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 #include <linux/page_ext.h>
  26 #include <linux/page_owner.h>
  27
  28 #include "internal.h"
  29
  30 #ifdef CONFIG_VM_EVENT_COUNTERS
  31 DEFINE_PER_CPU(struct vm_event_state, vm_event_states) = {{0}};
  32 EXPORT_PER_CPU_SYMBOL(vm_event_states);
  33
  34 static void sum_vm_events(unsigned long *ret)
  35 {
  36         int cpu;
  37         int i;
  38
  39         memset(ret, 0, NR_VM_EVENT_ITEMS * sizeof(unsigned long));
  40
  41         for_each_online_cpu(cpu) {
  42                 struct vm_event_state *this = &per_cpu(vm_event_states, cpu);
  43
  44                 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
  45                         ret[i] += this->event[i];
  46         }
  47 }
  48
  49 /*
  50  * Accumulate the vm event counters across all CPUs.
  51  * The result is unavoidably approximate - it can change
  52  * during and after execution of this function.
  53 */
  54 void all_vm_events(unsigned long *ret)
  55 {
  56         get_online_cpus();
  57         sum_vm_events(ret);
  58         put_online_cpus();
  59 }
  60 EXPORT_SYMBOL_GPL(all_vm_events);
  61
  62 /*
  63  * Fold the foreign cpu events into our own.
  64  *
  65  * This is adding to the events on one processor
  66  * but keeps the global counts constant.
  67  */
  68 void vm_events_fold_cpu(int cpu)
  69 {
  70         struct vm_event_state *fold_state = &per_cpu(vm_event_states, cpu);
  71         int i;
  72
  73         for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
  74                 count_vm_events(i, fold_state->event[i]);
  75                 fold_state->event[i] = 0;
  76         }
  77 }
  78
  79 #endif /* CONFIG_VM_EVENT_COUNTERS */
  80
  81 /*
  82  * Manage combined zone based / global counters
  83  *
  84  * vm_stat contains the global counters
  85  */
  86 atomic_long_t vm_stat[NR_VM_ZONE_STAT_ITEMS] __cacheline_aligned_in_smp;
  87 EXPORT_SYMBOL(vm_stat);
  88
  89 #ifdef CONFIG_SMP
  90
  91 int calculate_pressure_threshold(struct zone *zone)
  92 {
  93         int threshold;
  94         int watermark_distance;
  95
  96         /*
  97          * As vmstats are not up to date, there is drift between the estimated
  98          * and real values. For high thresholds and a high number of CPUs, it
  99          * is possible for the min watermark to be breached while the estimated
 100          * value looks fine. The pressure threshold is a reduced value such
 101          * that even the maximum amount of drift will not accidentally breach
 102          * the min watermark
 103          */
 104         watermark_distance = low_wmark_pages(zone) - min_wmark_pages(zone);
 105         threshold = max(1, (int)(watermark_distance / num_online_cpus()));
 106
 107         /*
 108          * Maximum threshold is 125
 109          */
 110         threshold = min(125, threshold);
 111
 112         return threshold;
 113 }
 114
 115 int calculate_normal_threshold(struct zone *zone)
 116 {
 117         int threshold;
 118         int mem;        /* memory in 128 MB units */
 119
 120         /*
 121          * The threshold scales with the number of processors and the amount
 122          * of memory per zone. More memory means that we can defer updates for
 123          * longer, more processors could lead to more contention.
 124          * fls() is used to have a cheap way of logarithmic scaling.
 125          *
 126          * Some sample thresholds:
 127          *
 128          * Threshold    Processors      (fls)   Zonesize        fls(mem+1)
 129          * ------------------------------------------------------------------
 130          * 8            1               1       0.9-1 GB        4
 131          * 16           2               2       0.9-1 GB        4
 132          * 20           2               2       1-2 GB          5
 133          * 24           2               2       2-4 GB          6
 134          * 28           2               2       4-8 GB          7
 135          * 32           2               2       8-16 GB         8
 136          * 4            2               2       <128M           1
 137          * 30           4               3       2-4 GB          5
 138          * 48           4               3       8-16 GB         8
 139          * 32           8               4       1-2 GB          4
 140          * 32           8               4       0.9-1GB         4
 141          * 10           16              5       <128M           1
 142          * 40           16              5       900M            4
 143          * 70           64              7       2-4 GB          5
 144          * 84           64              7       4-8 GB          6
 145          * 108          512             9       4-8 GB          6
 146          * 125          1024            10      8-16 GB         8
 147          * 125          1024            10      16-32 GB        9
 148          */
 149
 150         mem = zone->managed_pages >> (27 - PAGE_SHIFT);
 151
 152         threshold = 2 * fls(num_online_cpus()) * (1 + fls(mem));
 153
 154         /*
 155          * Maximum threshold is 125
 156          */
 157         threshold = min(125, threshold);
 158
 159         return threshold;
 160 }
 161
 162 /*
 163  * Refresh the thresholds for each zone.
 164  */
 165 void refresh_zone_stat_thresholds(void)
 166 {
 167         struct zone *zone;
 168         int cpu;
 169         int threshold;
 170
 171         for_each_populated_zone(zone) {
 172                 unsigned long max_drift, tolerate_drift;
 173
 174                 threshold = calculate_normal_threshold(zone);
 175
 176                 for_each_online_cpu(cpu)
 177                         per_cpu_ptr(zone->pageset, cpu)->stat_threshold
 178                                                         = threshold;
 179
 180                 /*
 181                  * Only set percpu_drift_mark if there is a danger that
 182                  * NR_FREE_PAGES reports the low watermark is ok when in fact
 183                  * the min watermark could be breached by an allocation
 184                  */
 185                 tolerate_drift = low_wmark_pages(zone) - min_wmark_pages(zone);
 186                 max_drift = num_online_cpus() * threshold;
 187                 if (max_drift > tolerate_drift)
 188                         zone->percpu_drift_mark = high_wmark_pages(zone) +
 189                                         max_drift;
 190         }
 191 }
 192
 193 void set_pgdat_percpu_threshold(pg_data_t *pgdat,
 194                                 int (*calculate_pressure)(struct zone *))
 195 {
 196         struct zone *zone;
 197         int cpu;
 198         int threshold;
 199         int i;
 200
 201         for (i = 0; i < pgdat->nr_zones; i++) {
 202                 zone = &pgdat->node_zones[i];
 203                 if (!zone->percpu_drift_mark)
 204                         continue;
 205
 206                 threshold = (*calculate_pressure)(zone);
 207                 for_each_online_cpu(cpu)
 208                         per_cpu_ptr(zone->pageset, cpu)->stat_threshold
 209                                                         = threshold;
 210         }
 211 }
 212
 213 /*
 214  * For use when we know that interrupts are disabled,
 215  * or when we know that preemption is disabled and that
 216  * particular counter cannot be updated from interrupt context.
 217  */
 218 void __mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 219                                 int delta)
 220 {
 221         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 222         s8 __percpu *p = pcp->vm_stat_diff + item;
 223         long x;
 224         long t;
 225
 226         x = delta + __this_cpu_read(*p);
 227
 228         t = __this_cpu_read(pcp->stat_threshold);
 229
 230         if (unlikely(x > t || x < -t)) {
 231                 zone_page_state_add(x, zone, item);
 232                 x = 0;
 233         }
 234         __this_cpu_write(*p, x);
 235 }
 236 EXPORT_SYMBOL(__mod_zone_page_state);
 237
 238 /*
 239  * Optimized increment and decrement functions.
 240  *
 241  * These are only for a single page and therefore can take a struct page *
 242  * argument instead of struct zone *. This allows the inclusion of the code
 243  * generated for page_zone(page) into the optimized functions.
 244  *
 245  * No overflow check is necessary and therefore the differential can be
 246  * incremented or decremented in place which may allow the compilers to
 247  * generate better code.
 248  * The increment or decrement is known and therefore one boundary check can
 249  * be omitted.
 250  *
 251  * NOTE: These functions are very performance sensitive. Change only
 252  * with care.
 253  *
 254  * Some processors have inc/dec instructions that are atomic vs an interrupt.
 255  * However, the code must first determine the differential location in a zone
 256  * based on the processor number and then inc/dec the counter. There is no
 257  * guarantee without disabling preemption that the processor will not change
 258  * in between and therefore the atomicity vs. interrupt cannot be exploited
 259  * in a useful way here.
 260  */
 261 void __inc_zone_state(struct zone *zone, enum zone_stat_item item)
 262 {
 263         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 264         s8 __percpu *p = pcp->vm_stat_diff + item;
 265         s8 v, t;
 266
 267         v = __this_cpu_inc_return(*p);
 268         t = __this_cpu_read(pcp->stat_threshold);
 269         if (unlikely(v > t)) {
 270                 s8 overstep = t >> 1;
 271
 272                 zone_page_state_add(v + overstep, zone, item);
 273                 __this_cpu_write(*p, -overstep);
 274         }
 275 }
 276
 277 void __inc_zone_page_state(struct page *page, enum zone_stat_item item)
 278 {
 279         __inc_zone_state(page_zone(page), item);
 280 }
 281 EXPORT_SYMBOL(__inc_zone_page_state);
 282
 283 void __dec_zone_state(struct zone *zone, enum zone_stat_item item)
 284 {
 285         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 286         s8 __percpu *p = pcp->vm_stat_diff + item;
 287         s8 v, t;
 288
 289         v = __this_cpu_dec_return(*p);
 290         t = __this_cpu_read(pcp->stat_threshold);
 291         if (unlikely(v < - t)) {
 292                 s8 overstep = t >> 1;
 293
 294                 zone_page_state_add(v - overstep, zone, item);
 295                 __this_cpu_write(*p, overstep);
 296         }
 297 }
 298
 299 void __dec_zone_page_state(struct page *page, enum zone_stat_item item)
 300 {
 301         __dec_zone_state(page_zone(page), item);
 302 }
 303 EXPORT_SYMBOL(__dec_zone_page_state);
 304
 305 #ifdef CONFIG_HAVE_CMPXCHG_LOCAL
 306 /*
 307  * If we have cmpxchg_local support then we do not need to incur the overhead
 308  * that comes with local_irq_save/restore if we use this_cpu_cmpxchg.
 309  *
 310  * mod_state() modifies the zone counter state through atomic per cpu
 311  * operations.
 312  *
 313  * Overstep mode specifies how overstep should handled:
 314  *     0       No overstepping
 315  *     1       Overstepping half of threshold
 316  *     -1      Overstepping minus half of threshold
 317 */
 318 static inline void mod_state(struct zone *zone,
 319        enum zone_stat_item item, int delta, int overstep_mode)
 320 {
 321         struct per_cpu_pageset __percpu *pcp = zone->pageset;
 322         s8 __percpu *p = pcp->vm_stat_diff + item;
 323         long o, n, t, z;
 324
 325         do {
 326                 z = 0;  /* overflow to zone counters */
 327
 328                 /*
 329                  * The fetching of the stat_threshold is racy. We may apply
 330                  * a counter threshold to the wrong the cpu if we get
 331                  * rescheduled while executing here. However, the next
 332                  * counter update will apply the threshold again and
 333                  * therefore bring the counter under the threshold again.
 334                  *
 335                  * Most of the time the thresholds are the same anyways
 336                  * for all cpus in a zone.
 337                  */
 338                 t = this_cpu_read(pcp->stat_threshold);
 339
 340                 o = this_cpu_read(*p);
 341                 n = delta + o;
 342
 343                 if (n > t || n < -t) {
 344                         int os = overstep_mode * (t >> 1) ;
 345
 346                         /* Overflow must be added to zone counters */
 347                         z = n + os;
 348                         n = -os;
 349                 }
 350         } while (this_cpu_cmpxchg(*p, o, n) != o);
 351
 352         if (z)
 353                 zone_page_state_add(z, zone, item);
 354 }
 355
 356 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 357                                         int delta)
 358 {
 359         mod_state(zone, item, delta, 0);
 360 }
 361 EXPORT_SYMBOL(mod_zone_page_state);
 362
 363 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
 364 {
 365         mod_state(zone, item, 1, 1);
 366 }
 367
 368 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 369 {
 370         mod_state(page_zone(page), item, 1, 1);
 371 }
 372 EXPORT_SYMBOL(inc_zone_page_state);
 373
 374 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 375 {
 376         mod_state(page_zone(page), item, -1, -1);
 377 }
 378 EXPORT_SYMBOL(dec_zone_page_state);
 379 #else
 380 /*
 381  * Use interrupt disable to serialize counter updates
 382  */
 383 void mod_zone_page_state(struct zone *zone, enum zone_stat_item item,
 384                                         int delta)
 385 {
 386         unsigned long flags;
 387
 388         local_irq_save(flags);
 389         __mod_zone_page_state(zone, item, delta);
 390         local_irq_restore(flags);
 391 }
 392 EXPORT_SYMBOL(mod_zone_page_state);
 393
 394 void inc_zone_state(struct zone *zone, enum zone_stat_item item)
 395 {
 396         unsigned long flags;
 397
 398         local_irq_save(flags);
 399         __inc_zone_state(zone, item);
 400         local_irq_restore(flags);
 401 }
 402
 403 void inc_zone_page_state(struct page *page, enum zone_stat_item item)
 404 {
 405         unsigned long flags;
 406         struct zone *zone;
 407
 408         zone = page_zone(page);
 409         local_irq_save(flags);
 410         __inc_zone_state(zone, item);
 411         local_irq_restore(flags);
 412 }
 413 EXPORT_SYMBOL(inc_zone_page_state);
 414
 415 void dec_zone_page_state(struct page *page, enum zone_stat_item item)
 416 {
 417         unsigned long flags;
 418
 419         local_irq_save(flags);
 420         __dec_zone_page_state(page, item);
 421         local_irq_restore(flags);
 422 }
 423 EXPORT_SYMBOL(dec_zone_page_state);
 424 #endif
 425
 426
 427 /*
 428  * Fold a differential into the global counters.
 429  * Returns the number of counters updated.
 430  */
 431 static int fold_diff(int *diff)
 432 {
 433         int i;
 434         int changes = 0;
 435
 436         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 437                 if (diff[i]) {
 438                         atomic_long_add(diff[i], &vm_stat[i]);
 439                         changes++;
 440         }
 441         return changes;
 442 }
 443
 444 /*
 445  * Update the zone counters for the current cpu.
 446  *
 447  * Note that refresh_cpu_vm_stats strives to only access
 448  * node local memory. The per cpu pagesets on remote zones are placed
 449  * in the memory local to the processor using that pageset. So the
 450  * loop over all zones will access a series of cachelines local to
 451  * the processor.
 452  *
 453  * The call to zone_page_state_add updates the cachelines with the
 454  * statistics in the remote zone struct as well as the global cachelines
 455  * with the global counters. These could cause remote node cache line
 456  * bouncing and will have to be only done when necessary.
 457  *
 458  * The function returns the number of global counters updated.
 459  */
 460 static int refresh_cpu_vm_stats(void)
 461 {
 462         struct zone *zone;
 463         int i;
 464         int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 465         int changes = 0;
 466
 467         for_each_populated_zone(zone) {
 468                 struct per_cpu_pageset __percpu *p = zone->pageset;
 469
 470                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++) {
 471                         int v;
 472
 473                         v = this_cpu_xchg(p->vm_stat_diff[i], 0);
 474                         if (v) {
 475
 476                                 atomic_long_add(v, &zone->vm_stat[i]);
 477                                 global_diff[i] += v;
 478 #ifdef CONFIG_NUMA
 479                                 /* 3 seconds idle till flush */
 480                                 __this_cpu_write(p->expire, 3);
 481 #endif
 482                         }
 483                 }
 484                 cond_resched();
 485 #ifdef CONFIG_NUMA
 486                 /*
 487                  * Deal with draining the remote pageset of this
 488                  * processor
 489                  *
 490                  * Check if there are pages remaining in this pageset
 491                  * if not then there is nothing to expire.
 492                  */
 493                 if (!__this_cpu_read(p->expire) ||
 494                                !__this_cpu_read(p->pcp.count))
 495                         continue;
 496
 497                 /*
 498                  * We never drain zones local to this processor.
 499                  */
 500                 if (zone_to_nid(zone) == numa_node_id()) {
 501                         __this_cpu_write(p->expire, 0);
 502                         continue;
 503                 }
 504
 505                 if (__this_cpu_dec_return(p->expire))
 506                         continue;
 507
 508                 if (__this_cpu_read(p->pcp.count)) {
 509                         drain_zone_pages(zone, this_cpu_ptr(&p->pcp));
 510                         changes++;
 511                 }
 512 #endif
 513         }
 514         changes += fold_diff(global_diff);
 515         return changes;
 516 }
 517
 518 /*
 519  * Fold the data for an offline cpu into the global array.
 520  * There cannot be any access by the offline cpu and therefore
 521  * synchronization is simplified.
 522  */
 523 void cpu_vm_stats_fold(int cpu)
 524 {
 525         struct zone *zone;
 526         int i;
 527         int global_diff[NR_VM_ZONE_STAT_ITEMS] = { 0, };
 528
 529         for_each_populated_zone(zone) {
 530                 struct per_cpu_pageset *p;
 531
 532                 p = per_cpu_ptr(zone->pageset, cpu);
 533
 534                 for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 535                         if (p->vm_stat_diff[i]) {
 536                                 int v;
 537
 538                                 v = p->vm_stat_diff[i];
 539                                 p->vm_stat_diff[i] = 0;
 540                                 atomic_long_add(v, &zone->vm_stat[i]);
 541                                 global_diff[i] += v;
 542                         }
 543         }
 544
 545         fold_diff(global_diff);
 546 }
 547
 548 /*
 549  * this is only called if !populated_zone(zone), which implies no other users of
 550  * pset->vm_stat_diff[] exsist.
 551  */
 552 void drain_zonestat(struct zone *zone, struct per_cpu_pageset *pset)
 553 {
 554         int i;
 555
 556         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
 557                 if (pset->vm_stat_diff[i]) {
 558                         int v = pset->vm_stat_diff[i];
 559                         pset->vm_stat_diff[i] = 0;
 560                         atomic_long_add(v, &zone->vm_stat[i]);
 561                         atomic_long_add(v, &vm_stat[i]);
 562                 }
 563 }
 564 #endif
 565
 566 #ifdef CONFIG_NUMA
 567 /*
 568  * zonelist = the list of zones passed to the allocator
 569  * z        = the zone from which the allocation occurred.
 570  *
 571  * Must be called with interrupts disabled.
 572  *
 573  * When __GFP_OTHER_NODE is set assume the node of the preferred
 574  * zone is the local node. This is useful for daemons who allocate
 575  * memory on behalf of other processes.
 576  */
 577 void zone_statistics(struct zone *preferred_zone, struct zone *z, gfp_t flags)
 578 {
 579         if (z->zone_pgdat == preferred_zone->zone_pgdat) {
 580                 __inc_zone_state(z, NUMA_HIT);
 581         } else {
 582                 __inc_zone_state(z, NUMA_MISS);
 583                 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
 584         }
 585         if (z->node == ((flags & __GFP_OTHER_NODE) ?
 586                         preferred_zone->node : numa_node_id()))
 587                 __inc_zone_state(z, NUMA_LOCAL);
 588         else
 589                 __inc_zone_state(z, NUMA_OTHER);
 590 }
 591 #endif
 592
 593 #ifdef CONFIG_COMPACTION
 594
 595 struct contig_page_info {
 596         unsigned long free_pages;
 597         unsigned long free_blocks_total;
 598         unsigned long free_blocks_suitable;
 599 };
 600
 601 /*
 602  * Calculate the number of free pages in a zone, how many contiguous
 603  * pages are free and how many are large enough to satisfy an allocation of
 604  * the target size. Note that this function makes no attempt to estimate
 605  * how many suitable free blocks there *might* be if MOVABLE pages were
 606  * migrated. Calculating that is possible, but expensive and can be
 607  * figured out from userspace
 608  */
 609 static void fill_contig_page_info(struct zone *zone,
 610                                 unsigned int suitable_order,
 611                                 struct contig_page_info *info)
 612 {
 613         unsigned int order;
 614
 615         info->free_pages = 0;
 616         info->free_blocks_total = 0;
 617         info->free_blocks_suitable = 0;
 618
 619         for (order = 0; order < MAX_ORDER; order++) {
 620                 unsigned long blocks;
 621
 622                 /* Count number of free blocks */
 623                 blocks = zone->free_area[order].nr_free;
 624                 info->free_blocks_total += blocks;
 625
 626                 /* Count free base pages */
 627                 info->free_pages += blocks << order;
 628
 629                 /* Count the suitable free blocks */
 630                 if (order >= suitable_order)
 631                         info->free_blocks_suitable += blocks <<
 632                                                 (order - suitable_order);
 633         }
 634 }
 635
 636 /*
 637  * A fragmentation index only makes sense if an allocation of a requested
 638  * size would fail. If that is true, the fragmentation index indicates
 639  * whether external fragmentation or a lack of memory was the problem.
 640  * The value can be used to determine if page reclaim or compaction
 641  * should be used
 642  */
 643 static int __fragmentation_index(unsigned int order, struct contig_page_info *info)
 644 {
 645         unsigned long requested = 1UL << order;
 646
 647         if (!info->free_blocks_total)
 648                 return 0;
 649
 650         /* Fragmentation index only makes sense when a request would fail */
 651         if (info->free_blocks_suitable)
 652                 return -1000;
 653
 654         /*
 655          * Index is between 0 and 1 so return within 3 decimal places
 656          *
 657          * 0 => allocation would fail due to lack of memory
 658          * 1 => allocation would fail due to fragmentation
 659          */
 660         return 1000 - div_u64( (1000+(div_u64(info->free_pages * 1000ULL, requested))), info->free_blocks_total);
 661 }
 662
 663 /* Same as __fragmentation index but allocs contig_page_info on stack */
 664 int fragmentation_index(struct zone *zone, unsigned int order)
 665 {
 666         struct contig_page_info info;
 667
 668         fill_contig_page_info(zone, order, &info);
 669         return __fragmentation_index(order, &info);
 670 }
 671 #endif
 672
 673 #if defined(CONFIG_PROC_FS) || defined(CONFIG_COMPACTION)
 674 #include <linux/proc_fs.h>
 675 #include <linux/seq_file.h>
 676
 677 static char * const migratetype_names[MIGRATE_TYPES] = {
 678         "Unmovable",
 679         "Reclaimable",
 680         "Movable",
 681         "Reserve",
 682 #ifdef CONFIG_CMA
 683         "CMA",
 684 #endif
 685 #ifdef CONFIG_MEMORY_ISOLATION
 686         "Isolate",
 687 #endif
 688 };
 689
 690 static void *frag_start(struct seq_file *m, loff_t *pos)
 691 {
 692         pg_data_t *pgdat;
 693         loff_t node = *pos;
 694         for (pgdat = first_online_pgdat();
 695              pgdat && node;
 696              pgdat = next_online_pgdat(pgdat))
 697                 --node;
 698
 699         return pgdat;
 700 }
 701
 702 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
 703 {
 704         pg_data_t *pgdat = (pg_data_t *)arg;
 705
 706         (*pos)++;
 707         return next_online_pgdat(pgdat);
 708 }
 709
 710 static void frag_stop(struct seq_file *m, void *arg)
 711 {
 712 }
 713
 714 /* Walk all the zones in a node and print using a callback */
 715 static void walk_zones_in_node(struct seq_file *m, pg_data_t *pgdat,
 716                 void (*print)(struct seq_file *m, pg_data_t *, struct zone *))
 717 {
 718         struct zone *zone;
 719         struct zone *node_zones = pgdat->node_zones;
 720         unsigned long flags;
 721
 722         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
 723                 if (!populated_zone(zone))
 724                         continue;
 725
 726                 spin_lock_irqsave(&zone->lock, flags);
 727                 print(m, pgdat, zone);
 728                 spin_unlock_irqrestore(&zone->lock, flags);
 729         }
 730 }
 731 #endif
 732
 733 #if defined(CONFIG_PROC_FS) || defined(CONFIG_SYSFS) || defined(CONFIG_NUMA)
 734 #ifdef CONFIG_ZONE_DMA
 735 #define TEXT_FOR_DMA(xx) xx "_dma",
 736 #else
 737 #define TEXT_FOR_DMA(xx)
 738 #endif
 739
 740 #ifdef CONFIG_ZONE_DMA32
 741 #define TEXT_FOR_DMA32(xx) xx "_dma32",
 742 #else
 743 #define TEXT_FOR_DMA32(xx)
 744 #endif
 745
 746 #ifdef CONFIG_HIGHMEM
 747 #define TEXT_FOR_HIGHMEM(xx) xx "_high",
 748 #else
 749 #define TEXT_FOR_HIGHMEM(xx)
 750 #endif
 751
 752 #define TEXTS_FOR_ZONES(xx) TEXT_FOR_DMA(xx) TEXT_FOR_DMA32(xx) xx "_normal", \
 753                                         TEXT_FOR_HIGHMEM(xx) xx "_movable",
 754
 755 const char * const vmstat_text[] = {
 756         /* enum zone_stat_item countes */
 757         "nr_free_pages",
 758         "nr_alloc_batch",
 759         "nr_inactive_anon",
 760         "nr_active_anon",
 761         "nr_inactive_file",
 762         "nr_active_file",
 763         "nr_unevictable",
 764         "nr_mlock",
 765         "nr_anon_pages",
 766         "nr_mapped",
 767         "nr_file_pages",
 768         "nr_dirty",
 769         "nr_writeback",
 770         "nr_slab_reclaimable",
 771         "nr_slab_unreclaimable",
 772         "nr_page_table_pages",
 773         "nr_kernel_stack",
 774         "nr_unstable",
 775         "nr_bounce",
 776         "nr_vmscan_write",
 777         "nr_vmscan_immediate_reclaim",
 778         "nr_writeback_temp",
 779         "nr_isolated_anon",
 780         "nr_isolated_file",
 781         "nr_shmem",
 782         "nr_dirtied",
 783         "nr_written",
 784         "nr_pages_scanned",
 785
 786 #ifdef CONFIG_NUMA
 787         "numa_hit",
 788         "numa_miss",
 789         "numa_foreign",
 790         "numa_interleave",
 791         "numa_local",
 792         "numa_other",
 793 #endif
 794         "workingset_refault",
 795         "workingset_activate",
 796         "workingset_nodereclaim",
 797         "nr_anon_transparent_hugepages",
 798         "nr_free_cma",
 799
 800         /* enum writeback_stat_item counters */
 801         "nr_dirty_threshold",
 802         "nr_dirty_background_threshold",
 803
 804 #ifdef CONFIG_VM_EVENT_COUNTERS
 805         /* enum vm_event_item counters */
 806         "pgpgin",
 807         "pgpgout",
 808         "pswpin",
 809         "pswpout",
 810
 811         TEXTS_FOR_ZONES("pgalloc")
 812
 813         "pgfree",
 814         "pgactivate",
 815         "pgdeactivate",
 816
 817         "pgfault",
 818         "pgmajfault",
 819
 820         TEXTS_FOR_ZONES("pgrefill")
 821         TEXTS_FOR_ZONES("pgsteal_kswapd")
 822         TEXTS_FOR_ZONES("pgsteal_direct")
 823         TEXTS_FOR_ZONES("pgscan_kswapd")
 824         TEXTS_FOR_ZONES("pgscan_direct")
 825         "pgscan_direct_throttle",
 826
 827 #ifdef CONFIG_NUMA
 828         "zone_reclaim_failed",
 829 #endif
 830         "pginodesteal",
 831         "slabs_scanned",
 832         "kswapd_inodesteal",
 833         "kswapd_low_wmark_hit_quickly",
 834         "kswapd_high_wmark_hit_quickly",
 835         "pageoutrun",
 836         "allocstall",
 837
 838         "pgrotated",
 839
 840         "drop_pagecache",
 841         "drop_slab",
 842
 843 #ifdef CONFIG_NUMA_BALANCING
 844         "numa_pte_updates",
 845         "numa_huge_pte_updates",
 846         "numa_hint_faults",
 847         "numa_hint_faults_local",
 848         "numa_pages_migrated",
 849 #endif
 850 #ifdef CONFIG_MIGRATION
 851         "pgmigrate_success",
 852         "pgmigrate_fail",
 853 #endif
 854 #ifdef CONFIG_COMPACTION
 855         "compact_migrate_scanned",
 856         "compact_free_scanned",
 857         "compact_isolated",
 858         "compact_stall",
 859         "compact_fail",
 860         "compact_success",
 861 #endif
 862
 863 #ifdef CONFIG_HUGETLB_PAGE
 864         "htlb_buddy_alloc_success",
 865         "htlb_buddy_alloc_fail",
 866 #endif
 867         "unevictable_pgs_culled",
 868         "unevictable_pgs_scanned",
 869         "unevictable_pgs_rescued",
 870         "unevictable_pgs_mlocked",
 871         "unevictable_pgs_munlocked",
 872         "unevictable_pgs_cleared",
 873         "unevictable_pgs_stranded",
 874
 875 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 876         "thp_fault_alloc",
 877         "thp_fault_fallback",
 878         "thp_collapse_alloc",
 879         "thp_collapse_alloc_failed",
 880         "thp_split",
 881         "thp_zero_page_alloc",
 882         "thp_zero_page_alloc_failed",
 883 #endif
 884 #ifdef CONFIG_MEMORY_BALLOON
 885         "balloon_inflate",
 886         "balloon_deflate",
 887 #ifdef CONFIG_BALLOON_COMPACTION
 888         "balloon_migrate",
 889 #endif
 890 #endif /* CONFIG_MEMORY_BALLOON */
 891 #ifdef CONFIG_DEBUG_TLBFLUSH
 892 #ifdef CONFIG_SMP
 893         "nr_tlb_remote_flush",
 894         "nr_tlb_remote_flush_received",
 895 #endif /* CONFIG_SMP */
 896         "nr_tlb_local_flush_all",
 897         "nr_tlb_local_flush_one",
 898 #endif /* CONFIG_DEBUG_TLBFLUSH */
 899
 900 #ifdef CONFIG_DEBUG_VM_VMACACHE
 901         "vmacache_find_calls",
 902         "vmacache_find_hits",
 903         "vmacache_full_flushes",
 904 #endif
 905 #endif /* CONFIG_VM_EVENTS_COUNTERS */
 906 };
 907 #endif /* CONFIG_PROC_FS || CONFIG_SYSFS || CONFIG_NUMA */
 908
 909
 910 #ifdef CONFIG_PROC_FS
 911 static void frag_show_print(struct seq_file *m, pg_data_t *pgdat,
 912                                                 struct zone *zone)
 913 {
 914         int order;
 915
 916         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 917         for (order = 0; order < MAX_ORDER; ++order)
 918                 seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
 919         seq_putc(m, '\n');
 920 }
 921
 922 /*
 923  * This walks the free areas for each zone.
 924  */
 925 static int frag_show(struct seq_file *m, void *arg)
 926 {
 927         pg_data_t *pgdat = (pg_data_t *)arg;
 928         walk_zones_in_node(m, pgdat, frag_show_print);
 929         return 0;
 930 }
 931
 932 static void pagetypeinfo_showfree_print(struct seq_file *m,
 933                                         pg_data_t *pgdat, struct zone *zone)
 934 {
 935         int order, mtype;
 936
 937         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++) {
 938                 seq_printf(m, "Node %4d, zone %8s, type %12s ",
 939                                         pgdat->node_id,
 940                                         zone->name,
 941                                         migratetype_names[mtype]);
 942                 for (order = 0; order < MAX_ORDER; ++order) {
 943                         unsigned long freecount = 0;
 944                         struct free_area *area;
 945                         struct list_head *curr;
 946
 947                         area = &(zone->free_area[order]);
 948
 949                         list_for_each(curr, &area->free_list[mtype])
 950                                 freecount++;
 951                         seq_printf(m, "%6lu ", freecount);
 952                 }
 953                 seq_putc(m, '\n');
 954         }
 955 }
 956
 957 /* Print out the free pages at each order for each migatetype */
 958 static int pagetypeinfo_showfree(struct seq_file *m, void *arg)
 959 {
 960         int order;
 961         pg_data_t *pgdat = (pg_data_t *)arg;
 962
 963         /* Print header */
 964         seq_printf(m, "%-43s ", "Free pages count per migrate type at order");
 965         for (order = 0; order < MAX_ORDER; ++order)
 966                 seq_printf(m, "%6d ", order);
 967         seq_putc(m, '\n');
 968
 969         walk_zones_in_node(m, pgdat, pagetypeinfo_showfree_print);
 970
 971         return 0;
 972 }
 973
 974 static void pagetypeinfo_showblockcount_print(struct seq_file *m,
 975                                         pg_data_t *pgdat, struct zone *zone)
 976 {
 977         int mtype;
 978         unsigned long pfn;
 979         unsigned long start_pfn = zone->zone_start_pfn;
 980         unsigned long end_pfn = zone_end_pfn(zone);
 981         unsigned long count[MIGRATE_TYPES] = { 0, };
 982
 983         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
 984                 struct page *page;
 985
 986                 if (!pfn_valid(pfn))
 987                         continue;
 988
 989                 page = pfn_to_page(pfn);
 990
 991                 /* Watch for unexpected holes punched in the memmap */
 992                 if (!memmap_valid_within(pfn, page, zone))
 993                         continue;
 994
 995                 mtype = get_pageblock_migratetype(page);
 996
 997                 if (mtype < MIGRATE_TYPES)
 998                         count[mtype]++;
 999         }
1000
1001         /* Print counts */
1002         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1003         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1004                 seq_printf(m, "%12lu ", count[mtype]);
1005         seq_putc(m, '\n');
1006 }
1007
1008 /* Print out the free pages at each order for each migratetype */
1009 static int pagetypeinfo_showblockcount(struct seq_file *m, void *arg)
1010 {
1011         int mtype;
1012         pg_data_t *pgdat = (pg_data_t *)arg;
1013
1014         seq_printf(m, "\n%-23s", "Number of blocks type ");
1015         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1016                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1017         seq_putc(m, '\n');
1018         walk_zones_in_node(m, pgdat, pagetypeinfo_showblockcount_print);
1019
1020         return 0;
1021 }
1022
1023 #ifdef CONFIG_PAGE_OWNER
1024 static void pagetypeinfo_showmixedcount_print(struct seq_file *m,
1025                                                         pg_data_t *pgdat,
1026                                                         struct zone *zone)
1027 {
1028         struct page *page;
1029         struct page_ext *page_ext;
1030         unsigned long pfn = zone->zone_start_pfn, block_end_pfn;
1031         unsigned long end_pfn = pfn + zone->spanned_pages;
1032         unsigned long count[MIGRATE_TYPES] = { 0, };
1033         int pageblock_mt, page_mt;
1034         int i;
1035
1036         /* Scan block by block. First and last block may be incomplete */
1037         pfn = zone->zone_start_pfn;
1038
1039         /*
1040          * Walk the zone in pageblock_nr_pages steps. If a page block spans
1041          * a zone boundary, it will be double counted between zones. This does
1042          * not matter as the mixed block count will still be correct
1043          */
1044         for (; pfn < end_pfn; ) {
1045                 if (!pfn_valid(pfn)) {
1046                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
1047                         continue;
1048                 }
1049
1050                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
1051                 block_end_pfn = min(block_end_pfn, end_pfn);
1052
1053                 page = pfn_to_page(pfn);
1054                 pageblock_mt = get_pfnblock_migratetype(page, pfn);
1055
1056                 for (; pfn < block_end_pfn; pfn++) {
1057                         if (!pfn_valid_within(pfn))
1058                                 continue;
1059
1060                         page = pfn_to_page(pfn);
1061                         if (PageBuddy(page)) {
1062                                 pfn += (1UL << page_order(page)) - 1;
1063                                 continue;
1064                         }
1065
1066                         if (PageReserved(page))
1067                                 continue;
1068
1069                         page_ext = lookup_page_ext(page);
1070
1071                         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
1072                                 continue;
1073
1074                         page_mt = gfpflags_to_migratetype(page_ext->gfp_mask);
1075                         if (pageblock_mt != page_mt) {
1076                                 if (is_migrate_cma(pageblock_mt))
1077                                         count[MIGRATE_MOVABLE]++;
1078                                 else
1079                                         count[pageblock_mt]++;
1080
1081                                 pfn = block_end_pfn;
1082                                 break;
1083                         }
1084                         pfn += (1UL << page_ext->order) - 1;
1085                 }
1086         }
1087
1088         /* Print counts */
1089         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1090         for (i = 0; i < MIGRATE_TYPES; i++)
1091                 seq_printf(m, "%12lu ", count[i]);
1092         seq_putc(m, '\n');
1093 }
1094 #endif /* CONFIG_PAGE_OWNER */
1095
1096 /*
1097  * Print out the number of pageblocks for each migratetype that contain pages
1098  * of other types. This gives an indication of how well fallbacks are being
1099  * contained by rmqueue_fallback(). It requires information from PAGE_OWNER
1100  * to determine what is going on
1101  */
1102 static void pagetypeinfo_showmixedcount(struct seq_file *m, pg_data_t *pgdat)
1103 {
1104 #ifdef CONFIG_PAGE_OWNER
1105         int mtype;
1106
1107         if (!page_owner_inited)
1108                 return;
1109
1110         drain_all_pages(NULL);
1111
1112         seq_printf(m, "\n%-23s", "Number of mixed blocks ");
1113         for (mtype = 0; mtype < MIGRATE_TYPES; mtype++)
1114                 seq_printf(m, "%12s ", migratetype_names[mtype]);
1115         seq_putc(m, '\n');
1116
1117         walk_zones_in_node(m, pgdat, pagetypeinfo_showmixedcount_print);
1118 #endif /* CONFIG_PAGE_OWNER */
1119 }
1120
1121 /*
1122  * This prints out statistics in relation to grouping pages by mobility.
1123  * It is expensive to collect so do not constantly read the file.
1124  */
1125 static int pagetypeinfo_show(struct seq_file *m, void *arg)
1126 {
1127         pg_data_t *pgdat = (pg_data_t *)arg;
1128
1129         /* check memoryless node */
1130         if (!node_state(pgdat->node_id, N_MEMORY))
1131                 return 0;
1132
1133         seq_printf(m, "Page block order: %d\n", pageblock_order);
1134         seq_printf(m, "Pages per block:  %lu\n", pageblock_nr_pages);
1135         seq_putc(m, '\n');
1136         pagetypeinfo_showfree(m, pgdat);
1137         pagetypeinfo_showblockcount(m, pgdat);
1138         pagetypeinfo_showmixedcount(m, pgdat);
1139
1140         return 0;
1141 }
1142
1143 static const struct seq_operations fragmentation_op = {
1144         .start  = frag_start,
1145         .next   = frag_next,
1146         .stop   = frag_stop,
1147         .show   = frag_show,
1148 };
1149
1150 static int fragmentation_open(struct inode *inode, struct file *file)
1151 {
1152         return seq_open(file, &fragmentation_op);
1153 }
1154
1155 static const struct file_operations fragmentation_file_operations = {
1156         .open           = fragmentation_open,
1157         .read           = seq_read,
1158         .llseek         = seq_lseek,
1159         .release        = seq_release,
1160 };
1161
1162 static const struct seq_operations pagetypeinfo_op = {
1163         .start  = frag_start,
1164         .next   = frag_next,
1165         .stop   = frag_stop,
1166         .show   = pagetypeinfo_show,
1167 };
1168
1169 static int pagetypeinfo_open(struct inode *inode, struct file *file)
1170 {
1171         return seq_open(file, &pagetypeinfo_op);
1172 }
1173
1174 static const struct file_operations pagetypeinfo_file_ops = {
1175         .open           = pagetypeinfo_open,
1176         .read           = seq_read,
1177         .llseek         = seq_lseek,
1178         .release        = seq_release,
1179 };
1180
1181 static void zoneinfo_show_print(struct seq_file *m, pg_data_t *pgdat,
1182                                                         struct zone *zone)
1183 {
1184         int i;
1185         seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1186         seq_printf(m,
1187                    "\n  pages free     %lu"
1188                    "\n        min      %lu"
1189                    "\n        low      %lu"
1190                    "\n        high     %lu"
1191                    "\n        scanned  %lu"
1192                    "\n        spanned  %lu"
1193                    "\n        present  %lu"
1194                    "\n        managed  %lu",
1195                    zone_page_state(zone, NR_FREE_PAGES),
1196                    min_wmark_pages(zone),
1197                    low_wmark_pages(zone),
1198                    high_wmark_pages(zone),
1199                    zone_page_state(zone, NR_PAGES_SCANNED),
1200                    zone->spanned_pages,
1201                    zone->present_pages,
1202                    zone->managed_pages);
1203
1204         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1205                 seq_printf(m, "\n    %-12s %lu", vmstat_text[i],
1206                                 zone_page_state(zone, i));
1207
1208         seq_printf(m,
1209                    "\n        protection: (%ld",
1210                    zone->lowmem_reserve[0]);
1211         for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1212                 seq_printf(m, ", %ld", zone->lowmem_reserve[i]);
1213         seq_printf(m,
1214                    ")"
1215                    "\n  pagesets");
1216         for_each_online_cpu(i) {
1217                 struct per_cpu_pageset *pageset;
1218
1219                 pageset = per_cpu_ptr(zone->pageset, i);
1220                 seq_printf(m,
1221                            "\n    cpu: %i"
1222                            "\n              count: %i"
1223                            "\n              high:  %i"
1224                            "\n              batch: %i",
1225                            i,
1226                            pageset->pcp.count,
1227                            pageset->pcp.high,
1228                            pageset->pcp.batch);
1229 #ifdef CONFIG_SMP
1230                 seq_printf(m, "\n  vm stats threshold: %d",
1231                                 pageset->stat_threshold);
1232 #endif
1233         }
1234         seq_printf(m,
1235                    "\n  all_unreclaimable: %u"
1236                    "\n  start_pfn:         %lu"
1237                    "\n  inactive_ratio:    %u",
1238                    !zone_reclaimable(zone),
1239                    zone->zone_start_pfn,
1240                    zone->inactive_ratio);
1241         seq_putc(m, '\n');
1242 }
1243
1244 /*
1245  * Output information about zones in @pgdat.
1246  */
1247 static int zoneinfo_show(struct seq_file *m, void *arg)
1248 {
1249         pg_data_t *pgdat = (pg_data_t *)arg;
1250         walk_zones_in_node(m, pgdat, zoneinfo_show_print);
1251         return 0;
1252 }
1253
1254 static const struct seq_operations zoneinfo_op = {
1255         .start  = frag_start, /* iterate over all zones. The same as in
1256                                * fragmentation. */
1257         .next   = frag_next,
1258         .stop   = frag_stop,
1259         .show   = zoneinfo_show,
1260 };
1261
1262 static int zoneinfo_open(struct inode *inode, struct file *file)
1263 {
1264         return seq_open(file, &zoneinfo_op);
1265 }
1266
1267 static const struct file_operations proc_zoneinfo_file_operations = {
1268         .open           = zoneinfo_open,
1269         .read           = seq_read,
1270         .llseek         = seq_lseek,
1271         .release        = seq_release,
1272 };
1273
1274 enum writeback_stat_item {
1275         NR_DIRTY_THRESHOLD,
1276         NR_DIRTY_BG_THRESHOLD,
1277         NR_VM_WRITEBACK_STAT_ITEMS,
1278 };
1279
1280 static void *vmstat_start(struct seq_file *m, loff_t *pos)
1281 {
1282         unsigned long *v;
1283         int i, stat_items_size;
1284
1285         if (*pos >= ARRAY_SIZE(vmstat_text))
1286                 return NULL;
1287         stat_items_size = NR_VM_ZONE_STAT_ITEMS * sizeof(unsigned long) +
1288                           NR_VM_WRITEBACK_STAT_ITEMS * sizeof(unsigned long);
1289
1290 #ifdef CONFIG_VM_EVENT_COUNTERS
1291         stat_items_size += sizeof(struct vm_event_state);
1292 #endif
1293
1294         v = kmalloc(stat_items_size, GFP_KERNEL);
1295         m->private = v;
1296         if (!v)
1297                 return ERR_PTR(-ENOMEM);
1298         for (i = 0; i < NR_VM_ZONE_STAT_ITEMS; i++)
1299                 v[i] = global_page_state(i);
1300         v += NR_VM_ZONE_STAT_ITEMS;
1301
1302         global_dirty_limits(v + NR_DIRTY_BG_THRESHOLD,
1303                             v + NR_DIRTY_THRESHOLD);
1304         v += NR_VM_WRITEBACK_STAT_ITEMS;
1305
1306 #ifdef CONFIG_VM_EVENT_COUNTERS
1307         all_vm_events(v);
1308         v[PGPGIN] /= 2;         /* sectors -> kbytes */
1309         v[PGPGOUT] /= 2;
1310 #endif
1311         return (unsigned long *)m->private + *pos;
1312 }
1313
1314 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
1315 {
1316         (*pos)++;
1317         if (*pos >= ARRAY_SIZE(vmstat_text))
1318                 return NULL;
1319         return (unsigned long *)m->private + *pos;
1320 }
1321
1322 static int vmstat_show(struct seq_file *m, void *arg)
1323 {
1324         unsigned long *l = arg;
1325         unsigned long off = l - (unsigned long *)m->private;
1326
1327         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
1328         return 0;
1329 }
1330
1331 static void vmstat_stop(struct seq_file *m, void *arg)
1332 {
1333         kfree(m->private);
1334         m->private = NULL;
1335 }
1336
1337 static const struct seq_operations vmstat_op = {
1338         .start  = vmstat_start,
1339         .next   = vmstat_next,
1340         .stop   = vmstat_stop,
1341         .show   = vmstat_show,
1342 };
1343
1344 static int vmstat_open(struct inode *inode, struct file *file)
1345 {
1346         return seq_open(file, &vmstat_op);
1347 }
1348
1349 static const struct file_operations proc_vmstat_file_operations = {
1350         .open           = vmstat_open,
1351         .read           = seq_read,
1352         .llseek         = seq_lseek,
1353         .release        = seq_release,
1354 };
1355 #endif /* CONFIG_PROC_FS */
1356
1357 #ifdef CONFIG_SMP
1358 static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
1359 int sysctl_stat_interval __read_mostly = HZ;
1360 static cpumask_var_t cpu_stat_off;
1361
1362 static void vmstat_update(struct work_struct *w)
1363 {
1364         if (refresh_cpu_vm_stats())
1365                 /*
1366                  * Counters were updated so we expect more updates
1367                  * to occur in the future. Keep on running the
1368                  * update worker thread.
1369                  */
1370                 schedule_delayed_work(this_cpu_ptr(&vmstat_work),
1371                         round_jiffies_relative(sysctl_stat_interval));
1372         else {
1373                 /*
1374                  * We did not update any counters so the app may be in
1375                  * a mode where it does not cause counter updates.
1376                  * We may be uselessly running vmstat_update.
1377                  * Defer the checking for differentials to the
1378                  * shepherd thread on a different processor.
1379                  */
1380                 int r;
1381                 /*
1382                  * Shepherd work thread does not race since it never
1383                  * changes the bit if its zero but the cpu
1384                  * online / off line code may race if
1385                  * worker threads are still allowed during
1386                  * shutdown / startup.
1387                  */
1388                 r = cpumask_test_and_set_cpu(smp_processor_id(),
1389                         cpu_stat_off);
1390                 VM_BUG_ON(r);
1391         }
1392 }
1393
1394 /*
1395  * Check if the diffs for a certain cpu indicate that
1396  * an update is needed.
1397  */
1398 static bool need_update(int cpu)
1399 {
1400         struct zone *zone;
1401
1402         for_each_populated_zone(zone) {
1403                 struct per_cpu_pageset *p = per_cpu_ptr(zone->pageset, cpu);
1404
1405                 BUILD_BUG_ON(sizeof(p->vm_stat_diff[0]) != 1);
1406                 /*
1407                  * The fast way of checking if there are any vmstat diffs.
1408                  * This works because the diffs are byte sized items.
1409                  */
1410                 if (memchr_inv(p->vm_stat_diff, 0, NR_VM_ZONE_STAT_ITEMS))
1411                         return true;
1412
1413         }
1414         return false;
1415 }
1416
1417
1418 /*
1419  * Shepherd worker thread that checks the
1420  * differentials of processors that have their worker
1421  * threads for vm statistics updates disabled because of
1422  * inactivity.
1423  */
1424 static void vmstat_shepherd(struct work_struct *w);
1425
1426 static DECLARE_DELAYED_WORK(shepherd, vmstat_shepherd);
1427
1428 static void vmstat_shepherd(struct work_struct *w)
1429 {
1430         int cpu;
1431
1432         get_online_cpus();
1433         /* Check processors whose vmstat worker threads have been disabled */
1434         for_each_cpu(cpu, cpu_stat_off)
1435                 if (need_update(cpu) &&
1436                         cpumask_test_and_clear_cpu(cpu, cpu_stat_off))
1437
1438                         schedule_delayed_work_on(cpu, &per_cpu(vmstat_work, cpu),
1439                                 __round_jiffies_relative(sysctl_stat_interval, cpu));
1440
1441         put_online_cpus();
1442
1443         schedule_delayed_work(&shepherd,
1444                 round_jiffies_relative(sysctl_stat_interval));
1445
1446 }
1447
1448 static void __init start_shepherd_timer(void)
1449 {
1450         int cpu;
1451
1452         for_each_possible_cpu(cpu)
1453                 INIT_DEFERRABLE_WORK(per_cpu_ptr(&vmstat_work, cpu),
1454                         vmstat_update);
1455
1456         if (!alloc_cpumask_var(&cpu_stat_off, GFP_KERNEL))
1457                 BUG();
1458         cpumask_copy(cpu_stat_off, cpu_online_mask);
1459
1460         schedule_delayed_work(&shepherd,
1461                 round_jiffies_relative(sysctl_stat_interval));
1462 }
1463
1464 static void vmstat_cpu_dead(int node)
1465 {
1466         int cpu;
1467
1468         get_online_cpus();
1469         for_each_online_cpu(cpu)
1470                 if (cpu_to_node(cpu) == node)
1471                         goto end;
1472
1473         node_clear_state(node, N_CPU);
1474 end:
1475         put_online_cpus();
1476 }
1477
1478 /*
1479  * Use the cpu notifier to insure that the thresholds are recalculated
1480  * when necessary.
1481  */
1482 static int vmstat_cpuup_callback(struct notifier_block *nfb,
1483                 unsigned long action,
1484                 void *hcpu)
1485 {
1486         long cpu = (long)hcpu;
1487
1488         switch (action) {
1489         case CPU_ONLINE:
1490         case CPU_ONLINE_FROZEN:
1491                 refresh_zone_stat_thresholds();
1492                 node_set_state(cpu_to_node(cpu), N_CPU);
1493                 cpumask_set_cpu(cpu, cpu_stat_off);
1494                 break;
1495         case CPU_DOWN_PREPARE:
1496         case CPU_DOWN_PREPARE_FROZEN:
1497                 cancel_delayed_work_sync(&per_cpu(vmstat_work, cpu));
1498                 cpumask_clear_cpu(cpu, cpu_stat_off);
1499                 break;
1500         case CPU_DOWN_FAILED:
1501         case CPU_DOWN_FAILED_FROZEN:
1502                 cpumask_set_cpu(cpu, cpu_stat_off);
1503                 break;
1504         case CPU_DEAD:
1505         case CPU_DEAD_FROZEN:
1506                 refresh_zone_stat_thresholds();
1507                 vmstat_cpu_dead(cpu_to_node(cpu));
1508                 break;
1509         default:
1510                 break;
1511         }
1512         return NOTIFY_OK;
1513 }
1514
1515 static struct notifier_block vmstat_notifier =
1516         { &vmstat_cpuup_callback, NULL, 0 };
1517 #endif
1518
1519 static int __init setup_vmstat(void)
1520 {
1521 #ifdef CONFIG_SMP
1522         cpu_notifier_register_begin();
1523         __register_cpu_notifier(&vmstat_notifier);
1524
1525         start_shepherd_timer();
1526         cpu_notifier_register_done();
1527 #endif
1528 #ifdef CONFIG_PROC_FS
1529         proc_create("buddyinfo", S_IRUGO, NULL, &fragmentation_file_operations);
1530         proc_create("pagetypeinfo", S_IRUGO, NULL, &pagetypeinfo_file_ops);
1531         proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
1532         proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
1533 #endif
1534         return 0;
1535 }
1536 module_init(setup_vmstat)
1537
1538 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_COMPACTION)
1539 #include <linux/debugfs.h>
1540
1541
1542 /*
1543  * Return an index indicating how much of the available free memory is
1544  * unusable for an allocation of the requested size.
1545  */
1546 static int unusable_free_index(unsigned int order,
1547                                 struct contig_page_info *info)
1548 {
1549         /* No free memory is interpreted as all free memory is unusable */
1550         if (info->free_pages == 0)
1551                 return 1000;
1552
1553         /*
1554          * Index should be a value between 0 and 1. Return a value to 3
1555          * decimal places.
1556          *
1557          * 0 => no fragmentation
1558          * 1 => high fragmentation
1559          */
1560         return div_u64((info->free_pages - (info->free_blocks_suitable << order)) * 1000ULL, info->free_pages);
1561
1562 }
1563
1564 static void unusable_show_print(struct seq_file *m,
1565                                         pg_data_t *pgdat, struct zone *zone)
1566 {
1567         unsigned int order;
1568         int index;
1569         struct contig_page_info info;
1570
1571         seq_printf(m, "Node %d, zone %8s ",
1572                                 pgdat->node_id,
1573                                 zone->name);
1574         for (order = 0; order < MAX_ORDER; ++order) {
1575                 fill_contig_page_info(zone, order, &info);
1576                 index = unusable_free_index(order, &info);
1577                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1578         }
1579
1580         seq_putc(m, '\n');
1581 }
1582
1583 /*
1584  * Display unusable free space index
1585  *
1586  * The unusable free space index measures how much of the available free
1587  * memory cannot be used to satisfy an allocation of a given size and is a
1588  * value between 0 and 1. The higher the value, the more of free memory is
1589  * unusable and by implication, the worse the external fragmentation is. This
1590  * can be expressed as a percentage by multiplying by 100.
1591  */
1592 static int unusable_show(struct seq_file *m, void *arg)
1593 {
1594         pg_data_t *pgdat = (pg_data_t *)arg;
1595
1596         /* check memoryless node */
1597         if (!node_state(pgdat->node_id, N_MEMORY))
1598                 return 0;
1599
1600         walk_zones_in_node(m, pgdat, unusable_show_print);
1601
1602         return 0;
1603 }
1604
1605 static const struct seq_operations unusable_op = {
1606         .start  = frag_start,
1607         .next   = frag_next,
1608         .stop   = frag_stop,
1609         .show   = unusable_show,
1610 };
1611
1612 static int unusable_open(struct inode *inode, struct file *file)
1613 {
1614         return seq_open(file, &unusable_op);
1615 }
1616
1617 static const struct file_operations unusable_file_ops = {
1618         .open           = unusable_open,
1619         .read           = seq_read,
1620         .llseek         = seq_lseek,
1621         .release        = seq_release,
1622 };
1623
1624 static void extfrag_show_print(struct seq_file *m,
1625                                         pg_data_t *pgdat, struct zone *zone)
1626 {
1627         unsigned int order;
1628         int index;
1629
1630         /* Alloc on stack as interrupts are disabled for zone walk */
1631         struct contig_page_info info;
1632
1633         seq_printf(m, "Node %d, zone %8s ",
1634                                 pgdat->node_id,
1635                                 zone->name);
1636         for (order = 0; order < MAX_ORDER; ++order) {
1637                 fill_contig_page_info(zone, order, &info);
1638                 index = __fragmentation_index(order, &info);
1639                 seq_printf(m, "%d.%03d ", index / 1000, index % 1000);
1640         }
1641
1642         seq_putc(m, '\n');
1643 }
1644
1645 /*
1646  * Display fragmentation index for orders that allocations would fail for
1647  */
1648 static int extfrag_show(struct seq_file *m, void *arg)
1649 {
1650         pg_data_t *pgdat = (pg_data_t *)arg;
1651
1652         walk_zones_in_node(m, pgdat, extfrag_show_print);
1653
1654         return 0;
1655 }
1656
1657 static const struct seq_operations extfrag_op = {
1658         .start  = frag_start,
1659         .next   = frag_next,
1660         .stop   = frag_stop,
1661         .show   = extfrag_show,
1662 };
1663
1664 static int extfrag_open(struct inode *inode, struct file *file)
1665 {
1666         return seq_open(file, &extfrag_op);
1667 }
1668
1669 static const struct file_operations extfrag_file_ops = {
1670         .open           = extfrag_open,
1671         .read           = seq_read,
1672         .llseek         = seq_lseek,
1673         .release        = seq_release,
1674 };
1675
1676 static int __init extfrag_debug_init(void)
1677 {
1678         struct dentry *extfrag_debug_root;
1679
1680         extfrag_debug_root = debugfs_create_dir("extfrag", NULL);
1681         if (!extfrag_debug_root)
1682                 return -ENOMEM;
1683
1684         if (!debugfs_create_file("unusable_index", 0444,
1685                         extfrag_debug_root, NULL, &unusable_file_ops))
1686                 goto fail;
1687
1688         if (!debugfs_create_file("extfrag_index", 0444,
1689                         extfrag_debug_root, NULL, &extfrag_file_ops))
1690                 goto fail;
1691
1692         return 0;
1693 fail:
1694         debugfs_remove_recursive(extfrag_debug_root);
1695         return -ENOMEM;
1696 }
1697
1698 module_init(extfrag_debug_init);
1699 #endif