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