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