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