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         x = delta + __this_cpu_read(*p);
 325
 326         t = __this_cpu_read(pcp->stat_threshold);
 327
 328         if (unlikely(x > t || x < -t)) {
 329                 zone_page_state_add(x, zone, item);
 330                 x = 0;
 331         }
 332         __this_cpu_write(*p, x);
 333 }
 334 EXPORT_SYMBOL(__mod_zone_page_state);
 335
 336 void __mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
 337                                 long delta)
 338 {
 339         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 340         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 341         long x;
 342         long t;
 343
 344         x = delta + __this_cpu_read(*p);
 345
 346         t = __this_cpu_read(pcp->stat_threshold);
 347
 348         if (unlikely(x > t || x < -t)) {
 349                 node_page_state_add(x, pgdat, item);
 350                 x = 0;
 351         }
 352         __this_cpu_write(*p, x);
 353 }
 354 EXPORT_SYMBOL(__mod_node_page_state);
 355
 356 /*
 357  * Optimized increment and decrement functions.
 358  *
 359  * These are only for a single page and therefore can take a struct page *
 360  * argument instead of struct zone *. This allows the inclusion of the code
 361  * generated for page_zone(page) into the optimized functions.
 362  *
 363  * No overflow check is necessary and therefore the differential can be
 364  * incremented or decremented in place which may allow the compilers to
 365  * generate better code.
 366  * The increment or decrement is known and therefore one boundary check can
 367  * be omitted.
 368  *
 369  * NOTE: These functions are very performance sensitive. Change only
 370  * with care.
 371  *
 372  * Some processors have inc/dec instructions that are atomic vs an interrupt.
 373  * However, the code must first determine the differential location in a zone
 374  * based on the processor number and then inc/dec the counter. There is no
 375  * guarantee without disabling preemption that the processor will not change
 376  * in between and therefore the atomicity vs. interrupt cannot be exploited
 377  * in a useful way here.
 378  */
 379 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
 380 {
 381         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 382         s8 __percpu *p = pcp->vm_stat_diff + item;
 383         s8 v, t;
 384
 385         v = __this_cpu_inc_return(*p);
 386         t = __this_cpu_read(pcp->stat_threshold);
 387         if (unlikely(v > t)) {
 388                 s8 overstep = t >> 1;
 389
 390                 zone_page_state_add(v + overstep, zone, item);
 391                 __this_cpu_write(*p, -overstep);
 392         }
 393 }
 394
 395 void __inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 396 {
 397         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 398         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 399         s8 v, t;
 400
 401         v = __this_cpu_inc_return(*p);
 402         t = __this_cpu_read(pcp->stat_threshold);
 403         if (unlikely(v > t)) {
 404                 s8 overstep = t >> 1;
 405
 406                 node_page_state_add(v + overstep, pgdat, item);
 407                 __this_cpu_write(*p, -overstep);
 408         }
 409 }
 410
 411 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
 412 {
 413         __inc_zone_state(page_zone(page), item);
 414 }
 415 EXPORT_SYMBOL(__inc_zone_page_state);
 416
 417 void __inc_node_page_state(struct page *page, enum node_stat_item item)
 418 {
 419         __inc_node_state(page_pgdat(page), item);
 420 }
 421 EXPORT_SYMBOL(__inc_node_page_state);
 422
 423 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
 424 {
 425         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 426         s8 __percpu *p = pcp->vm_stat_diff + item;
 427         s8 v, t;
 428
 429         v = __this_cpu_dec_return(*p);
 430         t = __this_cpu_read(pcp->stat_threshold);
 431         if (unlikely(v < - t)) {
 432                 s8 overstep = t >> 1;
 433
 434                 zone_page_state_add(v - overstep, zone, item);
 435                 __this_cpu_write(*p, overstep);
 436         }
 437 }
 438
 439 void __dec_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 440 {
 441         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 442         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 443         s8 v, t;
 444
 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                 node_page_state_add(v - overstep, pgdat, item);
 451                 __this_cpu_write(*p, overstep);
 452         }
 453 }
 454
 455 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
 456 {
 457         __dec_zone_state(page_zone(page), item);
 458 }
 459 EXPORT_SYMBOL(__dec_zone_page_state);
 460
 461 void __dec_node_page_state(struct page *page, enum node_stat_item item)
 462 {
 463         __dec_node_state(page_pgdat(page), item);
 464 }
 465 EXPORT_SYMBOL(__dec_node_page_state);
 466
 467 #ifdef CONFIG_HAVE_CMPXCHG_LOCAL
 468 /*
 469  * If we have cmpxchg_local support then we do not need to incur the overhead
 470  * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
 471  *
 472  * mod_state() modifies the zone counter state through atomic per cpu
 473  * operations.
 474  *
 475  * Overstep mode specifies how overstep should handled:
 476  *     0       No overstepping
 477  *     1       Overstepping half of threshold
 478  *     -1      Overstepping minus half of threshold
 479 */
 480 static inline void mod_zone_state(struct zone *zone,
 481        enum zone_stat_item item, long delta, int overstep_mode)
 482 {
 483         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 484         s8 __percpu *p = pcp->vm_stat_diff + item;
 485         long o, n, t, z;
 486
 487         do {
 488                 z = 0;  /* overflow to zone counters */
 489
 490                 /*
 491                  * The fetching of the stat_threshold is racy. We may apply
 492                  * a counter threshold to the wrong the cpu if we get
 493                  * rescheduled while executing here. However, the next
 494                  * counter update will apply the threshold again and
 495                  * therefore bring the counter under the threshold again.
 496                  *
 497                  * Most of the time the thresholds are the same anyways
 498                  * for all cpus in a zone.
 499                  */
 500                 t = this_cpu_read(pcp->stat_threshold);
 501
 502                 o = this_cpu_read(*p);
 503                 n = delta + o;
 504
 505                 if (n > t || n < -t) {
 506                         int os = overstep_mode * (t >> 1) ;
 507
 508                         /* Overflow must be added to zone counters */
 509                         z = n + os;
 510                         n = -os;
 511                 }
 512         } while (this_cpu_cmpxchg(*p, o, n) != o);
 513
 514         if (z)
 515                 zone_page_state_add(z, zone, item);
 516 }
 517
 518 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 519                          long delta)
 520 {
 521         mod_zone_state(zone, item, delta, 0);
 522 }
 523 EXPORT_SYMBOL(mod_zone_page_state);
 524
 525 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 526 {
 527         mod_zone_state(page_zone(page), item, 1, 1);
 528 }
 529 EXPORT_SYMBOL(inc_zone_page_state);
 530
 531 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 532 {
 533         mod_zone_state(page_zone(page), item, -1, -1);
 534 }
 535 EXPORT_SYMBOL(dec_zone_page_state);
 536
 537 static inline void mod_node_state(struct pglist_data *pgdat,
 538        enum node_stat_item item, int delta, int overstep_mode)
 539 {
 540         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 541         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 542         long o, n, t, z;
 543
 544         do {
 545                 z = 0;  /* overflow to node counters */
 546
 547                 /*
 548                  * The fetching of the stat_threshold is racy. We may apply
 549                  * a counter threshold to the wrong the cpu if we get
 550                  * rescheduled while executing here. However, the next
 551                  * counter update will apply the threshold again and
 552                  * therefore bring the counter under the threshold again.
 553                  *
 554                  * Most of the time the thresholds are the same anyways
 555                  * for all cpus in a node.
 556                  */
 557                 t = this_cpu_read(pcp->stat_threshold);
 558
 559                 o = this_cpu_read(*p);
 560                 n = delta + o;
 561
 562                 if (n > t || n < -t) {
 563                         int os = overstep_mode * (t >> 1) ;
 564
 565                         /* Overflow must be added to node counters */
 566                         z = n + os;
 567                         n = -os;
 568                 }
 569         } while (this_cpu_cmpxchg(*p, o, n) != o);
 570
 571         if (z)
 572                 node_page_state_add(z, pgdat, item);
 573 }
 574
 575 void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
 576                                         long delta)
 577 {
 578         mod_node_state(pgdat, item, delta, 0);
 579 }
 580 EXPORT_SYMBOL(mod_node_page_state);
 581
 582 void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 583 {
 584         mod_node_state(pgdat, item, 1, 1);
 585 }
 586
 587 void inc_node_page_state(struct page *page, enum node_stat_item item)
 588 {
 589         mod_node_state(page_pgdat(page), item, 1, 1);
 590 }
 591 EXPORT_SYMBOL(inc_node_page_state);
 592
 593 void dec_node_page_state(struct page *page, enum node_stat_item item)
 594 {
 595         mod_node_state(page_pgdat(page), item, -1, -1);
 596 }
 597 EXPORT_SYMBOL(dec_node_page_state);
 598 #else
 599 /*
 600  * Use interrupt disable to serialize counter updates
 601  */
 602 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 603                          long delta)
 604 {
 605         unsigned long flags;
 606
 607         local_irq_save(flags);
 608         __mod_zone_page_state(zone, item, delta);
 609         local_irq_restore(flags);
 610 }
 611 EXPORT_SYMBOL(mod_zone_page_state);
 612
 613 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 614 {
 615         unsigned long flags;
 616         struct zone *zone;
 617
 618         zone = page_zone(page);
 619         local_irq_save(flags);
 620         __inc_zone_state(zone, item);
 621         local_irq_restore(flags);
 622 }
 623 EXPORT_SYMBOL(inc_zone_page_state);
 624
 625 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 626 {
 627         unsigned long flags;
 628
 629         local_irq_save(flags);
 630         __dec_zone_page_state(page, item);
 631         local_irq_restore(flags);
 632 }
 633 EXPORT_SYMBOL(dec_zone_page_state);
 634
 635 void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 636 {
 637         unsigned long flags;
 638
 639         local_irq_save(flags);
 640         __inc_node_state(pgdat, item);
 641         local_irq_restore(flags);
 642 }
 643 EXPORT_SYMBOL(inc_node_state);
 644
 645 void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
 646                                         long delta)
 647 {
 648         unsigned long flags;
 649
 650         local_irq_save(flags);
 651         __mod_node_page_state(pgdat, item, delta);
 652         local_irq_restore(flags);
 653 }
 654 EXPORT_SYMBOL(mod_node_page_state);
 655
 656 void inc_node_page_state(struct page *page, enum node_stat_item item)
 657 {
 658         unsigned long flags;
 659         struct pglist_data *pgdat;
 660
 661         pgdat = page_pgdat(page);
 662         local_irq_save(flags);
 663         __inc_node_state(pgdat, item);
 664         local_irq_restore(flags);
 665 }
 666 EXPORT_SYMBOL(inc_node_page_state);
 667
 668 void dec_node_page_state(struct page *page, enum node_stat_item item)
 669 {
 670         unsigned long flags;
 671
 672         local_irq_save(flags);
 673         __dec_node_page_state(page, item);
 674         local_irq_restore(flags);
 675 }
 676 EXPORT_SYMBOL(dec_node_page_state);
 677 #endif
 678
 679 /*
 680  * Fold a differential into the global counters.
 681  * Returns the number of counters updated.
 682  */
 683 #ifdef CONFIG_NUMA
 684 static int fold_diff(int *zone_diff, int *numa_diff, int *node_diff)
 685 {
 686         int i;
 687         int changes = 0;
 688
 689         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 690                 if (zone_diff[i]) {
 691                         atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
 692                         changes++;
 693         }
 694
 695         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
 696                 if (numa_diff[i]) {
 697                         atomic_long_add(numa_diff[i], &vm_numa_stat[i]);
 698                         changes++;
 699         }
 700
 701         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 702                 if (node_diff[i]) {
 703                         atomic_long_add(node_diff[i], &vm_node_stat[i]);
 704                         changes++;
 705         }
 706         return changes;
 707 }
 708 #else
 709 static int fold_diff(int *zone_diff, int *node_diff)
 710 {
 711         int i;
 712         int changes = 0;
 713
 714         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 715                 if (zone_diff[i]) {
 716                         atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
 717                         changes++;
 718         }
 719
 720         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 721                 if (node_diff[i]) {
 722                         atomic_long_add(node_diff[i], &vm_node_stat[i]);
 723                         changes++;
 724         }
 725         return changes;
 726 }
 727 #endif /* CONFIG_NUMA */
 728
 729 /*
 730  * Update the zone counters for the current cpu.
 731  *
 732  * Note that refresh_cpu_vm_stats strives to only access
 733  * node local memory. The per cpu pagesets on remote zones are placed
 734  * in the memory local to the processor using that pageset. So the
 735  * loop over all zones will access a series of cachelines local to
 736  * the processor.
 737  *
 738  * The call to zone_page_state_add updates the cachelines with the
 739  * statistics in the remote zone struct as well as the global cachelines
 740  * with the global counters. These could cause remote node cache line
 741  * bouncing and will have to be only done when necessary.
 742  *
 743  * The function returns the number of global counters updated.
 744  */
 745 static int refresh_cpu_vm_stats(bool do_pagesets)
 746 {
 747         struct pglist_data *pgdat;
 748         struct zone *zone;
 749         int i;
 750         int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 751 #ifdef CONFIG_NUMA
 752         int global_numa_diff[NR_VM_NUMA_STAT_ITEMS] = { 0, };
 753 #endif
 754         int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
 755         int changes = 0;
 756
 757         for_each_populated_zone(zone) {
 758                 struct per_cpu_pageset __percpu *p = zone->pageset;
 759
 760                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
 761                         int v;
 762
 763                         v = this_cpu_xchg(p->vm_stat_diff[i], 0);
 764                         if (v) {
 765
 766                                 atomic_long_add(v, &zone->vm_stat[i]);
 767                                 global_zone_diff[i] += v;
 768 #ifdef CONFIG_NUMA
 769                                 /* 3 seconds idle till flush */
 770                                 __this_cpu_write(p->expire, 3);
 771 #endif
 772                         }
 773                 }
 774 #ifdef CONFIG_NUMA
 775                 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) {
 776                         int v;
 777
 778                         v = this_cpu_xchg(p->vm_numa_stat_diff[i], 0);
 779                         if (v) {
 780
 781                                 atomic_long_add(v, &zone->vm_numa_stat[i]);
 782                                 global_numa_diff[i] += v;
 783                                 __this_cpu_write(p->expire, 3);
 784                         }
 785                 }
 786
 787                 if (do_pagesets) {
 788                         cond_resched();
 789                         /*
 790                          * Deal with draining the remote pageset of this
 791                          * processor
 792                          *
 793                          * Check if there are pages remaining in this pageset
 794                          * if not then there is nothing to expire.
 795                          */
 796                         if (!__this_cpu_read(p->expire) ||
 797                                !__this_cpu_read(p->pcp.count))
 798                                 continue;
 799
 800                         /*
 801                          * We never drain zones local to this processor.
 802                          */
 803                         if (zone_to_nid(zone) == numa_node_id()) {
 804                                 __this_cpu_write(p->expire, 0);
 805                                 continue;
 806                         }
 807
 808                         if (__this_cpu_dec_return(p->expire))
 809                                 continue;
 810
 811                         if (__this_cpu_read(p->pcp.count)) {
 812                                 drain_zone_pages(zone, this_cpu_ptr(&p->pcp));
 813                                 changes++;
 814                         }
 815                 }
 816 #endif
 817         }
 818
 819         for_each_online_pgdat(pgdat) {
 820                 struct per_cpu_nodestat __percpu *p = pgdat->per_cpu_nodestats;
 821
 822                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
 823                         int v;
 824
 825                         v = this_cpu_xchg(p->vm_node_stat_diff[i], 0);
 826                         if (v) {
 827                                 atomic_long_add(v, &pgdat->vm_stat[i]);
 828                                 global_node_diff[i] += v;
 829                         }
 830                 }
 831         }
 832
 833 #ifdef CONFIG_NUMA
 834         changes += fold_diff(global_zone_diff, global_numa_diff,
 835                              global_node_diff);
 836 #else
 837         changes += fold_diff(global_zone_diff, global_node_diff);
 838 #endif
 839         return changes;
 840 }
 841
 842 /*
 843  * Fold the data for an offline cpu into the global array.
 844  * There cannot be any access by the offline cpu and therefore
 845  * synchronization is simplified.
 846  */
 847 void cpu_vm_stats_fold(int cpu)
 848 {
 849         struct pglist_data *pgdat;
 850         struct zone *zone;
 851         int i;
 852         int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 853 #ifdef CONFIG_NUMA
 854         int global_numa_diff[NR_VM_NUMA_STAT_ITEMS] = { 0, };
 855 #endif
 856         int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
 857
 858         for_each_populated_zone(zone) {
 859                 struct per_cpu_pageset *p;
 860
 861                 p = per_cpu_ptr(zone->pageset, cpu);
 862
 863                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 864                         if (p->vm_stat_diff[i]) {
 865                                 int v;
 866
 867                                 v = p->vm_stat_diff[i];
 868                                 p->vm_stat_diff[i] = 0;
 869                                 atomic_long_add(v, &zone->vm_stat[i]);
 870                                 global_zone_diff[i] += v;
 871                         }
 872
 873 #ifdef CONFIG_NUMA
 874                 for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
 875                         if (p->vm_numa_stat_diff[i]) {
 876                                 int v;
 877
 878                                 v = p->vm_numa_stat_diff[i];
 879                                 p->vm_numa_stat_diff[i] = 0;
 880                                 atomic_long_add(v, &zone->vm_numa_stat[i]);
 881                                 global_numa_diff[i] += v;
 882                         }
 883 #endif
 884         }
 885
 886         for_each_online_pgdat(pgdat) {
 887                 struct per_cpu_nodestat *p;
 888
 889                 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
 890
 891                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 892                         if (p->vm_node_stat_diff[i]) {
 893                                 int v;
 894
 895                                 v = p->vm_node_stat_diff[i];
 896                                 p->vm_node_stat_diff[i] = 0;
 897                                 atomic_long_add(v, &pgdat->vm_stat[i]);
 898                                 global_node_diff[i] += v;
 899                         }
 900         }
 901
 902 #ifdef CONFIG_NUMA
 903         fold_diff(global_zone_diff, global_numa_diff, global_node_diff);
 904 #else
 905         fold_diff(global_zone_diff, global_node_diff);
 906 #endif
 907 }
 908
 909 /*
 910  * this is only called if !populated_zone(zone), which implies no other users of
 911  * pset->vm_stat_diff[] exsist.
 912  */
 913 void drain_zonestat(struct zone *zone, struct per_cpu_pageset *pset)
 914 {
 915         int i;
 916
 917         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 918                 if (pset->vm_stat_diff[i]) {
 919                         int v = pset->vm_stat_diff[i];
 920                         pset->vm_stat_diff[i] = 0;
 921                         atomic_long_add(v, &zone->vm_stat[i]);
 922                         atomic_long_add(v, &vm_zone_stat[i]);
 923                 }
 924
 925 #ifdef CONFIG_NUMA
 926         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
 927                 if (pset->vm_numa_stat_diff[i]) {
 928                         int v = pset->vm_numa_stat_diff[i];
 929
 930                         pset->vm_numa_stat_diff[i] = 0;
 931                         atomic_long_add(v, &zone->vm_numa_stat[i]);
 932                         atomic_long_add(v, &vm_numa_stat[i]);
 933                 }
 934 #endif
 935 }
 936 #endif
 937
 938 #ifdef CONFIG_NUMA
 939 void __inc_numa_state(struct zone *zone,
 940                                  enum numa_stat_item item)
 941 {
 942         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 943         u16 __percpu *p = pcp->vm_numa_stat_diff + item;
 944         u16 v;
 945
 946         v = __this_cpu_inc_return(*p);
 947
 948         if (unlikely(v > NUMA_STATS_THRESHOLD)) {
 949                 zone_numa_state_add(v, zone, item);
 950                 __this_cpu_write(*p, 0);
 951         }
 952 }
 953
 954 /*
 955  * Determine the per node value of a stat item. This function
 956  * is called frequently in a NUMA machine, so try to be as
 957  * frugal as possible.
 958  */
 959 unsigned long sum_zone_node_page_state(int node,
 960                                  enum zone_stat_item item)
 961 {
 962         struct zone *zones = NODE_DATA(node)->node_zones;
 963         int i;
 964         unsigned long count = 0;
 965
 966         for (i = 0; i < MAX_NR_ZONES; i++)
 967                 count += zone_page_state(zones + i, item);
 968
 969         return count;
 970 }
 971
 972 /*
 973  * Determine the per node value of a numa stat item. To avoid deviation,
 974  * the per cpu stat number in vm_numa_stat_diff[] is also included.
 975  */
 976 unsigned long sum_zone_numa_state(int node,
 977                                  enum numa_stat_item item)
 978 {
 979         struct zone *zones = NODE_DATA(node)->node_zones;
 980         int i;
 981         unsigned long count = 0;
 982
 983         for (i = 0; i < MAX_NR_ZONES; i++)
 984                 count += zone_numa_state_snapshot(zones + i, item);
 985
 986         return count;
 987 }
 988
 989 /*
 990  * Determine the per node value of a stat item.
 991  */
 992 unsigned long node_page_state(struct pglist_data *pgdat,
 993                                 enum node_stat_item item)
 994 {
 995         long x = atomic_long_read(&pgdat->vm_stat[item]);
 996 #ifdef CONFIG_SMP
 997         if (x < 0)
 998                 x = 0;
 999 #endif
1000         return x;
1001 }
1002 #endif
1003
1004 #ifdef CONFIG_COMPACTION
1005
1006 struct contig_page_info {
1007         unsigned long free_pages;
1008         unsigned long free_blocks_total;
1009         unsigned long free_blocks_suitable;
1010 };
1011
1012 /*
1013  * Calculate the number of free pages in a zone, how many contiguous
1014  * pages are free and how many are large enough to satisfy an allocation of
1015  * the target size. Note that this function makes no attempt to estimate
1016  * how many suitable free blocks there *might* be if MOVABLE pages were
1017  * migrated. Calculating that is possible, but expensive and can be
1018  * figured out from userspace
1019  */
1020 static void fill_contig_page_info(struct zone *zone,
1021                                 unsigned int suitable_order,
1022                                 struct contig_page_info *info)
1023 {
1024         unsigned int order;
1025
1026         info->free_pages = 0;
1027         info->free_blocks_total = 0;
1028         info->free_blocks_suitable = 0;
1029
1030         for (order = 0; order < MAX_ORDER; order++) {
1031                 unsigned long blocks;
1032
1033                 /* Count number of free blocks */
1034                 blocks = zone->free_area[order].nr_free;
1035                 info->free_blocks_total += blocks;
1036
1037                 /* Count free base pages */
1038                 info->free_pages += blocks << order;
1039
1040                 /* Count the suitable free blocks */
1041                 if (order >= suitable_order)
1042                         info->free_blocks_suitable += blocks <<
1043                                                 (order - suitable_order);
1044         }
1045 }
1046
1047 /*
1048  * A fragmentation index only makes sense if an allocation of a requested
1049  * size would fail. If that is true, the fragmentation index indicates
1050  * whether external fragmentation or a lack of memory was the problem.
1051  * The value can be used to determine if page reclaim or compaction
1052  * should be used
1053  */
1054 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
1055 {
1056         unsigned long requested = 1UL << order;
1057
1058         if (WARN_ON_ONCE(order >= MAX_ORDER))
1059                 return 0;
1060
1061         if (!info->free_blocks_total)
1062                 return 0;
1063
1064         /* Fragmentation index only makes sense when a request would fail */
1065         if (info->free_blocks_suitable)
1066                 return -1000;
1067
1068         /*
1069          * Index is between 0 and 1 so return within 3 decimal places
1070          *
1071          * 0 => allocation would fail due to lack of memory
1072          * 1 => allocation would fail due to fragmentation
1073          */
1074         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
1075 }
1076
1077 /* Same as __fragmentation index but allocs contig_page_info on stack */
1078 int fragmentation_index(struct zone *zone, unsigned int order)
1079 {
1080         struct contig_page_info info;
1081
1082         fill_contig_page_info(zone, order, &info);
1083         return __fragmentation_index(order, &info);
1084 }
1085 #endif
1086
1087 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || \
1088     defined(CONFIG_NUMA) || defined(CONFIG_MEMCG)
1089 #ifdef CONFIG_ZONE_DMA
1090 #define TEXT_FOR_DMA(xx) xx "_dma",
1091 #else
1092 #define TEXT_FOR_DMA(xx)
1093 #endif
1094
1095 #ifdef CONFIG_ZONE_DMA32
1096 #define TEXT_FOR_DMA32(xx) xx "_dma32",
1097 #else
1098 #define TEXT_FOR_DMA32(xx)
1099 #endif
1100
1101 #ifdef CONFIG_HIGHMEM
1102 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
1103 #else
1104 #define TEXT_FOR_HIGHMEM(xx)
1105 #endif
1106
1107 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1108                                         TEXT_FOR_HIGHMEM(xx) xx "_movable",
1109
1110 const char * const vmstat_text[] = {
1111         /* enum zone_stat_item counters */
1112         "nr_free_pages",
1113         "nr_zone_inactive_anon",
1114         "nr_zone_active_anon",
1115         "nr_zone_inactive_file",
1116         "nr_zone_active_file",
1117         "nr_zone_unevictable",
1118         "nr_zone_write_pending",
1119         "nr_mlock",
1120         "nr_page_table_pages",
1121         "nr_kernel_stack",
1122 #if IS_ENABLED(CONFIG_SHADOW_CALL_STACK)
1123         "nr_shadow_call_stack",
1124 #endif
1125         "nr_bounce",
1126 #if IS_ENABLED(CONFIG_ZSMALLOC)
1127         "nr_zspages",
1128 #endif
1129         "nr_free_cma",
1130
1131         /* enum numa_stat_item counters */
1132 #ifdef CONFIG_NUMA
1133         "numa_hit",
1134         "numa_miss",
1135         "numa_foreign",
1136         "numa_interleave",
1137         "numa_local",
1138         "numa_other",
1139 #endif
1140
1141         /* enum node_stat_item counters */
1142         "nr_inactive_anon",
1143         "nr_active_anon",
1144         "nr_inactive_file",
1145         "nr_active_file",
1146         "nr_unevictable",
1147         "nr_slab_reclaimable",
1148         "nr_slab_unreclaimable",
1149         "nr_isolated_anon",
1150         "nr_isolated_file",
1151         "workingset_nodes",
1152         "workingset_refault",
1153         "workingset_activate",
1154         "workingset_restore",
1155         "workingset_nodereclaim",
1156         "nr_anon_pages",
1157         "nr_mapped",
1158         "nr_file_pages",
1159         "nr_dirty",
1160         "nr_writeback",
1161         "nr_writeback_temp",
1162         "nr_shmem",
1163         "nr_shmem_hugepages",
1164         "nr_shmem_pmdmapped",
1165         "nr_file_hugepages",
1166         "nr_file_pmdmapped",
1167         "nr_anon_transparent_hugepages",
1168         "nr_vmscan_write",
1169         "nr_vmscan_immediate_reclaim",
1170         "nr_dirtied",
1171         "nr_written",
1172         "nr_kernel_misc_reclaimable",
1173         "nr_foll_pin_acquired",
1174         "nr_foll_pin_released",
1175
1176         /* enum writeback_stat_item counters */
1177         "nr_dirty_threshold",
1178         "nr_dirty_background_threshold",
1179
1180 #if defined(CONFIG_VM_EVENT_COUNTERS) || defined(CONFIG_MEMCG)
1181         /* enum vm_event_item counters */
1182         "pgpgin",
1183         "pgpgout",
1184         "pswpin",
1185         "pswpout",
1186
1187         TEXTS_FOR_ZONES("pgalloc")
1188         TEXTS_FOR_ZONES("allocstall")
1189         TEXTS_FOR_ZONES("pgskip")
1190
1191         "pgfree",
1192         "pgactivate",
1193         "pgdeactivate",
1194         "pglazyfree",
1195
1196         "pgfault",
1197         "pgmajfault",
1198         "pglazyfreed",
1199
1200         "pgrefill",
1201         "pgsteal_kswapd",
1202         "pgsteal_direct",
1203         "pgscan_kswapd",
1204         "pgscan_direct",
1205         "pgscan_direct_throttle",
1206         "pgscan_anon",
1207         "pgscan_file",
1208         "pgsteal_anon",
1209         "pgsteal_file",
1210
1211 #ifdef CONFIG_NUMA
1212         "zone_reclaim_failed",
1213 #endif
1214         "pginodesteal",
1215         "slabs_scanned",
1216         "kswapd_inodesteal",
1217         "kswapd_low_wmark_hit_quickly",
1218         "kswapd_high_wmark_hit_quickly",
1219         "pageoutrun",
1220
1221         "pgrotated",
1222
1223         "drop_pagecache",
1224         "drop_slab",
1225         "oom_kill",
1226
1227 #ifdef CONFIG_NUMA_BALANCING
1228         "numa_pte_updates",
1229         "numa_huge_pte_updates",
1230         "numa_hint_faults",
1231         "numa_hint_faults_local",
1232         "numa_pages_migrated",
1233 #endif
1234 #ifdef CONFIG_MIGRATION
1235         "pgmigrate_success",
1236         "pgmigrate_fail",
1237 #endif
1238 #ifdef CONFIG_COMPACTION
1239         "compact_migrate_scanned",
1240         "compact_free_scanned",
1241         "compact_isolated",
1242         "compact_stall",
1243         "compact_fail",
1244         "compact_success",
1245         "compact_daemon_wake",
1246         "compact_daemon_migrate_scanned",
1247         "compact_daemon_free_scanned",
1248 #endif
1249
1250 #ifdef CONFIG_HUGETLB_PAGE
1251         "htlb_buddy_alloc_success",
1252         "htlb_buddy_alloc_fail",
1253 #endif
1254         "unevictable_pgs_culled",
1255         "unevictable_pgs_scanned",
1256         "unevictable_pgs_rescued",
1257         "unevictable_pgs_mlocked",
1258         "unevictable_pgs_munlocked",
1259         "unevictable_pgs_cleared",
1260         "unevictable_pgs_stranded",
1261
1262 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1263         "thp_fault_alloc",
1264         "thp_fault_fallback",
1265         "thp_fault_fallback_charge",
1266         "thp_collapse_alloc",
1267         "thp_collapse_alloc_failed",
1268         "thp_file_alloc",
1269         "thp_file_fallback",
1270         "thp_file_fallback_charge",
1271         "thp_file_mapped",
1272         "thp_split_page",
1273         "thp_split_page_failed",
1274         "thp_deferred_split_page",
1275         "thp_split_pmd",
1276 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1277         "thp_split_pud",
1278 #endif
1279         "thp_zero_page_alloc",
1280         "thp_zero_page_alloc_failed",
1281         "thp_swpout",
1282         "thp_swpout_fallback",
1283 #endif
1284 #ifdef CONFIG_MEMORY_BALLOON
1285         "balloon_inflate",
1286         "balloon_deflate",
1287 #ifdef CONFIG_BALLOON_COMPACTION
1288         "balloon_migrate",
1289 #endif
1290 #endif /* CONFIG_MEMORY_BALLOON */
1291 #ifdef CONFIG_DEBUG_TLBFLUSH
1292         "nr_tlb_remote_flush",
1293         "nr_tlb_remote_flush_received",
1294         "nr_tlb_local_flush_all",
1295         "nr_tlb_local_flush_one",
1296 #endif /* CONFIG_DEBUG_TLBFLUSH */
1297
1298 #ifdef CONFIG_DEBUG_VM_VMACACHE
1299         "vmacache_find_calls",
1300         "vmacache_find_hits",
1301 #endif
1302 #ifdef CONFIG_SWAP
1303         "swap_ra",
1304         "swap_ra_hit",
1305 #endif
1306 #endif /* CONFIG_VM_EVENT_COUNTERS || CONFIG_MEMCG */
1307 };
1308 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA || CONFIG_MEMCG */
1309
1310 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1311      defined(CONFIG_PROC_FS)
1312 static void *frag_start(struct seq_file *m, loff_t *pos)
1313 {
1314         pg_data_t *pgdat;
1315         loff_t node = *pos;
1316
1317         for (pgdat = first_online_pgdat();
1318              pgdat && node;
1319              pgdat = next_online_pgdat(pgdat))
1320                 --node;
1321
1322         return pgdat;
1323 }
1324
1325 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1326 {
1327         pg_data_t *pgdat = (pg_data_t *)arg;
1328
1329         (*pos)++;
1330         return next_online_pgdat(pgdat);
1331 }
1332
1333 static void frag_stop(struct seq_file *m, void *arg)
1334 {
1335 }
1336
1337 /*
1338  * Walk zones in a node and print using a callback.
1339  * If @assert_populated is true, only use callback for zones that are populated.
1340  */
1341 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1342                 bool assert_populated, bool nolock,
1343                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1344 {
1345         struct zone *zone;
1346         struct zone *node_zones = pgdat->node_zones;
1347         unsigned long flags;
1348
1349         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1350                 if (assert_populated && !populated_zone(zone))
1351                         continue;
1352
1353                 if (!nolock)
1354                         spin_lock_irqsave(&zone->lock, flags);
1355                 print(m, pgdat, zone);
1356                 if (!nolock)
1357                         spin_unlock_irqrestore(&zone->lock, flags);
1358         }
1359 }
1360 #endif
1361
1362 #ifdef CONFIG_PROC_FS
1363 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1364                                                 struct zone *zone)
1365 {
1366         int order;
1367
1368         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1369         for (order = 0; order < MAX_ORDER; ++order)
1370                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
1371         seq_putc(m, '\n');
1372 }
1373
1374 /*
1375  * This walks the free areas for each zone.
1376  */
1377 static int frag_show(struct seq_file *m, void *arg)
1378 {
1379         pg_data_t *pgdat = (pg_data_t *)arg;
1380         walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1381         return 0;
1382 }
1383
1384 static void pagetypeinfo_showfree_print(struct seq_file *m,
1385                                         pg_data_t *pgdat, struct zone *zone)
1386 {
1387         int order, mtype;
1388
1389         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1390                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
1391                                         pgdat->node_id,
1392                                         zone->name,
1393                                         migratetype_names[mtype]);
1394                 for (order = 0; order < MAX_ORDER; ++order) {
1395                         unsigned long freecount = 0;
1396                         struct free_area *area;
1397                         struct list_head *curr;
1398                         bool overflow = false;
1399
1400                         area = &(zone->free_area[order]);
1401
1402                         list_for_each(curr, &area->free_list[mtype]) {
1403                                 /*
1404                                  * Cap the free_list iteration because it might
1405                                  * be really large and we are under a spinlock
1406                                  * so a long time spent here could trigger a
1407                                  * hard lockup detector. Anyway this is a
1408                                  * debugging tool so knowing there is a handful
1409                                  * of pages of this order should be more than
1410                                  * sufficient.
1411                                  */
1412                                 if (++freecount >= 100000) {
1413                                         overflow = true;
1414                                         break;
1415                                 }
1416                         }
1417                         seq_printf(m, "%s%6lu ", overflow ? ">" : "", freecount);
1418                         spin_unlock_irq(&zone->lock);
1419                         cond_resched();
1420                         spin_lock_irq(&zone->lock);
1421                 }
1422                 seq_putc(m, '\n');
1423         }
1424 }
1425
1426 /* Print out the free pages at each order for each migatetype */
1427 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
1428 {
1429         int order;
1430         pg_data_t *pgdat = (pg_data_t *)arg;
1431
1432         /* Print header */
1433         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1434         for (order = 0; order < MAX_ORDER; ++order)
1435                 seq_printf(m, "%6d ", order);
1436         seq_putc(m, '\n');
1437
1438         walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1439
1440         return 0;
1441 }
1442
1443 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1444                                         pg_data_t *pgdat, struct zone *zone)
1445 {
1446         int mtype;
1447         unsigned long pfn;
1448         unsigned long start_pfn = zone->zone_start_pfn;
1449         unsigned long end_pfn = zone_end_pfn(zone);
1450         unsigned long count[MIGRATE_TYPES] = { 0, };
1451
1452         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1453                 struct page *page;
1454
1455                 page = pfn_to_online_page(pfn);
1456                 if (!page)
1457                         continue;
1458
1459                 /* Watch for unexpected holes punched in the memmap */
1460                 if (!memmap_valid_within(pfn, page, zone))
1461                         continue;
1462
1463                 if (page_zone(page) != zone)
1464                         continue;
1465
1466                 mtype = get_pageblock_migratetype(page);
1467
1468                 if (mtype < MIGRATE_TYPES)
1469                         count[mtype]++;
1470         }
1471
1472         /* Print counts */
1473         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1474         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1475                 seq_printf(m, "%12lu ", count[mtype]);
1476         seq_putc(m, '\n');
1477 }
1478
1479 /* Print out the number of pageblocks for each migratetype */
1480 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1481 {
1482         int mtype;
1483         pg_data_t *pgdat = (pg_data_t *)arg;
1484
1485         seq_printf(m, "\n%-23s", "Number of blocks type ");
1486         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1487                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1488         seq_putc(m, '\n');
1489         walk_zones_in_node(m, pgdat, true, false,
1490                 pagetypeinfo_showblockcount_print);
1491
1492         return 0;
1493 }
1494
1495 /*
1496  * Print out the number of pageblocks for each migratetype that contain pages
1497  * of other types. This gives an indication of how well fallbacks are being
1498  * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1499  * to determine what is going on
1500  */
1501 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1502 {
1503 #ifdef CONFIG_PAGE_OWNER
1504         int mtype;
1505
1506         if (!static_branch_unlikely(&page_owner_inited))
1507                 return;
1508
1509         drain_all_pages(NULL);
1510
1511         seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1512         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1513                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1514         seq_putc(m, '\n');
1515
1516         walk_zones_in_node(m, pgdat, true, true,
1517                 pagetypeinfo_showmixedcount_print);
1518 #endif /* CONFIG_PAGE_OWNER */
1519 }
1520
1521 /*
1522  * This prints out statistics in relation to grouping pages by mobility.
1523  * It is expensive to collect so do not constantly read the file.
1524  */
1525 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1526 {
1527         pg_data_t *pgdat = (pg_data_t *)arg;
1528
1529         /* check memoryless node */
1530         if (!node_state(pgdat->node_id, N_MEMORY))
1531                 return 0;
1532
1533         seq_printf(m, "Page block order: %d\n", pageblock_order);
1534         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
1535         seq_putc(m, '\n');
1536         pagetypeinfo_showfree(m, pgdat);
1537         pagetypeinfo_showblockcount(m, pgdat);
1538         pagetypeinfo_showmixedcount(m, pgdat);
1539
1540         return 0;
1541 }
1542
1543 static const struct seq_operations fragmentation_op = {
1544         .start  = frag_start,
1545         .next   = frag_next,
1546         .stop   = frag_stop,
1547         .show   = frag_show,
1548 };
1549
1550 static const struct seq_operations pagetypeinfo_op = {
1551         .start  = frag_start,
1552         .next   = frag_next,
1553         .stop   = frag_stop,
1554         .show   = pagetypeinfo_show,
1555 };
1556
1557 static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1558 {
1559         int zid;
1560
1561         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1562                 struct zone *compare = &pgdat->node_zones[zid];
1563
1564                 if (populated_zone(compare))
1565                         return zone == compare;
1566         }
1567
1568         return false;
1569 }
1570
1571 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1572                                                         struct zone *zone)
1573 {
1574         int i;
1575         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1576         if (is_zone_first_populated(pgdat, zone)) {
1577                 seq_printf(m, "\n  per-node stats");
1578                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1579                         seq_printf(m, "\n      %-12s %lu", node_stat_name(i),
1580                                    node_page_state(pgdat, i));
1581                 }
1582         }
1583         seq_printf(m,
1584                    "\n  pages free     %lu"
1585                    "\n        min      %lu"
1586                    "\n        low      %lu"
1587                    "\n        high     %lu"
1588                    "\n        spanned  %lu"
1589                    "\n        present  %lu"
1590                    "\n        managed  %lu",
1591                    zone_page_state(zone, NR_FREE_PAGES),
1592                    min_wmark_pages(zone),
1593                    low_wmark_pages(zone),
1594                    high_wmark_pages(zone),
1595                    zone->spanned_pages,
1596                    zone->present_pages,
1597                    zone_managed_pages(zone));
1598
1599         /* If unpopulated, no other information is useful */
1600         if (!populated_zone(zone)) {
1601                 seq_putc(m, '\n');
1602                 return;
1603         }
1604
1605         seq_printf(m,
1606                    "\n        protection: (%ld",
1607                    zone->lowmem_reserve[0]);
1608         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1609                 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1610         seq_putc(m, ')');
1611
1612         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1613                 seq_printf(m, "\n      %-12s %lu", zone_stat_name(i),
1614                            zone_page_state(zone, i));
1615
1616 #ifdef CONFIG_NUMA
1617         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1618                 seq_printf(m, "\n      %-12s %lu", numa_stat_name(i),
1619                            zone_numa_state_snapshot(zone, i));
1620 #endif
1621
1622         seq_printf(m, "\n  pagesets");
1623         for_each_online_cpu(i) {
1624                 struct per_cpu_pageset *pageset;
1625
1626                 pageset = per_cpu_ptr(zone->pageset, i);
1627                 seq_printf(m,
1628                            "\n    cpu: %i"
1629                            "\n              count: %i"
1630                            "\n              high:  %i"
1631                            "\n              batch: %i",
1632                            i,
1633                            pageset->pcp.count,
1634                            pageset->pcp.high,
1635                            pageset->pcp.batch);
1636 #ifdef CONFIG_SMP
1637                 seq_printf(m, "\n  vm stats threshold: %d",
1638                                 pageset->stat_threshold);
1639 #endif
1640         }
1641         seq_printf(m,
1642                    "\n  node_unreclaimable:  %u"
1643                    "\n  start_pfn:           %lu",
1644                    pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1645                    zone->zone_start_pfn);
1646         seq_putc(m, '\n');
1647 }
1648
1649 /*
1650  * Output information about zones in @pgdat.  All zones are printed regardless
1651  * of whether they are populated or not: lowmem_reserve_ratio operates on the
1652  * set of all zones and userspace would not be aware of such zones if they are
1653  * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1654  */
1655 static int zoneinfo_show(struct seq_file *m, void *arg)
1656 {
1657         pg_data_t *pgdat = (pg_data_t *)arg;
1658         walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1659         return 0;
1660 }
1661
1662 static const struct seq_operations zoneinfo_op = {
1663         .start  = frag_start, /* iterate over all zones. The same as in
1664                                * fragmentation. */
1665         .next   = frag_next,
1666         .stop   = frag_stop,
1667         .show   = zoneinfo_show,
1668 };
1669
1670 #define NR_VMSTAT_ITEMS (NR_VM_ZONE_STAT_ITEMS + \
1671                          NR_VM_NUMA_STAT_ITEMS + \
1672                          NR_VM_NODE_STAT_ITEMS + \
1673                          NR_VM_WRITEBACK_STAT_ITEMS + \
1674                          (IS_ENABLED(CONFIG_VM_EVENT_COUNTERS) ? \
1675                           NR_VM_EVENT_ITEMS : 0))
1676
1677 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1678 {
1679         unsigned long *v;
1680         int i;
1681
1682         if (*pos >= NR_VMSTAT_ITEMS)
1683                 return NULL;
1684
1685         BUILD_BUG_ON(ARRAY_SIZE(vmstat_text) < NR_VMSTAT_ITEMS);
1686         v = kmalloc_array(NR_VMSTAT_ITEMS, sizeof(unsigned long), GFP_KERNEL);
1687         m->private = v;
1688         if (!v)
1689                 return ERR_PTR(-ENOMEM);
1690         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1691                 v[i] = global_zone_page_state(i);
1692         v += NR_VM_ZONE_STAT_ITEMS;
1693
1694 #ifdef CONFIG_NUMA
1695         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1696                 v[i] = global_numa_state(i);
1697         v += NR_VM_NUMA_STAT_ITEMS;
1698 #endif
1699
1700         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
1701                 v[i] = global_node_page_state(i);
1702         v += NR_VM_NODE_STAT_ITEMS;
1703
1704         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1705                             v + NR_DIRTY_THRESHOLD);
1706         v += NR_VM_WRITEBACK_STAT_ITEMS;
1707
1708 #ifdef CONFIG_VM_EVENT_COUNTERS
1709         all_vm_events(v);
1710         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1711         v[PGPGOUT] /= 2;
1712 #endif
1713         return (unsigned long *)m->private + *pos;
1714 }
1715
1716 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1717 {
1718         (*pos)++;
1719         if (*pos >= NR_VMSTAT_ITEMS)
1720                 return NULL;
1721         return (unsigned long *)m->private + *pos;
1722 }
1723
1724 static int vmstat_show(struct seq_file *m, void *arg)
1725 {
1726         unsigned long *l = arg;
1727         unsigned long off = l - (unsigned long *)m->private;
1728
1729         seq_puts(m, vmstat_text[off]);
1730         seq_put_decimal_ull(m, " ", *l);
1731         seq_putc(m, '\n');
1732
1733         if (off == NR_VMSTAT_ITEMS - 1) {
1734                 /*
1735                  * We've come to the end - add any deprecated counters to avoid
1736                  * breaking userspace which might depend on them being present.
1737                  */
1738                 seq_puts(m, "nr_unstable 0\n");
1739         }
1740         return 0;
1741 }
1742
1743 static void vmstat_stop(struct seq_file *m, void *arg)
1744 {
1745         kfree(m->private);
1746         m->private = NULL;
1747 }
1748
1749 static const struct seq_operations vmstat_op = {
1750         .start  = vmstat_start,
1751         .next   = vmstat_next,
1752         .stop   = vmstat_stop,
1753         .show   = vmstat_show,
1754 };
1755 #endif /* CONFIG_PROC_FS */
1756
1757 #ifdef CONFIG_SMP
1758 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1759 int sysctl_stat_interval __read_mostly = HZ;
1760
1761 #ifdef CONFIG_PROC_FS
1762 static void refresh_vm_stats(struct work_struct *work)
1763 {
1764         refresh_cpu_vm_stats(true);
1765 }
1766
1767 int vmstat_refresh(struct ctl_table *table, int write,
1768                    void *buffer, size_t *lenp, loff_t *ppos)
1769 {
1770         long val;
1771         int err;
1772         int i;
1773
1774         /*
1775          * The regular update, every sysctl_stat_interval, may come later
1776          * than expected: leaving a significant amount in per_cpu buckets.
1777          * This is particularly misleading when checking a quantity of HUGE
1778          * pages, immediately after running a test.  /proc/sys/vm/stat_refresh,
1779          * which can equally be echo'ed to or cat'ted from (by root),
1780          * can be used to update the stats just before reading them.
1781          *
1782          * Oh, and since global_zone_page_state() etc. are so careful to hide
1783          * transiently negative values, report an error here if any of
1784          * the stats is negative, so we know to go looking for imbalance.
1785          */
1786         err = schedule_on_each_cpu(refresh_vm_stats);
1787         if (err)
1788                 return err;
1789         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1790                 val = atomic_long_read(&vm_zone_stat[i]);
1791                 if (val < 0) {
1792                         pr_warn("%s: %s %ld\n",
1793                                 __func__, zone_stat_name(i), val);
1794                         err = -EINVAL;
1795                 }
1796         }
1797 #ifdef CONFIG_NUMA
1798         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) {
1799                 val = atomic_long_read(&vm_numa_stat[i]);
1800                 if (val < 0) {
1801                         pr_warn("%s: %s %ld\n",
1802                                 __func__, numa_stat_name(i), val);
1803                         err = -EINVAL;
1804                 }
1805         }
1806 #endif
1807         if (err)
1808                 return err;
1809         if (write)
1810                 *ppos += *lenp;
1811         else
1812                 *lenp = 0;
1813         return 0;
1814 }
1815 #endif /* CONFIG_PROC_FS */
1816
1817 static void vmstat_update(struct work_struct *w)
1818 {
1819         if (refresh_cpu_vm_stats(true)) {
1820                 /*
1821                  * Counters were updated so we expect more updates
1822                  * to occur in the future. Keep on running the
1823                  * update worker thread.
1824                  */
1825                 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
1826                                 this_cpu_ptr(&vmstat_work),
1827                                 round_jiffies_relative(sysctl_stat_interval));
1828         }
1829 }
1830
1831 /*
1832  * Switch off vmstat processing and then fold all the remaining differentials
1833  * until the diffs stay at zero. The function is used by NOHZ and can only be
1834  * invoked when tick processing is not active.
1835  */
1836 /*
1837  * Check if the diffs for a certain cpu indicate that
1838  * an update is needed.
1839  */
1840 static bool need_update(int cpu)
1841 {
1842         struct zone *zone;
1843
1844         for_each_populated_zone(zone) {
1845                 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1846
1847                 BUILD_BUG_ON(sizeof(p->vm_stat_diff[0]) != 1);
1848 #ifdef CONFIG_NUMA
1849                 BUILD_BUG_ON(sizeof(p->vm_numa_stat_diff[0]) != 2);
1850 #endif
1851
1852                 /*
1853                  * The fast way of checking if there are any vmstat diffs.
1854                  */
1855                 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS *
1856                                sizeof(p->vm_stat_diff[0])))
1857                         return true;
1858 #ifdef CONFIG_NUMA
1859                 if (memchr_inv(p->vm_numa_stat_diff, 0, NR_VM_NUMA_STAT_ITEMS *
1860                                sizeof(p->vm_numa_stat_diff[0])))
1861                         return true;
1862 #endif
1863         }
1864         return false;
1865 }
1866
1867 /*
1868  * Switch off vmstat processing and then fold all the remaining differentials
1869  * until the diffs stay at zero. The function is used by NOHZ and can only be
1870  * invoked when tick processing is not active.
1871  */
1872 void quiet_vmstat(void)
1873 {
1874         if (system_state != SYSTEM_RUNNING)
1875                 return;
1876
1877         if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
1878                 return;
1879
1880         if (!need_update(smp_processor_id()))
1881                 return;
1882
1883         /*
1884          * Just refresh counters and do not care about the pending delayed
1885          * vmstat_update. It doesn't fire that often to matter and canceling
1886          * it would be too expensive from this path.
1887          * vmstat_shepherd will take care about that for us.
1888          */
1889         refresh_cpu_vm_stats(false);
1890 }
1891
1892 /*
1893  * Shepherd worker thread that checks the
1894  * differentials of processors that have their worker
1895  * threads for vm statistics updates disabled because of
1896  * inactivity.
1897  */
1898 static void vmstat_shepherd(struct work_struct *w);
1899
1900 static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
1901
1902 static void vmstat_shepherd(struct work_struct *w)
1903 {
1904         int cpu;
1905
1906         get_online_cpus();
1907         /* Check processors whose vmstat worker threads have been disabled */
1908         for_each_online_cpu(cpu) {
1909                 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
1910
1911                 if (!delayed_work_pending(dw) && need_update(cpu))
1912                         queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
1913         }
1914         put_online_cpus();
1915
1916         schedule_delayed_work(&shepherd,
1917                 round_jiffies_relative(sysctl_stat_interval));
1918 }
1919
1920 static void __init start_shepherd_timer(void)
1921 {
1922         int cpu;
1923
1924         for_each_possible_cpu(cpu)
1925                 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
1926                         vmstat_update);
1927
1928         schedule_delayed_work(&shepherd,
1929                 round_jiffies_relative(sysctl_stat_interval));
1930 }
1931
1932 static void __init init_cpu_node_state(void)
1933 {
1934         int node;
1935
1936         for_each_online_node(node) {
1937                 if (cpumask_weight(cpumask_of_node(node)) > 0)
1938                         node_set_state(node, N_CPU);
1939         }
1940 }
1941
1942 static int vmstat_cpu_online(unsigned int cpu)
1943 {
1944         refresh_zone_stat_thresholds();
1945         node_set_state(cpu_to_node(cpu), N_CPU);
1946         return 0;
1947 }
1948
1949 static int vmstat_cpu_down_prep(unsigned int cpu)
1950 {
1951         cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1952         return 0;
1953 }
1954
1955 static int vmstat_cpu_dead(unsigned int cpu)
1956 {
1957         const struct cpumask *node_cpus;
1958         int node;
1959
1960         node = cpu_to_node(cpu);
1961
1962         refresh_zone_stat_thresholds();
1963         node_cpus = cpumask_of_node(node);
1964         if (cpumask_weight(node_cpus) > 0)
1965                 return 0;
1966
1967         node_clear_state(node, N_CPU);
1968         return 0;
1969 }
1970
1971 #endif
1972
1973 struct workqueue_struct *mm_percpu_wq;
1974
1975 void __init init_mm_internals(void)
1976 {
1977         int ret __maybe_unused;
1978
1979         mm_percpu_wq = alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM, 0);
1980
1981 #ifdef CONFIG_SMP
1982         ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
1983                                         NULL, vmstat_cpu_dead);
1984         if (ret < 0)
1985                 pr_err("vmstat: failed to register 'dead' hotplug state\n");
1986
1987         ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "mm/vmstat:online",
1988                                         vmstat_cpu_online,
1989                                         vmstat_cpu_down_prep);
1990         if (ret < 0)
1991                 pr_err("vmstat: failed to register 'online' hotplug state\n");
1992
1993         get_online_cpus();
1994         init_cpu_node_state();
1995         put_online_cpus();
1996
1997         start_shepherd_timer();
1998 #endif
1999 #ifdef CONFIG_PROC_FS
2000         proc_create_seq("buddyinfo", 0444, NULL, &fragmentation_op);
2001         proc_create_seq("pagetypeinfo", 0400, NULL, &pagetypeinfo_op);
2002         proc_create_seq("vmstat", 0444, NULL, &vmstat_op);
2003         proc_create_seq("zoneinfo", 0444, NULL, &zoneinfo_op);
2004 #endif
2005 }
2006
2007 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
2008
2009 /*
2010  * Return an index indicating how much of the available free memory is
2011  * unusable for an allocation of the requested size.
2012  */
2013 static int unusable_free_index(unsigned int order,
2014                                 struct contig_page_info *info)
2015 {
2016         /* No free memory is interpreted as all free memory is unusable */
2017         if (info->free_pages == 0)
2018                 return 1000;
2019
2020         /*
2021          * Index should be a value between 0 and 1. Return a value to 3
2022          * decimal places.
2023          *
2024          * 0 => no fragmentation
2025          * 1 => high fragmentation
2026          */
2027         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
2028
2029 }
2030
2031 static void unusable_show_print(struct seq_file *m,
2032                                         pg_data_t *pgdat, struct zone *zone)
2033 {
2034         unsigned int order;
2035         int index;
2036         struct contig_page_info info;
2037
2038         seq_printf(m, "Node %d, zone %8s ",
2039                                 pgdat->node_id,
2040                                 zone->name);
2041         for (order = 0; order < MAX_ORDER; ++order) {
2042                 fill_contig_page_info(zone, order, &info);
2043                 index = unusable_free_index(order, &info);
2044                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2045         }
2046
2047         seq_putc(m, '\n');
2048 }
2049
2050 /*
2051  * Display unusable free space index
2052  *
2053  * The unusable free space index measures how much of the available free
2054  * memory cannot be used to satisfy an allocation of a given size and is a
2055  * value between 0 and 1. The higher the value, the more of free memory is
2056  * unusable and by implication, the worse the external fragmentation is. This
2057  * can be expressed as a percentage by multiplying by 100.
2058  */
2059 static int unusable_show(struct seq_file *m, void *arg)
2060 {
2061         pg_data_t *pgdat = (pg_data_t *)arg;
2062
2063         /* check memoryless node */
2064         if (!node_state(pgdat->node_id, N_MEMORY))
2065                 return 0;
2066
2067         walk_zones_in_node(m, pgdat, true, false, unusable_show_print);
2068
2069         return 0;
2070 }
2071
2072 static const struct seq_operations unusable_sops = {
2073         .start  = frag_start,
2074         .next   = frag_next,
2075         .stop   = frag_stop,
2076         .show   = unusable_show,
2077 };
2078
2079 DEFINE_SEQ_ATTRIBUTE(unusable);
2080
2081 static void extfrag_show_print(struct seq_file *m,
2082                                         pg_data_t *pgdat, struct zone *zone)
2083 {
2084         unsigned int order;
2085         int index;
2086
2087         /* Alloc on stack as interrupts are disabled for zone walk */
2088         struct contig_page_info info;
2089
2090         seq_printf(m, "Node %d, zone %8s ",
2091                                 pgdat->node_id,
2092                                 zone->name);
2093         for (order = 0; order < MAX_ORDER; ++order) {
2094                 fill_contig_page_info(zone, order, &info);
2095                 index = __fragmentation_index(order, &info);
2096                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2097         }
2098
2099         seq_putc(m, '\n');
2100 }
2101
2102 /*
2103  * Display fragmentation index for orders that allocations would fail for
2104  */
2105 static int extfrag_show(struct seq_file *m, void *arg)
2106 {
2107         pg_data_t *pgdat = (pg_data_t *)arg;
2108
2109         walk_zones_in_node(m, pgdat, true, false, extfrag_show_print);
2110
2111         return 0;
2112 }
2113
2114 static const struct seq_operations extfrag_sops = {
2115         .start  = frag_start,
2116         .next   = frag_next,
2117         .stop   = frag_stop,
2118         .show   = extfrag_show,
2119 };
2120
2121 DEFINE_SEQ_ATTRIBUTE(extfrag);
2122
2123 static int __init extfrag_debug_init(void)
2124 {
2125         struct dentry *extfrag_debug_root;
2126
2127         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
2128
2129         debugfs_create_file("unusable_index", 0444, extfrag_debug_root, NULL,
2130                             &unusable_fops);
2131
2132         debugfs_create_file("extfrag_index", 0444, extfrag_debug_root, NULL,
2133                             &extfrag_fops);
2134
2135         return 0;
2136 }
2137
2138 module_init(extfrag_debug_init);
2139 #endif