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