 global_zone_diff[i] += v;
 799                         }
 800
 801 #ifdef CONFIG_NUMA
 802                 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
 803                         if (p->vm_numa_stat_diff[i]) {
 804                                 int v;
 805
 806                                 v = p->vm_numa_stat_diff[i];
 807                                 p->vm_numa_stat_diff[i] = 0;
 808                                 atomic_long_add(v, &zone->vm_numa_stat[i]);
 809                                 global_numa_diff[i] += v;
 810                         }
 811 #endif
 812         }
 813
 814         for_each_online_pgdat(pgdat) {
 815                 struct per_cpu_nodestat *p;
 816
 817                 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
 818
 819                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 820                         if (p->vm_node_stat_diff[i]) {
 821                                 int v;
 822
 823                                 v = p->vm_node_stat_diff[i];
 824                                 p->vm_node_stat_diff[i] = 0;
 825                                 atomic_long_add(v, &pgdat->vm_stat[i]);
 826                                 global_node_diff[i] += v;
 827                         }
 828         }
 829
 830 #ifdef CONFIG_NUMA
 831         fold_diff(global_zone_diff, global_numa_diff, global_node_diff);
 832 #else
 833         fold_diff(global_zone_diff, global_node_diff);
 834 #endif
 835 }
 836
 837 /*
 838  * this is only called if !populated_zone(zone), which implies no other users of
 839  * pset->vm_stat_diff[] exsist.
 840  */
 841 void drain_zonestat(struct zone *zone, struct per_cpu_pageset *pset)
 842 {
 843         int i;
 844
 845         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 846                 if (pset->vm_stat_diff[i]) {
 847                         int v = pset->vm_stat_diff[i];
 848                         pset->vm_stat_diff[i] = 0;
 849                         atomic_long_add(v, &zone->vm_stat[i]);
 850                         atomic_long_add(v, &vm_zone_stat[i]);
 851                 }
 852
 853 #ifdef CONFIG_NUMA
 854         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
 855                 if (pset->vm_numa_stat_diff[i]) {
 856                         int v = pset->vm_numa_stat_diff[i];
 857
 858                         pset->vm_numa_stat_diff[i] = 0;
 859                         atomic_long_add(v, &zone->vm_numa_stat[i]);
 860                         atomic_long_add(v, &vm_numa_stat[i]);
 861                 }
 862 #endif
 863 }
 864 #endif
 865
 866 #ifdef CONFIG_NUMA
 867 void __inc_numa_state(struct zone *zone,
 868                                  enum numa_stat_item item)
 869 {
 870         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 871         u16 __percpu *p = pcp->vm_numa_stat_diff + item;
 872         u16 v;
 873
 874         v = __this_cpu_inc_return(*p);
 875
 876         if (unlikely(v > NUMA_STATS_THRESHOLD)) {
 877                 zone_numa_state_add(v, zone, item);
 878                 __this_cpu_write(*p, 0);
 879         }
 880 }
 881
 882 /*
 883  * Determine the per node value of a stat item. This function
 884  * is called frequently in a NUMA machine, so try to be as
 885  * frugal as possible.
 886  */
 887 unsigned long sum_zone_node_page_state(int node,
 888                                  enum zone_stat_item item)
 889 {
 890         struct zone *zones = NODE_DATA(node)->node_zones;
 891         int i;
 892         unsigned long count = 0;
 893
 894         for (i = 0; i < MAX_NR_ZONES; i++)
 895                 count += zone_page_state(zones + i, item);
 896
 897         return count;
 898 }
 899
 900 /*
 901  * Determine the per node value of a numa stat item. To avoid deviation,
 902  * the per cpu stat number in vm_numa_stat_diff[] is also included.
 903  */
 904 unsigned long sum_zone_numa_state(int node,
 905                                  enum numa_stat_item item)
 906 {
 907         struct zone *zones = NODE_DATA(node)->node_zones;
 908         int i;
 909         unsigned long count = 0;
 910
 911         for (i = 0; i < MAX_NR_ZONES; i++)
 912                 count += zone_numa_state_snapshot(zones + i, item);
 913
 914         return count;
 915 }
 916
 917 /*
 918  * Determine the per node value of a stat item.
 919  */
 920 unsigned long node_page_state(struct pglist_data *pgdat,
 921                                 enum node_stat_item item)
 922 {
 923         long x = atomic_long_read(&pgdat->vm_stat[item]);
 924 #ifdef CONFIG_SMP
 925         if (x < 0)
 926                 x = 0;
 927 #endif
 928         return x;
 929 }
 930 #endif
 931
 932 #ifdef CONFIG_COMPACTION
 933
 934 struct contig_page_info {
 935         unsigned long free_pages;
 936         unsigned long free_blocks_total;
 937         unsigned long free_blocks_suitable;
 938 };
 939
 940 /*
 941  * Calculate the number of free pages in a zone, how many contiguous
 942  * pages are free and how many are large enough to satisfy an allocation of
 943  * the target size. Note that this function makes no attempt to estimate
 944  * how many suitable free blocks there *might* be if MOVABLE pages were
 945  * migrated. Calculating that is possible, but expensive and can be
 946  * figured out from userspace
 947  */
 948 static void fill_contig_page_info(struct zone *zone,
 949                                 unsigned int suitable_order,
 950                                 struct contig_page_info *info)
 951 {
 952         unsigned int order;
 953
 954         info->free_pages = 0;
 955         info->free_blocks_total = 0;
 956         info->free_blocks_suitable = 0;
 957
 958         for (order = 0; order < MAX_ORDER; order++) {
 959                 unsigned long blocks;
 960
 961                 /* Count number of free blocks */
 962                 blocks = zone->free_area[order].nr_free;
 963                 info->free_blocks_total += blocks;
 964
 965                 /* Count free base pages */
 966                 info->free_pages += blocks << order;
 967
 968                 /* Count the suitable free blocks */
 969                 if (order >= suitable_order)
 970                         info->free_blocks_suitable += blocks <<
 971                                                 (order - suitable_order);
 972         }
 973 }
 974
 975 /*
 976  * A fragmentation index only makes sense if an allocation of a requested
 977  * size would fail. If that is true, the fragmentation index indicates
 978  * whether external fragmentation or a lack of memory was the problem.
 979  * The value can be used to determine if page reclaim or compaction
 980  * should be used
 981  */
 982 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
 983 {
 984         unsigned long requested = 1UL << order;
 985
 986         if (WARN_ON_ONCE(order >= MAX_ORDER))
 987                 return 0;
 988
 989         if (!info->free_blocks_total)
 990                 return 0;
 991
 992         /* Fragmentation index only makes sense when a request would fail */
 993         if (info->free_blocks_suitable)
 994                 return -1000;
 995
 996         /*
 997          * Index is between 0 and 1 so return within 3 decimal places
 998          *
 999          * 0 => allocation would fail due to lack of memory
1000          * 1 => allocation would fail due to fragmentation
1001          */
1002         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
1003 }
1004
1005 /* Same as __fragmentation index but allocs contig_page_info on stack */
1006 int fragmentation_index(struct zone *zone, unsigned int order)
1007 {
1008         struct contig_page_info info;
1009
1010         fill_contig_page_info(zone, order, &info);
1011         return __fragmentation_index(order, &info);
1012 }
1013 #endif
1014
1015 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || defined(CONFIG_NUMA)
1016 #ifdef CONFIG_ZONE_DMA
1017 #define TEXT_FOR_DMA(xx) xx "_dma",
1018 #else
1019 #define TEXT_FOR_DMA(xx)
1020 #endif
1021
1022 #ifdef CONFIG_ZONE_DMA32
1023 #define TEXT_FOR_DMA32(xx) xx "_dma32",
1024 #else
1025 #define TEXT_FOR_DMA32(xx)
1026 #endif
1027
1028 #ifdef CONFIG_HIGHMEM
1029 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
1030 #else
1031 #define TEXT_FOR_HIGHMEM(xx)
1032 #endif
1033
1034 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1035                                         TEXT_FOR_HIGHMEM(xx) xx "_movable",
1036
1037 const char * const vmstat_text[] = {
1038         /* enum zone_stat_item countes */
1039         "nr_free_pages",
1040         "nr_zone_inactive_anon",
1041         "nr_zone_active_anon",
1042         "nr_zone_inactive_file",
1043         "nr_zone_active_file",
1044         "nr_zone_unevictable",
1045         "nr_zone_write_pending",
1046         "nr_mlock",
1047         "nr_page_table_pages",
1048         "nr_kernel_stack",
1049         "nr_bounce",
1050 #if IS_ENABLED(CONFIG_ZSMALLOC)
1051         "nr_zspages",
1052 #endif
1053         "nr_free_cma",
1054
1055         /* enum numa_stat_item counters */
1056 #ifdef CONFIG_NUMA
1057         "numa_hit",
1058         "numa_miss",
1059         "numa_foreign",
1060         "numa_interleave",
1061         "numa_local",
1062         "numa_other",
1063 #endif
1064
1065         /* Node-based counters */
1066         "nr_inactive_anon",
1067         "nr_active_anon",
1068         "nr_inactive_file",
1069         "nr_active_file",
1070         "nr_unevictable",
1071         "nr_slab_reclaimable",
1072         "nr_slab_unreclaimable",
1073         "nr_isolated_anon",
1074         "nr_isolated_file",
1075         "workingset_refault",
1076         "workingset_activate",
1077         "workingset_nodereclaim",
1078         "nr_anon_pages",
1079         "nr_mapped",
1080         "nr_file_pages",
1081         "nr_dirty",
1082         "nr_writeback",
1083         "nr_writeback_temp",
1084         "nr_shmem",
1085         "nr_shmem_hugepages",
1086         "nr_shmem_pmdmapped",
1087         "nr_anon_transparent_hugepages",
1088         "nr_unstable",
1089         "nr_vmscan_write",
1090         "nr_vmscan_immediate_reclaim",
1091         "nr_dirtied",
1092         "nr_written",
1093
1094         /* enum writeback_stat_item counters */
1095         "nr_dirty_threshold",
1096         "nr_dirty_background_threshold",
1097
1098 #ifdef CONFIG_VM_EVENT_COUNTERS
1099         /* enum vm_event_item counters */
1100         "pgpgin",
1101         "pgpgout",
1102         "pswpin",
1103         "pswpout",
1104
1105         TEXTS_FOR_ZONES("pgalloc")
1106         TEXTS_FOR_ZONES("allocstall")
1107         TEXTS_FOR_ZONES("pgskip")
1108
1109         "pgfree",
1110         "pgactivate",
1111         "pgdeactivate",
1112         "pglazyfree",
1113
1114         "pgfault",
1115         "pgmajfault",
1116         "pglazyfreed",
1117
1118         "pgrefill",
1119         "pgsteal_kswapd",
1120         "pgsteal_direct",
1121         "pgscan_kswapd",
1122         "pgscan_direct",
1123         "pgscan_direct_throttle",
1124
1125 #ifdef CONFIG_NUMA
1126         "zone_reclaim_failed",
1127 #endif
1128         "pginodesteal",
1129         "slabs_scanned",
1130         "kswapd_inodesteal",
1131         "kswapd_low_wmark_hit_quickly",
1132         "kswapd_high_wmark_hit_quickly",
1133         "pageoutrun",
1134
1135         "pgrotated",
1136
1137         "drop_pagecache",
1138         "drop_slab",
1139         "oom_kill",
1140
1141 #ifdef CONFIG_NUMA_BALANCING
1142         "numa_pte_updates",
1143         "numa_huge_pte_updates",
1144         "numa_hint_faults",
1145         "numa_hint_faults_local",
1146         "numa_pages_migrated",
1147 #endif
1148 #ifdef CONFIG_MIGRATION
1149         "pgmigrate_success",
1150         "pgmigrate_fail",
1151 #endif
1152 #ifdef CONFIG_COMPACTION
1153         "compact_migrate_scanned",
1154         "compact_free_scanned",
1155         "compact_isolated",
1156         "compact_stall",
1157         "compact_fail",
1158         "compact_success",
1159         "compact_daemon_wake",
1160         "compact_daemon_migrate_scanned",
1161         "compact_daemon_free_scanned",
1162 #endif
1163
1164 #ifdef CONFIG_HUGETLB_PAGE
1165         "htlb_buddy_alloc_success",
1166         "htlb_buddy_alloc_fail",
1167 #endif
1168         "unevictable_pgs_culled",
1169         "unevictable_pgs_scanned",
1170         "unevictable_pgs_rescued",
1171         "unevictable_pgs_mlocked",
1172         "unevictable_pgs_munlocked",
1173         "unevictable_pgs_cleared",
1174         "unevictable_pgs_stranded",
1175
1176 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1177         "thp_fault_alloc",
1178         "thp_fault_fallback",
1179         "thp_collapse_alloc",
1180         "thp_collapse_alloc_failed",
1181         "thp_file_alloc",
1182         "thp_file_mapped",
1183         "thp_split_page",
1184         "thp_split_page_failed",
1185         "thp_deferred_split_page",
1186         "thp_split_pmd",
1187 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1188         "thp_split_pud",
1189 #endif
1190         "thp_zero_page_alloc",
1191         "thp_zero_page_alloc_failed",
1192         "thp_swpout",
1193         "thp_swpout_fallback",
1194 #endif
1195 #ifdef CONFIG_MEMORY_BALLOON
1196         "balloon_inflate",
1197         "balloon_deflate",
1198 #ifdef CONFIG_BALLOON_COMPACTION
1199         "balloon_migrate",
1200 #endif
1201 #endif /* CONFIG_MEMORY_BALLOON */
1202 #ifdef CONFIG_DEBUG_TLBFLUSH
1203 #ifdef CONFIG_SMP
1204         "nr_tlb_remote_flush",
1205         "nr_tlb_remote_flush_received",
1206 #endif /* CONFIG_SMP */
1207         "nr_tlb_local_flush_all",
1208         "nr_tlb_local_flush_one",
1209 #endif /* CONFIG_DEBUG_TLBFLUSH */
1210
1211 #ifdef CONFIG_DEBUG_VM_VMACACHE
1212         "vmacache_find_calls",
1213         "vmacache_find_hits",
1214         "vmacache_full_flushes",
1215 #endif
1216 #ifdef CONFIG_SWAP
1217         "swap_ra",
1218         "swap_ra_hit",
1219 #endif
1220 #endif /* CONFIG_VM_EVENTS_COUNTERS */
1221 };
1222 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
1223
1224 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1225      defined(CONFIG_PROC_FS)
1226 static void *frag_start(struct seq_file *m, loff_t *pos)
1227 {
1228         pg_data_t *pgdat;
1229         loff_t node = *pos;
1230
1231         for (pgdat = first_online_pgdat();
1232              pgdat && node;
1233              pgdat = next_online_pgdat(pgdat))
1234                 --node;
1235
1236         return pgdat;
1237 }
1238
1239 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1240 {
1241         pg_data_t *pgdat = (pg_data_t *)arg;
1242
1243         (*pos)++;
1244         return next_online_pgdat(pgdat);
1245 }
1246
1247 static void frag_stop(struct seq_file *m, void *arg)
1248 {
1249 }
1250
1251 /*
1252  * Walk zones in a node and print using a callback.
1253  * If @assert_populated is true, only use callback for zones that are populated.
1254  */
1255 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1256                 bool assert_populated, bool nolock,
1257                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1258 {
1259         struct zone *zone;
1260         struct zone *node_zones = pgdat->node_zones;
1261         unsigned long flags;
1262
1263         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1264                 if (assert_populated && !populated_zone(zone))
1265                         continue;
1266
1267                 if (!nolock)
1268                         spin_lock_irqsave(&zone->lock, flags);
1269                 print(m, pgdat, zone);
1270                 if (!nolock)
1271                         spin_unlock_irqrestore(&zone->lock, flags);
1272         }
1273 }
1274 #endif
1275
1276 #ifdef CONFIG_PROC_FS
1277 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1278                                                 struct zone *zone)
1279 {
1280         int order;
1281
1282         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1283         for (order = 0; order < MAX_ORDER; ++order)
1284                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
1285         seq_putc(m, '\n');
1286 }
1287
1288 /*
1289  * This walks the free areas for each zone.
1290  */
1291 static int frag_show(struct seq_file *m, void *arg)
1292 {
1293         pg_data_t *pgdat = (pg_data_t *)arg;
1294         walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1295         return 0;
1296 }
1297
1298 static void pagetypeinfo_showfree_print(struct seq_file *m,
1299                                         pg_data_t *pgdat, struct zone *zone)
1300 {
1301         int order, mtype;
1302
1303         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1304                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
1305                                         pgdat->node_id,
1306                                         zone->name,
1307                                         migratetype_names[mtype]);
1308                 for (order = 0; order < MAX_ORDER; ++order) {
1309                         unsigned long freecount = 0;
1310                         struct free_area *area;
1311                         struct list_head *curr;
1312
1313                         area = &(zone->free_area[order]);
1314
1315                         list_for_each(curr, &area->free_list[mtype])
1316                                 freecount++;
1317                         seq_printf(m, "%6lu ", freecount);
1318                 }
1319                 seq_putc(m, '\n');
1320         }
1321 }
1322
1323 /* Print out the free pages at each order for each migatetype */
1324 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
1325 {
1326         int order;
1327         pg_data_t *pgdat = (pg_data_t *)arg;
1328
1329         /* Print header */
1330         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1331         for (order = 0; order < MAX_ORDER; ++order)
1332                 seq_printf(m, "%6d ", order);
1333         seq_putc(m, '\n');
1334
1335         walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1336
1337         return 0;
1338 }
1339
1340 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1341                                         pg_data_t *pgdat, struct zone *zone)
1342 {
1343         int mtype;
1344         unsigned long pfn;
1345         unsigned long start_pfn = zone->zone_start_pfn;
1346         unsigned long end_pfn = zone_end_pfn(zone);
1347         unsigned long count[MIGRATE_TYPES] = { 0, };
1348
1349         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1350                 struct page *page;
1351
1352                 page = pfn_to_online_page(pfn);
1353                 if (!page)
1354                         continue;
1355
1356                 /* Watch for unexpected holes punched in the memmap */
1357                 if (!memmap_valid_within(pfn, page, zone))
1358                         continue;
1359
1360                 if (page_zone(page) != zone)
1361                         continue;
1362
1363                 mtype = get_pageblock_migratetype(page);
1364
1365                 if (mtype < MIGRATE_TYPES)
1366                         count[mtype]++;
1367         }
1368
1369         /* Print counts */
1370         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1371         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1372                 seq_printf(m, "%12lu ", count[mtype]);
1373         seq_putc(m, '\n');
1374 }
1375
1376 /* Print out the number of pageblocks for each migratetype */
1377 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1378 {
1379         int mtype;
1380         pg_data_t *pgdat = (pg_data_t *)arg;
1381
1382         seq_printf(m, "\n%-23s", "Number of blocks type ");
1383         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1384                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1385         seq_putc(m, '\n');
1386         walk_zones_in_node(m, pgdat, true, false,
1387                 pagetypeinfo_showblockcount_print);
1388
1389         return 0;
1390 }
1391
1392 /*
1393  * Print out the number of pageblocks for each migratetype that contain pages
1394  * of other types. This gives an indication of how well fallbacks are being
1395  * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1396  * to determine what is going on
1397  */
1398 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1399 {
1400 #ifdef CONFIG_PAGE_OWNER
1401         int mtype;
1402
1403         if (!static_branch_unlikely(&page_owner_inited))
1404                 return;
1405
1406         drain_all_pages(NULL);
1407
1408         seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1409         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1410                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1411         seq_putc(m, '\n');
1412
1413         walk_zones_in_node(m, pgdat, true, true,
1414                 pagetypeinfo_showmixedcount_print);
1415 #endif /* CONFIG_PAGE_OWNER */
1416 }
1417
1418 /*
1419  * This prints out statistics in relation to grouping pages by mobility.
1420  * It is expensive to collect so do not constantly read the file.
1421  */
1422 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1423 {
1424         pg_data_t *pgdat = (pg_data_t *)arg;
1425
1426         /* check memoryless node */
1427         if (!node_state(pgdat->node_id, N_MEMORY))
1428                 return 0;
1429
1430         seq_printf(m, "Page block order: %d\n", pageblock_order);
1431         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
1432         seq_putc(m, '\n');
1433         pagetypeinfo_showfree(m, pgdat);
1434         pagetypeinfo_showblockcount(m, pgdat);
1435         pagetypeinfo_showmixedcount(m, pgdat);
1436
1437         return 0;
1438 }
1439
1440 static const struct seq_operations fragmentation_op = {
1441         .start  = frag_start,
1442         .next   = frag_next,
1443         .stop   = frag_stop,
1444         .show   = frag_show,
1445 };
1446
1447 static int fragmentation_open(struct inode *inode, struct file *file)
1448 {
1449         return seq_open(file, &fragmentation_op);
1450 }
1451
1452 static const struct file_operations buddyinfo_file_operations = {
1453         .open           = fragmentation_open,
1454         .read           = seq_read,
1455         .llseek         = seq_lseek,
1456         .release        = seq_release,
1457 };
1458
1459 static const struct seq_operations pagetypeinfo_op = {
1460         .start  = frag_start,
1461         .next   = frag_next,
1462         .stop   = frag_stop,
1463         .show   = pagetypeinfo_show,
1464 };
1465
1466 static int pagetypeinfo_open(struct inode *inode, struct file *file)
1467 {
1468         return seq_open(file, &pagetypeinfo_op);
1469 }
1470
1471 static const struct file_operations pagetypeinfo_file_operations = {
1472         .open           = pagetypeinfo_open,
1473         .read           = seq_read,
1474         .llseek         = seq_lseek,
1475         .release        = seq_release,
1476 };
1477
1478 static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1479 {
1480         int zid;
1481
1482         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1483                 struct zone *compare = &pgdat->node_zones[zid];
1484
1485                 if (populated_zone(compare))
1486                         return zone == compare;
1487         }
1488
1489         return false;
1490 }
1491
1492 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1493                                                         struct zone *zone)
1494 {
1495         int i;
1496         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1497         if (is_zone_first_populated(pgdat, zone)) {
1498                 seq_printf(m, "\n  per-node stats");
1499                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1500                         seq_printf(m, "\n      %-12s %lu",
1501                                 vmstat_text[i + NR_VM_ZONE_STAT_ITEMS +
1502                                 NR_VM_NUMA_STAT_ITEMS],
1503                                 node_page_state(pgdat, i));
1504                 }
1505         }
1506         seq_printf(m,
1507                    "\n  pages free     %lu"
1508                    "\n        min      %lu"
1509                    "\n        low      %lu"
1510                    "\n        high     %lu"
1511                    "\n        spanned  %lu"
1512                    "\n        present  %lu"
1513                    "\n        managed  %lu",
1514                    zone_page_state(zone, NR_FREE_PAGES),
1515                    min_wmark_pages(zone),
1516                    low_wmark_pages(zone),
1517                    high_wmark_pages(zone),
1518                    zone->spanned_pages,
1519                    zone->present_pages,
1520                    zone->managed_pages);
1521
1522         seq_printf(m,
1523                    "\n        protection: (%ld",
1524                    zone->lowmem_reserve[0]);
1525         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1526                 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1527         seq_putc(m, ')');
1528
1529         /* If unpopulated, no other information is useful */
1530         if (!populated_zone(zone)) {
1531                 seq_putc(m, '\n');
1532                 return;
1533         }
1534
1535         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1536                 seq_printf(m, "\n      %-12s %lu", vmstat_text[i],
1537                                 zone_page_state(zone, i));
1538
1539 #ifdef CONFIG_NUMA
1540         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1541                 seq_printf(m, "\n      %-12s %lu",
1542                                 vmstat_text[i + NR_VM_ZONE_STAT_ITEMS],
1543                                 zone_numa_state_snapshot(zone, i));
1544 #endif
1545
1546         seq_printf(m, "\n  pagesets");
1547         for_each_online_cpu(i) {
1548                 struct per_cpu_pageset *pageset;
1549
1550                 pageset = per_cpu_ptr(zone->pageset, i);
1551                 seq_printf(m,
1552                            "\n    cpu: %i"
1553                            "\n              count: %i"
1554                            "\n              high:  %i"
1555                            "\n              batch: %i",
1556                            i,
1557                            pageset->pcp.count,
1558                            pageset->pcp.high,
1559                            pageset->pcp.batch);
1560 #ifdef CONFIG_SMP
1561                 seq_printf(m, "\n  vm stats threshold: %d",
1562                                 pageset->stat_threshold);
1563 #endif
1564         }
1565         seq_printf(m,
1566                    "\n  node_unreclaimable:  %u"
1567                    "\n  start_pfn:           %lu",
1568                    pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1569                    zone->zone_start_pfn);
1570         seq_putc(m, '\n');
1571 }
1572
1573 /*
1574  * Output information about zones in @pgdat.  All zones are printed regardless
1575  * of whether they are populated or not: lowmem_reserve_ratio operates on the
1576  * set of all zones and userspace would not be aware of such zones if they are
1577  * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1578  */
1579 static int zoneinfo_show(struct seq_file *m, void *arg)
1580 {
1581         pg_data_t *pgdat = (pg_data_t *)arg;
1582         walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1583         return 0;
1584 }
1585
1586 static const struct seq_operations zoneinfo_op = {
1587         .start  = frag_start, /* iterate over all zones. The same as in
1588                                * fragmentation. */
1589         .next   = frag_next,
1590         .stop   = frag_stop,
1591         .show   = zoneinfo_show,
1592 };
1593
1594 static int zoneinfo_open(struct inode *inode, struct file *file)
1595 {
1596         return seq_open(file, &zoneinfo_op);
1597 }
1598
1599 static const struct file_operations zoneinfo_file_operations = {
1600         .open           = zoneinfo_open,
1601         .read           = seq_read,
1602         .llseek         = seq_lseek,
1603         .release        = seq_release,
1604 };
1605
1606 enum writeback_stat_item {
1607         NR_DIRTY_THRESHOLD,
1608         NR_DIRTY_BG_THRESHOLD,
1609         NR_VM_WRITEBACK_STAT_ITEMS,
1610 };
1611
1612 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1613 {
1614         unsigned long *v;
1615         int i, stat_items_size;
1616
1617         if (*pos >= ARRAY_SIZE(vmstat_text))
1618                 return NULL;
1619         stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1620                           NR_VM_NUMA_STAT_ITEMS * sizeof(unsigned long) +
1621                           NR_VM_NODE_STAT_ITEMS * sizeof(unsigned long) +
1622                           NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1623
1624 #ifdef CONFIG_VM_EVENT_COUNTERS
1625         stat_items_size += sizeof(struct vm_event_state);
1626 #endif
1627
1628         v = kmalloc(stat_items_size, GFP_KERNEL);
1629         m->private = v;
1630         if (!v)
1631                 return ERR_PTR(-ENOMEM);
1632         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1633                 v[i] = global_zone_page_state(i);
1634         v += NR_VM_ZONE_STAT_ITEMS;
1635
1636 #ifdef CONFIG_NUMA
1637         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1638                 v[i] = global_numa_state(i);
1639         v += NR_VM_NUMA_STAT_ITEMS;
1640 #endif
1641
1642         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
1643                 v[i] = global_node_page_state(i);
1644         v += NR_VM_NODE_STAT_ITEMS;
1645
1646         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1647                             v + NR_DIRTY_THRESHOLD);
1648         v += NR_VM_WRITEBACK_STAT_ITEMS;
1649
1650 #ifdef CONFIG_VM_EVENT_COUNTERS
1651         all_vm_events(v);
1652         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1653         v[PGPGOUT] /= 2;
1654 #endif
1655         return (unsigned long *)m->private + *pos;
1656 }
1657
1658 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1659 {
1660         (*pos)++;
1661         if (*pos >= ARRAY_SIZE(vmstat_text))
1662                 return NULL;
1663         return (unsigned long *)m->private + *pos;
1664 }
1665
1666 static int vmstat_show(struct seq_file *m, void *arg)
1667 {
1668         unsigned long *l = arg;
1669         unsigned long off = l - (unsigned long *)m->private;
1670
1671         seq_puts(m, vmstat_text[off]);
1672         seq_put_decimal_ull(m, " ", *l);
1673         seq_putc(m, '\n');
1674         return 0;
1675 }
1676
1677 static void vmstat_stop(struct seq_file *m, void *arg)
1678 {
1679         kfree(m->private);
1680         m->private = NULL;
1681 }
1682
1683 static const struct seq_operations vmstat_op = {
1684         .start  = vmstat_start,
1685         .next   = vmstat_next,
1686         .stop   = vmstat_stop,
1687         .show   = vmstat_show,
1688 };
1689
1690 static int vmstat_open(struct inode *inode, struct file *file)
1691 {
1692         return seq_open(file, &vmstat_op);
1693 }
1694
1695 static const struct file_operations vmstat_file_operations = {
1696         .open           = vmstat_open,
1697         .read           = seq_read,
1698         .llseek         = seq_lseek,
1699         .release        = seq_release,
1700 };
1701 #endif /* CONFIG_PROC_FS */
1702
1703 #ifdef CONFIG_SMP
1704 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1705 int sysctl_stat_interval __read_mostly = HZ;
1706
1707 #ifdef CONFIG_PROC_FS
1708 static void refresh_vm_stats(struct work_struct *work)
1709 {
1710         refresh_cpu_vm_stats(true);
1711 }
1712
1713 int vmstat_refresh(struct ctl_table *table, int write,
1714                    void __user *buffer, size_t *lenp, loff_t *ppos)
1715 {
1716         long val;
1717         int err;
1718         int i;
1719
1720         /*
1721          * The regular update, every sysctl_stat_interval, may come later
1722          * than expected: leaving a significant amount in per_cpu buckets.
1723          * This is particularly misleading when checking a quantity of HUGE
1724          * pages, immediately after running a test.  /proc/sys/vm/stat_refresh,
1725          * which can equally be echo'ed to or cat'ted from (by root),
1726          * can be used to update the stats just before reading them.
1727          *
1728          * Oh, and since global_zone_page_state() etc. are so careful to hide
1729          * transiently negative values, report an error here if any of
1730          * the stats is negative, so we know to go looking for imbalance.
1731          */
1732         err = schedule_on_each_cpu(refresh_vm_stats);
1733         if (err)
1734                 return err;
1735         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1736                 val = atomic_long_read(&vm_zone_stat[i]);
1737                 if (val < 0) {
1738                         pr_warn("%s: %s %ld\n",
1739                                 __func__, vmstat_text[i], val);
1740                         err = -EINVAL;
1741                 }
1742         }
1743 #ifdef CONFIG_NUMA
1744         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) {
1745                 val = atomic_long_read(&vm_numa_stat[i]);
1746                 if (val < 0) {
1747                         pr_warn("%s: %s %ld\n",
1748                                 __func__, vmstat_text[i + NR_VM_ZONE_STAT_ITEMS], val);
1749                         err = -EINVAL;
1750                 }
1751         }
1752 #endif
1753         if (err)
1754                 return err;
1755         if (write)
1756                 *ppos += *lenp;
1757         else
1758                 *lenp = 0;
1759         return 0;
1760 }
1761 #endif /* CONFIG_PROC_FS */
1762
1763 static void vmstat_update(struct work_struct *w)
1764 {
1765         if (refresh_cpu_vm_stats(true)) {
1766                 /*
1767                  * Counters were updated so we expect more updates
1768                  * to occur in the future. Keep on running the
1769                  * update worker thread.
1770                  */
1771                 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
1772                                 this_cpu_ptr(&vmstat_work),
1773                                 round_jiffies_relative(sysctl_stat_interval));
1774         }
1775 }
1776
1777 /*
1778  * Switch off vmstat processing and then fold all the remaining differentials
1779  * until the diffs stay at zero. The function is used by NOHZ and can only be
1780  * invoked when tick processing is not active.
1781  */
1782 /*
1783  * Check if the diffs for a certain cpu indicate that
1784  * an update is needed.
1785  */
1786 static bool need_update(int cpu)
1787 {
1788         struct zone *zone;
1789
1790         for_each_populated_zone(zone) {
1791                 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1792
1793                 BUILD_BUG_ON(sizeof(p->vm_stat_diff[0]) != 1);
1794 #ifdef CONFIG_NUMA
1795                 BUILD_BUG_ON(sizeof(p->vm_numa_stat_diff[0]) != 2);
1796 #endif
1797
1798                 /*
1799                  * The fast way of checking if there are any vmstat diffs.
1800                  * This works because the diffs are byte sized items.
1801                  */
1802                 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS))
1803                         return true;
1804 #ifdef CONFIG_NUMA
1805                 if (memchr_inv(p->vm_numa_stat_diff, 0, NR_VM_NUMA_STAT_ITEMS))
1806                         return true;
1807 #endif
1808         }
1809         return false;
1810 }
1811
1812 /*
1813  * Switch off vmstat processing and then fold all the remaining differentials
1814  * until the diffs stay at zero. The function is used by NOHZ and can only be
1815  * invoked when tick processing is not active.
1816  */
1817 void quiet_vmstat(void)
1818 {
1819         if (system_state != SYSTEM_RUNNING)
1820                 return;
1821
1822         if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
1823                 return;
1824
1825         if (!need_update(smp_processor_id()))
1826                 return;
1827
1828         /*
1829          * Just refresh counters and do not care about the pending delayed
1830          * vmstat_update. It doesn't fire that often to matter and canceling
1831          * it would be too expensive from this path.
1832          * vmstat_shepherd will take care about that for us.
1833          */
1834         refresh_cpu_vm_stats(false);
1835 }
1836
1837 /*
1838  * Shepherd worker thread that checks the
1839  * differentials of processors that have their worker
1840  * threads for vm statistics updates disabled because of
1841  * inactivity.
1842  */
1843 static void vmstat_shepherd(struct work_struct *w);
1844
1845 static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
1846
1847 static void vmstat_shepherd(struct work_struct *w)
1848 {
1849         int cpu;
1850
1851         get_online_cpus();
1852         /* Check processors whose vmstat worker threads have been disabled */
1853         for_each_online_cpu(cpu) {
1854                 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
1855
1856                 if (!delayed_work_pending(dw) && need_update(cpu))
1857                         queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
1858         }
1859         put_online_cpus();
1860
1861         schedule_delayed_work(&shepherd,
1862                 round_jiffies_relative(sysctl_stat_interval));
1863 }
1864
1865 static void __init start_shepherd_timer(void)
1866 {
1867         int cpu;
1868
1869         for_each_possible_cpu(cpu)
1870                 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
1871                         vmstat_update);
1872
1873         schedule_delayed_work(&shepherd,
1874                 round_jiffies_relative(sysctl_stat_interval));
1875 }
1876
1877 static void __init init_cpu_node_state(void)
1878 {
1879         int node;
1880
1881         for_each_online_node(node) {
1882                 if (cpumask_weight(cpumask_of_node(node)) > 0)
1883                         node_set_state(node, N_CPU);
1884         }
1885 }
1886
1887 static int vmstat_cpu_online(unsigned int cpu)
1888 {
1889         refresh_zone_stat_thresholds();
1890         node_set_state(cpu_to_node(cpu), N_CPU);
1891         return 0;
1892 }
1893
1894 static int vmstat_cpu_down_prep(unsigned int cpu)
1895 {
1896         cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1897         return 0;
1898 }
1899
1900 static int vmstat_cpu_dead(unsigned int cpu)
1901 {
1902         const struct cpumask *node_cpus;
1903         int node;
1904
1905         node = cpu_to_node(cpu);
1906
1907         refresh_zone_stat_thresholds();
1908         node_cpus = cpumask_of_node(node);
1909         if (cpumask_weight(node_cpus) > 0)
1910                 return 0;
1911
1912         node_clear_state(node, N_CPU);
1913         return 0;
1914 }
1915
1916 #endif
1917
1918 struct workqueue_struct *mm_percpu_wq;
1919
1920 void __init init_mm_internals(void)
1921 {
1922         int ret __maybe_unused;
1923
1924         mm_percpu_wq = alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM, 0);
1925
1926 #ifdef CONFIG_SMP
1927         ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
1928                                         NULL, vmstat_cpu_dead);
1929         if (ret < 0)
1930                 pr_err("vmstat: failed to register 'dead' hotplug state\n");
1931
1932         ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "mm/vmstat:online",
1933                                         vmstat_cpu_online,
1934                                         vmstat_cpu_down_prep);
1935         if (ret < 0)
1936                 pr_err("vmstat: failed to register 'online' hotplug state\n");
1937
1938         get_online_cpus();
1939         init_cpu_node_state();
1940         put_online_cpus();
1941
1942         start_shepherd_timer();
1943 #endif
1944 #ifdef CONFIG_PROC_FS
1945         proc_create("buddyinfo", 0444, NULL, &buddyinfo_file_operations);
1946         proc_create("pagetypeinfo", 0444, NULL, &pagetypeinfo_file_operations);
1947         proc_create("vmstat", 0444, NULL, &vmstat_file_operations);
1948         proc_create("zoneinfo", 0444, NULL, &zoneinfo_file_operations);
1949 #endif
1950 }
1951
1952 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1953
1954 /*
1955  * Return an index indicating how much of the available free memory is
1956  * unusable for an allocation of the requested size.
1957  */
1958 static int unusable_free_index(unsigned int order,
1959                                 struct contig_page_info *info)
1960 {
1961         /* No free memory is interpreted as all free memory is unusable */
1962         if (info->free_pages == 0)
1963                 return 1000;
1964
1965         /*
1966          * Index should be a value between 0 and 1. Return a value to 3
1967          * decimal places.
1968          *
1969          * 0 => no fragmentation
1970          * 1 => high fragmentation
1971          */
1972         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1973
1974 }
1975
1976 static void unusable_show_print(struct seq_file *m,
1977                                         pg_data_t *pgdat, struct zone *zone)
1978 {
1979         unsigned int order;
1980         int index;
1981         struct contig_page_info info;
1982
1983         seq_printf(m, "Node %d, zone %8s ",
1984                                 pgdat->node_id,
1985                                 zone->name);
1986         for (order = 0; order < MAX_ORDER; ++order) {
1987                 fill_contig_page_info(zone, order, &info);
1988                 index = unusable_free_index(order, &info);
1989                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1990         }
1991
1992         seq_putc(m, '\n');
1993 }
1994
1995 /*
1996  * Display unusable free space index
1997  *
1998  * The unusable free space index measures how much of the available free
1999  * memory cannot be used to satisfy an allocation of a given size and is a
2000  * value between 0 and 1. The higher the value, the more of free memory is
2001  * unusable and by implication, the worse the external fragmentation is. This
2002  * can be expressed as a percentage by multiplying by 100.
2003  */
2004 static int unusable_show(struct seq_file *m, void *arg)
2005 {
2006         pg_data_t *pgdat = (pg_data_t *)arg;
2007
2008         /* check memoryless node */
2009         if (!node_state(pgdat->node_id, N_MEMORY))
2010                 return 0;
2011
2012         walk_zones_in_node(m, pgdat, true, false, unusable_show_print);
2013
2014         return 0;
2015 }
2016
2017 static const struct seq_operations unusable_op = {
2018         .start  = frag_start,
2019         .next   = frag_next,
2020         .stop   = frag_stop,
2021         .show   = unusable_show,
2022 };
2023
2024 static int unusable_open(struct inode *inode, struct file *file)
2025 {
2026         return seq_open(file, &unusable_op);
2027 }
2028
2029 static const struct file_operations unusable_file_ops = {
2030         .open           = unusable_open,
2031         .read           = seq_read,
2032         .llseek         = seq_lseek,
2033         .release        = seq_release,
2034 };
2035
2036 static void extfrag_show_print(struct seq_file *m,
2037                                         pg_data_t *pgdat, struct zone *zone)
2038 {
2039         unsigned int order;
2040         int index;
2041
2042         /* Alloc on stack as interrupts are disabled for zone walk */
2043         struct contig_page_info info;
2044
2045         seq_printf(m, "Node %d, zone %8s ",
2046                                 pgdat->node_id,
2047                                 zone->name);
2048         for (order = 0; order < MAX_ORDER; ++order) {
2049                 fill_contig_page_info(zone, order, &info);
2050                 index = __fragmentation_index(order, &info);
2051                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2052         }
2053
2054         seq_putc(m, '\n');
2055 }
2056
2057 /*
2058  * Display fragmentation index for orders that allocations would fail for
2059  */
2060 static int extfrag_show(struct seq_file *m, void *arg)
2061 {
2062         pg_data_t *pgdat = (pg_data_t *)arg;
2063
2064         walk_zones_in_node(m, pgdat, true, false, extfrag_show_print);
2065
2066         return 0;
2067 }
2068
2069 static const struct seq_operations extfrag_op = {
2070         .start  = frag_start,
2071         .next   = frag_next,
2072         .stop   = frag_stop,
2073         .show   = extfrag_show,
2074 };
2075
2076 static int extfrag_open(struct inode *inode, struct file *file)
2077 {
2078         return seq_open(file, &extfrag_op);
2079 }
2080
2081 static const struct file_operations extfrag_file_ops = {
2082         .open           = extfrag_open,
2083         .read           = seq_read,
2084         .llseek         = seq_lseek,
2085         .release        = seq_release,
2086 };
2087
2088 static int __init extfrag_debug_init(void)
2089 {
2090         struct dentry *extfrag_debug_root;
2091
2092         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
2093         if (!extfrag_debug_root)
2094                 return -ENOMEM;
2095
2096         if (!debugfs_create_file("unusable_index", 0444,
2097                         extfrag_debug_root, NULL, &unusable_file_ops))
2098                 goto fail;
2099
2100         if (!debugfs_create_file("extfrag_index", 0444,
2101                         extfrag_debug_root, NULL, &extfrag_file_ops))
2102                 goto fail;
2103
2104         return 0;
2105 fail:
2106         debugfs_remove_recursive(extfrag_debug_root);
2107         return -ENOMEM;
2108 }
2109
2110 module_init(extfrag_debug_init);
2111 #endif