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 <[email protected]>
  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_owner.h>
  30 #include <linux/sched/isolation.h>
  31
  32 #include "internal.h"
  33
  34 #ifdef CONFIG_NUMA
  35 int sysctl_vm_numa_stat = ENABLE_NUMA_STAT;
  36
  37 /* zero numa counters within a zone */
  38 static void zero_zone_numa_counters(struct zone *zone)
  39 {
  40         int item, cpu;
  41
  42         for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++) {
  43                 atomic_long_set(&zone->vm_numa_event[item], 0);
  44                 for_each_online_cpu(cpu) {
  45                         per_cpu_ptr(zone->per_cpu_zonestats, cpu)->vm_numa_event[item]
  46                                                 = 0;
  47                 }
  48         }
  49 }
  50
  51 /* zero numa counters of all the populated zones */
  52 static void zero_zones_numa_counters(void)
  53 {
  54         struct zone *zone;
  55
  56         for_each_populated_zone(zone)
  57                 zero_zone_numa_counters(zone);
  58 }
  59
  60 /* zero global numa counters */
  61 static void zero_global_numa_counters(void)
  62 {
  63         int item;
  64
  65         for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++)
  66                 atomic_long_set(&vm_numa_event[item], 0);
  67 }
  68
  69 static void invalid_numa_statistics(void)
  70 {
  71         zero_zones_numa_counters();
  72         zero_global_numa_counters();
  73 }
  74
  75 static DEFINE_MUTEX(vm_numa_stat_lock);
  76
  77 int sysctl_vm_numa_stat_handler(const struct ctl_table *table, int write,
  78                 void *buffer, size_t *length, loff_t *ppos)
  79 {
  80         int ret, oldval;
  81
  82         mutex_lock(&vm_numa_stat_lock);
  83         if (write)
  84                 oldval = sysctl_vm_numa_stat;
  85         ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
  86         if (ret || !write)
  87                 goto out;
  88
  89         if (oldval == sysctl_vm_numa_stat)
  90                 goto out;
  91         else if (sysctl_vm_numa_stat == ENABLE_NUMA_STAT) {
  92                 static_branch_enable(&vm_numa_stat_key);
  93                 pr_info("enable numa statistics\n");
  94         } else {
  95                 static_branch_disable(&vm_numa_stat_key);
  96                 invalid_numa_statistics();
  97                 pr_info("disable numa statistics, and clear numa counters\n");
  98         }
  99
 100 out:
 101         mutex_unlock(&vm_numa_stat_lock);
 102         return ret;
 103 }
 104 #endif
 105
 106 #ifdef CONFIG_VM_EVENT_COUNTERS
 107 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
 108 EXPORT_PER_CPU_SYMBOL(vm_event_states);
 109
 110 static void sum_vm_events(unsigned long *ret)
 111 {
 112         int cpu;
 113         int i;
 114
 115         memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
 116
 117         for_each_online_cpu(cpu) {
 118                 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
 119
 120                 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
 121                         ret[i] += this->event[i];
 122         }
 123 }
 124
 125 /*
 126  * Accumulate the vm event counters across all CPUs.
 127  * The result is unavoidably approximate - it can change
 128  * during and after execution of this function.
 129 */
 130 void all_vm_events(unsigned long *ret)
 131 {
 132         cpus_read_lock();
 133         sum_vm_events(ret);
 134         cpus_read_unlock();
 135 }
 136 EXPORT_SYMBOL_GPL(all_vm_events);
 137
 138 /*
 139  * Fold the foreign cpu events into our own.
 140  *
 141  * This is adding to the events on one processor
 142  * but keeps the global counts constant.
 143  */
 144 void vm_events_fold_cpu(int cpu)
 145 {
 146         struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
 147         int i;
 148
 149         for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
 150                 count_vm_events(i, fold_state->event[i]);
 151                 fold_state->event[i] = 0;
 152         }
 153 }
 154
 155 #endif /* CONFIG_VM_EVENT_COUNTERS */
 156
 157 /*
 158  * Manage combined zone based / global counters
 159  *
 160  * vm_stat contains the global counters
 161  */
 162 atomic_long_t vm_zone_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
 163 atomic_long_t vm_node_stat[NR_VM_NODE_STAT_ITEMS] __cacheline_aligned_in_smp;
 164 atomic_long_t vm_numa_event[NR_VM_NUMA_EVENT_ITEMS] __cacheline_aligned_in_smp;
 165 EXPORT_SYMBOL(vm_zone_stat);
 166 EXPORT_SYMBOL(vm_node_stat);
 167
 168 #ifdef CONFIG_NUMA
 169 static void fold_vm_zone_numa_events(struct zone *zone)
 170 {
 171         unsigned long zone_numa_events[NR_VM_NUMA_EVENT_ITEMS] = { 0, };
 172         int cpu;
 173         enum numa_stat_item item;
 174
 175         for_each_online_cpu(cpu) {
 176                 struct per_cpu_zonestat *pzstats;
 177
 178                 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
 179                 for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++)
 180                         zone_numa_events[item] += xchg(&pzstats->vm_numa_event[item], 0);
 181         }
 182
 183         for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++)
 184                 zone_numa_event_add(zone_numa_events[item], zone, item);
 185 }
 186
 187 void fold_vm_numa_events(void)
 188 {
 189         struct zone *zone;
 190
 191         for_each_populated_zone(zone)
 192                 fold_vm_zone_numa_events(zone);
 193 }
 194 #endif
 195
 196 #ifdef CONFIG_SMP
 197
 198 int calculate_pressure_threshold(struct zone *zone)
 199 {
 200         int threshold;
 201         int watermark_distance;
 202
 203         /*
 204          * As vmstats are not up to date, there is drift between the estimated
 205          * and real values. For high thresholds and a high number of CPUs, it
 206          * is possible for the min watermark to be breached while the estimated
 207          * value looks fine. The pressure threshold is a reduced value such
 208          * that even the maximum amount of drift will not accidentally breach
 209          * the min watermark
 210          */
 211         watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
 212         threshold = max(1, (int)(watermark_distance / num_online_cpus()));
 213
 214         /*
 215          * Maximum threshold is 125
 216          */
 217         threshold = min(125, threshold);
 218
 219         return threshold;
 220 }
 221
 222 int calculate_normal_threshold(struct zone *zone)
 223 {
 224         int threshold;
 225         int mem;        /* memory in 128 MB units */
 226
 227         /*
 228          * The threshold scales with the number of processors and the amount
 229          * of memory per zone. More memory means that we can defer updates for
 230          * longer, more processors could lead to more contention.
 231          * fls() is used to have a cheap way of logarithmic scaling.
 232          *
 233          * Some sample thresholds:
 234          *
 235          * Threshold    Processors      (fls)   Zonesize        fls(mem)+1
 236          * ------------------------------------------------------------------
 237          * 8            1               1       0.9-1 GB        4
 238          * 16           2               2       0.9-1 GB        4
 239          * 20           2               2       1-2 GB          5
 240          * 24           2               2       2-4 GB          6
 241          * 28           2               2       4-8 GB          7
 242          * 32           2               2       8-16 GB         8
 243          * 4            2               2       <128M           1
 244          * 30           4               3       2-4 GB          5
 245          * 48           4               3       8-16 GB         8
 246          * 32           8               4       1-2 GB          4
 247          * 32           8               4       0.9-1GB         4
 248          * 10           16              5       <128M           1
 249          * 40           16              5       900M            4
 250          * 70           64              7       2-4 GB          5
 251          * 84           64              7       4-8 GB          6
 252          * 108          512             9       4-8 GB          6
 253          * 125          1024            10      8-16 GB         8
 254          * 125          1024            10      16-32 GB        9
 255          */
 256
 257         mem = zone_managed_pages(zone) >> (27 - PAGE_SHIFT);
 258
 259         threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
 260
 261         /*
 262          * Maximum threshold is 125
 263          */
 264         threshold = min(125, threshold);
 265
 266         return threshold;
 267 }
 268
 269 /*
 270  * Refresh the thresholds for each zone.
 271  */
 272 void refresh_zone_stat_thresholds(void)
 273 {
 274         struct pglist_data *pgdat;
 275         struct zone *zone;
 276         int cpu;
 277         int threshold;
 278
 279         /* Zero current pgdat thresholds */
 280         for_each_online_pgdat(pgdat) {
 281                 for_each_online_cpu(cpu) {
 282                         per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold = 0;
 283                 }
 284         }
 285
 286         for_each_populated_zone(zone) {
 287                 struct pglist_data *pgdat = zone->zone_pgdat;
 288                 unsigned long max_drift, tolerate_drift;
 289
 290                 threshold = calculate_normal_threshold(zone);
 291
 292                 for_each_online_cpu(cpu) {
 293                         int pgdat_threshold;
 294
 295                         per_cpu_ptr(zone->per_cpu_zonestats, cpu)->stat_threshold
 296                                                         = threshold;
 297
 298                         /* Base nodestat threshold on the largest populated zone. */
 299                         pgdat_threshold = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold;
 300                         per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold
 301                                 = max(threshold, pgdat_threshold);
 302                 }
 303
 304                 /*
 305                  * Only set percpu_drift_mark if there is a danger that
 306                  * NR_FREE_PAGES reports the low watermark is ok when in fact
 307                  * the min watermark could be breached by an allocation
 308                  */
 309                 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
 310                 max_drift = num_online_cpus() * threshold;
 311                 if (max_drift > tolerate_drift)
 312                         zone->percpu_drift_mark = high_wmark_pages(zone) +
 313                                         max_drift;
 314         }
 315 }
 316
 317 void set_pgdat_percpu_threshold(pg_data_t *pgdat,
 318                                 int (*calculate_pressure)(struct zone *))
 319 {
 320         struct zone *zone;
 321         int cpu;
 322         int threshold;
 323         int i;
 324
 325         for (i = 0; i < pgdat->nr_zones; i++) {
 326                 zone = &pgdat->node_zones[i];
 327                 if (!zone->percpu_drift_mark)
 328                         continue;
 329
 330                 threshold = (*calculate_pressure)(zone);
 331                 for_each_online_cpu(cpu)
 332                         per_cpu_ptr(zone->per_cpu_zonestats, cpu)->stat_threshold
 333                                                         = threshold;
 334         }
 335 }
 336
 337 /*
 338  * For use when we know that interrupts are disabled,
 339  * or when we know that preemption is disabled and that
 340  * particular counter cannot be updated from interrupt context.
 341  */
 342 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 343                            long delta)
 344 {
 345         struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
 346         s8 __percpu *p = pcp->vm_stat_diff + item;
 347         long x;
 348         long t;
 349
 350         /*
 351          * Accurate vmstat updates require a RMW. On !PREEMPT_RT kernels,
 352          * atomicity is provided by IRQs being disabled -- either explicitly
 353          * or via local_lock_irq. On PREEMPT_RT, local_lock_irq only disables
 354          * CPU migrations and preemption potentially corrupts a counter so
 355          * disable preemption.
 356          */
 357         preempt_disable_nested();
 358
 359         x = delta + __this_cpu_read(*p);
 360
 361         t = __this_cpu_read(pcp->stat_threshold);
 362
 363         if (unlikely(abs(x) > t)) {
 364                 zone_page_state_add(x, zone, item);
 365                 x = 0;
 366         }
 367         __this_cpu_write(*p, x);
 368
 369         preempt_enable_nested();
 370 }
 371 EXPORT_SYMBOL(__mod_zone_page_state);
 372
 373 void __mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
 374                                 long delta)
 375 {
 376         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 377         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 378         long x;
 379         long t;
 380
 381         if (vmstat_item_in_bytes(item)) {
 382                 /*
 383                  * Only cgroups use subpage accounting right now; at
 384                  * the global level, these items still change in
 385                  * multiples of whole pages. Store them as pages
 386                  * internally to keep the per-cpu counters compact.
 387                  */
 388                 VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
 389                 delta >>= PAGE_SHIFT;
 390         }
 391
 392         /* See __mod_node_page_state */
 393         preempt_disable_nested();
 394
 395         x = delta + __this_cpu_read(*p);
 396
 397         t = __this_cpu_read(pcp->stat_threshold);
 398
 399         if (unlikely(abs(x) > t)) {
 400                 node_page_state_add(x, pgdat, item);
 401                 x = 0;
 402         }
 403         __this_cpu_write(*p, x);
 404
 405         preempt_enable_nested();
 406 }
 407 EXPORT_SYMBOL(__mod_node_page_state);
 408
 409 /*
 410  * Optimized increment and decrement functions.
 411  *
 412  * These are only for a single page and therefore can take a struct page *
 413  * argument instead of struct zone *. This allows the inclusion of the code
 414  * generated for page_zone(page) into the optimized functions.
 415  *
 416  * No overflow check is necessary and therefore the differential can be
 417  * incremented or decremented in place which may allow the compilers to
 418  * generate better code.
 419  * The increment or decrement is known and therefore one boundary check can
 420  * be omitted.
 421  *
 422  * NOTE: These functions are very performance sensitive. Change only
 423  * with care.
 424  *
 425  * Some processors have inc/dec instructions that are atomic vs an interrupt.
 426  * However, the code must first determine the differential location in a zone
 427  * based on the processor number and then inc/dec the counter. There is no
 428  * guarantee without disabling preemption that the processor will not change
 429  * in between and therefore the atomicity vs. interrupt cannot be exploited
 430  * in a useful way here.
 431  */
 432 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
 433 {
 434         struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
 435         s8 __percpu *p = pcp->vm_stat_diff + item;
 436         s8 v, t;
 437
 438         /* See __mod_node_page_state */
 439         preempt_disable_nested();
 440
 441         v = __this_cpu_inc_return(*p);
 442         t = __this_cpu_read(pcp->stat_threshold);
 443         if (unlikely(v > t)) {
 444                 s8 overstep = t >> 1;
 445
 446                 zone_page_state_add(v + overstep, zone, item);
 447                 __this_cpu_write(*p, -overstep);
 448         }
 449
 450         preempt_enable_nested();
 451 }
 452
 453 void __inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 454 {
 455         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 456         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 457         s8 v, t;
 458
 459         VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
 460
 461         /* See __mod_node_page_state */
 462         preempt_disable_nested();
 463
 464         v = __this_cpu_inc_return(*p);
 465         t = __this_cpu_read(pcp->stat_threshold);
 466         if (unlikely(v > t)) {
 467                 s8 overstep = t >> 1;
 468
 469                 node_page_state_add(v + overstep, pgdat, item);
 470                 __this_cpu_write(*p, -overstep);
 471         }
 472
 473         preempt_enable_nested();
 474 }
 475
 476 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
 477 {
 478         __inc_zone_state(page_zone(page), item);
 479 }
 480 EXPORT_SYMBOL(__inc_zone_page_state);
 481
 482 void __inc_node_page_state(struct page *page, enum node_stat_item item)
 483 {
 484         __inc_node_state(page_pgdat(page), item);
 485 }
 486 EXPORT_SYMBOL(__inc_node_page_state);
 487
 488 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
 489 {
 490         struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
 491         s8 __percpu *p = pcp->vm_stat_diff + item;
 492         s8 v, t;
 493
 494         /* See __mod_node_page_state */
 495         preempt_disable_nested();
 496
 497         v = __this_cpu_dec_return(*p);
 498         t = __this_cpu_read(pcp->stat_threshold);
 499         if (unlikely(v < - t)) {
 500                 s8 overstep = t >> 1;
 501
 502                 zone_page_state_add(v - overstep, zone, item);
 503                 __this_cpu_write(*p, overstep);
 504         }
 505
 506         preempt_enable_nested();
 507 }
 508
 509 void __dec_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 510 {
 511         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 512         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 513         s8 v, t;
 514
 515         VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
 516
 517         /* See __mod_node_page_state */
 518         preempt_disable_nested();
 519
 520         v = __this_cpu_dec_return(*p);
 521         t = __this_cpu_read(pcp->stat_threshold);
 522         if (unlikely(v < - t)) {
 523                 s8 overstep = t >> 1;
 524
 525                 node_page_state_add(v - overstep, pgdat, item);
 526                 __this_cpu_write(*p, overstep);
 527         }
 528
 529         preempt_enable_nested();
 530 }
 531
 532 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
 533 {
 534         __dec_zone_state(page_zone(page), item);
 535 }
 536 EXPORT_SYMBOL(__dec_zone_page_state);
 537
 538 void __dec_node_page_state(struct page *page, enum node_stat_item item)
 539 {
 540         __dec_node_state(page_pgdat(page), item);
 541 }
 542 EXPORT_SYMBOL(__dec_node_page_state);
 543
 544 #ifdef CONFIG_HAVE_CMPXCHG_LOCAL
 545 /*
 546  * If we have cmpxchg_local support then we do not need to incur the overhead
 547  * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
 548  *
 549  * mod_state() modifies the zone counter state through atomic per cpu
 550  * operations.
 551  *
 552  * Overstep mode specifies how overstep should handled:
 553  *     0       No overstepping
 554  *     1       Overstepping half of threshold
 555  *     -1      Overstepping minus half of threshold
 556 */
 557 static inline void mod_zone_state(struct zone *zone,
 558        enum zone_stat_item item, long delta, int overstep_mode)
 559 {
 560         struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
 561         s8 __percpu *p = pcp->vm_stat_diff + item;
 562         long n, t, z;
 563         s8 o;
 564
 565         o = this_cpu_read(*p);
 566         do {
 567                 z = 0;  /* overflow to zone counters */
 568
 569                 /*
 570                  * The fetching of the stat_threshold is racy. We may apply
 571                  * a counter threshold to the wrong the cpu if we get
 572                  * rescheduled while executing here. However, the next
 573                  * counter update will apply the threshold again and
 574                  * therefore bring the counter under the threshold again.
 575                  *
 576                  * Most of the time the thresholds are the same anyways
 577                  * for all cpus in a zone.
 578                  */
 579                 t = this_cpu_read(pcp->stat_threshold);
 580
 581                 n = delta + (long)o;
 582
 583                 if (abs(n) > t) {
 584                         int os = overstep_mode * (t >> 1) ;
 585
 586                         /* Overflow must be added to zone counters */
 587                         z = n + os;
 588                         n = -os;
 589                 }
 590         } while (!this_cpu_try_cmpxchg(*p, &o, n));
 591
 592         if (z)
 593                 zone_page_state_add(z, zone, item);
 594 }
 595
 596 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 597                          long delta)
 598 {
 599         mod_zone_state(zone, item, delta, 0);
 600 }
 601 EXPORT_SYMBOL(mod_zone_page_state);
 602
 603 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 604 {
 605         mod_zone_state(page_zone(page), item, 1, 1);
 606 }
 607 EXPORT_SYMBOL(inc_zone_page_state);
 608
 609 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 610 {
 611         mod_zone_state(page_zone(page), item, -1, -1);
 612 }
 613 EXPORT_SYMBOL(dec_zone_page_state);
 614
 615 static inline void mod_node_state(struct pglist_data *pgdat,
 616        enum node_stat_item item, int delta, int overstep_mode)
 617 {
 618         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 619         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 620         long n, t, z;
 621         s8 o;
 622
 623         if (vmstat_item_in_bytes(item)) {
 624                 /*
 625                  * Only cgroups use subpage accounting right now; at
 626                  * the global level, these items still change in
 627                  * multiples of whole pages. Store them as pages
 628                  * internally to keep the per-cpu counters compact.
 629                  */
 630                 VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
 631                 delta >>= PAGE_SHIFT;
 632         }
 633
 634         o = this_cpu_read(*p);
 635         do {
 636                 z = 0;  /* overflow to node counters */
 637
 638                 /*
 639                  * The fetching of the stat_threshold is racy. We may apply
 640                  * a counter threshold to the wrong the cpu if we get
 641                  * rescheduled while executing here. However, the next
 642                  * counter update will apply the threshold again and
 643                  * therefore bring the counter under the threshold again.
 644                  *
 645                  * Most of the time the thresholds are the same anyways
 646                  * for all cpus in a node.
 647                  */
 648                 t = this_cpu_read(pcp->stat_threshold);
 649
 650                 n = delta + (long)o;
 651
 652                 if (abs(n) > t) {
 653                         int os = overstep_mode * (t >> 1) ;
 654
 655                         /* Overflow must be added to node counters */
 656                         z = n + os;
 657                         n = -os;
 658                 }
 659         } while (!this_cpu_try_cmpxchg(*p, &o, n));
 660
 661         if (z)
 662                 node_page_state_add(z, pgdat, item);
 663 }
 664
 665 void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
 666                                         long delta)
 667 {
 668         mod_node_state(pgdat, item, delta, 0);
 669 }
 670 EXPORT_SYMBOL(mod_node_page_state);
 671
 672 void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 673 {
 674         mod_node_state(pgdat, item, 1, 1);
 675 }
 676
 677 void inc_node_page_state(struct page *page, enum node_stat_item item)
 678 {
 679         mod_node_state(page_pgdat(page), item, 1, 1);
 680 }
 681 EXPORT_SYMBOL(inc_node_page_state);
 682
 683 void dec_node_page_state(struct page *page, enum node_stat_item item)
 684 {
 685         mod_node_state(page_pgdat(page), item, -1, -1);
 686 }
 687 EXPORT_SYMBOL(dec_node_page_state);
 688 #else
 689 /*
 690  * Use interrupt disable to serialize counter updates
 691  */
 692 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 693                          long delta)
 694 {
 695         unsigned long flags;
 696
 697         local_irq_save(flags);
 698         __mod_zone_page_state(zone, item, delta);
 699         local_irq_restore(flags);
 700 }
 701 EXPORT_SYMBOL(mod_zone_page_state);
 702
 703 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 704 {
 705         unsigned long flags;
 706         struct zone *zone;
 707
 708         zone = page_zone(page);
 709         local_irq_save(flags);
 710         __inc_zone_state(zone, item);
 711         local_irq_restore(flags);
 712 }
 713 EXPORT_SYMBOL(inc_zone_page_state);
 714
 715 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 716 {
 717         unsigned long flags;
 718
 719         local_irq_save(flags);
 720         __dec_zone_page_state(page, item);
 721         local_irq_restore(flags);
 722 }
 723 EXPORT_SYMBOL(dec_zone_page_state);
 724
 725 void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 726 {
 727         unsigned long flags;
 728
 729         local_irq_save(flags);
 730         __inc_node_state(pgdat, item);
 731         local_irq_restore(flags);
 732 }
 733 EXPORT_SYMBOL(inc_node_state);
 734
 735 void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
 736                                         long delta)
 737 {
 738         unsigned long flags;
 739
 740         local_irq_save(flags);
 741         __mod_node_page_state(pgdat, item, delta);
 742         local_irq_restore(flags);
 743 }
 744 EXPORT_SYMBOL(mod_node_page_state);
 745
 746 void inc_node_page_state(struct page *page, enum node_stat_item item)
 747 {
 748         unsigned long flags;
 749         struct pglist_data *pgdat;
 750
 751         pgdat = page_pgdat(page);
 752         local_irq_save(flags);
 753         __inc_node_state(pgdat, item);
 754         local_irq_restore(flags);
 755 }
 756 EXPORT_SYMBOL(inc_node_page_state);
 757
 758 void dec_node_page_state(struct page *page, enum node_stat_item item)
 759 {
 760         unsigned long flags;
 761
 762         local_irq_save(flags);
 763         __dec_node_page_state(page, item);
 764         local_irq_restore(flags);
 765 }
 766 EXPORT_SYMBOL(dec_node_page_state);
 767 #endif
 768
 769 /*
 770  * Fold a differential into the global counters.
 771  * Returns the number of counters updated.
 772  */
 773 static int fold_diff(int *zone_diff, int *node_diff)
 774 {
 775         int i;
 776         int changes = 0;
 777
 778         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 779                 if (zone_diff[i]) {
 780                         atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
 781                         changes++;
 782         }
 783
 784         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 785                 if (node_diff[i]) {
 786                         atomic_long_add(node_diff[i], &vm_node_stat[i]);
 787                         changes++;
 788         }
 789         return changes;
 790 }
 791
 792 /*
 793  * Update the zone counters for the current cpu.
 794  *
 795  * Note that refresh_cpu_vm_stats strives to only access
 796  * node local memory. The per cpu pagesets on remote zones are placed
 797  * in the memory local to the processor using that pageset. So the
 798  * loop over all zones will access a series of cachelines local to
 799  * the processor.
 800  *
 801  * The call to zone_page_state_add updates the cachelines with the
 802  * statistics in the remote zone struct as well as the global cachelines
 803  * with the global counters. These could cause remote node cache line
 804  * bouncing and will have to be only done when necessary.
 805  *
 806  * The function returns the number of global counters updated.
 807  */
 808 static int refresh_cpu_vm_stats(bool do_pagesets)
 809 {
 810         struct pglist_data *pgdat;
 811         struct zone *zone;
 812         int i;
 813         int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 814         int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
 815         int changes = 0;
 816
 817         for_each_populated_zone(zone) {
 818                 struct per_cpu_zonestat __percpu *pzstats = zone->per_cpu_zonestats;
 819                 struct per_cpu_pages __percpu *pcp = zone->per_cpu_pageset;
 820
 821                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
 822                         int v;
 823
 824                         v = this_cpu_xchg(pzstats->vm_stat_diff[i], 0);
 825                         if (v) {
 826
 827                                 atomic_long_add(v, &zone->vm_stat[i]);
 828                                 global_zone_diff[i] += v;
 829 #ifdef CONFIG_NUMA
 830                                 /* 3 seconds idle till flush */
 831                                 __this_cpu_write(pcp->expire, 3);
 832 #endif
 833                         }
 834                 }
 835
 836                 if (do_pagesets) {
 837                         cond_resched();
 838
 839                         changes += decay_pcp_high(zone, this_cpu_ptr(pcp));
 840 #ifdef CONFIG_NUMA
 841                         /*
 842                          * Deal with draining the remote pageset of this
 843                          * processor
 844                          *
 845                          * Check if there are pages remaining in this pageset
 846                          * if not then there is nothing to expire.
 847                          */
 848                         if (!__this_cpu_read(pcp->expire) ||
 849                                !__this_cpu_read(pcp->count))
 850                                 continue;
 851
 852                         /*
 853                          * We never drain zones local to this processor.
 854                          */
 855                         if (zone_to_nid(zone) == numa_node_id()) {
 856                                 __this_cpu_write(pcp->expire, 0);
 857                                 continue;
 858                         }
 859
 860                         if (__this_cpu_dec_return(pcp->expire)) {
 861                                 changes++;
 862                                 continue;
 863                         }
 864
 865                         if (__this_cpu_read(pcp->count)) {
 866                                 drain_zone_pages(zone, this_cpu_ptr(pcp));
 867                                 changes++;
 868                         }
 869 #endif
 870                 }
 871         }
 872
 873         for_each_online_pgdat(pgdat) {
 874                 struct per_cpu_nodestat __percpu *p = pgdat->per_cpu_nodestats;
 875
 876                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
 877                         int v;
 878
 879                         v = this_cpu_xchg(p->vm_node_stat_diff[i], 0);
 880                         if (v) {
 881                                 atomic_long_add(v, &pgdat->vm_stat[i]);
 882                                 global_node_diff[i] += v;
 883                         }
 884                 }
 885         }
 886
 887         changes += fold_diff(global_zone_diff, global_node_diff);
 888         return changes;
 889 }
 890
 891 /*
 892  * Fold the data for an offline cpu into the global array.
 893  * There cannot be any access by the offline cpu and therefore
 894  * synchronization is simplified.
 895  */
 896 void cpu_vm_stats_fold(int cpu)
 897 {
 898         struct pglist_data *pgdat;
 899         struct zone *zone;
 900         int i;
 901         int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 902         int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
 903
 904         for_each_populated_zone(zone) {
 905                 struct per_cpu_zonestat *pzstats;
 906
 907                 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
 908
 909                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
 910                         if (pzstats->vm_stat_diff[i]) {
 911                                 int v;
 912
 913                                 v = pzstats->vm_stat_diff[i];
 914                                 pzstats->vm_stat_diff[i] = 0;
 915                                 atomic_long_add(v, &zone->vm_stat[i]);
 916                                 global_zone_diff[i] += v;
 917                         }
 918                 }
 919 #ifdef CONFIG_NUMA
 920                 for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++) {
 921                         if (pzstats->vm_numa_event[i]) {
 922                                 unsigned long v;
 923
 924                                 v = pzstats->vm_numa_event[i];
 925                                 pzstats->vm_numa_event[i] = 0;
 926                                 zone_numa_event_add(v, zone, i);
 927                         }
 928                 }
 929 #endif
 930         }
 931
 932         for_each_online_pgdat(pgdat) {
 933                 struct per_cpu_nodestat *p;
 934
 935                 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
 936
 937                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 938                         if (p->vm_node_stat_diff[i]) {
 939                                 int v;
 940
 941                                 v = p->vm_node_stat_diff[i];
 942                                 p->vm_node_stat_diff[i] = 0;
 943                                 atomic_long_add(v, &pgdat->vm_stat[i]);
 944                                 global_node_diff[i] += v;
 945                         }
 946         }
 947
 948         fold_diff(global_zone_diff, global_node_diff);
 949 }
 950
 951 /*
 952  * this is only called if !populated_zone(zone), which implies no other users of
 953  * pset->vm_stat_diff[] exist.
 954  */
 955 void drain_zonestat(struct zone *zone, struct per_cpu_zonestat *pzstats)
 956 {
 957         unsigned long v;
 958         int i;
 959
 960         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
 961                 if (pzstats->vm_stat_diff[i]) {
 962                         v = pzstats->vm_stat_diff[i];
 963                         pzstats->vm_stat_diff[i] = 0;
 964                         zone_page_state_add(v, zone, i);
 965                 }
 966         }
 967
 968 #ifdef CONFIG_NUMA
 969         for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++) {
 970                 if (pzstats->vm_numa_event[i]) {
 971                         v = pzstats->vm_numa_event[i];
 972                         pzstats->vm_numa_event[i] = 0;
 973                         zone_numa_event_add(v, zone, i);
 974                 }
 975         }
 976 #endif
 977 }
 978 #endif
 979
 980 #ifdef CONFIG_NUMA
 981 /*
 982  * Determine the per node value of a stat item. This function
 983  * is called frequently in a NUMA machine, so try to be as
 984  * frugal as possible.
 985  */
 986 unsigned long sum_zone_node_page_state(int node,
 987                                  enum zone_stat_item item)
 988 {
 989         struct zone *zones = NODE_DATA(node)->node_zones;
 990         int i;
 991         unsigned long count = 0;
 992
 993         for (i = 0; i < MAX_NR_ZONES; i++)
 994                 count += zone_page_state(zones + i, item);
 995
 996         return count;
 997 }
 998
 999 /* Determine the per node value of a numa stat item. */
1000 unsigned long sum_zone_numa_event_state(int node,
1001                                  enum numa_stat_item item)
1002 {
1003         struct zone *zones = NODE_DATA(node)->node_zones;
1004         unsigned long count = 0;
1005         int i;
1006
1007         for (i = 0; i < MAX_NR_ZONES; i++)
1008                 count += zone_numa_event_state(zones + i, item);
1009
1010         return count;
1011 }
1012
1013 /*
1014  * Determine the per node value of a stat item.
1015  */
1016 unsigned long node_page_state_pages(struct pglist_data *pgdat,
1017                                     enum node_stat_item item)
1018 {
1019         long x = atomic_long_read(&pgdat->vm_stat[item]);
1020 #ifdef CONFIG_SMP
1021         if (x < 0)
1022                 x = 0;
1023 #endif
1024         return x;
1025 }
1026
1027 unsigned long node_page_state(struct pglist_data *pgdat,
1028                               enum node_stat_item item)
1029 {
1030         VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
1031
1032         return node_page_state_pages(pgdat, item);
1033 }
1034 #endif
1035
1036 /*
1037  * Count number of pages "struct page" and "struct page_ext" consume.
1038  * nr_memmap_boot_pages: # of pages allocated by boot allocator
1039  * nr_memmap_pages: # of pages that were allocated by buddy allocator
1040  */
1041 static atomic_long_t nr_memmap_boot_pages = ATOMIC_LONG_INIT(0);
1042 static atomic_long_t nr_memmap_pages = ATOMIC_LONG_INIT(0);
1043
1044 void memmap_boot_pages_add(long delta)
1045 {
1046         atomic_long_add(delta, &nr_memmap_boot_pages);
1047 }
1048
1049 void memmap_pages_add(long delta)
1050 {
1051         atomic_long_add(delta, &nr_memmap_pages);
1052 }
1053
1054 #ifdef CONFIG_COMPACTION
1055
1056 struct contig_page_info {
1057         unsigned long free_pages;
1058         unsigned long free_blocks_total;
1059         unsigned long free_blocks_suitable;
1060 };
1061
1062 /*
1063  * Calculate the number of free pages in a zone, how many contiguous
1064  * pages are free and how many are large enough to satisfy an allocation of
1065  * the target size. Note that this function makes no attempt to estimate
1066  * how many suitable free blocks there *might* be if MOVABLE pages were
1067  * migrated. Calculating that is possible, but expensive and can be
1068  * figured out from userspace
1069  */
1070 static void fill_contig_page_info(struct zone *zone,
1071                                 unsigned int suitable_order,
1072                                 struct contig_page_info *info)
1073 {
1074         unsigned int order;
1075
1076         info->free_pages = 0;
1077         info->free_blocks_total = 0;
1078         info->free_blocks_suitable = 0;
1079
1080         for (order = 0; order < NR_PAGE_ORDERS; order++) {
1081                 unsigned long blocks;
1082
1083                 /*
1084                  * Count number of free blocks.
1085                  *
1086                  * Access to nr_free is lockless as nr_free is used only for
1087                  * diagnostic purposes. Use data_race to avoid KCSAN warning.
1088                  */
1089                 blocks = data_race(zone->free_area[order].nr_free);
1090                 info->free_blocks_total += blocks;
1091
1092                 /* Count free base pages */
1093                 info->free_pages += blocks << order;
1094
1095                 /* Count the suitable free blocks */
1096                 if (order >= suitable_order)
1097                         info->free_blocks_suitable += blocks <<
1098                                                 (order - suitable_order);
1099         }
1100 }
1101
1102 /*
1103  * A fragmentation index only makes sense if an allocation of a requested
1104  * size would fail. If that is true, the fragmentation index indicates
1105  * whether external fragmentation or a lack of memory was the problem.
1106  * The value can be used to determine if page reclaim or compaction
1107  * should be used
1108  */
1109 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
1110 {
1111         unsigned long requested = 1UL << order;
1112
1113         if (WARN_ON_ONCE(order > MAX_PAGE_ORDER))
1114                 return 0;
1115
1116         if (!info->free_blocks_total)
1117                 return 0;
1118
1119         /* Fragmentation index only makes sense when a request would fail */
1120         if (info->free_blocks_suitable)
1121                 return -1000;
1122
1123         /*
1124          * Index is between 0 and 1 so return within 3 decimal places
1125          *
1126          * 0 => allocation would fail due to lack of memory
1127          * 1 => allocation would fail due to fragmentation
1128          */
1129         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
1130 }
1131
1132 /*
1133  * Calculates external fragmentation within a zone wrt the given order.
1134  * It is defined as the percentage of pages found in blocks of size
1135  * less than 1 << order. It returns values in range [0, 100].
1136  */
1137 unsigned int extfrag_for_order(struct zone *zone, unsigned int order)
1138 {
1139         struct contig_page_info info;
1140
1141         fill_contig_page_info(zone, order, &info);
1142         if (info.free_pages == 0)
1143                 return 0;
1144
1145         return div_u64((info.free_pages -
1146                         (info.free_blocks_suitable << order)) * 100,
1147                         info.free_pages);
1148 }
1149
1150 /* Same as __fragmentation index but allocs contig_page_info on stack */
1151 int fragmentation_index(struct zone *zone, unsigned int order)
1152 {
1153         struct contig_page_info info;
1154
1155         fill_contig_page_info(zone, order, &info);
1156         return __fragmentation_index(order, &info);
1157 }
1158 #endif
1159
1160 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || \
1161     defined(CONFIG_NUMA) || defined(CONFIG_MEMCG)
1162 #ifdef CONFIG_ZONE_DMA
1163 #define TEXT_FOR_DMA(xx) xx "_dma",
1164 #else
1165 #define TEXT_FOR_DMA(xx)
1166 #endif
1167
1168 #ifdef CONFIG_ZONE_DMA32
1169 #define TEXT_FOR_DMA32(xx) xx "_dma32",
1170 #else
1171 #define TEXT_FOR_DMA32(xx)
1172 #endif
1173
1174 #ifdef CONFIG_HIGHMEM
1175 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
1176 #else
1177 #define TEXT_FOR_HIGHMEM(xx)
1178 #endif
1179
1180 #ifdef CONFIG_ZONE_DEVICE
1181 #define TEXT_FOR_DEVICE(xx) xx "_device",
1182 #else
1183 #define TEXT_FOR_DEVICE(xx)
1184 #endif
1185
1186 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1187                                         TEXT_FOR_HIGHMEM(xx) xx "_movable", \
1188                                         TEXT_FOR_DEVICE(xx)
1189
1190 const char * const vmstat_text[] = {
1191         /* enum zone_stat_item counters */
1192         "nr_free_pages",
1193         "nr_zone_inactive_anon",
1194         "nr_zone_active_anon",
1195         "nr_zone_inactive_file",
1196         "nr_zone_active_file",
1197         "nr_zone_unevictable",
1198         "nr_zone_write_pending",
1199         "nr_mlock",
1200         "nr_bounce",
1201 #if IS_ENABLED(CONFIG_ZSMALLOC)
1202         "nr_zspages",
1203 #endif
1204         "nr_free_cma",
1205 #ifdef CONFIG_UNACCEPTED_MEMORY
1206         "nr_unaccepted",
1207 #endif
1208
1209         /* enum numa_stat_item counters */
1210 #ifdef CONFIG_NUMA
1211         "numa_hit",
1212         "numa_miss",
1213         "numa_foreign",
1214         "numa_interleave",
1215         "numa_local",
1216         "numa_other",
1217 #endif
1218
1219         /* enum node_stat_item counters */
1220         "nr_inactive_anon",
1221         "nr_active_anon",
1222         "nr_inactive_file",
1223         "nr_active_file",
1224         "nr_unevictable",
1225         "nr_slab_reclaimable",
1226         "nr_slab_unreclaimable",
1227         "nr_isolated_anon",
1228         "nr_isolated_file",
1229         "workingset_nodes",
1230         "workingset_refault_anon",
1231         "workingset_refault_file",
1232         "workingset_activate_anon",
1233         "workingset_activate_file",
1234         "workingset_restore_anon",
1235         "workingset_restore_file",
1236         "workingset_nodereclaim",
1237         "nr_anon_pages",
1238         "nr_mapped",
1239         "nr_file_pages",
1240         "nr_dirty",
1241         "nr_writeback",
1242         "nr_writeback_temp",
1243         "nr_shmem",
1244         "nr_shmem_hugepages",
1245         "nr_shmem_pmdmapped",
1246         "nr_file_hugepages",
1247         "nr_file_pmdmapped",
1248         "nr_anon_transparent_hugepages",
1249         "nr_vmscan_write",
1250         "nr_vmscan_immediate_reclaim",
1251         "nr_dirtied",
1252         "nr_written",
1253         "nr_throttled_written",
1254         "nr_kernel_misc_reclaimable",
1255         "nr_foll_pin_acquired",
1256         "nr_foll_pin_released",
1257         "nr_kernel_stack",
1258 #if IS_ENABLED(CONFIG_SHADOW_CALL_STACK)
1259         "nr_shadow_call_stack",
1260 #endif
1261         "nr_page_table_pages",
1262         "nr_sec_page_table_pages",
1263 #ifdef CONFIG_IOMMU_SUPPORT
1264         "nr_iommu_pages",
1265 #endif
1266 #ifdef CONFIG_SWAP
1267         "nr_swapcached",
1268 #endif
1269 #ifdef CONFIG_NUMA_BALANCING
1270         "pgpromote_success",
1271         "pgpromote_candidate",
1272 #endif
1273         "pgdemote_kswapd",
1274         "pgdemote_direct",
1275         "pgdemote_khugepaged",
1276 #ifdef CONFIG_HUGETLB_PAGE
1277         "nr_hugetlb",
1278 #endif
1279         /* system-wide enum vm_stat_item counters */
1280         "nr_dirty_threshold",
1281         "nr_dirty_background_threshold",
1282         "nr_memmap_pages",
1283         "nr_memmap_boot_pages",
1284
1285 #if defined(CONFIG_VM_EVENT_COUNTERS) || defined(CONFIG_MEMCG)
1286         /* enum vm_event_item counters */
1287         "pgpgin",
1288         "pgpgout",
1289         "pswpin",
1290         "pswpout",
1291
1292         TEXTS_FOR_ZONES("pgalloc")
1293         TEXTS_FOR_ZONES("allocstall")
1294         TEXTS_FOR_ZONES("pgskip")
1295
1296         "pgfree",
1297         "pgactivate",
1298         "pgdeactivate",
1299         "pglazyfree",
1300
1301         "pgfault",
1302         "pgmajfault",
1303         "pglazyfreed",
1304
1305         "pgrefill",
1306         "pgreuse",
1307         "pgsteal_kswapd",
1308         "pgsteal_direct",
1309         "pgsteal_khugepaged",
1310         "pgscan_kswapd",
1311         "pgscan_direct",
1312         "pgscan_khugepaged",
1313         "pgscan_direct_throttle",
1314         "pgscan_anon",
1315         "pgscan_file",
1316         "pgsteal_anon",
1317         "pgsteal_file",
1318
1319 #ifdef CONFIG_NUMA
1320         "zone_reclaim_success",
1321         "zone_reclaim_failed",
1322 #endif
1323         "pginodesteal",
1324         "slabs_scanned",
1325         "kswapd_inodesteal",
1326         "kswapd_low_wmark_hit_quickly",
1327         "kswapd_high_wmark_hit_quickly",
1328         "pageoutrun",
1329
1330         "pgrotated",
1331
1332         "drop_pagecache",
1333         "drop_slab",
1334         "oom_kill",
1335
1336 #ifdef CONFIG_NUMA_BALANCING
1337         "numa_pte_updates",
1338         "numa_huge_pte_updates",
1339         "numa_hint_faults",
1340         "numa_hint_faults_local",
1341         "numa_pages_migrated",
1342 #endif
1343 #ifdef CONFIG_MIGRATION
1344         "pgmigrate_success",
1345         "pgmigrate_fail",
1346         "thp_migration_success",
1347         "thp_migration_fail",
1348         "thp_migration_split",
1349 #endif
1350 #ifdef CONFIG_COMPACTION
1351         "compact_migrate_scanned",
1352         "compact_free_scanned",
1353         "compact_isolated",
1354         "compact_stall",
1355         "compact_fail",
1356         "compact_success",
1357         "compact_daemon_wake",
1358         "compact_daemon_migrate_scanned",
1359         "compact_daemon_free_scanned",
1360 #endif
1361
1362 #ifdef CONFIG_HUGETLB_PAGE
1363         "htlb_buddy_alloc_success",
1364         "htlb_buddy_alloc_fail",
1365 #endif
1366 #ifdef CONFIG_CMA
1367         "cma_alloc_success",
1368         "cma_alloc_fail",
1369 #endif
1370         "unevictable_pgs_culled",
1371         "unevictable_pgs_scanned",
1372         "unevictable_pgs_rescued",
1373         "unevictable_pgs_mlocked",
1374         "unevictable_pgs_munlocked",
1375         "unevictable_pgs_cleared",
1376         "unevictable_pgs_stranded",
1377
1378 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1379         "thp_fault_alloc",
1380         "thp_fault_fallback",
1381         "thp_fault_fallback_charge",
1382         "thp_collapse_alloc",
1383         "thp_collapse_alloc_failed",
1384         "thp_file_alloc",
1385         "thp_file_fallback",
1386         "thp_file_fallback_charge",
1387         "thp_file_mapped",
1388         "thp_split_page",
1389         "thp_split_page_failed",
1390         "thp_deferred_split_page",
1391         "thp_underused_split_page",
1392         "thp_split_pmd",
1393         "thp_scan_exceed_none_pte",
1394         "thp_scan_exceed_swap_pte",
1395         "thp_scan_exceed_share_pte",
1396 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1397         "thp_split_pud",
1398 #endif
1399         "thp_zero_page_alloc",
1400         "thp_zero_page_alloc_failed",
1401         "thp_swpout",
1402         "thp_swpout_fallback",
1403 #endif
1404 #ifdef CONFIG_MEMORY_BALLOON
1405         "balloon_inflate",
1406         "balloon_deflate",
1407 #ifdef CONFIG_BALLOON_COMPACTION
1408         "balloon_migrate",
1409 #endif
1410 #endif /* CONFIG_MEMORY_BALLOON */
1411 #ifdef CONFIG_DEBUG_TLBFLUSH
1412         "nr_tlb_remote_flush",
1413         "nr_tlb_remote_flush_received",
1414         "nr_tlb_local_flush_all",
1415         "nr_tlb_local_flush_one",
1416 #endif /* CONFIG_DEBUG_TLBFLUSH */
1417
1418 #ifdef CONFIG_SWAP
1419         "swap_ra",
1420         "swap_ra_hit",
1421         "swpin_zero",
1422         "swpout_zero",
1423 #ifdef CONFIG_KSM
1424         "ksm_swpin_copy",
1425 #endif
1426 #endif
1427 #ifdef CONFIG_KSM
1428         "cow_ksm",
1429 #endif
1430 #ifdef CONFIG_ZSWAP
1431         "zswpin",
1432         "zswpout",
1433         "zswpwb",
1434 #endif
1435 #ifdef CONFIG_X86
1436         "direct_map_level2_splits",
1437         "direct_map_level3_splits",
1438 #endif
1439 #ifdef CONFIG_PER_VMA_LOCK_STATS
1440         "vma_lock_success",
1441         "vma_lock_abort",
1442         "vma_lock_retry",
1443         "vma_lock_miss",
1444 #endif
1445 #ifdef CONFIG_DEBUG_STACK_USAGE
1446         "kstack_1k",
1447 #if THREAD_SIZE > 1024
1448         "kstack_2k",
1449 #endif
1450 #if THREAD_SIZE > 2048
1451         "kstack_4k",
1452 #endif
1453 #if THREAD_SIZE > 4096
1454         "kstack_8k",
1455 #endif
1456 #if THREAD_SIZE > 8192
1457         "kstack_16k",
1458 #endif
1459 #if THREAD_SIZE > 16384
1460         "kstack_32k",
1461 #endif
1462 #if THREAD_SIZE > 32768
1463         "kstack_64k",
1464 #endif
1465 #if THREAD_SIZE > 65536
1466         "kstack_rest",
1467 #endif
1468 #endif
1469 #endif /* CONFIG_VM_EVENT_COUNTERS || CONFIG_MEMCG */
1470 };
1471 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA || CONFIG_MEMCG */
1472
1473 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1474      defined(CONFIG_PROC_FS)
1475 static void *frag_start(struct seq_file *m, loff_t *pos)
1476 {
1477         pg_data_t *pgdat;
1478         loff_t node = *pos;
1479
1480         for (pgdat = first_online_pgdat();
1481              pgdat && node;
1482              pgdat = next_online_pgdat(pgdat))
1483                 --node;
1484
1485         return pgdat;
1486 }
1487
1488 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1489 {
1490         pg_data_t *pgdat = (pg_data_t *)arg;
1491
1492         (*pos)++;
1493         return next_online_pgdat(pgdat);
1494 }
1495
1496 static void frag_stop(struct seq_file *m, void *arg)
1497 {
1498 }
1499
1500 /*
1501  * Walk zones in a node and print using a callback.
1502  * If @assert_populated is true, only use callback for zones that are populated.
1503  */
1504 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1505                 bool assert_populated, bool nolock,
1506                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1507 {
1508         struct zone *zone;
1509         struct zone *node_zones = pgdat->node_zones;
1510         unsigned long flags;
1511
1512         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1513                 if (assert_populated && !populated_zone(zone))
1514                         continue;
1515
1516                 if (!nolock)
1517                         spin_lock_irqsave(&zone->lock, flags);
1518                 print(m, pgdat, zone);
1519                 if (!nolock)
1520                         spin_unlock_irqrestore(&zone->lock, flags);
1521         }
1522 }
1523 #endif
1524
1525 #ifdef CONFIG_PROC_FS
1526 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1527                                                 struct zone *zone)
1528 {
1529         int order;
1530
1531         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1532         for (order = 0; order < NR_PAGE_ORDERS; ++order)
1533                 /*
1534                  * Access to nr_free is lockless as nr_free is used only for
1535                  * printing purposes. Use data_race to avoid KCSAN warning.
1536                  */
1537                 seq_printf(m, "%6lu ", data_race(zone->free_area[order].nr_free));
1538         seq_putc(m, '\n');
1539 }
1540
1541 /*
1542  * This walks the free areas for each zone.
1543  */
1544 static int frag_show(struct seq_file *m, void *arg)
1545 {
1546         pg_data_t *pgdat = (pg_data_t *)arg;
1547         walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1548         return 0;
1549 }
1550
1551 static void pagetypeinfo_showfree_print(struct seq_file *m,
1552                                         pg_data_t *pgdat, struct zone *zone)
1553 {
1554         int order, mtype;
1555
1556         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1557                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
1558                                         pgdat->node_id,
1559                                         zone->name,
1560                                         migratetype_names[mtype]);
1561                 for (order = 0; order < NR_PAGE_ORDERS; ++order) {
1562                         unsigned long freecount = 0;
1563                         struct free_area *area;
1564                         struct list_head *curr;
1565                         bool overflow = false;
1566
1567                         area = &(zone->free_area[order]);
1568
1569                         list_for_each(curr, &area->free_list[mtype]) {
1570                                 /*
1571                                  * Cap the free_list iteration because it might
1572                                  * be really large and we are under a spinlock
1573                                  * so a long time spent here could trigger a
1574                                  * hard lockup detector. Anyway this is a
1575                                  * debugging tool so knowing there is a handful
1576                                  * of pages of this order should be more than
1577                                  * sufficient.
1578                                  */
1579                                 if (++freecount >= 100000) {
1580                                         overflow = true;
1581                                         break;
1582                                 }
1583                         }
1584                         seq_printf(m, "%s%6lu ", overflow ? ">" : "", freecount);
1585                         spin_unlock_irq(&zone->lock);
1586                         cond_resched();
1587                         spin_lock_irq(&zone->lock);
1588                 }
1589                 seq_putc(m, '\n');
1590         }
1591 }
1592
1593 /* Print out the free pages at each order for each migatetype */
1594 static void pagetypeinfo_showfree(struct seq_file *m, void *arg)
1595 {
1596         int order;
1597         pg_data_t *pgdat = (pg_data_t *)arg;
1598
1599         /* Print header */
1600         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1601         for (order = 0; order < NR_PAGE_ORDERS; ++order)
1602                 seq_printf(m, "%6d ", order);
1603         seq_putc(m, '\n');
1604
1605         walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1606 }
1607
1608 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1609                                         pg_data_t *pgdat, struct zone *zone)
1610 {
1611         int mtype;
1612         unsigned long pfn;
1613         unsigned long start_pfn = zone->zone_start_pfn;
1614         unsigned long end_pfn = zone_end_pfn(zone);
1615         unsigned long count[MIGRATE_TYPES] = { 0, };
1616
1617         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1618                 struct page *page;
1619
1620                 page = pfn_to_online_page(pfn);
1621                 if (!page)
1622                         continue;
1623
1624                 if (page_zone(page) != zone)
1625                         continue;
1626
1627                 mtype = get_pageblock_migratetype(page);
1628
1629                 if (mtype < MIGRATE_TYPES)
1630                         count[mtype]++;
1631         }
1632
1633         /* Print counts */
1634         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1635         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1636                 seq_printf(m, "%12lu ", count[mtype]);
1637         seq_putc(m, '\n');
1638 }
1639
1640 /* Print out the number of pageblocks for each migratetype */
1641 static void pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1642 {
1643         int mtype;
1644         pg_data_t *pgdat = (pg_data_t *)arg;
1645
1646         seq_printf(m, "\n%-23s", "Number of blocks type ");
1647         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1648                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1649         seq_putc(m, '\n');
1650         walk_zones_in_node(m, pgdat, true, false,
1651                 pagetypeinfo_showblockcount_print);
1652 }
1653
1654 /*
1655  * Print out the number of pageblocks for each migratetype that contain pages
1656  * of other types. This gives an indication of how well fallbacks are being
1657  * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1658  * to determine what is going on
1659  */
1660 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1661 {
1662 #ifdef CONFIG_PAGE_OWNER
1663         int mtype;
1664
1665         if (!static_branch_unlikely(&page_owner_inited))
1666                 return;
1667
1668         drain_all_pages(NULL);
1669
1670         seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1671         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1672                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1673         seq_putc(m, '\n');
1674
1675         walk_zones_in_node(m, pgdat, true, true,
1676                 pagetypeinfo_showmixedcount_print);
1677 #endif /* CONFIG_PAGE_OWNER */
1678 }
1679
1680 /*
1681  * This prints out statistics in relation to grouping pages by mobility.
1682  * It is expensive to collect so do not constantly read the file.
1683  */
1684 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1685 {
1686         pg_data_t *pgdat = (pg_data_t *)arg;
1687
1688         /* check memoryless node */
1689         if (!node_state(pgdat->node_id, N_MEMORY))
1690                 return 0;
1691
1692         seq_printf(m, "Page block order: %d\n", pageblock_order);
1693         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
1694         seq_putc(m, '\n');
1695         pagetypeinfo_showfree(m, pgdat);
1696         pagetypeinfo_showblockcount(m, pgdat);
1697         pagetypeinfo_showmixedcount(m, pgdat);
1698
1699         return 0;
1700 }
1701
1702 static const struct seq_operations fragmentation_op = {
1703         .start  = frag_start,
1704         .next   = frag_next,
1705         .stop   = frag_stop,
1706         .show   = frag_show,
1707 };
1708
1709 static const struct seq_operations pagetypeinfo_op = {
1710         .start  = frag_start,
1711         .next   = frag_next,
1712         .stop   = frag_stop,
1713         .show   = pagetypeinfo_show,
1714 };
1715
1716 static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1717 {
1718         int zid;
1719
1720         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1721                 struct zone *compare = &pgdat->node_zones[zid];
1722
1723                 if (populated_zone(compare))
1724                         return zone == compare;
1725         }
1726
1727         return false;
1728 }
1729
1730 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1731                                                         struct zone *zone)
1732 {
1733         int i;
1734         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1735         if (is_zone_first_populated(pgdat, zone)) {
1736                 seq_printf(m, "\n  per-node stats");
1737                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1738                         unsigned long pages = node_page_state_pages(pgdat, i);
1739
1740                         if (vmstat_item_print_in_thp(i))
1741                                 pages /= HPAGE_PMD_NR;
1742                         seq_printf(m, "\n      %-12s %lu", node_stat_name(i),
1743                                    pages);
1744                 }
1745         }
1746         seq_printf(m,
1747                    "\n  pages free     %lu"
1748                    "\n        boost    %lu"
1749                    "\n        min      %lu"
1750                    "\n        low      %lu"
1751                    "\n        high     %lu"
1752                    "\n        promo    %lu"
1753                    "\n        spanned  %lu"
1754                    "\n        present  %lu"
1755                    "\n        managed  %lu"
1756                    "\n        cma      %lu",
1757                    zone_page_state(zone, NR_FREE_PAGES),
1758                    zone->watermark_boost,
1759                    min_wmark_pages(zone),
1760                    low_wmark_pages(zone),
1761                    high_wmark_pages(zone),
1762                    promo_wmark_pages(zone),
1763                    zone->spanned_pages,
1764                    zone->present_pages,
1765                    zone_managed_pages(zone),
1766                    zone_cma_pages(zone));
1767
1768         seq_printf(m,
1769                    "\n        protection: (%ld",
1770                    zone->lowmem_reserve[0]);
1771         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1772                 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1773         seq_putc(m, ')');
1774
1775         /* If unpopulated, no other information is useful */
1776         if (!populated_zone(zone)) {
1777                 seq_putc(m, '\n');
1778                 return;
1779         }
1780
1781         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1782                 seq_printf(m, "\n      %-12s %lu", zone_stat_name(i),
1783                            zone_page_state(zone, i));
1784
1785 #ifdef CONFIG_NUMA
1786         fold_vm_zone_numa_events(zone);
1787         for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++)
1788                 seq_printf(m, "\n      %-12s %lu", numa_stat_name(i),
1789                            zone_numa_event_state(zone, i));
1790 #endif
1791
1792         seq_printf(m, "\n  pagesets");
1793         for_each_online_cpu(i) {
1794                 struct per_cpu_pages *pcp;
1795                 struct per_cpu_zonestat __maybe_unused *pzstats;
1796
1797                 pcp = per_cpu_ptr(zone->per_cpu_pageset, i);
1798                 seq_printf(m,
1799                            "\n    cpu: %i"
1800                            "\n              count:    %i"
1801                            "\n              high:     %i"
1802                            "\n              batch:    %i"
1803                            "\n              high_min: %i"
1804                            "\n              high_max: %i",
1805                            i,
1806                            pcp->count,
1807                            pcp->high,
1808                            pcp->batch,
1809                            pcp->high_min,
1810                            pcp->high_max);
1811 #ifdef CONFIG_SMP
1812                 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, i);
1813                 seq_printf(m, "\n  vm stats threshold: %d",
1814                                 pzstats->stat_threshold);
1815 #endif
1816         }
1817         seq_printf(m,
1818                    "\n  node_unreclaimable:  %u"
1819                    "\n  start_pfn:           %lu",
1820                    pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1821                    zone->zone_start_pfn);
1822         seq_putc(m, '\n');
1823 }
1824
1825 /*
1826  * Output information about zones in @pgdat.  All zones are printed regardless
1827  * of whether they are populated or not: lowmem_reserve_ratio operates on the
1828  * set of all zones and userspace would not be aware of such zones if they are
1829  * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1830  */
1831 static int zoneinfo_show(struct seq_file *m, void *arg)
1832 {
1833         pg_data_t *pgdat = (pg_data_t *)arg;
1834         walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1835         return 0;
1836 }
1837
1838 static const struct seq_operations zoneinfo_op = {
1839         .start  = frag_start, /* iterate over all zones. The same as in
1840                                * fragmentation. */
1841         .next   = frag_next,
1842         .stop   = frag_stop,
1843         .show   = zoneinfo_show,
1844 };
1845
1846 #define NR_VMSTAT_ITEMS (NR_VM_ZONE_STAT_ITEMS + \
1847                          NR_VM_NUMA_EVENT_ITEMS + \
1848                          NR_VM_NODE_STAT_ITEMS + \
1849                          NR_VM_STAT_ITEMS + \
1850                          (IS_ENABLED(CONFIG_VM_EVENT_COUNTERS) ? \
1851                           NR_VM_EVENT_ITEMS : 0))
1852
1853 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1854 {
1855         unsigned long *v;
1856         int i;
1857
1858         if (*pos >= NR_VMSTAT_ITEMS)
1859                 return NULL;
1860
1861         BUILD_BUG_ON(ARRAY_SIZE(vmstat_text) < NR_VMSTAT_ITEMS);
1862         fold_vm_numa_events();
1863         v = kmalloc_array(NR_VMSTAT_ITEMS, sizeof(unsigned long), GFP_KERNEL);
1864         m->private = v;
1865         if (!v)
1866                 return ERR_PTR(-ENOMEM);
1867         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1868                 v[i] = global_zone_page_state(i);
1869         v += NR_VM_ZONE_STAT_ITEMS;
1870
1871 #ifdef CONFIG_NUMA
1872         for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++)
1873                 v[i] = global_numa_event_state(i);
1874         v += NR_VM_NUMA_EVENT_ITEMS;
1875 #endif
1876
1877         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1878                 v[i] = global_node_page_state_pages(i);
1879                 if (vmstat_item_print_in_thp(i))
1880                         v[i] /= HPAGE_PMD_NR;
1881         }
1882         v += NR_VM_NODE_STAT_ITEMS;
1883
1884         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1885                             v + NR_DIRTY_THRESHOLD);
1886         v[NR_MEMMAP_PAGES] = atomic_long_read(&nr_memmap_pages);
1887         v[NR_MEMMAP_BOOT_PAGES] = atomic_long_read(&nr_memmap_boot_pages);
1888         v += NR_VM_STAT_ITEMS;
1889
1890 #ifdef CONFIG_VM_EVENT_COUNTERS
1891         all_vm_events(v);
1892         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1893         v[PGPGOUT] /= 2;
1894 #endif
1895         return (unsigned long *)m->private + *pos;
1896 }
1897
1898 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1899 {
1900         (*pos)++;
1901         if (*pos >= NR_VMSTAT_ITEMS)
1902                 return NULL;
1903         return (unsigned long *)m->private + *pos;
1904 }
1905
1906 static int vmstat_show(struct seq_file *m, void *arg)
1907 {
1908         unsigned long *l = arg;
1909         unsigned long off = l - (unsigned long *)m->private;
1910
1911         seq_puts(m, vmstat_text[off]);
1912         seq_put_decimal_ull(m, " ", *l);
1913         seq_putc(m, '\n');
1914
1915         if (off == NR_VMSTAT_ITEMS - 1) {
1916                 /*
1917                  * We've come to the end - add any deprecated counters to avoid
1918                  * breaking userspace which might depend on them being present.
1919                  */
1920                 seq_puts(m, "nr_unstable 0\n");
1921         }
1922         return 0;
1923 }
1924
1925 static void vmstat_stop(struct seq_file *m, void *arg)
1926 {
1927         kfree(m->private);
1928         m->private = NULL;
1929 }
1930
1931 static const struct seq_operations vmstat_op = {
1932         .start  = vmstat_start,
1933         .next   = vmstat_next,
1934         .stop   = vmstat_stop,
1935         .show   = vmstat_show,
1936 };
1937 #endif /* CONFIG_PROC_FS */
1938
1939 #ifdef CONFIG_SMP
1940 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1941 int sysctl_stat_interval __read_mostly = HZ;
1942 static int vmstat_late_init_done;
1943
1944 #ifdef CONFIG_PROC_FS
1945 static void refresh_vm_stats(struct work_struct *work)
1946 {
1947         refresh_cpu_vm_stats(true);
1948 }
1949
1950 int vmstat_refresh(const struct ctl_table *table, int write,
1951                    void *buffer, size_t *lenp, loff_t *ppos)
1952 {
1953         long val;
1954         int err;
1955         int i;
1956
1957         /*
1958          * The regular update, every sysctl_stat_interval, may come later
1959          * than expected: leaving a significant amount in per_cpu buckets.
1960          * This is particularly misleading when checking a quantity of HUGE
1961          * pages, immediately after running a test.  /proc/sys/vm/stat_refresh,
1962          * which can equally be echo'ed to or cat'ted from (by root),
1963          * can be used to update the stats just before reading them.
1964          *
1965          * Oh, and since global_zone_page_state() etc. are so careful to hide
1966          * transiently negative values, report an error here if any of
1967          * the stats is negative, so we know to go looking for imbalance.
1968          */
1969         err = schedule_on_each_cpu(refresh_vm_stats);
1970         if (err)
1971                 return err;
1972         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1973                 /*
1974                  * Skip checking stats known to go negative occasionally.
1975                  */
1976                 switch (i) {
1977                 case NR_ZONE_WRITE_PENDING:
1978                 case NR_FREE_CMA_PAGES:
1979                         continue;
1980                 }
1981                 val = atomic_long_read(&vm_zone_stat[i]);
1982                 if (val < 0) {
1983                         pr_warn("%s: %s %ld\n",
1984                                 __func__, zone_stat_name(i), val);
1985                 }
1986         }
1987         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1988                 /*
1989                  * Skip checking stats known to go negative occasionally.
1990                  */
1991                 switch (i) {
1992                 case NR_WRITEBACK:
1993                         continue;
1994                 }
1995                 val = atomic_long_read(&vm_node_stat[i]);
1996                 if (val < 0) {
1997                         pr_warn("%s: %s %ld\n",
1998                                 __func__, node_stat_name(i), val);
1999                 }
2000         }
2001         if (write)
2002                 *ppos += *lenp;
2003         else
2004                 *lenp = 0;
2005         return 0;
2006 }
2007 #endif /* CONFIG_PROC_FS */
2008
2009 static void vmstat_update(struct work_struct *w)
2010 {
2011         if (refresh_cpu_vm_stats(true)) {
2012                 /*
2013                  * Counters were updated so we expect more updates
2014                  * to occur in the future. Keep on running the
2015                  * update worker thread.
2016                  */
2017                 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
2018                                 this_cpu_ptr(&vmstat_work),
2019                                 round_jiffies_relative(sysctl_stat_interval));
2020         }
2021 }
2022
2023 /*
2024  * Check if the diffs for a certain cpu indicate that
2025  * an update is needed.
2026  */
2027 static bool need_update(int cpu)
2028 {
2029         pg_data_t *last_pgdat = NULL;
2030         struct zone *zone;
2031
2032         for_each_populated_zone(zone) {
2033                 struct per_cpu_zonestat *pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
2034                 struct per_cpu_nodestat *n;
2035
2036                 /*
2037                  * The fast way of checking if there are any vmstat diffs.
2038                  */
2039                 if (memchr_inv(pzstats->vm_stat_diff, 0, sizeof(pzstats->vm_stat_diff)))
2040                         return true;
2041
2042                 if (last_pgdat == zone->zone_pgdat)
2043                         continue;
2044                 last_pgdat = zone->zone_pgdat;
2045                 n = per_cpu_ptr(zone->zone_pgdat->per_cpu_nodestats, cpu);
2046                 if (memchr_inv(n->vm_node_stat_diff, 0, sizeof(n->vm_node_stat_diff)))
2047                         return true;
2048         }
2049         return false;
2050 }
2051
2052 /*
2053  * Switch off vmstat processing and then fold all the remaining differentials
2054  * until the diffs stay at zero. The function is used by NOHZ and can only be
2055  * invoked when tick processing is not active.
2056  */
2057 void quiet_vmstat(void)
2058 {
2059         if (system_state != SYSTEM_RUNNING)
2060                 return;
2061
2062         if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
2063                 return;
2064
2065         if (!need_update(smp_processor_id()))
2066                 return;
2067
2068         /*
2069          * Just refresh counters and do not care about the pending delayed
2070          * vmstat_update. It doesn't fire that often to matter and canceling
2071          * it would be too expensive from this path.
2072          * vmstat_shepherd will take care about that for us.
2073          */
2074         refresh_cpu_vm_stats(false);
2075 }
2076
2077 /*
2078  * Shepherd worker thread that checks the
2079  * differentials of processors that have their worker
2080  * threads for vm statistics updates disabled because of
2081  * inactivity.
2082  */
2083 static void vmstat_shepherd(struct work_struct *w);
2084
2085 static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
2086
2087 static void vmstat_shepherd(struct work_struct *w)
2088 {
2089         int cpu;
2090
2091         cpus_read_lock();
2092         /* Check processors whose vmstat worker threads have been disabled */
2093         for_each_online_cpu(cpu) {
2094                 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
2095
2096                 /*
2097                  * In kernel users of vmstat counters either require the precise value and
2098                  * they are using zone_page_state_snapshot interface or they can live with
2099                  * an imprecision as the regular flushing can happen at arbitrary time and
2100                  * cumulative error can grow (see calculate_normal_threshold).
2101                  *
2102                  * From that POV the regular flushing can be postponed for CPUs that have
2103                  * been isolated from the kernel interference without critical
2104                  * infrastructure ever noticing. Skip regular flushing from vmstat_shepherd
2105                  * for all isolated CPUs to avoid interference with the isolated workload.
2106                  */
2107                 if (cpu_is_isolated(cpu))
2108                         continue;
2109
2110                 if (!delayed_work_pending(dw) && need_update(cpu))
2111                         queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
2112
2113                 cond_resched();
2114         }
2115         cpus_read_unlock();
2116
2117         schedule_delayed_work(&shepherd,
2118                 round_jiffies_relative(sysctl_stat_interval));
2119 }
2120
2121 static void __init start_shepherd_timer(void)
2122 {
2123         int cpu;
2124
2125         for_each_possible_cpu(cpu)
2126                 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
2127                         vmstat_update);
2128
2129         schedule_delayed_work(&shepherd,
2130                 round_jiffies_relative(sysctl_stat_interval));
2131 }
2132
2133 static void __init init_cpu_node_state(void)
2134 {
2135         int node;
2136
2137         for_each_online_node(node) {
2138                 if (!cpumask_empty(cpumask_of_node(node)))
2139                         node_set_state(node, N_CPU);
2140         }
2141 }
2142
2143 static int vmstat_cpu_online(unsigned int cpu)
2144 {
2145         if (vmstat_late_init_done)
2146                 refresh_zone_stat_thresholds();
2147
2148         if (!node_state(cpu_to_node(cpu), N_CPU)) {
2149                 node_set_state(cpu_to_node(cpu), N_CPU);
2150         }
2151
2152         return 0;
2153 }
2154
2155 static int vmstat_cpu_down_prep(unsigned int cpu)
2156 {
2157         cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
2158         return 0;
2159 }
2160
2161 static int vmstat_cpu_dead(unsigned int cpu)
2162 {
2163         const struct cpumask *node_cpus;
2164         int node;
2165
2166         node = cpu_to_node(cpu);
2167
2168         refresh_zone_stat_thresholds();
2169         node_cpus = cpumask_of_node(node);
2170         if (!cpumask_empty(node_cpus))
2171                 return 0;
2172
2173         node_clear_state(node, N_CPU);
2174
2175         return 0;
2176 }
2177
2178 static int __init vmstat_late_init(void)
2179 {
2180         refresh_zone_stat_thresholds();
2181         vmstat_late_init_done = 1;
2182
2183         return 0;
2184 }
2185 late_initcall(vmstat_late_init);
2186 #endif
2187
2188 struct workqueue_struct *mm_percpu_wq;
2189
2190 void __init init_mm_internals(void)
2191 {
2192         int ret __maybe_unused;
2193
2194         mm_percpu_wq = alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM, 0);
2195
2196 #ifdef CONFIG_SMP
2197         ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
2198                                         NULL, vmstat_cpu_dead);
2199         if (ret < 0)
2200                 pr_err("vmstat: failed to register 'dead' hotplug state\n");
2201
2202         ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "mm/vmstat:online",
2203                                         vmstat_cpu_online,
2204                                         vmstat_cpu_down_prep);
2205         if (ret < 0)
2206                 pr_err("vmstat: failed to register 'online' hotplug state\n");
2207
2208         cpus_read_lock();
2209         init_cpu_node_state();
2210         cpus_read_unlock();
2211
2212         start_shepherd_timer();
2213 #endif
2214 #ifdef CONFIG_PROC_FS
2215         proc_create_seq("buddyinfo", 0444, NULL, &fragmentation_op);
2216         proc_create_seq("pagetypeinfo", 0400, NULL, &pagetypeinfo_op);
2217         proc_create_seq("vmstat", 0444, NULL, &vmstat_op);
2218         proc_create_seq("zoneinfo", 0444, NULL, &zoneinfo_op);
2219 #endif
2220 }
2221
2222 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
2223
2224 /*
2225  * Return an index indicating how much of the available free memory is
2226  * unusable for an allocation of the requested size.
2227  */
2228 static int unusable_free_index(unsigned int order,
2229                                 struct contig_page_info *info)
2230 {
2231         /* No free memory is interpreted as all free memory is unusable */
2232         if (info->free_pages == 0)
2233                 return 1000;
2234
2235         /*
2236          * Index should be a value between 0 and 1. Return a value to 3
2237          * decimal places.
2238          *
2239          * 0 => no fragmentation
2240          * 1 => high fragmentation
2241          */
2242         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
2243
2244 }
2245
2246 static void unusable_show_print(struct seq_file *m,
2247                                         pg_data_t *pgdat, struct zone *zone)
2248 {
2249         unsigned int order;
2250         int index;
2251         struct contig_page_info info;
2252
2253         seq_printf(m, "Node %d, zone %8s ",
2254                                 pgdat->node_id,
2255                                 zone->name);
2256         for (order = 0; order < NR_PAGE_ORDERS; ++order) {
2257                 fill_contig_page_info(zone, order, &info);
2258                 index = unusable_free_index(order, &info);
2259                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2260         }
2261
2262         seq_putc(m, '\n');
2263 }
2264
2265 /*
2266  * Display unusable free space index
2267  *
2268  * The unusable free space index measures how much of the available free
2269  * memory cannot be used to satisfy an allocation of a given size and is a
2270  * value between 0 and 1. The higher the value, the more of free memory is
2271  * unusable and by implication, the worse the external fragmentation is. This
2272  * can be expressed as a percentage by multiplying by 100.
2273  */
2274 static int unusable_show(struct seq_file *m, void *arg)
2275 {
2276         pg_data_t *pgdat = (pg_data_t *)arg;
2277
2278         /* check memoryless node */
2279         if (!node_state(pgdat->node_id, N_MEMORY))
2280                 return 0;
2281
2282         walk_zones_in_node(m, pgdat, true, false, unusable_show_print);
2283
2284         return 0;
2285 }
2286
2287 static const struct seq_operations unusable_sops = {
2288         .start  = frag_start,
2289         .next   = frag_next,
2290         .stop   = frag_stop,
2291         .show   = unusable_show,
2292 };
2293
2294 DEFINE_SEQ_ATTRIBUTE(unusable);
2295
2296 static void extfrag_show_print(struct seq_file *m,
2297                                         pg_data_t *pgdat, struct zone *zone)
2298 {
2299         unsigned int order;
2300         int index;
2301
2302         /* Alloc on stack as interrupts are disabled for zone walk */
2303         struct contig_page_info info;
2304
2305         seq_printf(m, "Node %d, zone %8s ",
2306                                 pgdat->node_id,
2307                                 zone->name);
2308         for (order = 0; order < NR_PAGE_ORDERS; ++order) {
2309                 fill_contig_page_info(zone, order, &info);
2310                 index = __fragmentation_index(order, &info);
2311                 seq_printf(m, "%2d.%03d ", index / 1000, index % 1000);
2312         }
2313
2314         seq_putc(m, '\n');
2315 }
2316
2317 /*
2318  * Display fragmentation index for orders that allocations would fail for
2319  */
2320 static int extfrag_show(struct seq_file *m, void *arg)
2321 {
2322         pg_data_t *pgdat = (pg_data_t *)arg;
2323
2324         walk_zones_in_node(m, pgdat, true, false, extfrag_show_print);
2325
2326         return 0;
2327 }
2328
2329 static const struct seq_operations extfrag_sops = {
2330         .start  = frag_start,
2331         .next   = frag_next,
2332         .stop   = frag_stop,
2333         .show   = extfrag_show,
2334 };
2335
2336 DEFINE_SEQ_ATTRIBUTE(extfrag);
2337
2338 static int __init extfrag_debug_init(void)
2339 {
2340         struct dentry *extfrag_debug_root;
2341
2342         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
2343
2344         debugfs_create_file("unusable_index", 0444, extfrag_debug_root, NULL,
2345                             &unusable_fops);
2346
2347         debugfs_create_file("extfrag_index", 0444, extfrag_debug_root, NULL,
2348                             &extfrag_fops);
2349
2350         return 0;
2351 }
2352
2353 module_init(extfrag_debug_init);
2354
2355 #endif