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