mm/vmstat.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  linux/mm/vmstat.c
   4  *
   5  *  Manages VM statistics
   6  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   7  *
   8  *  zoned VM statistics
   9  *  Copyright (C) 2006 Silicon Graphics, Inc.,
  10  *              Christoph Lameter <christoph@lameter.com>
  11  *  Copyright (C) 2008-2014 Christoph Lameter
  12  */
  13 #include <linux/fs.h>
  14 #include <linux/mm.h>
  15 #include <linux/err.h>
  16 #include <linux/module.h>
  17 #include <linux/slab.h>
  18 #include <linux/cpu.h>
  19 #include <linux/cpumask.h>
  20 #include <linux/vmstat.h>
  21 #include <linux/proc_fs.h>
  22 #include <linux/seq_file.h>
  23 #include <linux/debugfs.h>
  24 #include <linux/sched.h>
  25 #include <linux/math64.h>
  26 #include <linux/writeback.h>
  27 #include <linux/compaction.h>
  28 #include <linux/mm_inline.h>
  29 #include <linux/page_ext.h>
  30 #include <linux/page_owner.h>
  31
  32 #include "internal.h"
  33
  34 #ifdef CONFIG_NUMA
  35 int sysctl_vm_numa_stat = ENABLE_NUMA_STAT;
  36
  37 /* zero numa counters within a zone */
  38 static void zero_zone_numa_counters(struct zone *zone)
  39 {
  40         int item, cpu;
  41
  42         for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++) {
  43                 atomic_long_set(&zone->vm_numa_event[item], 0);
  44                 for_each_online_cpu(cpu) {
  45                         per_cpu_ptr(zone->per_cpu_zonestats, cpu)->vm_numa_event[item]
  46                                                 = 0;
  47                 }
  48         }
  49 }
  50
  51 /* zero numa counters of all the populated zones */
  52 static void zero_zones_numa_counters(void)
  53 {
  54         struct zone *zone;
  55
  56         for_each_populated_zone(zone)
  57                 zero_zone_numa_counters(zone);
  58 }
  59
  60 /* zero global numa counters */
  61 static void zero_global_numa_counters(void)
  62 {
  63         int item;
  64
  65         for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++)
  66                 atomic_long_set(&vm_numa_event[item], 0);
  67 }
  68
  69 static void invalid_numa_statistics(void)
  70 {
  71         zero_zones_numa_counters();
  72         zero_global_numa_counters();
  73 }
  74
  75 static DEFINE_MUTEX(vm_numa_stat_lock);
  76
  77 int sysctl_vm_numa_stat_handler(struct ctl_table *table, int write,
  78                 void *buffer, size_t *length, loff_t *ppos)
  79 {
  80         int ret, oldval;
  81
  82         mutex_lock(&vm_numa_stat_lock);
  83         if (write)
  84                 oldval = sysctl_vm_numa_stat;
  85         ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
  86         if (ret || !write)
  87                 goto out;
  88
  89         if (oldval == sysctl_vm_numa_stat)
  90                 goto out;
  91         else if (sysctl_vm_numa_stat == ENABLE_NUMA_STAT) {
  92                 static_branch_enable(&vm_numa_stat_key);
  93                 pr_info("enable numa statistics\n");
  94         } else {
  95                 static_branch_disable(&vm_numa_stat_key);
  96                 invalid_numa_statistics();
  97                 pr_info("disable numa statistics, and clear numa counters\n");
  98         }
  99
 100 out:
 101         mutex_unlock(&vm_numa_stat_lock);
 102         return ret;
 103 }
 104 #endif
 105
 106 #ifdef CONFIG_VM_EVENT_COUNTERS
 107 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
 108 EXPORT_PER_CPU_SYMBOL(vm_event_states);
 109
 110 static void sum_vm_events(unsigned long *ret)
 111 {
 112         int cpu;
 113         int i;
 114
 115         memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
 116
 117         for_each_online_cpu(cpu) {
 118                 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
 119
 120                 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
 121                         ret[i] += this->event[i];
 122         }
 123 }
 124
 125 /*
 126  * Accumulate the vm event counters across all CPUs.
 127  * The result is unavoidably approximate - it can change
 128  * during and after execution of this function.
 129 */
 130 void all_vm_events(unsigned long *ret)
 131 {
 132         cpus_read_lock();
 133         sum_vm_events(ret);
 134         cpus_read_unlock();
 135 }
 136 EXPORT_SYMBOL_GPL(all_vm_events);
 137
 138 /*
 139  * Fold the foreign cpu events into our own.
 140  *
 141  * This is adding to the events on one processor
 142  * but keeps the global counts constant.
 143  */
 144 void vm_events_fold_cpu(int cpu)
 145 {
 146         struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
 147         int i;
 148
 149         for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
 150                 count_vm_events(i, fold_state->event[i]);
 151                 fold_state->event[i] = 0;
 152         }
 153 }
 154
 155 #endif /* CONFIG_VM_EVENT_COUNTERS */
 156
 157 /*
 158  * Manage combined zone based / global counters
 159  *
 160  * vm_stat contains the global counters
 161  */
 162 atomic_long_t vm_zone_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
 163 atomic_long_t vm_node_stat[NR_VM_NODE_STAT_ITEMS] __cacheline_aligned_in_smp;
 164 atomic_long_t vm_numa_event[NR_VM_NUMA_EVENT_ITEMS] __cacheline_aligned_in_smp;
 165 EXPORT_SYMBOL(vm_zone_stat);
 166 EXPORT_SYMBOL(vm_node_stat);
 167
 168 #ifdef CONFIG_NUMA
 169 static void fold_vm_zone_numa_events(struct zone *zone)
 170 {
 171         unsigned long zone_numa_events[NR_VM_NUMA_EVENT_ITEMS] = { 0, };
 172         int cpu;
 173         enum numa_stat_item item;
 174
 175         for_each_online_cpu(cpu) {
 176                 struct per_cpu_zonestat *pzstats;
 177
 178                 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
 179                 for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++)
 180                         zone_numa_events[item] += xchg(&pzstats->vm_numa_event[item], 0);
 181         }
 182
 183         for (item = 0; item < NR_VM_NUMA_EVENT_ITEMS; item++)
 184                 zone_numa_event_add(zone_numa_events[item], zone, item);
 185 }
 186
 187 void fold_vm_numa_events(void)
 188 {
 189         struct zone *zone;
 190
 191         for_each_populated_zone(zone)
 192                 fold_vm_zone_numa_events(zone);
 193 }
 194 #endif
 195
 196 #ifdef CONFIG_SMP
 197
 198 int calculate_pressure_threshold(struct zone *zone)
 199 {
 200         int threshold;
 201         int watermark_distance;
 202
 203         /*
 204          * As vmstats are not up to date, there is drift between the estimated
 205          * and real values. For high thresholds and a high number of CPUs, it
 206          * is possible for the min watermark to be breached while the estimated
 207          * value looks fine. The pressure threshold is a reduced value such
 208          * that even the maximum amount of drift will not accidentally breach
 209          * the min watermark
 210          */
 211         watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
 212         threshold = max(1, (int)(watermark_distance / num_online_cpus()));
 213
 214         /*
 215          * Maximum threshold is 125
 216          */
 217         threshold = min(125, threshold);
 218
 219         return threshold;
 220 }
 221
 222 int calculate_normal_threshold(struct zone *zone)
 223 {
 224         int threshold;
 225         int mem;        /* memory in 128 MB units */
 226
 227         /*
 228          * The threshold scales with the number of processors and the amount
 229          * of memory per zone. More memory means that we can defer updates for
 230          * longer, more processors could lead to more contention.
 231          * fls() is used to have a cheap way of logarithmic scaling.
 232          *
 233          * Some sample thresholds:
 234          *
 235          * Threshold    Processors      (fls)   Zonesize        fls(mem)+1
 236          * ------------------------------------------------------------------
 237          * 8            1               1       0.9-1 GB        4
 238          * 16           2               2       0.9-1 GB        4
 239          * 20           2               2       1-2 GB          5
 240          * 24           2               2       2-4 GB          6
 241          * 28           2               2       4-8 GB          7
 242          * 32           2               2       8-16 GB         8
 243          * 4            2               2       <128M           1
 244          * 30           4               3       2-4 GB          5
 245          * 48           4               3       8-16 GB         8
 246          * 32           8               4       1-2 GB          4
 247          * 32           8               4       0.9-1GB         4
 248          * 10           16              5       <128M           1
 249          * 40           16              5       900M            4
 250          * 70           64              7       2-4 GB          5
 251          * 84           64              7       4-8 GB          6
 252          * 108          512             9       4-8 GB          6
 253          * 125          1024            10      8-16 GB         8
 254          * 125          1024            10      16-32 GB        9
 255          */
 256
 257         mem = zone_managed_pages(zone) >> (27 - PAGE_SHIFT);
 258
 259         threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
 260
 261         /*
 262          * Maximum threshold is 125
 263          */
 264         threshold = min(125, threshold);
 265
 266         return threshold;
 267 }
 268
 269 /*
 270  * Refresh the thresholds for each zone.
 271  */
 272 void refresh_zone_stat_thresholds(void)
 273 {
 274         struct pglist_data *pgdat;
 275         struct zone *zone;
 276         int cpu;
 277         int threshold;
 278
 279         /* Zero current pgdat thresholds */
 280         for_each_online_pgdat(pgdat) {
 281                 for_each_online_cpu(cpu) {
 282                         per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold = 0;
 283                 }
 284         }
 285
 286         for_each_populated_zone(zone) {
 287                 struct pglist_data *pgdat = zone->zone_pgdat;
 288                 unsigned long max_drift, tolerate_drift;
 289
 290                 threshold = calculate_normal_threshold(zone);
 291
 292                 for_each_online_cpu(cpu) {
 293                         int pgdat_threshold;
 294
 295                         per_cpu_ptr(zone->per_cpu_zonestats, cpu)->stat_threshold
 296                                                         = threshold;
 297
 298                         /* Base nodestat threshold on the largest populated zone. */
 299                         pgdat_threshold = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold;
 300                         per_cpu_ptr(pgdat->per_cpu_nodestats, cpu)->stat_threshold
 301                                 = max(threshold, pgdat_threshold);
 302                 }
 303
 304                 /*
 305                  * Only set percpu_drift_mark if there is a danger that
 306                  * NR_FREE_PAGES reports the low watermark is ok when in fact
 307                  * the min watermark could be breached by an allocation
 308                  */
 309                 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
 310                 max_drift = num_online_cpus() * threshold;
 311                 if (max_drift > tolerate_drift)
 312                         zone->percpu_drift_mark = high_wmark_pages(zone) +
 313                                         max_drift;
 314         }
 315 }
 316
 317 void set_pgdat_percpu_threshold(pg_data_t *pgdat,
 318                                 int (*calculate_pressure)(struct zone *))
 319 {
 320         struct zone *zone;
 321         int cpu;
 322         int threshold;
 323         int i;
 324
 325         for (i = 0; i < pgdat->nr_zones; i++) {
 326                 zone = &pgdat->node_zones[i];
 327                 if (!zone->percpu_drift_mark)
 328                         continue;
 329
 330                 threshold = (*calculate_pressure)(zone);
 331                 for_each_online_cpu(cpu)
 332                         per_cpu_ptr(zone->per_cpu_zonestats, cpu)->stat_threshold
 333                                                         = threshold;
 334         }
 335 }
 336
 337 /*
 338  * For use when we know that interrupts are disabled,
 339  * or when we know that preemption is disabled and that
 340  * particular counter cannot be updated from interrupt context.
 341  */
 342 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 343                            long delta)
 344 {
 345         struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
 346         s8 __percpu *p = pcp->vm_stat_diff + item;
 347         long x;
 348         long t;
 349
 350         /*
 351          * Accurate vmstat updates require a RMW. On !PREEMPT_RT kernels,
 352          * atomicity is provided by IRQs being disabled -- either explicitly
 353          * or via local_lock_irq. On PREEMPT_RT, local_lock_irq only disables
 354          * CPU migrations and preemption potentially corrupts a counter so
 355          * disable preemption.
 356          */
 357         if (IS_ENABLED(CONFIG_PREEMPT_RT))
 358                 preempt_disable();
 359
 360         x = delta + __this_cpu_read(*p);
 361
 362         t = __this_cpu_read(pcp->stat_threshold);
 363
 364         if (unlikely(abs(x) > t)) {
 365                 zone_page_state_add(x, zone, item);
 366                 x = 0;
 367         }
 368         __this_cpu_write(*p, x);
 369
 370         if (IS_ENABLED(CONFIG_PREEMPT_RT))
 371                 preempt_enable();
 372 }
 373 EXPORT_SYMBOL(__mod_zone_page_state);
 374
 375 void __mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
 376                                 long delta)
 377 {
 378         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 379         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 380         long x;
 381         long t;
 382
 383         if (vmstat_item_in_bytes(item)) {
 384                 /*
 385                  * Only cgroups use subpage accounting right now; at
 386                  * the global level, these items still change in
 387                  * multiples of whole pages. Store them as pages
 388                  * internally to keep the per-cpu counters compact.
 389                  */
 390                 VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
 391                 delta >>= PAGE_SHIFT;
 392         }
 393
 394         /* See __mod_node_page_state */
 395         if (IS_ENABLED(CONFIG_PREEMPT_RT))
 396                 preempt_disable();
 397
 398         x = delta + __this_cpu_read(*p);
 399
 400         t = __this_cpu_read(pcp->stat_threshold);
 401
 402         if (unlikely(abs(x) > t)) {
 403                 node_page_state_add(x, pgdat, item);
 404                 x = 0;
 405         }
 406         __this_cpu_write(*p, x);
 407
 408         if (IS_ENABLED(CONFIG_PREEMPT_RT))
 409                 preempt_enable();
 410 }
 411 EXPORT_SYMBOL(__mod_node_page_state);
 412
 413 /*
 414  * Optimized increment and decrement functions.
 415  *
 416  * These are only for a single page and therefore can take a struct page *
 417  * argument instead of struct zone *. This allows the inclusion of the code
 418  * generated for page_zone(page) into the optimized functions.
 419  *
 420  * No overflow check is necessary and therefore the differential can be
 421  * incremented or decremented in place which may allow the compilers to
 422  * generate better code.
 423  * The increment or decrement is known and therefore one boundary check can
 424  * be omitted.
 425  *
 426  * NOTE: These functions are very performance sensitive. Change only
 427  * with care.
 428  *
 429  * Some processors have inc/dec instructions that are atomic vs an interrupt.
 430  * However, the code must first determine the differential location in a zone
 431  * based on the processor number and then inc/dec the counter. There is no
 432  * guarantee without disabling preemption that the processor will not change
 433  * in between and therefore the atomicity vs. interrupt cannot be exploited
 434  * in a useful way here.
 435  */
 436 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
 437 {
 438         struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
 439         s8 __percpu *p = pcp->vm_stat_diff + item;
 440         s8 v, t;
 441
 442         /* See __mod_node_page_state */
 443         if (IS_ENABLED(CONFIG_PREEMPT_RT))
 444                 preempt_disable();
 445
 446         v = __this_cpu_inc_return(*p);
 447         t = __this_cpu_read(pcp->stat_threshold);
 448         if (unlikely(v > t)) {
 449                 s8 overstep = t >> 1;
 450
 451                 zone_page_state_add(v + overstep, zone, item);
 452                 __this_cpu_write(*p, -overstep);
 453         }
 454
 455         if (IS_ENABLED(CONFIG_PREEMPT_RT))
 456                 preempt_enable();
 457 }
 458
 459 void __inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 460 {
 461         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 462         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 463         s8 v, t;
 464
 465         VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
 466
 467         /* See __mod_node_page_state */
 468         if (IS_ENABLED(CONFIG_PREEMPT_RT))
 469                 preempt_disable();
 470
 471         v = __this_cpu_inc_return(*p);
 472         t = __this_cpu_read(pcp->stat_threshold);
 473         if (unlikely(v > t)) {
 474                 s8 overstep = t >> 1;
 475
 476                 node_page_state_add(v + overstep, pgdat, item);
 477                 __this_cpu_write(*p, -overstep);
 478         }
 479
 480         if (IS_ENABLED(CONFIG_PREEMPT_RT))
 481                 preempt_enable();
 482 }
 483
 484 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
 485 {
 486         __inc_zone_state(page_zone(page), item);
 487 }
 488 EXPORT_SYMBOL(__inc_zone_page_state);
 489
 490 void __inc_node_page_state(struct page *page, enum node_stat_item item)
 491 {
 492         __inc_node_state(page_pgdat(page), item);
 493 }
 494 EXPORT_SYMBOL(__inc_node_page_state);
 495
 496 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
 497 {
 498         struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
 499         s8 __percpu *p = pcp->vm_stat_diff + item;
 500         s8 v, t;
 501
 502         /* See __mod_node_page_state */
 503         if (IS_ENABLED(CONFIG_PREEMPT_RT))
 504                 preempt_disable();
 505
 506         v = __this_cpu_dec_return(*p);
 507         t = __this_cpu_read(pcp->stat_threshold);
 508         if (unlikely(v < - t)) {
 509                 s8 overstep = t >> 1;
 510
 511                 zone_page_state_add(v - overstep, zone, item);
 512                 __this_cpu_write(*p, overstep);
 513         }
 514
 515         if (IS_ENABLED(CONFIG_PREEMPT_RT))
 516                 preempt_enable();
 517 }
 518
 519 void __dec_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 520 {
 521         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 522         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 523         s8 v, t;
 524
 525         VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
 526
 527         /* See __mod_node_page_state */
 528         if (IS_ENABLED(CONFIG_PREEMPT_RT))
 529                 preempt_disable();
 530
 531         v = __this_cpu_dec_return(*p);
 532         t = __this_cpu_read(pcp->stat_threshold);
 533         if (unlikely(v < - t)) {
 534                 s8 overstep = t >> 1;
 535
 536                 node_page_state_add(v - overstep, pgdat, item);
 537                 __this_cpu_write(*p, overstep);
 538         }
 539
 540         if (IS_ENABLED(CONFIG_PREEMPT_RT))
 541                 preempt_enable();
 542 }
 543
 544 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
 545 {
 546         __dec_zone_state(page_zone(page), item);
 547 }
 548 EXPORT_SYMBOL(__dec_zone_page_state);
 549
 550 void __dec_node_page_state(struct page *page, enum node_stat_item item)
 551 {
 552         __dec_node_state(page_pgdat(page), item);
 553 }
 554 EXPORT_SYMBOL(__dec_node_page_state);
 555
 556 #ifdef CONFIG_HAVE_CMPXCHG_LOCAL
 557 /*
 558  * If we have cmpxchg_local support then we do not need to incur the overhead
 559  * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
 560  *
 561  * mod_state() modifies the zone counter state through atomic per cpu
 562  * operations.
 563  *
 564  * Overstep mode specifies how overstep should handled:
 565  *     0       No overstepping
 566  *     1       Overstepping half of threshold
 567  *     -1      Overstepping minus half of threshold
 568 */
 569 static inline void mod_zone_state(struct zone *zone,
 570        enum zone_stat_item item, long delta, int overstep_mode)
 571 {
 572         struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
 573         s8 __percpu *p = pcp->vm_stat_diff + item;
 574         long o, n, t, z;
 575
 576         do {
 577                 z = 0;  /* overflow to zone counters */
 578
 579                 /*
 580                  * The fetching of the stat_threshold is racy. We may apply
 581                  * a counter threshold to the wrong the cpu if we get
 582                  * rescheduled while executing here. However, the next
 583                  * counter update will apply the threshold again and
 584                  * therefore bring the counter under the threshold again.
 585                  *
 586                  * Most of the time the thresholds are the same anyways
 587                  * for all cpus in a zone.
 588                  */
 589                 t = this_cpu_read(pcp->stat_threshold);
 590
 591                 o = this_cpu_read(*p);
 592                 n = delta + o;
 593
 594                 if (abs(n) > t) {
 595                         int os = overstep_mode * (t >> 1) ;
 596
 597                         /* Overflow must be added to zone counters */
 598                         z = n + os;
 599                         n = -os;
 600                 }
 601         } while (this_cpu_cmpxchg(*p, o, n) != o);
 602
 603         if (z)
 604                 zone_page_state_add(z, zone, item);
 605 }
 606
 607 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 608                          long delta)
 609 {
 610         mod_zone_state(zone, item, delta, 0);
 611 }
 612 EXPORT_SYMBOL(mod_zone_page_state);
 613
 614 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 615 {
 616         mod_zone_state(page_zone(page), item, 1, 1);
 617 }
 618 EXPORT_SYMBOL(inc_zone_page_state);
 619
 620 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 621 {
 622         mod_zone_state(page_zone(page), item, -1, -1);
 623 }
 624 EXPORT_SYMBOL(dec_zone_page_state);
 625
 626 static inline void mod_node_state(struct pglist_data *pgdat,
 627        enum node_stat_item item, int delta, int overstep_mode)
 628 {
 629         struct per_cpu_nodestat __percpu *pcp = pgdat->per_cpu_nodestats;
 630         s8 __percpu *p = pcp->vm_node_stat_diff + item;
 631         long o, n, t, z;
 632
 633         if (vmstat_item_in_bytes(item)) {
 634                 /*
 635                  * Only cgroups use subpage accounting right now; at
 636                  * the global level, these items still change in
 637                  * multiples of whole pages. Store them as pages
 638                  * internally to keep the per-cpu counters compact.
 639                  */
 640                 VM_WARN_ON_ONCE(delta & (PAGE_SIZE - 1));
 641                 delta >>= PAGE_SHIFT;
 642         }
 643
 644         do {
 645                 z = 0;  /* overflow to node counters */
 646
 647                 /*
 648                  * The fetching of the stat_threshold is racy. We may apply
 649                  * a counter threshold to the wrong the cpu if we get
 650                  * rescheduled while executing here. However, the next
 651                  * counter update will apply the threshold again and
 652                  * therefore bring the counter under the threshold again.
 653                  *
 654                  * Most of the time the thresholds are the same anyways
 655                  * for all cpus in a node.
 656                  */
 657                 t = this_cpu_read(pcp->stat_threshold);
 658
 659                 o = this_cpu_read(*p);
 660                 n = delta + o;
 661
 662                 if (abs(n) > t) {
 663                         int os = overstep_mode * (t >> 1) ;
 664
 665                         /* Overflow must be added to node counters */
 666                         z = n + os;
 667                         n = -os;
 668                 }
 669         } while (this_cpu_cmpxchg(*p, o, n) != o);
 670
 671         if (z)
 672                 node_page_state_add(z, pgdat, item);
 673 }
 674
 675 void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
 676                                         long delta)
 677 {
 678         mod_node_state(pgdat, item, delta, 0);
 679 }
 680 EXPORT_SYMBOL(mod_node_page_state);
 681
 682 void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 683 {
 684         mod_node_state(pgdat, item, 1, 1);
 685 }
 686
 687 void inc_node_page_state(struct page *page, enum node_stat_item item)
 688 {
 689         mod_node_state(page_pgdat(page), item, 1, 1);
 690 }
 691 EXPORT_SYMBOL(inc_node_page_state);
 692
 693 void dec_node_page_state(struct page *page, enum node_stat_item item)
 694 {
 695         mod_node_state(page_pgdat(page), item, -1, -1);
 696 }
 697 EXPORT_SYMBOL(dec_node_page_state);
 698 #else
 699 /*
 700  * Use interrupt disable to serialize counter updates
 701  */
 702 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 703                          long delta)
 704 {
 705         unsigned long flags;
 706
 707         local_irq_save(flags);
 708         __mod_zone_page_state(zone, item, delta);
 709         local_irq_restore(flags);
 710 }
 711 EXPORT_SYMBOL(mod_zone_page_state);
 712
 713 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 714 {
 715         unsigned long flags;
 716         struct zone *zone;
 717
 718         zone = page_zone(page);
 719         local_irq_save(flags);
 720         __inc_zone_state(zone, item);
 721         local_irq_restore(flags);
 722 }
 723 EXPORT_SYMBOL(inc_zone_page_state);
 724
 725 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 726 {
 727         unsigned long flags;
 728
 729         local_irq_save(flags);
 730         __dec_zone_page_state(page, item);
 731         local_irq_restore(flags);
 732 }
 733 EXPORT_SYMBOL(dec_zone_page_state);
 734
 735 void inc_node_state(struct pglist_data *pgdat, enum node_stat_item item)
 736 {
 737         unsigned long flags;
 738
 739         local_irq_save(flags);
 740         __inc_node_state(pgdat, item);
 741         local_irq_restore(flags);
 742 }
 743 EXPORT_SYMBOL(inc_node_state);
 744
 745 void mod_node_page_state(struct pglist_data *pgdat, enum node_stat_item item,
 746                                         long delta)
 747 {
 748         unsigned long flags;
 749
 750         local_irq_save(flags);
 751         __mod_node_page_state(pgdat, item, delta);
 752         local_irq_restore(flags);
 753 }
 754 EXPORT_SYMBOL(mod_node_page_state);
 755
 756 void inc_node_page_state(struct page *page, enum node_stat_item item)
 757 {
 758         unsigned long flags;
 759         struct pglist_data *pgdat;
 760
 761         pgdat = page_pgdat(page);
 762         local_irq_save(flags);
 763         __inc_node_state(pgdat, item);
 764         local_irq_restore(flags);
 765 }
 766 EXPORT_SYMBOL(inc_node_page_state);
 767
 768 void dec_node_page_state(struct page *page, enum node_stat_item item)
 769 {
 770         unsigned long flags;
 771
 772         local_irq_save(flags);
 773         __dec_node_page_state(page, item);
 774         local_irq_restore(flags);
 775 }
 776 EXPORT_SYMBOL(dec_node_page_state);
 777 #endif
 778
 779 /*
 780  * Fold a differential into the global counters.
 781  * Returns the number of counters updated.
 782  */
 783 static int fold_diff(int *zone_diff, int *node_diff)
 784 {
 785         int i;
 786         int changes = 0;
 787
 788         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 789                 if (zone_diff[i]) {
 790                         atomic_long_add(zone_diff[i], &vm_zone_stat[i]);
 791                         changes++;
 792         }
 793
 794         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 795                 if (node_diff[i]) {
 796                         atomic_long_add(node_diff[i], &vm_node_stat[i]);
 797                         changes++;
 798         }
 799         return changes;
 800 }
 801
 802 /*
 803  * Update the zone counters for the current cpu.
 804  *
 805  * Note that refresh_cpu_vm_stats strives to only access
 806  * node local memory. The per cpu pagesets on remote zones are placed
 807  * in the memory local to the processor using that pageset. So the
 808  * loop over all zones will access a series of cachelines local to
 809  * the processor.
 810  *
 811  * The call to zone_page_state_add updates the cachelines with the
 812  * statistics in the remote zone struct as well as the global cachelines
 813  * with the global counters. These could cause remote node cache line
 814  * bouncing and will have to be only done when necessary.
 815  *
 816  * The function returns the number of global counters updated.
 817  */
 818 static int refresh_cpu_vm_stats(bool do_pagesets)
 819 {
 820         struct pglist_data *pgdat;
 821         struct zone *zone;
 822         int i;
 823         int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 824         int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
 825         int changes = 0;
 826
 827         for_each_populated_zone(zone) {
 828                 struct per_cpu_zonestat __percpu *pzstats = zone->per_cpu_zonestats;
 829 #ifdef CONFIG_NUMA
 830                 struct per_cpu_pages __percpu *pcp = zone->per_cpu_pageset;
 831 #endif
 832
 833                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
 834                         int v;
 835
 836                         v = this_cpu_xchg(pzstats->vm_stat_diff[i], 0);
 837                         if (v) {
 838
 839                                 atomic_long_add(v, &zone->vm_stat[i]);
 840                                 global_zone_diff[i] += v;
 841 #ifdef CONFIG_NUMA
 842                                 /* 3 seconds idle till flush */
 843                                 __this_cpu_write(pcp->expire, 3);
 844 #endif
 845                         }
 846                 }
 847 #ifdef CONFIG_NUMA
 848
 849                 if (do_pagesets) {
 850                         cond_resched();
 851                         /*
 852                          * Deal with draining the remote pageset of this
 853                          * processor
 854                          *
 855                          * Check if there are pages remaining in this pageset
 856                          * if not then there is nothing to expire.
 857                          */
 858                         if (!__this_cpu_read(pcp->expire) ||
 859                                !__this_cpu_read(pcp->count))
 860                                 continue;
 861
 862                         /*
 863                          * We never drain zones local to this processor.
 864                          */
 865                         if (zone_to_nid(zone) == numa_node_id()) {
 866                                 __this_cpu_write(pcp->expire, 0);
 867                                 continue;
 868                         }
 869
 870                         if (__this_cpu_dec_return(pcp->expire))
 871                                 continue;
 872
 873                         if (__this_cpu_read(pcp->count)) {
 874                                 drain_zone_pages(zone, this_cpu_ptr(pcp));
 875                                 changes++;
 876                         }
 877                 }
 878 #endif
 879         }
 880
 881         for_each_online_pgdat(pgdat) {
 882                 struct per_cpu_nodestat __percpu *p = pgdat->per_cpu_nodestats;
 883
 884                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
 885                         int v;
 886
 887                         v = this_cpu_xchg(p->vm_node_stat_diff[i], 0);
 888                         if (v) {
 889                                 atomic_long_add(v, &pgdat->vm_stat[i]);
 890                                 global_node_diff[i] += v;
 891                         }
 892                 }
 893         }
 894
 895         changes += fold_diff(global_zone_diff, global_node_diff);
 896         return changes;
 897 }
 898
 899 /*
 900  * Fold the data for an offline cpu into the global array.
 901  * There cannot be any access by the offline cpu and therefore
 902  * synchronization is simplified.
 903  */
 904 void cpu_vm_stats_fold(int cpu)
 905 {
 906         struct pglist_data *pgdat;
 907         struct zone *zone;
 908         int i;
 909         int global_zone_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 910         int global_node_diff[NR_VM_NODE_STAT_ITEMS] = { 0, };
 911
 912         for_each_populated_zone(zone) {
 913                 struct per_cpu_zonestat *pzstats;
 914
 915                 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
 916
 917                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
 918                         if (pzstats->vm_stat_diff[i]) {
 919                                 int v;
 920
 921                                 v = pzstats->vm_stat_diff[i];
 922                                 pzstats->vm_stat_diff[i] = 0;
 923                                 atomic_long_add(v, &zone->vm_stat[i]);
 924                                 global_zone_diff[i] += v;
 925                         }
 926                 }
 927 #ifdef CONFIG_NUMA
 928                 for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++) {
 929                         if (pzstats->vm_numa_event[i]) {
 930                                 unsigned long v;
 931
 932                                 v = pzstats->vm_numa_event[i];
 933                                 pzstats->vm_numa_event[i] = 0;
 934                                 zone_numa_event_add(v, zone, i);
 935                         }
 936                 }
 937 #endif
 938         }
 939
 940         for_each_online_pgdat(pgdat) {
 941                 struct per_cpu_nodestat *p;
 942
 943                 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
 944
 945                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
 946                         if (p->vm_node_stat_diff[i]) {
 947                                 int v;
 948
 949                                 v = p->vm_node_stat_diff[i];
 950                                 p->vm_node_stat_diff[i] = 0;
 951                                 atomic_long_add(v, &pgdat->vm_stat[i]);
 952                                 global_node_diff[i] += v;
 953                         }
 954         }
 955
 956         fold_diff(global_zone_diff, global_node_diff);
 957 }
 958
 959 /*
 960  * this is only called if !populated_zone(zone), which implies no other users of
 961  * pset->vm_stat_diff[] exist.
 962  */
 963 void drain_zonestat(struct zone *zone, struct per_cpu_zonestat *pzstats)
 964 {
 965         unsigned long v;
 966         int i;
 967
 968         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
 969                 if (pzstats->vm_stat_diff[i]) {
 970                         v = pzstats->vm_stat_diff[i];
 971                         pzstats->vm_stat_diff[i] = 0;
 972                         zone_page_state_add(v, zone, i);
 973                 }
 974         }
 975
 976 #ifdef CONFIG_NUMA
 977         for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++) {
 978                 if (pzstats->vm_numa_event[i]) {
 979                         v = pzstats->vm_numa_event[i];
 980                         pzstats->vm_numa_event[i] = 0;
 981                         zone_numa_event_add(v, zone, i);
 982                 }
 983         }
 984 #endif
 985 }
 986 #endif
 987
 988 #ifdef CONFIG_NUMA
 989 /*
 990  * Determine the per node value of a stat item. This function
 991  * is called frequently in a NUMA machine, so try to be as
 992  * frugal as possible.
 993  */
 994 unsigned long sum_zone_node_page_state(int node,
 995                                  enum zone_stat_item item)
 996 {
 997         struct zone *zones = NODE_DATA(node)->node_zones;
 998         int i;
 999         unsigned long count = 0;
1000
1001         for (i = 0; i < MAX_NR_ZONES; i++)
1002                 count += zone_page_state(zones + i, item);
1003
1004         return count;
1005 }
1006
1007 /* Determine the per node value of a numa stat item. */
1008 unsigned long sum_zone_numa_event_state(int node,
1009                                  enum numa_stat_item item)
1010 {
1011         struct zone *zones = NODE_DATA(node)->node_zones;
1012         unsigned long count = 0;
1013         int i;
1014
1015         for (i = 0; i < MAX_NR_ZONES; i++)
1016                 count += zone_numa_event_state(zones + i, item);
1017
1018         return count;
1019 }
1020
1021 /*
1022  * Determine the per node value of a stat item.
1023  */
1024 unsigned long node_page_state_pages(struct pglist_data *pgdat,
1025                                     enum node_stat_item item)
1026 {
1027         long x = atomic_long_read(&pgdat->vm_stat[item]);
1028 #ifdef CONFIG_SMP
1029         if (x < 0)
1030                 x = 0;
1031 #endif
1032         return x;
1033 }
1034
1035 unsigned long node_page_state(struct pglist_data *pgdat,
1036                               enum node_stat_item item)
1037 {
1038         VM_WARN_ON_ONCE(vmstat_item_in_bytes(item));
1039
1040         return node_page_state_pages(pgdat, item);
1041 }
1042 #endif
1043
1044 #ifdef CONFIG_COMPACTION
1045
1046 struct contig_page_info {
1047         unsigned long free_pages;
1048         unsigned long free_blocks_total;
1049         unsigned long free_blocks_suitable;
1050 };
1051
1052 /*
1053  * Calculate the number of free pages in a zone, how many contiguous
1054  * pages are free and how many are large enough to satisfy an allocation of
1055  * the target size. Note that this function makes no attempt to estimate
1056  * how many suitable free blocks there *might* be if MOVABLE pages were
1057  * migrated. Calculating that is possible, but expensive and can be
1058  * figured out from userspace
1059  */
1060 static void fill_contig_page_info(struct zone *zone,
1061                                 unsigned int suitable_order,
1062                                 struct contig_page_info *info)
1063 {
1064         unsigned int order;
1065
1066         info->free_pages = 0;
1067         info->free_blocks_total = 0;
1068         info->free_blocks_suitable = 0;
1069
1070         for (order = 0; order < MAX_ORDER; order++) {
1071                 unsigned long blocks;
1072
1073                 /*
1074                  * Count number of free blocks.
1075                  *
1076                  * Access to nr_free is lockless as nr_free is used only for
1077                  * diagnostic purposes. Use data_race to avoid KCSAN warning.
1078                  */
1079                 blocks = data_race(zone->free_area[order].nr_free);
1080                 info->free_blocks_total += blocks;
1081
1082                 /* Count free base pages */
1083                 info->free_pages += blocks << order;
1084
1085                 /* Count the suitable free blocks */
1086                 if (order >= suitable_order)
1087                         info->free_blocks_suitable += blocks <<
1088                                                 (order - suitable_order);
1089         }
1090 }
1091
1092 /*
1093  * A fragmentation index only makes sense if an allocation of a requested
1094  * size would fail. If that is true, the fragmentation index indicates
1095  * whether external fragmentation or a lack of memory was the problem.
1096  * The value can be used to determine if page reclaim or compaction
1097  * should be used
1098  */
1099 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
1100 {
1101         unsigned long requested = 1UL << order;
1102
1103         if (WARN_ON_ONCE(order >= MAX_ORDER))
1104                 return 0;
1105
1106         if (!info->free_blocks_total)
1107                 return 0;
1108
1109         /* Fragmentation index only makes sense when a request would fail */
1110         if (info->free_blocks_suitable)
1111                 return -1000;
1112
1113         /*
1114          * Index is between 0 and 1 so return within 3 decimal places
1115          *
1116          * 0 => allocation would fail due to lack of memory
1117          * 1 => allocation would fail due to fragmentation
1118          */
1119         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
1120 }
1121
1122 /*
1123  * Calculates external fragmentation within a zone wrt the given order.
1124  * It is defined as the percentage of pages found in blocks of size
1125  * less than 1 << order. It returns values in range [0, 100].
1126  */
1127 unsigned int extfrag_for_order(struct zone *zone, unsigned int order)
1128 {
1129         struct contig_page_info info;
1130
1131         fill_contig_page_info(zone, order, &info);
1132         if (info.free_pages == 0)
1133                 return 0;
1134
1135         return div_u64((info.free_pages -
1136                         (info.free_blocks_suitable << order)) * 100,
1137                         info.free_pages);
1138 }
1139
1140 /* Same as __fragmentation index but allocs contig_page_info on stack */
1141 int fragmentation_index(struct zone *zone, unsigned int order)
1142 {
1143         struct contig_page_info info;
1144
1145         fill_contig_page_info(zone, order, &info);
1146         return __fragmentation_index(order, &info);
1147 }
1148 #endif
1149
1150 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || \
1151     defined(CONFIG_NUMA) || defined(CONFIG_MEMCG)
1152 #ifdef CONFIG_ZONE_DMA
1153 #define TEXT_FOR_DMA(xx) xx "_dma",
1154 #else
1155 #define TEXT_FOR_DMA(xx)
1156 #endif
1157
1158 #ifdef CONFIG_ZONE_DMA32
1159 #define TEXT_FOR_DMA32(xx) xx "_dma32",
1160 #else
1161 #define TEXT_FOR_DMA32(xx)
1162 #endif
1163
1164 #ifdef CONFIG_HIGHMEM
1165 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
1166 #else
1167 #define TEXT_FOR_HIGHMEM(xx)
1168 #endif
1169
1170 #ifdef CONFIG_ZONE_DEVICE
1171 #define TEXT_FOR_DEVICE(xx) xx "_device",
1172 #else
1173 #define TEXT_FOR_DEVICE(xx)
1174 #endif
1175
1176 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
1177                                         TEXT_FOR_HIGHMEM(xx) xx "_movable", \
1178                                         TEXT_FOR_DEVICE(xx)
1179
1180 const char * const vmstat_text[] = {
1181         /* enum zone_stat_item counters */
1182         "nr_free_pages",
1183         "nr_zone_inactive_anon",
1184         "nr_zone_active_anon",
1185         "nr_zone_inactive_file",
1186         "nr_zone_active_file",
1187         "nr_zone_unevictable",
1188         "nr_zone_write_pending",
1189         "nr_mlock",
1190         "nr_bounce",
1191 #if IS_ENABLED(CONFIG_ZSMALLOC)
1192         "nr_zspages",
1193 #endif
1194         "nr_free_cma",
1195
1196         /* enum numa_stat_item counters */
1197 #ifdef CONFIG_NUMA
1198         "numa_hit",
1199         "numa_miss",
1200         "numa_foreign",
1201         "numa_interleave",
1202         "numa_local",
1203         "numa_other",
1204 #endif
1205
1206         /* enum node_stat_item counters */
1207         "nr_inactive_anon",
1208         "nr_active_anon",
1209         "nr_inactive_file",
1210         "nr_active_file",
1211         "nr_unevictable",
1212         "nr_slab_reclaimable",
1213         "nr_slab_unreclaimable",
1214         "nr_isolated_anon",
1215         "nr_isolated_file",
1216         "workingset_nodes",
1217         "workingset_refault_anon",
1218         "workingset_refault_file",
1219         "workingset_activate_anon",
1220         "workingset_activate_file",
1221         "workingset_restore_anon",
1222         "workingset_restore_file",
1223         "workingset_nodereclaim",
1224         "nr_anon_pages",
1225         "nr_mapped",
1226         "nr_file_pages",
1227         "nr_dirty",
1228         "nr_writeback",
1229         "nr_writeback_temp",
1230         "nr_shmem",
1231         "nr_shmem_hugepages",
1232         "nr_shmem_pmdmapped",
1233         "nr_file_hugepages",
1234         "nr_file_pmdmapped",
1235         "nr_anon_transparent_hugepages",
1236         "nr_vmscan_write",
1237         "nr_vmscan_immediate_reclaim",
1238         "nr_dirtied",
1239         "nr_written",
1240         "nr_throttled_written",
1241         "nr_kernel_misc_reclaimable",
1242         "nr_foll_pin_acquired",
1243         "nr_foll_pin_released",
1244         "nr_kernel_stack",
1245 #if IS_ENABLED(CONFIG_SHADOW_CALL_STACK)
1246         "nr_shadow_call_stack",
1247 #endif
1248         "nr_page_table_pages",
1249 #ifdef CONFIG_SWAP
1250         "nr_swapcached",
1251 #endif
1252 #ifdef CONFIG_NUMA_BALANCING
1253         "pgpromote_success",
1254         "pgpromote_candidate",
1255 #endif
1256
1257         /* enum writeback_stat_item counters */
1258         "nr_dirty_threshold",
1259         "nr_dirty_background_threshold",
1260
1261 #if defined(CONFIG_VM_EVENT_COUNTERS) || defined(CONFIG_MEMCG)
1262         /* enum vm_event_item counters */
1263         "pgpgin",
1264         "pgpgout",
1265         "pswpin",
1266         "pswpout",
1267
1268         TEXTS_FOR_ZONES("pgalloc")
1269         TEXTS_FOR_ZONES("allocstall")
1270         TEXTS_FOR_ZONES("pgskip")
1271
1272         "pgfree",
1273         "pgactivate",
1274         "pgdeactivate",
1275         "pglazyfree",
1276
1277         "pgfault",
1278         "pgmajfault",
1279         "pglazyfreed",
1280
1281         "pgrefill",
1282         "pgreuse",
1283         "pgsteal_kswapd",
1284         "pgsteal_direct",
1285         "pgdemote_kswapd",
1286         "pgdemote_direct",
1287         "pgscan_kswapd",
1288         "pgscan_direct",
1289         "pgscan_direct_throttle",
1290         "pgscan_anon",
1291         "pgscan_file",
1292         "pgsteal_anon",
1293         "pgsteal_file",
1294
1295 #ifdef CONFIG_NUMA
1296         "zone_reclaim_failed",
1297 #endif
1298         "pginodesteal",
1299         "slabs_scanned",
1300         "kswapd_inodesteal",
1301         "kswapd_low_wmark_hit_quickly",
1302         "kswapd_high_wmark_hit_quickly",
1303         "pageoutrun",
1304
1305         "pgrotated",
1306
1307         "drop_pagecache",
1308         "drop_slab",
1309         "oom_kill",
1310
1311 #ifdef CONFIG_NUMA_BALANCING
1312         "numa_pte_updates",
1313         "numa_huge_pte_updates",
1314         "numa_hint_faults",
1315         "numa_hint_faults_local",
1316         "numa_pages_migrated",
1317 #endif
1318 #ifdef CONFIG_MIGRATION
1319         "pgmigrate_success",
1320         "pgmigrate_fail",
1321         "thp_migration_success",
1322         "thp_migration_fail",
1323         "thp_migration_split",
1324 #endif
1325 #ifdef CONFIG_COMPACTION
1326         "compact_migrate_scanned",
1327         "compact_free_scanned",
1328         "compact_isolated",
1329         "compact_stall",
1330         "compact_fail",
1331         "compact_success",
1332         "compact_daemon_wake",
1333         "compact_daemon_migrate_scanned",
1334         "compact_daemon_free_scanned",
1335 #endif
1336
1337 #ifdef CONFIG_HUGETLB_PAGE
1338         "htlb_buddy_alloc_success",
1339         "htlb_buddy_alloc_fail",
1340 #endif
1341 #ifdef CONFIG_CMA
1342         "cma_alloc_success",
1343         "cma_alloc_fail",
1344 #endif
1345         "unevictable_pgs_culled",
1346         "unevictable_pgs_scanned",
1347         "unevictable_pgs_rescued",
1348         "unevictable_pgs_mlocked",
1349         "unevictable_pgs_munlocked",
1350         "unevictable_pgs_cleared",
1351         "unevictable_pgs_stranded",
1352
1353 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1354         "thp_fault_alloc",
1355         "thp_fault_fallback",
1356         "thp_fault_fallback_charge",
1357         "thp_collapse_alloc",
1358         "thp_collapse_alloc_failed",
1359         "thp_file_alloc",
1360         "thp_file_fallback",
1361         "thp_file_fallback_charge",
1362         "thp_file_mapped",
1363         "thp_split_page",
1364         "thp_split_page_failed",
1365         "thp_deferred_split_page",
1366         "thp_split_pmd",
1367         "thp_scan_exceed_none_pte",
1368         "thp_scan_exceed_swap_pte",
1369         "thp_scan_exceed_share_pte",
1370 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1371         "thp_split_pud",
1372 #endif
1373         "thp_zero_page_alloc",
1374         "thp_zero_page_alloc_failed",
1375         "thp_swpout",
1376         "thp_swpout_fallback",
1377 #endif
1378 #ifdef CONFIG_MEMORY_BALLOON
1379         "balloon_inflate",
1380         "balloon_deflate",
1381 #ifdef CONFIG_BALLOON_COMPACTION
1382         "balloon_migrate",
1383 #endif
1384 #endif /* CONFIG_MEMORY_BALLOON */
1385 #ifdef CONFIG_DEBUG_TLBFLUSH
1386         "nr_tlb_remote_flush",
1387         "nr_tlb_remote_flush_received",
1388         "nr_tlb_local_flush_all",
1389         "nr_tlb_local_flush_one",
1390 #endif /* CONFIG_DEBUG_TLBFLUSH */
1391
1392 #ifdef CONFIG_SWAP
1393         "swap_ra",
1394         "swap_ra_hit",
1395 #ifdef CONFIG_KSM
1396         "ksm_swpin_copy",
1397 #endif
1398 #endif
1399 #ifdef CONFIG_KSM
1400         "cow_ksm",
1401 #endif
1402 #ifdef CONFIG_ZSWAP
1403         "zswpin",
1404         "zswpout",
1405 #endif
1406 #ifdef CONFIG_X86
1407         "direct_map_level2_splits",
1408         "direct_map_level3_splits",
1409 #endif
1410 #endif /* CONFIG_VM_EVENT_COUNTERS || CONFIG_MEMCG */
1411 };
1412 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA || CONFIG_MEMCG */
1413
1414 #if (defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)) || \
1415      defined(CONFIG_PROC_FS)
1416 static void *frag_start(struct seq_file *m, loff_t *pos)
1417 {
1418         pg_data_t *pgdat;
1419         loff_t node = *pos;
1420
1421         for (pgdat = first_online_pgdat();
1422              pgdat && node;
1423              pgdat = next_online_pgdat(pgdat))
1424                 --node;
1425
1426         return pgdat;
1427 }
1428
1429 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1430 {
1431         pg_data_t *pgdat = (pg_data_t *)arg;
1432
1433         (*pos)++;
1434         return next_online_pgdat(pgdat);
1435 }
1436
1437 static void frag_stop(struct seq_file *m, void *arg)
1438 {
1439 }
1440
1441 /*
1442  * Walk zones in a node and print using a callback.
1443  * If @assert_populated is true, only use callback for zones that are populated.
1444  */
1445 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
1446                 bool assert_populated, bool nolock,
1447                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
1448 {
1449         struct zone *zone;
1450         struct zone *node_zones = pgdat->node_zones;
1451         unsigned long flags;
1452
1453         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1454                 if (assert_populated && !populated_zone(zone))
1455                         continue;
1456
1457                 if (!nolock)
1458                         spin_lock_irqsave(&zone->lock, flags);
1459                 print(m, pgdat, zone);
1460                 if (!nolock)
1461                         spin_unlock_irqrestore(&zone->lock, flags);
1462         }
1463 }
1464 #endif
1465
1466 #ifdef CONFIG_PROC_FS
1467 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
1468                                                 struct zone *zone)
1469 {
1470         int order;
1471
1472         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1473         for (order = 0; order < MAX_ORDER; ++order)
1474                 /*
1475                  * Access to nr_free is lockless as nr_free is used only for
1476                  * printing purposes. Use data_race to avoid KCSAN warning.
1477                  */
1478                 seq_printf(m, "%6lu ", data_race(zone->free_area[order].nr_free));
1479         seq_putc(m, '\n');
1480 }
1481
1482 /*
1483  * This walks the free areas for each zone.
1484  */
1485 static int frag_show(struct seq_file *m, void *arg)
1486 {
1487         pg_data_t *pgdat = (pg_data_t *)arg;
1488         walk_zones_in_node(m, pgdat, true, false, frag_show_print);
1489         return 0;
1490 }
1491
1492 static void pagetypeinfo_showfree_print(struct seq_file *m,
1493                                         pg_data_t *pgdat, struct zone *zone)
1494 {
1495         int order, mtype;
1496
1497         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
1498                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
1499                                         pgdat->node_id,
1500                                         zone->name,
1501                                         migratetype_names[mtype]);
1502                 for (order = 0; order < MAX_ORDER; ++order) {
1503                         unsigned long freecount = 0;
1504                         struct free_area *area;
1505                         struct list_head *curr;
1506                         bool overflow = false;
1507
1508                         area = &(zone->free_area[order]);
1509
1510                         list_for_each(curr, &area->free_list[mtype]) {
1511                                 /*
1512                                  * Cap the free_list iteration because it might
1513                                  * be really large and we are under a spinlock
1514                                  * so a long time spent here could trigger a
1515                                  * hard lockup detector. Anyway this is a
1516                                  * debugging tool so knowing there is a handful
1517                                  * of pages of this order should be more than
1518                                  * sufficient.
1519                                  */
1520                                 if (++freecount >= 100000) {
1521                                         overflow = true;
1522                                         break;
1523                                 }
1524                         }
1525                         seq_printf(m, "%s%6lu ", overflow ? ">" : "", freecount);
1526                         spin_unlock_irq(&zone->lock);
1527                         cond_resched();
1528                         spin_lock_irq(&zone->lock);
1529                 }
1530                 seq_putc(m, '\n');
1531         }
1532 }
1533
1534 /* Print out the free pages at each order for each migatetype */
1535 static void pagetypeinfo_showfree(struct seq_file *m, void *arg)
1536 {
1537         int order;
1538         pg_data_t *pgdat = (pg_data_t *)arg;
1539
1540         /* Print header */
1541         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
1542         for (order = 0; order < MAX_ORDER; ++order)
1543                 seq_printf(m, "%6d ", order);
1544         seq_putc(m, '\n');
1545
1546         walk_zones_in_node(m, pgdat, true, false, pagetypeinfo_showfree_print);
1547 }
1548
1549 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
1550                                         pg_data_t *pgdat, struct zone *zone)
1551 {
1552         int mtype;
1553         unsigned long pfn;
1554         unsigned long start_pfn = zone->zone_start_pfn;
1555         unsigned long end_pfn = zone_end_pfn(zone);
1556         unsigned long count[MIGRATE_TYPES] = { 0, };
1557
1558         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
1559                 struct page *page;
1560
1561                 page = pfn_to_online_page(pfn);
1562                 if (!page)
1563                         continue;
1564
1565                 if (page_zone(page) != zone)
1566                         continue;
1567
1568                 mtype = get_pageblock_migratetype(page);
1569
1570                 if (mtype < MIGRATE_TYPES)
1571                         count[mtype]++;
1572         }
1573
1574         /* Print counts */
1575         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1576         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1577                 seq_printf(m, "%12lu ", count[mtype]);
1578         seq_putc(m, '\n');
1579 }
1580
1581 /* Print out the number of pageblocks for each migratetype */
1582 static void pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1583 {
1584         int mtype;
1585         pg_data_t *pgdat = (pg_data_t *)arg;
1586
1587         seq_printf(m, "\n%-23s", "Number of blocks type ");
1588         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1589                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1590         seq_putc(m, '\n');
1591         walk_zones_in_node(m, pgdat, true, false,
1592                 pagetypeinfo_showblockcount_print);
1593 }
1594
1595 /*
1596  * Print out the number of pageblocks for each migratetype that contain pages
1597  * of other types. This gives an indication of how well fallbacks are being
1598  * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1599  * to determine what is going on
1600  */
1601 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1602 {
1603 #ifdef CONFIG_PAGE_OWNER
1604         int mtype;
1605
1606         if (!static_branch_unlikely(&page_owner_inited))
1607                 return;
1608
1609         drain_all_pages(NULL);
1610
1611         seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1612         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1613                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1614         seq_putc(m, '\n');
1615
1616         walk_zones_in_node(m, pgdat, true, true,
1617                 pagetypeinfo_showmixedcount_print);
1618 #endif /* CONFIG_PAGE_OWNER */
1619 }
1620
1621 /*
1622  * This prints out statistics in relation to grouping pages by mobility.
1623  * It is expensive to collect so do not constantly read the file.
1624  */
1625 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1626 {
1627         pg_data_t *pgdat = (pg_data_t *)arg;
1628
1629         /* check memoryless node */
1630         if (!node_state(pgdat->node_id, N_MEMORY))
1631                 return 0;
1632
1633         seq_printf(m, "Page block order: %d\n", pageblock_order);
1634         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
1635         seq_putc(m, '\n');
1636         pagetypeinfo_showfree(m, pgdat);
1637         pagetypeinfo_showblockcount(m, pgdat);
1638         pagetypeinfo_showmixedcount(m, pgdat);
1639
1640         return 0;
1641 }
1642
1643 static const struct seq_operations fragmentation_op = {
1644         .start  = frag_start,
1645         .next   = frag_next,
1646         .stop   = frag_stop,
1647         .show   = frag_show,
1648 };
1649
1650 static const struct seq_operations pagetypeinfo_op = {
1651         .start  = frag_start,
1652         .next   = frag_next,
1653         .stop   = frag_stop,
1654         .show   = pagetypeinfo_show,
1655 };
1656
1657 static bool is_zone_first_populated(pg_data_t *pgdat, struct zone *zone)
1658 {
1659         int zid;
1660
1661         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1662                 struct zone *compare = &pgdat->node_zones[zid];
1663
1664                 if (populated_zone(compare))
1665                         return zone == compare;
1666         }
1667
1668         return false;
1669 }
1670
1671 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1672                                                         struct zone *zone)
1673 {
1674         int i;
1675         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1676         if (is_zone_first_populated(pgdat, zone)) {
1677                 seq_printf(m, "\n  per-node stats");
1678                 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1679                         unsigned long pages = node_page_state_pages(pgdat, i);
1680
1681                         if (vmstat_item_print_in_thp(i))
1682                                 pages /= HPAGE_PMD_NR;
1683                         seq_printf(m, "\n      %-12s %lu", node_stat_name(i),
1684                                    pages);
1685                 }
1686         }
1687         seq_printf(m,
1688                    "\n  pages free     %lu"
1689                    "\n        boost    %lu"
1690                    "\n        min      %lu"
1691                    "\n        low      %lu"
1692                    "\n        high     %lu"
1693                    "\n        spanned  %lu"
1694                    "\n        present  %lu"
1695                    "\n        managed  %lu"
1696                    "\n        cma      %lu",
1697                    zone_page_state(zone, NR_FREE_PAGES),
1698                    zone->watermark_boost,
1699                    min_wmark_pages(zone),
1700                    low_wmark_pages(zone),
1701                    high_wmark_pages(zone),
1702                    zone->spanned_pages,
1703                    zone->present_pages,
1704                    zone_managed_pages(zone),
1705                    zone_cma_pages(zone));
1706
1707         seq_printf(m,
1708                    "\n        protection: (%ld",
1709                    zone->lowmem_reserve[0]);
1710         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1711                 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1712         seq_putc(m, ')');
1713
1714         /* If unpopulated, no other information is useful */
1715         if (!populated_zone(zone)) {
1716                 seq_putc(m, '\n');
1717                 return;
1718         }
1719
1720         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1721                 seq_printf(m, "\n      %-12s %lu", zone_stat_name(i),
1722                            zone_page_state(zone, i));
1723
1724 #ifdef CONFIG_NUMA
1725         for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++)
1726                 seq_printf(m, "\n      %-12s %lu", numa_stat_name(i),
1727                            zone_numa_event_state(zone, i));
1728 #endif
1729
1730         seq_printf(m, "\n  pagesets");
1731         for_each_online_cpu(i) {
1732                 struct per_cpu_pages *pcp;
1733                 struct per_cpu_zonestat __maybe_unused *pzstats;
1734
1735                 pcp = per_cpu_ptr(zone->per_cpu_pageset, i);
1736                 seq_printf(m,
1737                            "\n    cpu: %i"
1738                            "\n              count: %i"
1739                            "\n              high:  %i"
1740                            "\n              batch: %i",
1741                            i,
1742                            pcp->count,
1743                            pcp->high,
1744                            pcp->batch);
1745 #ifdef CONFIG_SMP
1746                 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, i);
1747                 seq_printf(m, "\n  vm stats threshold: %d",
1748                                 pzstats->stat_threshold);
1749 #endif
1750         }
1751         seq_printf(m,
1752                    "\n  node_unreclaimable:  %u"
1753                    "\n  start_pfn:           %lu",
1754                    pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES,
1755                    zone->zone_start_pfn);
1756         seq_putc(m, '\n');
1757 }
1758
1759 /*
1760  * Output information about zones in @pgdat.  All zones are printed regardless
1761  * of whether they are populated or not: lowmem_reserve_ratio operates on the
1762  * set of all zones and userspace would not be aware of such zones if they are
1763  * suppressed here (zoneinfo displays the effect of lowmem_reserve_ratio).
1764  */
1765 static int zoneinfo_show(struct seq_file *m, void *arg)
1766 {
1767         pg_data_t *pgdat = (pg_data_t *)arg;
1768         walk_zones_in_node(m, pgdat, false, false, zoneinfo_show_print);
1769         return 0;
1770 }
1771
1772 static const struct seq_operations zoneinfo_op = {
1773         .start  = frag_start, /* iterate over all zones. The same as in
1774                                * fragmentation. */
1775         .next   = frag_next,
1776         .stop   = frag_stop,
1777         .show   = zoneinfo_show,
1778 };
1779
1780 #define NR_VMSTAT_ITEMS (NR_VM_ZONE_STAT_ITEMS + \
1781                          NR_VM_NUMA_EVENT_ITEMS + \
1782                          NR_VM_NODE_STAT_ITEMS + \
1783                          NR_VM_WRITEBACK_STAT_ITEMS + \
1784                          (IS_ENABLED(CONFIG_VM_EVENT_COUNTERS) ? \
1785                           NR_VM_EVENT_ITEMS : 0))
1786
1787 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1788 {
1789         unsigned long *v;
1790         int i;
1791
1792         if (*pos >= NR_VMSTAT_ITEMS)
1793                 return NULL;
1794
1795         BUILD_BUG_ON(ARRAY_SIZE(vmstat_text) < NR_VMSTAT_ITEMS);
1796         fold_vm_numa_events();
1797         v = kmalloc_array(NR_VMSTAT_ITEMS, sizeof(unsigned long), GFP_KERNEL);
1798         m->private = v;
1799         if (!v)
1800                 return ERR_PTR(-ENOMEM);
1801         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1802                 v[i] = global_zone_page_state(i);
1803         v += NR_VM_ZONE_STAT_ITEMS;
1804
1805 #ifdef CONFIG_NUMA
1806         for (i = 0; i < NR_VM_NUMA_EVENT_ITEMS; i++)
1807                 v[i] = global_numa_event_state(i);
1808         v += NR_VM_NUMA_EVENT_ITEMS;
1809 #endif
1810
1811         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1812                 v[i] = global_node_page_state_pages(i);
1813                 if (vmstat_item_print_in_thp(i))
1814                         v[i] /= HPAGE_PMD_NR;
1815         }
1816         v += NR_VM_NODE_STAT_ITEMS;
1817
1818         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1819                             v + NR_DIRTY_THRESHOLD);
1820         v += NR_VM_WRITEBACK_STAT_ITEMS;
1821
1822 #ifdef CONFIG_VM_EVENT_COUNTERS
1823         all_vm_events(v);
1824         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1825         v[PGPGOUT] /= 2;
1826 #endif
1827         return (unsigned long *)m->private + *pos;
1828 }
1829
1830 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1831 {
1832         (*pos)++;
1833         if (*pos >= NR_VMSTAT_ITEMS)
1834                 return NULL;
1835         return (unsigned long *)m->private + *pos;
1836 }
1837
1838 static int vmstat_show(struct seq_file *m, void *arg)
1839 {
1840         unsigned long *l = arg;
1841         unsigned long off = l - (unsigned long *)m->private;
1842
1843         seq_puts(m, vmstat_text[off]);
1844         seq_put_decimal_ull(m, " ", *l);
1845         seq_putc(m, '\n');
1846
1847         if (off == NR_VMSTAT_ITEMS - 1) {
1848                 /*
1849                  * We've come to the end - add any deprecated counters to avoid
1850                  * breaking userspace which might depend on them being present.
1851                  */
1852                 seq_puts(m, "nr_unstable 0\n");
1853         }
1854         return 0;
1855 }
1856
1857 static void vmstat_stop(struct seq_file *m, void *arg)
1858 {
1859         kfree(m->private);
1860         m->private = NULL;
1861 }
1862
1863 static const struct seq_operations vmstat_op = {
1864         .start  = vmstat_start,
1865         .next   = vmstat_next,
1866         .stop   = vmstat_stop,
1867         .show   = vmstat_show,
1868 };
1869 #endif /* CONFIG_PROC_FS */
1870
1871 #ifdef CONFIG_SMP
1872 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1873 int sysctl_stat_interval __read_mostly = HZ;
1874
1875 #ifdef CONFIG_PROC_FS
1876 static void refresh_vm_stats(struct work_struct *work)
1877 {
1878         refresh_cpu_vm_stats(true);
1879 }
1880
1881 int vmstat_refresh(struct ctl_table *table, int write,
1882                    void *buffer, size_t *lenp, loff_t *ppos)
1883 {
1884         long val;
1885         int err;
1886         int i;
1887
1888         /*
1889          * The regular update, every sysctl_stat_interval, may come later
1890          * than expected: leaving a significant amount in per_cpu buckets.
1891          * This is particularly misleading when checking a quantity of HUGE
1892          * pages, immediately after running a test.  /proc/sys/vm/stat_refresh,
1893          * which can equally be echo'ed to or cat'ted from (by root),
1894          * can be used to update the stats just before reading them.
1895          *
1896          * Oh, and since global_zone_page_state() etc. are so careful to hide
1897          * transiently negative values, report an error here if any of
1898          * the stats is negative, so we know to go looking for imbalance.
1899          */
1900         err = schedule_on_each_cpu(refresh_vm_stats);
1901         if (err)
1902                 return err;
1903         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
1904                 /*
1905                  * Skip checking stats known to go negative occasionally.
1906                  */
1907                 switch (i) {
1908                 case NR_ZONE_WRITE_PENDING:
1909                 case NR_FREE_CMA_PAGES:
1910                         continue;
1911                 }
1912                 val = atomic_long_read(&vm_zone_stat[i]);
1913                 if (val < 0) {
1914                         pr_warn("%s: %s %ld\n",
1915                                 __func__, zone_stat_name(i), val);
1916                 }
1917         }
1918         for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
1919                 /*
1920                  * Skip checking stats known to go negative occasionally.
1921                  */
1922                 switch (i) {
1923                 case NR_WRITEBACK:
1924                         continue;
1925                 }
1926                 val = atomic_long_read(&vm_node_stat[i]);
1927                 if (val < 0) {
1928                         pr_warn("%s: %s %ld\n",
1929                                 __func__, node_stat_name(i), val);
1930                 }
1931         }
1932         if (write)
1933                 *ppos += *lenp;
1934         else
1935                 *lenp = 0;
1936         return 0;
1937 }
1938 #endif /* CONFIG_PROC_FS */
1939
1940 static void vmstat_update(struct work_struct *w)
1941 {
1942         if (refresh_cpu_vm_stats(true)) {
1943                 /*
1944                  * Counters were updated so we expect more updates
1945                  * to occur in the future. Keep on running the
1946                  * update worker thread.
1947                  */
1948                 queue_delayed_work_on(smp_processor_id(), mm_percpu_wq,
1949                                 this_cpu_ptr(&vmstat_work),
1950                                 round_jiffies_relative(sysctl_stat_interval));
1951         }
1952 }
1953
1954 /*
1955  * Check if the diffs for a certain cpu indicate that
1956  * an update is needed.
1957  */
1958 static bool need_update(int cpu)
1959 {
1960         pg_data_t *last_pgdat = NULL;
1961         struct zone *zone;
1962
1963         for_each_populated_zone(zone) {
1964                 struct per_cpu_zonestat *pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
1965                 struct per_cpu_nodestat *n;
1966
1967                 /*
1968                  * The fast way of checking if there are any vmstat diffs.
1969                  */
1970                 if (memchr_inv(pzstats->vm_stat_diff, 0, sizeof(pzstats->vm_stat_diff)))
1971                         return true;
1972
1973                 if (last_pgdat == zone->zone_pgdat)
1974                         continue;
1975                 last_pgdat = zone->zone_pgdat;
1976                 n = per_cpu_ptr(zone->zone_pgdat->per_cpu_nodestats, cpu);
1977                 if (memchr_inv(n->vm_node_stat_diff, 0, sizeof(n->vm_node_stat_diff)))
1978                         return true;
1979         }
1980         return false;
1981 }
1982
1983 /*
1984  * Switch off vmstat processing and then fold all the remaining differentials
1985  * until the diffs stay at zero. The function is used by NOHZ and can only be
1986  * invoked when tick processing is not active.
1987  */
1988 void quiet_vmstat(void)
1989 {
1990         if (system_state != SYSTEM_RUNNING)
1991                 return;
1992
1993         if (!delayed_work_pending(this_cpu_ptr(&vmstat_work)))
1994                 return;
1995
1996         if (!need_update(smp_processor_id()))
1997                 return;
1998
1999         /*
2000          * Just refresh counters and do not care about the pending delayed
2001          * vmstat_update. It doesn't fire that often to matter and canceling
2002          * it would be too expensive from this path.
2003          * vmstat_shepherd will take care about that for us.
2004          */
2005         refresh_cpu_vm_stats(false);
2006 }
2007
2008 /*
2009  * Shepherd worker thread that checks the
2010  * differentials of processors that have their worker
2011  * threads for vm statistics updates disabled because of
2012  * inactivity.
2013  */
2014 static void vmstat_shepherd(struct work_struct *w);
2015
2016 static DECLARE_DEFERRABLE_WORK(shepherd, vmstat_shepherd);
2017
2018 static void vmstat_shepherd(struct work_struct *w)
2019 {
2020         int cpu;
2021
2022         cpus_read_lock();
2023         /* Check processors whose vmstat worker threads have been disabled */
2024         for_each_online_cpu(cpu) {
2025                 struct delayed_work *dw = &per_cpu(vmstat_work, cpu);
2026
2027                 if (!delayed_work_pending(dw) && need_update(cpu))
2028                         queue_delayed_work_on(cpu, mm_percpu_wq, dw, 0);
2029
2030                 cond_resched();
2031         }
2032         cpus_read_unlock();
2033
2034         schedule_delayed_work(&shepherd,
2035                 round_jiffies_relative(sysctl_stat_interval));
2036 }
2037
2038 static void __init start_shepherd_timer(void)
2039 {
2040         int cpu;
2041
2042         for_each_possible_cpu(cpu)
2043                 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
2044                         vmstat_update);
2045
2046         schedule_delayed_work(&shepherd,
2047                 round_jiffies_relative(sysctl_stat_interval));
2048 }
2049
2050 static void __init init_cpu_node_state(void)
2051 {
2052         int node;
2053
2054         for_each_online_node(node) {
2055                 if (!cpumask_empty(cpumask_of_node(node)))
2056                         node_set_state(node, N_CPU);
2057         }
2058 }
2059
2060 static int vmstat_cpu_online(unsigned int cpu)
2061 {
2062         refresh_zone_stat_thresholds();
2063
2064         if (!node_state(cpu_to_node(cpu), N_CPU)) {
2065                 node_set_state(cpu_to_node(cpu), N_CPU);
2066         }
2067
2068         return 0;
2069 }
2070
2071 static int vmstat_cpu_down_prep(unsigned int cpu)
2072 {
2073         cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
2074         return 0;
2075 }
2076
2077 static int vmstat_cpu_dead(unsigned int cpu)
2078 {
2079         const struct cpumask *node_cpus;
2080         int node;
2081
2082         node = cpu_to_node(cpu);
2083
2084         refresh_zone_stat_thresholds();
2085         node_cpus = cpumask_of_node(node);
2086         if (!cpumask_empty(node_cpus))
2087                 return 0;
2088
2089         node_clear_state(node, N_CPU);
2090
2091         return 0;
2092 }
2093
2094 #endif
2095
2096 struct workqueue_struct *mm_percpu_wq;
2097
2098 void __init init_mm_internals(void)
2099 {
2100         int ret __maybe_unused;
2101
2102         mm_percpu_wq = alloc_workqueue("mm_percpu_wq", WQ_MEM_RECLAIM, 0);
2103
2104 #ifdef CONFIG_SMP
2105         ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
2106                                         NULL, vmstat_cpu_dead);
2107         if (ret < 0)
2108                 pr_err("vmstat: failed to register 'dead' hotplug state\n");
2109
2110         ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "mm/vmstat:online",
2111                                         vmstat_cpu_online,
2112                                         vmstat_cpu_down_prep);
2113         if (ret < 0)
2114                 pr_err("vmstat: failed to register 'online' hotplug state\n");
2115
2116         cpus_read_lock();
2117         init_cpu_node_state();
2118         cpus_read_unlock();
2119
2120         start_shepherd_timer();
2121 #endif
2122 #ifdef CONFIG_PROC_FS
2123         proc_create_seq("buddyinfo", 0444, NULL, &fragmentation_op);
2124         proc_create_seq("pagetypeinfo", 0400, NULL, &pagetypeinfo_op);
2125         proc_create_seq("vmstat", 0444, NULL, &vmstat_op);
2126         proc_create_seq("zoneinfo", 0444, NULL, &zoneinfo_op);
2127 #endif
2128 }
2129
2130 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
2131
2132 /*
2133  * Return an index indicating how much of the available free memory is
2134  * unusable for an allocation of the requested size.
2135  */
2136 static int unusable_free_index(unsigned int order,
2137                                 struct contig_page_info *info)
2138 {
2139         /* No free memory is interpreted as all free memory is unusable */
2140         if (info->free_pages == 0)
2141                 return 1000;
2142
2143         /*
2144          * Index should be a value between 0 and 1. Return a value to 3
2145          * decimal places.
2146          *
2147          * 0 => no fragmentation
2148          * 1 => high fragmentation
2149          */
2150         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
2151
2152 }
2153
2154 static void unusable_show_print(struct seq_file *m,
2155                                         pg_data_t *pgdat, struct zone *zone)
2156 {
2157         unsigned int order;
2158         int index;
2159         struct contig_page_info info;
2160
2161         seq_printf(m, "Node %d, zone %8s ",
2162                                 pgdat->node_id,
2163                                 zone->name);
2164         for (order = 0; order < MAX_ORDER; ++order) {
2165                 fill_contig_page_info(zone, order, &info);
2166                 index = unusable_free_index(order, &info);
2167                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
2168         }
2169
2170         seq_putc(m, '\n');
2171 }
2172
2173 /*
2174  * Display unusable free space index
2175  *
2176  * The unusable free space index measures how much of the available free
2177  * memory cannot be used to satisfy an allocation of a given size and is a
2178  * value between 0 and 1. The higher the value, the more of free memory is
2179  * unusable and by implication, the worse the external fragmentation is. This
2180  * can be expressed as a percentage by multiplying by 100.
2181  */
2182 static int unusable_show(struct seq_file *m, void *arg)
2183 {
2184         pg_data_t *pgdat = (pg_data_t *)arg;
2185
2186         /* check memoryless node */
2187         if (!node_state(pgdat->node_id, N_MEMORY))
2188                 return 0;
2189
2190         walk_zones_in_node(m, pgdat, true, false, unusable_show_print);
2191
2192         return 0;
2193 }
2194
2195 static const struct seq_operations unusable_sops = {
2196         .start  = frag_start,
2197         .next   = frag_next,
2198         .stop   = frag_stop,
2199         .show   = unusable_show,
2200 };
2201
2202 DEFINE_SEQ_ATTRIBUTE(unusable);
2203
2204 static void extfrag_show_print(struct seq_file *m,
2205                                         pg_data_t *pgdat, struct zone *zone)
2206 {
2207         unsigned int order;
2208         int index;
2209
2210         /* Alloc on stack as interrupts are disabled for zone walk */
2211         struct contig_page_info info;
2212
2213         seq_printf(m, "Node %d, zone %8s ",
2214                                 pgdat->node_id,
2215                                 zone->name);
2216         for (order = 0; order < MAX_ORDER; ++order) {
2217                 fill_contig_page_info(zone, order, &info);
2218                 index = __fragmentation_index(order, &info);
2219                 seq_printf(m, "%2d.%03d ", index / 1000, index % 1000);
2220         }
2221
2222         seq_putc(m, '\n');
2223 }
2224
2225 /*
2226  * Display fragmentation index for orders that allocations would fail for
2227  */
2228 static int extfrag_show(struct seq_file *m, void *arg)
2229 {
2230         pg_data_t *pgdat = (pg_data_t *)arg;
2231
2232         walk_zones_in_node(m, pgdat, true, false, extfrag_show_print);
2233
2234         return 0;
2235 }
2236
2237 static const struct seq_operations extfrag_sops = {
2238         .start  = frag_start,
2239         .next   = frag_next,
2240         .stop   = frag_stop,
2241         .show   = extfrag_show,
2242 };
2243
2244 DEFINE_SEQ_ATTRIBUTE(extfrag);
2245
2246 static int __init extfrag_debug_init(void)
2247 {
2248         struct dentry *extfrag_debug_root;
2249
2250         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
2251
2252         debugfs_create_file("unusable_index", 0444, extfrag_debug_root, NULL,
2253                             &unusable_fops);
2254
2255         debugfs_create_file("extfrag_index", 0444, extfrag_debug_root, NULL,
2256                             &extfrag_fops);
2257
2258         return 0;
2259 }
2260
2261 module_init(extfrag_debug_init);
2262 #endif