mm/vmstat.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  linux/mm/vmstat.c
   4  *
   5  *  Manages VM statistics
   6  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   7  *
   8  *  zoned VM statistics
   9  *  Copyright (C) 2006 Silicon Graphics, Inc.,
  10  *              Christoph Lameter <christoph@lameter.com>
  11  *  Copyright (C) 2008-2014 Christoph Lameter
  12  */
  13 #include <linux/fs.h>
  14 #include <linux/mm.h>
  15 #include <linux/err.h>
  16 #include <linux/module.h>
  17 #include <linux/slab.h>
  18 #include <linux/cpu.h>
  19 #include <linux/cpumask.h>
  20 #include <linux/vmstat.h>
  21 #include <linux/proc_fs.h>
  22 #include <linux/seq_file.h>
  23 #include <linux/debugfs.h>
  24 #include <linux/sched.h>
  25 #include <linux/math64.h>
  26 #include <linux/writeback.h>
  27 #include <linux/compaction.h>
  28 #include <linux/mm_inline.h>
  29 #include <linux/page_ext.h>
  30 #include <linux/page_owner.h>
  31
  32 #include "internal.h"
  33
  34 #define NUMA_STATS_THRESHOLD (U16_MAX - 2)
  35
  36 #ifdef CONFIG_NUMA
  37 int sysctl_vm_numa_stat = ENABLE_NUMA_STAT;
  38
  39 /* zero numa counters within a zone */
  40 static void zero_zone_numa_counters(struct zone *zone)
  41 {
  42         int item, cpu;
  43
  44         for (item = 0; item < NR_VM_NUMA_STAT_ITEMS; item++) {
  45                 atomic_long_set(&zone->vm_numa_stat[item], 0);
  46                 for_each_online_cpu(cpu)
  47                         per_cpu_ptr(zone->pageset, cpu)->vm_numa_stat_diff[item]
  48                                                 = 0;
  49         }
  50 }
  51
  52 /* zero numa counters of all the populated zones */
  53 static void zero_zones_numa_counters(void)
  54 {
  55         struct zone *zone;
  56
  57         for_each_populated_zone(zone)
  58                 zero_zone_numa_counters(zone);
  59 }
  60
  61 /* zero global numa counters */
  62 static void zero_global_numa_counters(void)
  63 {
  64         int item;
  65
  66         for (item = 0; item < NR_VM_NUMA_STAT_ITEMS; item++)
  67                 atomic_long_set(&vm_numa_stat[item], 0);
  68 }
  69
  70 static void invalid_numa_statistics(void)
  71 {
  72         zero_zones_numa_counters();
  73         zero_global_numa_counters();
  74 }
  75
  76 static DEFINE_MUTEX(vm_numa_stat_lock);
  77
  78 int sysctl_vm_numa_stat_handler(struct ctl_table *table, int write,
  79                 void *buffer, size_t *length, loff_t *ppos)
  80 {
  81         int ret, oldval;
  82
  83         mutex_lock(&vm_numa_stat_lock);
  84         if (write)
  85                 oldval = sysctl_vm_numa_stat;
  86         ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
  87         if (ret || !write)
  88                 goto out;
  89
  90         if (oldval == sysctl_vm_numa_stat)
  91                 goto out;
  92         else if (sysctl_vm_numa_stat == ENABLE_NUMA_STAT) {
  93                 static_branch_enable(&vm_numa_stat_key);
  94                 pr_info("enable numa statistics\n");
  95         } else {
  96                 static_branch_disable(&vm_numa_stat_key);
  97                 invalid_numa_statistics();
  98                 pr_info("disable numa statistics, and clear numa counters\n");
  99         }
 100
 101 out:
 102         mutex_unlock(&vm_numa_stat_lock);
 103         return ret;
 104 }
 105 #endif
 106
 107 #ifdef CONFIG_VM_EVENT_COUNTERS
 108 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
 109 EXPORT_PER_CPU_SYMBOL(vm_event_states);
 110
 111 static void sum_vm_events(unsigned long *ret)
 112 {
 113         int cpu;
 114         int i;
 115
 116         memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
 117
 118         for_each_online_cpu(cpu) {
 119                 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
 120
 121                 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
 122                         ret[i] += this->event[i];
 123         }
 124 }
 125
 126 /*
 127  * Accumulate the vm event counters across all CPUs.
 128  * The result is unavoidably approximate - it can change
 129  * during and after execution of this function.
 130 */
 131 void all_vm_events(unsigned long *ret)
 132 {
 133         get_online_cpus();
 134         sum_vm_events(ret);
 135         put_online_cpus();
 136 }
 137 EXPORT_SYMBOL_GPL(all_vm_events);
 138
 139 /*
 140  * Fold the foreign cpu events into our own.
 141  *
 142  * This is adding to the events on one processor
 143  * but keeps the global counts constant.
 144  */
 145 void vm_events_fold_cpu(int cpu)
 146 {
 147         struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
 148         int i;
 149
 150         for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
 151                 count_vm_events(i, fold_state->event[i]);
 152                 fold_state->event[i] = 0;
 153         }
 154 }
 155
 156 #endif /* CONFIG_VM_EVENT_COUNTERS */
 157
 158 /*
 159  * Manage combined zone based / global counters
 160  *
 161  * vm_stat contains the global counters
 162  */
 163 atomic_long_t vm_zone_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
 164 atomic_long_t vm_numa_stat[NR_VM_NUMA_STAT_ITEMS] __cacheline_aligned_in_smp;
 165 atomic_long_t vm_node_stat[NR_VM_NODE_STAT_ITEMS] __cacheline_aligned_in_smp;
 166 EXPORT_SYMBOL(vm_zone_stat);
 167 EXPORT_SYMBOL(vm_numa_stat);
 168 EXPORT_SYMBOL(vm_node_stat);
 169
 170 #ifdef CONFIG_SMP
 171
 172 int calculate_pressure_threshold(struct zone *zone)
 173 {
 174         int threshold;
 175         int watermark_distance;
 176
 177         /*
 178          * As vmstats are not up to date, there is drift between the estimated
 179          * and real values. For high thresholds and a high number of CPUs, it
 180          * is possible for the min watermark to be breached while the estimated
 181          * value looks fine. The pressure threshold is a reduced value such
 182          * that even the maximum amount of drift will not accidentally breach
 183          * the min watermark
 184          */
 185         watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
 186         threshold = max(1, (int)(watermark_distance / num_online_cpus()));
 187
 188         /*
 189          * Maximum threshold is 125
 190          */
 191         threshold = min(125, threshold);
 192
 193         return threshold;
 194 }
 195
 196 int calculate_normal_threshold(struct zone *zone)
 197 {
 198         int threshold;
 199         int mem;        /* memory in 128 MB units */
 200
 201         /*
 202          * The threshold scales with the number of processors and the amount
 203          * of memory per zone. More memory means that we can defer updates for
 204          * longer, more processors could lead to more contention.
 205          * fls() is used to have a cheap way of logarithmic scaling.
 206          *
 207          * Some sample thresholds:
 208          *
 209          * Threshold    Processors      (fls)   Zonesize        fls(mem+1)
 210          * ------------------------------------------------------------------
 211          * 8            1               1       0.9-1 GB        4
 212          * 16           2               2       0.9-1 GB        4
 213          * 20           2               2       1-2 GB          5
 214          * 24           2               2       2-4 GB          6
 215          * 28           2               2       4-8 GB          7
 216          * 32           2               2       8-16 GB         8
 217          * 4            2               2       <128M           1
 218          * 30           4               3       2-4 GB          5
 219          * 48           4               3       8-16 GB         8
 220          * 32           8               4       1-2 GB          4
 221          * 32           8               4       0.9-1GB         4
 222          * 10           16              5       <128M           1
 223          * 40           16              5       900M            4
 224          * 70           64              7       2-4 GB          5
 225          * 84           64              7       4-8 GB          6
 226          * 108          512             9       4-8 GB          6
 227          * 125          1024            10      8-16 GB         8
 228          * 125          1024            10      16-32 GB        9
 229          */
 230
 231         mem = zone_managed_pages(zone) >> (27 - PAGE_SHIFT);
 232
 233         threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
 234
 235         /*
 236          * Maximum threshold is 125
 237          */
 238         threshold = min(125, threshold);
 239
 240         return threshold;
 241 }
 242
 243 /*
 244  * Refresh the thresholds for each zone.
 245  */
 246 void refresh_zone_stat_thresholds(void)
 247 {
 248         struct pglist_data *pgdat;
 249         struct zone *zone;
 250         int cpu;
 251         int threshold;
 252
 253         /* Zero current pgdat thresholds */
 254         for_each_online_pgdat(pgdat) {
 255                 for_each_online_cpu(cpu) {
 256                         per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold = 0;
 257                 }
 258         }
 259
 260         for_each_populated_zone(zone) {
 261                 struct pglist_data *pgdat = zone->zone_pgdat;
 262                 unsigned long max_drift, tolerate_drift;
 263
 264                 threshold = calculate_normal_threshold(zone);
 265
 266                 for_each_online_cpu(cpu) {
 267                         int pgdat_threshold;
 268
 269                         per_cpu_ptr(zone->pageset, cpu)->stat_threshold
 270                                                         = threshold;
 271
 272                         /* Base nodestat threshold on the largest populated zone. */
 273                         pgdat_threshold = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold;
 274                         per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold
 275                                 = max(threshold, pgdat_threshold);
 276                 }
 277
 278                 /*
 279                  * Only set percpu_drift_mark if there is a danger that
 280                  * NR_FREE_PAGES reports the low watermark is ok when in fact
 281                  * the min watermark could be breached by an allocation
 282                  */
 283                 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
 284                 max_drift = num_online_cpus() * threshold;
 285                 if (max_drift > tolerate_drift)
 286                         zone->percpu_drift_mark = high_wmark_pages(zone) +
 287                                         max_drift;
 288         }
 289 }
 290
 291 void set_pgdat_percpu_threshold(pg_data_t *pgdat,
 292                                 int (*calculate_pressure)(struct zone *))
 293 {
 294         struct zone *zone;
 295         int cpu;
 296         int threshold;
 297         int i;
 298
 299         for (i = 0; i < pgdat->nr_zones; i++) {
 300                 zone = &pgdat->node_zones[i];
 301                 if (!zone->percpu_drift_mark)
 302                         continue;
 303
 304                 threshold = (*calculate_pressure)(zone);
 305                 for_each_online_cpu(cpu)
 306                         per_cpu_ptr(zone->pageset, cpu)->stat_threshold
 307                                                         = threshold;
 308         }
 309 }
 310
 311 /*
 312  * For use when we know that interrupts are disabled,
 313  * or when we know that preemption is disabled and that
 314  * particular counter cannot be updated from interrupt context.
 315  */
 316 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 317                            long delta)
 318 {
 319         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 320         s8 __percpu *p = pcp->vm_stat_diff + item;
 321         long x;
 322         long t;
 323
 324         x = delta + __this_cpu_read(*p);
 325
 326         t = __this_cpu_read(pcp->stat_threshold);
 327
 328         if (unlikely(abs(x) > t)) {
 329                 zone_page_state_add(x, zone, item);
 330                 x = 0;
 331         }
 332         __this_cpu_write(*p, x);
 333 }
 334 EXPORT_SYMBOL(__mod_zone_page_state);
 335
 336 void __mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
 337                                 long delta)
 338 {
 339         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 340         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 341         long x;
 342         long t;
 343
 344         if (vmstat_item_in_bytes(item)) {
 345                 /*
 346                  * Only cgroups use subpage accounting right now; at
 347                  * the global level, these items still change in
 348                  * multiples of whole pages. Store them as pages
 349                  * internally to keep the per-cpu counters compact.
 350                  */
 351                 VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
 352                 delta >>= PAGE_SHIFT;
 353         }
 354
 355         x = delta + __this_cpu_read(*p);
 356
 357         t = __this_cpu_read(pcp->stat_threshold);
 358
 359         if (unlikely(abs(x) > t)) {
 360                 node_page_state_add(x, pgdat, item);
 361                 x = 0;
 362         }
 363         __this_cpu_write(*p, x);
 364 }
 365 EXPORT_SYMBOL(__mod_node_page_state);
 366
 367 /*
 368  * Optimized increment and decrement functions.
 369  *
 370  * These are only for a single page and therefore can take a struct page *
 371  * argument instead of struct zone *. This allows the inclusion of the code
 372  * generated for page_zone(page) into the optimized functions.
 373  *
 374  * No overflow check is necessary and therefore the differential can be
 375  * incremented or decremented in place which may allow the compilers to
 376  * generate better code.
 377  * The increment or decrement is known and therefore one boundary check can
 378  * be omitted.
 379  *
 380  * NOTE: These functions are very performance sensitive. Change only
 381  * with care.
 382  *
 383  * Some processors have inc/dec instructions that are atomic vs an interrupt.
 384  * However, the code must first determine the differential location in a zone
 385  * based on the processor number and then inc/dec the counter. There is no
 386  * guarantee without disabling preemption that the processor will not change
 387  * in between and therefore the atomicity vs. interrupt cannot be exploited
 388  * in a useful way here.
 389  */
 390 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
 391 {
 392         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 393         s8 __percpu *p = pcp->vm_stat_diff + item;
 394         s8 v, t;
 395
 396         v = __this_cpu_inc_return(*p);
 397         t = __this_cpu_read(pcp->stat_threshold);
 398         if (unlikely(v > t)) {
 399                 s8 overstep = t >> 1;
 400
 401                 zone_page_state_add(v + overstep, zone, item);
 402                 __this_cpu_write(*p, -overstep);
 403         }
 404 }
 405
 406 void __inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 407 {
 408         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 409         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 410         s8 v, t;
 411
 412         VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
 413
 414         v = __this_cpu_inc_return(*p);
 415         t = __this_cpu_read(pcp->stat_threshold);
 416         if (unlikely(v > t)) {
 417                 s8 overstep = t >> 1;
 418
 419                 node_page_state_add(v + overstep, pgdat, item);
 420                 __this_cpu_write(*p, -overstep);
 421         }
 422 }
 423
 424 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
 425 {
 426         __inc_zone_state(page_zone(page), item);
 427 }
 428 EXPORT_SYMBOL(__inc_zone_page_state);
 429
 430 void __inc_node_page_state(struct page *page, enum node_stat_item item)
 431 {
 432         __inc_node_state(page_pgdat(page), item);
 433 }
 434 EXPORT_SYMBOL(__inc_node_page_state);
 435
 436 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
 437 {
 438         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 439         s8 __percpu *p = pcp->vm_stat_diff + item;
 440         s8 v, t;
 441
 442         v = __this_cpu_dec_return(*p);
 443         t = __this_cpu_read(pcp->stat_threshold);
 444         if (unlikely(v < - t)) {
 445                 s8 overstep = t >> 1;
 446
 447                 zone_page_state_add(v - overstep, zone, item);
 448                 __this_cpu_write(*p, overstep);
 449         }
 450 }
 451
 452 void __dec_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 453 {
 454         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 455         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 456         s8 v, t;
 457
 458         VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
 459
 460         v = __this_cpu_dec_return(*p);
 461         t = __this_cpu_read(pcp->stat_threshold);
 462         if (unlikely(v < - t)) {
 463                 s8 overstep = t >> 1;
 464
 465                 node_page_state_add(v - overstep, pgdat, item);
 466                 __this_cpu_write(*p, overstep);
 467         }
 468 }
 469
 470 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
 471 {
 472         __dec_zone_state(page_zone(page), item);
 473 }
 474 EXPORT_SYMBOL(__dec_zone_page_state);
 475
 476 void __dec_node_page_state(struct page *page, enum node_stat_item item)
 477 {
 478         __dec_node_state(page_pgdat(page), item);
 479 }
 480 EXPORT_SYMBOL(__dec_node_page_state);
 481
 482 #ifdef CONFIG_HAVE_CMPXCHG_LOCAL
 483 /*
 484  * If we have cmpxchg_local support then we do not need to incur the overhead
 485  * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
 486  *
 487  * mod_state() modifies the zone counter state through atomic per cpu
 488  * operations.
 489  *
 490  * Overstep mode specifies how overstep should handled:
 491  *     0       No overstepping
 492  *     1       Overstepping half of threshold
 493  *     -1      Overstepping minus half of threshold
 494 */
 495 static inline void mod_zone_state(struct zone *zone,
 496        enum zone_stat_item item, long delta, int overstep_mode)
 497 {
 498         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 499         s8 __percpu *p = pcp->vm_stat_diff + item;
 500         long o, n, t, z;
 501
 502         do {
 503                 z = 0;  /* overflow to zone counters */
 504
 505                 /*
 506                  * The fetching of the stat_threshold is racy. We may apply
 507                  * a counter threshold to the wrong the cpu if we get
 508                  * rescheduled while executing here. However, the next
 509                  * counter update will apply the threshold again and
 510                  * therefore bring the counter under the threshold again.
 511                  *
 512                  * Most of the time the thresholds are the same anyways
 513                  * for all cpus in a zone.
 514                  */
 515                 t = this_cpu_read(pcp->stat_threshold);
 516
 517                 o = this_cpu_read(*p);
 518                 n = delta + o;
 519
 520                 if (abs(n) > t) {
 521                         int os = overstep_mode * (t >> 1) ;
 522
 523                         /* Overflow must be added to zone counters */
 524                         z = n + os;
 525                         n = -os;
 526                 }
 527         } while (this_cpu_cmpxchg(*p, o, n) != o);
 528
 529         if (z)
 530                 zone_page_state_add(z, zone, item);
 531 }
 532
 533 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 534                          long delta)
 535 {
 536         mod_zone_state(zone, item, delta, 0);
 537 }
 538 EXPORT_SYMBOL(mod_zone_page_state);
 539
 540 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 541 {
 542         mod_zone_state(page_zone(page), item, 1, 1);
 543 }
 544 EXPORT_SYMBOL(inc_zone_page_state);
 545
 546 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 547 {
 548         mod_zone_state(page_zone(page), item, -1, -1);
 549 }
 550 EXPORT_SYMBOL(dec_zone_page_state);
 551
 552 static inline void mod_node_state(struct pglist_data *pgdat,
 553        enum node_stat_item item, int delta, int overstep_mode)
 554 {
 555         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 556         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 557         long o, n, t, z;
 558
 559         if (vmstat_item_in_bytes(item)) {
 560                 /*
 561                  * Only cgroups use subpage accounting right now; at
 562                  * the global level, these items still change in
 563                  * multiples of whole pages. Store them as pages
 564                  * internally to keep the per-cpu counters compact.
 565                  */
 566                 VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
 567                 delta >>= PAGE_SHIFT;
 568         }
 569
 570         do {
 571                 z = 0;  /* overflow to node counters */
 572
 573                 /*
 574                  * The fetching of the stat_threshold is racy. We may apply
 575                  * a counter threshold to the wrong the cpu if we get
 576                  * rescheduled while executing here. However, the next
 577                  * counter update will apply the threshold again and
 578                  * therefore bring the counter under the threshold again.
 579                  *
 580                  * Most of the time the thresholds are the same anyways
 581                  * for all cpus in a node.
 582                  */
 583                 t = this_cpu_read(pcp->stat_threshold);
 584
 585                 o = this_cpu_read(*p);
 586                 n = delta + o;
 587
 588                 if (abs(n) > t) {
 589                         int os = overstep_mode * (t >> 1) ;
 590
 591                         /* Overflow must be added to node counters */
 592                         z = n + os;
 593                         n = -os;
 594                 }
 595         } while (this_cpu_cmpxchg(*p, o, n) != o);
 596
 597         if (z)
 598                 node_page_state_add(z, pgdat, item);
 599 }
 600
 601 void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
 602                                         long delta)
 603 {
 604         mod_node_state(pgdat, item, delta, 0);
 605 }
 606 EXPORT_SYMBOL(mod_node_page_state);
 607
 608 void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 609 {
 610         mod_node_state(pgdat, item, 1, 1);
 611 }
 612
 613 void inc_node_page_state(struct page *page, enum node_stat_item item)
 614 {
 615         mod_node_state(page_pgdat(page), item, 1, 1);
 616 }
 617 EXPORT_SYMBOL(inc_node_page_state);
 618
 619 void dec_node_page_state(struct page *page, enum node_stat_item item)
 620 {
 621         mod_node_state(page_pgdat(page), item, -1, -1);
 622 }
 623 EXPORT_SYMBOL(dec_node_page_state);
 624 #else
 625 /*
 626  * Use interrupt disable to serialize counter updates
 627  */
 628 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 629                          long delta)
 630 {
 631         unsigned long flags;
 632
 633         local_irq_save(flags);
 634         __mod_zone_page_state(zone, item, delta);
 635         local_irq_restore(flags);
 636 }
 637 EXPORT_SYMBOL(mod_zone_page_state);
 638
 639 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 640 {
 641         unsigned long flags;
 642         struct zone *zone;
 643
 644         zone = page_zone(page);
 645         local_irq_save(flags);
 646         __inc_zone_state(zone, item);
 647         local_irq_restore(flags);
 648 }
 649 EXPORT_SYMBOL(inc_zone_page_state);
 650
 651 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 652 {
 653         unsigned long flags;
 654
 655         local_irq_save(flags);
 656         __dec_zone_page_state(page, item);
 657         local_irq_restore(flags);
 658 }
 659 EXPORT_SYMBOL(dec_zone_page_state);
 660
 661 void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 662 {
 663         unsigned long flags;
 664
 665         local_irq_save(flags);
 666         __inc_node_state(pgdat, item);
 667         local_irq_restore(flags);
 668 }
 669 EXPORT_SYMBOL(inc_node_state);
 670
 671 void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
 672                                         long delta)
 673 {
 674         unsigned long flags;
 675
 676         local_irq_save(flags);
 677         __mod_node_page_state(pgdat, item, delta);
 678         local_irq_restore(flags);
 679 }
 680 EXPORT_SYMBOL(mod_node_page_state);
 681
 682 void inc_node_page_state(struct page *page, enum node_stat_item item)
 683 {
 684         unsigned long flags;
 685         struct pglist_data *pgdat;
 686
 687         pgdat = page_pgdat(page);
 688         local_irq_save(flags);
 689         __inc_node_state(pgdat, item);
 690         local_irq_restore(flags);
 691 }
 692 EXPORT_SYMBOL(inc_node_page_state);
 693
 694 void dec_node_page_state(struct page *page, enum node_stat_item item)
 695 {
 696         unsigned long flags;
 697
 698         local_irq_save(flags);
 699         __dec_node_page_state(page, item);
 700         local_irq_restore(flags);
 701 }
 702 EXPORT_SYMBOL(dec_node_page_state);
 703 #endif
 704
 705 /*
 706  * Fold a differential into the global counters.
 707  * Returns the number of counters updated.
 708  */
 709 #ifdef CONFIG_NUMA
 710 static int fold_diff(int *zone_diff, int *numa_diff, int *node_diff)
 711 {
 712         int i;
 713         int changes = 0;
 714
 715         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 716                 if (zone_diff[i]) {
 717                         atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
 718                         changes++;
 719         }
 720
 721         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
 722                 if (numa_diff[i]) {
 723                         atomic_long_add(numa_diff[i], &vm_numa_stat[i]);
 724                         changes++;
 725         }
 726
 727         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 728                 if (node_diff[i]) {
 729                         atomic_long_add(node_diff[i], &vm_node_stat[i]);
 730                         changes++;
 731         }
 732         return changes;
 733 }
 734 #else
 735 static int fold_diff(int *zone_diff, int *node_diff)
 736 {
 737         int i;
 738         int changes = 0;
 739
 740         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 741                 if (zone_diff[i]) {
 742                         atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
 743                         changes++;
 744         }
 745
 746         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 747                 if (node_diff[i]) {
 748                         atomic_long_add(node_diff[i], &vm_node_stat[i]);
 749                         changes++;
 750         }
 751         return changes;
 752 }
 753 #endif /* CONFIG_NUMA */
 754
 755 /*
 756  * Update the zone counters for the current cpu.
 757  *
 758  * Note that refresh_cpu_vm_stats strives to only access
 759  * node local memory. The per cpu pagesets on remote zones are placed
 760  * in the memory local to the processor using that pageset. So the
 761  * loop over all zones will access a series of cachelines local to
 762  * the processor.
 763  *
 764  * The call to zone_page_state_add updates the cachelines with the
 765  * statistics in the remote zone struct as well as the global cachelines
 766  * with the global counters. These could cause remote node cache line
 767  * bouncing and will have to be only done when necessary.
 768  *
 769  * The function returns the number of global counters updated.
 770  */
 771 static int refresh_cpu_vm_stats(bool do_pagesets)
 772 {
 773         struct pglist_data *pgdat;
 774         struct zone *zone;
 775         int i;
 776         int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 777 #ifdef CONFIG_NUMA
 778         int global_numa_diff[NR_VM_NUMA_STAT_ITEMS] = { 0, };
 779 #endif
 780         int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
 781         int changes = 0;
 782
 783         for_each_populated_zone(zone) {
 784                 struct per_cpu_pageset __percpu *p = zone->pageset;
 785
 786                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
 787                         int v;
 788
 789                         v = this_cpu_xchg(p->vm_stat_diff[i], 0);
 790                         if (v) {
 791
 792                                 atomic_long_add(v, &zone->vm_stat[i]);
 793                                 global_zone_diff[i] += v;
 794 #ifdef CONFIG_NUMA
 795                                 /* 3 seconds idle till flush */
 796                                 __this_cpu_write(p->expire, 3);
 797 #endif
 798                         }
 799                 }
 800 #ifdef CONFIG_NUMA
 801                 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) {
 802                         int v;
 803
 804                         v = this_cpu_xchg(p->vm_numa_stat_diff[i], 0);
 805                         if (v) {
 806
 807                                 atomic_long_add(v, &zone->vm_numa_stat[i]);
 808                                 global_numa_diff[i] += v;
 809                                 __this_cpu_write(p->expire, 3);
 810                         }
 811                 }
 812
 813                 if (do_pagesets) {
 814                         cond_resched();
 815                         /*
 816                          * Deal with draining the remote pageset of this
 817                          * processor
 818                          *
 819                          * Check if there are pages remaining in this pageset
 820                          * if not then there is nothing to expire.
 821                          */
 822                         if (!__this_cpu_read(p->expire) ||
 823                                !__this_cpu_read(p->pcp.count))
 824                                 continue;
 825
 826                         /*
 827                          * We never drain zones local to this processor.
 828                          */
 829                         if (zone_to_nid(zone) == numa_node_id()) {
 830                                 __this_cpu_write(p->expire, 0);
 831                                 continue;
 832                         }
 833
 834                         if (__this_cpu_dec_return(p->expire))
 835                                 continue;
 836
 837                         if (__this_cpu_read(p->pcp.count)) {
 838                                 drain_zone_pages(zone, this_cpu_ptr(&p->pcp));
 839                                 changes++;
 840                         }
 841                 }
 842 #endif
 843         }
 844
 845         for_each_online_pgdat(pgdat) {
 846                 struct per_cpu_nodestat __percpu *p = pgdat->per_cpu_nodestats;
 847
 848                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
 849                         int v;
 850
 851                         v = this_cpu_xchg(p->vm_node_stat_diff[i], 0);
 852                         if (v) {
 853                                 atomic_long_add(v, &pgdat->vm_stat[i]);
 854                                 global_node_diff[i] += v;
 855                         }
 856                 }
 857         }
 858
 859 #ifdef CONFIG_NUMA
 860         changes += fold_diff(global_zone_diff, global_numa_diff,
 861                              global_node_diff);
 862 #else
 863         changes += fold_diff(global_zone_diff, global_node_diff);
 864 #endif
 865         return changes;
 866 }
 867
 868 /*
 869  * Fold the data for an offline cpu into the global array.
 870  * There cannot be any access by the offline cpu and therefore
 871  * synchronization is simplified.
 872  */
 873 void cpu_vm_stats_fold(int cpu)
 874 {
 875         struct pglist_data *pgdat;
 876         struct zone *zone;
 877         int i;
 878         int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 879 #ifdef CONFIG_NUMA
 880         int global_numa_diff[NR_VM_NUMA_STAT_ITEMS] = { 0, };
 881 #endif
 882         int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
 883
 884         for_each_populated_zone(zone) {
 885                 struct per_cpu_pageset *p;
 886
 887                 p = per_cpu_ptr(zone->pageset, cpu);
 888
 889                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 890                         if (p->vm_stat_diff[i]) {
 891                                 int v;
 892
 893                                 v = p->vm_stat_diff[i];
 894                                 p->vm_stat_diff[i] = 0;
 895                                 atomic_long_add(v, &zone->vm_stat[i]);
 896                                 global_zone_diff[i] += v;
 897                         }
 898
 899 #ifdef CONFIG_NUMA
 900                 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
 901                         if (p->vm_numa_stat_diff[i]) {
 902                                 int v;
 903
 904                                 v = p->vm_numa_stat_diff[i];
 905                                 p->vm_numa_stat_diff[i] = 0;
 906                                 atomic_long_add(v, &zone->vm_numa_stat[i]);
 907                                 global_numa_diff[i] += v;
 908                         }
 909 #endif
 910         }
 911
 912         for_each_online_pgdat(pgdat) {
 913                 struct per_cpu_nodestat *p;
 914
 915                 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
 916
 917                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 918                         if (p->vm_node_stat_diff[i]) {
 919                                 int v;
 920
 921                                 v = p->vm_node_stat_diff[i];
 922                                 p->vm_node_stat_diff[i] = 0;
 923                                 atomic_long_add(v, &pgdat->vm_stat[i]);
 924                                 global_node_diff[i] += v;
 925                         }
 926         }
 927
 928 #ifdef CONFIG_NUMA
 929         fold_diff(global_zone_diff, global_numa_diff, global_node_diff);
 930 #else
 931         fold_diff(global_zone_diff, global_node_diff);
 932 #endif
 933 }
 934
 935 /*
 936  * this is only called if !populated_zone(zone), which implies no other users of
 937  * pset->vm_stat_diff[] exsist.
 938  */
 939 void drain_zonestat(struct zone *zone, struct per_cpu_pageset *pset)
 940 {
 941         int i;
 942
 943         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 944                 if (pset->vm_stat_diff[i]) {
 945                         int v = pset->vm_stat_diff[i];
 946                         pset->vm_stat_diff[i] = 0;
 947                         atomic_long_add(v, &zone->vm_stat[i]);
 948                         atomic_long_add(v, &vm_zone_stat[i]);
 949                 }
 950
 951 #ifdef CONFIG_NUMA
 952         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
 953                 if (pset->vm_numa_stat_diff[i]) {
 954                         int v = pset->vm_numa_stat_diff[i];
 955
 956                         pset->vm_numa_stat_diff[i] = 0;
 957                         atomic_long_add(v, &zone->vm_numa_stat[i]);
 958                         atomic_long_add(v, &vm_numa_stat[i]);
 959                 }
 960 #endif
 961 }
 962 #endif
 963
 964 #ifdef CONFIG_NUMA
 965 void __inc_numa_state(struct zone *zone,
 966                                  enum numa_stat_item item)
 967 {
 968         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 969         u16 __percpu *p = pcp->vm_numa_stat_diff + item;
 970         u16 v;
 971
 972         v = __this_cpu_inc_return(*p);
 973
 974         if (unlikely(v > NUMA_STATS_THRESHOLD)) {
 975                 zone_numa_state_add(v, zone, item);
 976                 __this_cpu_write(*p, 0);
 977         }
 978 }
 979
 980 /*
 981  * Determine the per node value of a stat item. This function
 982  * is called frequently in a NUMA machine, so try to be as
 983  * frugal as possible.
 984  */
 985 unsigned long sum_zone_node_page_state(int node,
 986                                  enum zone_stat_item item)
 987 {
 988         struct zone *zones = NODE_DATA(node)->node_zones;
 989         int i;
 990         unsigned long count = 0;
 991
 992         for (i = 0; i < MAX_NR_ZONES; i++)
 993                 count += zone_page_state(zones + i, item);
 994
 995         return count;
 996 }
 997
 998 /*
 999  * Determine the per node value of a numa stat item. To avoid deviation,
1000  * the per cpu stat number in vm_numa_stat_diff[] is also included.
1001  */
1002 unsigned long sum_zone_numa_state(int node,
1003                                  enum numa_stat_item item)
1004 {
1005         struct zone *zones = NODE_DATA(node)->node_zones;
1006         int i;
1007         unsigned long count = 0;
1008
1009         for (i = 0; i < MAX_NR_ZONES; i++)
1010                 count += zone_numa_state_snapshot(zones + i, item);
1011
1012         return count;
1013 }
1014
1015 /*
1016  * Determine the per node value of a stat item.
1017  */
1018 unsigned long node_page_state_pages(struct pglist_data *pgdat,
1019                                     enum node_stat_item item)
1020 {
1021         long x = atomic_long_read(&pgdat->vm_stat[item]);
1022 #ifdef CONFIG_SMP
1023         if (x < 0)
1024                 x = 0;
1025 #endif
1026         return x;
1027 }
1028
1029 unsigned long node_page_state(struct pglist_data *pgdat,
1030                               enum node_stat_item item)
1031 {
1032         VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
1033
1034         return node_page_state_pages(pgdat, item);
1035 }
1036 #endif
1037
1038 #ifdef CONFIG_COMPACTION
1039
1040 struct contig_page_info {
1041         unsigned long free_pages;
1042         unsigned long free_blocks_total;
1043         unsigned long free_blocks_suitable;
1044 };
1045
1046 /*
1047  * Calculate the number of free pages in a zone, how many contiguous
1048  * pages are free and how many are large enough to satisfy an allocation of
1049  * the target size. Note that this function makes no attempt to estimate
1050  * how many suitable free blocks there *might* be if MOVABLE pages were
1051  * migrated. Calculating that is possible, but expensive and can be
1052  * figured out from userspace
1053  */
1054 static void fill_contig_page_info(struct zone *zone,
1055                                 unsigned int suitable_order,
1056                                 struct contig_page_info *info)
1057 {
1058         unsigned int order;
1059
1060         info->free_pages = 0;
1061         info->free_blocks_total = 0;
1062         info->free_blocks_suitable = 0;
1063
1064         for (order = 0; order < MAX_ORDER; order++) {
1065                 unsigned long blocks;
1066
1067                 /* Count number of free blocks */
1068                 blocks = zone->free_area[order].nr_free;
1069                 info->free_blocks_total += blocks;
1070
1071                 /* Count free base pages */
1072                 info->free_pages += blocks << order;
1073
1074                 /* Count the suitable free blocks */
1075                 if (order >= suitable_order)
1076                         info->free_blocks_suitable += blocks <<
1077                                                 (order - suitable_order);
1078         }
1079 }
1080
1081 /*
1082  * A fragmentation index only makes sense if an allocation of a requested
1083  * size would fail. If that is true, the fragmentation index indicates
1084  * whether external fragmentation or a lack of memory was the problem.
1085  * The value can be used to determine if page reclaim or compaction
1086  * should be used
1087  */
1088 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
1089 {
1090         unsigned long requested = 1UL << order;
1091
1092         if (WARN_ON_ONCE(order >= MAX_ORDER))
1093                 return 0;
1094
1095         if (!info->free_blocks_total)
1096                 return 0;
1097
1098         /* Fragmentation index only makes sense when a request would fail */
1099         if (info->free_blocks_suitable)
1100                 return -1000;
1101
1102         /*
1103          * Index is between 0 and 1 so return within 3 decimal places
1104          *
1105          * 0 => allocation would fail due to lack of memory
1106          * 1 => allocation would fail due to fragmentation
1107          */
1108         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
1109 }
1110
1111 /*
1112  * Calculates external fragmentation within a zone wrt the given order.
1113  * It is defined as the percentage of pages found in blocks of size
1114  * less than 1 << order. It returns values in range [0, 100].
1115  */
1116 unsigned int extfrag_for_order(struct zone *zone, unsigned int order)
1117 {
1118         struct contig_page_info info;
1119
1120         fill_contig_page_info(zone, order, &info);
1121         if (info.free_pages == 0)
1122                 return 0;
1123
1124         return div_u64((info.free_pages -
1125                         (info.free_blocks_suitable << order)) * 100,
1126                         info.free_pages);
1127 }
1128
1129 /* Same as __fragmentation index but allocs contig_page_info on stack */
1130 int fragmentation_index(struct zone *zone, unsigned int order)
1131 {
1132         struct contig_page_info info;
1133
1134         fill_contig_page_info(zone, order, &info);
1135         return __fragmentation_index(order, &info);
1136 }
1137 #endif
1138
1139 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || \
1140     defined(CONFIG_NUMA) || defined(CONFIG_MEMCG)
1141 #ifdef CONFIG_ZONE_DMA
1142 #define TEXT_FOR_DMA(xx) xx "_dma",
1143 #else
1144 #define TEXT_FOR_DMA(xx)
1145 #endif
1146
1147 #ifdef CONFIG_ZONE_DMA32
1148 #define TEXT_FOR_DMA32(xx) xx "_dma32",
1149 #else
1150 #define TEXT_FOR_DMA32(xx)
1151 #endif
1152
1153 #ifdef CONFIG_HIGHMEM
1154 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
1155 #else
1156 #define TEXT_FOR_HIGHMEM(xx)
1157 #endif
1158
1159 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1160                                         TEXT_FOR_HIGHMEM(xx) xx "_movable",
1161
1162 const char * const vmstat_text[] = {
1163         /* enum zone_stat_item counters */
1164         "nr_free_pages",
1165         "nr_zone_inactive_anon",
1166         "nr_zone_active_anon",
1167         "nr_zone_inactive_file",
1168         "nr_zone_active_file",
1169         "nr_zone_unevictable",
1170         "nr_zone_write_pending",
1171         "nr_mlock",
1172         "nr_bounce",
1173 #if IS_ENABLED(CONFIG_ZSMALLOC)
1174         "nr_zspages",
1175 #endif
1176         "nr_free_cma",
1177
1178         /* enum numa_stat_item counters */
1179 #ifdef CONFIG_NUMA
1180         "numa_hit",
1181         "numa_miss",
1182         "numa_foreign",
1183         "numa_interleave",
1184         "numa_local",
1185         "numa_other",
1186 #endif
1187
1188         /* enum node_stat_item counters */
1189         "nr_inactive_anon",
1190         "nr_active_anon",
1191         "nr_inactive_file",
1192         "nr_active_file",
1193         "nr_unevictable",
1194         "nr_slab_reclaimable",
1195         "nr_slab_unreclaimable",
1196         "nr_isolated_anon",
1197         "nr_isolated_file",
1198         "workingset_nodes",
1199         "workingset_refault_anon",
1200         "workingset_refault_file",
1201         "workingset_activate_anon",
1202         "workingset_activate_file",
1203         "workingset_restore_anon",
1204         "workingset_restore_file",
1205         "workingset_nodereclaim",
1206         "nr_anon_pages",
1207         "nr_mapped",
1208         "nr_file_pages",
1209         "nr_dirty",
1210         "nr_writeback",
1211         "nr_writeback_temp",
1212         "nr_shmem",
1213         "nr_shmem_hugepages",
1214         "nr_shmem_pmdmapped",
1215         "nr_file_hugepages",
1216         "nr_file_pmdmapped",
1217         "nr_anon_transparent_hugepages",
1218         "nr_vmscan_write",
1219         "nr_vmscan_immediate_reclaim",
1220         "nr_dirtied",
1221         "nr_written",
1222         "nr_kernel_misc_reclaimable",
1223         "nr_foll_pin_acquired",
1224         "nr_foll_pin_released",
1225         "nr_kernel_stack",
1226 #if IS_ENABLED(CONFIG_SHADOW_CALL_STACK)
1227         "nr_shadow_call_stack",
1228 #endif
1229         "nr_page_table_pages",
1230 #ifdef CONFIG_SWAP
1231         "nr_swapcached",
1232 #endif
1233
1234         /* enum writeback_stat_item counters */
1235         "nr_dirty_threshold",
1236         "nr_dirty_background_threshold",
1237
1238 #if defined(CONFIG_VM_EVENT_COUNTERS) || defined(CONFIG_MEMCG)
1239         /* enum vm_event_item counters */
1240         "pgpgin",
1241         "pgpgout",
1242         "pswpin",
1243         "pswpout",
1244
1245         TEXTS_FOR_ZONES("pgalloc")
1246         TEXTS_FOR_ZONES("allocstall")
1247         TEXTS_FOR_ZONES("pgskip")
1248
1249         "pgfree",
1250         "pgactivate",
1251         "pgdeactivate",
1252         "pglazyfree",
1253
1254         "pgfault",
1255         "pgmajfault",
1256         "pglazyfreed",
1257
1258         "pgrefill",
1259         "pgreuse",
1260         "pgsteal_kswapd",
1261         "pgsteal_direct",
1262         "pgscan_kswapd",
1263         "pgscan_direct",
1264         "pgscan_direct_throttle",
1265         "pgscan_anon",
1266         "pgscan_file",
1267         "pgsteal_anon",
1268         "pgsteal_file",
1269
1270 #ifdef CONFIG_NUMA
1271         "zone_reclaim_failed",
1272 #endif
1273         "pginodesteal",
1274         "slabs_scanned",
1275         "kswapd_inodesteal",
1276         "kswapd_low_wmark_hit_quickly",
1277         "kswapd_high_wmark_hit_quickly",
1278         "pageoutrun",
1279
1280         "pgrotated",
1281
1282         "drop_pagecache",
1283         "drop_slab",
1284         "oom_kill",
1285
1286 #ifdef CONFIG_NUMA_BALANCING
1287         "numa_pte_updates",
1288         "numa_huge_pte_updates",
1289         "numa_hint_faults",
1290         "numa_hint_faults_local",
1291         "numa_pages_migrated",
1292 #endif
1293 #ifdef CONFIG_MIGRATION
1294         "pgmigrate_success",
1295         "pgmigrate_fail",
1296         "thp_migration_success",
1297         "thp_migration_fail",
1298         "thp_migration_split",
1299 #endif
1300 #ifdef CONFIG_COMPACTION
1301         "compact_migrate_scanned",
1302         "compact_free_scanned",
1303         "compact_isolated",
1304         "compact_stall",
1305         "compact_fail",
1306         "compact_success",
1307         "compact_daemon_wake",
1308         "compact_daemon_migrate_scanned",
1309         "compact_daemon_free_scanned",
1310 #endif
1311
1312 #ifdef CONFIG_HUGETLB_PAGE
1313         "htlb_buddy_alloc_success",
1314         "htlb_buddy_alloc_fail",
1315 #endif
1316 #ifdef CONFIG_CMA
1317         "cma_alloc_success",
1318         "cma_alloc_fail",
1319 #endif
1320         "unevictable_pgs_culled",
1321         "unevictable_pgs_scanned",
1322         "unevictable_pgs_rescued",
1323         "unevictable_pgs_mlocked",
1324         "unevictable_pgs_munlocked",
1325         "unevictable_pgs_cleared",
1326         "unevictable_pgs_stranded",
1327
1328 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1329         "thp_fault_alloc",
1330         "thp_fault_fallback",
1331         "thp_fault_fallback_charge",
1332         "thp_collapse_alloc",
1333         "thp_collapse_alloc_failed",
1334         "thp_file_alloc",
1335         "thp_file_fallback",
1336         "thp_file_fallback_charge",
1337         "thp_file_mapped",
1338         "thp_split_page",
1339         "thp_split_page_failed",
1340         "thp_deferred_split_page",
1341         "thp_split_pmd",
1342 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1343         "thp_split_pud",
1344 #endif
1345         "thp_zero_page_alloc",
1346         "thp_zero_page_alloc_failed",
1347         "thp_swpout",
1348         "thp_swpout_fallback",
1349 #endif
1350 #ifdef CONFIG_MEMORY_BALLOON
1351         "balloon_inflate",
1352         "balloon_deflate",
1353 #ifdef CONFIG_BALLOON_COMPACTION
1354         "balloon_migrate",
1355 #endif
1356 #endif /* CONFIG_MEMORY_BALLOON */
1357 #ifdef CONFIG_DEBUG_TLBFLUSH
1358         "nr_tlb_remote_flush",
1359         "nr_tlb_remote_flush_received",
1360         "nr_tlb_local_flush_all",
1361         "nr_tlb_local_flush_one",
1362 #endif /* CONFIG_DEBUG_TLBFLUSH */
1363
1364 #ifdef CONFIG_DEBUG_VM_VMACACHE
1365         "vmacache_find_calls",
1366         "vmacache_find_hits",
1367 #endif
1368 #ifdef CONFIG_SWAP
1369         "swap_ra",
1370         "swap_ra_hit",
1371 #endif
1372 #endif /* CONFIG_VM_EVENT_COUNTERS || CONFIG_MEMCG */
1373 };
1374 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA || CONFIG_MEMCG */
1375
1376 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1377      defined(CONFIG_PROC_FS)
1378 static void *frag_start(struct seq_file *m, loff_t *pos)
1379 {
1380         pg_data_t *pgdat;
1381         loff_t node = *pos;
1382
1383         for (pgdat = first_online_pgdat();
1384              pgdat && node;
1385              pgdat = next_online_pgdat(pgdat))
1386                 --node;
1387
1388         return pgdat;
1389 }
1390
1391 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1392 {
1393         pg_data_t *pgdat = (pg_data_t *)arg;
1394
1395         (*pos)++;
1396         return next_online_pgdat(pgdat);
1397 }
1398
1399 static void frag_stop(struct seq_file *m, void *arg)
1400 {
1401 }
1402
1403 /*
1404  * Walk zones in a node and print using a callback.
1405  * If @assert_populated is true, only use callback for zones that are populated.
1406  */
1407 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1408                 bool assert_populated, bool nolock,
1409                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1410 {
1411         struct zone *zone;
1412         struct zone *node_zones = pgdat->node_zones;
1413         unsigned long flags;
1414
1415         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1416                 if (assert_populated && !populated_zone(zone))
1417                         continue;
1418
1419                 if (!nolock)
1420                         spin_lock_irqsave(&zone->lock, flags);
1421                 print(m, pgdat, zone);
1422                 if (!nolock)
1423                         spin_unlock_irqrestore(&zone->lock, flags);
1424         }
1425 }
1426 #endif
1427
1428 #ifdef CONFIG_PROC_FS
1429 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1430                                                 struct zone *zone)
1431 {
1432         int order;
1433
1434         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1435         for (order = 0; order < MAX_ORDER; ++order)
1436                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
1437         seq_putc(m, '\n');
1438 }
1439
1440 /*
1441  * This walks the free areas for each zone.
1442  */
1443 static int frag_show(struct seq_file *m, void *arg)
1444 {
1445         pg_data_t *pgdat = (pg_data_t *)arg;
1446         walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1447         return 0;
1448 }
1449
1450 static void pagetypeinfo_showfree_print(struct seq_file *m,
1451                                         pg_data_t *pgdat, struct zone *zone)
1452 {
1453         int order, mtype;
1454
1455         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1456                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
1457                                         pgdat->node_id,
1458                                         zone->name,
1459                                         migratetype_names[mtype]);
1460                 for (order = 0; order < MAX_ORDER; ++order) {
1461                         unsigned long freecount = 0;
1462                         struct free_area *area;
1463                         struct list_head *curr;
1464                         bool overflow = false;
1465
1466                         area = &(zone->free_area[order]);
1467
1468                         list_for_each(curr, &area->free_list[mtype]) {
1469                                 /*
1470                                  * Cap the free_list iteration because it might
1471                                  * be really large and we are under a spinlock
1472                                  * so a long time spent here could trigger a
1473                                  * hard lockup detector. Anyway this is a
1474                                  * debugging tool so knowing there is a handful
1475                                  * of pages of this order should be more than
1476                                  * sufficient.
1477                                  */
1478                                 if (++freecount >= 100000) {
1479                                         overflow = true;
1480                                         break;
1481                                 }
1482                         }
1483                         seq_printf(m, "%s%6lu ", overflow ? ">" : "", freecount);
1484                         spin_unlock_irq(&zone->lock);
1485                         cond_resched();
1486                         spin_lock_irq(&zone->lock);
1487                 }
1488                 seq_putc(m, '\n');
1489         }
1490 }
1491
1492 /* Print out the free pages at each order for each migatetype */
1493 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
1494 {
1495         int order;
1496         pg_data_t *pgdat = (pg_data_t *)arg;
1497
1498         /* Print header */
1499         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1500         for (order = 0; order < MAX_ORDER; ++order)
1501                 seq_printf(m, "%6d ", order);
1502         seq_putc(m, '\n');
1503
1504         walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1505
1506         return 0;
1507 }
1508
1509 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1510                                         pg_data_t *pgdat, struct zone *zone)
1511 {
1512         int mtype;
1513         unsigned long pfn;
1514         unsigned long start_pfn = zone->zone_start_pfn;
1515         unsigned long end_pfn = zone_end_pfn(zone);
1516         unsigned long count[MIGRATE_TYPES] = { 0, };
1517
1518         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1519                 struct page *page;
1520
1521                 page = pfn_to_online_page(pfn);
1522                 if (!page)
1523                         continue;
1524
1525                 if (page_zone(page) != zone)
1526                         continue;
1527
1528                 mtype = get_pageblock_migratetype(page);
1529
1530                 if (mtype < MIGRATE_TYPES)
1531                         count[mtype]++;
1532         }
1533
1534         /* Print counts */
1535         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1536         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1537                 seq_printf(m, "%12lu ", count[mtype]);
1538         seq_putc(m, '\n');
1539 }
1540
1541 /* Print out the number of pageblocks for each migratetype */
1542 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1543 {
1544         int mtype;
1545         pg_data_t *pgdat = (pg_data_t *)arg;
1546
1547         seq_printf(m, "\n%-23s", "Number of blocks type ");
1548         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1549                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1550         seq_putc(m, '\n');
1551         walk_zones_in_node(m, pgdat, true, false,
1552                 pagetypeinfo_showblockcount_print);
1553
1554         return 0;
1555 }
1556
1557 /*
1558  * Print out the number of pageblocks for each migratetype that contain pages
1559  * of other types. This gives an indication of how well fallbacks are being
1560  * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1561  * to determine what is going on
1562  */
1563 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1564 {
1565 #ifdef CONFIG_PAGE_OWNER
1566         int mtype;
1567
1568         if (!static_branch_unlikely(&page_owner_inited))
1569                 return;
1570
1571         drain_all_pages(NULL);
1572
1573         seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1574         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1575                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1576         seq_putc(m, '\n');
1577
1578         walk_zones_in_node(m, pgdat, true, true,
1579                 pagetypeinfo_showmixedcount_print);
1580 #endif /* CONFIG_PAGE_OWNER */
1581 }
1582
1583 /*
1584  * This prints out statistics in relation to grouping pages by mobility.
1585  * It is expensive to collect so do not constantly read the file.
1586  */
1587 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1588 {
1589         pg_data_t *pgdat = (pg_data_t *)arg;
1590
1591         /* check memoryless node */
1592         if (!node_state(pgdat->node_id, N_MEMORY))
1593                 return 0;
1594
1595         seq_printf(m, "Page block order: %d\n", pageblock_order);
1596         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
1597         seq_putc(m, '\n');
1598         pagetypeinfo_showfree(m, pgdat);
1599         pagetypeinfo_showblockcount(m, pgdat);
1600         pagetypeinfo_showmixedcount(m, pgdat);
1601
1602         return 0;
1603 }
1604
1605 static const struct seq_operations fragmentation_op = {
1606         .start  = frag_start,
1607         .next   = frag_next,
1608         .stop   = frag_stop,
1609         .show   = frag_show,
1610 };
1611
1612 static const struct seq_operations pagetypeinfo_op = {
1613         .start  = frag_start,
1614         .next   = frag_next,
1615         .stop   = frag_stop,
1616         .show   = pagetypeinfo_show,
1617 };
1618
1619 static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1620 {
1621         int zid;
1622
1623         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1624                 struct zone *compare = &pgdat->node_zones[zid];
1625
1626                 if (populated_zone(compare))
1627                         return zone == compare;
1628         }
1629
1630         return false;
1631 }
1632
1633 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1634                                                         struct zone *zone)
1635 {
1636         int i;
1637         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1638         if (is_zone_first_populated(pgdat, zone)) {
1639                 seq_printf(m, "\n  per-node stats");
1640                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1641                         unsigned long pages = node_page_state_pages(pgdat, i);
1642
1643                         if (vmstat_item_print_in_thp(i))
1644                                 pages /= HPAGE_PMD_NR;
1645                         seq_printf(m, "\n      %-12s %lu", node_stat_name(i),
1646                                    pages);
1647                 }
1648         }
1649         seq_printf(m,
1650                    "\n  pages free     %lu"
1651                    "\n        min      %lu"
1652                    "\n        low      %lu"
1653                    "\n        high     %lu"
1654                    "\n        spanned  %lu"
1655                    "\n        present  %lu"
1656                    "\n        managed  %lu"
1657                    "\n        cma      %lu",
1658                    zone_page_state(zone, NR_FREE_PAGES),
1659                    min_wmark_pages(zone),
1660                    low_wmark_pages(zone),
1661                    high_wmark_pages(zone),
1662                    zone->spanned_pages,
1663                    zone->present_pages,
1664                    zone_managed_pages(zone),
1665                    zone_cma_pages(zone));
1666
1667         seq_printf(m,
1668                    "\n        protection: (%ld",
1669                    zone->lowmem_reserve[0]);
1670         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1671                 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1672         seq_putc(m, ')');
1673
1674         /* If unpopulated, no other information is useful */
1675         if (!populated_zone(zone)) {
1676                 seq_putc(m, '\n');
1677                 return;
1678         }
1679
1680         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1681                 seq_printf(m, "\n      %-12s %lu", zone_stat_name(i),
1682                            zone_page_state(zone, i));
1683
1684 #ifdef CONFIG_NUMA
1685         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1686                 seq_printf(m, "\n      %-12s %lu", numa_stat_name(i),
1687                            zone_numa_state_snapshot(zone, i));
1688 #endif
1689
1690         seq_printf(m, "\n  pagesets");
1691         for_each_online_cpu(i) {
1692                 struct per_cpu_pageset *pageset;
1693
1694                 pageset = per_cpu_ptr(zone->pageset, i);
1695                 seq_printf(m,
1696                            "\n    cpu: %i"
1697                            "\n              count: %i"
1698                            "\n              high:  %i"
1699                            "\n              batch: %i",
1700                            i,
1701                            pageset->pcp.count,
1702                            pageset->pcp.high,
1703                            pageset->pcp.batch);
1704 #ifdef CONFIG_SMP
1705                 seq_printf(m, "\n  vm stats threshold: %d",
1706                                 pageset->stat_threshold);
1707 #endif
1708         }
1709         seq_printf(m,
1710                    "\n  node_unreclaimable:  %u"
1711                    "\n  start_pfn:           %lu",
1712                    pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1713                    zone->zone_start_pfn);
1714         seq_putc(m, '\n');
1715 }
1716
1717 /*
1718  * Output information about zones in @pgdat.  All zones are printed regardless
1719  * of whether they are populated or not: lowmem_reserve_ratio operates on the
1720  * set of all zones and userspace would not be aware of such zones if they are
1721  * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1722  */
1723 static int zoneinfo_show(struct seq_file *m, void *arg)
1724 {
1725         pg_data_t *pgdat = (pg_data_t *)arg;
1726         walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1727         return 0;
1728 }
1729
1730 static const struct seq_operations zoneinfo_op = {
1731         .start  = frag_start, /* iterate over all zones. The same as in
1732                                * fragmentation. */
1733         .next   = frag_next,
1734         .stop   = frag_stop,
1735         .show   = zoneinfo_show,
1736 };
1737
1738 #define NR_VMSTAT_ITEMS (NR_VM_ZONE_STAT_ITEMS + \
1739                          NR_VM_NUMA_STAT_ITEMS + \
1740                          NR_VM_NODE_STAT_ITEMS + \
1741                          NR_VM_WRITEBACK_STAT_ITEMS + \
1742                          (IS_ENABLED(CONFIG_VM_EVENT_COUNTERS) ? \
1743                           NR_VM_EVENT_ITEMS : 0))
1744
1745 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1746 {
1747         unsigned long *v;
1748         int i;
1749
1750         if (*pos >= NR_VMSTAT_ITEMS)
1751                 return NULL;
1752
1753         BUILD_BUG_ON(ARRAY_SIZE(vmstat_text) < NR_VMSTAT_ITEMS);
1754         v = kmalloc_array(NR_VMSTAT_ITEMS, sizeof(unsigned long), GFP_KERNEL);
1755         m->private = v;
1756         if (!v)
1757                 return ERR_PTR(-ENOMEM);
1758         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1759                 v[i] = global_zone_page_state(i);
1760         v += NR_VM_ZONE_STAT_ITEMS;
1761
1762 #ifdef CONFIG_NUMA
1763         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1764                 v[i] = global_numa_state(i);
1765         v += NR_VM_NUMA_STAT_ITEMS;
1766 #endif
1767
1768         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1769                 v[i] = global_node_page_state_pages(i);
1770                 if (vmstat_item_print_in_thp(i))
1771                         v[i] /= HPAGE_PMD_NR;
1772         }
1773         v += NR_VM_NODE_STAT_ITEMS;
1774
1775         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1776                             v + NR_DIRTY_THRESHOLD);
1777         v += NR_VM_WRITEBACK_STAT_ITEMS;
1778
1779 #ifdef CONFIG_VM_EVENT_COUNTERS
1780         all_vm_events(v);
1781         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1782         v[PGPGOUT] /= 2;
1783 #endif
1784         return (unsigned long *)m->private + *pos;
1785 }
1786
1787 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1788 {
1789         (*pos)++;
1790         if (*pos >= NR_VMSTAT_ITEMS)
1791                 return NULL;
1792         return (unsigned long *)m->private + *pos;
1793 }
1794
1795 static int vmstat_show(struct seq_file *m, void *arg)
1796 {
1797         unsigned long *l = arg;
1798         unsigned long off = l - (unsigned long *)m->private;
1799
1800         seq_puts(m, vmstat_text[off]);
1801         seq_put_decimal_ull(m, " ", *l);
1802         seq_putc(m, '\n');
1803
1804         if (off == NR_VMSTAT_ITEMS - 1) {
1805                 /*
1806                  * We've come to the end - add any deprecated counters to avoid
1807                  * breaking userspace which might depend on them being present.
1808                  */
1809                 seq_puts(m, "nr_unstable 0\n");
1810         }
1811         return 0;
1812 }
1813
1814 static void vmstat_stop(struct seq_file *m, void *arg)
1815 {
1816         kfree(m->private);
1817         m->private = NULL;
1818 }
1819
1820 static const struct seq_operations vmstat_op = {
1821         .start  = vmstat_start,
1822         .next   = vmstat_next,
1823         .stop   = vmstat_stop,
1824         .show   = vmstat_show,
1825 };
1826 #endif /* CONFIG_PROC_FS */
1827
1828 #ifdef CONFIG_SMP
1829 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1830 int sysctl_stat_interval __read_mostly = HZ;
1831
1832 #ifdef CONFIG_PROC_FS
1833 static void refresh_vm_stats(struct work_struct *work)
1834 {
1835         refresh_cpu_vm_stats(true);
1836 }
1837
1838 int vmstat_refresh(struct ctl_table *table, int write,
1839                    void *buffer, size_t *lenp, loff_t *ppos)
1840 {
1841         long val;
1842         int err;
1843         int i;
1844
1845         /*
1846          * The regular update, every sysctl_stat_interval, may come later
1847          * than expected: leaving a significant amount in per_cpu buckets.
1848          * This is particularly misleading when checking a quantity of HUGE
1849          * pages, immediately after running a test.  /proc/sys/vm/stat_refresh,
1850          * which can equally be echo'ed to or cat'ted from (by root),
1851          * can be used to update the stats just before reading them.
1852          *
1853          * Oh, and since global_zone_page_state() etc. are so careful to hide
1854          * transiently negative values, report an error here if any of
1855          * the stats is negative, so we know to go looking for imbalance.
1856          */
1857         err = schedule_on_each_cpu(refresh_vm_stats);
1858         if (err)
1859                 return err;
1860         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1861                 val = atomic_long_read(&vm_zone_stat[i]);
1862                 if (val < 0) {
1863                         pr_warn("%s: %s %ld\n",
1864                                 __func__, zone_stat_name(i), val);
1865                         err = -EINVAL;
1866                 }
1867         }
1868 #ifdef CONFIG_NUMA
1869         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) {
1870                 val = atomic_long_read(&vm_numa_stat[i]);
1871                 if (val < 0) {
1872                         pr_warn("%s: %s %ld\n",
1873                                 __func__, numa_stat_name(i), val);
1874                         err = -EINVAL;
1875                 }
1876         }
1877 #endif
1878         if (err)
1879                 return err;
1880         if (write)
1881                 *ppos += *lenp;
1882         else
1883                 *lenp = 0;
1884         return 0;
1885 }
1886 #endif /* CONFIG_PROC_FS */
1887
1888 static void vmstat_update(struct work_struct *w)
1889 {
1890         if (refresh_cpu_vm_stats(true)) {
1891                 /*
1892                  * Counters were updated so we expect more updates
1893                  * to occur in the future. Keep on running the
1894                  * update worker thread.
1895                  */
1896                 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
1897                                 this_cpu_ptr(&vmstat_work),
1898                                 round_jiffies_relative(sysctl_stat_interval));
1899         }
1900 }
1901
1902 /*
1903  * Switch off vmstat processing and then fold all the remaining differentials
1904  * until the diffs stay at zero. The function is used by NOHZ and can only be
1905  * invoked when tick processing is not active.
1906  */
1907 /*
1908  * Check if the diffs for a certain cpu indicate that
1909  * an update is needed.
1910  */
1911 static bool need_update(int cpu)
1912 {
1913         pg_data_t *last_pgdat = NULL;
1914         struct zone *zone;
1915
1916         for_each_populated_zone(zone) {
1917                 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1918                 struct per_cpu_nodestat *n;
1919                 /*
1920                  * The fast way of checking if there are any vmstat diffs.
1921                  */
1922                 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS *
1923                                sizeof(p->vm_stat_diff[0])))
1924                         return true;
1925 #ifdef CONFIG_NUMA
1926                 if (memchr_inv(p->vm_numa_stat_diff, 0, NR_VM_NUMA_STAT_ITEMS *
1927                                sizeof(p->vm_numa_stat_diff[0])))
1928                         return true;
1929 #endif
1930                 if (last_pgdat == zone->zone_pgdat)
1931                         continue;
1932                 last_pgdat = zone->zone_pgdat;
1933                 n = per_cpu_ptr(zone->zone_pgdat->per_cpu_nodestats, cpu);
1934                 if (memchr_inv(n->vm_node_stat_diff, 0, NR_VM_NODE_STAT_ITEMS *
1935                                sizeof(n->vm_node_stat_diff[0])))
1936                     return true;
1937         }
1938         return false;
1939 }
1940
1941 /*
1942  * Switch off vmstat processing and then fold all the remaining differentials
1943  * until the diffs stay at zero. The function is used by NOHZ and can only be
1944  * invoked when tick processing is not active.
1945  */
1946 void quiet_vmstat(void)
1947 {
1948         if (system_state != SYSTEM_RUNNING)
1949                 return;
1950
1951         if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
1952                 return;
1953
1954         if (!need_update(smp_processor_id()))
1955                 return;
1956
1957         /*
1958          * Just refresh counters and do not care about the pending delayed
1959          * vmstat_update. It doesn't fire that often to matter and canceling
1960          * it would be too expensive from this path.
1961          * vmstat_shepherd will take care about that for us.
1962          */
1963         refresh_cpu_vm_stats(false);
1964 }
1965
1966 /*
1967  * Shepherd worker thread that checks the
1968  * differentials of processors that have their worker
1969  * threads for vm statistics updates disabled because of
1970  * inactivity.
1971  */
1972 static void vmstat_shepherd(struct work_struct *w);
1973
1974 static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
1975
1976 static void vmstat_shepherd(struct work_struct *w)
1977 {
1978         int cpu;
1979
1980         get_online_cpus();
1981         /* Check processors whose vmstat worker threads have been disabled */
1982         for_each_online_cpu(cpu) {
1983                 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
1984
1985                 if (!delayed_work_pending(dw) && need_update(cpu))
1986                         queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
1987
1988                 cond_resched();
1989         }
1990         put_online_cpus();
1991
1992         schedule_delayed_work(&shepherd,
1993                 round_jiffies_relative(sysctl_stat_interval));
1994 }
1995
1996 static void __init start_shepherd_timer(void)
1997 {
1998         int cpu;
1999
2000         for_each_possible_cpu(cpu)
2001                 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
2002                         vmstat_update);
2003
2004         schedule_delayed_work(&shepherd,
2005                 round_jiffies_relative(sysctl_stat_interval));
2006 }
2007
2008 static void __init init_cpu_node_state(void)
2009 {
2010         int node;
2011
2012         for_each_online_node(node) {
2013                 if (cpumask_weight(cpumask_of_node(node)) > 0)
2014                         node_set_state(node, N_CPU);
2015         }
2016 }
2017
2018 static int vmstat_cpu_online(unsigned int cpu)
2019 {
2020         refresh_zone_stat_thresholds();
2021         node_set_state(cpu_to_node(cpu), N_CPU);
2022         return 0;
2023 }
2024
2025 static int vmstat_cpu_down_prep(unsigned int cpu)
2026 {
2027         cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
2028         return 0;
2029 }
2030
2031 static int vmstat_cpu_dead(unsigned int cpu)
2032 {
2033         const struct cpumask *node_cpus;
2034         int node;
2035
2036         node = cpu_to_node(cpu);
2037
2038         refresh_zone_stat_thresholds();
2039         node_cpus = cpumask_of_node(node);
2040         if (cpumask_weight(node_cpus) > 0)
2041                 return 0;
2042
2043         node_clear_state(node, N_CPU);
2044         return 0;
2045 }
2046
2047 #endif
2048
2049 struct workqueue_struct *mm_percpu_wq;
2050
2051 void __init init_mm_internals(void)
2052 {
2053         int ret __maybe_unused;
2054
2055         mm_percpu_wq = alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM, 0);
2056
2057 #ifdef CONFIG_SMP
2058         ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
2059                                         NULL, vmstat_cpu_dead);
2060         if (ret < 0)
2061                 pr_err("vmstat: failed to register 'dead' hotplug state\n");
2062
2063         ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "mm/vmstat:online",
2064                                         vmstat_cpu_online,
2065                                         vmstat_cpu_down_prep);
2066         if (ret < 0)
2067                 pr_err("vmstat: failed to register 'online' hotplug state\n");
2068
2069         get_online_cpus();
2070         init_cpu_node_state();
2071         put_online_cpus();
2072
2073         start_shepherd_timer();
2074 #endif
2075 #ifdef CONFIG_PROC_FS
2076         proc_create_seq("buddyinfo", 0444, NULL, &fragmentation_op);
2077         proc_create_seq("pagetypeinfo", 0400, NULL, &pagetypeinfo_op);
2078         proc_create_seq("vmstat", 0444, NULL, &vmstat_op);
2079         proc_create_seq("zoneinfo", 0444, NULL, &zoneinfo_op);
2080 #endif
2081 }
2082
2083 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
2084
2085 /*
2086  * Return an index indicating how much of the available free memory is
2087  * unusable for an allocation of the requested size.
2088  */
2089 static int unusable_free_index(unsigned int order,
2090                                 struct contig_page_info *info)
2091 {
2092         /* No free memory is interpreted as all free memory is unusable */
2093         if (info->free_pages == 0)
2094                 return 1000;
2095
2096         /*
2097          * Index should be a value between 0 and 1. Return a value to 3
2098          * decimal places.
2099          *
2100          * 0 => no fragmentation
2101          * 1 => high fragmentation
2102          */
2103         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
2104
2105 }
2106
2107 static void unusable_show_print(struct seq_file *m,
2108                                         pg_data_t *pgdat, struct zone *zone)
2109 {
2110         unsigned int order;
2111         int index;
2112         struct contig_page_info info;
2113
2114         seq_printf(m, "Node %d, zone %8s ",
2115                                 pgdat->node_id,
2116                                 zone->name);
2117         for (order = 0; order < MAX_ORDER; ++order) {
2118                 fill_contig_page_info(zone, order, &info);
2119                 index = unusable_free_index(order, &info);
2120                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2121         }
2122
2123         seq_putc(m, '\n');
2124 }
2125
2126 /*
2127  * Display unusable free space index
2128  *
2129  * The unusable free space index measures how much of the available free
2130  * memory cannot be used to satisfy an allocation of a given size and is a
2131  * value between 0 and 1. The higher the value, the more of free memory is
2132  * unusable and by implication, the worse the external fragmentation is. This
2133  * can be expressed as a percentage by multiplying by 100.
2134  */
2135 static int unusable_show(struct seq_file *m, void *arg)
2136 {
2137         pg_data_t *pgdat = (pg_data_t *)arg;
2138
2139         /* check memoryless node */
2140         if (!node_state(pgdat->node_id, N_MEMORY))
2141                 return 0;
2142
2143         walk_zones_in_node(m, pgdat, true, false, unusable_show_print);
2144
2145         return 0;
2146 }
2147
2148 static const struct seq_operations unusable_sops = {
2149         .start  = frag_start,
2150         .next   = frag_next,
2151         .stop   = frag_stop,
2152         .show   = unusable_show,
2153 };
2154
2155 DEFINE_SEQ_ATTRIBUTE(unusable);
2156
2157 static void extfrag_show_print(struct seq_file *m,
2158                                         pg_data_t *pgdat, struct zone *zone)
2159 {
2160         unsigned int order;
2161         int index;
2162
2163         /* Alloc on stack as interrupts are disabled for zone walk */
2164         struct contig_page_info info;
2165
2166         seq_printf(m, "Node %d, zone %8s ",
2167                                 pgdat->node_id,
2168                                 zone->name);
2169         for (order = 0; order < MAX_ORDER; ++order) {
2170                 fill_contig_page_info(zone, order, &info);
2171                 index = __fragmentation_index(order, &info);
2172                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2173         }
2174
2175         seq_putc(m, '\n');
2176 }
2177
2178 /*
2179  * Display fragmentation index for orders that allocations would fail for
2180  */
2181 static int extfrag_show(struct seq_file *m, void *arg)
2182 {
2183         pg_data_t *pgdat = (pg_data_t *)arg;
2184
2185         walk_zones_in_node(m, pgdat, true, false, extfrag_show_print);
2186
2187         return 0;
2188 }
2189
2190 static const struct seq_operations extfrag_sops = {
2191         .start  = frag_start,
2192         .next   = frag_next,
2193         .stop   = frag_stop,
2194         .show   = extfrag_show,
2195 };
2196
2197 DEFINE_SEQ_ATTRIBUTE(extfrag);
2198
2199 static int __init extfrag_debug_init(void)
2200 {
2201         struct dentry *extfrag_debug_root;
2202
2203         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
2204
2205         debugfs_create_file("unusable_index", 0444, extfrag_debug_root, NULL,
2206                             &unusable_fops);
2207
2208         debugfs_create_file("extfrag_index", 0444, extfrag_debug_root, NULL,
2209                             &extfrag_fops);
2210
2211         return 0;
2212 }
2213
2214 module_init(extfrag_debug_init);
2215 #endif