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 #ifdef CONFIG_COMPACTION
1037
1038 struct contig_page_info {
1039         unsigned long free_pages;
1040         unsigned long free_blocks_total;
1041         unsigned long free_blocks_suitable;
1042 };
1043
1044 /*
1045  * Calculate the number of free pages in a zone, how many contiguous
1046  * pages are free and how many are large enough to satisfy an allocation of
1047  * the target size. Note that this function makes no attempt to estimate
1048  * how many suitable free blocks there *might* be if MOVABLE pages were
1049  * migrated. Calculating that is possible, but expensive and can be
1050  * figured out from userspace
1051  */
1052 static void fill_contig_page_info(struct zone *zone,
1053                                 unsigned int suitable_order,
1054                                 struct contig_page_info *info)
1055 {
1056         unsigned int order;
1057
1058         info->free_pages = 0;
1059         info->free_blocks_total = 0;
1060         info->free_blocks_suitable = 0;
1061
1062         for (order = 0; order < NR_PAGE_ORDERS; order++) {
1063                 unsigned long blocks;
1064
1065                 /*
1066                  * Count number of free blocks.
1067                  *
1068                  * Access to nr_free is lockless as nr_free is used only for
1069                  * diagnostic purposes. Use data_race to avoid KCSAN warning.
1070                  */
1071                 blocks = data_race(zone->free_area[order].nr_free);
1072                 info->free_blocks_total += blocks;
1073
1074                 /* Count free base pages */
1075                 info->free_pages += blocks << order;
1076
1077                 /* Count the suitable free blocks */
1078                 if (order >= suitable_order)
1079                         info->free_blocks_suitable += blocks <<
1080                                                 (order - suitable_order);
1081         }
1082 }
1083
1084 /*
1085  * A fragmentation index only makes sense if an allocation of a requested
1086  * size would fail. If that is true, the fragmentation index indicates
1087  * whether external fragmentation or a lack of memory was the problem.
1088  * The value can be used to determine if page reclaim or compaction
1089  * should be used
1090  */
1091 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
1092 {
1093         unsigned long requested = 1UL << order;
1094
1095         if (WARN_ON_ONCE(order > MAX_PAGE_ORDER))
1096                 return 0;
1097
1098         if (!info->free_blocks_total)
1099                 return 0;
1100
1101         /* Fragmentation index only makes sense when a request would fail */
1102         if (info->free_blocks_suitable)
1103                 return -1000;
1104
1105         /*
1106          * Index is between 0 and 1 so return within 3 decimal places
1107          *
1108          * 0 => allocation would fail due to lack of memory
1109          * 1 => allocation would fail due to fragmentation
1110          */
1111         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
1112 }
1113
1114 /*
1115  * Calculates external fragmentation within a zone wrt the given order.
1116  * It is defined as the percentage of pages found in blocks of size
1117  * less than 1 << order. It returns values in range [0, 100].
1118  */
1119 unsigned int extfrag_for_order(struct zone *zone, unsigned int order)
1120 {
1121         struct contig_page_info info;
1122
1123         fill_contig_page_info(zone, order, &info);
1124         if (info.free_pages == 0)
1125                 return 0;
1126
1127         return div_u64((info.free_pages -
1128                         (info.free_blocks_suitable << order)) * 100,
1129                         info.free_pages);
1130 }
1131
1132 /* Same as __fragmentation index but allocs contig_page_info on stack */
1133 int fragmentation_index(struct zone *zone, unsigned int order)
1134 {
1135         struct contig_page_info info;
1136
1137         fill_contig_page_info(zone, order, &info);
1138         return __fragmentation_index(order, &info);
1139 }
1140 #endif
1141
1142 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || \
1143     defined(CONFIG_NUMA) || defined(CONFIG_MEMCG)
1144 #ifdef CONFIG_ZONE_DMA
1145 #define TEXT_FOR_DMA(xx) xx "_dma",
1146 #else
1147 #define TEXT_FOR_DMA(xx)
1148 #endif
1149
1150 #ifdef CONFIG_ZONE_DMA32
1151 #define TEXT_FOR_DMA32(xx) xx "_dma32",
1152 #else
1153 #define TEXT_FOR_DMA32(xx)
1154 #endif
1155
1156 #ifdef CONFIG_HIGHMEM
1157 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
1158 #else
1159 #define TEXT_FOR_HIGHMEM(xx)
1160 #endif
1161
1162 #ifdef CONFIG_ZONE_DEVICE
1163 #define TEXT_FOR_DEVICE(xx) xx "_device",
1164 #else
1165 #define TEXT_FOR_DEVICE(xx)
1166 #endif
1167
1168 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1169                                         TEXT_FOR_HIGHMEM(xx) xx "_movable", \
1170                                         TEXT_FOR_DEVICE(xx)
1171
1172 const char * const vmstat_text[] = {
1173         /* enum zone_stat_item counters */
1174         "nr_free_pages",
1175         "nr_zone_inactive_anon",
1176         "nr_zone_active_anon",
1177         "nr_zone_inactive_file",
1178         "nr_zone_active_file",
1179         "nr_zone_unevictable",
1180         "nr_zone_write_pending",
1181         "nr_mlock",
1182         "nr_bounce",
1183 #if IS_ENABLED(CONFIG_ZSMALLOC)
1184         "nr_zspages",
1185 #endif
1186         "nr_free_cma",
1187 #ifdef CONFIG_UNACCEPTED_MEMORY
1188         "nr_unaccepted",
1189 #endif
1190
1191         /* enum numa_stat_item counters */
1192 #ifdef CONFIG_NUMA
1193         "numa_hit",
1194         "numa_miss",
1195         "numa_foreign",
1196         "numa_interleave",
1197         "numa_local",
1198         "numa_other",
1199 #endif
1200
1201         /* enum node_stat_item counters */
1202         "nr_inactive_anon",
1203         "nr_active_anon",
1204         "nr_inactive_file",
1205         "nr_active_file",
1206         "nr_unevictable",
1207         "nr_slab_reclaimable",
1208         "nr_slab_unreclaimable",
1209         "nr_isolated_anon",
1210         "nr_isolated_file",
1211         "workingset_nodes",
1212         "workingset_refault_anon",
1213         "workingset_refault_file",
1214         "workingset_activate_anon",
1215         "workingset_activate_file",
1216         "workingset_restore_anon",
1217         "workingset_restore_file",
1218         "workingset_nodereclaim",
1219         "nr_anon_pages",
1220         "nr_mapped",
1221         "nr_file_pages",
1222         "nr_dirty",
1223         "nr_writeback",
1224         "nr_writeback_temp",
1225         "nr_shmem",
1226         "nr_shmem_hugepages",
1227         "nr_shmem_pmdmapped",
1228         "nr_file_hugepages",
1229         "nr_file_pmdmapped",
1230         "nr_anon_transparent_hugepages",
1231         "nr_vmscan_write",
1232         "nr_vmscan_immediate_reclaim",
1233         "nr_dirtied",
1234         "nr_written",
1235         "nr_throttled_written",
1236         "nr_kernel_misc_reclaimable",
1237         "nr_foll_pin_acquired",
1238         "nr_foll_pin_released",
1239         "nr_kernel_stack",
1240 #if IS_ENABLED(CONFIG_SHADOW_CALL_STACK)
1241         "nr_shadow_call_stack",
1242 #endif
1243         "nr_page_table_pages",
1244         "nr_sec_page_table_pages",
1245 #ifdef CONFIG_IOMMU_SUPPORT
1246         "nr_iommu_pages",
1247 #endif
1248 #ifdef CONFIG_SWAP
1249         "nr_swapcached",
1250 #endif
1251 #ifdef CONFIG_NUMA_BALANCING
1252         "pgpromote_success",
1253         "pgpromote_candidate",
1254 #endif
1255         "pgdemote_kswapd",
1256         "pgdemote_direct",
1257         "pgdemote_khugepaged",
1258         "nr_memmap",
1259         "nr_memmap_boot",
1260         /* enum writeback_stat_item counters */
1261         "nr_dirty_threshold",
1262         "nr_dirty_background_threshold",
1263
1264 #if defined(CONFIG_VM_EVENT_COUNTERS) || defined(CONFIG_MEMCG)
1265         /* enum vm_event_item counters */
1266         "pgpgin",
1267         "pgpgout",
1268         "pswpin",
1269         "pswpout",
1270
1271         TEXTS_FOR_ZONES("pgalloc")
1272         TEXTS_FOR_ZONES("allocstall")
1273         TEXTS_FOR_ZONES("pgskip")
1274
1275         "pgfree",
1276         "pgactivate",
1277         "pgdeactivate",
1278         "pglazyfree",
1279
1280         "pgfault",
1281         "pgmajfault",
1282         "pglazyfreed",
1283
1284         "pgrefill",
1285         "pgreuse",
1286         "pgsteal_kswapd",
1287         "pgsteal_direct",
1288         "pgsteal_khugepaged",
1289         "pgscan_kswapd",
1290         "pgscan_direct",
1291         "pgscan_khugepaged",
1292         "pgscan_direct_throttle",
1293         "pgscan_anon",
1294         "pgscan_file",
1295         "pgsteal_anon",
1296         "pgsteal_file",
1297
1298 #ifdef CONFIG_NUMA
1299         "zone_reclaim_failed",
1300 #endif
1301         "pginodesteal",
1302         "slabs_scanned",
1303         "kswapd_inodesteal",
1304         "kswapd_low_wmark_hit_quickly",
1305         "kswapd_high_wmark_hit_quickly",
1306         "pageoutrun",
1307
1308         "pgrotated",
1309
1310         "drop_pagecache",
1311         "drop_slab",
1312         "oom_kill",
1313
1314 #ifdef CONFIG_NUMA_BALANCING
1315         "numa_pte_updates",
1316         "numa_huge_pte_updates",
1317         "numa_hint_faults",
1318         "numa_hint_faults_local",
1319         "numa_pages_migrated",
1320 #endif
1321 #ifdef CONFIG_MIGRATION
1322         "pgmigrate_success",
1323         "pgmigrate_fail",
1324         "thp_migration_success",
1325         "thp_migration_fail",
1326         "thp_migration_split",
1327 #endif
1328 #ifdef CONFIG_COMPACTION
1329         "compact_migrate_scanned",
1330         "compact_free_scanned",
1331         "compact_isolated",
1332         "compact_stall",
1333         "compact_fail",
1334         "compact_success",
1335         "compact_daemon_wake",
1336         "compact_daemon_migrate_scanned",
1337         "compact_daemon_free_scanned",
1338 #endif
1339
1340 #ifdef CONFIG_HUGETLB_PAGE
1341         "htlb_buddy_alloc_success",
1342         "htlb_buddy_alloc_fail",
1343 #endif
1344 #ifdef CONFIG_CMA
1345         "cma_alloc_success",
1346         "cma_alloc_fail",
1347 #endif
1348         "unevictable_pgs_culled",
1349         "unevictable_pgs_scanned",
1350         "unevictable_pgs_rescued",
1351         "unevictable_pgs_mlocked",
1352         "unevictable_pgs_munlocked",
1353         "unevictable_pgs_cleared",
1354         "unevictable_pgs_stranded",
1355
1356 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1357         "thp_fault_alloc",
1358         "thp_fault_fallback",
1359         "thp_fault_fallback_charge",
1360         "thp_collapse_alloc",
1361         "thp_collapse_alloc_failed",
1362         "thp_file_alloc",
1363         "thp_file_fallback",
1364         "thp_file_fallback_charge",
1365         "thp_file_mapped",
1366         "thp_split_page",
1367         "thp_split_page_failed",
1368         "thp_deferred_split_page",
1369         "thp_split_pmd",
1370         "thp_scan_exceed_none_pte",
1371         "thp_scan_exceed_swap_pte",
1372         "thp_scan_exceed_share_pte",
1373 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1374         "thp_split_pud",
1375 #endif
1376         "thp_zero_page_alloc",
1377         "thp_zero_page_alloc_failed",
1378         "thp_swpout",
1379         "thp_swpout_fallback",
1380 #endif
1381 #ifdef CONFIG_MEMORY_BALLOON
1382         "balloon_inflate",
1383         "balloon_deflate",
1384 #ifdef CONFIG_BALLOON_COMPACTION
1385         "balloon_migrate",
1386 #endif
1387 #endif /* CONFIG_MEMORY_BALLOON */
1388 #ifdef CONFIG_DEBUG_TLBFLUSH
1389         "nr_tlb_remote_flush",
1390         "nr_tlb_remote_flush_received",
1391         "nr_tlb_local_flush_all",
1392         "nr_tlb_local_flush_one",
1393 #endif /* CONFIG_DEBUG_TLBFLUSH */
1394
1395 #ifdef CONFIG_SWAP
1396         "swap_ra",
1397         "swap_ra_hit",
1398 #ifdef CONFIG_KSM
1399         "ksm_swpin_copy",
1400 #endif
1401 #endif
1402 #ifdef CONFIG_KSM
1403         "cow_ksm",
1404 #endif
1405 #ifdef CONFIG_ZSWAP
1406         "zswpin",
1407         "zswpout",
1408         "zswpwb",
1409 #endif
1410 #ifdef CONFIG_X86
1411         "direct_map_level2_splits",
1412         "direct_map_level3_splits",
1413 #endif
1414 #ifdef CONFIG_PER_VMA_LOCK_STATS
1415         "vma_lock_success",
1416         "vma_lock_abort",
1417         "vma_lock_retry",
1418         "vma_lock_miss",
1419 #endif
1420 #endif /* CONFIG_VM_EVENT_COUNTERS || CONFIG_MEMCG */
1421 };
1422 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA || CONFIG_MEMCG */
1423
1424 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1425      defined(CONFIG_PROC_FS)
1426 static void *frag_start(struct seq_file *m, loff_t *pos)
1427 {
1428         pg_data_t *pgdat;
1429         loff_t node = *pos;
1430
1431         for (pgdat = first_online_pgdat();
1432              pgdat && node;
1433              pgdat = next_online_pgdat(pgdat))
1434                 --node;
1435
1436         return pgdat;
1437 }
1438
1439 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1440 {
1441         pg_data_t *pgdat = (pg_data_t *)arg;
1442
1443         (*pos)++;
1444         return next_online_pgdat(pgdat);
1445 }
1446
1447 static void frag_stop(struct seq_file *m, void *arg)
1448 {
1449 }
1450
1451 /*
1452  * Walk zones in a node and print using a callback.
1453  * If @assert_populated is true, only use callback for zones that are populated.
1454  */
1455 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1456                 bool assert_populated, bool nolock,
1457                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1458 {
1459         struct zone *zone;
1460         struct zone *node_zones = pgdat->node_zones;
1461         unsigned long flags;
1462
1463         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1464                 if (assert_populated && !populated_zone(zone))
1465                         continue;
1466
1467                 if (!nolock)
1468                         spin_lock_irqsave(&zone->lock, flags);
1469                 print(m, pgdat, zone);
1470                 if (!nolock)
1471                         spin_unlock_irqrestore(&zone->lock, flags);
1472         }
1473 }
1474 #endif
1475
1476 #ifdef CONFIG_PROC_FS
1477 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1478                                                 struct zone *zone)
1479 {
1480         int order;
1481
1482         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1483         for (order = 0; order < NR_PAGE_ORDERS; ++order)
1484                 /*
1485                  * Access to nr_free is lockless as nr_free is used only for
1486                  * printing purposes. Use data_race to avoid KCSAN warning.
1487                  */
1488                 seq_printf(m, "%6lu ", data_race(zone->free_area[order].nr_free));
1489         seq_putc(m, '\n');
1490 }
1491
1492 /*
1493  * This walks the free areas for each zone.
1494  */
1495 static int frag_show(struct seq_file *m, void *arg)
1496 {
1497         pg_data_t *pgdat = (pg_data_t *)arg;
1498         walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1499         return 0;
1500 }
1501
1502 static void pagetypeinfo_showfree_print(struct seq_file *m,
1503                                         pg_data_t *pgdat, struct zone *zone)
1504 {
1505         int order, mtype;
1506
1507         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1508                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
1509                                         pgdat->node_id,
1510                                         zone->name,
1511                                         migratetype_names[mtype]);
1512                 for (order = 0; order < NR_PAGE_ORDERS; ++order) {
1513                         unsigned long freecount = 0;
1514                         struct free_area *area;
1515                         struct list_head *curr;
1516                         bool overflow = false;
1517
1518                         area = &(zone->free_area[order]);
1519
1520                         list_for_each(curr, &area->free_list[mtype]) {
1521                                 /*
1522                                  * Cap the free_list iteration because it might
1523                                  * be really large and we are under a spinlock
1524                                  * so a long time spent here could trigger a
1525                                  * hard lockup detector. Anyway this is a
1526                                  * debugging tool so knowing there is a handful
1527                                  * of pages of this order should be more than
1528                                  * sufficient.
1529                                  */
1530                                 if (++freecount >= 100000) {
1531                                         overflow = true;
1532                                         break;
1533                                 }
1534                         }
1535                         seq_printf(m, "%s%6lu ", overflow ? ">" : "", freecount);
1536                         spin_unlock_irq(&zone->lock);
1537                         cond_resched();
1538                         spin_lock_irq(&zone->lock);
1539                 }
1540                 seq_putc(m, '\n');
1541         }
1542 }
1543
1544 /* Print out the free pages at each order for each migatetype */
1545 static void pagetypeinfo_showfree(struct seq_file *m, void *arg)
1546 {
1547         int order;
1548         pg_data_t *pgdat = (pg_data_t *)arg;
1549
1550         /* Print header */
1551         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1552         for (order = 0; order < NR_PAGE_ORDERS; ++order)
1553                 seq_printf(m, "%6d ", order);
1554         seq_putc(m, '\n');
1555
1556         walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1557 }
1558
1559 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1560                                         pg_data_t *pgdat, struct zone *zone)
1561 {
1562         int mtype;
1563         unsigned long pfn;
1564         unsigned long start_pfn = zone->zone_start_pfn;
1565         unsigned long end_pfn = zone_end_pfn(zone);
1566         unsigned long count[MIGRATE_TYPES] = { 0, };
1567
1568         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1569                 struct page *page;
1570
1571                 page = pfn_to_online_page(pfn);
1572                 if (!page)
1573                         continue;
1574
1575                 if (page_zone(page) != zone)
1576                         continue;
1577
1578                 mtype = get_pageblock_migratetype(page);
1579
1580                 if (mtype < MIGRATE_TYPES)
1581                         count[mtype]++;
1582         }
1583
1584         /* Print counts */
1585         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1586         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1587                 seq_printf(m, "%12lu ", count[mtype]);
1588         seq_putc(m, '\n');
1589 }
1590
1591 /* Print out the number of pageblocks for each migratetype */
1592 static void pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1593 {
1594         int mtype;
1595         pg_data_t *pgdat = (pg_data_t *)arg;
1596
1597         seq_printf(m, "\n%-23s", "Number of blocks type ");
1598         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1599                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1600         seq_putc(m, '\n');
1601         walk_zones_in_node(m, pgdat, true, false,
1602                 pagetypeinfo_showblockcount_print);
1603 }
1604
1605 /*
1606  * Print out the number of pageblocks for each migratetype that contain pages
1607  * of other types. This gives an indication of how well fallbacks are being
1608  * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1609  * to determine what is going on
1610  */
1611 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1612 {
1613 #ifdef CONFIG_PAGE_OWNER
1614         int mtype;
1615
1616         if (!static_branch_unlikely(&page_owner_inited))
1617                 return;
1618
1619         drain_all_pages(NULL);
1620
1621         seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1622         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1623                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1624         seq_putc(m, '\n');
1625
1626         walk_zones_in_node(m, pgdat, true, true,
1627                 pagetypeinfo_showmixedcount_print);
1628 #endif /* CONFIG_PAGE_OWNER */
1629 }
1630
1631 /*
1632  * This prints out statistics in relation to grouping pages by mobility.
1633  * It is expensive to collect so do not constantly read the file.
1634  */
1635 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1636 {
1637         pg_data_t *pgdat = (pg_data_t *)arg;
1638
1639         /* check memoryless node */
1640         if (!node_state(pgdat->node_id, N_MEMORY))
1641                 return 0;
1642
1643         seq_printf(m, "Page block order: %d\n", pageblock_order);
1644         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
1645         seq_putc(m, '\n');
1646         pagetypeinfo_showfree(m, pgdat);
1647         pagetypeinfo_showblockcount(m, pgdat);
1648         pagetypeinfo_showmixedcount(m, pgdat);
1649
1650         return 0;
1651 }
1652
1653 static const struct seq_operations fragmentation_op = {
1654         .start  = frag_start,
1655         .next   = frag_next,
1656         .stop   = frag_stop,
1657         .show   = frag_show,
1658 };
1659
1660 static const struct seq_operations pagetypeinfo_op = {
1661         .start  = frag_start,
1662         .next   = frag_next,
1663         .stop   = frag_stop,
1664         .show   = pagetypeinfo_show,
1665 };
1666
1667 static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1668 {
1669         int zid;
1670
1671         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1672                 struct zone *compare = &pgdat->node_zones[zid];
1673
1674                 if (populated_zone(compare))
1675                         return zone == compare;
1676         }
1677
1678         return false;
1679 }
1680
1681 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1682                                                         struct zone *zone)
1683 {
1684         int i;
1685         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1686         if (is_zone_first_populated(pgdat, zone)) {
1687                 seq_printf(m, "\n  per-node stats");
1688                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1689                         unsigned long pages = node_page_state_pages(pgdat, i);
1690
1691                         if (vmstat_item_print_in_thp(i))
1692                                 pages /= HPAGE_PMD_NR;
1693                         seq_printf(m, "\n      %-12s %lu", node_stat_name(i),
1694                                    pages);
1695                 }
1696         }
1697         seq_printf(m,
1698                    "\n  pages free     %lu"
1699                    "\n        boost    %lu"
1700                    "\n        min      %lu"
1701                    "\n        low      %lu"
1702                    "\n        high     %lu"
1703                    "\n        spanned  %lu"
1704                    "\n        present  %lu"
1705                    "\n        managed  %lu"
1706                    "\n        cma      %lu",
1707                    zone_page_state(zone, NR_FREE_PAGES),
1708                    zone->watermark_boost,
1709                    min_wmark_pages(zone),
1710                    low_wmark_pages(zone),
1711                    high_wmark_pages(zone),
1712                    zone->spanned_pages,
1713                    zone->present_pages,
1714                    zone_managed_pages(zone),
1715                    zone_cma_pages(zone));
1716
1717         seq_printf(m,
1718                    "\n        protection: (%ld",
1719                    zone->lowmem_reserve[0]);
1720         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1721                 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1722         seq_putc(m, ')');
1723
1724         /* If unpopulated, no other information is useful */
1725         if (!populated_zone(zone)) {
1726                 seq_putc(m, '\n');
1727                 return;
1728         }
1729
1730         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1731                 seq_printf(m, "\n      %-12s %lu", zone_stat_name(i),
1732                            zone_page_state(zone, i));
1733
1734 #ifdef CONFIG_NUMA
1735         for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++)
1736                 seq_printf(m, "\n      %-12s %lu", numa_stat_name(i),
1737                            zone_numa_event_state(zone, i));
1738 #endif
1739
1740         seq_printf(m, "\n  pagesets");
1741         for_each_online_cpu(i) {
1742                 struct per_cpu_pages *pcp;
1743                 struct per_cpu_zonestat __maybe_unused *pzstats;
1744
1745                 pcp = per_cpu_ptr(zone->per_cpu_pageset, i);
1746                 seq_printf(m,
1747                            "\n    cpu: %i"
1748                            "\n              count: %i"
1749                            "\n              high:  %i"
1750                            "\n              batch: %i",
1751                            i,
1752                            pcp->count,
1753                            pcp->high,
1754                            pcp->batch);
1755 #ifdef CONFIG_SMP
1756                 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, i);
1757                 seq_printf(m, "\n  vm stats threshold: %d",
1758                                 pzstats->stat_threshold);
1759 #endif
1760         }
1761         seq_printf(m,
1762                    "\n  node_unreclaimable:  %u"
1763                    "\n  start_pfn:           %lu",
1764                    pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1765                    zone->zone_start_pfn);
1766         seq_putc(m, '\n');
1767 }
1768
1769 /*
1770  * Output information about zones in @pgdat.  All zones are printed regardless
1771  * of whether they are populated or not: lowmem_reserve_ratio operates on the
1772  * set of all zones and userspace would not be aware of such zones if they are
1773  * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1774  */
1775 static int zoneinfo_show(struct seq_file *m, void *arg)
1776 {
1777         pg_data_t *pgdat = (pg_data_t *)arg;
1778         walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1779         return 0;
1780 }
1781
1782 static const struct seq_operations zoneinfo_op = {
1783         .start  = frag_start, /* iterate over all zones. The same as in
1784                                * fragmentation. */
1785         .next   = frag_next,
1786         .stop   = frag_stop,
1787         .show   = zoneinfo_show,
1788 };
1789
1790 #define NR_VMSTAT_ITEMS (NR_VM_ZONE_STAT_ITEMS + \
1791                          NR_VM_NUMA_EVENT_ITEMS + \
1792                          NR_VM_NODE_STAT_ITEMS + \
1793                          NR_VM_WRITEBACK_STAT_ITEMS + \
1794                          (IS_ENABLED(CONFIG_VM_EVENT_COUNTERS) ? \
1795                           NR_VM_EVENT_ITEMS : 0))
1796
1797 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1798 {
1799         unsigned long *v;
1800         int i;
1801
1802         if (*pos >= NR_VMSTAT_ITEMS)
1803                 return NULL;
1804
1805         BUILD_BUG_ON(ARRAY_SIZE(vmstat_text) < NR_VMSTAT_ITEMS);
1806         fold_vm_numa_events();
1807         v = kmalloc_array(NR_VMSTAT_ITEMS, sizeof(unsigned long), GFP_KERNEL);
1808         m->private = v;
1809         if (!v)
1810                 return ERR_PTR(-ENOMEM);
1811         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1812                 v[i] = global_zone_page_state(i);
1813         v += NR_VM_ZONE_STAT_ITEMS;
1814
1815 #ifdef CONFIG_NUMA
1816         for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++)
1817                 v[i] = global_numa_event_state(i);
1818         v += NR_VM_NUMA_EVENT_ITEMS;
1819 #endif
1820
1821         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1822                 v[i] = global_node_page_state_pages(i);
1823                 if (vmstat_item_print_in_thp(i))
1824                         v[i] /= HPAGE_PMD_NR;
1825         }
1826         v += NR_VM_NODE_STAT_ITEMS;
1827
1828         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1829                             v + NR_DIRTY_THRESHOLD);
1830         v += NR_VM_WRITEBACK_STAT_ITEMS;
1831
1832 #ifdef CONFIG_VM_EVENT_COUNTERS
1833         all_vm_events(v);
1834         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1835         v[PGPGOUT] /= 2;
1836 #endif
1837         return (unsigned long *)m->private + *pos;
1838 }
1839
1840 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1841 {
1842         (*pos)++;
1843         if (*pos >= NR_VMSTAT_ITEMS)
1844                 return NULL;
1845         return (unsigned long *)m->private + *pos;
1846 }
1847
1848 static int vmstat_show(struct seq_file *m, void *arg)
1849 {
1850         unsigned long *l = arg;
1851         unsigned long off = l - (unsigned long *)m->private;
1852
1853         seq_puts(m, vmstat_text[off]);
1854         seq_put_decimal_ull(m, " ", *l);
1855         seq_putc(m, '\n');
1856
1857         if (off == NR_VMSTAT_ITEMS - 1) {
1858                 /*
1859                  * We've come to the end - add any deprecated counters to avoid
1860                  * breaking userspace which might depend on them being present.
1861                  */
1862                 seq_puts(m, "nr_unstable 0\n");
1863         }
1864         return 0;
1865 }
1866
1867 static void vmstat_stop(struct seq_file *m, void *arg)
1868 {
1869         kfree(m->private);
1870         m->private = NULL;
1871 }
1872
1873 static const struct seq_operations vmstat_op = {
1874         .start  = vmstat_start,
1875         .next   = vmstat_next,
1876         .stop   = vmstat_stop,
1877         .show   = vmstat_show,
1878 };
1879 #endif /* CONFIG_PROC_FS */
1880
1881 #ifdef CONFIG_SMP
1882 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1883 int sysctl_stat_interval __read_mostly = HZ;
1884
1885 #ifdef CONFIG_PROC_FS
1886 static void refresh_vm_stats(struct work_struct *work)
1887 {
1888         refresh_cpu_vm_stats(true);
1889 }
1890
1891 int vmstat_refresh(const struct ctl_table *table, int write,
1892                    void *buffer, size_t *lenp, loff_t *ppos)
1893 {
1894         long val;
1895         int err;
1896         int i;
1897
1898         /*
1899          * The regular update, every sysctl_stat_interval, may come later
1900          * than expected: leaving a significant amount in per_cpu buckets.
1901          * This is particularly misleading when checking a quantity of HUGE
1902          * pages, immediately after running a test.  /proc/sys/vm/stat_refresh,
1903          * which can equally be echo'ed to or cat'ted from (by root),
1904          * can be used to update the stats just before reading them.
1905          *
1906          * Oh, and since global_zone_page_state() etc. are so careful to hide
1907          * transiently negative values, report an error here if any of
1908          * the stats is negative, so we know to go looking for imbalance.
1909          */
1910         err = schedule_on_each_cpu(refresh_vm_stats);
1911         if (err)
1912                 return err;
1913         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1914                 /*
1915                  * Skip checking stats known to go negative occasionally.
1916                  */
1917                 switch (i) {
1918                 case NR_ZONE_WRITE_PENDING:
1919                 case NR_FREE_CMA_PAGES:
1920                         continue;
1921                 }
1922                 val = atomic_long_read(&vm_zone_stat[i]);
1923                 if (val < 0) {
1924                         pr_warn("%s: %s %ld\n",
1925                                 __func__, zone_stat_name(i), val);
1926                 }
1927         }
1928         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1929                 /*
1930                  * Skip checking stats known to go negative occasionally.
1931                  */
1932                 switch (i) {
1933                 case NR_WRITEBACK:
1934                         continue;
1935                 }
1936                 val = atomic_long_read(&vm_node_stat[i]);
1937                 if (val < 0) {
1938                         pr_warn("%s: %s %ld\n",
1939                                 __func__, node_stat_name(i), val);
1940                 }
1941         }
1942         if (write)
1943                 *ppos += *lenp;
1944         else
1945                 *lenp = 0;
1946         return 0;
1947 }
1948 #endif /* CONFIG_PROC_FS */
1949
1950 static void vmstat_update(struct work_struct *w)
1951 {
1952         if (refresh_cpu_vm_stats(true)) {
1953                 /*
1954                  * Counters were updated so we expect more updates
1955                  * to occur in the future. Keep on running the
1956                  * update worker thread.
1957                  */
1958                 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
1959                                 this_cpu_ptr(&vmstat_work),
1960                                 round_jiffies_relative(sysctl_stat_interval));
1961         }
1962 }
1963
1964 /*
1965  * Check if the diffs for a certain cpu indicate that
1966  * an update is needed.
1967  */
1968 static bool need_update(int cpu)
1969 {
1970         pg_data_t *last_pgdat = NULL;
1971         struct zone *zone;
1972
1973         for_each_populated_zone(zone) {
1974                 struct per_cpu_zonestat *pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
1975                 struct per_cpu_nodestat *n;
1976
1977                 /*
1978                  * The fast way of checking if there are any vmstat diffs.
1979                  */
1980                 if (memchr_inv(pzstats->vm_stat_diff, 0, sizeof(pzstats->vm_stat_diff)))
1981                         return true;
1982
1983                 if (last_pgdat == zone->zone_pgdat)
1984                         continue;
1985                 last_pgdat = zone->zone_pgdat;
1986                 n = per_cpu_ptr(zone->zone_pgdat->per_cpu_nodestats, cpu);
1987                 if (memchr_inv(n->vm_node_stat_diff, 0, sizeof(n->vm_node_stat_diff)))
1988                         return true;
1989         }
1990         return false;
1991 }
1992
1993 /*
1994  * Switch off vmstat processing and then fold all the remaining differentials
1995  * until the diffs stay at zero. The function is used by NOHZ and can only be
1996  * invoked when tick processing is not active.
1997  */
1998 void quiet_vmstat(void)
1999 {
2000         if (system_state != SYSTEM_RUNNING)
2001                 return;
2002
2003         if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
2004                 return;
2005
2006         if (!need_update(smp_processor_id()))
2007                 return;
2008
2009         /*
2010          * Just refresh counters and do not care about the pending delayed
2011          * vmstat_update. It doesn't fire that often to matter and canceling
2012          * it would be too expensive from this path.
2013          * vmstat_shepherd will take care about that for us.
2014          */
2015         refresh_cpu_vm_stats(false);
2016 }
2017
2018 /*
2019  * Shepherd worker thread that checks the
2020  * differentials of processors that have their worker
2021  * threads for vm statistics updates disabled because of
2022  * inactivity.
2023  */
2024 static void vmstat_shepherd(struct work_struct *w);
2025
2026 static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
2027
2028 static void vmstat_shepherd(struct work_struct *w)
2029 {
2030         int cpu;
2031
2032         cpus_read_lock();
2033         /* Check processors whose vmstat worker threads have been disabled */
2034         for_each_online_cpu(cpu) {
2035                 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
2036
2037                 /*
2038                  * In kernel users of vmstat counters either require the precise value and
2039                  * they are using zone_page_state_snapshot interface or they can live with
2040                  * an imprecision as the regular flushing can happen at arbitrary time and
2041                  * cumulative error can grow (see calculate_normal_threshold).
2042                  *
2043                  * From that POV the regular flushing can be postponed for CPUs that have
2044                  * been isolated from the kernel interference without critical
2045                  * infrastructure ever noticing. Skip regular flushing from vmstat_shepherd
2046                  * for all isolated CPUs to avoid interference with the isolated workload.
2047                  */
2048                 if (cpu_is_isolated(cpu))
2049                         continue;
2050
2051                 if (!delayed_work_pending(dw) && need_update(cpu))
2052                         queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
2053
2054                 cond_resched();
2055         }
2056         cpus_read_unlock();
2057
2058         schedule_delayed_work(&shepherd,
2059                 round_jiffies_relative(sysctl_stat_interval));
2060 }
2061
2062 static void __init start_shepherd_timer(void)
2063 {
2064         int cpu;
2065
2066         for_each_possible_cpu(cpu)
2067                 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
2068                         vmstat_update);
2069
2070         schedule_delayed_work(&shepherd,
2071                 round_jiffies_relative(sysctl_stat_interval));
2072 }
2073
2074 static void __init init_cpu_node_state(void)
2075 {
2076         int node;
2077
2078         for_each_online_node(node) {
2079                 if (!cpumask_empty(cpumask_of_node(node)))
2080                         node_set_state(node, N_CPU);
2081         }
2082 }
2083
2084 static int vmstat_cpu_online(unsigned int cpu)
2085 {
2086         refresh_zone_stat_thresholds();
2087
2088         if (!node_state(cpu_to_node(cpu), N_CPU)) {
2089                 node_set_state(cpu_to_node(cpu), N_CPU);
2090         }
2091
2092         return 0;
2093 }
2094
2095 static int vmstat_cpu_down_prep(unsigned int cpu)
2096 {
2097         cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
2098         return 0;
2099 }
2100
2101 static int vmstat_cpu_dead(unsigned int cpu)
2102 {
2103         const struct cpumask *node_cpus;
2104         int node;
2105
2106         node = cpu_to_node(cpu);
2107
2108         refresh_zone_stat_thresholds();
2109         node_cpus = cpumask_of_node(node);
2110         if (!cpumask_empty(node_cpus))
2111                 return 0;
2112
2113         node_clear_state(node, N_CPU);
2114
2115         return 0;
2116 }
2117
2118 #endif
2119
2120 struct workqueue_struct *mm_percpu_wq;
2121
2122 void __init init_mm_internals(void)
2123 {
2124         int ret __maybe_unused;
2125
2126         mm_percpu_wq = alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM, 0);
2127
2128 #ifdef CONFIG_SMP
2129         ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
2130                                         NULL, vmstat_cpu_dead);
2131         if (ret < 0)
2132                 pr_err("vmstat: failed to register 'dead' hotplug state\n");
2133
2134         ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "mm/vmstat:online",
2135                                         vmstat_cpu_online,
2136                                         vmstat_cpu_down_prep);
2137         if (ret < 0)
2138                 pr_err("vmstat: failed to register 'online' hotplug state\n");
2139
2140         cpus_read_lock();
2141         init_cpu_node_state();
2142         cpus_read_unlock();
2143
2144         start_shepherd_timer();
2145 #endif
2146 #ifdef CONFIG_PROC_FS
2147         proc_create_seq("buddyinfo", 0444, NULL, &fragmentation_op);
2148         proc_create_seq("pagetypeinfo", 0400, NULL, &pagetypeinfo_op);
2149         proc_create_seq("vmstat", 0444, NULL, &vmstat_op);
2150         proc_create_seq("zoneinfo", 0444, NULL, &zoneinfo_op);
2151 #endif
2152 }
2153
2154 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
2155
2156 /*
2157  * Return an index indicating how much of the available free memory is
2158  * unusable for an allocation of the requested size.
2159  */
2160 static int unusable_free_index(unsigned int order,
2161                                 struct contig_page_info *info)
2162 {
2163         /* No free memory is interpreted as all free memory is unusable */
2164         if (info->free_pages == 0)
2165                 return 1000;
2166
2167         /*
2168          * Index should be a value between 0 and 1. Return a value to 3
2169          * decimal places.
2170          *
2171          * 0 => no fragmentation
2172          * 1 => high fragmentation
2173          */
2174         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
2175
2176 }
2177
2178 static void unusable_show_print(struct seq_file *m,
2179                                         pg_data_t *pgdat, struct zone *zone)
2180 {
2181         unsigned int order;
2182         int index;
2183         struct contig_page_info info;
2184
2185         seq_printf(m, "Node %d, zone %8s ",
2186                                 pgdat->node_id,
2187                                 zone->name);
2188         for (order = 0; order < NR_PAGE_ORDERS; ++order) {
2189                 fill_contig_page_info(zone, order, &info);
2190                 index = unusable_free_index(order, &info);
2191                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2192         }
2193
2194         seq_putc(m, '\n');
2195 }
2196
2197 /*
2198  * Display unusable free space index
2199  *
2200  * The unusable free space index measures how much of the available free
2201  * memory cannot be used to satisfy an allocation of a given size and is a
2202  * value between 0 and 1. The higher the value, the more of free memory is
2203  * unusable and by implication, the worse the external fragmentation is. This
2204  * can be expressed as a percentage by multiplying by 100.
2205  */
2206 static int unusable_show(struct seq_file *m, void *arg)
2207 {
2208         pg_data_t *pgdat = (pg_data_t *)arg;
2209
2210         /* check memoryless node */
2211         if (!node_state(pgdat->node_id, N_MEMORY))
2212                 return 0;
2213
2214         walk_zones_in_node(m, pgdat, true, false, unusable_show_print);
2215
2216         return 0;
2217 }
2218
2219 static const struct seq_operations unusable_sops = {
2220         .start  = frag_start,
2221         .next   = frag_next,
2222         .stop   = frag_stop,
2223         .show   = unusable_show,
2224 };
2225
2226 DEFINE_SEQ_ATTRIBUTE(unusable);
2227
2228 static void extfrag_show_print(struct seq_file *m,
2229                                         pg_data_t *pgdat, struct zone *zone)
2230 {
2231         unsigned int order;
2232         int index;
2233
2234         /* Alloc on stack as interrupts are disabled for zone walk */
2235         struct contig_page_info info;
2236
2237         seq_printf(m, "Node %d, zone %8s ",
2238                                 pgdat->node_id,
2239                                 zone->name);
2240         for (order = 0; order < NR_PAGE_ORDERS; ++order) {
2241                 fill_contig_page_info(zone, order, &info);
2242                 index = __fragmentation_index(order, &info);
2243                 seq_printf(m, "%2d.%03d ", index / 1000, index % 1000);
2244         }
2245
2246         seq_putc(m, '\n');
2247 }
2248
2249 /*
2250  * Display fragmentation index for orders that allocations would fail for
2251  */
2252 static int extfrag_show(struct seq_file *m, void *arg)
2253 {
2254         pg_data_t *pgdat = (pg_data_t *)arg;
2255
2256         walk_zones_in_node(m, pgdat, true, false, extfrag_show_print);
2257
2258         return 0;
2259 }
2260
2261 static const struct seq_operations extfrag_sops = {
2262         .start  = frag_start,
2263         .next   = frag_next,
2264         .stop   = frag_stop,
2265         .show   = extfrag_show,
2266 };
2267
2268 DEFINE_SEQ_ATTRIBUTE(extfrag);
2269
2270 static int __init extfrag_debug_init(void)
2271 {
2272         struct dentry *extfrag_debug_root;
2273
2274         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
2275
2276         debugfs_create_file("unusable_index", 0444, extfrag_debug_root, NULL,
2277                             &unusable_fops);
2278
2279         debugfs_create_file("extfrag_index", 0444, extfrag_debug_root, NULL,
2280                             &extfrag_fops);
2281
2282         return 0;
2283 }
2284
2285 module_init(extfrag_debug_init);
2286
2287 #endif
2288
2289 /*
2290  * Page metadata size (struct page and page_ext) in pages
2291  */
2292 static unsigned long early_perpage_metadata[MAX_NUMNODES] __meminitdata;
2293
2294 void __meminit mod_node_early_perpage_metadata(int nid, long delta)
2295 {
2296         early_perpage_metadata[nid] += delta;
2297 }
2298
2299 void __meminit store_early_perpage_metadata(void)
2300 {
2301         int nid;
2302         struct pglist_data *pgdat;
2303
2304         for_each_online_pgdat(pgdat) {
2305                 nid = pgdat->node_id;
2306                 mod_node_page_state(NODE_DATA(nid), NR_MEMMAP_BOOT,
2307                                     early_perpage_metadata[nid]);
2308         }
2309 }