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 #define NUMA_STATS_THRESHOLD (U16_MAX - 2)
  35
  36 #ifdef CONFIG_NUMA
  37 int sysctl_vm_numa_stat = ENABLE_NUMA_STAT;
  38
  39 /* zero numa counters within a zone */
  40 static void zero_zone_numa_counters(struct zone *zone)
  41 {
  42         int item, cpu;
  43
  44         for (item = 0; item < NR_VM_NUMA_STAT_ITEMS; item++) {
  45                 atomic_long_set(&zone->vm_numa_stat[item], 0);
  46                 for_each_online_cpu(cpu)
  47                         per_cpu_ptr(zone->pageset, cpu)->vm_numa_stat_diff[item]
  48                                                 = 0;
  49         }
  50 }
  51
  52 /* zero numa counters of all the populated zones */
  53 static void zero_zones_numa_counters(void)
  54 {
  55         struct zone *zone;
  56
  57         for_each_populated_zone(zone)
  58                 zero_zone_numa_counters(zone);
  59 }
  60
  61 /* zero global numa counters */
  62 static void zero_global_numa_counters(void)
  63 {
  64         int item;
  65
  66         for (item = 0; item < NR_VM_NUMA_STAT_ITEMS; item++)
  67                 atomic_long_set(&vm_numa_stat[item], 0);
  68 }
  69
  70 static void invalid_numa_statistics(void)
  71 {
  72         zero_zones_numa_counters();
  73         zero_global_numa_counters();
  74 }
  75
  76 static DEFINE_MUTEX(vm_numa_stat_lock);
  77
  78 int sysctl_vm_numa_stat_handler(struct ctl_table *table, int write,
  79                 void *buffer, size_t *length, loff_t *ppos)
  80 {
  81         int ret, oldval;
  82
  83         mutex_lock(&vm_numa_stat_lock);
  84         if (write)
  85                 oldval = sysctl_vm_numa_stat;
  86         ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
  87         if (ret || !write)
  88                 goto out;
  89
  90         if (oldval == sysctl_vm_numa_stat)
  91                 goto out;
  92         else if (sysctl_vm_numa_stat == ENABLE_NUMA_STAT) {
  93                 static_branch_enable(&vm_numa_stat_key);
  94                 pr_info("enable numa statistics\n");
  95         } else {
  96                 static_branch_disable(&vm_numa_stat_key);
  97                 invalid_numa_statistics();
  98                 pr_info("disable numa statistics, and clear numa counters\n");
  99         }
 100
 101 out:
 102         mutex_unlock(&vm_numa_stat_lock);
 103         return ret;
 104 }
 105 #endif
 106
 107 #ifdef CONFIG_VM_EVENT_COUNTERS
 108 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
 109 EXPORT_PER_CPU_SYMBOL(vm_event_states);
 110
 111 static void sum_vm_events(unsigned long *ret)
 112 {
 113         int cpu;
 114         int i;
 115
 116         memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
 117
 118         for_each_online_cpu(cpu) {
 119                 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
 120
 121                 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
 122                         ret[i] += this->event[i];
 123         }
 124 }
 125
 126 /*
 127  * Accumulate the vm event counters across all CPUs.
 128  * The result is unavoidably approximate - it can change
 129  * during and after execution of this function.
 130 */
 131 void all_vm_events(unsigned long *ret)
 132 {
 133         get_online_cpus();
 134         sum_vm_events(ret);
 135         put_online_cpus();
 136 }
 137 EXPORT_SYMBOL_GPL(all_vm_events);
 138
 139 /*
 140  * Fold the foreign cpu events into our own.
 141  *
 142  * This is adding to the events on one processor
 143  * but keeps the global counts constant.
 144  */
 145 void vm_events_fold_cpu(int cpu)
 146 {
 147         struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
 148         int i;
 149
 150         for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
 151                 count_vm_events(i, fold_state->event[i]);
 152                 fold_state->event[i] = 0;
 153         }
 154 }
 155
 156 #endif /* CONFIG_VM_EVENT_COUNTERS */
 157
 158 /*
 159  * Manage combined zone based / global counters
 160  *
 161  * vm_stat contains the global counters
 162  */
 163 atomic_long_t vm_zone_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
 164 atomic_long_t vm_numa_stat[NR_VM_NUMA_STAT_ITEMS] __cacheline_aligned_in_smp;
 165 atomic_long_t vm_node_stat[NR_VM_NODE_STAT_ITEMS] __cacheline_aligned_in_smp;
 166 EXPORT_SYMBOL(vm_zone_stat);
 167 EXPORT_SYMBOL(vm_numa_stat);
 168 EXPORT_SYMBOL(vm_node_stat);
 169
 170 #ifdef CONFIG_SMP
 171
 172 int calculate_pressure_threshold(struct zone *zone)
 173 {
 174         int threshold;
 175         int watermark_distance;
 176
 177         /*
 178          * As vmstats are not up to date, there is drift between the estimated
 179          * and real values. For high thresholds and a high number of CPUs, it
 180          * is possible for the min watermark to be breached while the estimated
 181          * value looks fine. The pressure threshold is a reduced value such
 182          * that even the maximum amount of drift will not accidentally breach
 183          * the min watermark
 184          */
 185         watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
 186         threshold = max(1, (int)(watermark_distance / num_online_cpus()));
 187
 188         /*
 189          * Maximum threshold is 125
 190          */
 191         threshold = min(125, threshold);
 192
 193         return threshold;
 194 }
 195
 196 int calculate_normal_threshold(struct zone *zone)
 197 {
 198         int threshold;
 199         int mem;        /* memory in 128 MB units */
 200
 201         /*
 202          * The threshold scales with the number of processors and the amount
 203          * of memory per zone. More memory means that we can defer updates for
 204          * longer, more processors could lead to more contention.
 205          * fls() is used to have a cheap way of logarithmic scaling.
 206          *
 207          * Some sample thresholds:
 208          *
 209          * Threshold    Processors      (fls)   Zonesize        fls(mem+1)
 210          * ------------------------------------------------------------------
 211          * 8            1               1       0.9-1 GB        4
 212          * 16           2               2       0.9-1 GB        4
 213          * 20           2               2       1-2 GB          5
 214          * 24           2               2       2-4 GB          6
 215          * 28           2               2       4-8 GB          7
 216          * 32           2               2       8-16 GB         8
 217          * 4            2               2       <128M           1
 218          * 30           4               3       2-4 GB          5
 219          * 48           4               3       8-16 GB         8
 220          * 32           8               4       1-2 GB          4
 221          * 32           8               4       0.9-1GB         4
 222          * 10           16              5       <128M           1
 223          * 40           16              5       900M            4
 224          * 70           64              7       2-4 GB          5
 225          * 84           64              7       4-8 GB          6
 226          * 108          512             9       4-8 GB          6
 227          * 125          1024            10      8-16 GB         8
 228          * 125          1024            10      16-32 GB        9
 229          */
 230
 231         mem = zone_managed_pages(zone) >> (27 - PAGE_SHIFT);
 232
 233         threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
 234
 235         /*
 236          * Maximum threshold is 125
 237          */
 238         threshold = min(125, threshold);
 239
 240         return threshold;
 241 }
 242
 243 /*
 244  * Refresh the thresholds for each zone.
 245  */
 246 void refresh_zone_stat_thresholds(void)
 247 {
 248         struct pglist_data *pgdat;
 249         struct zone *zone;
 250         int cpu;
 251         int threshold;
 252
 253         /* Zero current pgdat thresholds */
 254         for_each_online_pgdat(pgdat) {
 255                 for_each_online_cpu(cpu) {
 256                         per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold = 0;
 257                 }
 258         }
 259
 260         for_each_populated_zone(zone) {
 261                 struct pglist_data *pgdat = zone->zone_pgdat;
 262                 unsigned long max_drift, tolerate_drift;
 263
 264                 threshold = calculate_normal_threshold(zone);
 265
 266                 for_each_online_cpu(cpu) {
 267                         int pgdat_threshold;
 268
 269                         per_cpu_ptr(zone->pageset, cpu)->stat_threshold
 270                                                         = threshold;
 271
 272                         /* Base nodestat threshold on the largest populated zone. */
 273                         pgdat_threshold = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold;
 274                         per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold
 275                                 = max(threshold, pgdat_threshold);
 276                 }
 277
 278                 /*
 279                  * Only set percpu_drift_mark if there is a danger that
 280                  * NR_FREE_PAGES reports the low watermark is ok when in fact
 281                  * the min watermark could be breached by an allocation
 282                  */
 283                 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
 284                 max_drift = num_online_cpus() * threshold;
 285                 if (max_drift > tolerate_drift)
 286                         zone->percpu_drift_mark = high_wmark_pages(zone) +
 287                                         max_drift;
 288         }
 289 }
 290
 291 void set_pgdat_percpu_threshold(pg_data_t *pgdat,
 292                                 int (*calculate_pressure)(struct zone *))
 293 {
 294         struct zone *zone;
 295         int cpu;
 296         int threshold;
 297         int i;
 298
 299         for (i = 0; i < pgdat->nr_zones; i++) {
 300                 zone = &pgdat->node_zones[i];
 301                 if (!zone->percpu_drift_mark)
 302                         continue;
 303
 304                 threshold = (*calculate_pressure)(zone);
 305                 for_each_online_cpu(cpu)
 306                         per_cpu_ptr(zone->pageset, cpu)->stat_threshold
 307                                                         = threshold;
 308         }
 309 }
 310
 311 /*
 312  * For use when we know that interrupts are disabled,
 313  * or when we know that preemption is disabled and that
 314  * particular counter cannot be updated from interrupt context.
 315  */
 316 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 317                            long delta)
 318 {
 319         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 320         s8 __percpu *p = pcp->vm_stat_diff + item;
 321         long x;
 322         long t;
 323
 324         preempt_disable_rt();
 325         x = delta + __this_cpu_read(*p);
 326
 327         t = __this_cpu_read(pcp->stat_threshold);
 328
 329         if (unlikely(abs(x) > t)) {
 330                 zone_page_state_add(x, zone, item);
 331                 x = 0;
 332         }
 333         __this_cpu_write(*p, x);
 334         preempt_enable_rt();
 335 }
 336 EXPORT_SYMBOL(__mod_zone_page_state);
 337
 338 void __mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
 339                                 long delta)
 340 {
 341         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 342         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 343         long x;
 344         long t;
 345
 346         if (vmstat_item_in_bytes(item)) {
 347                 VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
 348                 delta >>= PAGE_SHIFT;
 349         }
 350
 351         preempt_disable_rt();
 352         x = delta + __this_cpu_read(*p);
 353
 354         t = __this_cpu_read(pcp->stat_threshold);
 355
 356         if (unlikely(abs(x) > t)) {
 357                 node_page_state_add(x, pgdat, item);
 358                 x = 0;
 359         }
 360         __this_cpu_write(*p, x);
 361         preempt_enable_rt();
 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_pageset __percpu *pcp = zone->pageset;
 391         s8 __percpu *p = pcp->vm_stat_diff + item;
 392         s8 v, t;
 393
 394         preempt_disable_rt();
 395         v = __this_cpu_inc_return(*p);
 396         t = __this_cpu_read(pcp->stat_threshold);
 397         if (unlikely(v > t)) {
 398                 s8 overstep = t >> 1;
 399
 400                 zone_page_state_add(v + overstep, zone, item);
 401                 __this_cpu_write(*p, -overstep);
 402         }
 403         preempt_enable_rt();
 404 }
 405
 406 void __inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 407 {
 408         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 409         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 410         s8 v, t;
 411
 412         VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
 413
 414         preempt_disable_rt();
 415         v = __this_cpu_inc_return(*p);
 416         t = __this_cpu_read(pcp->stat_threshold);
 417         if (unlikely(v > t)) {
 418                 s8 overstep = t >> 1;
 419
 420                 node_page_state_add(v + overstep, pgdat, item);
 421                 __this_cpu_write(*p, -overstep);
 422         }
 423         preempt_enable_rt();
 424 }
 425
 426 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
 427 {
 428         __inc_zone_state(page_zone(page), item);
 429 }
 430 EXPORT_SYMBOL(__inc_zone_page_state);
 431
 432 void __inc_node_page_state(struct page *page, enum node_stat_item item)
 433 {
 434         __inc_node_state(page_pgdat(page), item);
 435 }
 436 EXPORT_SYMBOL(__inc_node_page_state);
 437
 438 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
 439 {
 440         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 441         s8 __percpu *p = pcp->vm_stat_diff + item;
 442         s8 v, t;
 443
 444         preempt_disable_rt();
 445         v = __this_cpu_dec_return(*p);
 446         t = __this_cpu_read(pcp->stat_threshold);
 447         if (unlikely(v < - t)) {
 448                 s8 overstep = t >> 1;
 449
 450                 zone_page_state_add(v - overstep, zone, item);
 451                 __this_cpu_write(*p, overstep);
 452         }
 453         preempt_enable_rt();
 454 }
 455
 456 void __dec_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 457 {
 458         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 459         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 460         s8 v, t;
 461
 462         VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
 463
 464         preempt_disable_rt();
 465         v = __this_cpu_dec_return(*p);
 466         t = __this_cpu_read(pcp->stat_threshold);
 467         if (unlikely(v < - t)) {
 468                 s8 overstep = t >> 1;
 469
 470                 node_page_state_add(v - overstep, pgdat, item);
 471                 __this_cpu_write(*p, overstep);
 472         }
 473         preempt_enable_rt();
 474 }
 475
 476 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
 477 {
 478         __dec_zone_state(page_zone(page), item);
 479 }
 480 EXPORT_SYMBOL(__dec_zone_page_state);
 481
 482 void __dec_node_page_state(struct page *page, enum node_stat_item item)
 483 {
 484         __dec_node_state(page_pgdat(page), item);
 485 }
 486 EXPORT_SYMBOL(__dec_node_page_state);
 487
 488 #ifdef CONFIG_HAVE_CMPXCHG_LOCAL
 489 /*
 490  * If we have cmpxchg_local support then we do not need to incur the overhead
 491  * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
 492  *
 493  * mod_state() modifies the zone counter state through atomic per cpu
 494  * operations.
 495  *
 496  * Overstep mode specifies how overstep should handled:
 497  *     0       No overstepping
 498  *     1       Overstepping half of threshold
 499  *     -1      Overstepping minus half of threshold
 500 */
 501 static inline void mod_zone_state(struct zone *zone,
 502        enum zone_stat_item item, long delta, int overstep_mode)
 503 {
 504         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 505         s8 __percpu *p = pcp->vm_stat_diff + item;
 506         long o, n, t, z;
 507
 508         do {
 509                 z = 0;  /* overflow to zone counters */
 510
 511                 /*
 512                  * The fetching of the stat_threshold is racy. We may apply
 513                  * a counter threshold to the wrong the cpu if we get
 514                  * rescheduled while executing here. However, the next
 515                  * counter update will apply the threshold again and
 516                  * therefore bring the counter under the threshold again.
 517                  *
 518                  * Most of the time the thresholds are the same anyways
 519                  * for all cpus in a zone.
 520                  */
 521                 t = this_cpu_read(pcp->stat_threshold);
 522
 523                 o = this_cpu_read(*p);
 524                 n = delta + o;
 525
 526                 if (abs(n) > t) {
 527                         int os = overstep_mode * (t >> 1) ;
 528
 529                         /* Overflow must be added to zone counters */
 530                         z = n + os;
 531                         n = -os;
 532                 }
 533         } while (this_cpu_cmpxchg(*p, o, n) != o);
 534
 535         if (z)
 536                 zone_page_state_add(z, zone, item);
 537 }
 538
 539 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 540                          long delta)
 541 {
 542         mod_zone_state(zone, item, delta, 0);
 543 }
 544 EXPORT_SYMBOL(mod_zone_page_state);
 545
 546 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 547 {
 548         mod_zone_state(page_zone(page), item, 1, 1);
 549 }
 550 EXPORT_SYMBOL(inc_zone_page_state);
 551
 552 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 553 {
 554         mod_zone_state(page_zone(page), item, -1, -1);
 555 }
 556 EXPORT_SYMBOL(dec_zone_page_state);
 557
 558 static inline void mod_node_state(struct pglist_data *pgdat,
 559        enum node_stat_item item, int delta, int overstep_mode)
 560 {
 561         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 562         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 563         long o, n, t, z;
 564
 565         if (vmstat_item_in_bytes(item)) {
 566                 VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
 567                 delta >>= PAGE_SHIFT;
 568         }
 569
 570         do {
 571                 z = 0;  /* overflow to node counters */
 572
 573                 /*
 574                  * The fetching of the stat_threshold is racy. We may apply
 575                  * a counter threshold to the wrong the cpu if we get
 576                  * rescheduled while executing here. However, the next
 577                  * counter update will apply the threshold again and
 578                  * therefore bring the counter under the threshold again.
 579                  *
 580                  * Most of the time the thresholds are the same anyways
 581                  * for all cpus in a node.
 582                  */
 583                 t = this_cpu_read(pcp->stat_threshold);
 584
 585                 o = this_cpu_read(*p);
 586                 n = delta + o;
 587
 588                 if (abs(n) > t) {
 589                         int os = overstep_mode * (t >> 1) ;
 590
 591                         /* Overflow must be added to node counters */
 592                         z = n + os;
 593                         n = -os;
 594                 }
 595         } while (this_cpu_cmpxchg(*p, o, n) != o);
 596
 597         if (z)
 598                 node_page_state_add(z, pgdat, item);
 599 }
 600
 601 void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
 602                                         long delta)
 603 {
 604         mod_node_state(pgdat, item, delta, 0);
 605 }
 606 EXPORT_SYMBOL(mod_node_page_state);
 607
 608 void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 609 {
 610         mod_node_state(pgdat, item, 1, 1);
 611 }
 612
 613 void inc_node_page_state(struct page *page, enum node_stat_item item)
 614 {
 615         mod_node_state(page_pgdat(page), item, 1, 1);
 616 }
 617 EXPORT_SYMBOL(inc_node_page_state);
 618
 619 void dec_node_page_state(struct page *page, enum node_stat_item item)
 620 {
 621         mod_node_state(page_pgdat(page), item, -1, -1);
 622 }
 623 EXPORT_SYMBOL(dec_node_page_state);
 624 #else
 625 /*
 626  * Use interrupt disable to serialize counter updates
 627  */
 628 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 629                          long delta)
 630 {
 631         unsigned long flags;
 632
 633         local_irq_save(flags);
 634         __mod_zone_page_state(zone, item, delta);
 635         local_irq_restore(flags);
 636 }
 637 EXPORT_SYMBOL(mod_zone_page_state);
 638
 639 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 640 {
 641         unsigned long flags;
 642         struct zone *zone;
 643
 644         zone = page_zone(page);
 645         local_irq_save(flags);
 646         __inc_zone_state(zone, item);
 647         local_irq_restore(flags);
 648 }
 649 EXPORT_SYMBOL(inc_zone_page_state);
 650
 651 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 652 {
 653         unsigned long flags;
 654
 655         local_irq_save(flags);
 656         __dec_zone_page_state(page, item);
 657         local_irq_restore(flags);
 658 }
 659 EXPORT_SYMBOL(dec_zone_page_state);
 660
 661 void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 662 {
 663         unsigned long flags;
 664
 665         local_irq_save(flags);
 666         __inc_node_state(pgdat, item);
 667         local_irq_restore(flags);
 668 }
 669 EXPORT_SYMBOL(inc_node_state);
 670
 671 void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
 672                                         long delta)
 673 {
 674         unsigned long flags;
 675
 676         local_irq_save(flags);
 677         __mod_node_page_state(pgdat, item, delta);
 678         local_irq_restore(flags);
 679 }
 680 EXPORT_SYMBOL(mod_node_page_state);
 681
 682 void inc_node_page_state(struct page *page, enum node_stat_item item)
 683 {
 684         unsigned long flags;
 685         struct pglist_data *pgdat;
 686
 687         pgdat = page_pgdat(page);
 688         local_irq_save(flags);
 689         __inc_node_state(pgdat, item);
 690         local_irq_restore(flags);
 691 }
 692 EXPORT_SYMBOL(inc_node_page_state);
 693
 694 void dec_node_page_state(struct page *page, enum node_stat_item item)
 695 {
 696         unsigned long flags;
 697
 698         local_irq_save(flags);
 699         __dec_node_page_state(page, item);
 700         local_irq_restore(flags);
 701 }
 702 EXPORT_SYMBOL(dec_node_page_state);
 703 #endif
 704
 705 /*
 706  * Fold a differential into the global counters.
 707  * Returns the number of counters updated.
 708  */
 709 #ifdef CONFIG_NUMA
 710 static int fold_diff(int *zone_diff, int *numa_diff, int *node_diff)
 711 {
 712         int i;
 713         int changes = 0;
 714
 715         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 716                 if (zone_diff[i]) {
 717                         atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
 718                         changes++;
 719         }
 720
 721         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
 722                 if (numa_diff[i]) {
 723                         atomic_long_add(numa_diff[i], &vm_numa_stat[i]);
 724                         changes++;
 725         }
 726
 727         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 728                 if (node_diff[i]) {
 729                         atomic_long_add(node_diff[i], &vm_node_stat[i]);
 730                         changes++;
 731         }
 732         return changes;
 733 }
 734 #else
 735 static int fold_diff(int *zone_diff, int *node_diff)
 736 {
 737         int i;
 738         int changes = 0;
 739
 740         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 741                 if (zone_diff[i]) {
 742                         atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
 743                         changes++;
 744         }
 745
 746         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 747                 if (node_diff[i]) {
 748                         atomic_long_add(node_diff[i], &vm_node_stat[i]);
 749                         changes++;
 750         }
 751         return changes;
 752 }
 753 #endif /* CONFIG_NUMA */
 754
 755 /*
 756  * Update the zone counters for the current cpu.
 757  *
 758  * Note that refresh_cpu_vm_stats strives to only access
 759  * node local memory. The per cpu pagesets on remote zones are placed
 760  * in the memory local to the processor using that pageset. So the
 761  * loop over all zones will access a series of cachelines local to
 762  * the processor.
 763  *
 764  * The call to zone_page_state_add updates the cachelines with the
 765  * statistics in the remote zone struct as well as the global cachelines
 766  * with the global counters. These could cause remote node cache line
 767  * bouncing and will have to be only done when necessary.
 768  *
 769  * The function returns the number of global counters updated.
 770  */
 771 static int refresh_cpu_vm_stats(bool do_pagesets)
 772 {
 773         struct pglist_data *pgdat;
 774         struct zone *zone;
 775         int i;
 776         int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 777 #ifdef CONFIG_NUMA
 778         int global_numa_diff[NR_VM_NUMA_STAT_ITEMS] = { 0, };
 779 #endif
 780         int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
 781         int changes = 0;
 782
 783         for_each_populated_zone(zone) {
 784                 struct per_cpu_pageset __percpu *p = zone->pageset;
 785
 786                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
 787                         int v;
 788
 789                         v = this_cpu_xchg(p->vm_stat_diff[i], 0);
 790                         if (v) {
 791
 792                                 atomic_long_add(v, &zone->vm_stat[i]);
 793                                 global_zone_diff[i] += v;
 794 #ifdef CONFIG_NUMA
 795                                 /* 3 seconds idle till flush */
 796                                 __this_cpu_write(p->expire, 3);
 797 #endif
 798                         }
 799                 }
 800 #ifdef CONFIG_NUMA
 801                 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) {
 802                         int v;
 803
 804                         v = this_cpu_xchg(p->vm_numa_stat_diff[i], 0);
 805                         if (v) {
 806
 807                                 atomic_long_add(v, &zone->vm_numa_stat[i]);
 808                                 global_numa_diff[i] += v;
 809                                 __this_cpu_write(p->expire, 3);
 810                         }
 811                 }
 812
 813                 if (do_pagesets) {
 814                         cond_resched();
 815                         /*
 816                          * Deal with draining the remote pageset of this
 817                          * processor
 818                          *
 819                          * Check if there are pages remaining in this pageset
 820                          * if not then there is nothing to expire.
 821                          */
 822                         if (!__this_cpu_read(p->expire) ||
 823                                !__this_cpu_read(p->pcp.count))
 824                                 continue;
 825
 826                         /*
 827                          * We never drain zones local to this processor.
 828                          */
 829                         if (zone_to_nid(zone) == numa_node_id()) {
 830                                 __this_cpu_write(p->expire, 0);
 831                                 continue;
 832                         }
 833
 834                         if (__this_cpu_dec_return(p->expire))
 835                                 continue;
 836
 837                         if (__this_cpu_read(p->pcp.count)) {
 838                                 drain_zone_pages(zone, this_cpu_ptr(&p->pcp));
 839                                 changes++;
 840                         }
 841                 }
 842 #endif
 843         }
 844
 845         for_each_online_pgdat(pgdat) {
 846                 struct per_cpu_nodestat __percpu *p = pgdat->per_cpu_nodestats;
 847
 848                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
 849                         int v;
 850
 851                         v = this_cpu_xchg(p->vm_node_stat_diff[i], 0);
 852                         if (v) {
 853                                 atomic_long_add(v, &pgdat->vm_stat[i]);
 854                                 global_node_diff[i] += v;
 855                         }
 856                 }
 857         }
 858
 859 #ifdef CONFIG_NUMA
 860         changes += fold_diff(global_zone_diff, global_numa_diff,
 861                              global_node_diff);
 862 #else
 863         changes += fold_diff(global_zone_diff, global_node_diff);
 864 #endif
 865         return changes;
 866 }
 867
 868 /*
 869  * Fold the data for an offline cpu into the global array.
 870  * There cannot be any access by the offline cpu and therefore
 871  * synchronization is simplified.
 872  */
 873 void cpu_vm_stats_fold(int cpu)
 874 {
 875         struct pglist_data *pgdat;
 876         struct zone *zone;
 877         int i;
 878         int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 879 #ifdef CONFIG_NUMA
 880         int global_numa_diff[NR_VM_NUMA_STAT_ITEMS] = { 0, };
 881 #endif
 882         int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
 883
 884         for_each_populated_zone(zone) {
 885                 struct per_cpu_pageset *p;
 886
 887                 p = per_cpu_ptr(zone->pageset, cpu);
 888
 889                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 890                         if (p->vm_stat_diff[i]) {
 891                                 int v;
 892
 893                                 v = p->vm_stat_diff[i];
 894                                 p->vm_stat_diff[i] = 0;
 895                                 atomic_long_add(v, &zone->vm_stat[i]);
 896                                 global_zone_diff[i] += v;
 897                         }
 898
 899 #ifdef CONFIG_NUMA
 900                 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
 901                         if (p->vm_numa_stat_diff[i]) {
 902                                 int v;
 903
 904                                 v = p->vm_numa_stat_diff[i];
 905                                 p->vm_numa_stat_diff[i] = 0;
 906                                 atomic_long_add(v, &zone->vm_numa_stat[i]);
 907                                 global_numa_diff[i] += v;
 908                         }
 909 #endif
 910         }
 911
 912         for_each_online_pgdat(pgdat) {
 913                 struct per_cpu_nodestat *p;
 914
 915                 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
 916
 917                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 918                         if (p->vm_node_stat_diff[i]) {
 919                                 int v;
 920
 921                                 v = p->vm_node_stat_diff[i];
 922                                 p->vm_node_stat_diff[i] = 0;
 923                                 atomic_long_add(v, &pgdat->vm_stat[i]);
 924                                 global_node_diff[i] += v;
 925                         }
 926         }
 927
 928 #ifdef CONFIG_NUMA
 929         fold_diff(global_zone_diff, global_numa_diff, global_node_diff);
 930 #else
 931         fold_diff(global_zone_diff, global_node_diff);
 932 #endif
 933 }
 934
 935 /*
 936  * this is only called if !populated_zone(zone), which implies no other users of
 937  * pset->vm_stat_diff[] exsist.
 938  */
 939 void drain_zonestat(struct zone *zone, struct per_cpu_pageset *pset)
 940 {
 941         int i;
 942
 943         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 944                 if (pset->vm_stat_diff[i]) {
 945                         int v = pset->vm_stat_diff[i];
 946                         pset->vm_stat_diff[i] = 0;
 947                         atomic_long_add(v, &zone->vm_stat[i]);
 948                         atomic_long_add(v, &vm_zone_stat[i]);
 949                 }
 950
 951 #ifdef CONFIG_NUMA
 952         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
 953                 if (pset->vm_numa_stat_diff[i]) {
 954                         int v = pset->vm_numa_stat_diff[i];
 955
 956                         pset->vm_numa_stat_diff[i] = 0;
 957                         atomic_long_add(v, &zone->vm_numa_stat[i]);
 958                         atomic_long_add(v, &vm_numa_stat[i]);
 959                 }
 960 #endif
 961 }
 962 #endif
 963
 964 #ifdef CONFIG_NUMA
 965 void __inc_numa_state(struct zone *zone,
 966                                  enum numa_stat_item item)
 967 {
 968         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 969         u16 __percpu *p = pcp->vm_numa_stat_diff + item;
 970         u16 v;
 971
 972         v = __this_cpu_inc_return(*p);
 973
 974         if (unlikely(v > NUMA_STATS_THRESHOLD)) {
 975                 zone_numa_state_add(v, zone, item);
 976                 __this_cpu_write(*p, 0);
 977         }
 978 }
 979
 980 /*
 981  * Determine the per node value of a stat item. This function
 982  * is called frequently in a NUMA machine, so try to be as
 983  * frugal as possible.
 984  */
 985 unsigned long sum_zone_node_page_state(int node,
 986                                  enum zone_stat_item item)
 987 {
 988         struct zone *zones = NODE_DATA(node)->node_zones;
 989         int i;
 990         unsigned long count = 0;
 991
 992         for (i = 0; i < MAX_NR_ZONES; i++)
 993                 count += zone_page_state(zones + i, item);
 994
 995         return count;
 996 }
 997
 998 /*
 999  * Determine the per node value of a numa stat item. To avoid deviation,
1000  * the per cpu stat number in vm_numa_stat_diff[] is also included.
1001  */
1002 unsigned long sum_zone_numa_state(int node,
1003                                  enum numa_stat_item item)
1004 {
1005         struct zone *zones = NODE_DATA(node)->node_zones;
1006         int i;
1007         unsigned long count = 0;
1008
1009         for (i = 0; i < MAX_NR_ZONES; i++)
1010                 count += zone_numa_state_snapshot(zones + i, item);
1011
1012         return count;
1013 }
1014
1015 /*
1016  * Determine the per node value of a stat item.
1017  */
1018 unsigned long node_page_state_pages(struct pglist_data *pgdat,
1019                                     enum node_stat_item item)
1020 {
1021         long x = atomic_long_read(&pgdat->vm_stat[item]);
1022 #ifdef CONFIG_SMP
1023         if (x < 0)
1024                 x = 0;
1025 #endif
1026         return x;
1027 }
1028
1029 unsigned long node_page_state(struct pglist_data *pgdat,
1030                               enum node_stat_item item)
1031 {
1032         VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
1033
1034         return node_page_state_pages(pgdat, item);
1035 }
1036 #endif
1037
1038 #ifdef CONFIG_COMPACTION
1039
1040 struct contig_page_info {
1041         unsigned long free_pages;
1042         unsigned long free_blocks_total;
1043         unsigned long free_blocks_suitable;
1044 };
1045
1046 /*
1047  * Calculate the number of free pages in a zone, how many contiguous
1048  * pages are free and how many are large enough to satisfy an allocation of
1049  * the target size. Note that this function makes no attempt to estimate
1050  * how many suitable free blocks there *might* be if MOVABLE pages were
1051  * migrated. Calculating that is possible, but expensive and can be
1052  * figured out from userspace
1053  */
1054 static void fill_contig_page_info(struct zone *zone,
1055                                 unsigned int suitable_order,
1056                                 struct contig_page_info *info)
1057 {
1058         unsigned int order;
1059
1060         info->free_pages = 0;
1061         info->free_blocks_total = 0;
1062         info->free_blocks_suitable = 0;
1063
1064         for (order = 0; order < MAX_ORDER; order++) {
1065                 unsigned long blocks;
1066
1067                 /* Count number of free blocks */
1068                 blocks = zone->free_area[order].nr_free;
1069                 info->free_blocks_total += blocks;
1070
1071                 /* Count free base pages */
1072                 info->free_pages += blocks << order;
1073
1074                 /* Count the suitable free blocks */
1075                 if (order >= suitable_order)
1076                         info->free_blocks_suitable += blocks <<
1077                                                 (order - suitable_order);
1078         }
1079 }
1080
1081 /*
1082  * A fragmentation index only makes sense if an allocation of a requested
1083  * size would fail. If that is true, the fragmentation index indicates
1084  * whether external fragmentation or a lack of memory was the problem.
1085  * The value can be used to determine if page reclaim or compaction
1086  * should be used
1087  */
1088 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
1089 {
1090         unsigned long requested = 1UL << order;
1091
1092         if (WARN_ON_ONCE(order >= MAX_ORDER))
1093                 return 0;
1094
1095         if (!info->free_blocks_total)
1096                 return 0;
1097
1098         /* Fragmentation index only makes sense when a request would fail */
1099         if (info->free_blocks_suitable)
1100                 return -1000;
1101
1102         /*
1103          * Index is between 0 and 1 so return within 3 decimal places
1104          *
1105          * 0 => allocation would fail due to lack of memory
1106          * 1 => allocation would fail due to fragmentation
1107          */
1108         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
1109 }
1110
1111 /*
1112  * Calculates external fragmentation within a zone wrt the given order.
1113  * It is defined as the percentage of pages found in blocks of size
1114  * less than 1 << order. It returns values in range [0, 100].
1115  */
1116 unsigned int extfrag_for_order(struct zone *zone, unsigned int order)
1117 {
1118         struct contig_page_info info;
1119
1120         fill_contig_page_info(zone, order, &info);
1121         if (info.free_pages == 0)
1122                 return 0;
1123
1124         return div_u64((info.free_pages -
1125                         (info.free_blocks_suitable << order)) * 100,
1126                         info.free_pages);
1127 }
1128
1129 /* Same as __fragmentation index but allocs contig_page_info on stack */
1130 int fragmentation_index(struct zone *zone, unsigned int order)
1131 {
1132         struct contig_page_info info;
1133
1134         fill_contig_page_info(zone, order, &info);
1135         return __fragmentation_index(order, &info);
1136 }
1137 #endif
1138
1139 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || \
1140     defined(CONFIG_NUMA) || defined(CONFIG_MEMCG)
1141 #ifdef CONFIG_ZONE_DMA
1142 #define TEXT_FOR_DMA(xx) xx "_dma",
1143 #else
1144 #define TEXT_FOR_DMA(xx)
1145 #endif
1146
1147 #ifdef CONFIG_ZONE_DMA32
1148 #define TEXT_FOR_DMA32(xx) xx "_dma32",
1149 #else
1150 #define TEXT_FOR_DMA32(xx)
1151 #endif
1152
1153 #ifdef CONFIG_HIGHMEM
1154 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
1155 #else
1156 #define TEXT_FOR_HIGHMEM(xx)
1157 #endif
1158
1159 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1160                                         TEXT_FOR_HIGHMEM(xx) xx "_movable",
1161
1162 const char * const vmstat_text[] = {
1163         /* enum zone_stat_item counters */
1164         "nr_free_pages",
1165         "nr_zone_inactive_anon",
1166         "nr_zone_active_anon",
1167         "nr_zone_inactive_file",
1168         "nr_zone_active_file",
1169         "nr_zone_unevictable",
1170         "nr_zone_write_pending",
1171         "nr_mlock",
1172         "nr_page_table_pages",
1173         "nr_bounce",
1174 #if IS_ENABLED(CONFIG_ZSMALLOC)
1175         "nr_zspages",
1176 #endif
1177         "nr_free_cma",
1178
1179         /* enum numa_stat_item counters */
1180 #ifdef CONFIG_NUMA
1181         "numa_hit",
1182         "numa_miss",
1183         "numa_foreign",
1184         "numa_interleave",
1185         "numa_local",
1186         "numa_other",
1187 #endif
1188
1189         /* enum node_stat_item counters */
1190         "nr_inactive_anon",
1191         "nr_active_anon",
1192         "nr_inactive_file",
1193         "nr_active_file",
1194         "nr_unevictable",
1195         "nr_slab_reclaimable",
1196         "nr_slab_unreclaimable",
1197         "nr_isolated_anon",
1198         "nr_isolated_file",
1199         "workingset_nodes",
1200         "workingset_refault_anon",
1201         "workingset_refault_file",
1202         "workingset_activate_anon",
1203         "workingset_activate_file",
1204         "workingset_restore_anon",
1205         "workingset_restore_file",
1206         "workingset_nodereclaim",
1207         "nr_anon_pages",
1208         "nr_mapped",
1209         "nr_file_pages",
1210         "nr_dirty",
1211         "nr_writeback",
1212         "nr_writeback_temp",
1213         "nr_shmem",
1214         "nr_shmem_hugepages",
1215 #ifdef CONFIG_FINEGRAINED_THP
1216         "nr_shmem_64kb_hugepages",
1217 #endif
1218         "nr_shmem_pmdmapped",
1219 #ifdef CONFIG_FINEGRAINED_THP
1220         "nr_shmem_ptemapped",
1221         "nr_file_64kb_hugepages",
1222 #endif
1223         "nr_file_hugepages",
1224         "nr_file_pmdmapped",
1225 #ifdef CONFIG_FINEGRAINED_THP
1226         "nr_file_ptemapped",
1227 #endif
1228         "nr_anon_transparent_hugepages",
1229 #ifdef CONFIG_FINEGRAINED_THP
1230         "nr_anon_64KB_transparent_hugepages",
1231 #endif
1232         "nr_vmscan_write",
1233         "nr_vmscan_immediate_reclaim",
1234         "nr_dirtied",
1235         "nr_written",
1236         "nr_kernel_misc_reclaimable",
1237         "nr_foll_pin_acquired",
1238         "nr_foll_pin_released",
1239         "nr_kernel_stack",
1240 #if IS_ENABLED(CONFIG_SHADOW_CALL_STACK)
1241         "nr_shadow_call_stack",
1242 #endif
1243
1244         /* enum writeback_stat_item counters */
1245         "nr_dirty_threshold",
1246         "nr_dirty_background_threshold",
1247
1248 #if defined(CONFIG_VM_EVENT_COUNTERS) || defined(CONFIG_MEMCG)
1249         /* enum vm_event_item counters */
1250         "pgpgin",
1251         "pgpgout",
1252         "pswpin",
1253         "pswpout",
1254
1255         TEXTS_FOR_ZONES("pgalloc")
1256         TEXTS_FOR_ZONES("allocstall")
1257         TEXTS_FOR_ZONES("pgskip")
1258
1259         "pgfree",
1260         "pgactivate",
1261         "pgdeactivate",
1262         "pglazyfree",
1263
1264         "pgfault",
1265         "pgmajfault",
1266         "pglazyfreed",
1267
1268         "pgrefill",
1269         "pgreuse",
1270         "pgsteal_kswapd",
1271         "pgsteal_direct",
1272         "pgscan_kswapd",
1273         "pgscan_direct",
1274         "pgscan_direct_throttle",
1275         "pgscan_anon",
1276         "pgscan_file",
1277         "pgsteal_anon",
1278         "pgsteal_file",
1279
1280 #ifdef CONFIG_NUMA
1281         "zone_reclaim_failed",
1282 #endif
1283         "pginodesteal",
1284         "slabs_scanned",
1285         "kswapd_inodesteal",
1286         "kswapd_low_wmark_hit_quickly",
1287         "kswapd_high_wmark_hit_quickly",
1288         "pageoutrun",
1289
1290         "pgrotated",
1291
1292         "drop_pagecache",
1293         "drop_slab",
1294         "oom_kill",
1295
1296 #ifdef CONFIG_NUMA_BALANCING
1297         "numa_pte_updates",
1298         "numa_huge_pte_updates",
1299         "numa_hint_faults",
1300         "numa_hint_faults_local",
1301         "numa_pages_migrated",
1302 #endif
1303 #ifdef CONFIG_MIGRATION
1304         "pgmigrate_success",
1305         "pgmigrate_fail",
1306         "thp_migration_success",
1307         "thp_migration_fail",
1308         "thp_migration_split",
1309 #endif
1310 #ifdef CONFIG_COMPACTION
1311         "compact_migrate_scanned",
1312         "compact_free_scanned",
1313         "compact_isolated",
1314         "compact_stall",
1315         "compact_fail",
1316         "compact_success",
1317         "compact_daemon_wake",
1318         "compact_daemon_migrate_scanned",
1319         "compact_daemon_free_scanned",
1320 #endif
1321
1322 #ifdef CONFIG_HUGETLB_PAGE
1323         "htlb_buddy_alloc_success",
1324         "htlb_buddy_alloc_fail",
1325 #endif
1326 #ifdef CONFIG_CMA
1327         "cma_alloc_success",
1328         "cma_alloc_fail",
1329 #endif
1330         "unevictable_pgs_culled",
1331         "unevictable_pgs_scanned",
1332         "unevictable_pgs_rescued",
1333         "unevictable_pgs_mlocked",
1334         "unevictable_pgs_munlocked",
1335         "unevictable_pgs_cleared",
1336         "unevictable_pgs_stranded",
1337
1338 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1339         "thp_fault_alloc",
1340         "thp_fault_fallback",
1341         "thp_fault_fallback_charge",
1342         "thp_collapse_alloc",
1343         "thp_collapse_alloc_failed",
1344         "thp_file_alloc",
1345         "thp_file_fallback",
1346         "thp_file_fallback_charge",
1347         "thp_file_mapped",
1348         "thp_split_page",
1349         "thp_split_page_failed",
1350         "thp_deferred_split_page",
1351 #ifdef CONFIG_FINEGRAINED_THP
1352         "thp_split_cont_pte",
1353 #endif
1354         "thp_split_pmd",
1355 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1356         "thp_split_pud",
1357 #endif
1358         "thp_zero_page_alloc",
1359         "thp_zero_page_alloc_failed",
1360         "thp_swpout",
1361         "thp_swpout_fallback",
1362 #endif
1363 #ifdef CONFIG_MEMORY_BALLOON
1364         "balloon_inflate",
1365         "balloon_deflate",
1366 #ifdef CONFIG_BALLOON_COMPACTION
1367         "balloon_migrate",
1368 #endif
1369 #endif /* CONFIG_MEMORY_BALLOON */
1370 #ifdef CONFIG_DEBUG_TLBFLUSH
1371         "nr_tlb_remote_flush",
1372         "nr_tlb_remote_flush_received",
1373         "nr_tlb_local_flush_all",
1374         "nr_tlb_local_flush_one",
1375 #endif /* CONFIG_DEBUG_TLBFLUSH */
1376
1377 #ifdef CONFIG_DEBUG_VM_VMACACHE
1378         "vmacache_find_calls",
1379         "vmacache_find_hits",
1380 #endif
1381 #ifdef CONFIG_SWAP
1382         "swap_ra",
1383         "swap_ra_hit",
1384 #endif
1385 #endif /* CONFIG_VM_EVENT_COUNTERS || CONFIG_MEMCG */
1386 };
1387 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA || CONFIG_MEMCG */
1388
1389 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1390      defined(CONFIG_PROC_FS)
1391 static void *frag_start(struct seq_file *m, loff_t *pos)
1392 {
1393         pg_data_t *pgdat;
1394         loff_t node = *pos;
1395
1396         for (pgdat = first_online_pgdat();
1397              pgdat && node;
1398              pgdat = next_online_pgdat(pgdat))
1399                 --node;
1400
1401         return pgdat;
1402 }
1403
1404 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1405 {
1406         pg_data_t *pgdat = (pg_data_t *)arg;
1407
1408         (*pos)++;
1409         return next_online_pgdat(pgdat);
1410 }
1411
1412 static void frag_stop(struct seq_file *m, void *arg)
1413 {
1414 }
1415
1416 /*
1417  * Walk zones in a node and print using a callback.
1418  * If @assert_populated is true, only use callback for zones that are populated.
1419  */
1420 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1421                 bool assert_populated, bool nolock,
1422                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1423 {
1424         struct zone *zone;
1425         struct zone *node_zones = pgdat->node_zones;
1426         unsigned long flags;
1427
1428         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1429                 if (assert_populated && !populated_zone(zone))
1430                         continue;
1431
1432                 if (!nolock)
1433                         spin_lock_irqsave(&zone->lock, flags);
1434                 print(m, pgdat, zone);
1435                 if (!nolock)
1436                         spin_unlock_irqrestore(&zone->lock, flags);
1437         }
1438 }
1439 #endif
1440
1441 #ifdef CONFIG_PROC_FS
1442 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1443                                                 struct zone *zone)
1444 {
1445         int order;
1446
1447         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1448         for (order = 0; order < MAX_ORDER; ++order)
1449                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
1450         seq_putc(m, '\n');
1451 }
1452
1453 /*
1454  * This walks the free areas for each zone.
1455  */
1456 static int frag_show(struct seq_file *m, void *arg)
1457 {
1458         pg_data_t *pgdat = (pg_data_t *)arg;
1459         walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1460         return 0;
1461 }
1462
1463 static void pagetypeinfo_showfree_print(struct seq_file *m,
1464                                         pg_data_t *pgdat, struct zone *zone)
1465 {
1466         int order, mtype;
1467
1468         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1469                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
1470                                         pgdat->node_id,
1471                                         zone->name,
1472                                         migratetype_names[mtype]);
1473                 for (order = 0; order < MAX_ORDER; ++order) {
1474                         unsigned long freecount = 0;
1475                         struct free_area *area;
1476                         struct list_head *curr;
1477                         bool overflow = false;
1478
1479                         area = &(zone->free_area[order]);
1480
1481                         list_for_each(curr, &area->free_list[mtype]) {
1482                                 /*
1483                                  * Cap the free_list iteration because it might
1484                                  * be really large and we are under a spinlock
1485                                  * so a long time spent here could trigger a
1486                                  * hard lockup detector. Anyway this is a
1487                                  * debugging tool so knowing there is a handful
1488                                  * of pages of this order should be more than
1489                                  * sufficient.
1490                                  */
1491                                 if (++freecount >= 100000) {
1492                                         overflow = true;
1493                                         break;
1494                                 }
1495                         }
1496                         seq_printf(m, "%s%6lu ", overflow ? ">" : "", freecount);
1497                         spin_unlock_irq(&zone->lock);
1498                         cond_resched();
1499                         spin_lock_irq(&zone->lock);
1500                 }
1501                 seq_putc(m, '\n');
1502         }
1503 }
1504
1505 /* Print out the free pages at each order for each migatetype */
1506 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
1507 {
1508         int order;
1509         pg_data_t *pgdat = (pg_data_t *)arg;
1510
1511         /* Print header */
1512         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1513         for (order = 0; order < MAX_ORDER; ++order)
1514                 seq_printf(m, "%6d ", order);
1515         seq_putc(m, '\n');
1516
1517         walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1518
1519         return 0;
1520 }
1521
1522 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1523                                         pg_data_t *pgdat, struct zone *zone)
1524 {
1525         int mtype;
1526         unsigned long pfn;
1527         unsigned long start_pfn = zone->zone_start_pfn;
1528         unsigned long end_pfn = zone_end_pfn(zone);
1529         unsigned long count[MIGRATE_TYPES] = { 0, };
1530
1531         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1532                 struct page *page;
1533
1534                 page = pfn_to_online_page(pfn);
1535                 if (!page)
1536                         continue;
1537
1538                 /* Watch for unexpected holes punched in the memmap */
1539                 if (!memmap_valid_within(pfn, page, zone))
1540                         continue;
1541
1542                 if (page_zone(page) != zone)
1543                         continue;
1544
1545                 mtype = get_pageblock_migratetype(page);
1546
1547                 if (mtype < MIGRATE_TYPES)
1548                         count[mtype]++;
1549         }
1550
1551         /* Print counts */
1552         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1553         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1554                 seq_printf(m, "%12lu ", count[mtype]);
1555         seq_putc(m, '\n');
1556 }
1557
1558 /* Print out the number of pageblocks for each migratetype */
1559 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1560 {
1561         int mtype;
1562         pg_data_t *pgdat = (pg_data_t *)arg;
1563
1564         seq_printf(m, "\n%-23s", "Number of blocks type ");
1565         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1566                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1567         seq_putc(m, '\n');
1568         walk_zones_in_node(m, pgdat, true, false,
1569                 pagetypeinfo_showblockcount_print);
1570
1571         return 0;
1572 }
1573
1574 /*
1575  * Print out the number of pageblocks for each migratetype that contain pages
1576  * of other types. This gives an indication of how well fallbacks are being
1577  * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1578  * to determine what is going on
1579  */
1580 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1581 {
1582 #ifdef CONFIG_PAGE_OWNER
1583         int mtype;
1584
1585         if (!static_branch_unlikely(&page_owner_inited))
1586                 return;
1587
1588         drain_all_pages(NULL);
1589
1590         seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1591         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1592                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1593         seq_putc(m, '\n');
1594
1595         walk_zones_in_node(m, pgdat, true, true,
1596                 pagetypeinfo_showmixedcount_print);
1597 #endif /* CONFIG_PAGE_OWNER */
1598 }
1599
1600 /*
1601  * This prints out statistics in relation to grouping pages by mobility.
1602  * It is expensive to collect so do not constantly read the file.
1603  */
1604 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1605 {
1606         pg_data_t *pgdat = (pg_data_t *)arg;
1607
1608         /* check memoryless node */
1609         if (!node_state(pgdat->node_id, N_MEMORY))
1610                 return 0;
1611
1612         seq_printf(m, "Page block order: %d\n", pageblock_order);
1613         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
1614         seq_putc(m, '\n');
1615         pagetypeinfo_showfree(m, pgdat);
1616         pagetypeinfo_showblockcount(m, pgdat);
1617         pagetypeinfo_showmixedcount(m, pgdat);
1618
1619         return 0;
1620 }
1621
1622 static const struct seq_operations fragmentation_op = {
1623         .start  = frag_start,
1624         .next   = frag_next,
1625         .stop   = frag_stop,
1626         .show   = frag_show,
1627 };
1628
1629 static const struct seq_operations pagetypeinfo_op = {
1630         .start  = frag_start,
1631         .next   = frag_next,
1632         .stop   = frag_stop,
1633         .show   = pagetypeinfo_show,
1634 };
1635
1636 static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1637 {
1638         int zid;
1639
1640         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1641                 struct zone *compare = &pgdat->node_zones[zid];
1642
1643                 if (populated_zone(compare))
1644                         return zone == compare;
1645         }
1646
1647         return false;
1648 }
1649
1650 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1651                                                         struct zone *zone)
1652 {
1653         int i;
1654         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1655         if (is_zone_first_populated(pgdat, zone)) {
1656                 seq_printf(m, "\n  per-node stats");
1657                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1658                         seq_printf(m, "\n      %-12s %lu", node_stat_name(i),
1659                                    node_page_state_pages(pgdat, i));
1660                 }
1661         }
1662         seq_printf(m,
1663                    "\n  pages free     %lu"
1664                    "\n        min      %lu"
1665                    "\n        low      %lu"
1666                    "\n        high     %lu"
1667                    "\n        spanned  %lu"
1668                    "\n        present  %lu"
1669                    "\n        managed  %lu",
1670                    zone_page_state(zone, NR_FREE_PAGES),
1671                    min_wmark_pages(zone),
1672                    low_wmark_pages(zone),
1673                    high_wmark_pages(zone),
1674                    zone->spanned_pages,
1675                    zone->present_pages,
1676                    zone_managed_pages(zone));
1677
1678         seq_printf(m,
1679                    "\n        protection: (%ld",
1680                    zone->lowmem_reserve[0]);
1681         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1682                 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1683         seq_putc(m, ')');
1684
1685         /* If unpopulated, no other information is useful */
1686         if (!populated_zone(zone)) {
1687                 seq_putc(m, '\n');
1688                 return;
1689         }
1690
1691         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1692                 seq_printf(m, "\n      %-12s %lu", zone_stat_name(i),
1693                            zone_page_state(zone, i));
1694
1695 #ifdef CONFIG_NUMA
1696         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1697                 seq_printf(m, "\n      %-12s %lu", numa_stat_name(i),
1698                            zone_numa_state_snapshot(zone, i));
1699 #endif
1700
1701         seq_printf(m, "\n  pagesets");
1702         for_each_online_cpu(i) {
1703                 struct per_cpu_pageset *pageset;
1704
1705                 pageset = per_cpu_ptr(zone->pageset, i);
1706                 seq_printf(m,
1707                            "\n    cpu: %i"
1708                            "\n              count: %i"
1709                            "\n              high:  %i"
1710                            "\n              batch: %i",
1711                            i,
1712                            pageset->pcp.count,
1713                            pageset->pcp.high,
1714                            pageset->pcp.batch);
1715 #ifdef CONFIG_SMP
1716                 seq_printf(m, "\n  vm stats threshold: %d",
1717                                 pageset->stat_threshold);
1718 #endif
1719         }
1720         seq_printf(m,
1721                    "\n  node_unreclaimable:  %u"
1722                    "\n  start_pfn:           %lu",
1723                    pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1724                    zone->zone_start_pfn);
1725         seq_putc(m, '\n');
1726 }
1727
1728 /*
1729  * Output information about zones in @pgdat.  All zones are printed regardless
1730  * of whether they are populated or not: lowmem_reserve_ratio operates on the
1731  * set of all zones and userspace would not be aware of such zones if they are
1732  * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1733  */
1734 static int zoneinfo_show(struct seq_file *m, void *arg)
1735 {
1736         pg_data_t *pgdat = (pg_data_t *)arg;
1737         walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1738         return 0;
1739 }
1740
1741 static const struct seq_operations zoneinfo_op = {
1742         .start  = frag_start, /* iterate over all zones. The same as in
1743                                * fragmentation. */
1744         .next   = frag_next,
1745         .stop   = frag_stop,
1746         .show   = zoneinfo_show,
1747 };
1748
1749 #define NR_VMSTAT_ITEMS (NR_VM_ZONE_STAT_ITEMS + \
1750                          NR_VM_NUMA_STAT_ITEMS + \
1751                          NR_VM_NODE_STAT_ITEMS + \
1752                          NR_VM_WRITEBACK_STAT_ITEMS + \
1753                          (IS_ENABLED(CONFIG_VM_EVENT_COUNTERS) ? \
1754                           NR_VM_EVENT_ITEMS : 0))
1755
1756 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1757 {
1758         unsigned long *v;
1759         int i;
1760
1761         if (*pos >= NR_VMSTAT_ITEMS)
1762                 return NULL;
1763
1764         BUILD_BUG_ON(ARRAY_SIZE(vmstat_text) < NR_VMSTAT_ITEMS);
1765         v = kmalloc_array(NR_VMSTAT_ITEMS, sizeof(unsigned long), GFP_KERNEL);
1766         m->private = v;
1767         if (!v)
1768                 return ERR_PTR(-ENOMEM);
1769         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1770                 v[i] = global_zone_page_state(i);
1771         v += NR_VM_ZONE_STAT_ITEMS;
1772
1773 #ifdef CONFIG_NUMA
1774         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1775                 v[i] = global_numa_state(i);
1776         v += NR_VM_NUMA_STAT_ITEMS;
1777 #endif
1778
1779         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
1780                 v[i] = global_node_page_state_pages(i);
1781         v += NR_VM_NODE_STAT_ITEMS;
1782
1783         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1784                             v + NR_DIRTY_THRESHOLD);
1785         v += NR_VM_WRITEBACK_STAT_ITEMS;
1786
1787 #ifdef CONFIG_VM_EVENT_COUNTERS
1788         all_vm_events(v);
1789         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1790         v[PGPGOUT] /= 2;
1791 #endif
1792         return (unsigned long *)m->private + *pos;
1793 }
1794
1795 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1796 {
1797         (*pos)++;
1798         if (*pos >= NR_VMSTAT_ITEMS)
1799                 return NULL;
1800         return (unsigned long *)m->private + *pos;
1801 }
1802
1803 static int vmstat_show(struct seq_file *m, void *arg)
1804 {
1805         unsigned long *l = arg;
1806         unsigned long off = l - (unsigned long *)m->private;
1807
1808         seq_puts(m, vmstat_text[off]);
1809         seq_put_decimal_ull(m, " ", *l);
1810         seq_putc(m, '\n');
1811
1812         if (off == NR_VMSTAT_ITEMS - 1) {
1813                 /*
1814                  * We've come to the end - add any deprecated counters to avoid
1815                  * breaking userspace which might depend on them being present.
1816                  */
1817                 seq_puts(m, "nr_unstable 0\n");
1818         }
1819         return 0;
1820 }
1821
1822 static void vmstat_stop(struct seq_file *m, void *arg)
1823 {
1824         kfree(m->private);
1825         m->private = NULL;
1826 }
1827
1828 static const struct seq_operations vmstat_op = {
1829         .start  = vmstat_start,
1830         .next   = vmstat_next,
1831         .stop   = vmstat_stop,
1832         .show   = vmstat_show,
1833 };
1834 #endif /* CONFIG_PROC_FS */
1835
1836 #ifdef CONFIG_SMP
1837 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1838 int sysctl_stat_interval __read_mostly = HZ;
1839
1840 #ifdef CONFIG_PROC_FS
1841 static void refresh_vm_stats(struct work_struct *work)
1842 {
1843         refresh_cpu_vm_stats(true);
1844 }
1845
1846 int vmstat_refresh(struct ctl_table *table, int write,
1847                    void *buffer, size_t *lenp, loff_t *ppos)
1848 {
1849         long val;
1850         int err;
1851         int i;
1852
1853         /*
1854          * The regular update, every sysctl_stat_interval, may come later
1855          * than expected: leaving a significant amount in per_cpu buckets.
1856          * This is particularly misleading when checking a quantity of HUGE
1857          * pages, immediately after running a test.  /proc/sys/vm/stat_refresh,
1858          * which can equally be echo'ed to or cat'ted from (by root),
1859          * can be used to update the stats just before reading them.
1860          *
1861          * Oh, and since global_zone_page_state() etc. are so careful to hide
1862          * transiently negative values, report an error here if any of
1863          * the stats is negative, so we know to go looking for imbalance.
1864          */
1865         err = schedule_on_each_cpu(refresh_vm_stats);
1866         if (err)
1867                 return err;
1868         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1869                 val = atomic_long_read(&vm_zone_stat[i]);
1870                 if (val < 0) {
1871                         pr_warn("%s: %s %ld\n",
1872                                 __func__, zone_stat_name(i), val);
1873                         err = -EINVAL;
1874                 }
1875         }
1876 #ifdef CONFIG_NUMA
1877         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) {
1878                 val = atomic_long_read(&vm_numa_stat[i]);
1879                 if (val < 0) {
1880                         pr_warn("%s: %s %ld\n",
1881                                 __func__, numa_stat_name(i), val);
1882                         err = -EINVAL;
1883                 }
1884         }
1885 #endif
1886         if (err)
1887                 return err;
1888         if (write)
1889                 *ppos += *lenp;
1890         else
1891                 *lenp = 0;
1892         return 0;
1893 }
1894 #endif /* CONFIG_PROC_FS */
1895
1896 static void vmstat_update(struct work_struct *w)
1897 {
1898         if (refresh_cpu_vm_stats(true)) {
1899                 /*
1900                  * Counters were updated so we expect more updates
1901                  * to occur in the future. Keep on running the
1902                  * update worker thread.
1903                  */
1904                 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
1905                                 this_cpu_ptr(&vmstat_work),
1906                                 round_jiffies_relative(sysctl_stat_interval));
1907         }
1908 }
1909
1910 /*
1911  * Switch off vmstat processing and then fold all the remaining differentials
1912  * until the diffs stay at zero. The function is used by NOHZ and can only be
1913  * invoked when tick processing is not active.
1914  */
1915 /*
1916  * Check if the diffs for a certain cpu indicate that
1917  * an update is needed.
1918  */
1919 static bool need_update(int cpu)
1920 {
1921         struct zone *zone;
1922
1923         for_each_populated_zone(zone) {
1924                 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1925
1926                 BUILD_BUG_ON(sizeof(p->vm_stat_diff[0]) != 1);
1927 #ifdef CONFIG_NUMA
1928                 BUILD_BUG_ON(sizeof(p->vm_numa_stat_diff[0]) != 2);
1929 #endif
1930
1931                 /*
1932                  * The fast way of checking if there are any vmstat diffs.
1933                  */
1934                 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS *
1935                                sizeof(p->vm_stat_diff[0])))
1936                         return true;
1937 #ifdef CONFIG_NUMA
1938                 if (memchr_inv(p->vm_numa_stat_diff, 0, NR_VM_NUMA_STAT_ITEMS *
1939                                sizeof(p->vm_numa_stat_diff[0])))
1940                         return true;
1941 #endif
1942         }
1943         return false;
1944 }
1945
1946 /*
1947  * Switch off vmstat processing and then fold all the remaining differentials
1948  * until the diffs stay at zero. The function is used by NOHZ and can only be
1949  * invoked when tick processing is not active.
1950  */
1951 void quiet_vmstat(void)
1952 {
1953         if (system_state != SYSTEM_RUNNING)
1954                 return;
1955
1956         if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
1957                 return;
1958
1959         if (!need_update(smp_processor_id()))
1960                 return;
1961
1962         /*
1963          * Just refresh counters and do not care about the pending delayed
1964          * vmstat_update. It doesn't fire that often to matter and canceling
1965          * it would be too expensive from this path.
1966          * vmstat_shepherd will take care about that for us.
1967          */
1968         refresh_cpu_vm_stats(false);
1969 }
1970
1971 /*
1972  * Shepherd worker thread that checks the
1973  * differentials of processors that have their worker
1974  * threads for vm statistics updates disabled because of
1975  * inactivity.
1976  */
1977 static void vmstat_shepherd(struct work_struct *w);
1978
1979 static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
1980
1981 static void vmstat_shepherd(struct work_struct *w)
1982 {
1983         int cpu;
1984
1985         get_online_cpus();
1986         /* Check processors whose vmstat worker threads have been disabled */
1987         for_each_online_cpu(cpu) {
1988                 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
1989
1990                 if (!delayed_work_pending(dw) && need_update(cpu))
1991                         queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
1992         }
1993         put_online_cpus();
1994
1995         schedule_delayed_work(&shepherd,
1996                 round_jiffies_relative(sysctl_stat_interval));
1997 }
1998
1999 static void __init start_shepherd_timer(void)
2000 {
2001         int cpu;
2002
2003         for_each_possible_cpu(cpu)
2004                 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
2005                         vmstat_update);
2006
2007         schedule_delayed_work(&shepherd,
2008                 round_jiffies_relative(sysctl_stat_interval));
2009 }
2010
2011 static void __init init_cpu_node_state(void)
2012 {
2013         int node;
2014
2015         for_each_online_node(node) {
2016                 if (cpumask_weight(cpumask_of_node(node)) > 0)
2017                         node_set_state(node, N_CPU);
2018         }
2019 }
2020
2021 static int vmstat_cpu_online(unsigned int cpu)
2022 {
2023         refresh_zone_stat_thresholds();
2024         node_set_state(cpu_to_node(cpu), N_CPU);
2025         return 0;
2026 }
2027
2028 static int vmstat_cpu_down_prep(unsigned int cpu)
2029 {
2030         cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
2031         return 0;
2032 }
2033
2034 static int vmstat_cpu_dead(unsigned int cpu)
2035 {
2036         const struct cpumask *node_cpus;
2037         int node;
2038
2039         node = cpu_to_node(cpu);
2040
2041         refresh_zone_stat_thresholds();
2042         node_cpus = cpumask_of_node(node);
2043         if (cpumask_weight(node_cpus) > 0)
2044                 return 0;
2045
2046         node_clear_state(node, N_CPU);
2047         return 0;
2048 }
2049
2050 #endif
2051
2052 struct workqueue_struct *mm_percpu_wq;
2053
2054 void __init init_mm_internals(void)
2055 {
2056         int ret __maybe_unused;
2057
2058         mm_percpu_wq = alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM, 0);
2059
2060 #ifdef CONFIG_SMP
2061         ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
2062                                         NULL, vmstat_cpu_dead);
2063         if (ret < 0)
2064                 pr_err("vmstat: failed to register 'dead' hotplug state\n");
2065
2066         ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "mm/vmstat:online",
2067                                         vmstat_cpu_online,
2068                                         vmstat_cpu_down_prep);
2069         if (ret < 0)
2070                 pr_err("vmstat: failed to register 'online' hotplug state\n");
2071
2072         get_online_cpus();
2073         init_cpu_node_state();
2074         put_online_cpus();
2075
2076         start_shepherd_timer();
2077 #endif
2078 #ifdef CONFIG_PROC_FS
2079         proc_create_seq("buddyinfo", 0444, NULL, &fragmentation_op);
2080         proc_create_seq("pagetypeinfo", 0400, NULL, &pagetypeinfo_op);
2081         proc_create_seq("vmstat", 0444, NULL, &vmstat_op);
2082         proc_create_seq("zoneinfo", 0444, NULL, &zoneinfo_op);
2083 #endif
2084 }
2085
2086 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
2087
2088 /*
2089  * Return an index indicating how much of the available free memory is
2090  * unusable for an allocation of the requested size.
2091  */
2092 static int unusable_free_index(unsigned int order,
2093                                 struct contig_page_info *info)
2094 {
2095         /* No free memory is interpreted as all free memory is unusable */
2096         if (info->free_pages == 0)
2097                 return 1000;
2098
2099         /*
2100          * Index should be a value between 0 and 1. Return a value to 3
2101          * decimal places.
2102          *
2103          * 0 => no fragmentation
2104          * 1 => high fragmentation
2105          */
2106         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
2107
2108 }
2109
2110 static void unusable_show_print(struct seq_file *m,
2111                                         pg_data_t *pgdat, struct zone *zone)
2112 {
2113         unsigned int order;
2114         int index;
2115         struct contig_page_info info;
2116
2117         seq_printf(m, "Node %d, zone %8s ",
2118                                 pgdat->node_id,
2119                                 zone->name);
2120         for (order = 0; order < MAX_ORDER; ++order) {
2121                 fill_contig_page_info(zone, order, &info);
2122                 index = unusable_free_index(order, &info);
2123                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2124         }
2125
2126         seq_putc(m, '\n');
2127 }
2128
2129 /*
2130  * Display unusable free space index
2131  *
2132  * The unusable free space index measures how much of the available free
2133  * memory cannot be used to satisfy an allocation of a given size and is a
2134  * value between 0 and 1. The higher the value, the more of free memory is
2135  * unusable and by implication, the worse the external fragmentation is. This
2136  * can be expressed as a percentage by multiplying by 100.
2137  */
2138 static int unusable_show(struct seq_file *m, void *arg)
2139 {
2140         pg_data_t *pgdat = (pg_data_t *)arg;
2141
2142         /* check memoryless node */
2143         if (!node_state(pgdat->node_id, N_MEMORY))
2144                 return 0;
2145
2146         walk_zones_in_node(m, pgdat, true, false, unusable_show_print);
2147
2148         return 0;
2149 }
2150
2151 static const struct seq_operations unusable_sops = {
2152         .start  = frag_start,
2153         .next   = frag_next,
2154         .stop   = frag_stop,
2155         .show   = unusable_show,
2156 };
2157
2158 DEFINE_SEQ_ATTRIBUTE(unusable);
2159
2160 static void extfrag_show_print(struct seq_file *m,
2161                                         pg_data_t *pgdat, struct zone *zone)
2162 {
2163         unsigned int order;
2164         int index;
2165
2166         /* Alloc on stack as interrupts are disabled for zone walk */
2167         struct contig_page_info info;
2168
2169         seq_printf(m, "Node %d, zone %8s ",
2170                                 pgdat->node_id,
2171                                 zone->name);
2172         for (order = 0; order < MAX_ORDER; ++order) {
2173                 fill_contig_page_info(zone, order, &info);
2174                 index = __fragmentation_index(order, &info);
2175                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2176         }
2177
2178         seq_putc(m, '\n');
2179 }
2180
2181 /*
2182  * Display fragmentation index for orders that allocations would fail for
2183  */
2184 static int extfrag_show(struct seq_file *m, void *arg)
2185 {
2186         pg_data_t *pgdat = (pg_data_t *)arg;
2187
2188         walk_zones_in_node(m, pgdat, true, false, extfrag_show_print);
2189
2190         return 0;
2191 }
2192
2193 static const struct seq_operations extfrag_sops = {
2194         .start  = frag_start,
2195         .next   = frag_next,
2196         .stop   = frag_stop,
2197         .show   = extfrag_show,
2198 };
2199
2200 DEFINE_SEQ_ATTRIBUTE(extfrag);
2201
2202 static int __init extfrag_debug_init(void)
2203 {
2204         struct dentry *extfrag_debug_root;
2205
2206         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
2207
2208         debugfs_create_file("unusable_index", 0444, extfrag_debug_root, NULL,
2209                             &unusable_fops);
2210
2211         debugfs_create_file("extfrag_index", 0444, extfrag_debug_root, NULL,
2212                             &extfrag_fops);
2213
2214         return 0;
2215 }
2216
2217 module_init(extfrag_debug_init);
2218 #endif