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