mm/vmstat.c

   1 /*
   2  *  linux/mm/vmstat.c
   3  *
   4  *  Manages VM statistics
   5  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   6  *
   7  *  zoned VM statistics
   8  *  Copyright (C) 2006 Silicon Graphics, Inc.,
   9  *              Christoph Lameter <christoph@lameter.com>
  10  *  Copyright (C) 2008-2014 Christoph Lameter
  11  */
  12 #include <linux/fs.h>
  13 #include <linux/mm.h>
  14 #include <linux/err.h>
  15 #include <linux/module.h>
  16 #include <linux/slab.h>
  17 #include <linux/cpu.h>
  18 #include <linux/cpumask.h>
  19 #include <linux/vmstat.h>
  20 #include <linux/sched.h>
  21 #include <linux/math64.h>
  22 #include <linux/writeback.h>
  23 #include <linux/compaction.h>
  24 #include <linux/mm_inline.h>
  25
  26 #include "internal.h"
  27
  28 #ifdef CONFIG_VM_EVENT_COUNTERS
  29 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
  30 EXPORT_PER_CPU_SYMBOL(vm_event_states);
  31
  32 static void sum_vm_events(unsigned long *ret)
  33 {
  34         int cpu;
  35         int i;
  36
  37         memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
  38
  39         for_each_online_cpu(cpu) {
  40                 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
  41
  42                 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
  43                         ret[i] += this->event[i];
  44         }
  45 }
  46
  47 /*
  48  * Accumulate the vm event counters across all CPUs.
  49  * The result is unavoidably approximate - it can change
  50  * during and after execution of this function.
  51 */
  52 void all_vm_events(unsigned long *ret)
  53 {
  54         get_online_cpus();
  55         sum_vm_events(ret);
  56         put_online_cpus();
  57 }
  58 EXPORT_SYMBOL_GPL(all_vm_events);
  59
  60 /*
  61  * Fold the foreign cpu events into our own.
  62  *
  63  * This is adding to the events on one processor
  64  * but keeps the global counts constant.
  65  */
  66 void vm_events_fold_cpu(int cpu)
  67 {
  68         struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
  69         int i;
  70
  71         for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
  72                 count_vm_events(i, fold_state->event[i]);
  73                 fold_state->event[i] = 0;
  74         }
  75 }
  76
  77 #endif /* CONFIG_VM_EVENT_COUNTERS */
  78
  79 /*
  80  * Manage combined zone based / global counters
  81  *
  82  * vm_stat contains the global counters
  83  */
  84 atomic_long_t vm_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
  85 EXPORT_SYMBOL(vm_stat);
  86
  87 #ifdef CONFIG_SMP
  88
  89 int calculate_pressure_threshold(struct zone *zone)
  90 {
  91         int threshold;
  92         int watermark_distance;
  93
  94         /*
  95          * As vmstats are not up to date, there is drift between the estimated
  96          * and real values. For high thresholds and a high number of CPUs, it
  97          * is possible for the min watermark to be breached while the estimated
  98          * value looks fine. The pressure threshold is a reduced value such
  99          * that even the maximum amount of drift will not accidentally breach
 100          * the min watermark
 101          */
 102         watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
 103         threshold = max(1, (int)(watermark_distance / num_online_cpus()));
 104
 105         /*
 106          * Maximum threshold is 125
 107          */
 108         threshold = min(125, threshold);
 109
 110         return threshold;
 111 }
 112
 113 int calculate_normal_threshold(struct zone *zone)
 114 {
 115         int threshold;
 116         int mem;        /* memory in 128 MB units */
 117
 118         /*
 119          * The threshold scales with the number of processors and the amount
 120          * of memory per zone. More memory means that we can defer updates for
 121          * longer, more processors could lead to more contention.
 122          * fls() is used to have a cheap way of logarithmic scaling.
 123          *
 124          * Some sample thresholds:
 125          *
 126          * Threshold    Processors      (fls)   Zonesize        fls(mem+1)
 127          * ------------------------------------------------------------------
 128          * 8            1               1       0.9-1 GB        4
 129          * 16           2               2       0.9-1 GB        4
 130          * 20           2               2       1-2 GB          5
 131          * 24           2               2       2-4 GB          6
 132          * 28           2               2       4-8 GB          7
 133          * 32           2               2       8-16 GB         8
 134          * 4            2               2       <128M           1
 135          * 30           4               3       2-4 GB          5
 136          * 48           4               3       8-16 GB         8
 137          * 32           8               4       1-2 GB          4
 138          * 32           8               4       0.9-1GB         4
 139          * 10           16              5       <128M           1
 140          * 40           16              5       900M            4
 141          * 70           64              7       2-4 GB          5
 142          * 84           64              7       4-8 GB          6
 143          * 108          512             9       4-8 GB          6
 144          * 125          1024            10      8-16 GB         8
 145          * 125          1024            10      16-32 GB        9
 146          */
 147
 148         mem = zone->managed_pages >> (27 - PAGE_SHIFT);
 149
 150         threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
 151
 152         /*
 153          * Maximum threshold is 125
 154          */
 155         threshold = min(125, threshold);
 156
 157         return threshold;
 158 }
 159
 160 /*
 161  * Refresh the thresholds for each zone.
 162  */
 163 void refresh_zone_stat_thresholds(void)
 164 {
 165         struct zone *zone;
 166         int cpu;
 167         int threshold;
 168
 169         for_each_populated_zone(zone) {
 170                 unsigned long max_drift, tolerate_drift;
 171
 172                 threshold = calculate_normal_threshold(zone);
 173
 174                 for_each_online_cpu(cpu)
 175                         per_cpu_ptr(zone->pageset, cpu)->stat_threshold
 176                                                         = threshold;
 177
 178                 /*
 179                  * Only set percpu_drift_mark if there is a danger that
 180                  * NR_FREE_PAGES reports the low watermark is ok when in fact
 181                  * the min watermark could be breached by an allocation
 182                  */
 183                 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
 184                 max_drift = num_online_cpus() * threshold;
 185                 if (max_drift > tolerate_drift)
 186                         zone->percpu_drift_mark = high_wmark_pages(zone) +
 187                                         max_drift;
 188         }
 189 }
 190
 191 void set_pgdat_percpu_threshold(pg_data_t *pgdat,
 192                                 int (*calculate_pressure)(struct zone *))
 193 {
 194         struct zone *zone;
 195         int cpu;
 196         int threshold;
 197         int i;
 198
 199         for (i = 0; i < pgdat->nr_zones; i++) {
 200                 zone = &pgdat->node_zones[i];
 201                 if (!zone->percpu_drift_mark)
 202                         continue;
 203
 204                 threshold = (*calculate_pressure)(zone);
 205                 for_each_online_cpu(cpu)
 206                         per_cpu_ptr(zone->pageset, cpu)->stat_threshold
 207                                                         = threshold;
 208         }
 209 }
 210
 211 /*
 212  * For use when we know that interrupts are disabled,
 213  * or when we know that preemption is disabled and that
 214  * particular counter cannot be updated from interrupt context.
 215  */
 216 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 217                                 int delta)
 218 {
 219         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 220         s8 __percpu *p = pcp->vm_stat_diff + item;
 221         long x;
 222         long t;
 223
 224         x = delta + __this_cpu_read(*p);
 225
 226         t = __this_cpu_read(pcp->stat_threshold);
 227
 228         if (unlikely(x > t || x < -t)) {
 229                 zone_page_state_add(x, zone, item);
 230                 x = 0;
 231         }
 232         __this_cpu_write(*p, x);
 233 }
 234 EXPORT_SYMBOL(__mod_zone_page_state);
 235
 236 /*
 237  * Optimized increment and decrement functions.
 238  *
 239  * These are only for a single page and therefore can take a struct page *
 240  * argument instead of struct zone *. This allows the inclusion of the code
 241  * generated for page_zone(page) into the optimized functions.
 242  *
 243  * No overflow check is necessary and therefore the differential can be
 244  * incremented or decremented in place which may allow the compilers to
 245  * generate better code.
 246  * The increment or decrement is known and therefore one boundary check can
 247  * be omitted.
 248  *
 249  * NOTE: These functions are very performance sensitive. Change only
 250  * with care.
 251  *
 252  * Some processors have inc/dec instructions that are atomic vs an interrupt.
 253  * However, the code must first determine the differential location in a zone
 254  * based on the processor number and then inc/dec the counter. There is no
 255  * guarantee without disabling preemption that the processor will not change
 256  * in between and therefore the atomicity vs. interrupt cannot be exploited
 257  * in a useful way here.
 258  */
 259 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
 260 {
 261         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 262         s8 __percpu *p = pcp->vm_stat_diff + item;
 263         s8 v, t;
 264
 265         v = __this_cpu_inc_return(*p);
 266         t = __this_cpu_read(pcp->stat_threshold);
 267         if (unlikely(v > t)) {
 268                 s8 overstep = t >> 1;
 269
 270                 zone_page_state_add(v + overstep, zone, item);
 271                 __this_cpu_write(*p, -overstep);
 272         }
 273 }
 274
 275 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
 276 {
 277         __inc_zone_state(page_zone(page), item);
 278 }
 279 EXPORT_SYMBOL(__inc_zone_page_state);
 280
 281 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
 282 {
 283         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 284         s8 __percpu *p = pcp->vm_stat_diff + item;
 285         s8 v, t;
 286
 287         v = __this_cpu_dec_return(*p);
 288         t = __this_cpu_read(pcp->stat_threshold);
 289         if (unlikely(v < - t)) {
 290                 s8 overstep = t >> 1;
 291
 292                 zone_page_state_add(v - overstep, zone, item);
 293                 __this_cpu_write(*p, overstep);
 294         }
 295 }
 296
 297 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
 298 {
 299         __dec_zone_state(page_zone(page), item);
 300 }
 301 EXPORT_SYMBOL(__dec_zone_page_state);
 302
 303 #ifdef CONFIG_HAVE_CMPXCHG_LOCAL
 304 /*
 305  * If we have cmpxchg_local support then we do not need to incur the overhead
 306  * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
 307  *
 308  * mod_state() modifies the zone counter state through atomic per cpu
 309  * operations.
 310  *
 311  * Overstep mode specifies how overstep should handled:
 312  *     0       No overstepping
 313  *     1       Overstepping half of threshold
 314  *     -1      Overstepping minus half of threshold
 315 */
 316 static inline void mod_state(struct zone *zone,
 317        enum zone_stat_item item, int delta, int overstep_mode)
 318 {
 319         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 320         s8 __percpu *p = pcp->vm_stat_diff + item;
 321         long o, n, t, z;
 322
 323         do {
 324                 z = 0;  /* overflow to zone counters */
 325
 326                 /*
 327                  * The fetching of the stat_threshold is racy. We may apply
 328                  * a counter threshold to the wrong the cpu if we get
 329                  * rescheduled while executing here. However, the next
 330                  * counter update will apply the threshold again and
 331                  * therefore bring the counter under the threshold again.
 332                  *
 333                  * Most of the time the thresholds are the same anyways
 334                  * for all cpus in a zone.
 335                  */
 336                 t = this_cpu_read(pcp->stat_threshold);
 337
 338                 o = this_cpu_read(*p);
 339                 n = delta + o;
 340
 341                 if (n > t || n < -t) {
 342                         int os = overstep_mode * (t >> 1) ;
 343
 344                         /* Overflow must be added to zone counters */
 345                         z = n + os;
 346                         n = -os;
 347                 }
 348         } while (this_cpu_cmpxchg(*p, o, n) != o);
 349
 350         if (z)
 351                 zone_page_state_add(z, zone, item);
 352 }
 353
 354 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 355                                         int delta)
 356 {
 357         mod_state(zone, item, delta, 0);
 358 }
 359 EXPORT_SYMBOL(mod_zone_page_state);
 360
 361 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
 362 {
 363         mod_state(zone, item, 1, 1);
 364 }
 365
 366 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 367 {
 368         mod_state(page_zone(page), item, 1, 1);
 369 }
 370 EXPORT_SYMBOL(inc_zone_page_state);
 371
 372 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 373 {
 374         mod_state(page_zone(page), item, -1, -1);
 375 }
 376 EXPORT_SYMBOL(dec_zone_page_state);
 377 #else
 378 /*
 379  * Use interrupt disable to serialize counter updates
 380  */
 381 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 382                                         int delta)
 383 {
 384         unsigned long flags;
 385
 386         local_irq_save(flags);
 387         __mod_zone_page_state(zone, item, delta);
 388         local_irq_restore(flags);
 389 }
 390 EXPORT_SYMBOL(mod_zone_page_state);
 391
 392 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
 393 {
 394         unsigned long flags;
 395
 396         local_irq_save(flags);
 397         __inc_zone_state(zone, item);
 398         local_irq_restore(flags);
 399 }
 400
 401 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 402 {
 403         unsigned long flags;
 404         struct zone *zone;
 405
 406         zone = page_zone(page);
 407         local_irq_save(flags);
 408         __inc_zone_state(zone, item);
 409         local_irq_restore(flags);
 410 }
 411 EXPORT_SYMBOL(inc_zone_page_state);
 412
 413 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 414 {
 415         unsigned long flags;
 416
 417         local_irq_save(flags);
 418         __dec_zone_page_state(page, item);
 419         local_irq_restore(flags);
 420 }
 421 EXPORT_SYMBOL(dec_zone_page_state);
 422 #endif
 423
 424
 425 /*
 426  * Fold a differential into the global counters.
 427  * Returns the number of counters updated.
 428  */
 429 static int fold_diff(int *diff)
 430 {
 431         int i;
 432         int changes = 0;
 433
 434         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 435                 if (diff[i]) {
 436                         atomic_long_add(diff[i], &vm_stat[i]);
 437                         changes++;
 438         }
 439         return changes;
 440 }
 441
 442 /*
 443  * Update the zone counters for the current cpu.
 444  *
 445  * Note that refresh_cpu_vm_stats strives to only access
 446  * node local memory. The per cpu pagesets on remote zones are placed
 447  * in the memory local to the processor using that pageset. So the
 448  * loop over all zones will access a series of cachelines local to
 449  * the processor.
 450  *
 451  * The call to zone_page_state_add updates the cachelines with the
 452  * statistics in the remote zone struct as well as the global cachelines
 453  * with the global counters. These could cause remote node cache line
 454  * bouncing and will have to be only done when necessary.
 455  *
 456  * The function returns the number of global counters updated.
 457  */
 458 static int refresh_cpu_vm_stats(void)
 459 {
 460         struct zone *zone;
 461         int i;
 462         int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 463         int changes = 0;
 464
 465         for_each_populated_zone(zone) {
 466                 struct per_cpu_pageset __percpu *p = zone->pageset;
 467
 468                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
 469                         int v;
 470
 471                         v = this_cpu_xchg(p->vm_stat_diff[i], 0);
 472                         if (v) {
 473
 474                                 atomic_long_add(v, &zone->vm_stat[i]);
 475                                 global_diff[i] += v;
 476 #ifdef CONFIG_NUMA
 477                                 /* 3 seconds idle till flush */
 478                                 __this_cpu_write(p->expire, 3);
 479 #endif
 480                         }
 481                 }
 482                 cond_resched();
 483 #ifdef CONFIG_NUMA
 484                 /*
 485                  * Deal with draining the remote pageset of this
 486                  * processor
 487                  *
 488                  * Check if there are pages remaining in this pageset
 489                  * if not then there is nothing to expire.
 490                  */
 491                 if (!__this_cpu_read(p->expire) ||
 492                                !__this_cpu_read(p->pcp.count))
 493                         continue;
 494
 495                 /*
 496                  * We never drain zones local to this processor.
 497                  */
 498                 if (zone_to_nid(zone) == numa_node_id()) {
 499                         __this_cpu_write(p->expire, 0);
 500                         continue;
 501                 }
 502
 503                 if (__this_cpu_dec_return(p->expire))
 504                         continue;
 505
 506                 if (__this_cpu_read(p->pcp.count)) {
 507                         drain_zone_pages(zone, this_cpu_ptr(&p->pcp));
 508                         changes++;
 509                 }
 510 #endif
 511         }
 512         changes += fold_diff(global_diff);
 513         return changes;
 514 }
 515
 516 /*
 517  * Fold the data for an offline cpu into the global array.
 518  * There cannot be any access by the offline cpu and therefore
 519  * synchronization is simplified.
 520  */
 521 void cpu_vm_stats_fold(int cpu)
 522 {
 523         struct zone *zone;
 524         int i;
 525         int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 526
 527         for_each_populated_zone(zone) {
 528                 struct per_cpu_pageset *p;
 529
 530                 p = per_cpu_ptr(zone->pageset, cpu);
 531
 532                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 533                         if (p->vm_stat_diff[i]) {
 534                                 int v;
 535
 536                                 v = p->vm_stat_diff[i];
 537                                 p->vm_stat_diff[i] = 0;
 538                                 atomic_long_add(v, &zone->vm_stat[i]);
 539                                 global_diff[i] += v;
 540                         }
 541         }
 542
 543         fold_diff(global_diff);
 544 }
 545
 546 /*
 547  * this is only called if !populated_zone(zone), which implies no other users of
 548  * pset->vm_stat_diff[] exsist.
 549  */
 550 void drain_zonestat(struct zone *zone, struct per_cpu_pageset *pset)
 551 {
 552         int i;
 553
 554         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 555                 if (pset->vm_stat_diff[i]) {
 556                         int v = pset->vm_stat_diff[i];
 557                         pset->vm_stat_diff[i] = 0;
 558                         atomic_long_add(v, &zone->vm_stat[i]);
 559                         atomic_long_add(v, &vm_stat[i]);
 560                 }
 561 }
 562 #endif
 563
 564 #ifdef CONFIG_NUMA
 565 /*
 566  * zonelist = the list of zones passed to the allocator
 567  * z        = the zone from which the allocation occurred.
 568  *
 569  * Must be called with interrupts disabled.
 570  *
 571  * When __GFP_OTHER_NODE is set assume the node of the preferred
 572  * zone is the local node. This is useful for daemons who allocate
 573  * memory on behalf of other processes.
 574  */
 575 void zone_statistics(struct zone *preferred_zone, struct zone *z, gfp_t flags)
 576 {
 577         if (z->zone_pgdat == preferred_zone->zone_pgdat) {
 578                 __inc_zone_state(z, NUMA_HIT);
 579         } else {
 580                 __inc_zone_state(z, NUMA_MISS);
 581                 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
 582         }
 583         if (z->node == ((flags & __GFP_OTHER_NODE) ?
 584                         preferred_zone->node : numa_node_id()))
 585                 __inc_zone_state(z, NUMA_LOCAL);
 586         else
 587                 __inc_zone_state(z, NUMA_OTHER);
 588 }
 589 #endif
 590
 591 #ifdef CONFIG_COMPACTION
 592
 593 struct contig_page_info {
 594         unsigned long free_pages;
 595         unsigned long free_blocks_total;
 596         unsigned long free_blocks_suitable;
 597 };
 598
 599 /*
 600  * Calculate the number of free pages in a zone, how many contiguous
 601  * pages are free and how many are large enough to satisfy an allocation of
 602  * the target size. Note that this function makes no attempt to estimate
 603  * how many suitable free blocks there *might* be if MOVABLE pages were
 604  * migrated. Calculating that is possible, but expensive and can be
 605  * figured out from userspace
 606  */
 607 static void fill_contig_page_info(struct zone *zone,
 608                                 unsigned int suitable_order,
 609                                 struct contig_page_info *info)
 610 {
 611         unsigned int order;
 612
 613         info->free_pages = 0;
 614         info->free_blocks_total = 0;
 615         info->free_blocks_suitable = 0;
 616
 617         for (order = 0; order < MAX_ORDER; order++) {
 618                 unsigned long blocks;
 619
 620                 /* Count number of free blocks */
 621                 blocks = zone->free_area[order].nr_free;
 622                 info->free_blocks_total += blocks;
 623
 624                 /* Count free base pages */
 625                 info->free_pages += blocks << order;
 626
 627                 /* Count the suitable free blocks */
 628                 if (order >= suitable_order)
 629                         info->free_blocks_suitable += blocks <<
 630                                                 (order - suitable_order);
 631         }
 632 }
 633
 634 /*
 635  * A fragmentation index only makes sense if an allocation of a requested
 636  * size would fail. If that is true, the fragmentation index indicates
 637  * whether external fragmentation or a lack of memory was the problem.
 638  * The value can be used to determine if page reclaim or compaction
 639  * should be used
 640  */
 641 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
 642 {
 643         unsigned long requested = 1UL << order;
 644
 645         if (!info->free_blocks_total)
 646                 return 0;
 647
 648         /* Fragmentation index only makes sense when a request would fail */
 649         if (info->free_blocks_suitable)
 650                 return -1000;
 651
 652         /*
 653          * Index is between 0 and 1 so return within 3 decimal places
 654          *
 655          * 0 => allocation would fail due to lack of memory
 656          * 1 => allocation would fail due to fragmentation
 657          */
 658         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
 659 }
 660
 661 /* Same as __fragmentation index but allocs contig_page_info on stack */
 662 int fragmentation_index(struct zone *zone, unsigned int order)
 663 {
 664         struct contig_page_info info;
 665
 666         fill_contig_page_info(zone, order, &info);
 667         return __fragmentation_index(order, &info);
 668 }
 669 #endif
 670
 671 #if defined(CONFIG_PROC_FS) || defined(CONFIG_COMPACTION)
 672 #include <linux/proc_fs.h>
 673 #include <linux/seq_file.h>
 674
 675 static char * const migratetype_names[MIGRATE_TYPES] = {
 676         "Unmovable",
 677         "Reclaimable",
 678         "Movable",
 679         "Reserve",
 680 #ifdef CONFIG_CMA
 681         "CMA",
 682 #endif
 683 #ifdef CONFIG_MEMORY_ISOLATION
 684         "Isolate",
 685 #endif
 686 };
 687
 688 static void *frag_start(struct seq_file *m, loff_t *pos)
 689 {
 690         pg_data_t *pgdat;
 691         loff_t node = *pos;
 692         for (pgdat = first_online_pgdat();
 693              pgdat && node;
 694              pgdat = next_online_pgdat(pgdat))
 695                 --node;
 696
 697         return pgdat;
 698 }
 699
 700 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
 701 {
 702         pg_data_t *pgdat = (pg_data_t *)arg;
 703
 704         (*pos)++;
 705         return next_online_pgdat(pgdat);
 706 }
 707
 708 static void frag_stop(struct seq_file *m, void *arg)
 709 {
 710 }
 711
 712 /* Walk all the zones in a node and print using a callback */
 713 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
 714                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
 715 {
 716         struct zone *zone;
 717         struct zone *node_zones = pgdat->node_zones;
 718         unsigned long flags;
 719
 720         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 721                 if (!populated_zone(zone))
 722                         continue;
 723
 724                 spin_lock_irqsave(&zone->lock, flags);
 725                 print(m, pgdat, zone);
 726                 spin_unlock_irqrestore(&zone->lock, flags);
 727         }
 728 }
 729 #endif
 730
 731 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || defined(CONFIG_NUMA)
 732 #ifdef CONFIG_ZONE_DMA
 733 #define TEXT_FOR_DMA(xx) xx "_dma",
 734 #else
 735 #define TEXT_FOR_DMA(xx)
 736 #endif
 737
 738 #ifdef CONFIG_ZONE_DMA32
 739 #define TEXT_FOR_DMA32(xx) xx "_dma32",
 740 #else
 741 #define TEXT_FOR_DMA32(xx)
 742 #endif
 743
 744 #ifdef CONFIG_HIGHMEM
 745 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
 746 #else
 747 #define TEXT_FOR_HIGHMEM(xx)
 748 #endif
 749
 750 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
 751                                         TEXT_FOR_HIGHMEM(xx) xx "_movable",
 752
 753 const char * const vmstat_text[] = {
 754         /* enum zone_stat_item countes */
 755         "nr_free_pages",
 756         "nr_alloc_batch",
 757         "nr_inactive_anon",
 758         "nr_active_anon",
 759         "nr_inactive_file",
 760         "nr_active_file",
 761         "nr_unevictable",
 762         "nr_mlock",
 763         "nr_anon_pages",
 764         "nr_mapped",
 765         "nr_file_pages",
 766         "nr_dirty",
 767         "nr_writeback",
 768         "nr_slab_reclaimable",
 769         "nr_slab_unreclaimable",
 770         "nr_page_table_pages",
 771         "nr_kernel_stack",
 772         "nr_unstable",
 773         "nr_bounce",
 774         "nr_vmscan_write",
 775         "nr_vmscan_immediate_reclaim",
 776         "nr_writeback_temp",
 777         "nr_isolated_anon",
 778         "nr_isolated_file",
 779         "nr_shmem",
 780         "nr_dirtied",
 781         "nr_written",
 782         "nr_pages_scanned",
 783
 784 #ifdef CONFIG_NUMA
 785         "numa_hit",
 786         "numa_miss",
 787         "numa_foreign",
 788         "numa_interleave",
 789         "numa_local",
 790         "numa_other",
 791 #endif
 792         "workingset_refault",
 793         "workingset_activate",
 794         "workingset_nodereclaim",
 795         "nr_anon_transparent_hugepages",
 796         "nr_free_cma",
 797
 798         /* enum writeback_stat_item counters */
 799         "nr_dirty_threshold",
 800         "nr_dirty_background_threshold",
 801
 802 #ifdef CONFIG_VM_EVENT_COUNTERS
 803         /* enum vm_event_item counters */
 804         "pgpgin",
 805         "pgpgout",
 806         "pswpin",
 807         "pswpout",
 808
 809         TEXTS_FOR_ZONES("pgalloc")
 810
 811         "pgfree",
 812         "pgactivate",
 813         "pgdeactivate",
 814
 815         "pgfault",
 816         "pgmajfault",
 817
 818         TEXTS_FOR_ZONES("pgrefill")
 819         TEXTS_FOR_ZONES("pgsteal_kswapd")
 820         TEXTS_FOR_ZONES("pgsteal_direct")
 821         TEXTS_FOR_ZONES("pgscan_kswapd")
 822         TEXTS_FOR_ZONES("pgscan_direct")
 823         "pgscan_direct_throttle",
 824
 825 #ifdef CONFIG_NUMA
 826         "zone_reclaim_failed",
 827 #endif
 828         "pginodesteal",
 829         "slabs_scanned",
 830         "kswapd_inodesteal",
 831         "kswapd_low_wmark_hit_quickly",
 832         "kswapd_high_wmark_hit_quickly",
 833         "pageoutrun",
 834         "allocstall",
 835
 836         "pgrotated",
 837
 838         "drop_pagecache",
 839         "drop_slab",
 840
 841 #ifdef CONFIG_NUMA_BALANCING
 842         "numa_pte_updates",
 843         "numa_huge_pte_updates",
 844         "numa_hint_faults",
 845         "numa_hint_faults_local",
 846         "numa_pages_migrated",
 847 #endif
 848 #ifdef CONFIG_MIGRATION
 849         "pgmigrate_success",
 850         "pgmigrate_fail",
 851 #endif
 852 #ifdef CONFIG_COMPACTION
 853         "compact_migrate_scanned",
 854         "compact_free_scanned",
 855         "compact_isolated",
 856         "compact_stall",
 857         "compact_fail",
 858         "compact_success",
 859 #endif
 860
 861 #ifdef CONFIG_HUGETLB_PAGE
 862         "htlb_buddy_alloc_success",
 863         "htlb_buddy_alloc_fail",
 864 #endif
 865         "unevictable_pgs_culled",
 866         "unevictable_pgs_scanned",
 867         "unevictable_pgs_rescued",
 868         "unevictable_pgs_mlocked",
 869         "unevictable_pgs_munlocked",
 870         "unevictable_pgs_cleared",
 871         "unevictable_pgs_stranded",
 872
 873 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 874         "thp_fault_alloc",
 875         "thp_fault_fallback",
 876         "thp_collapse_alloc",
 877         "thp_collapse_alloc_failed",
 878         "thp_split",
 879         "thp_zero_page_alloc",
 880         "thp_zero_page_alloc_failed",
 881 #endif
 882 #ifdef CONFIG_MEMORY_BALLOON
 883         "balloon_inflate",
 884         "balloon_deflate",
 885 #ifdef CONFIG_BALLOON_COMPACTION
 886         "balloon_migrate",
 887 #endif
 888 #endif /* CONFIG_MEMORY_BALLOON */
 889 #ifdef CONFIG_DEBUG_TLBFLUSH
 890 #ifdef CONFIG_SMP
 891         "nr_tlb_remote_flush",
 892         "nr_tlb_remote_flush_received",
 893 #endif /* CONFIG_SMP */
 894         "nr_tlb_local_flush_all",
 895         "nr_tlb_local_flush_one",
 896 #endif /* CONFIG_DEBUG_TLBFLUSH */
 897
 898 #ifdef CONFIG_DEBUG_VM_VMACACHE
 899         "vmacache_find_calls",
 900         "vmacache_find_hits",
 901 #endif
 902 #endif /* CONFIG_VM_EVENTS_COUNTERS */
 903 };
 904 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
 905
 906
 907 #ifdef CONFIG_PROC_FS
 908 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
 909                                                 struct zone *zone)
 910 {
 911         int order;
 912
 913         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 914         for (order = 0; order < MAX_ORDER; ++order)
 915                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
 916         seq_putc(m, '\n');
 917 }
 918
 919 /*
 920  * This walks the free areas for each zone.
 921  */
 922 static int frag_show(struct seq_file *m, void *arg)
 923 {
 924         pg_data_t *pgdat = (pg_data_t *)arg;
 925         walk_zones_in_node(m, pgdat, frag_show_print);
 926         return 0;
 927 }
 928
 929 static void pagetypeinfo_showfree_print(struct seq_file *m,
 930                                         pg_data_t *pgdat, struct zone *zone)
 931 {
 932         int order, mtype;
 933
 934         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
 935                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
 936                                         pgdat->node_id,
 937                                         zone->name,
 938                                         migratetype_names[mtype]);
 939                 for (order = 0; order < MAX_ORDER; ++order) {
 940                         unsigned long freecount = 0;
 941                         struct free_area *area;
 942                         struct list_head *curr;
 943
 944                         area = &(zone->free_area[order]);
 945
 946                         list_for_each(curr, &area->free_list[mtype])
 947                                 freecount++;
 948                         seq_printf(m, "%6lu ", freecount);
 949                 }
 950                 seq_putc(m, '\n');
 951         }
 952 }
 953
 954 /* Print out the free pages at each order for each migatetype */
 955 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
 956 {
 957         int order;
 958         pg_data_t *pgdat = (pg_data_t *)arg;
 959
 960         /* Print header */
 961         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
 962         for (order = 0; order < MAX_ORDER; ++order)
 963                 seq_printf(m, "%6d ", order);
 964         seq_putc(m, '\n');
 965
 966         walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
 967
 968         return 0;
 969 }
 970
 971 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
 972                                         pg_data_t *pgdat, struct zone *zone)
 973 {
 974         int mtype;
 975         unsigned long pfn;
 976         unsigned long start_pfn = zone->zone_start_pfn;
 977         unsigned long end_pfn = zone_end_pfn(zone);
 978         unsigned long count[MIGRATE_TYPES] = { 0, };
 979
 980         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
 981                 struct page *page;
 982
 983                 if (!pfn_valid(pfn))
 984                         continue;
 985
 986                 page = pfn_to_page(pfn);
 987
 988                 /* Watch for unexpected holes punched in the memmap */
 989                 if (!memmap_valid_within(pfn, page, zone))
 990                         continue;
 991
 992                 mtype = get_pageblock_migratetype(page);
 993
 994                 if (mtype < MIGRATE_TYPES)
 995                         count[mtype]++;
 996         }
 997
 998         /* Print counts */
 999         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1000         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1001                 seq_printf(m, "%12lu ", count[mtype]);
1002         seq_putc(m, '\n');
1003 }
1004
1005 /* Print out the free pages at each order for each migratetype */
1006 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1007 {
1008         int mtype;
1009         pg_data_t *pgdat = (pg_data_t *)arg;
1010
1011         seq_printf(m, "\n%-23s", "Number of blocks type ");
1012         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1013                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1014         seq_putc(m, '\n');
1015         walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
1016
1017         return 0;
1018 }
1019
1020 /*
1021  * This prints out statistics in relation to grouping pages by mobility.
1022  * It is expensive to collect so do not constantly read the file.
1023  */
1024 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1025 {
1026         pg_data_t *pgdat = (pg_data_t *)arg;
1027
1028         /* check memoryless node */
1029         if (!node_state(pgdat->node_id, N_MEMORY))
1030                 return 0;
1031
1032         seq_printf(m, "Page block order: %d\n", pageblock_order);
1033         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
1034         seq_putc(m, '\n');
1035         pagetypeinfo_showfree(m, pgdat);
1036         pagetypeinfo_showblockcount(m, pgdat);
1037
1038         return 0;
1039 }
1040
1041 static const struct seq_operations fragmentation_op = {
1042         .start  = frag_start,
1043         .next   = frag_next,
1044         .stop   = frag_stop,
1045         .show   = frag_show,
1046 };
1047
1048 static int fragmentation_open(struct inode *inode, struct file *file)
1049 {
1050         return seq_open(file, &fragmentation_op);
1051 }
1052
1053 static const struct file_operations fragmentation_file_operations = {
1054         .open           = fragmentation_open,
1055         .read           = seq_read,
1056         .llseek         = seq_lseek,
1057         .release        = seq_release,
1058 };
1059
1060 static const struct seq_operations pagetypeinfo_op = {
1061         .start  = frag_start,
1062         .next   = frag_next,
1063         .stop   = frag_stop,
1064         .show   = pagetypeinfo_show,
1065 };
1066
1067 static int pagetypeinfo_open(struct inode *inode, struct file *file)
1068 {
1069         return seq_open(file, &pagetypeinfo_op);
1070 }
1071
1072 static const struct file_operations pagetypeinfo_file_ops = {
1073         .open           = pagetypeinfo_open,
1074         .read           = seq_read,
1075         .llseek         = seq_lseek,
1076         .release        = seq_release,
1077 };
1078
1079 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1080                                                         struct zone *zone)
1081 {
1082         int i;
1083         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1084         seq_printf(m,
1085                    "\n  pages free     %lu"
1086                    "\n        min      %lu"
1087                    "\n        low      %lu"
1088                    "\n        high     %lu"
1089                    "\n        scanned  %lu"
1090                    "\n        spanned  %lu"
1091                    "\n        present  %lu"
1092                    "\n        managed  %lu",
1093                    zone_page_state(zone, NR_FREE_PAGES),
1094                    min_wmark_pages(zone),
1095                    low_wmark_pages(zone),
1096                    high_wmark_pages(zone),
1097                    zone_page_state(zone, NR_PAGES_SCANNED),
1098                    zone->spanned_pages,
1099                    zone->present_pages,
1100                    zone->managed_pages);
1101
1102         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1103                 seq_printf(m, "\n    %-12s %lu", vmstat_text[i],
1104                                 zone_page_state(zone, i));
1105
1106         seq_printf(m,
1107                    "\n        protection: (%ld",
1108                    zone->lowmem_reserve[0]);
1109         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1110                 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1111         seq_printf(m,
1112                    ")"
1113                    "\n  pagesets");
1114         for_each_online_cpu(i) {
1115                 struct per_cpu_pageset *pageset;
1116
1117                 pageset = per_cpu_ptr(zone->pageset, i);
1118                 seq_printf(m,
1119                            "\n    cpu: %i"
1120                            "\n              count: %i"
1121                            "\n              high:  %i"
1122                            "\n              batch: %i",
1123                            i,
1124                            pageset->pcp.count,
1125                            pageset->pcp.high,
1126                            pageset->pcp.batch);
1127 #ifdef CONFIG_SMP
1128                 seq_printf(m, "\n  vm stats threshold: %d",
1129                                 pageset->stat_threshold);
1130 #endif
1131         }
1132         seq_printf(m,
1133                    "\n  all_unreclaimable: %u"
1134                    "\n  start_pfn:         %lu"
1135                    "\n  inactive_ratio:    %u",
1136                    !zone_reclaimable(zone),
1137                    zone->zone_start_pfn,
1138                    zone->inactive_ratio);
1139         seq_putc(m, '\n');
1140 }
1141
1142 /*
1143  * Output information about zones in @pgdat.
1144  */
1145 static int zoneinfo_show(struct seq_file *m, void *arg)
1146 {
1147         pg_data_t *pgdat = (pg_data_t *)arg;
1148         walk_zones_in_node(m, pgdat, zoneinfo_show_print);
1149         return 0;
1150 }
1151
1152 static const struct seq_operations zoneinfo_op = {
1153         .start  = frag_start, /* iterate over all zones. The same as in
1154                                * fragmentation. */
1155         .next   = frag_next,
1156         .stop   = frag_stop,
1157         .show   = zoneinfo_show,
1158 };
1159
1160 static int zoneinfo_open(struct inode *inode, struct file *file)
1161 {
1162         return seq_open(file, &zoneinfo_op);
1163 }
1164
1165 static const struct file_operations proc_zoneinfo_file_operations = {
1166         .open           = zoneinfo_open,
1167         .read           = seq_read,
1168         .llseek         = seq_lseek,
1169         .release        = seq_release,
1170 };
1171
1172 enum writeback_stat_item {
1173         NR_DIRTY_THRESHOLD,
1174         NR_DIRTY_BG_THRESHOLD,
1175         NR_VM_WRITEBACK_STAT_ITEMS,
1176 };
1177
1178 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1179 {
1180         unsigned long *v;
1181         int i, stat_items_size;
1182
1183         if (*pos >= ARRAY_SIZE(vmstat_text))
1184                 return NULL;
1185         stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1186                           NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1187
1188 #ifdef CONFIG_VM_EVENT_COUNTERS
1189         stat_items_size += sizeof(struct vm_event_state);
1190 #endif
1191
1192         v = kmalloc(stat_items_size, GFP_KERNEL);
1193         m->private = v;
1194         if (!v)
1195                 return ERR_PTR(-ENOMEM);
1196         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1197                 v[i] = global_page_state(i);
1198         v += NR_VM_ZONE_STAT_ITEMS;
1199
1200         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1201                             v + NR_DIRTY_THRESHOLD);
1202         v += NR_VM_WRITEBACK_STAT_ITEMS;
1203
1204 #ifdef CONFIG_VM_EVENT_COUNTERS
1205         all_vm_events(v);
1206         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1207         v[PGPGOUT] /= 2;
1208 #endif
1209         return (unsigned long *)m->private + *pos;
1210 }
1211
1212 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1213 {
1214         (*pos)++;
1215         if (*pos >= ARRAY_SIZE(vmstat_text))
1216                 return NULL;
1217         return (unsigned long *)m->private + *pos;
1218 }
1219
1220 static int vmstat_show(struct seq_file *m, void *arg)
1221 {
1222         unsigned long *l = arg;
1223         unsigned long off = l - (unsigned long *)m->private;
1224
1225         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1226         return 0;
1227 }
1228
1229 static void vmstat_stop(struct seq_file *m, void *arg)
1230 {
1231         kfree(m->private);
1232         m->private = NULL;
1233 }
1234
1235 static const struct seq_operations vmstat_op = {
1236         .start  = vmstat_start,
1237         .next   = vmstat_next,
1238         .stop   = vmstat_stop,
1239         .show   = vmstat_show,
1240 };
1241
1242 static int vmstat_open(struct inode *inode, struct file *file)
1243 {
1244         return seq_open(file, &vmstat_op);
1245 }
1246
1247 static const struct file_operations proc_vmstat_file_operations = {
1248         .open           = vmstat_open,
1249         .read           = seq_read,
1250         .llseek         = seq_lseek,
1251         .release        = seq_release,
1252 };
1253 #endif /* CONFIG_PROC_FS */
1254
1255 #ifdef CONFIG_SMP
1256 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1257 int sysctl_stat_interval __read_mostly = HZ;
1258 static cpumask_var_t cpu_stat_off;
1259
1260 static void vmstat_update(struct work_struct *w)
1261 {
1262         if (refresh_cpu_vm_stats())
1263                 /*
1264                  * Counters were updated so we expect more updates
1265                  * to occur in the future. Keep on running the
1266                  * update worker thread.
1267                  */
1268                 schedule_delayed_work(this_cpu_ptr(&vmstat_work),
1269                         round_jiffies_relative(sysctl_stat_interval));
1270         else {
1271                 /*
1272                  * We did not update any counters so the app may be in
1273                  * a mode where it does not cause counter updates.
1274                  * We may be uselessly running vmstat_update.
1275                  * Defer the checking for differentials to the
1276                  * shepherd thread on a different processor.
1277                  */
1278                 int r;
1279                 /*
1280                  * Shepherd work thread does not race since it never
1281                  * changes the bit if its zero but the cpu
1282                  * online / off line code may race if
1283                  * worker threads are still allowed during
1284                  * shutdown / startup.
1285                  */
1286                 r = cpumask_test_and_set_cpu(smp_processor_id(),
1287                         cpu_stat_off);
1288                 VM_BUG_ON(r);
1289         }
1290 }
1291
1292 /*
1293  * Check if the diffs for a certain cpu indicate that
1294  * an update is needed.
1295  */
1296 static bool need_update(int cpu)
1297 {
1298         struct zone *zone;
1299
1300         for_each_populated_zone(zone) {
1301                 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1302
1303                 BUILD_BUG_ON(sizeof(p->vm_stat_diff[0]) != 1);
1304                 /*
1305                  * The fast way of checking if there are any vmstat diffs.
1306                  * This works because the diffs are byte sized items.
1307                  */
1308                 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS))
1309                         return true;
1310
1311         }
1312         return false;
1313 }
1314
1315
1316 /*
1317  * Shepherd worker thread that checks the
1318  * differentials of processors that have their worker
1319  * threads for vm statistics updates disabled because of
1320  * inactivity.
1321  */
1322 static void vmstat_shepherd(struct work_struct *w);
1323
1324 static DECLARE_DELAYED_WORK(shepherd, vmstat_shepherd);
1325
1326 static void vmstat_shepherd(struct work_struct *w)
1327 {
1328         int cpu;
1329
1330         get_online_cpus();
1331         /* Check processors whose vmstat worker threads have been disabled */
1332         for_each_cpu(cpu, cpu_stat_off)
1333                 if (need_update(cpu) &&
1334                         cpumask_test_and_clear_cpu(cpu, cpu_stat_off))
1335
1336                         schedule_delayed_work_on(cpu, &per_cpu(vmstat_work, cpu),
1337                                 __round_jiffies_relative(sysctl_stat_interval, cpu));
1338
1339         put_online_cpus();
1340
1341         schedule_delayed_work(&shepherd,
1342                 round_jiffies_relative(sysctl_stat_interval));
1343
1344 }
1345
1346 static void __init start_shepherd_timer(void)
1347 {
1348         int cpu;
1349
1350         for_each_possible_cpu(cpu)
1351                 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
1352                         vmstat_update);
1353
1354         if (!alloc_cpumask_var(&cpu_stat_off, GFP_KERNEL))
1355                 BUG();
1356         cpumask_copy(cpu_stat_off, cpu_online_mask);
1357
1358         schedule_delayed_work(&shepherd,
1359                 round_jiffies_relative(sysctl_stat_interval));
1360 }
1361
1362 static void vmstat_cpu_dead(int node)
1363 {
1364         int cpu;
1365
1366         get_online_cpus();
1367         for_each_online_cpu(cpu)
1368                 if (cpu_to_node(cpu) == node)
1369                         goto end;
1370
1371         node_clear_state(node, N_CPU);
1372 end:
1373         put_online_cpus();
1374 }
1375
1376 /*
1377  * Use the cpu notifier to insure that the thresholds are recalculated
1378  * when necessary.
1379  */
1380 static int vmstat_cpuup_callback(struct notifier_block *nfb,
1381                 unsigned long action,
1382                 void *hcpu)
1383 {
1384         long cpu = (long)hcpu;
1385
1386         switch (action) {
1387         case CPU_ONLINE:
1388         case CPU_ONLINE_FROZEN:
1389                 refresh_zone_stat_thresholds();
1390                 node_set_state(cpu_to_node(cpu), N_CPU);
1391                 cpumask_set_cpu(cpu, cpu_stat_off);
1392                 break;
1393         case CPU_DOWN_PREPARE:
1394         case CPU_DOWN_PREPARE_FROZEN:
1395                 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1396                 cpumask_clear_cpu(cpu, cpu_stat_off);
1397                 break;
1398         case CPU_DOWN_FAILED:
1399         case CPU_DOWN_FAILED_FROZEN:
1400                 cpumask_set_cpu(cpu, cpu_stat_off);
1401                 break;
1402         case CPU_DEAD:
1403         case CPU_DEAD_FROZEN:
1404                 refresh_zone_stat_thresholds();
1405                 vmstat_cpu_dead(cpu_to_node(cpu));
1406                 break;
1407         default:
1408                 break;
1409         }
1410         return NOTIFY_OK;
1411 }
1412
1413 static struct notifier_block vmstat_notifier =
1414         { &vmstat_cpuup_callback, NULL, 0 };
1415 #endif
1416
1417 static int __init setup_vmstat(void)
1418 {
1419 #ifdef CONFIG_SMP
1420         cpu_notifier_register_begin();
1421         __register_cpu_notifier(&vmstat_notifier);
1422
1423         start_shepherd_timer();
1424         cpu_notifier_register_done();
1425 #endif
1426 #ifdef CONFIG_PROC_FS
1427         proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1428         proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1429         proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1430         proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1431 #endif
1432         return 0;
1433 }
1434 module_init(setup_vmstat)
1435
1436 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1437 #include <linux/debugfs.h>
1438
1439
1440 /*
1441  * Return an index indicating how much of the available free memory is
1442  * unusable for an allocation of the requested size.
1443  */
1444 static int unusable_free_index(unsigned int order,
1445                                 struct contig_page_info *info)
1446 {
1447         /* No free memory is interpreted as all free memory is unusable */
1448         if (info->free_pages == 0)
1449                 return 1000;
1450
1451         /*
1452          * Index should be a value between 0 and 1. Return a value to 3
1453          * decimal places.
1454          *
1455          * 0 => no fragmentation
1456          * 1 => high fragmentation
1457          */
1458         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1459
1460 }
1461
1462 static void unusable_show_print(struct seq_file *m,
1463                                         pg_data_t *pgdat, struct zone *zone)
1464 {
1465         unsigned int order;
1466         int index;
1467         struct contig_page_info info;
1468
1469         seq_printf(m, "Node %d, zone %8s ",
1470                                 pgdat->node_id,
1471                                 zone->name);
1472         for (order = 0; order < MAX_ORDER; ++order) {
1473                 fill_contig_page_info(zone, order, &info);
1474                 index = unusable_free_index(order, &info);
1475                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1476         }
1477
1478         seq_putc(m, '\n');
1479 }
1480
1481 /*
1482  * Display unusable free space index
1483  *
1484  * The unusable free space index measures how much of the available free
1485  * memory cannot be used to satisfy an allocation of a given size and is a
1486  * value between 0 and 1. The higher the value, the more of free memory is
1487  * unusable and by implication, the worse the external fragmentation is. This
1488  * can be expressed as a percentage by multiplying by 100.
1489  */
1490 static int unusable_show(struct seq_file *m, void *arg)
1491 {
1492         pg_data_t *pgdat = (pg_data_t *)arg;
1493
1494         /* check memoryless node */
1495         if (!node_state(pgdat->node_id, N_MEMORY))
1496                 return 0;
1497
1498         walk_zones_in_node(m, pgdat, unusable_show_print);
1499
1500         return 0;
1501 }
1502
1503 static const struct seq_operations unusable_op = {
1504         .start  = frag_start,
1505         .next   = frag_next,
1506         .stop   = frag_stop,
1507         .show   = unusable_show,
1508 };
1509
1510 static int unusable_open(struct inode *inode, struct file *file)
1511 {
1512         return seq_open(file, &unusable_op);
1513 }
1514
1515 static const struct file_operations unusable_file_ops = {
1516         .open           = unusable_open,
1517         .read           = seq_read,
1518         .llseek         = seq_lseek,
1519         .release        = seq_release,
1520 };
1521
1522 static void extfrag_show_print(struct seq_file *m,
1523                                         pg_data_t *pgdat, struct zone *zone)
1524 {
1525         unsigned int order;
1526         int index;
1527
1528         /* Alloc on stack as interrupts are disabled for zone walk */
1529         struct contig_page_info info;
1530
1531         seq_printf(m, "Node %d, zone %8s ",
1532                                 pgdat->node_id,
1533                                 zone->name);
1534         for (order = 0; order < MAX_ORDER; ++order) {
1535                 fill_contig_page_info(zone, order, &info);
1536                 index = __fragmentation_index(order, &info);
1537                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1538         }
1539
1540         seq_putc(m, '\n');
1541 }
1542
1543 /*
1544  * Display fragmentation index for orders that allocations would fail for
1545  */
1546 static int extfrag_show(struct seq_file *m, void *arg)
1547 {
1548         pg_data_t *pgdat = (pg_data_t *)arg;
1549
1550         walk_zones_in_node(m, pgdat, extfrag_show_print);
1551
1552         return 0;
1553 }
1554
1555 static const struct seq_operations extfrag_op = {
1556         .start  = frag_start,
1557         .next   = frag_next,
1558         .stop   = frag_stop,
1559         .show   = extfrag_show,
1560 };
1561
1562 static int extfrag_open(struct inode *inode, struct file *file)
1563 {
1564         return seq_open(file, &extfrag_op);
1565 }
1566
1567 static const struct file_operations extfrag_file_ops = {
1568         .open           = extfrag_open,
1569         .read           = seq_read,
1570         .llseek         = seq_lseek,
1571         .release        = seq_release,
1572 };
1573
1574 static int __init extfrag_debug_init(void)
1575 {
1576         struct dentry *extfrag_debug_root;
1577
1578         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1579         if (!extfrag_debug_root)
1580                 return -ENOMEM;
1581
1582         if (!debugfs_create_file("unusable_index", 0444,
1583                         extfrag_debug_root, NULL, &unusable_file_ops))
1584                 goto fail;
1585
1586         if (!debugfs_create_file("extfrag_index", 0444,
1587                         extfrag_debug_root, NULL, &extfrag_file_ops))
1588                 goto fail;
1589
1590         return 0;
1591 fail:
1592         debugfs_remove_recursive(extfrag_debug_root);
1593         return -ENOMEM;
1594 }
1595
1596 module_init(extfrag_debug_init);
1597 #endif