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 <[email protected]>
  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 __user *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
1207 #ifdef CONFIG_NUMA
1208         "zone_reclaim_failed",
1209 #endif
1210         "pginodesteal",
1211         "slabs_scanned",
1212         "kswapd_inodesteal",
1213         "kswapd_low_wmark_hit_quickly",
1214         "kswapd_high_wmark_hit_quickly",
1215         "pageoutrun",
1216
1217         "pgrotated",
1218
1219         "drop_pagecache",
1220         "drop_slab",
1221         "oom_kill",
1222
1223 #ifdef CONFIG_NUMA_BALANCING
1224         "numa_pte_updates",
1225         "numa_huge_pte_updates",
1226         "numa_hint_faults",
1227         "numa_hint_faults_local",
1228         "numa_pages_migrated",
1229 #endif
1230 #ifdef CONFIG_MIGRATION
1231         "pgmigrate_success",
1232         "pgmigrate_fail",
1233 #endif
1234 #ifdef CONFIG_COMPACTION
1235         "compact_migrate_scanned",
1236         "compact_free_scanned",
1237         "compact_isolated",
1238         "compact_stall",
1239         "compact_fail",
1240         "compact_success",
1241         "compact_daemon_wake",
1242         "compact_daemon_migrate_scanned",
1243         "compact_daemon_free_scanned",
1244 #endif
1245
1246 #ifdef CONFIG_HUGETLB_PAGE
1247         "htlb_buddy_alloc_success",
1248         "htlb_buddy_alloc_fail",
1249 #endif
1250         "unevictable_pgs_culled",
1251         "unevictable_pgs_scanned",
1252         "unevictable_pgs_rescued",
1253         "unevictable_pgs_mlocked",
1254         "unevictable_pgs_munlocked",
1255         "unevictable_pgs_cleared",
1256         "unevictable_pgs_stranded",
1257
1258 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1259         "thp_fault_alloc",
1260         "thp_fault_fallback",
1261         "thp_fault_fallback_charge",
1262         "thp_collapse_alloc",
1263         "thp_collapse_alloc_failed",
1264         "thp_file_alloc",
1265         "thp_file_fallback",
1266         "thp_file_fallback_charge",
1267         "thp_file_mapped",
1268         "thp_split_page",
1269         "thp_split_page_failed",
1270         "thp_deferred_split_page",
1271         "thp_split_pmd",
1272 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1273         "thp_split_pud",
1274 #endif
1275         "thp_zero_page_alloc",
1276         "thp_zero_page_alloc_failed",
1277         "thp_swpout",
1278         "thp_swpout_fallback",
1279 #endif
1280 #ifdef CONFIG_MEMORY_BALLOON
1281         "balloon_inflate",
1282         "balloon_deflate",
1283 #ifdef CONFIG_BALLOON_COMPACTION
1284         "balloon_migrate",
1285 #endif
1286 #endif /* CONFIG_MEMORY_BALLOON */
1287 #ifdef CONFIG_DEBUG_TLBFLUSH
1288         "nr_tlb_remote_flush",
1289         "nr_tlb_remote_flush_received",
1290         "nr_tlb_local_flush_all",
1291         "nr_tlb_local_flush_one",
1292 #endif /* CONFIG_DEBUG_TLBFLUSH */
1293
1294 #ifdef CONFIG_DEBUG_VM_VMACACHE
1295         "vmacache_find_calls",
1296         "vmacache_find_hits",
1297 #endif
1298 #ifdef CONFIG_SWAP
1299         "swap_ra",
1300         "swap_ra_hit",
1301 #endif
1302 #endif /* CONFIG_VM_EVENT_COUNTERS || CONFIG_MEMCG */
1303 };
1304 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA || CONFIG_MEMCG */
1305
1306 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1307      defined(CONFIG_PROC_FS)
1308 static void *frag_start(struct seq_file *m, loff_t *pos)
1309 {
1310         pg_data_t *pgdat;
1311         loff_t node = *pos;
1312
1313         for (pgdat = first_online_pgdat();
1314              pgdat && node;
1315              pgdat = next_online_pgdat(pgdat))
1316                 --node;
1317
1318         return pgdat;
1319 }
1320
1321 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1322 {
1323         pg_data_t *pgdat = (pg_data_t *)arg;
1324
1325         (*pos)++;
1326         return next_online_pgdat(pgdat);
1327 }
1328
1329 static void frag_stop(struct seq_file *m, void *arg)
1330 {
1331 }
1332
1333 /*
1334  * Walk zones in a node and print using a callback.
1335  * If @assert_populated is true, only use callback for zones that are populated.
1336  */
1337 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1338                 bool assert_populated, bool nolock,
1339                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1340 {
1341         struct zone *zone;
1342         struct zone *node_zones = pgdat->node_zones;
1343         unsigned long flags;
1344
1345         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1346                 if (assert_populated && !populated_zone(zone))
1347                         continue;
1348
1349                 if (!nolock)
1350                         spin_lock_irqsave(&zone->lock, flags);
1351                 print(m, pgdat, zone);
1352                 if (!nolock)
1353                         spin_unlock_irqrestore(&zone->lock, flags);
1354         }
1355 }
1356 #endif
1357
1358 #ifdef CONFIG_PROC_FS
1359 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1360                                                 struct zone *zone)
1361 {
1362         int order;
1363
1364         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1365         for (order = 0; order < MAX_ORDER; ++order)
1366                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
1367         seq_putc(m, '\n');
1368 }
1369
1370 /*
1371  * This walks the free areas for each zone.
1372  */
1373 static int frag_show(struct seq_file *m, void *arg)
1374 {
1375         pg_data_t *pgdat = (pg_data_t *)arg;
1376         walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1377         return 0;
1378 }
1379
1380 static void pagetypeinfo_showfree_print(struct seq_file *m,
1381                                         pg_data_t *pgdat, struct zone *zone)
1382 {
1383         int order, mtype;
1384
1385         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1386                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
1387                                         pgdat->node_id,
1388                                         zone->name,
1389                                         migratetype_names[mtype]);
1390                 for (order = 0; order < MAX_ORDER; ++order) {
1391                         unsigned long freecount = 0;
1392                         struct free_area *area;
1393                         struct list_head *curr;
1394                         bool overflow = false;
1395
1396                         area = &(zone->free_area[order]);
1397
1398                         list_for_each(curr, &area->free_list[mtype]) {
1399                                 /*
1400                                  * Cap the free_list iteration because it might
1401                                  * be really large and we are under a spinlock
1402                                  * so a long time spent here could trigger a
1403                                  * hard lockup detector. Anyway this is a
1404                                  * debugging tool so knowing there is a handful
1405                                  * of pages of this order should be more than
1406                                  * sufficient.
1407                                  */
1408                                 if (++freecount >= 100000) {
1409                                         overflow = true;
1410                                         break;
1411                                 }
1412                         }
1413                         seq_printf(m, "%s%6lu ", overflow ? ">" : "", freecount);
1414                         spin_unlock_irq(&zone->lock);
1415                         cond_resched();
1416                         spin_lock_irq(&zone->lock);
1417                 }
1418                 seq_putc(m, '\n');
1419         }
1420 }
1421
1422 /* Print out the free pages at each order for each migatetype */
1423 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
1424 {
1425         int order;
1426         pg_data_t *pgdat = (pg_data_t *)arg;
1427
1428         /* Print header */
1429         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1430         for (order = 0; order < MAX_ORDER; ++order)
1431                 seq_printf(m, "%6d ", order);
1432         seq_putc(m, '\n');
1433
1434         walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1435
1436         return 0;
1437 }
1438
1439 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1440                                         pg_data_t *pgdat, struct zone *zone)
1441 {
1442         int mtype;
1443         unsigned long pfn;
1444         unsigned long start_pfn = zone->zone_start_pfn;
1445         unsigned long end_pfn = zone_end_pfn(zone);
1446         unsigned long count[MIGRATE_TYPES] = { 0, };
1447
1448         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1449                 struct page *page;
1450
1451                 page = pfn_to_online_page(pfn);
1452                 if (!page)
1453                         continue;
1454
1455                 /* Watch for unexpected holes punched in the memmap */
1456                 if (!memmap_valid_within(pfn, page, zone))
1457                         continue;
1458
1459                 if (page_zone(page) != zone)
1460                         continue;
1461
1462                 mtype = get_pageblock_migratetype(page);
1463
1464                 if (mtype < MIGRATE_TYPES)
1465                         count[mtype]++;
1466         }
1467
1468         /* Print counts */
1469         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1470         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1471                 seq_printf(m, "%12lu ", count[mtype]);
1472         seq_putc(m, '\n');
1473 }
1474
1475 /* Print out the number of pageblocks for each migratetype */
1476 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1477 {
1478         int mtype;
1479         pg_data_t *pgdat = (pg_data_t *)arg;
1480
1481         seq_printf(m, "\n%-23s", "Number of blocks type ");
1482         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1483                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1484         seq_putc(m, '\n');
1485         walk_zones_in_node(m, pgdat, true, false,
1486                 pagetypeinfo_showblockcount_print);
1487
1488         return 0;
1489 }
1490
1491 /*
1492  * Print out the number of pageblocks for each migratetype that contain pages
1493  * of other types. This gives an indication of how well fallbacks are being
1494  * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1495  * to determine what is going on
1496  */
1497 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1498 {
1499 #ifdef CONFIG_PAGE_OWNER
1500         int mtype;
1501
1502         if (!static_branch_unlikely(&page_owner_inited))
1503                 return;
1504
1505         drain_all_pages(NULL);
1506
1507         seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1508         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1509                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1510         seq_putc(m, '\n');
1511
1512         walk_zones_in_node(m, pgdat, true, true,
1513                 pagetypeinfo_showmixedcount_print);
1514 #endif /* CONFIG_PAGE_OWNER */
1515 }
1516
1517 /*
1518  * This prints out statistics in relation to grouping pages by mobility.
1519  * It is expensive to collect so do not constantly read the file.
1520  */
1521 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1522 {
1523         pg_data_t *pgdat = (pg_data_t *)arg;
1524
1525         /* check memoryless node */
1526         if (!node_state(pgdat->node_id, N_MEMORY))
1527                 return 0;
1528
1529         seq_printf(m, "Page block order: %d\n", pageblock_order);
1530         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
1531         seq_putc(m, '\n');
1532         pagetypeinfo_showfree(m, pgdat);
1533         pagetypeinfo_showblockcount(m, pgdat);
1534         pagetypeinfo_showmixedcount(m, pgdat);
1535
1536         return 0;
1537 }
1538
1539 static const struct seq_operations fragmentation_op = {
1540         .start  = frag_start,
1541         .next   = frag_next,
1542         .stop   = frag_stop,
1543         .show   = frag_show,
1544 };
1545
1546 static const struct seq_operations pagetypeinfo_op = {
1547         .start  = frag_start,
1548         .next   = frag_next,
1549         .stop   = frag_stop,
1550         .show   = pagetypeinfo_show,
1551 };
1552
1553 static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1554 {
1555         int zid;
1556
1557         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1558                 struct zone *compare = &pgdat->node_zones[zid];
1559
1560                 if (populated_zone(compare))
1561                         return zone == compare;
1562         }
1563
1564         return false;
1565 }
1566
1567 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1568                                                         struct zone *zone)
1569 {
1570         int i;
1571         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1572         if (is_zone_first_populated(pgdat, zone)) {
1573                 seq_printf(m, "\n  per-node stats");
1574                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1575                         seq_printf(m, "\n      %-12s %lu", node_stat_name(i),
1576                                    node_page_state(pgdat, i));
1577                 }
1578         }
1579         seq_printf(m,
1580                    "\n  pages free     %lu"
1581                    "\n        min      %lu"
1582                    "\n        low      %lu"
1583                    "\n        high     %lu"
1584                    "\n        spanned  %lu"
1585                    "\n        present  %lu"
1586                    "\n        managed  %lu",
1587                    zone_page_state(zone, NR_FREE_PAGES),
1588                    min_wmark_pages(zone),
1589                    low_wmark_pages(zone),
1590                    high_wmark_pages(zone),
1591                    zone->spanned_pages,
1592                    zone->present_pages,
1593                    zone_managed_pages(zone));
1594
1595         /* If unpopulated, no other information is useful */
1596         if (!populated_zone(zone)) {
1597                 seq_putc(m, '\n');
1598                 return;
1599         }
1600
1601         seq_printf(m,
1602                    "\n        protection: (%ld",
1603                    zone->lowmem_reserve[0]);
1604         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1605                 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1606         seq_putc(m, ')');
1607
1608         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1609                 seq_printf(m, "\n      %-12s %lu", zone_stat_name(i),
1610                            zone_page_state(zone, i));
1611
1612 #ifdef CONFIG_NUMA
1613         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1614                 seq_printf(m, "\n      %-12s %lu", numa_stat_name(i),
1615                            zone_numa_state_snapshot(zone, i));
1616 #endif
1617
1618         seq_printf(m, "\n  pagesets");
1619         for_each_online_cpu(i) {
1620                 struct per_cpu_pageset *pageset;
1621
1622                 pageset = per_cpu_ptr(zone->pageset, i);
1623                 seq_printf(m,
1624                            "\n    cpu: %i"
1625                            "\n              count: %i"
1626                            "\n              high:  %i"
1627                            "\n              batch: %i",
1628                            i,
1629                            pageset->pcp.count,
1630                            pageset->pcp.high,
1631                            pageset->pcp.batch);
1632 #ifdef CONFIG_SMP
1633                 seq_printf(m, "\n  vm stats threshold: %d",
1634                                 pageset->stat_threshold);
1635 #endif
1636         }
1637         seq_printf(m,
1638                    "\n  node_unreclaimable:  %u"
1639                    "\n  start_pfn:           %lu",
1640                    pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1641                    zone->zone_start_pfn);
1642         seq_putc(m, '\n');
1643 }
1644
1645 /*
1646  * Output information about zones in @pgdat.  All zones are printed regardless
1647  * of whether they are populated or not: lowmem_reserve_ratio operates on the
1648  * set of all zones and userspace would not be aware of such zones if they are
1649  * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1650  */
1651 static int zoneinfo_show(struct seq_file *m, void *arg)
1652 {
1653         pg_data_t *pgdat = (pg_data_t *)arg;
1654         walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1655         return 0;
1656 }
1657
1658 static const struct seq_operations zoneinfo_op = {
1659         .start  = frag_start, /* iterate over all zones. The same as in
1660                                * fragmentation. */
1661         .next   = frag_next,
1662         .stop   = frag_stop,
1663         .show   = zoneinfo_show,
1664 };
1665
1666 #define NR_VMSTAT_ITEMS (NR_VM_ZONE_STAT_ITEMS + \
1667                          NR_VM_NUMA_STAT_ITEMS + \
1668                          NR_VM_NODE_STAT_ITEMS + \
1669                          NR_VM_WRITEBACK_STAT_ITEMS + \
1670                          (IS_ENABLED(CONFIG_VM_EVENT_COUNTERS) ? \
1671                           NR_VM_EVENT_ITEMS : 0))
1672
1673 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1674 {
1675         unsigned long *v;
1676         int i;
1677
1678         if (*pos >= NR_VMSTAT_ITEMS)
1679                 return NULL;
1680
1681         BUILD_BUG_ON(ARRAY_SIZE(vmstat_text) < NR_VMSTAT_ITEMS);
1682         v = kmalloc_array(NR_VMSTAT_ITEMS, sizeof(unsigned long), GFP_KERNEL);
1683         m->private = v;
1684         if (!v)
1685                 return ERR_PTR(-ENOMEM);
1686         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1687                 v[i] = global_zone_page_state(i);
1688         v += NR_VM_ZONE_STAT_ITEMS;
1689
1690 #ifdef CONFIG_NUMA
1691         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++)
1692                 v[i] = global_numa_state(i);
1693         v += NR_VM_NUMA_STAT_ITEMS;
1694 #endif
1695
1696         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
1697                 v[i] = global_node_page_state(i);
1698         v += NR_VM_NODE_STAT_ITEMS;
1699
1700         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1701                             v + NR_DIRTY_THRESHOLD);
1702         v += NR_VM_WRITEBACK_STAT_ITEMS;
1703
1704 #ifdef CONFIG_VM_EVENT_COUNTERS
1705         all_vm_events(v);
1706         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1707         v[PGPGOUT] /= 2;
1708 #endif
1709         return (unsigned long *)m->private + *pos;
1710 }
1711
1712 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1713 {
1714         (*pos)++;
1715         if (*pos >= NR_VMSTAT_ITEMS)
1716                 return NULL;
1717         return (unsigned long *)m->private + *pos;
1718 }
1719
1720 static int vmstat_show(struct seq_file *m, void *arg)
1721 {
1722         unsigned long *l = arg;
1723         unsigned long off = l - (unsigned long *)m->private;
1724
1725         seq_puts(m, vmstat_text[off]);
1726         seq_put_decimal_ull(m, " ", *l);
1727         seq_putc(m, '\n');
1728
1729         if (off == NR_VMSTAT_ITEMS - 1) {
1730                 /*
1731                  * We've come to the end - add any deprecated counters to avoid
1732                  * breaking userspace which might depend on them being present.
1733                  */
1734                 seq_puts(m, "nr_unstable 0\n");
1735         }
1736         return 0;
1737 }
1738
1739 static void vmstat_stop(struct seq_file *m, void *arg)
1740 {
1741         kfree(m->private);
1742         m->private = NULL;
1743 }
1744
1745 static const struct seq_operations vmstat_op = {
1746         .start  = vmstat_start,
1747         .next   = vmstat_next,
1748         .stop   = vmstat_stop,
1749         .show   = vmstat_show,
1750 };
1751 #endif /* CONFIG_PROC_FS */
1752
1753 #ifdef CONFIG_SMP
1754 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1755 int sysctl_stat_interval __read_mostly = HZ;
1756
1757 #ifdef CONFIG_PROC_FS
1758 static void refresh_vm_stats(struct work_struct *work)
1759 {
1760         refresh_cpu_vm_stats(true);
1761 }
1762
1763 int vmstat_refresh(struct ctl_table *table, int write,
1764                    void __user *buffer, size_t *lenp, loff_t *ppos)
1765 {
1766         long val;
1767         int err;
1768         int i;
1769
1770         /*
1771          * The regular update, every sysctl_stat_interval, may come later
1772          * than expected: leaving a significant amount in per_cpu buckets.
1773          * This is particularly misleading when checking a quantity of HUGE
1774          * pages, immediately after running a test.  /proc/sys/vm/stat_refresh,
1775          * which can equally be echo'ed to or cat'ted from (by root),
1776          * can be used to update the stats just before reading them.
1777          *
1778          * Oh, and since global_zone_page_state() etc. are so careful to hide
1779          * transiently negative values, report an error here if any of
1780          * the stats is negative, so we know to go looking for imbalance.
1781          */
1782         err = schedule_on_each_cpu(refresh_vm_stats);
1783         if (err)
1784                 return err;
1785         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1786                 val = atomic_long_read(&vm_zone_stat[i]);
1787                 if (val < 0) {
1788                         pr_warn("%s: %s %ld\n",
1789                                 __func__, zone_stat_name(i), val);
1790                         err = -EINVAL;
1791                 }
1792         }
1793 #ifdef CONFIG_NUMA
1794         for (i = 0; i < NR_VM_NUMA_STAT_ITEMS; i++) {
1795                 val = atomic_long_read(&vm_numa_stat[i]);
1796                 if (val < 0) {
1797                         pr_warn("%s: %s %ld\n",
1798                                 __func__, numa_stat_name(i), val);
1799                         err = -EINVAL;
1800                 }
1801         }
1802 #endif
1803         if (err)
1804                 return err;
1805         if (write)
1806                 *ppos += *lenp;
1807         else
1808                 *lenp = 0;
1809         return 0;
1810 }
1811 #endif /* CONFIG_PROC_FS */
1812
1813 static void vmstat_update(struct work_struct *w)
1814 {
1815         if (refresh_cpu_vm_stats(true)) {
1816                 /*
1817                  * Counters were updated so we expect more updates
1818                  * to occur in the future. Keep on running the
1819                  * update worker thread.
1820                  */
1821                 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
1822                                 this_cpu_ptr(&vmstat_work),
1823                                 round_jiffies_relative(sysctl_stat_interval));
1824         }
1825 }
1826
1827 /*
1828  * Switch off vmstat processing and then fold all the remaining differentials
1829  * until the diffs stay at zero. The function is used by NOHZ and can only be
1830  * invoked when tick processing is not active.
1831  */
1832 /*
1833  * Check if the diffs for a certain cpu indicate that
1834  * an update is needed.
1835  */
1836 static bool need_update(int cpu)
1837 {
1838         struct zone *zone;
1839
1840         for_each_populated_zone(zone) {
1841                 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1842
1843                 BUILD_BUG_ON(sizeof(p->vm_stat_diff[0]) != 1);
1844 #ifdef CONFIG_NUMA
1845                 BUILD_BUG_ON(sizeof(p->vm_numa_stat_diff[0]) != 2);
1846 #endif
1847
1848                 /*
1849                  * The fast way of checking if there are any vmstat diffs.
1850                  */
1851                 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS *
1852                                sizeof(p->vm_stat_diff[0])))
1853                         return true;
1854 #ifdef CONFIG_NUMA
1855                 if (memchr_inv(p->vm_numa_stat_diff, 0, NR_VM_NUMA_STAT_ITEMS *
1856                                sizeof(p->vm_numa_stat_diff[0])))
1857                         return true;
1858 #endif
1859         }
1860         return false;
1861 }
1862
1863 /*
1864  * Switch off vmstat processing and then fold all the remaining differentials
1865  * until the diffs stay at zero. The function is used by NOHZ and can only be
1866  * invoked when tick processing is not active.
1867  */
1868 void quiet_vmstat(void)
1869 {
1870         if (system_state != SYSTEM_RUNNING)
1871                 return;
1872
1873         if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
1874                 return;
1875
1876         if (!need_update(smp_processor_id()))
1877                 return;
1878
1879         /*
1880          * Just refresh counters and do not care about the pending delayed
1881          * vmstat_update. It doesn't fire that often to matter and canceling
1882          * it would be too expensive from this path.
1883          * vmstat_shepherd will take care about that for us.
1884          */
1885         refresh_cpu_vm_stats(false);
1886 }
1887
1888 /*
1889  * Shepherd worker thread that checks the
1890  * differentials of processors that have their worker
1891  * threads for vm statistics updates disabled because of
1892  * inactivity.
1893  */
1894 static void vmstat_shepherd(struct work_struct *w);
1895
1896 static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
1897
1898 static void vmstat_shepherd(struct work_struct *w)
1899 {
1900         int cpu;
1901
1902         get_online_cpus();
1903         /* Check processors whose vmstat worker threads have been disabled */
1904         for_each_online_cpu(cpu) {
1905                 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
1906
1907                 if (!delayed_work_pending(dw) && need_update(cpu))
1908                         queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
1909         }
1910         put_online_cpus();
1911
1912         schedule_delayed_work(&shepherd,
1913                 round_jiffies_relative(sysctl_stat_interval));
1914 }
1915
1916 static void __init start_shepherd_timer(void)
1917 {
1918         int cpu;
1919
1920         for_each_possible_cpu(cpu)
1921                 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
1922                         vmstat_update);
1923
1924         schedule_delayed_work(&shepherd,
1925                 round_jiffies_relative(sysctl_stat_interval));
1926 }
1927
1928 static void __init init_cpu_node_state(void)
1929 {
1930         int node;
1931
1932         for_each_online_node(node) {
1933                 if (cpumask_weight(cpumask_of_node(node)) > 0)
1934                         node_set_state(node, N_CPU);
1935         }
1936 }
1937
1938 static int vmstat_cpu_online(unsigned int cpu)
1939 {
1940         refresh_zone_stat_thresholds();
1941         node_set_state(cpu_to_node(cpu), N_CPU);
1942         return 0;
1943 }
1944
1945 static int vmstat_cpu_down_prep(unsigned int cpu)
1946 {
1947         cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1948         return 0;
1949 }
1950
1951 static int vmstat_cpu_dead(unsigned int cpu)
1952 {
1953         const struct cpumask *node_cpus;
1954         int node;
1955
1956         node = cpu_to_node(cpu);
1957
1958         refresh_zone_stat_thresholds();
1959         node_cpus = cpumask_of_node(node);
1960         if (cpumask_weight(node_cpus) > 0)
1961                 return 0;
1962
1963         node_clear_state(node, N_CPU);
1964         return 0;
1965 }
1966
1967 #endif
1968
1969 struct workqueue_struct *mm_percpu_wq;
1970
1971 void __init init_mm_internals(void)
1972 {
1973         int ret __maybe_unused;
1974
1975         mm_percpu_wq = alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM, 0);
1976
1977 #ifdef CONFIG_SMP
1978         ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
1979                                         NULL, vmstat_cpu_dead);
1980         if (ret < 0)
1981                 pr_err("vmstat: failed to register 'dead' hotplug state\n");
1982
1983         ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "mm/vmstat:online",
1984                                         vmstat_cpu_online,
1985                                         vmstat_cpu_down_prep);
1986         if (ret < 0)
1987                 pr_err("vmstat: failed to register 'online' hotplug state\n");
1988
1989         get_online_cpus();
1990         init_cpu_node_state();
1991         put_online_cpus();
1992
1993         start_shepherd_timer();
1994 #endif
1995 #ifdef CONFIG_PROC_FS
1996         proc_create_seq("buddyinfo", 0444, NULL, &fragmentation_op);
1997         proc_create_seq("pagetypeinfo", 0400, NULL, &pagetypeinfo_op);
1998         proc_create_seq("vmstat", 0444, NULL, &vmstat_op);
1999         proc_create_seq("zoneinfo", 0444, NULL, &zoneinfo_op);
2000 #endif
2001 }
2002
2003 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
2004
2005 /*
2006  * Return an index indicating how much of the available free memory is
2007  * unusable for an allocation of the requested size.
2008  */
2009 static int unusable_free_index(unsigned int order,
2010                                 struct contig_page_info *info)
2011 {
2012         /* No free memory is interpreted as all free memory is unusable */
2013         if (info->free_pages == 0)
2014                 return 1000;
2015
2016         /*
2017          * Index should be a value between 0 and 1. Return a value to 3
2018          * decimal places.
2019          *
2020          * 0 => no fragmentation
2021          * 1 => high fragmentation
2022          */
2023         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
2024
2025 }
2026
2027 static void unusable_show_print(struct seq_file *m,
2028                                         pg_data_t *pgdat, struct zone *zone)
2029 {
2030         unsigned int order;
2031         int index;
2032         struct contig_page_info info;
2033
2034         seq_printf(m, "Node %d, zone %8s ",
2035                                 pgdat->node_id,
2036                                 zone->name);
2037         for (order = 0; order < MAX_ORDER; ++order) {
2038                 fill_contig_page_info(zone, order, &info);
2039                 index = unusable_free_index(order, &info);
2040                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2041         }
2042
2043         seq_putc(m, '\n');
2044 }
2045
2046 /*
2047  * Display unusable free space index
2048  *
2049  * The unusable free space index measures how much of the available free
2050  * memory cannot be used to satisfy an allocation of a given size and is a
2051  * value between 0 and 1. The higher the value, the more of free memory is
2052  * unusable and by implication, the worse the external fragmentation is. This
2053  * can be expressed as a percentage by multiplying by 100.
2054  */
2055 static int unusable_show(struct seq_file *m, void *arg)
2056 {
2057         pg_data_t *pgdat = (pg_data_t *)arg;
2058
2059         /* check memoryless node */
2060         if (!node_state(pgdat->node_id, N_MEMORY))
2061                 return 0;
2062
2063         walk_zones_in_node(m, pgdat, true, false, unusable_show_print);
2064
2065         return 0;
2066 }
2067
2068 static const struct seq_operations unusable_op = {
2069         .start  = frag_start,
2070         .next   = frag_next,
2071         .stop   = frag_stop,
2072         .show   = unusable_show,
2073 };
2074
2075 static int unusable_open(struct inode *inode, struct file *file)
2076 {
2077         return seq_open(file, &unusable_op);
2078 }
2079
2080 static const struct file_operations unusable_file_ops = {
2081         .open           = unusable_open,
2082         .read           = seq_read,
2083         .llseek         = seq_lseek,
2084         .release        = seq_release,
2085 };
2086
2087 static void extfrag_show_print(struct seq_file *m,
2088                                         pg_data_t *pgdat, struct zone *zone)
2089 {
2090         unsigned int order;
2091         int index;
2092
2093         /* Alloc on stack as interrupts are disabled for zone walk */
2094         struct contig_page_info info;
2095
2096         seq_printf(m, "Node %d, zone %8s ",
2097                                 pgdat->node_id,
2098                                 zone->name);
2099         for (order = 0; order < MAX_ORDER; ++order) {
2100                 fill_contig_page_info(zone, order, &info);
2101                 index = __fragmentation_index(order, &info);
2102                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2103         }
2104
2105         seq_putc(m, '\n');
2106 }
2107
2108 /*
2109  * Display fragmentation index for orders that allocations would fail for
2110  */
2111 static int extfrag_show(struct seq_file *m, void *arg)
2112 {
2113         pg_data_t *pgdat = (pg_data_t *)arg;
2114
2115         walk_zones_in_node(m, pgdat, true, false, extfrag_show_print);
2116
2117         return 0;
2118 }
2119
2120 static const struct seq_operations extfrag_op = {
2121         .start  = frag_start,
2122         .next   = frag_next,
2123         .stop   = frag_stop,
2124         .show   = extfrag_show,
2125 };
2126
2127 static int extfrag_open(struct inode *inode, struct file *file)
2128 {
2129         return seq_open(file, &extfrag_op);
2130 }
2131
2132 static const struct file_operations extfrag_file_ops = {
2133         .open           = extfrag_open,
2134         .read           = seq_read,
2135         .llseek         = seq_lseek,
2136         .release        = seq_release,
2137 };
2138
2139 static int __init extfrag_debug_init(void)
2140 {
2141         struct dentry *extfrag_debug_root;
2142
2143         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
2144
2145         debugfs_create_file("unusable_index", 0444, extfrag_debug_root, NULL,
2146                             &unusable_file_ops);
2147
2148         debugfs_create_file("extfrag_index", 0444, extfrag_debug_root, NULL,
2149                             &extfrag_file_ops);
2150
2151         return 0;
2152 }
2153
2154 module_init(extfrag_debug_init);
2155 #endif