mm/vmstat.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  linux/mm/vmstat.c
   4  *
   5  *  Manages VM statistics
   6  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   7  *
   8  *  zoned VM statistics
   9  *  Copyright (C) 2006 Silicon Graphics, Inc.,
  10  *              Christoph Lameter <christoph@lameter.com>
  11  *  Copyright (C) 2008-2014 Christoph Lameter
  12  */
  13 #include <linux/fs.h>
  14 #include <linux/mm.h>
  15 #include <linux/err.h>
  16 #include <linux/module.h>
  17 #include <linux/slab.h>
  18 #include <linux/cpu.h>
  19 #include <linux/cpumask.h>
  20 #include <linux/vmstat.h>
  21 #include <linux/proc_fs.h>
  22 #include <linux/seq_file.h>
  23 #include <linux/debugfs.h>
  24 #include <linux/sched.h>
  25 #include <linux/math64.h>
  26 #include <linux/writeback.h>
  27 #include <linux/compaction.h>
  28 #include <linux/mm_inline.h>
  29 #include <linux/page_ext.h>
  30 #include <linux/page_owner.h>
  31 #include <linux/sched/isolation.h>
  32
  33 #include "internal.h"
  34
  35 #ifdef CONFIG_NUMA
  36 int sysctl_vm_numa_stat = ENABLE_NUMA_STAT;
  37
  38 /* zero numa counters within a zone */
  39 static void zero_zone_numa_counters(struct zone *zone)
  40 {
  41         int item, cpu;
  42
  43         for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++) {
  44                 atomic_long_set(&zone->vm_numa_event[item], 0);
  45                 for_each_online_cpu(cpu) {
  46                         per_cpu_ptr(zone->per_cpu_zonestats, cpu)->vm_numa_event[item]
  47                                                 = 0;
  48                 }
  49         }
  50 }
  51
  52 /* zero numa counters of all the populated zones */
  53 static void zero_zones_numa_counters(void)
  54 {
  55         struct zone *zone;
  56
  57         for_each_populated_zone(zone)
  58                 zero_zone_numa_counters(zone);
  59 }
  60
  61 /* zero global numa counters */
  62 static void zero_global_numa_counters(void)
  63 {
  64         int item;
  65
  66         for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++)
  67                 atomic_long_set(&vm_numa_event[item], 0);
  68 }
  69
  70 static void invalid_numa_statistics(void)
  71 {
  72         zero_zones_numa_counters();
  73         zero_global_numa_counters();
  74 }
  75
  76 static DEFINE_MUTEX(vm_numa_stat_lock);
  77
  78 int sysctl_vm_numa_stat_handler(struct ctl_table *table, int write,
  79                 void *buffer, size_t *length, loff_t *ppos)
  80 {
  81         int ret, oldval;
  82
  83         mutex_lock(&vm_numa_stat_lock);
  84         if (write)
  85                 oldval = sysctl_vm_numa_stat;
  86         ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
  87         if (ret || !write)
  88                 goto out;
  89
  90         if (oldval == sysctl_vm_numa_stat)
  91                 goto out;
  92         else if (sysctl_vm_numa_stat == ENABLE_NUMA_STAT) {
  93                 static_branch_enable(&vm_numa_stat_key);
  94                 pr_info("enable numa statistics\n");
  95         } else {
  96                 static_branch_disable(&vm_numa_stat_key);
  97                 invalid_numa_statistics();
  98                 pr_info("disable numa statistics, and clear numa counters\n");
  99         }
 100
 101 out:
 102         mutex_unlock(&vm_numa_stat_lock);
 103         return ret;
 104 }
 105 #endif
 106
 107 #ifdef CONFIG_VM_EVENT_COUNTERS
 108 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
 109 EXPORT_PER_CPU_SYMBOL(vm_event_states);
 110
 111 static void sum_vm_events(unsigned long *ret)
 112 {
 113         int cpu;
 114         int i;
 115
 116         memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
 117
 118         for_each_online_cpu(cpu) {
 119                 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
 120
 121                 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
 122                         ret[i] += this->event[i];
 123         }
 124 }
 125
 126 /*
 127  * Accumulate the vm event counters across all CPUs.
 128  * The result is unavoidably approximate - it can change
 129  * during and after execution of this function.
 130 */
 131 void all_vm_events(unsigned long *ret)
 132 {
 133         cpus_read_lock();
 134         sum_vm_events(ret);
 135         cpus_read_unlock();
 136 }
 137 EXPORT_SYMBOL_GPL(all_vm_events);
 138
 139 /*
 140  * Fold the foreign cpu events into our own.
 141  *
 142  * This is adding to the events on one processor
 143  * but keeps the global counts constant.
 144  */
 145 void vm_events_fold_cpu(int cpu)
 146 {
 147         struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
 148         int i;
 149
 150         for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
 151                 count_vm_events(i, fold_state->event[i]);
 152                 fold_state->event[i] = 0;
 153         }
 154 }
 155
 156 #endif /* CONFIG_VM_EVENT_COUNTERS */
 157
 158 /*
 159  * Manage combined zone based / global counters
 160  *
 161  * vm_stat contains the global counters
 162  */
 163 atomic_long_t vm_zone_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
 164 atomic_long_t vm_node_stat[NR_VM_NODE_STAT_ITEMS] __cacheline_aligned_in_smp;
 165 atomic_long_t vm_numa_event[NR_VM_NUMA_EVENT_ITEMS] __cacheline_aligned_in_smp;
 166 EXPORT_SYMBOL(vm_zone_stat);
 167 EXPORT_SYMBOL(vm_node_stat);
 168
 169 #ifdef CONFIG_NUMA
 170 static void fold_vm_zone_numa_events(struct zone *zone)
 171 {
 172         unsigned long zone_numa_events[NR_VM_NUMA_EVENT_ITEMS] = { 0, };
 173         int cpu;
 174         enum numa_stat_item item;
 175
 176         for_each_online_cpu(cpu) {
 177                 struct per_cpu_zonestat *pzstats;
 178
 179                 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
 180                 for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++)
 181                         zone_numa_events[item] += xchg(&pzstats->vm_numa_event[item], 0);
 182         }
 183
 184         for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++)
 185                 zone_numa_event_add(zone_numa_events[item], zone, item);
 186 }
 187
 188 void fold_vm_numa_events(void)
 189 {
 190         struct zone *zone;
 191
 192         for_each_populated_zone(zone)
 193                 fold_vm_zone_numa_events(zone);
 194 }
 195 #endif
 196
 197 #ifdef CONFIG_SMP
 198
 199 int calculate_pressure_threshold(struct zone *zone)
 200 {
 201         int threshold;
 202         int watermark_distance;
 203
 204         /*
 205          * As vmstats are not up to date, there is drift between the estimated
 206          * and real values. For high thresholds and a high number of CPUs, it
 207          * is possible for the min watermark to be breached while the estimated
 208          * value looks fine. The pressure threshold is a reduced value such
 209          * that even the maximum amount of drift will not accidentally breach
 210          * the min watermark
 211          */
 212         watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
 213         threshold = max(1, (int)(watermark_distance / num_online_cpus()));
 214
 215         /*
 216          * Maximum threshold is 125
 217          */
 218         threshold = min(125, threshold);
 219
 220         return threshold;
 221 }
 222
 223 int calculate_normal_threshold(struct zone *zone)
 224 {
 225         int threshold;
 226         int mem;        /* memory in 128 MB units */
 227
 228         /*
 229          * The threshold scales with the number of processors and the amount
 230          * of memory per zone. More memory means that we can defer updates for
 231          * longer, more processors could lead to more contention.
 232          * fls() is used to have a cheap way of logarithmic scaling.
 233          *
 234          * Some sample thresholds:
 235          *
 236          * Threshold    Processors      (fls)   Zonesize        fls(mem)+1
 237          * ------------------------------------------------------------------
 238          * 8            1               1       0.9-1 GB        4
 239          * 16           2               2       0.9-1 GB        4
 240          * 20           2               2       1-2 GB          5
 241          * 24           2               2       2-4 GB          6
 242          * 28           2               2       4-8 GB          7
 243          * 32           2               2       8-16 GB         8
 244          * 4            2               2       <128M           1
 245          * 30           4               3       2-4 GB          5
 246          * 48           4               3       8-16 GB         8
 247          * 32           8               4       1-2 GB          4
 248          * 32           8               4       0.9-1GB         4
 249          * 10           16              5       <128M           1
 250          * 40           16              5       900M            4
 251          * 70           64              7       2-4 GB          5
 252          * 84           64              7       4-8 GB          6
 253          * 108          512             9       4-8 GB          6
 254          * 125          1024            10      8-16 GB         8
 255          * 125          1024            10      16-32 GB        9
 256          */
 257
 258         mem = zone_managed_pages(zone) >> (27 - PAGE_SHIFT);
 259
 260         threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
 261
 262         /*
 263          * Maximum threshold is 125
 264          */
 265         threshold = min(125, threshold);
 266
 267         return threshold;
 268 }
 269
 270 /*
 271  * Refresh the thresholds for each zone.
 272  */
 273 void refresh_zone_stat_thresholds(void)
 274 {
 275         struct pglist_data *pgdat;
 276         struct zone *zone;
 277         int cpu;
 278         int threshold;
 279
 280         /* Zero current pgdat thresholds */
 281         for_each_online_pgdat(pgdat) {
 282                 for_each_online_cpu(cpu) {
 283                         per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold = 0;
 284                 }
 285         }
 286
 287         for_each_populated_zone(zone) {
 288                 struct pglist_data *pgdat = zone->zone_pgdat;
 289                 unsigned long max_drift, tolerate_drift;
 290
 291                 threshold = calculate_normal_threshold(zone);
 292
 293                 for_each_online_cpu(cpu) {
 294                         int pgdat_threshold;
 295
 296                         per_cpu_ptr(zone->per_cpu_zonestats, cpu)->stat_threshold
 297                                                         = threshold;
 298
 299                         /* Base nodestat threshold on the largest populated zone. */
 300                         pgdat_threshold = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold;
 301                         per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold
 302                                 = max(threshold, pgdat_threshold);
 303                 }
 304
 305                 /*
 306                  * Only set percpu_drift_mark if there is a danger that
 307                  * NR_FREE_PAGES reports the low watermark is ok when in fact
 308                  * the min watermark could be breached by an allocation
 309                  */
 310                 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
 311                 max_drift = num_online_cpus() * threshold;
 312                 if (max_drift > tolerate_drift)
 313                         zone->percpu_drift_mark = high_wmark_pages(zone) +
 314                                         max_drift;
 315         }
 316 }
 317
 318 void set_pgdat_percpu_threshold(pg_data_t *pgdat,
 319                                 int (*calculate_pressure)(struct zone *))
 320 {
 321         struct zone *zone;
 322         int cpu;
 323         int threshold;
 324         int i;
 325
 326         for (i = 0; i < pgdat->nr_zones; i++) {
 327                 zone = &pgdat->node_zones[i];
 328                 if (!zone->percpu_drift_mark)
 329                         continue;
 330
 331                 threshold = (*calculate_pressure)(zone);
 332                 for_each_online_cpu(cpu)
 333                         per_cpu_ptr(zone->per_cpu_zonestats, cpu)->stat_threshold
 334                                                         = threshold;
 335         }
 336 }
 337
 338 /*
 339  * For use when we know that interrupts are disabled,
 340  * or when we know that preemption is disabled and that
 341  * particular counter cannot be updated from interrupt context.
 342  */
 343 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 344                            long delta)
 345 {
 346         struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
 347         s8 __percpu *p = pcp->vm_stat_diff + item;
 348         long x;
 349         long t;
 350
 351         /*
 352          * Accurate vmstat updates require a RMW. On !PREEMPT_RT kernels,
 353          * atomicity is provided by IRQs being disabled -- either explicitly
 354          * or via local_lock_irq. On PREEMPT_RT, local_lock_irq only disables
 355          * CPU migrations and preemption potentially corrupts a counter so
 356          * disable preemption.
 357          */
 358         preempt_disable_nested();
 359
 360         x = delta + __this_cpu_read(*p);
 361
 362         t = __this_cpu_read(pcp->stat_threshold);
 363
 364         if (unlikely(abs(x) > t)) {
 365                 zone_page_state_add(x, zone, item);
 366                 x = 0;
 367         }
 368         __this_cpu_write(*p, x);
 369
 370         preempt_enable_nested();
 371 }
 372 EXPORT_SYMBOL(__mod_zone_page_state);
 373
 374 void __mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
 375                                 long delta)
 376 {
 377         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 378         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 379         long x;
 380         long t;
 381
 382         if (vmstat_item_in_bytes(item)) {
 383                 /*
 384                  * Only cgroups use subpage accounting right now; at
 385                  * the global level, these items still change in
 386                  * multiples of whole pages. Store them as pages
 387                  * internally to keep the per-cpu counters compact.
 388                  */
 389                 VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
 390                 delta >>= PAGE_SHIFT;
 391         }
 392
 393         /* See __mod_node_page_state */
 394         preempt_disable_nested();
 395
 396         x = delta + __this_cpu_read(*p);
 397
 398         t = __this_cpu_read(pcp->stat_threshold);
 399
 400         if (unlikely(abs(x) > t)) {
 401                 node_page_state_add(x, pgdat, item);
 402                 x = 0;
 403         }
 404         __this_cpu_write(*p, x);
 405
 406         preempt_enable_nested();
 407 }
 408 EXPORT_SYMBOL(__mod_node_page_state);
 409
 410 /*
 411  * Optimized increment and decrement functions.
 412  *
 413  * These are only for a single page and therefore can take a struct page *
 414  * argument instead of struct zone *. This allows the inclusion of the code
 415  * generated for page_zone(page) into the optimized functions.
 416  *
 417  * No overflow check is necessary and therefore the differential can be
 418  * incremented or decremented in place which may allow the compilers to
 419  * generate better code.
 420  * The increment or decrement is known and therefore one boundary check can
 421  * be omitted.
 422  *
 423  * NOTE: These functions are very performance sensitive. Change only
 424  * with care.
 425  *
 426  * Some processors have inc/dec instructions that are atomic vs an interrupt.
 427  * However, the code must first determine the differential location in a zone
 428  * based on the processor number and then inc/dec the counter. There is no
 429  * guarantee without disabling preemption that the processor will not change
 430  * in between and therefore the atomicity vs. interrupt cannot be exploited
 431  * in a useful way here.
 432  */
 433 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
 434 {
 435         struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
 436         s8 __percpu *p = pcp->vm_stat_diff + item;
 437         s8 v, t;
 438
 439         /* See __mod_node_page_state */
 440         preempt_disable_nested();
 441
 442         v = __this_cpu_inc_return(*p);
 443         t = __this_cpu_read(pcp->stat_threshold);
 444         if (unlikely(v > t)) {
 445                 s8 overstep = t >> 1;
 446
 447                 zone_page_state_add(v + overstep, zone, item);
 448                 __this_cpu_write(*p, -overstep);
 449         }
 450
 451         preempt_enable_nested();
 452 }
 453
 454 void __inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 455 {
 456         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 457         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 458         s8 v, t;
 459
 460         VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
 461
 462         /* See __mod_node_page_state */
 463         preempt_disable_nested();
 464
 465         v = __this_cpu_inc_return(*p);
 466         t = __this_cpu_read(pcp->stat_threshold);
 467         if (unlikely(v > t)) {
 468                 s8 overstep = t >> 1;
 469
 470                 node_page_state_add(v + overstep, pgdat, item);
 471                 __this_cpu_write(*p, -overstep);
 472         }
 473
 474         preempt_enable_nested();
 475 }
 476
 477 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
 478 {
 479         __inc_zone_state(page_zone(page), item);
 480 }
 481 EXPORT_SYMBOL(__inc_zone_page_state);
 482
 483 void __inc_node_page_state(struct page *page, enum node_stat_item item)
 484 {
 485         __inc_node_state(page_pgdat(page), item);
 486 }
 487 EXPORT_SYMBOL(__inc_node_page_state);
 488
 489 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
 490 {
 491         struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
 492         s8 __percpu *p = pcp->vm_stat_diff + item;
 493         s8 v, t;
 494
 495         /* See __mod_node_page_state */
 496         preempt_disable_nested();
 497
 498         v = __this_cpu_dec_return(*p);
 499         t = __this_cpu_read(pcp->stat_threshold);
 500         if (unlikely(v < - t)) {
 501                 s8 overstep = t >> 1;
 502
 503                 zone_page_state_add(v - overstep, zone, item);
 504                 __this_cpu_write(*p, overstep);
 505         }
 506
 507         preempt_enable_nested();
 508 }
 509
 510 void __dec_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 511 {
 512         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 513         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 514         s8 v, t;
 515
 516         VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
 517
 518         /* See __mod_node_page_state */
 519         preempt_disable_nested();
 520
 521         v = __this_cpu_dec_return(*p);
 522         t = __this_cpu_read(pcp->stat_threshold);
 523         if (unlikely(v < - t)) {
 524                 s8 overstep = t >> 1;
 525
 526                 node_page_state_add(v - overstep, pgdat, item);
 527                 __this_cpu_write(*p, overstep);
 528         }
 529
 530         preempt_enable_nested();
 531 }
 532
 533 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
 534 {
 535         __dec_zone_state(page_zone(page), item);
 536 }
 537 EXPORT_SYMBOL(__dec_zone_page_state);
 538
 539 void __dec_node_page_state(struct page *page, enum node_stat_item item)
 540 {
 541         __dec_node_state(page_pgdat(page), item);
 542 }
 543 EXPORT_SYMBOL(__dec_node_page_state);
 544
 545 #ifdef CONFIG_HAVE_CMPXCHG_LOCAL
 546 /*
 547  * If we have cmpxchg_local support then we do not need to incur the overhead
 548  * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
 549  *
 550  * mod_state() modifies the zone counter state through atomic per cpu
 551  * operations.
 552  *
 553  * Overstep mode specifies how overstep should handled:
 554  *     0       No overstepping
 555  *     1       Overstepping half of threshold
 556  *     -1      Overstepping minus half of threshold
 557 */
 558 static inline void mod_zone_state(struct zone *zone,
 559        enum zone_stat_item item, long delta, int overstep_mode)
 560 {
 561         struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
 562         s8 __percpu *p = pcp->vm_stat_diff + item;
 563         long o, n, t, z;
 564
 565         do {
 566                 z = 0;  /* overflow to zone counters */
 567
 568                 /*
 569                  * The fetching of the stat_threshold is racy. We may apply
 570                  * a counter threshold to the wrong the cpu if we get
 571                  * rescheduled while executing here. However, the next
 572                  * counter update will apply the threshold again and
 573                  * therefore bring the counter under the threshold again.
 574                  *
 575                  * Most of the time the thresholds are the same anyways
 576                  * for all cpus in a zone.
 577                  */
 578                 t = this_cpu_read(pcp->stat_threshold);
 579
 580                 o = this_cpu_read(*p);
 581                 n = delta + 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_cmpxchg(*p, o, n) != o);
 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 o, n, t, z;
 621
 622         if (vmstat_item_in_bytes(item)) {
 623                 /*
 624                  * Only cgroups use subpage accounting right now; at
 625                  * the global level, these items still change in
 626                  * multiples of whole pages. Store them as pages
 627                  * internally to keep the per-cpu counters compact.
 628                  */
 629                 VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
 630                 delta >>= PAGE_SHIFT;
 631         }
 632
 633         do {
 634                 z = 0;  /* overflow to node counters */
 635
 636                 /*
 637                  * The fetching of the stat_threshold is racy. We may apply
 638                  * a counter threshold to the wrong the cpu if we get
 639                  * rescheduled while executing here. However, the next
 640                  * counter update will apply the threshold again and
 641                  * therefore bring the counter under the threshold again.
 642                  *
 643                  * Most of the time the thresholds are the same anyways
 644                  * for all cpus in a node.
 645                  */
 646                 t = this_cpu_read(pcp->stat_threshold);
 647
 648                 o = this_cpu_read(*p);
 649                 n = delta + o;
 650
 651                 if (abs(n) > t) {
 652                         int os = overstep_mode * (t >> 1) ;
 653
 654                         /* Overflow must be added to node counters */
 655                         z = n + os;
 656                         n = -os;
 657                 }
 658         } while (this_cpu_cmpxchg(*p, o, n) != o);
 659
 660         if (z)
 661                 node_page_state_add(z, pgdat, item);
 662 }
 663
 664 void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
 665                                         long delta)
 666 {
 667         mod_node_state(pgdat, item, delta, 0);
 668 }
 669 EXPORT_SYMBOL(mod_node_page_state);
 670
 671 void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 672 {
 673         mod_node_state(pgdat, item, 1, 1);
 674 }
 675
 676 void inc_node_page_state(struct page *page, enum node_stat_item item)
 677 {
 678         mod_node_state(page_pgdat(page), item, 1, 1);
 679 }
 680 EXPORT_SYMBOL(inc_node_page_state);
 681
 682 void dec_node_page_state(struct page *page, enum node_stat_item item)
 683 {
 684         mod_node_state(page_pgdat(page), item, -1, -1);
 685 }
 686 EXPORT_SYMBOL(dec_node_page_state);
 687 #else
 688 /*
 689  * Use interrupt disable to serialize counter updates
 690  */
 691 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 692                          long delta)
 693 {
 694         unsigned long flags;
 695
 696         local_irq_save(flags);
 697         __mod_zone_page_state(zone, item, delta);
 698         local_irq_restore(flags);
 699 }
 700 EXPORT_SYMBOL(mod_zone_page_state);
 701
 702 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 703 {
 704         unsigned long flags;
 705         struct zone *zone;
 706
 707         zone = page_zone(page);
 708         local_irq_save(flags);
 709         __inc_zone_state(zone, item);
 710         local_irq_restore(flags);
 711 }
 712 EXPORT_SYMBOL(inc_zone_page_state);
 713
 714 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 715 {
 716         unsigned long flags;
 717
 718         local_irq_save(flags);
 719         __dec_zone_page_state(page, item);
 720         local_irq_restore(flags);
 721 }
 722 EXPORT_SYMBOL(dec_zone_page_state);
 723
 724 void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 725 {
 726         unsigned long flags;
 727
 728         local_irq_save(flags);
 729         __inc_node_state(pgdat, item);
 730         local_irq_restore(flags);
 731 }
 732 EXPORT_SYMBOL(inc_node_state);
 733
 734 void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
 735                                         long delta)
 736 {
 737         unsigned long flags;
 738
 739         local_irq_save(flags);
 740         __mod_node_page_state(pgdat, item, delta);
 741         local_irq_restore(flags);
 742 }
 743 EXPORT_SYMBOL(mod_node_page_state);
 744
 745 void inc_node_page_state(struct page *page, enum node_stat_item item)
 746 {
 747         unsigned long flags;
 748         struct pglist_data *pgdat;
 749
 750         pgdat = page_pgdat(page);
 751         local_irq_save(flags);
 752         __inc_node_state(pgdat, item);
 753         local_irq_restore(flags);
 754 }
 755 EXPORT_SYMBOL(inc_node_page_state);
 756
 757 void dec_node_page_state(struct page *page, enum node_stat_item item)
 758 {
 759         unsigned long flags;
 760
 761         local_irq_save(flags);
 762         __dec_node_page_state(page, item);
 763         local_irq_restore(flags);
 764 }
 765 EXPORT_SYMBOL(dec_node_page_state);
 766 #endif
 767
 768 /*
 769  * Fold a differential into the global counters.
 770  * Returns the number of counters updated.
 771  */
 772 static int fold_diff(int *zone_diff, int *node_diff)
 773 {
 774         int i;
 775         int changes = 0;
 776
 777         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 778                 if (zone_diff[i]) {
 779                         atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
 780                         changes++;
 781         }
 782
 783         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 784                 if (node_diff[i]) {
 785                         atomic_long_add(node_diff[i], &vm_node_stat[i]);
 786                         changes++;
 787         }
 788         return changes;
 789 }
 790
 791 /*
 792  * Update the zone counters for the current cpu.
 793  *
 794  * Note that refresh_cpu_vm_stats strives to only access
 795  * node local memory. The per cpu pagesets on remote zones are placed
 796  * in the memory local to the processor using that pageset. So the
 797  * loop over all zones will access a series of cachelines local to
 798  * the processor.
 799  *
 800  * The call to zone_page_state_add updates the cachelines with the
 801  * statistics in the remote zone struct as well as the global cachelines
 802  * with the global counters. These could cause remote node cache line
 803  * bouncing and will have to be only done when necessary.
 804  *
 805  * The function returns the number of global counters updated.
 806  */
 807 static int refresh_cpu_vm_stats(bool do_pagesets)
 808 {
 809         struct pglist_data *pgdat;
 810         struct zone *zone;
 811         int i;
 812         int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 813         int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
 814         int changes = 0;
 815
 816         for_each_populated_zone(zone) {
 817                 struct per_cpu_zonestat __percpu *pzstats = zone->per_cpu_zonestats;
 818 #ifdef CONFIG_NUMA
 819                 struct per_cpu_pages __percpu *pcp = zone->per_cpu_pageset;
 820 #endif
 821
 822                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
 823                         int v;
 824
 825                         v = this_cpu_xchg(pzstats->vm_stat_diff[i], 0);
 826                         if (v) {
 827
 828                                 atomic_long_add(v, &zone->vm_stat[i]);
 829                                 global_zone_diff[i] += v;
 830 #ifdef CONFIG_NUMA
 831                                 /* 3 seconds idle till flush */
 832                                 __this_cpu_write(pcp->expire, 3);
 833 #endif
 834                         }
 835                 }
 836 #ifdef CONFIG_NUMA
 837
 838                 if (do_pagesets) {
 839                         cond_resched();
 840                         /*
 841                          * Deal with draining the remote pageset of this
 842                          * processor
 843                          *
 844                          * Check if there are pages remaining in this pageset
 845                          * if not then there is nothing to expire.
 846                          */
 847                         if (!__this_cpu_read(pcp->expire) ||
 848                                !__this_cpu_read(pcp->count))
 849                                 continue;
 850
 851                         /*
 852                          * We never drain zones local to this processor.
 853                          */
 854                         if (zone_to_nid(zone) == numa_node_id()) {
 855                                 __this_cpu_write(pcp->expire, 0);
 856                                 continue;
 857                         }
 858
 859                         if (__this_cpu_dec_return(pcp->expire))
 860                                 continue;
 861
 862                         if (__this_cpu_read(pcp->count)) {
 863                                 drain_zone_pages(zone, this_cpu_ptr(pcp));
 864                                 changes++;
 865                         }
 866                 }
 867 #endif
 868         }
 869
 870         for_each_online_pgdat(pgdat) {
 871                 struct per_cpu_nodestat __percpu *p = pgdat->per_cpu_nodestats;
 872
 873                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
 874                         int v;
 875
 876                         v = this_cpu_xchg(p->vm_node_stat_diff[i], 0);
 877                         if (v) {
 878                                 atomic_long_add(v, &pgdat->vm_stat[i]);
 879                                 global_node_diff[i] += v;
 880                         }
 881                 }
 882         }
 883
 884         changes += fold_diff(global_zone_diff, global_node_diff);
 885         return changes;
 886 }
 887
 888 /*
 889  * Fold the data for an offline cpu into the global array.
 890  * There cannot be any access by the offline cpu and therefore
 891  * synchronization is simplified.
 892  */
 893 void cpu_vm_stats_fold(int cpu)
 894 {
 895         struct pglist_data *pgdat;
 896         struct zone *zone;
 897         int i;
 898         int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 899         int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
 900
 901         for_each_populated_zone(zone) {
 902                 struct per_cpu_zonestat *pzstats;
 903
 904                 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
 905
 906                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
 907                         if (pzstats->vm_stat_diff[i]) {
 908                                 int v;
 909
 910                                 v = pzstats->vm_stat_diff[i];
 911                                 pzstats->vm_stat_diff[i] = 0;
 912                                 atomic_long_add(v, &zone->vm_stat[i]);
 913                                 global_zone_diff[i] += v;
 914                         }
 915                 }
 916 #ifdef CONFIG_NUMA
 917                 for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++) {
 918                         if (pzstats->vm_numa_event[i]) {
 919                                 unsigned long v;
 920
 921                                 v = pzstats->vm_numa_event[i];
 922                                 pzstats->vm_numa_event[i] = 0;
 923                                 zone_numa_event_add(v, zone, i);
 924                         }
 925                 }
 926 #endif
 927         }
 928
 929         for_each_online_pgdat(pgdat) {
 930                 struct per_cpu_nodestat *p;
 931
 932                 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
 933
 934                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 935                         if (p->vm_node_stat_diff[i]) {
 936                                 int v;
 937
 938                                 v = p->vm_node_stat_diff[i];
 939                                 p->vm_node_stat_diff[i] = 0;
 940                                 atomic_long_add(v, &pgdat->vm_stat[i]);
 941                                 global_node_diff[i] += v;
 942                         }
 943         }
 944
 945         fold_diff(global_zone_diff, global_node_diff);
 946 }
 947
 948 /*
 949  * this is only called if !populated_zone(zone), which implies no other users of
 950  * pset->vm_stat_diff[] exist.
 951  */
 952 void drain_zonestat(struct zone *zone, struct per_cpu_zonestat *pzstats)
 953 {
 954         unsigned long v;
 955         int i;
 956
 957         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
 958                 if (pzstats->vm_stat_diff[i]) {
 959                         v = pzstats->vm_stat_diff[i];
 960                         pzstats->vm_stat_diff[i] = 0;
 961                         zone_page_state_add(v, zone, i);
 962                 }
 963         }
 964
 965 #ifdef CONFIG_NUMA
 966         for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++) {
 967                 if (pzstats->vm_numa_event[i]) {
 968                         v = pzstats->vm_numa_event[i];
 969                         pzstats->vm_numa_event[i] = 0;
 970                         zone_numa_event_add(v, zone, i);
 971                 }
 972         }
 973 #endif
 974 }
 975 #endif
 976
 977 #ifdef CONFIG_NUMA
 978 /*
 979  * Determine the per node value of a stat item. This function
 980  * is called frequently in a NUMA machine, so try to be as
 981  * frugal as possible.
 982  */
 983 unsigned long sum_zone_node_page_state(int node,
 984                                  enum zone_stat_item item)
 985 {
 986         struct zone *zones = NODE_DATA(node)->node_zones;
 987         int i;
 988         unsigned long count = 0;
 989
 990         for (i = 0; i < MAX_NR_ZONES; i++)
 991                 count += zone_page_state(zones + i, item);
 992
 993         return count;
 994 }
 995
 996 /* Determine the per node value of a numa stat item. */
 997 unsigned long sum_zone_numa_event_state(int node,
 998                                  enum numa_stat_item item)
 999 {
1000         struct zone *zones = NODE_DATA(node)->node_zones;
1001         unsigned long count = 0;
1002         int i;
1003
1004         for (i = 0; i < MAX_NR_ZONES; i++)
1005                 count += zone_numa_event_state(zones + i, item);
1006
1007         return count;
1008 }
1009
1010 /*
1011  * Determine the per node value of a stat item.
1012  */
1013 unsigned long node_page_state_pages(struct pglist_data *pgdat,
1014                                     enum node_stat_item item)
1015 {
1016         long x = atomic_long_read(&pgdat->vm_stat[item]);
1017 #ifdef CONFIG_SMP
1018         if (x < 0)
1019                 x = 0;
1020 #endif
1021         return x;
1022 }
1023
1024 unsigned long node_page_state(struct pglist_data *pgdat,
1025                               enum node_stat_item item)
1026 {
1027         VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
1028
1029         return node_page_state_pages(pgdat, item);
1030 }
1031 #endif
1032
1033 #ifdef CONFIG_COMPACTION
1034
1035 struct contig_page_info {
1036         unsigned long free_pages;
1037         unsigned long free_blocks_total;
1038         unsigned long free_blocks_suitable;
1039 };
1040
1041 /*
1042  * Calculate the number of free pages in a zone, how many contiguous
1043  * pages are free and how many are large enough to satisfy an allocation of
1044  * the target size. Note that this function makes no attempt to estimate
1045  * how many suitable free blocks there *might* be if MOVABLE pages were
1046  * migrated. Calculating that is possible, but expensive and can be
1047  * figured out from userspace
1048  */
1049 static void fill_contig_page_info(struct zone *zone,
1050                                 unsigned int suitable_order,
1051                                 struct contig_page_info *info)
1052 {
1053         unsigned int order;
1054
1055         info->free_pages = 0;
1056         info->free_blocks_total = 0;
1057         info->free_blocks_suitable = 0;
1058
1059         for (order = 0; order <= MAX_ORDER; order++) {
1060                 unsigned long blocks;
1061
1062                 /*
1063                  * Count number of free blocks.
1064                  *
1065                  * Access to nr_free is lockless as nr_free is used only for
1066                  * diagnostic purposes. Use data_race to avoid KCSAN warning.
1067                  */
1068                 blocks = data_race(zone->free_area[order].nr_free);
1069                 info->free_blocks_total += blocks;
1070
1071                 /* Count free base pages */
1072                 info->free_pages += blocks << order;
1073
1074                 /* Count the suitable free blocks */
1075                 if (order >= suitable_order)
1076                         info->free_blocks_suitable += blocks <<
1077                                                 (order - suitable_order);
1078         }
1079 }
1080
1081 /*
1082  * A fragmentation index only makes sense if an allocation of a requested
1083  * size would fail. If that is true, the fragmentation index indicates
1084  * whether external fragmentation or a lack of memory was the problem.
1085  * The value can be used to determine if page reclaim or compaction
1086  * should be used
1087  */
1088 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
1089 {
1090         unsigned long requested = 1UL << order;
1091
1092         if (WARN_ON_ONCE(order > MAX_ORDER))
1093                 return 0;
1094
1095         if (!info->free_blocks_total)
1096                 return 0;
1097
1098         /* Fragmentation index only makes sense when a request would fail */
1099         if (info->free_blocks_suitable)
1100                 return -1000;
1101
1102         /*
1103          * Index is between 0 and 1 so return within 3 decimal places
1104          *
1105          * 0 => allocation would fail due to lack of memory
1106          * 1 => allocation would fail due to fragmentation
1107          */
1108         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
1109 }
1110
1111 /*
1112  * Calculates external fragmentation within a zone wrt the given order.
1113  * It is defined as the percentage of pages found in blocks of size
1114  * less than 1 << order. It returns values in range [0, 100].
1115  */
1116 unsigned int extfrag_for_order(struct zone *zone, unsigned int order)
1117 {
1118         struct contig_page_info info;
1119
1120         fill_contig_page_info(zone, order, &info);
1121         if (info.free_pages == 0)
1122                 return 0;
1123
1124         return div_u64((info.free_pages -
1125                         (info.free_blocks_suitable << order)) * 100,
1126                         info.free_pages);
1127 }
1128
1129 /* Same as __fragmentation index but allocs contig_page_info on stack */
1130 int fragmentation_index(struct zone *zone, unsigned int order)
1131 {
1132         struct contig_page_info info;
1133
1134         fill_contig_page_info(zone, order, &info);
1135         return __fragmentation_index(order, &info);
1136 }
1137 #endif
1138
1139 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || \
1140     defined(CONFIG_NUMA) || defined(CONFIG_MEMCG)
1141 #ifdef CONFIG_ZONE_DMA
1142 #define TEXT_FOR_DMA(xx) xx "_dma",
1143 #else
1144 #define TEXT_FOR_DMA(xx)
1145 #endif
1146
1147 #ifdef CONFIG_ZONE_DMA32
1148 #define TEXT_FOR_DMA32(xx) xx "_dma32",
1149 #else
1150 #define TEXT_FOR_DMA32(xx)
1151 #endif
1152
1153 #ifdef CONFIG_HIGHMEM
1154 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
1155 #else
1156 #define TEXT_FOR_HIGHMEM(xx)
1157 #endif
1158
1159 #ifdef CONFIG_ZONE_DEVICE
1160 #define TEXT_FOR_DEVICE(xx) xx "_device",
1161 #else
1162 #define TEXT_FOR_DEVICE(xx)
1163 #endif
1164
1165 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1166                                         TEXT_FOR_HIGHMEM(xx) xx "_movable", \
1167                                         TEXT_FOR_DEVICE(xx)
1168
1169 const char * const vmstat_text[] = {
1170         /* enum zone_stat_item counters */
1171         "nr_free_pages",
1172         "nr_zone_inactive_anon",
1173         "nr_zone_active_anon",
1174         "nr_zone_inactive_file",
1175         "nr_zone_active_file",
1176         "nr_zone_unevictable",
1177         "nr_zone_write_pending",
1178         "nr_mlock",
1179         "nr_bounce",
1180 #if IS_ENABLED(CONFIG_ZSMALLOC)
1181         "nr_zspages",
1182 #endif
1183         "nr_free_cma",
1184 #ifdef CONFIG_UNACCEPTED_MEMORY
1185         "nr_unaccepted",
1186 #endif
1187
1188         /* enum numa_stat_item counters */
1189 #ifdef CONFIG_NUMA
1190         "numa_hit",
1191         "numa_miss",
1192         "numa_foreign",
1193         "numa_interleave",
1194         "numa_local",
1195         "numa_other",
1196 #endif
1197
1198         /* enum node_stat_item counters */
1199         "nr_inactive_anon",
1200         "nr_active_anon",
1201         "nr_inactive_file",
1202         "nr_active_file",
1203         "nr_unevictable",
1204         "nr_slab_reclaimable",
1205         "nr_slab_unreclaimable",
1206         "nr_isolated_anon",
1207         "nr_isolated_file",
1208         "workingset_nodes",
1209         "workingset_refault_anon",
1210         "workingset_refault_file",
1211         "workingset_activate_anon",
1212         "workingset_activate_file",
1213         "workingset_restore_anon",
1214         "workingset_restore_file",
1215         "workingset_nodereclaim",
1216         "nr_anon_pages",
1217         "nr_mapped",
1218         "nr_file_pages",
1219         "nr_dirty",
1220         "nr_writeback",
1221         "nr_writeback_temp",
1222         "nr_shmem",
1223         "nr_shmem_hugepages",
1224         "nr_shmem_pmdmapped",
1225         "nr_file_hugepages",
1226         "nr_file_pmdmapped",
1227         "nr_anon_transparent_hugepages",
1228         "nr_vmscan_write",
1229         "nr_vmscan_immediate_reclaim",
1230         "nr_dirtied",
1231         "nr_written",
1232         "nr_throttled_written",
1233         "nr_kernel_misc_reclaimable",
1234         "nr_foll_pin_acquired",
1235         "nr_foll_pin_released",
1236         "nr_kernel_stack",
1237 #if IS_ENABLED(CONFIG_SHADOW_CALL_STACK)
1238         "nr_shadow_call_stack",
1239 #endif
1240         "nr_page_table_pages",
1241         "nr_sec_page_table_pages",
1242 #ifdef CONFIG_SWAP
1243         "nr_swapcached",
1244 #endif
1245 #ifdef CONFIG_NUMA_BALANCING
1246         "pgpromote_success",
1247         "pgpromote_candidate",
1248 #endif
1249
1250         /* enum writeback_stat_item counters */
1251         "nr_dirty_threshold",
1252         "nr_dirty_background_threshold",
1253
1254 #if defined(CONFIG_VM_EVENT_COUNTERS) || defined(CONFIG_MEMCG)
1255         /* enum vm_event_item counters */
1256         "pgpgin",
1257         "pgpgout",
1258         "pswpin",
1259         "pswpout",
1260
1261         TEXTS_FOR_ZONES("pgalloc")
1262         TEXTS_FOR_ZONES("allocstall")
1263         TEXTS_FOR_ZONES("pgskip")
1264
1265         "pgfree",
1266         "pgactivate",
1267         "pgdeactivate",
1268         "pglazyfree",
1269
1270         "pgfault",
1271         "pgmajfault",
1272         "pglazyfreed",
1273
1274         "pgrefill",
1275         "pgreuse",
1276         "pgsteal_kswapd",
1277         "pgsteal_direct",
1278         "pgsteal_khugepaged",
1279         "pgdemote_kswapd",
1280         "pgdemote_direct",
1281         "pgdemote_khugepaged",
1282         "pgscan_kswapd",
1283         "pgscan_direct",
1284         "pgscan_khugepaged",
1285         "pgscan_direct_throttle",
1286         "pgscan_anon",
1287         "pgscan_file",
1288         "pgsteal_anon",
1289         "pgsteal_file",
1290
1291 #ifdef CONFIG_NUMA
1292         "zone_reclaim_failed",
1293 #endif
1294         "pginodesteal",
1295         "slabs_scanned",
1296         "kswapd_inodesteal",
1297         "kswapd_low_wmark_hit_quickly",
1298         "kswapd_high_wmark_hit_quickly",
1299         "pageoutrun",
1300
1301         "pgrotated",
1302
1303         "drop_pagecache",
1304         "drop_slab",
1305         "oom_kill",
1306
1307 #ifdef CONFIG_NUMA_BALANCING
1308         "numa_pte_updates",
1309         "numa_huge_pte_updates",
1310         "numa_hint_faults",
1311         "numa_hint_faults_local",
1312         "numa_pages_migrated",
1313 #endif
1314 #ifdef CONFIG_MIGRATION
1315         "pgmigrate_success",
1316         "pgmigrate_fail",
1317         "thp_migration_success",
1318         "thp_migration_fail",
1319         "thp_migration_split",
1320 #endif
1321 #ifdef CONFIG_COMPACTION
1322         "compact_migrate_scanned",
1323         "compact_free_scanned",
1324         "compact_isolated",
1325         "compact_stall",
1326         "compact_fail",
1327         "compact_success",
1328         "compact_daemon_wake",
1329         "compact_daemon_migrate_scanned",
1330         "compact_daemon_free_scanned",
1331 #endif
1332
1333 #ifdef CONFIG_HUGETLB_PAGE
1334         "htlb_buddy_alloc_success",
1335         "htlb_buddy_alloc_fail",
1336 #endif
1337 #ifdef CONFIG_CMA
1338         "cma_alloc_success",
1339         "cma_alloc_fail",
1340 #endif
1341         "unevictable_pgs_culled",
1342         "unevictable_pgs_scanned",
1343         "unevictable_pgs_rescued",
1344         "unevictable_pgs_mlocked",
1345         "unevictable_pgs_munlocked",
1346         "unevictable_pgs_cleared",
1347         "unevictable_pgs_stranded",
1348
1349 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1350         "thp_fault_alloc",
1351         "thp_fault_fallback",
1352         "thp_fault_fallback_charge",
1353         "thp_collapse_alloc",
1354         "thp_collapse_alloc_failed",
1355         "thp_file_alloc",
1356         "thp_file_fallback",
1357         "thp_file_fallback_charge",
1358         "thp_file_mapped",
1359         "thp_split_page",
1360         "thp_split_page_failed",
1361         "thp_deferred_split_page",
1362         "thp_split_pmd",
1363         "thp_scan_exceed_none_pte",
1364         "thp_scan_exceed_swap_pte",
1365         "thp_scan_exceed_share_pte",
1366 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1367         "thp_split_pud",
1368 #endif
1369         "thp_zero_page_alloc",
1370         "thp_zero_page_alloc_failed",
1371         "thp_swpout",
1372         "thp_swpout_fallback",
1373 #endif
1374 #ifdef CONFIG_MEMORY_BALLOON
1375         "balloon_inflate",
1376         "balloon_deflate",
1377 #ifdef CONFIG_BALLOON_COMPACTION
1378         "balloon_migrate",
1379 #endif
1380 #endif /* CONFIG_MEMORY_BALLOON */
1381 #ifdef CONFIG_DEBUG_TLBFLUSH
1382         "nr_tlb_remote_flush",
1383         "nr_tlb_remote_flush_received",
1384         "nr_tlb_local_flush_all",
1385         "nr_tlb_local_flush_one",
1386 #endif /* CONFIG_DEBUG_TLBFLUSH */
1387
1388 #ifdef CONFIG_SWAP
1389         "swap_ra",
1390         "swap_ra_hit",
1391 #ifdef CONFIG_KSM
1392         "ksm_swpin_copy",
1393 #endif
1394 #endif
1395 #ifdef CONFIG_KSM
1396         "cow_ksm",
1397 #endif
1398 #ifdef CONFIG_ZSWAP
1399         "zswpin",
1400         "zswpout",
1401 #endif
1402 #ifdef CONFIG_X86
1403         "direct_map_level2_splits",
1404         "direct_map_level3_splits",
1405 #endif
1406 #ifdef CONFIG_PER_VMA_LOCK_STATS
1407         "vma_lock_success",
1408         "vma_lock_abort",
1409         "vma_lock_retry",
1410         "vma_lock_miss",
1411 #endif
1412 #endif /* CONFIG_VM_EVENT_COUNTERS || CONFIG_MEMCG */
1413 };
1414 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA || CONFIG_MEMCG */
1415
1416 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1417      defined(CONFIG_PROC_FS)
1418 static void *frag_start(struct seq_file *m, loff_t *pos)
1419 {
1420         pg_data_t *pgdat;
1421         loff_t node = *pos;
1422
1423         for (pgdat = first_online_pgdat();
1424              pgdat && node;
1425              pgdat = next_online_pgdat(pgdat))
1426                 --node;
1427
1428         return pgdat;
1429 }
1430
1431 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1432 {
1433         pg_data_t *pgdat = (pg_data_t *)arg;
1434
1435         (*pos)++;
1436         return next_online_pgdat(pgdat);
1437 }
1438
1439 static void frag_stop(struct seq_file *m, void *arg)
1440 {
1441 }
1442
1443 /*
1444  * Walk zones in a node and print using a callback.
1445  * If @assert_populated is true, only use callback for zones that are populated.
1446  */
1447 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1448                 bool assert_populated, bool nolock,
1449                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1450 {
1451         struct zone *zone;
1452         struct zone *node_zones = pgdat->node_zones;
1453         unsigned long flags;
1454
1455         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1456                 if (assert_populated && !populated_zone(zone))
1457                         continue;
1458
1459                 if (!nolock)
1460                         spin_lock_irqsave(&zone->lock, flags);
1461                 print(m, pgdat, zone);
1462                 if (!nolock)
1463                         spin_unlock_irqrestore(&zone->lock, flags);
1464         }
1465 }
1466 #endif
1467
1468 #ifdef CONFIG_PROC_FS
1469 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1470                                                 struct zone *zone)
1471 {
1472         int order;
1473
1474         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1475         for (order = 0; order <= MAX_ORDER; ++order)
1476                 /*
1477                  * Access to nr_free is lockless as nr_free is used only for
1478                  * printing purposes. Use data_race to avoid KCSAN warning.
1479                  */
1480                 seq_printf(m, "%6lu ", data_race(zone->free_area[order].nr_free));
1481         seq_putc(m, '\n');
1482 }
1483
1484 /*
1485  * This walks the free areas for each zone.
1486  */
1487 static int frag_show(struct seq_file *m, void *arg)
1488 {
1489         pg_data_t *pgdat = (pg_data_t *)arg;
1490         walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1491         return 0;
1492 }
1493
1494 static void pagetypeinfo_showfree_print(struct seq_file *m,
1495                                         pg_data_t *pgdat, struct zone *zone)
1496 {
1497         int order, mtype;
1498
1499         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1500                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
1501                                         pgdat->node_id,
1502                                         zone->name,
1503                                         migratetype_names[mtype]);
1504                 for (order = 0; order <= MAX_ORDER; ++order) {
1505                         unsigned long freecount = 0;
1506                         struct free_area *area;
1507                         struct list_head *curr;
1508                         bool overflow = false;
1509
1510                         area = &(zone->free_area[order]);
1511
1512                         list_for_each(curr, &area->free_list[mtype]) {
1513                                 /*
1514                                  * Cap the free_list iteration because it might
1515                                  * be really large and we are under a spinlock
1516                                  * so a long time spent here could trigger a
1517                                  * hard lockup detector. Anyway this is a
1518                                  * debugging tool so knowing there is a handful
1519                                  * of pages of this order should be more than
1520                                  * sufficient.
1521                                  */
1522                                 if (++freecount >= 100000) {
1523                                         overflow = true;
1524                                         break;
1525                                 }
1526                         }
1527                         seq_printf(m, "%s%6lu ", overflow ? ">" : "", freecount);
1528                         spin_unlock_irq(&zone->lock);
1529                         cond_resched();
1530                         spin_lock_irq(&zone->lock);
1531                 }
1532                 seq_putc(m, '\n');
1533         }
1534 }
1535
1536 /* Print out the free pages at each order for each migatetype */
1537 static void pagetypeinfo_showfree(struct seq_file *m, void *arg)
1538 {
1539         int order;
1540         pg_data_t *pgdat = (pg_data_t *)arg;
1541
1542         /* Print header */
1543         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1544         for (order = 0; order <= MAX_ORDER; ++order)
1545                 seq_printf(m, "%6d ", order);
1546         seq_putc(m, '\n');
1547
1548         walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1549 }
1550
1551 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1552                                         pg_data_t *pgdat, struct zone *zone)
1553 {
1554         int mtype;
1555         unsigned long pfn;
1556         unsigned long start_pfn = zone->zone_start_pfn;
1557         unsigned long end_pfn = zone_end_pfn(zone);
1558         unsigned long count[MIGRATE_TYPES] = { 0, };
1559
1560         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1561                 struct page *page;
1562
1563                 page = pfn_to_online_page(pfn);
1564                 if (!page)
1565                         continue;
1566
1567                 if (page_zone(page) != zone)
1568                         continue;
1569
1570                 mtype = get_pageblock_migratetype(page);
1571
1572                 if (mtype < MIGRATE_TYPES)
1573                         count[mtype]++;
1574         }
1575
1576         /* Print counts */
1577         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1578         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1579                 seq_printf(m, "%12lu ", count[mtype]);
1580         seq_putc(m, '\n');
1581 }
1582
1583 /* Print out the number of pageblocks for each migratetype */
1584 static void pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1585 {
1586         int mtype;
1587         pg_data_t *pgdat = (pg_data_t *)arg;
1588
1589         seq_printf(m, "\n%-23s", "Number of blocks type ");
1590         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1591                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1592         seq_putc(m, '\n');
1593         walk_zones_in_node(m, pgdat, true, false,
1594                 pagetypeinfo_showblockcount_print);
1595 }
1596
1597 /*
1598  * Print out the number of pageblocks for each migratetype that contain pages
1599  * of other types. This gives an indication of how well fallbacks are being
1600  * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1601  * to determine what is going on
1602  */
1603 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1604 {
1605 #ifdef CONFIG_PAGE_OWNER
1606         int mtype;
1607
1608         if (!static_branch_unlikely(&page_owner_inited))
1609                 return;
1610
1611         drain_all_pages(NULL);
1612
1613         seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1614         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1615                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1616         seq_putc(m, '\n');
1617
1618         walk_zones_in_node(m, pgdat, true, true,
1619                 pagetypeinfo_showmixedcount_print);
1620 #endif /* CONFIG_PAGE_OWNER */
1621 }
1622
1623 /*
1624  * This prints out statistics in relation to grouping pages by mobility.
1625  * It is expensive to collect so do not constantly read the file.
1626  */
1627 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1628 {
1629         pg_data_t *pgdat = (pg_data_t *)arg;
1630
1631         /* check memoryless node */
1632         if (!node_state(pgdat->node_id, N_MEMORY))
1633                 return 0;
1634
1635         seq_printf(m, "Page block order: %d\n", pageblock_order);
1636         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
1637         seq_putc(m, '\n');
1638         pagetypeinfo_showfree(m, pgdat);
1639         pagetypeinfo_showblockcount(m, pgdat);
1640         pagetypeinfo_showmixedcount(m, pgdat);
1641
1642         return 0;
1643 }
1644
1645 static const struct seq_operations fragmentation_op = {
1646         .start  = frag_start,
1647         .next   = frag_next,
1648         .stop   = frag_stop,
1649         .show   = frag_show,
1650 };
1651
1652 static const struct seq_operations pagetypeinfo_op = {
1653         .start  = frag_start,
1654         .next   = frag_next,
1655         .stop   = frag_stop,
1656         .show   = pagetypeinfo_show,
1657 };
1658
1659 static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1660 {
1661         int zid;
1662
1663         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1664                 struct zone *compare = &pgdat->node_zones[zid];
1665
1666                 if (populated_zone(compare))
1667                         return zone == compare;
1668         }
1669
1670         return false;
1671 }
1672
1673 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1674                                                         struct zone *zone)
1675 {
1676         int i;
1677         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1678         if (is_zone_first_populated(pgdat, zone)) {
1679                 seq_printf(m, "\n  per-node stats");
1680                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1681                         unsigned long pages = node_page_state_pages(pgdat, i);
1682
1683                         if (vmstat_item_print_in_thp(i))
1684                                 pages /= HPAGE_PMD_NR;
1685                         seq_printf(m, "\n      %-12s %lu", node_stat_name(i),
1686                                    pages);
1687                 }
1688         }
1689         seq_printf(m,
1690                    "\n  pages free     %lu"
1691                    "\n        boost    %lu"
1692                    "\n        min      %lu"
1693                    "\n        low      %lu"
1694                    "\n        high     %lu"
1695                    "\n        spanned  %lu"
1696                    "\n        present  %lu"
1697                    "\n        managed  %lu"
1698                    "\n        cma      %lu",
1699                    zone_page_state(zone, NR_FREE_PAGES),
1700                    zone->watermark_boost,
1701                    min_wmark_pages(zone),
1702                    low_wmark_pages(zone),
1703                    high_wmark_pages(zone),
1704                    zone->spanned_pages,
1705                    zone->present_pages,
1706                    zone_managed_pages(zone),
1707                    zone_cma_pages(zone));
1708
1709         seq_printf(m,
1710                    "\n        protection: (%ld",
1711                    zone->lowmem_reserve[0]);
1712         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1713                 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1714         seq_putc(m, ')');
1715
1716         /* If unpopulated, no other information is useful */
1717         if (!populated_zone(zone)) {
1718                 seq_putc(m, '\n');
1719                 return;
1720         }
1721
1722         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1723                 seq_printf(m, "\n      %-12s %lu", zone_stat_name(i),
1724                            zone_page_state(zone, i));
1725
1726 #ifdef CONFIG_NUMA
1727         for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++)
1728                 seq_printf(m, "\n      %-12s %lu", numa_stat_name(i),
1729                            zone_numa_event_state(zone, i));
1730 #endif
1731
1732         seq_printf(m, "\n  pagesets");
1733         for_each_online_cpu(i) {
1734                 struct per_cpu_pages *pcp;
1735                 struct per_cpu_zonestat __maybe_unused *pzstats;
1736
1737                 pcp = per_cpu_ptr(zone->per_cpu_pageset, i);
1738                 seq_printf(m,
1739                            "\n    cpu: %i"
1740                            "\n              count: %i"
1741                            "\n              high:  %i"
1742                            "\n              batch: %i",
1743                            i,
1744                            pcp->count,
1745                            pcp->high,
1746                            pcp->batch);
1747 #ifdef CONFIG_SMP
1748                 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, i);
1749                 seq_printf(m, "\n  vm stats threshold: %d",
1750                                 pzstats->stat_threshold);
1751 #endif
1752         }
1753         seq_printf(m,
1754                    "\n  node_unreclaimable:  %u"
1755                    "\n  start_pfn:           %lu",
1756                    pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1757                    zone->zone_start_pfn);
1758         seq_putc(m, '\n');
1759 }
1760
1761 /*
1762  * Output information about zones in @pgdat.  All zones are printed regardless
1763  * of whether they are populated or not: lowmem_reserve_ratio operates on the
1764  * set of all zones and userspace would not be aware of such zones if they are
1765  * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1766  */
1767 static int zoneinfo_show(struct seq_file *m, void *arg)
1768 {
1769         pg_data_t *pgdat = (pg_data_t *)arg;
1770         walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1771         return 0;
1772 }
1773
1774 static const struct seq_operations zoneinfo_op = {
1775         .start  = frag_start, /* iterate over all zones. The same as in
1776                                * fragmentation. */
1777         .next   = frag_next,
1778         .stop   = frag_stop,
1779         .show   = zoneinfo_show,
1780 };
1781
1782 #define NR_VMSTAT_ITEMS (NR_VM_ZONE_STAT_ITEMS + \
1783                          NR_VM_NUMA_EVENT_ITEMS + \
1784                          NR_VM_NODE_STAT_ITEMS + \
1785                          NR_VM_WRITEBACK_STAT_ITEMS + \
1786                          (IS_ENABLED(CONFIG_VM_EVENT_COUNTERS) ? \
1787                           NR_VM_EVENT_ITEMS : 0))
1788
1789 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1790 {
1791         unsigned long *v;
1792         int i;
1793
1794         if (*pos >= NR_VMSTAT_ITEMS)
1795                 return NULL;
1796
1797         BUILD_BUG_ON(ARRAY_SIZE(vmstat_text) < NR_VMSTAT_ITEMS);
1798         fold_vm_numa_events();
1799         v = kmalloc_array(NR_VMSTAT_ITEMS, sizeof(unsigned long), GFP_KERNEL);
1800         m->private = v;
1801         if (!v)
1802                 return ERR_PTR(-ENOMEM);
1803         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1804                 v[i] = global_zone_page_state(i);
1805         v += NR_VM_ZONE_STAT_ITEMS;
1806
1807 #ifdef CONFIG_NUMA
1808         for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++)
1809                 v[i] = global_numa_event_state(i);
1810         v += NR_VM_NUMA_EVENT_ITEMS;
1811 #endif
1812
1813         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1814                 v[i] = global_node_page_state_pages(i);
1815                 if (vmstat_item_print_in_thp(i))
1816                         v[i] /= HPAGE_PMD_NR;
1817         }
1818         v += NR_VM_NODE_STAT_ITEMS;
1819
1820         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1821                             v + NR_DIRTY_THRESHOLD);
1822         v += NR_VM_WRITEBACK_STAT_ITEMS;
1823
1824 #ifdef CONFIG_VM_EVENT_COUNTERS
1825         all_vm_events(v);
1826         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1827         v[PGPGOUT] /= 2;
1828 #endif
1829         return (unsigned long *)m->private + *pos;
1830 }
1831
1832 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1833 {
1834         (*pos)++;
1835         if (*pos >= NR_VMSTAT_ITEMS)
1836                 return NULL;
1837         return (unsigned long *)m->private + *pos;
1838 }
1839
1840 static int vmstat_show(struct seq_file *m, void *arg)
1841 {
1842         unsigned long *l = arg;
1843         unsigned long off = l - (unsigned long *)m->private;
1844
1845         seq_puts(m, vmstat_text[off]);
1846         seq_put_decimal_ull(m, " ", *l);
1847         seq_putc(m, '\n');
1848
1849         if (off == NR_VMSTAT_ITEMS - 1) {
1850                 /*
1851                  * We've come to the end - add any deprecated counters to avoid
1852                  * breaking userspace which might depend on them being present.
1853                  */
1854                 seq_puts(m, "nr_unstable 0\n");
1855         }
1856         return 0;
1857 }
1858
1859 static void vmstat_stop(struct seq_file *m, void *arg)
1860 {
1861         kfree(m->private);
1862         m->private = NULL;
1863 }
1864
1865 static const struct seq_operations vmstat_op = {
1866         .start  = vmstat_start,
1867         .next   = vmstat_next,
1868         .stop   = vmstat_stop,
1869         .show   = vmstat_show,
1870 };
1871 #endif /* CONFIG_PROC_FS */
1872
1873 #ifdef CONFIG_SMP
1874 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1875 int sysctl_stat_interval __read_mostly = HZ;
1876
1877 #ifdef CONFIG_PROC_FS
1878 static void refresh_vm_stats(struct work_struct *work)
1879 {
1880         refresh_cpu_vm_stats(true);
1881 }
1882
1883 int vmstat_refresh(struct ctl_table *table, int write,
1884                    void *buffer, size_t *lenp, loff_t *ppos)
1885 {
1886         long val;
1887         int err;
1888         int i;
1889
1890         /*
1891          * The regular update, every sysctl_stat_interval, may come later
1892          * than expected: leaving a significant amount in per_cpu buckets.
1893          * This is particularly misleading when checking a quantity of HUGE
1894          * pages, immediately after running a test.  /proc/sys/vm/stat_refresh,
1895          * which can equally be echo'ed to or cat'ted from (by root),
1896          * can be used to update the stats just before reading them.
1897          *
1898          * Oh, and since global_zone_page_state() etc. are so careful to hide
1899          * transiently negative values, report an error here if any of
1900          * the stats is negative, so we know to go looking for imbalance.
1901          */
1902         err = schedule_on_each_cpu(refresh_vm_stats);
1903         if (err)
1904                 return err;
1905         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1906                 /*
1907                  * Skip checking stats known to go negative occasionally.
1908                  */
1909                 switch (i) {
1910                 case NR_ZONE_WRITE_PENDING:
1911                 case NR_FREE_CMA_PAGES:
1912                         continue;
1913                 }
1914                 val = atomic_long_read(&vm_zone_stat[i]);
1915                 if (val < 0) {
1916                         pr_warn("%s: %s %ld\n",
1917                                 __func__, zone_stat_name(i), val);
1918                 }
1919         }
1920         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1921                 /*
1922                  * Skip checking stats known to go negative occasionally.
1923                  */
1924                 switch (i) {
1925                 case NR_WRITEBACK:
1926                         continue;
1927                 }
1928                 val = atomic_long_read(&vm_node_stat[i]);
1929                 if (val < 0) {
1930                         pr_warn("%s: %s %ld\n",
1931                                 __func__, node_stat_name(i), val);
1932                 }
1933         }
1934         if (write)
1935                 *ppos += *lenp;
1936         else
1937                 *lenp = 0;
1938         return 0;
1939 }
1940 #endif /* CONFIG_PROC_FS */
1941
1942 static void vmstat_update(struct work_struct *w)
1943 {
1944         if (refresh_cpu_vm_stats(true)) {
1945                 /*
1946                  * Counters were updated so we expect more updates
1947                  * to occur in the future. Keep on running the
1948                  * update worker thread.
1949                  */
1950                 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
1951                                 this_cpu_ptr(&vmstat_work),
1952                                 round_jiffies_relative(sysctl_stat_interval));
1953         }
1954 }
1955
1956 /*
1957  * Check if the diffs for a certain cpu indicate that
1958  * an update is needed.
1959  */
1960 static bool need_update(int cpu)
1961 {
1962         pg_data_t *last_pgdat = NULL;
1963         struct zone *zone;
1964
1965         for_each_populated_zone(zone) {
1966                 struct per_cpu_zonestat *pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
1967                 struct per_cpu_nodestat *n;
1968
1969                 /*
1970                  * The fast way of checking if there are any vmstat diffs.
1971                  */
1972                 if (memchr_inv(pzstats->vm_stat_diff, 0, sizeof(pzstats->vm_stat_diff)))
1973                         return true;
1974
1975                 if (last_pgdat == zone->zone_pgdat)
1976                         continue;
1977                 last_pgdat = zone->zone_pgdat;
1978                 n = per_cpu_ptr(zone->zone_pgdat->per_cpu_nodestats, cpu);
1979                 if (memchr_inv(n->vm_node_stat_diff, 0, sizeof(n->vm_node_stat_diff)))
1980                         return true;
1981         }
1982         return false;
1983 }
1984
1985 /*
1986  * Switch off vmstat processing and then fold all the remaining differentials
1987  * until the diffs stay at zero. The function is used by NOHZ and can only be
1988  * invoked when tick processing is not active.
1989  */
1990 void quiet_vmstat(void)
1991 {
1992         if (system_state != SYSTEM_RUNNING)
1993                 return;
1994
1995         if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
1996                 return;
1997
1998         if (!need_update(smp_processor_id()))
1999                 return;
2000
2001         /*
2002          * Just refresh counters and do not care about the pending delayed
2003          * vmstat_update. It doesn't fire that often to matter and canceling
2004          * it would be too expensive from this path.
2005          * vmstat_shepherd will take care about that for us.
2006          */
2007         refresh_cpu_vm_stats(false);
2008 }
2009
2010 /*
2011  * Shepherd worker thread that checks the
2012  * differentials of processors that have their worker
2013  * threads for vm statistics updates disabled because of
2014  * inactivity.
2015  */
2016 static void vmstat_shepherd(struct work_struct *w);
2017
2018 static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
2019
2020 static void vmstat_shepherd(struct work_struct *w)
2021 {
2022         int cpu;
2023
2024         cpus_read_lock();
2025         /* Check processors whose vmstat worker threads have been disabled */
2026         for_each_online_cpu(cpu) {
2027                 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
2028
2029                 /*
2030                  * In kernel users of vmstat counters either require the precise value and
2031                  * they are using zone_page_state_snapshot interface or they can live with
2032                  * an imprecision as the regular flushing can happen at arbitrary time and
2033                  * cumulative error can grow (see calculate_normal_threshold).
2034                  *
2035                  * From that POV the regular flushing can be postponed for CPUs that have
2036                  * been isolated from the kernel interference without critical
2037                  * infrastructure ever noticing. Skip regular flushing from vmstat_shepherd
2038                  * for all isolated CPUs to avoid interference with the isolated workload.
2039                  */
2040                 if (cpu_is_isolated(cpu))
2041                         continue;
2042
2043                 if (!delayed_work_pending(dw) && need_update(cpu))
2044                         queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
2045
2046                 cond_resched();
2047         }
2048         cpus_read_unlock();
2049
2050         schedule_delayed_work(&shepherd,
2051                 round_jiffies_relative(sysctl_stat_interval));
2052 }
2053
2054 static void __init start_shepherd_timer(void)
2055 {
2056         int cpu;
2057
2058         for_each_possible_cpu(cpu)
2059                 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
2060                         vmstat_update);
2061
2062         schedule_delayed_work(&shepherd,
2063                 round_jiffies_relative(sysctl_stat_interval));
2064 }
2065
2066 static void __init init_cpu_node_state(void)
2067 {
2068         int node;
2069
2070         for_each_online_node(node) {
2071                 if (!cpumask_empty(cpumask_of_node(node)))
2072                         node_set_state(node, N_CPU);
2073         }
2074 }
2075
2076 static int vmstat_cpu_online(unsigned int cpu)
2077 {
2078         refresh_zone_stat_thresholds();
2079
2080         if (!node_state(cpu_to_node(cpu), N_CPU)) {
2081                 node_set_state(cpu_to_node(cpu), N_CPU);
2082         }
2083
2084         return 0;
2085 }
2086
2087 static int vmstat_cpu_down_prep(unsigned int cpu)
2088 {
2089         cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
2090         return 0;
2091 }
2092
2093 static int vmstat_cpu_dead(unsigned int cpu)
2094 {
2095         const struct cpumask *node_cpus;
2096         int node;
2097
2098         node = cpu_to_node(cpu);
2099
2100         refresh_zone_stat_thresholds();
2101         node_cpus = cpumask_of_node(node);
2102         if (!cpumask_empty(node_cpus))
2103                 return 0;
2104
2105         node_clear_state(node, N_CPU);
2106
2107         return 0;
2108 }
2109
2110 #endif
2111
2112 struct workqueue_struct *mm_percpu_wq;
2113
2114 void __init init_mm_internals(void)
2115 {
2116         int ret __maybe_unused;
2117
2118         mm_percpu_wq = alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM, 0);
2119
2120 #ifdef CONFIG_SMP
2121         ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
2122                                         NULL, vmstat_cpu_dead);
2123         if (ret < 0)
2124                 pr_err("vmstat: failed to register 'dead' hotplug state\n");
2125
2126         ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "mm/vmstat:online",
2127                                         vmstat_cpu_online,
2128                                         vmstat_cpu_down_prep);
2129         if (ret < 0)
2130                 pr_err("vmstat: failed to register 'online' hotplug state\n");
2131
2132         cpus_read_lock();
2133         init_cpu_node_state();
2134         cpus_read_unlock();
2135
2136         start_shepherd_timer();
2137 #endif
2138 #ifdef CONFIG_PROC_FS
2139         proc_create_seq("buddyinfo", 0444, NULL, &fragmentation_op);
2140         proc_create_seq("pagetypeinfo", 0400, NULL, &pagetypeinfo_op);
2141         proc_create_seq("vmstat", 0444, NULL, &vmstat_op);
2142         proc_create_seq("zoneinfo", 0444, NULL, &zoneinfo_op);
2143 #endif
2144 }
2145
2146 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
2147
2148 /*
2149  * Return an index indicating how much of the available free memory is
2150  * unusable for an allocation of the requested size.
2151  */
2152 static int unusable_free_index(unsigned int order,
2153                                 struct contig_page_info *info)
2154 {
2155         /* No free memory is interpreted as all free memory is unusable */
2156         if (info->free_pages == 0)
2157                 return 1000;
2158
2159         /*
2160          * Index should be a value between 0 and 1. Return a value to 3
2161          * decimal places.
2162          *
2163          * 0 => no fragmentation
2164          * 1 => high fragmentation
2165          */
2166         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
2167
2168 }
2169
2170 static void unusable_show_print(struct seq_file *m,
2171                                         pg_data_t *pgdat, struct zone *zone)
2172 {
2173         unsigned int order;
2174         int index;
2175         struct contig_page_info info;
2176
2177         seq_printf(m, "Node %d, zone %8s ",
2178                                 pgdat->node_id,
2179                                 zone->name);
2180         for (order = 0; order <= MAX_ORDER; ++order) {
2181                 fill_contig_page_info(zone, order, &info);
2182                 index = unusable_free_index(order, &info);
2183                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2184         }
2185
2186         seq_putc(m, '\n');
2187 }
2188
2189 /*
2190  * Display unusable free space index
2191  *
2192  * The unusable free space index measures how much of the available free
2193  * memory cannot be used to satisfy an allocation of a given size and is a
2194  * value between 0 and 1. The higher the value, the more of free memory is
2195  * unusable and by implication, the worse the external fragmentation is. This
2196  * can be expressed as a percentage by multiplying by 100.
2197  */
2198 static int unusable_show(struct seq_file *m, void *arg)
2199 {
2200         pg_data_t *pgdat = (pg_data_t *)arg;
2201
2202         /* check memoryless node */
2203         if (!node_state(pgdat->node_id, N_MEMORY))
2204                 return 0;
2205
2206         walk_zones_in_node(m, pgdat, true, false, unusable_show_print);
2207
2208         return 0;
2209 }
2210
2211 static const struct seq_operations unusable_sops = {
2212         .start  = frag_start,
2213         .next   = frag_next,
2214         .stop   = frag_stop,
2215         .show   = unusable_show,
2216 };
2217
2218 DEFINE_SEQ_ATTRIBUTE(unusable);
2219
2220 static void extfrag_show_print(struct seq_file *m,
2221                                         pg_data_t *pgdat, struct zone *zone)
2222 {
2223         unsigned int order;
2224         int index;
2225
2226         /* Alloc on stack as interrupts are disabled for zone walk */
2227         struct contig_page_info info;
2228
2229         seq_printf(m, "Node %d, zone %8s ",
2230                                 pgdat->node_id,
2231                                 zone->name);
2232         for (order = 0; order <= MAX_ORDER; ++order) {
2233                 fill_contig_page_info(zone, order, &info);
2234                 index = __fragmentation_index(order, &info);
2235                 seq_printf(m, "%2d.%03d ", index / 1000, index % 1000);
2236         }
2237
2238         seq_putc(m, '\n');
2239 }
2240
2241 /*
2242  * Display fragmentation index for orders that allocations would fail for
2243  */
2244 static int extfrag_show(struct seq_file *m, void *arg)
2245 {
2246         pg_data_t *pgdat = (pg_data_t *)arg;
2247
2248         walk_zones_in_node(m, pgdat, true, false, extfrag_show_print);
2249
2250         return 0;
2251 }
2252
2253 static const struct seq_operations extfrag_sops = {
2254         .start  = frag_start,
2255         .next   = frag_next,
2256         .stop   = frag_stop,
2257         .show   = extfrag_show,
2258 };
2259
2260 DEFINE_SEQ_ATTRIBUTE(extfrag);
2261
2262 static int __init extfrag_debug_init(void)
2263 {
2264         struct dentry *extfrag_debug_root;
2265
2266         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
2267
2268         debugfs_create_file("unusable_index", 0444, extfrag_debug_root, NULL,
2269                             &unusable_fops);
2270
2271         debugfs_create_file("extfrag_index", 0444, extfrag_debug_root, NULL,
2272                             &extfrag_fops);
2273
2274         return 0;
2275 }
2276
2277 module_init(extfrag_debug_init);
2278 #endif