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