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