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         get_online_cpus();
 133         sum_vm_events(ret);
 134         put_online_cpus();
 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 /* See __count_vm_event comment on why raw_cpu_inc is used. */
 942 void __count_numa_event(struct zone *zone,
 943                                  enum numa_stat_item item)
 944 {
 945         struct per_cpu_zonestat __percpu *pzstats = zone->per_cpu_zonestats;
 946
 947         raw_cpu_inc(pzstats->vm_numa_event[item]);
 948 }
 949
 950 /*
 951  * Determine the per node value of a stat item. This function
 952  * is called frequently in a NUMA machine, so try to be as
 953  * frugal as possible.
 954  */
 955 unsigned long sum_zone_node_page_state(int node,
 956                                  enum zone_stat_item item)
 957 {
 958         struct zone *zones = NODE_DATA(node)->node_zones;
 959         int i;
 960         unsigned long count = 0;
 961
 962         for (i = 0; i < MAX_NR_ZONES; i++)
 963                 count += zone_page_state(zones + i, item);
 964
 965         return count;
 966 }
 967
 968 /* Determine the per node value of a numa stat item. */
 969 unsigned long sum_zone_numa_event_state(int node,
 970                                  enum numa_stat_item item)
 971 {
 972         struct zone *zones = NODE_DATA(node)->node_zones;
 973         unsigned long count = 0;
 974         int i;
 975
 976         for (i = 0; i < MAX_NR_ZONES; i++)
 977                 count += zone_numa_event_state(zones + i, item);
 978
 979         return count;
 980 }
 981
 982 /*
 983  * Determine the per node value of a stat item.
 984  */
 985 unsigned long node_page_state_pages(struct pglist_data *pgdat,
 986                                     enum node_stat_item item)
 987 {
 988         long x = atomic_long_read(&pgdat->vm_stat[item]);
 989 #ifdef CONFIG_SMP
 990         if (x < 0)
 991                 x = 0;
 992 #endif
 993         return x;
 994 }
 995
 996 unsigned long node_page_state(struct pglist_data *pgdat,
 997                               enum node_stat_item item)
 998 {
 999         VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
1000
1001         return node_page_state_pages(pgdat, item);
1002 }
1003 #endif
1004
1005 #ifdef CONFIG_COMPACTION
1006
1007 struct contig_page_info {
1008         unsigned long free_pages;
1009         unsigned long free_blocks_total;
1010         unsigned long free_blocks_suitable;
1011 };
1012
1013 /*
1014  * Calculate the number of free pages in a zone, how many contiguous
1015  * pages are free and how many are large enough to satisfy an allocation of
1016  * the target size. Note that this function makes no attempt to estimate
1017  * how many suitable free blocks there *might* be if MOVABLE pages were
1018  * migrated. Calculating that is possible, but expensive and can be
1019  * figured out from userspace
1020  */
1021 static void fill_contig_page_info(struct zone *zone,
1022                                 unsigned int suitable_order,
1023                                 struct contig_page_info *info)
1024 {
1025         unsigned int order;
1026
1027         info->free_pages = 0;
1028         info->free_blocks_total = 0;
1029         info->free_blocks_suitable = 0;
1030
1031         for (order = 0; order < MAX_ORDER; order++) {
1032                 unsigned long blocks;
1033
1034                 /* Count number of free blocks */
1035                 blocks = zone->free_area[order].nr_free;
1036                 info->free_blocks_total += blocks;
1037
1038                 /* Count free base pages */
1039                 info->free_pages += blocks << order;
1040
1041                 /* Count the suitable free blocks */
1042                 if (order >= suitable_order)
1043                         info->free_blocks_suitable += blocks <<
1044                                                 (order - suitable_order);
1045         }
1046 }
1047
1048 /*
1049  * A fragmentation index only makes sense if an allocation of a requested
1050  * size would fail. If that is true, the fragmentation index indicates
1051  * whether external fragmentation or a lack of memory was the problem.
1052  * The value can be used to determine if page reclaim or compaction
1053  * should be used
1054  */
1055 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
1056 {
1057         unsigned long requested = 1UL << order;
1058
1059         if (WARN_ON_ONCE(order >= MAX_ORDER))
1060                 return 0;
1061
1062         if (!info->free_blocks_total)
1063                 return 0;
1064
1065         /* Fragmentation index only makes sense when a request would fail */
1066         if (info->free_blocks_suitable)
1067                 return -1000;
1068
1069         /*
1070          * Index is between 0 and 1 so return within 3 decimal places
1071          *
1072          * 0 => allocation would fail due to lack of memory
1073          * 1 => allocation would fail due to fragmentation
1074          */
1075         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
1076 }
1077
1078 /*
1079  * Calculates external fragmentation within a zone wrt the given order.
1080  * It is defined as the percentage of pages found in blocks of size
1081  * less than 1 << order. It returns values in range [0, 100].
1082  */
1083 unsigned int extfrag_for_order(struct zone *zone, unsigned int order)
1084 {
1085         struct contig_page_info info;
1086
1087         fill_contig_page_info(zone, order, &info);
1088         if (info.free_pages == 0)
1089                 return 0;
1090
1091         return div_u64((info.free_pages -
1092                         (info.free_blocks_suitable << order)) * 100,
1093                         info.free_pages);
1094 }
1095
1096 /* Same as __fragmentation index but allocs contig_page_info on stack */
1097 int fragmentation_index(struct zone *zone, unsigned int order)
1098 {
1099         struct contig_page_info info;
1100
1101         fill_contig_page_info(zone, order, &info);
1102         return __fragmentation_index(order, &info);
1103 }
1104 #endif
1105
1106 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || \
1107     defined(CONFIG_NUMA) || defined(CONFIG_MEMCG)
1108 #ifdef CONFIG_ZONE_DMA
1109 #define TEXT_FOR_DMA(xx) xx "_dma",
1110 #else
1111 #define TEXT_FOR_DMA(xx)
1112 #endif
1113
1114 #ifdef CONFIG_ZONE_DMA32
1115 #define TEXT_FOR_DMA32(xx) xx "_dma32",
1116 #else
1117 #define TEXT_FOR_DMA32(xx)
1118 #endif
1119
1120 #ifdef CONFIG_HIGHMEM
1121 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
1122 #else
1123 #define TEXT_FOR_HIGHMEM(xx)
1124 #endif
1125
1126 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1127                                         TEXT_FOR_HIGHMEM(xx) xx "_movable",
1128
1129 const char * const vmstat_text[] = {
1130         /* enum zone_stat_item counters */
1131         "nr_free_pages",
1132         "nr_zone_inactive_anon",
1133         "nr_zone_active_anon",
1134         "nr_zone_inactive_file",
1135         "nr_zone_active_file",
1136         "nr_zone_unevictable",
1137         "nr_zone_write_pending",
1138         "nr_mlock",
1139         "nr_bounce",
1140 #if IS_ENABLED(CONFIG_ZSMALLOC)
1141         "nr_zspages",
1142 #endif
1143         "nr_free_cma",
1144
1145         /* enum numa_stat_item counters */
1146 #ifdef CONFIG_NUMA
1147         "numa_hit",
1148         "numa_miss",
1149         "numa_foreign",
1150         "numa_interleave",
1151         "numa_local",
1152         "numa_other",
1153 #endif
1154
1155         /* enum node_stat_item counters */
1156         "nr_inactive_anon",
1157         "nr_active_anon",
1158         "nr_inactive_file",
1159         "nr_active_file",
1160         "nr_unevictable",
1161         "nr_slab_reclaimable",
1162         "nr_slab_unreclaimable",
1163         "nr_isolated_anon",
1164         "nr_isolated_file",
1165         "workingset_nodes",
1166         "workingset_refault_anon",
1167         "workingset_refault_file",
1168         "workingset_activate_anon",
1169         "workingset_activate_file",
1170         "workingset_restore_anon",
1171         "workingset_restore_file",
1172         "workingset_nodereclaim",
1173         "nr_anon_pages",
1174         "nr_mapped",
1175         "nr_file_pages",
1176         "nr_dirty",
1177         "nr_writeback",
1178         "nr_writeback_temp",
1179         "nr_shmem",
1180         "nr_shmem_hugepages",
1181         "nr_shmem_pmdmapped",
1182         "nr_file_hugepages",
1183         "nr_file_pmdmapped",
1184         "nr_anon_transparent_hugepages",
1185         "nr_vmscan_write",
1186         "nr_vmscan_immediate_reclaim",
1187         "nr_dirtied",
1188         "nr_written",
1189         "nr_kernel_misc_reclaimable",
1190         "nr_foll_pin_acquired",
1191         "nr_foll_pin_released",
1192         "nr_kernel_stack",
1193 #if IS_ENABLED(CONFIG_SHADOW_CALL_STACK)
1194         "nr_shadow_call_stack",
1195 #endif
1196         "nr_page_table_pages",
1197 #ifdef CONFIG_SWAP
1198         "nr_swapcached",
1199 #endif
1200
1201         /* enum writeback_stat_item counters */
1202         "nr_dirty_threshold",
1203         "nr_dirty_background_threshold",
1204
1205 #if defined(CONFIG_VM_EVENT_COUNTERS) || defined(CONFIG_MEMCG)
1206         /* enum vm_event_item counters */
1207         "pgpgin",
1208         "pgpgout",
1209         "pswpin",
1210         "pswpout",
1211
1212         TEXTS_FOR_ZONES("pgalloc")
1213         TEXTS_FOR_ZONES("allocstall")
1214         TEXTS_FOR_ZONES("pgskip")
1215
1216         "pgfree",
1217         "pgactivate",
1218         "pgdeactivate",
1219         "pglazyfree",
1220
1221         "pgfault",
1222         "pgmajfault",
1223         "pglazyfreed",
1224
1225         "pgrefill",
1226         "pgreuse",
1227         "pgsteal_kswapd",
1228         "pgsteal_direct",
1229         "pgscan_kswapd",
1230         "pgscan_direct",
1231         "pgscan_direct_throttle",
1232         "pgscan_anon",
1233         "pgscan_file",
1234         "pgsteal_anon",
1235         "pgsteal_file",
1236
1237 #ifdef CONFIG_NUMA
1238         "zone_reclaim_failed",
1239 #endif
1240         "pginodesteal",
1241         "slabs_scanned",
1242         "kswapd_inodesteal",
1243         "kswapd_low_wmark_hit_quickly",
1244         "kswapd_high_wmark_hit_quickly",
1245         "pageoutrun",
1246
1247         "pgrotated",
1248
1249         "drop_pagecache",
1250         "drop_slab",
1251         "oom_kill",
1252
1253 #ifdef CONFIG_NUMA_BALANCING
1254         "numa_pte_updates",
1255         "numa_huge_pte_updates",
1256         "numa_hint_faults",
1257         "numa_hint_faults_local",
1258         "numa_pages_migrated",
1259 #endif
1260 #ifdef CONFIG_MIGRATION
1261         "pgmigrate_success",
1262         "pgmigrate_fail",
1263         "thp_migration_success",
1264         "thp_migration_fail",
1265         "thp_migration_split",
1266 #endif
1267 #ifdef CONFIG_COMPACTION
1268         "compact_migrate_scanned",
1269         "compact_free_scanned",
1270         "compact_isolated",
1271         "compact_stall",
1272         "compact_fail",
1273         "compact_success",
1274         "compact_daemon_wake",
1275         "compact_daemon_migrate_scanned",
1276         "compact_daemon_free_scanned",
1277 #endif
1278
1279 #ifdef CONFIG_HUGETLB_PAGE
1280         "htlb_buddy_alloc_success",
1281         "htlb_buddy_alloc_fail",
1282 #endif
1283 #ifdef CONFIG_CMA
1284         "cma_alloc_success",
1285         "cma_alloc_fail",
1286 #endif
1287         "unevictable_pgs_culled",
1288         "unevictable_pgs_scanned",
1289         "unevictable_pgs_rescued",
1290         "unevictable_pgs_mlocked",
1291         "unevictable_pgs_munlocked",
1292         "unevictable_pgs_cleared",
1293         "unevictable_pgs_stranded",
1294
1295 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1296         "thp_fault_alloc",
1297         "thp_fault_fallback",
1298         "thp_fault_fallback_charge",
1299         "thp_collapse_alloc",
1300         "thp_collapse_alloc_failed",
1301         "thp_file_alloc",
1302         "thp_file_fallback",
1303         "thp_file_fallback_charge",
1304         "thp_file_mapped",
1305         "thp_split_page",
1306         "thp_split_page_failed",
1307         "thp_deferred_split_page",
1308         "thp_split_pmd",
1309 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1310         "thp_split_pud",
1311 #endif
1312         "thp_zero_page_alloc",
1313         "thp_zero_page_alloc_failed",
1314         "thp_swpout",
1315         "thp_swpout_fallback",
1316 #endif
1317 #ifdef CONFIG_MEMORY_BALLOON
1318         "balloon_inflate",
1319         "balloon_deflate",
1320 #ifdef CONFIG_BALLOON_COMPACTION
1321         "balloon_migrate",
1322 #endif
1323 #endif /* CONFIG_MEMORY_BALLOON */
1324 #ifdef CONFIG_DEBUG_TLBFLUSH
1325         "nr_tlb_remote_flush",
1326         "nr_tlb_remote_flush_received",
1327         "nr_tlb_local_flush_all",
1328         "nr_tlb_local_flush_one",
1329 #endif /* CONFIG_DEBUG_TLBFLUSH */
1330
1331 #ifdef CONFIG_DEBUG_VM_VMACACHE
1332         "vmacache_find_calls",
1333         "vmacache_find_hits",
1334 #endif
1335 #ifdef CONFIG_SWAP
1336         "swap_ra",
1337         "swap_ra_hit",
1338 #endif
1339 #ifdef CONFIG_X86
1340         "direct_map_level2_splits",
1341         "direct_map_level3_splits",
1342 #endif
1343 #endif /* CONFIG_VM_EVENT_COUNTERS || CONFIG_MEMCG */
1344 };
1345 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA || CONFIG_MEMCG */
1346
1347 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1348      defined(CONFIG_PROC_FS)
1349 static void *frag_start(struct seq_file *m, loff_t *pos)
1350 {
1351         pg_data_t *pgdat;
1352         loff_t node = *pos;
1353
1354         for (pgdat = first_online_pgdat();
1355              pgdat && node;
1356              pgdat = next_online_pgdat(pgdat))
1357                 --node;
1358
1359         return pgdat;
1360 }
1361
1362 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1363 {
1364         pg_data_t *pgdat = (pg_data_t *)arg;
1365
1366         (*pos)++;
1367         return next_online_pgdat(pgdat);
1368 }
1369
1370 static void frag_stop(struct seq_file *m, void *arg)
1371 {
1372 }
1373
1374 /*
1375  * Walk zones in a node and print using a callback.
1376  * If @assert_populated is true, only use callback for zones that are populated.
1377  */
1378 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1379                 bool assert_populated, bool nolock,
1380                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1381 {
1382         struct zone *zone;
1383         struct zone *node_zones = pgdat->node_zones;
1384         unsigned long flags;
1385
1386         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1387                 if (assert_populated && !populated_zone(zone))
1388                         continue;
1389
1390                 if (!nolock)
1391                         spin_lock_irqsave(&zone->lock, flags);
1392                 print(m, pgdat, zone);
1393                 if (!nolock)
1394                         spin_unlock_irqrestore(&zone->lock, flags);
1395         }
1396 }
1397 #endif
1398
1399 #ifdef CONFIG_PROC_FS
1400 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1401                                                 struct zone *zone)
1402 {
1403         int order;
1404
1405         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1406         for (order = 0; order < MAX_ORDER; ++order)
1407                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
1408         seq_putc(m, '\n');
1409 }
1410
1411 /*
1412  * This walks the free areas for each zone.
1413  */
1414 static int frag_show(struct seq_file *m, void *arg)
1415 {
1416         pg_data_t *pgdat = (pg_data_t *)arg;
1417         walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1418         return 0;
1419 }
1420
1421 static void pagetypeinfo_showfree_print(struct seq_file *m,
1422                                         pg_data_t *pgdat, struct zone *zone)
1423 {
1424         int order, mtype;
1425
1426         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1427                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
1428                                         pgdat->node_id,
1429                                         zone->name,
1430                                         migratetype_names[mtype]);
1431                 for (order = 0; order < MAX_ORDER; ++order) {
1432                         unsigned long freecount = 0;
1433                         struct free_area *area;
1434                         struct list_head *curr;
1435                         bool overflow = false;
1436
1437                         area = &(zone->free_area[order]);
1438
1439                         list_for_each(curr, &area->free_list[mtype]) {
1440                                 /*
1441                                  * Cap the free_list iteration because it might
1442                                  * be really large and we are under a spinlock
1443                                  * so a long time spent here could trigger a
1444                                  * hard lockup detector. Anyway this is a
1445                                  * debugging tool so knowing there is a handful
1446                                  * of pages of this order should be more than
1447                                  * sufficient.
1448                                  */
1449                                 if (++freecount >= 100000) {
1450                                         overflow = true;
1451                                         break;
1452                                 }
1453                         }
1454                         seq_printf(m, "%s%6lu ", overflow ? ">" : "", freecount);
1455                         spin_unlock_irq(&zone->lock);
1456                         cond_resched();
1457                         spin_lock_irq(&zone->lock);
1458                 }
1459                 seq_putc(m, '\n');
1460         }
1461 }
1462
1463 /* Print out the free pages at each order for each migatetype */
1464 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
1465 {
1466         int order;
1467         pg_data_t *pgdat = (pg_data_t *)arg;
1468
1469         /* Print header */
1470         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1471         for (order = 0; order < MAX_ORDER; ++order)
1472                 seq_printf(m, "%6d ", order);
1473         seq_putc(m, '\n');
1474
1475         walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1476
1477         return 0;
1478 }
1479
1480 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1481                                         pg_data_t *pgdat, struct zone *zone)
1482 {
1483         int mtype;
1484         unsigned long pfn;
1485         unsigned long start_pfn = zone->zone_start_pfn;
1486         unsigned long end_pfn = zone_end_pfn(zone);
1487         unsigned long count[MIGRATE_TYPES] = { 0, };
1488
1489         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1490                 struct page *page;
1491
1492                 page = pfn_to_online_page(pfn);
1493                 if (!page)
1494                         continue;
1495
1496                 if (page_zone(page) != zone)
1497                         continue;
1498
1499                 mtype = get_pageblock_migratetype(page);
1500
1501                 if (mtype < MIGRATE_TYPES)
1502                         count[mtype]++;
1503         }
1504
1505         /* Print counts */
1506         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1507         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1508                 seq_printf(m, "%12lu ", count[mtype]);
1509         seq_putc(m, '\n');
1510 }
1511
1512 /* Print out the number of pageblocks for each migratetype */
1513 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1514 {
1515         int mtype;
1516         pg_data_t *pgdat = (pg_data_t *)arg;
1517
1518         seq_printf(m, "\n%-23s", "Number of blocks type ");
1519         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1520                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1521         seq_putc(m, '\n');
1522         walk_zones_in_node(m, pgdat, true, false,
1523                 pagetypeinfo_showblockcount_print);
1524
1525         return 0;
1526 }
1527
1528 /*
1529  * Print out the number of pageblocks for each migratetype that contain pages
1530  * of other types. This gives an indication of how well fallbacks are being
1531  * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1532  * to determine what is going on
1533  */
1534 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1535 {
1536 #ifdef CONFIG_PAGE_OWNER
1537         int mtype;
1538
1539         if (!static_branch_unlikely(&page_owner_inited))
1540                 return;
1541
1542         drain_all_pages(NULL);
1543
1544         seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1545         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1546                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1547         seq_putc(m, '\n');
1548
1549         walk_zones_in_node(m, pgdat, true, true,
1550                 pagetypeinfo_showmixedcount_print);
1551 #endif /* CONFIG_PAGE_OWNER */
1552 }
1553
1554 /*
1555  * This prints out statistics in relation to grouping pages by mobility.
1556  * It is expensive to collect so do not constantly read the file.
1557  */
1558 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1559 {
1560         pg_data_t *pgdat = (pg_data_t *)arg;
1561
1562         /* check memoryless node */
1563         if (!node_state(pgdat->node_id, N_MEMORY))
1564                 return 0;
1565
1566         seq_printf(m, "Page block order: %d\n", pageblock_order);
1567         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
1568         seq_putc(m, '\n');
1569         pagetypeinfo_showfree(m, pgdat);
1570         pagetypeinfo_showblockcount(m, pgdat);
1571         pagetypeinfo_showmixedcount(m, pgdat);
1572
1573         return 0;
1574 }
1575
1576 static const struct seq_operations fragmentation_op = {
1577         .start  = frag_start,
1578         .next   = frag_next,
1579         .stop   = frag_stop,
1580         .show   = frag_show,
1581 };
1582
1583 static const struct seq_operations pagetypeinfo_op = {
1584         .start  = frag_start,
1585         .next   = frag_next,
1586         .stop   = frag_stop,
1587         .show   = pagetypeinfo_show,
1588 };
1589
1590 static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1591 {
1592         int zid;
1593
1594         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1595                 struct zone *compare = &pgdat->node_zones[zid];
1596
1597                 if (populated_zone(compare))
1598                         return zone == compare;
1599         }
1600
1601         return false;
1602 }
1603
1604 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1605                                                         struct zone *zone)
1606 {
1607         int i;
1608         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1609         if (is_zone_first_populated(pgdat, zone)) {
1610                 seq_printf(m, "\n  per-node stats");
1611                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1612                         unsigned long pages = node_page_state_pages(pgdat, i);
1613
1614                         if (vmstat_item_print_in_thp(i))
1615                                 pages /= HPAGE_PMD_NR;
1616                         seq_printf(m, "\n      %-12s %lu", node_stat_name(i),
1617                                    pages);
1618                 }
1619         }
1620         seq_printf(m,
1621                    "\n  pages free     %lu"
1622                    "\n        min      %lu"
1623                    "\n        low      %lu"
1624                    "\n        high     %lu"
1625                    "\n        spanned  %lu"
1626                    "\n        present  %lu"
1627                    "\n        managed  %lu"
1628                    "\n        cma      %lu",
1629                    zone_page_state(zone, NR_FREE_PAGES),
1630                    min_wmark_pages(zone),
1631                    low_wmark_pages(zone),
1632                    high_wmark_pages(zone),
1633                    zone->spanned_pages,
1634                    zone->present_pages,
1635                    zone_managed_pages(zone),
1636                    zone_cma_pages(zone));
1637
1638         seq_printf(m,
1639                    "\n        protection: (%ld",
1640                    zone->lowmem_reserve[0]);
1641         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1642                 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1643         seq_putc(m, ')');
1644
1645         /* If unpopulated, no other information is useful */
1646         if (!populated_zone(zone)) {
1647                 seq_putc(m, '\n');
1648                 return;
1649         }
1650
1651         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1652                 seq_printf(m, "\n      %-12s %lu", zone_stat_name(i),
1653                            zone_page_state(zone, i));
1654
1655 #ifdef CONFIG_NUMA
1656         for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++)
1657                 seq_printf(m, "\n      %-12s %lu", numa_stat_name(i),
1658                            zone_numa_event_state(zone, i));
1659 #endif
1660
1661         seq_printf(m, "\n  pagesets");
1662         for_each_online_cpu(i) {
1663                 struct per_cpu_pages *pcp;
1664                 struct per_cpu_zonestat __maybe_unused *pzstats;
1665
1666                 pcp = per_cpu_ptr(zone->per_cpu_pageset, i);
1667                 seq_printf(m,
1668                            "\n    cpu: %i"
1669                            "\n              count: %i"
1670                            "\n              high:  %i"
1671                            "\n              batch: %i",
1672                            i,
1673                            pcp->count,
1674                            pcp->high,
1675                            pcp->batch);
1676 #ifdef CONFIG_SMP
1677                 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, i);
1678                 seq_printf(m, "\n  vm stats threshold: %d",
1679                                 pzstats->stat_threshold);
1680 #endif
1681         }
1682         seq_printf(m,
1683                    "\n  node_unreclaimable:  %u"
1684                    "\n  start_pfn:           %lu",
1685                    pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1686                    zone->zone_start_pfn);
1687         seq_putc(m, '\n');
1688 }
1689
1690 /*
1691  * Output information about zones in @pgdat.  All zones are printed regardless
1692  * of whether they are populated or not: lowmem_reserve_ratio operates on the
1693  * set of all zones and userspace would not be aware of such zones if they are
1694  * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1695  */
1696 static int zoneinfo_show(struct seq_file *m, void *arg)
1697 {
1698         pg_data_t *pgdat = (pg_data_t *)arg;
1699         walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1700         return 0;
1701 }
1702
1703 static const struct seq_operations zoneinfo_op = {
1704         .start  = frag_start, /* iterate over all zones. The same as in
1705                                * fragmentation. */
1706         .next   = frag_next,
1707         .stop   = frag_stop,
1708         .show   = zoneinfo_show,
1709 };
1710
1711 #define NR_VMSTAT_ITEMS (NR_VM_ZONE_STAT_ITEMS + \
1712                          NR_VM_NUMA_EVENT_ITEMS + \
1713                          NR_VM_NODE_STAT_ITEMS + \
1714                          NR_VM_WRITEBACK_STAT_ITEMS + \
1715                          (IS_ENABLED(CONFIG_VM_EVENT_COUNTERS) ? \
1716                           NR_VM_EVENT_ITEMS : 0))
1717
1718 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1719 {
1720         unsigned long *v;
1721         int i;
1722
1723         if (*pos >= NR_VMSTAT_ITEMS)
1724                 return NULL;
1725
1726         BUILD_BUG_ON(ARRAY_SIZE(vmstat_text) < NR_VMSTAT_ITEMS);
1727         fold_vm_numa_events();
1728         v = kmalloc_array(NR_VMSTAT_ITEMS, sizeof(unsigned long), GFP_KERNEL);
1729         m->private = v;
1730         if (!v)
1731                 return ERR_PTR(-ENOMEM);
1732         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1733                 v[i] = global_zone_page_state(i);
1734         v += NR_VM_ZONE_STAT_ITEMS;
1735
1736 #ifdef CONFIG_NUMA
1737         for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++)
1738                 v[i] = global_numa_event_state(i);
1739         v += NR_VM_NUMA_EVENT_ITEMS;
1740 #endif
1741
1742         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1743                 v[i] = global_node_page_state_pages(i);
1744                 if (vmstat_item_print_in_thp(i))
1745                         v[i] /= HPAGE_PMD_NR;
1746         }
1747         v += NR_VM_NODE_STAT_ITEMS;
1748
1749         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1750                             v + NR_DIRTY_THRESHOLD);
1751         v += NR_VM_WRITEBACK_STAT_ITEMS;
1752
1753 #ifdef CONFIG_VM_EVENT_COUNTERS
1754         all_vm_events(v);
1755         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1756         v[PGPGOUT] /= 2;
1757 #endif
1758         return (unsigned long *)m->private + *pos;
1759 }
1760
1761 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1762 {
1763         (*pos)++;
1764         if (*pos >= NR_VMSTAT_ITEMS)
1765                 return NULL;
1766         return (unsigned long *)m->private + *pos;
1767 }
1768
1769 static int vmstat_show(struct seq_file *m, void *arg)
1770 {
1771         unsigned long *l = arg;
1772         unsigned long off = l - (unsigned long *)m->private;
1773
1774         seq_puts(m, vmstat_text[off]);
1775         seq_put_decimal_ull(m, " ", *l);
1776         seq_putc(m, '\n');
1777
1778         if (off == NR_VMSTAT_ITEMS - 1) {
1779                 /*
1780                  * We've come to the end - add any deprecated counters to avoid
1781                  * breaking userspace which might depend on them being present.
1782                  */
1783                 seq_puts(m, "nr_unstable 0\n");
1784         }
1785         return 0;
1786 }
1787
1788 static void vmstat_stop(struct seq_file *m, void *arg)
1789 {
1790         kfree(m->private);
1791         m->private = NULL;
1792 }
1793
1794 static const struct seq_operations vmstat_op = {
1795         .start  = vmstat_start,
1796         .next   = vmstat_next,
1797         .stop   = vmstat_stop,
1798         .show   = vmstat_show,
1799 };
1800 #endif /* CONFIG_PROC_FS */
1801
1802 #ifdef CONFIG_SMP
1803 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1804 int sysctl_stat_interval __read_mostly = HZ;
1805
1806 #ifdef CONFIG_PROC_FS
1807 static void refresh_vm_stats(struct work_struct *work)
1808 {
1809         refresh_cpu_vm_stats(true);
1810 }
1811
1812 int vmstat_refresh(struct ctl_table *table, int write,
1813                    void *buffer, size_t *lenp, loff_t *ppos)
1814 {
1815         long val;
1816         int err;
1817         int i;
1818
1819         /*
1820          * The regular update, every sysctl_stat_interval, may come later
1821          * than expected: leaving a significant amount in per_cpu buckets.
1822          * This is particularly misleading when checking a quantity of HUGE
1823          * pages, immediately after running a test.  /proc/sys/vm/stat_refresh,
1824          * which can equally be echo'ed to or cat'ted from (by root),
1825          * can be used to update the stats just before reading them.
1826          *
1827          * Oh, and since global_zone_page_state() etc. are so careful to hide
1828          * transiently negative values, report an error here if any of
1829          * the stats is negative, so we know to go looking for imbalance.
1830          */
1831         err = schedule_on_each_cpu(refresh_vm_stats);
1832         if (err)
1833                 return err;
1834         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1835                 /*
1836                  * Skip checking stats known to go negative occasionally.
1837                  */
1838                 switch (i) {
1839                 case NR_ZONE_WRITE_PENDING:
1840                 case NR_FREE_CMA_PAGES:
1841                         continue;
1842                 }
1843                 val = atomic_long_read(&vm_zone_stat[i]);
1844                 if (val < 0) {
1845                         pr_warn("%s: %s %ld\n",
1846                                 __func__, zone_stat_name(i), val);
1847                 }
1848         }
1849         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1850                 /*
1851                  * Skip checking stats known to go negative occasionally.
1852                  */
1853                 switch (i) {
1854                 case NR_WRITEBACK:
1855                         continue;
1856                 }
1857                 val = atomic_long_read(&vm_node_stat[i]);
1858                 if (val < 0) {
1859                         pr_warn("%s: %s %ld\n",
1860                                 __func__, node_stat_name(i), val);
1861                 }
1862         }
1863         if (write)
1864                 *ppos += *lenp;
1865         else
1866                 *lenp = 0;
1867         return 0;
1868 }
1869 #endif /* CONFIG_PROC_FS */
1870
1871 static void vmstat_update(struct work_struct *w)
1872 {
1873         if (refresh_cpu_vm_stats(true)) {
1874                 /*
1875                  * Counters were updated so we expect more updates
1876                  * to occur in the future. Keep on running the
1877                  * update worker thread.
1878                  */
1879                 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
1880                                 this_cpu_ptr(&vmstat_work),
1881                                 round_jiffies_relative(sysctl_stat_interval));
1882         }
1883 }
1884
1885 /*
1886  * Switch off vmstat processing and then fold all the remaining differentials
1887  * until the diffs stay at zero. The function is used by NOHZ and can only be
1888  * invoked when tick processing is not active.
1889  */
1890 /*
1891  * Check if the diffs for a certain cpu indicate that
1892  * an update is needed.
1893  */
1894 static bool need_update(int cpu)
1895 {
1896         pg_data_t *last_pgdat = NULL;
1897         struct zone *zone;
1898
1899         for_each_populated_zone(zone) {
1900                 struct per_cpu_zonestat *pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
1901                 struct per_cpu_nodestat *n;
1902
1903                 /*
1904                  * The fast way of checking if there are any vmstat diffs.
1905                  */
1906                 if (memchr_inv(pzstats->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS *
1907                                sizeof(pzstats->vm_stat_diff[0])))
1908                         return true;
1909
1910                 if (last_pgdat == zone->zone_pgdat)
1911                         continue;
1912                 last_pgdat = zone->zone_pgdat;
1913                 n = per_cpu_ptr(zone->zone_pgdat->per_cpu_nodestats, cpu);
1914                 if (memchr_inv(n->vm_node_stat_diff, 0, NR_VM_NODE_STAT_ITEMS *
1915                                sizeof(n->vm_node_stat_diff[0])))
1916                     return true;
1917         }
1918         return false;
1919 }
1920
1921 /*
1922  * Switch off vmstat processing and then fold all the remaining differentials
1923  * until the diffs stay at zero. The function is used by NOHZ and can only be
1924  * invoked when tick processing is not active.
1925  */
1926 void quiet_vmstat(void)
1927 {
1928         if (system_state != SYSTEM_RUNNING)
1929                 return;
1930
1931         if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
1932                 return;
1933
1934         if (!need_update(smp_processor_id()))
1935                 return;
1936
1937         /*
1938          * Just refresh counters and do not care about the pending delayed
1939          * vmstat_update. It doesn't fire that often to matter and canceling
1940          * it would be too expensive from this path.
1941          * vmstat_shepherd will take care about that for us.
1942          */
1943         refresh_cpu_vm_stats(false);
1944 }
1945
1946 /*
1947  * Shepherd worker thread that checks the
1948  * differentials of processors that have their worker
1949  * threads for vm statistics updates disabled because of
1950  * inactivity.
1951  */
1952 static void vmstat_shepherd(struct work_struct *w);
1953
1954 static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
1955
1956 static void vmstat_shepherd(struct work_struct *w)
1957 {
1958         int cpu;
1959
1960         get_online_cpus();
1961         /* Check processors whose vmstat worker threads have been disabled */
1962         for_each_online_cpu(cpu) {
1963                 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
1964
1965                 if (!delayed_work_pending(dw) && need_update(cpu))
1966                         queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
1967
1968                 cond_resched();
1969         }
1970         put_online_cpus();
1971
1972         schedule_delayed_work(&shepherd,
1973                 round_jiffies_relative(sysctl_stat_interval));
1974 }
1975
1976 static void __init start_shepherd_timer(void)
1977 {
1978         int cpu;
1979
1980         for_each_possible_cpu(cpu)
1981                 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
1982                         vmstat_update);
1983
1984         schedule_delayed_work(&shepherd,
1985                 round_jiffies_relative(sysctl_stat_interval));
1986 }
1987
1988 static void __init init_cpu_node_state(void)
1989 {
1990         int node;
1991
1992         for_each_online_node(node) {
1993                 if (cpumask_weight(cpumask_of_node(node)) > 0)
1994                         node_set_state(node, N_CPU);
1995         }
1996 }
1997
1998 static int vmstat_cpu_online(unsigned int cpu)
1999 {
2000         refresh_zone_stat_thresholds();
2001         node_set_state(cpu_to_node(cpu), N_CPU);
2002         return 0;
2003 }
2004
2005 static int vmstat_cpu_down_prep(unsigned int cpu)
2006 {
2007         cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
2008         return 0;
2009 }
2010
2011 static int vmstat_cpu_dead(unsigned int cpu)
2012 {
2013         const struct cpumask *node_cpus;
2014         int node;
2015
2016         node = cpu_to_node(cpu);
2017
2018         refresh_zone_stat_thresholds();
2019         node_cpus = cpumask_of_node(node);
2020         if (cpumask_weight(node_cpus) > 0)
2021                 return 0;
2022
2023         node_clear_state(node, N_CPU);
2024         return 0;
2025 }
2026
2027 #endif
2028
2029 struct workqueue_struct *mm_percpu_wq;
2030
2031 void __init init_mm_internals(void)
2032 {
2033         int ret __maybe_unused;
2034
2035         mm_percpu_wq = alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM, 0);
2036
2037 #ifdef CONFIG_SMP
2038         ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
2039                                         NULL, vmstat_cpu_dead);
2040         if (ret < 0)
2041                 pr_err("vmstat: failed to register 'dead' hotplug state\n");
2042
2043         ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "mm/vmstat:online",
2044                                         vmstat_cpu_online,
2045                                         vmstat_cpu_down_prep);
2046         if (ret < 0)
2047                 pr_err("vmstat: failed to register 'online' hotplug state\n");
2048
2049         get_online_cpus();
2050         init_cpu_node_state();
2051         put_online_cpus();
2052
2053         start_shepherd_timer();
2054 #endif
2055 #ifdef CONFIG_PROC_FS
2056         proc_create_seq("buddyinfo", 0444, NULL, &fragmentation_op);
2057         proc_create_seq("pagetypeinfo", 0400, NULL, &pagetypeinfo_op);
2058         proc_create_seq("vmstat", 0444, NULL, &vmstat_op);
2059         proc_create_seq("zoneinfo", 0444, NULL, &zoneinfo_op);
2060 #endif
2061 }
2062
2063 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
2064
2065 /*
2066  * Return an index indicating how much of the available free memory is
2067  * unusable for an allocation of the requested size.
2068  */
2069 static int unusable_free_index(unsigned int order,
2070                                 struct contig_page_info *info)
2071 {
2072         /* No free memory is interpreted as all free memory is unusable */
2073         if (info->free_pages == 0)
2074                 return 1000;
2075
2076         /*
2077          * Index should be a value between 0 and 1. Return a value to 3
2078          * decimal places.
2079          *
2080          * 0 => no fragmentation
2081          * 1 => high fragmentation
2082          */
2083         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
2084
2085 }
2086
2087 static void unusable_show_print(struct seq_file *m,
2088                                         pg_data_t *pgdat, struct zone *zone)
2089 {
2090         unsigned int order;
2091         int index;
2092         struct contig_page_info info;
2093
2094         seq_printf(m, "Node %d, zone %8s ",
2095                                 pgdat->node_id,
2096                                 zone->name);
2097         for (order = 0; order < MAX_ORDER; ++order) {
2098                 fill_contig_page_info(zone, order, &info);
2099                 index = unusable_free_index(order, &info);
2100                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2101         }
2102
2103         seq_putc(m, '\n');
2104 }
2105
2106 /*
2107  * Display unusable free space index
2108  *
2109  * The unusable free space index measures how much of the available free
2110  * memory cannot be used to satisfy an allocation of a given size and is a
2111  * value between 0 and 1. The higher the value, the more of free memory is
2112  * unusable and by implication, the worse the external fragmentation is. This
2113  * can be expressed as a percentage by multiplying by 100.
2114  */
2115 static int unusable_show(struct seq_file *m, void *arg)
2116 {
2117         pg_data_t *pgdat = (pg_data_t *)arg;
2118
2119         /* check memoryless node */
2120         if (!node_state(pgdat->node_id, N_MEMORY))
2121                 return 0;
2122
2123         walk_zones_in_node(m, pgdat, true, false, unusable_show_print);
2124
2125         return 0;
2126 }
2127
2128 static const struct seq_operations unusable_sops = {
2129         .start  = frag_start,
2130         .next   = frag_next,
2131         .stop   = frag_stop,
2132         .show   = unusable_show,
2133 };
2134
2135 DEFINE_SEQ_ATTRIBUTE(unusable);
2136
2137 static void extfrag_show_print(struct seq_file *m,
2138                                         pg_data_t *pgdat, struct zone *zone)
2139 {
2140         unsigned int order;
2141         int index;
2142
2143         /* Alloc on stack as interrupts are disabled for zone walk */
2144         struct contig_page_info info;
2145
2146         seq_printf(m, "Node %d, zone %8s ",
2147                                 pgdat->node_id,
2148                                 zone->name);
2149         for (order = 0; order < MAX_ORDER; ++order) {
2150                 fill_contig_page_info(zone, order, &info);
2151                 index = __fragmentation_index(order, &info);
2152                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2153         }
2154
2155         seq_putc(m, '\n');
2156 }
2157
2158 /*
2159  * Display fragmentation index for orders that allocations would fail for
2160  */
2161 static int extfrag_show(struct seq_file *m, void *arg)
2162 {
2163         pg_data_t *pgdat = (pg_data_t *)arg;
2164
2165         walk_zones_in_node(m, pgdat, true, false, extfrag_show_print);
2166
2167         return 0;
2168 }
2169
2170 static const struct seq_operations extfrag_sops = {
2171         .start  = frag_start,
2172         .next   = frag_next,
2173         .stop   = frag_stop,
2174         .show   = extfrag_show,
2175 };
2176
2177 DEFINE_SEQ_ATTRIBUTE(extfrag);
2178
2179 static int __init extfrag_debug_init(void)
2180 {
2181         struct dentry *extfrag_debug_root;
2182
2183         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
2184
2185         debugfs_create_file("unusable_index", 0444, extfrag_debug_root, NULL,
2186                             &unusable_fops);
2187
2188         debugfs_create_file("extfrag_index", 0444, extfrag_debug_root, NULL,
2189                             &extfrag_fops);
2190
2191         return 0;
2192 }
2193
2194 module_init(extfrag_debug_init);
2195 #endif