drivers/md/dm-stats.c

   1 #include <linux/errno.h>
   2 #include <linux/numa.h>
   3 #include <linux/slab.h>
   4 #include <linux/rculist.h>
   5 #include <linux/threads.h>
   6 #include <linux/preempt.h>
   7 #include <linux/irqflags.h>
   8 #include <linux/vmalloc.h>
   9 #include <linux/mm.h>
  10 #include <linux/module.h>
  11 #include <linux/device-mapper.h>
  12
  13 #include "dm.h"
  14 #include "dm-stats.h"
  15
  16 #define DM_MSG_PREFIX "stats"
  17
  18 static int dm_stat_need_rcu_barrier;
  19
  20 /*
  21  * Using 64-bit values to avoid overflow (which is a
  22  * problem that block/genhd.c's IO accounting has).
  23  */
  24 struct dm_stat_percpu {
  25         unsigned long long sectors[2];
  26         unsigned long long ios[2];
  27         unsigned long long merges[2];
  28         unsigned long long ticks[2];
  29         unsigned long long io_ticks[2];
  30         unsigned long long io_ticks_total;
  31         unsigned long long time_in_queue;
  32 };
  33
  34 struct dm_stat_shared {
  35         atomic_t in_flight[2];
  36         unsigned long stamp;
  37         struct dm_stat_percpu tmp;
  38 };
  39
  40 struct dm_stat {
  41         struct list_head list_entry;
  42         int id;
  43         size_t n_entries;
  44         sector_t start;
  45         sector_t end;
  46         sector_t step;
  47         const char *program_id;
  48         const char *aux_data;
  49         struct rcu_head rcu_head;
  50         size_t shared_alloc_size;
  51         size_t percpu_alloc_size;
  52         struct dm_stat_percpu *stat_percpu[NR_CPUS];
  53         struct dm_stat_shared stat_shared[0];
  54 };
  55
  56 struct dm_stats_last_position {
  57         sector_t last_sector;
  58         unsigned last_rw;
  59 };
  60
  61 /*
  62  * A typo on the command line could possibly make the kernel run out of memory
  63  * and crash. To prevent the crash we account all used memory. We fail if we
  64  * exhaust 1/4 of all memory or 1/2 of vmalloc space.
  65  */
  66 #define DM_STATS_MEMORY_FACTOR          4
  67 #define DM_STATS_VMALLOC_FACTOR         2
  68
  69 static DEFINE_SPINLOCK(shared_memory_lock);
  70
  71 static unsigned long shared_memory_amount;
  72
  73 static bool __check_shared_memory(size_t alloc_size)
  74 {
  75         size_t a;
  76
  77         a = shared_memory_amount + alloc_size;
  78         if (a < shared_memory_amount)
  79                 return false;
  80         if (a >> PAGE_SHIFT > totalram_pages / DM_STATS_MEMORY_FACTOR)
  81                 return false;
  82 #ifdef CONFIG_MMU
  83         if (a > (VMALLOC_END - VMALLOC_START) / DM_STATS_VMALLOC_FACTOR)
  84                 return false;
  85 #endif
  86         return true;
  87 }
  88
  89 static bool check_shared_memory(size_t alloc_size)
  90 {
  91         bool ret;
  92
  93         spin_lock_irq(&shared_memory_lock);
  94
  95         ret = __check_shared_memory(alloc_size);
  96
  97         spin_unlock_irq(&shared_memory_lock);
  98
  99         return ret;
 100 }
 101
 102 static bool claim_shared_memory(size_t alloc_size)
 103 {
 104         spin_lock_irq(&shared_memory_lock);
 105
 106         if (!__check_shared_memory(alloc_size)) {
 107                 spin_unlock_irq(&shared_memory_lock);
 108                 return false;
 109         }
 110
 111         shared_memory_amount += alloc_size;
 112
 113         spin_unlock_irq(&shared_memory_lock);
 114
 115         return true;
 116 }
 117
 118 static void free_shared_memory(size_t alloc_size)
 119 {
 120         unsigned long flags;
 121
 122         spin_lock_irqsave(&shared_memory_lock, flags);
 123
 124         if (WARN_ON_ONCE(shared_memory_amount < alloc_size)) {
 125                 spin_unlock_irqrestore(&shared_memory_lock, flags);
 126                 DMCRIT("Memory usage accounting bug.");
 127                 return;
 128         }
 129
 130         shared_memory_amount -= alloc_size;
 131
 132         spin_unlock_irqrestore(&shared_memory_lock, flags);
 133 }
 134
 135 static void *dm_kvzalloc(size_t alloc_size, int node)
 136 {
 137         void *p;
 138
 139         if (!claim_shared_memory(alloc_size))
 140                 return NULL;
 141
 142         if (alloc_size <= KMALLOC_MAX_SIZE) {
 143                 p = kzalloc_node(alloc_size, GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN, node);
 144                 if (p)
 145                         return p;
 146         }
 147         p = vzalloc_node(alloc_size, node);
 148         if (p)
 149                 return p;
 150
 151         free_shared_memory(alloc_size);
 152
 153         return NULL;
 154 }
 155
 156 static void dm_kvfree(void *ptr, size_t alloc_size)
 157 {
 158         if (!ptr)
 159                 return;
 160
 161         free_shared_memory(alloc_size);
 162
 163         kvfree(ptr);
 164 }
 165
 166 static void dm_stat_free(struct rcu_head *head)
 167 {
 168         int cpu;
 169         struct dm_stat *s = container_of(head, struct dm_stat, rcu_head);
 170
 171         kfree(s->program_id);
 172         kfree(s->aux_data);
 173         for_each_possible_cpu(cpu)
 174                 dm_kvfree(s->stat_percpu[cpu], s->percpu_alloc_size);
 175         dm_kvfree(s, s->shared_alloc_size);
 176 }
 177
 178 static int dm_stat_in_flight(struct dm_stat_shared *shared)
 179 {
 180         return atomic_read(&shared->in_flight[READ]) +
 181                atomic_read(&shared->in_flight[WRITE]);
 182 }
 183
 184 void dm_stats_init(struct dm_stats *stats)
 185 {
 186         int cpu;
 187         struct dm_stats_last_position *last;
 188
 189         mutex_init(&stats->mutex);
 190         INIT_LIST_HEAD(&stats->list);
 191         stats->last = alloc_percpu(struct dm_stats_last_position);
 192         for_each_possible_cpu(cpu) {
 193                 last = per_cpu_ptr(stats->last, cpu);
 194                 last->last_sector = (sector_t)ULLONG_MAX;
 195                 last->last_rw = UINT_MAX;
 196         }
 197 }
 198
 199 void dm_stats_cleanup(struct dm_stats *stats)
 200 {
 201         size_t ni;
 202         struct dm_stat *s;
 203         struct dm_stat_shared *shared;
 204
 205         while (!list_empty(&stats->list)) {
 206                 s = container_of(stats->list.next, struct dm_stat, list_entry);
 207                 list_del(&s->list_entry);
 208                 for (ni = 0; ni < s->n_entries; ni++) {
 209                         shared = &s->stat_shared[ni];
 210                         if (WARN_ON(dm_stat_in_flight(shared))) {
 211                                 DMCRIT("leaked in-flight counter at index %lu "
 212                                        "(start %llu, end %llu, step %llu): reads %d, writes %d",
 213                                        (unsigned long)ni,
 214                                        (unsigned long long)s->start,
 215                                        (unsigned long long)s->end,
 216                                        (unsigned long long)s->step,
 217                                        atomic_read(&shared->in_flight[READ]),
 218                                        atomic_read(&shared->in_flight[WRITE]));
 219                         }
 220                 }
 221                 dm_stat_free(&s->rcu_head);
 222         }
 223         free_percpu(stats->last);
 224 }
 225
 226 static int dm_stats_create(struct dm_stats *stats, sector_t start, sector_t end,
 227                            sector_t step, const char *program_id, const char *aux_data,
 228                            void (*suspend_callback)(struct mapped_device *),
 229                            void (*resume_callback)(struct mapped_device *),
 230                            struct mapped_device *md)
 231 {
 232         struct list_head *l;
 233         struct dm_stat *s, *tmp_s;
 234         sector_t n_entries;
 235         size_t ni;
 236         size_t shared_alloc_size;
 237         size_t percpu_alloc_size;
 238         struct dm_stat_percpu *p;
 239         int cpu;
 240         int ret_id;
 241         int r;
 242
 243         if (end < start || !step)
 244                 return -EINVAL;
 245
 246         n_entries = end - start;
 247         if (dm_sector_div64(n_entries, step))
 248                 n_entries++;
 249
 250         if (n_entries != (size_t)n_entries || !(size_t)(n_entries + 1))
 251                 return -EOVERFLOW;
 252
 253         shared_alloc_size = sizeof(struct dm_stat) + (size_t)n_entries * sizeof(struct dm_stat_shared);
 254         if ((shared_alloc_size - sizeof(struct dm_stat)) / sizeof(struct dm_stat_shared) != n_entries)
 255                 return -EOVERFLOW;
 256
 257         percpu_alloc_size = (size_t)n_entries * sizeof(struct dm_stat_percpu);
 258         if (percpu_alloc_size / sizeof(struct dm_stat_percpu) != n_entries)
 259                 return -EOVERFLOW;
 260
 261         if (!check_shared_memory(shared_alloc_size + num_possible_cpus() * percpu_alloc_size))
 262                 return -ENOMEM;
 263
 264         s = dm_kvzalloc(shared_alloc_size, NUMA_NO_NODE);
 265         if (!s)
 266                 return -ENOMEM;
 267
 268         s->n_entries = n_entries;
 269         s->start = start;
 270         s->end = end;
 271         s->step = step;
 272         s->shared_alloc_size = shared_alloc_size;
 273         s->percpu_alloc_size = percpu_alloc_size;
 274
 275         s->program_id = kstrdup(program_id, GFP_KERNEL);
 276         if (!s->program_id) {
 277                 r = -ENOMEM;
 278                 goto out;
 279         }
 280         s->aux_data = kstrdup(aux_data, GFP_KERNEL);
 281         if (!s->aux_data) {
 282                 r = -ENOMEM;
 283                 goto out;
 284         }
 285
 286         for (ni = 0; ni < n_entries; ni++) {
 287                 atomic_set(&s->stat_shared[ni].in_flight[READ], 0);
 288                 atomic_set(&s->stat_shared[ni].in_flight[WRITE], 0);
 289         }
 290
 291         for_each_possible_cpu(cpu) {
 292                 p = dm_kvzalloc(percpu_alloc_size, cpu_to_node(cpu));
 293                 if (!p) {
 294                         r = -ENOMEM;
 295                         goto out;
 296                 }
 297                 s->stat_percpu[cpu] = p;
 298         }
 299
 300         /*
 301          * Suspend/resume to make sure there is no i/o in flight,
 302          * so that newly created statistics will be exact.
 303          *
 304          * (note: we couldn't suspend earlier because we must not
 305          * allocate memory while suspended)
 306          */
 307         suspend_callback(md);
 308
 309         mutex_lock(&stats->mutex);
 310         s->id = 0;
 311         list_for_each(l, &stats->list) {
 312                 tmp_s = container_of(l, struct dm_stat, list_entry);
 313                 if (WARN_ON(tmp_s->id < s->id)) {
 314                         r = -EINVAL;
 315                         goto out_unlock_resume;
 316                 }
 317                 if (tmp_s->id > s->id)
 318                         break;
 319                 if (unlikely(s->id == INT_MAX)) {
 320                         r = -ENFILE;
 321                         goto out_unlock_resume;
 322                 }
 323                 s->id++;
 324         }
 325         ret_id = s->id;
 326         list_add_tail_rcu(&s->list_entry, l);
 327         mutex_unlock(&stats->mutex);
 328
 329         resume_callback(md);
 330
 331         return ret_id;
 332
 333 out_unlock_resume:
 334         mutex_unlock(&stats->mutex);
 335         resume_callback(md);
 336 out:
 337         dm_stat_free(&s->rcu_head);
 338         return r;
 339 }
 340
 341 static struct dm_stat *__dm_stats_find(struct dm_stats *stats, int id)
 342 {
 343         struct dm_stat *s;
 344
 345         list_for_each_entry(s, &stats->list, list_entry) {
 346                 if (s->id > id)
 347                         break;
 348                 if (s->id == id)
 349                         return s;
 350         }
 351
 352         return NULL;
 353 }
 354
 355 static int dm_stats_delete(struct dm_stats *stats, int id)
 356 {
 357         struct dm_stat *s;
 358         int cpu;
 359
 360         mutex_lock(&stats->mutex);
 361
 362         s = __dm_stats_find(stats, id);
 363         if (!s) {
 364                 mutex_unlock(&stats->mutex);
 365                 return -ENOENT;
 366         }
 367
 368         list_del_rcu(&s->list_entry);
 369         mutex_unlock(&stats->mutex);
 370
 371         /*
 372          * vfree can't be called from RCU callback
 373          */
 374         for_each_possible_cpu(cpu)
 375                 if (is_vmalloc_addr(s->stat_percpu))
 376                         goto do_sync_free;
 377         if (is_vmalloc_addr(s)) {
 378 do_sync_free:
 379                 synchronize_rcu_expedited();
 380                 dm_stat_free(&s->rcu_head);
 381         } else {
 382                 ACCESS_ONCE(dm_stat_need_rcu_barrier) = 1;
 383                 call_rcu(&s->rcu_head, dm_stat_free);
 384         }
 385         return 0;
 386 }
 387
 388 static int dm_stats_list(struct dm_stats *stats, const char *program,
 389                          char *result, unsigned maxlen)
 390 {
 391         struct dm_stat *s;
 392         sector_t len;
 393         unsigned sz = 0;
 394
 395         /*
 396          * Output format:
 397          *   <region_id>: <start_sector>+<length> <step> <program_id> <aux_data>
 398          */
 399
 400         mutex_lock(&stats->mutex);
 401         list_for_each_entry(s, &stats->list, list_entry) {
 402                 if (!program || !strcmp(program, s->program_id)) {
 403                         len = s->end - s->start;
 404                         DMEMIT("%d: %llu+%llu %llu %s %s\n", s->id,
 405                                 (unsigned long long)s->start,
 406                                 (unsigned long long)len,
 407                                 (unsigned long long)s->step,
 408                                 s->program_id,
 409                                 s->aux_data);
 410                 }
 411         }
 412         mutex_unlock(&stats->mutex);
 413
 414         return 1;
 415 }
 416
 417 static void dm_stat_round(struct dm_stat_shared *shared, struct dm_stat_percpu *p)
 418 {
 419         /*
 420          * This is racy, but so is part_round_stats_single.
 421          */
 422         unsigned long now = jiffies;
 423         unsigned in_flight_read;
 424         unsigned in_flight_write;
 425         unsigned long difference = now - shared->stamp;
 426
 427         if (!difference)
 428                 return;
 429         in_flight_read = (unsigned)atomic_read(&shared->in_flight[READ]);
 430         in_flight_write = (unsigned)atomic_read(&shared->in_flight[WRITE]);
 431         if (in_flight_read)
 432                 p->io_ticks[READ] += difference;
 433         if (in_flight_write)
 434                 p->io_ticks[WRITE] += difference;
 435         if (in_flight_read + in_flight_write) {
 436                 p->io_ticks_total += difference;
 437                 p->time_in_queue += (in_flight_read + in_flight_write) * difference;
 438         }
 439         shared->stamp = now;
 440 }
 441
 442 static void dm_stat_for_entry(struct dm_stat *s, size_t entry,
 443                               unsigned long bi_rw, sector_t len, bool merged,
 444                               bool end, unsigned long duration)
 445 {
 446         unsigned long idx = bi_rw & REQ_WRITE;
 447         struct dm_stat_shared *shared = &s->stat_shared[entry];
 448         struct dm_stat_percpu *p;
 449
 450         /*
 451          * For strict correctness we should use local_irq_save/restore
 452          * instead of preempt_disable/enable.
 453          *
 454          * preempt_disable/enable is racy if the driver finishes bios
 455          * from non-interrupt context as well as from interrupt context
 456          * or from more different interrupts.
 457          *
 458          * On 64-bit architectures the race only results in not counting some
 459          * events, so it is acceptable.  On 32-bit architectures the race could
 460          * cause the counter going off by 2^32, so we need to do proper locking
 461          * there.
 462          *
 463          * part_stat_lock()/part_stat_unlock() have this race too.
 464          */
 465 #if BITS_PER_LONG == 32
 466         unsigned long flags;
 467         local_irq_save(flags);
 468 #else
 469         preempt_disable();
 470 #endif
 471         p = &s->stat_percpu[smp_processor_id()][entry];
 472
 473         if (!end) {
 474                 dm_stat_round(shared, p);
 475                 atomic_inc(&shared->in_flight[idx]);
 476         } else {
 477                 dm_stat_round(shared, p);
 478                 atomic_dec(&shared->in_flight[idx]);
 479                 p->sectors[idx] += len;
 480                 p->ios[idx] += 1;
 481                 p->merges[idx] += merged;
 482                 p->ticks[idx] += duration;
 483         }
 484
 485 #if BITS_PER_LONG == 32
 486         local_irq_restore(flags);
 487 #else
 488         preempt_enable();
 489 #endif
 490 }
 491
 492 static void __dm_stat_bio(struct dm_stat *s, unsigned long bi_rw,
 493                           sector_t bi_sector, sector_t end_sector,
 494                           bool end, unsigned long duration,
 495                           struct dm_stats_aux *stats_aux)
 496 {
 497         sector_t rel_sector, offset, todo, fragment_len;
 498         size_t entry;
 499
 500         if (end_sector <= s->start || bi_sector >= s->end)
 501                 return;
 502         if (unlikely(bi_sector < s->start)) {
 503                 rel_sector = 0;
 504                 todo = end_sector - s->start;
 505         } else {
 506                 rel_sector = bi_sector - s->start;
 507                 todo = end_sector - bi_sector;
 508         }
 509         if (unlikely(end_sector > s->end))
 510                 todo -= (end_sector - s->end);
 511
 512         offset = dm_sector_div64(rel_sector, s->step);
 513         entry = rel_sector;
 514         do {
 515                 if (WARN_ON_ONCE(entry >= s->n_entries)) {
 516                         DMCRIT("Invalid area access in region id %d", s->id);
 517                         return;
 518                 }
 519                 fragment_len = todo;
 520                 if (fragment_len > s->step - offset)
 521                         fragment_len = s->step - offset;
 522                 dm_stat_for_entry(s, entry, bi_rw, fragment_len,
 523                                   stats_aux->merged, end, duration);
 524                 todo -= fragment_len;
 525                 entry++;
 526                 offset = 0;
 527         } while (unlikely(todo != 0));
 528 }
 529
 530 void dm_stats_account_io(struct dm_stats *stats, unsigned long bi_rw,
 531                          sector_t bi_sector, unsigned bi_sectors, bool end,
 532                          unsigned long duration, struct dm_stats_aux *stats_aux)
 533 {
 534         struct dm_stat *s;
 535         sector_t end_sector;
 536         struct dm_stats_last_position *last;
 537
 538         if (unlikely(!bi_sectors))
 539                 return;
 540
 541         end_sector = bi_sector + bi_sectors;
 542
 543         if (!end) {
 544                 /*
 545                  * A race condition can at worst result in the merged flag being
 546                  * misrepresented, so we don't have to disable preemption here.
 547                  */
 548                 last = raw_cpu_ptr(stats->last);
 549                 stats_aux->merged =
 550                         (bi_sector == (ACCESS_ONCE(last->last_sector) &&
 551                                        ((bi_rw & (REQ_WRITE | REQ_DISCARD)) ==
 552                                         (ACCESS_ONCE(last->last_rw) & (REQ_WRITE | REQ_DISCARD)))
 553                                        ));
 554                 ACCESS_ONCE(last->last_sector) = end_sector;
 555                 ACCESS_ONCE(last->last_rw) = bi_rw;
 556         }
 557
 558         rcu_read_lock();
 559
 560         list_for_each_entry_rcu(s, &stats->list, list_entry)
 561                 __dm_stat_bio(s, bi_rw, bi_sector, end_sector, end, duration, stats_aux);
 562
 563         rcu_read_unlock();
 564 }
 565
 566 static void __dm_stat_init_temporary_percpu_totals(struct dm_stat_shared *shared,
 567                                                    struct dm_stat *s, size_t x)
 568 {
 569         int cpu;
 570         struct dm_stat_percpu *p;
 571
 572         local_irq_disable();
 573         p = &s->stat_percpu[smp_processor_id()][x];
 574         dm_stat_round(shared, p);
 575         local_irq_enable();
 576
 577         memset(&shared->tmp, 0, sizeof(shared->tmp));
 578         for_each_possible_cpu(cpu) {
 579                 p = &s->stat_percpu[cpu][x];
 580                 shared->tmp.sectors[READ] += ACCESS_ONCE(p->sectors[READ]);
 581                 shared->tmp.sectors[WRITE] += ACCESS_ONCE(p->sectors[WRITE]);
 582                 shared->tmp.ios[READ] += ACCESS_ONCE(p->ios[READ]);
 583                 shared->tmp.ios[WRITE] += ACCESS_ONCE(p->ios[WRITE]);
 584                 shared->tmp.merges[READ] += ACCESS_ONCE(p->merges[READ]);
 585                 shared->tmp.merges[WRITE] += ACCESS_ONCE(p->merges[WRITE]);
 586                 shared->tmp.ticks[READ] += ACCESS_ONCE(p->ticks[READ]);
 587                 shared->tmp.ticks[WRITE] += ACCESS_ONCE(p->ticks[WRITE]);
 588                 shared->tmp.io_ticks[READ] += ACCESS_ONCE(p->io_ticks[READ]);
 589                 shared->tmp.io_ticks[WRITE] += ACCESS_ONCE(p->io_ticks[WRITE]);
 590                 shared->tmp.io_ticks_total += ACCESS_ONCE(p->io_ticks_total);
 591                 shared->tmp.time_in_queue += ACCESS_ONCE(p->time_in_queue);
 592         }
 593 }
 594
 595 static void __dm_stat_clear(struct dm_stat *s, size_t idx_start, size_t idx_end,
 596                             bool init_tmp_percpu_totals)
 597 {
 598         size_t x;
 599         struct dm_stat_shared *shared;
 600         struct dm_stat_percpu *p;
 601
 602         for (x = idx_start; x < idx_end; x++) {
 603                 shared = &s->stat_shared[x];
 604                 if (init_tmp_percpu_totals)
 605                         __dm_stat_init_temporary_percpu_totals(shared, s, x);
 606                 local_irq_disable();
 607                 p = &s->stat_percpu[smp_processor_id()][x];
 608                 p->sectors[READ] -= shared->tmp.sectors[READ];
 609                 p->sectors[WRITE] -= shared->tmp.sectors[WRITE];
 610                 p->ios[READ] -= shared->tmp.ios[READ];
 611                 p->ios[WRITE] -= shared->tmp.ios[WRITE];
 612                 p->merges[READ] -= shared->tmp.merges[READ];
 613                 p->merges[WRITE] -= shared->tmp.merges[WRITE];
 614                 p->ticks[READ] -= shared->tmp.ticks[READ];
 615                 p->ticks[WRITE] -= shared->tmp.ticks[WRITE];
 616                 p->io_ticks[READ] -= shared->tmp.io_ticks[READ];
 617                 p->io_ticks[WRITE] -= shared->tmp.io_ticks[WRITE];
 618                 p->io_ticks_total -= shared->tmp.io_ticks_total;
 619                 p->time_in_queue -= shared->tmp.time_in_queue;
 620                 local_irq_enable();
 621         }
 622 }
 623
 624 static int dm_stats_clear(struct dm_stats *stats, int id)
 625 {
 626         struct dm_stat *s;
 627
 628         mutex_lock(&stats->mutex);
 629
 630         s = __dm_stats_find(stats, id);
 631         if (!s) {
 632                 mutex_unlock(&stats->mutex);
 633                 return -ENOENT;
 634         }
 635
 636         __dm_stat_clear(s, 0, s->n_entries, true);
 637
 638         mutex_unlock(&stats->mutex);
 639
 640         return 1;
 641 }
 642
 643 /*
 644  * This is like jiffies_to_msec, but works for 64-bit values.
 645  */
 646 static unsigned long long dm_jiffies_to_msec64(unsigned long long j)
 647 {
 648         unsigned long long result = 0;
 649         unsigned mult;
 650
 651         if (j)
 652                 result = jiffies_to_msecs(j & 0x3fffff);
 653         if (j >= 1 << 22) {
 654                 mult = jiffies_to_msecs(1 << 22);
 655                 result += (unsigned long long)mult * (unsigned long long)jiffies_to_msecs((j >> 22) & 0x3fffff);
 656         }
 657         if (j >= 1ULL << 44)
 658                 result += (unsigned long long)mult * (unsigned long long)mult * (unsigned long long)jiffies_to_msecs(j >> 44);
 659
 660         return result;
 661 }
 662
 663 static int dm_stats_print(struct dm_stats *stats, int id,
 664                           size_t idx_start, size_t idx_len,
 665                           bool clear, char *result, unsigned maxlen)
 666 {
 667         unsigned sz = 0;
 668         struct dm_stat *s;
 669         size_t x;
 670         sector_t start, end, step;
 671         size_t idx_end;
 672         struct dm_stat_shared *shared;
 673
 674         /*
 675          * Output format:
 676          *   <start_sector>+<length> counters
 677          */
 678
 679         mutex_lock(&stats->mutex);
 680
 681         s = __dm_stats_find(stats, id);
 682         if (!s) {
 683                 mutex_unlock(&stats->mutex);
 684                 return -ENOENT;
 685         }
 686
 687         idx_end = idx_start + idx_len;
 688         if (idx_end < idx_start ||
 689             idx_end > s->n_entries)
 690                 idx_end = s->n_entries;
 691
 692         if (idx_start > idx_end)
 693                 idx_start = idx_end;
 694
 695         step = s->step;
 696         start = s->start + (step * idx_start);
 697
 698         for (x = idx_start; x < idx_end; x++, start = end) {
 699                 shared = &s->stat_shared[x];
 700                 end = start + step;
 701                 if (unlikely(end > s->end))
 702                         end = s->end;
 703
 704                 __dm_stat_init_temporary_percpu_totals(shared, s, x);
 705
 706                 DMEMIT("%llu+%llu %llu %llu %llu %llu %llu %llu %llu %llu %d %llu %llu %llu %llu\n",
 707                        (unsigned long long)start,
 708                        (unsigned long long)step,
 709                        shared->tmp.ios[READ],
 710                        shared->tmp.merges[READ],
 711                        shared->tmp.sectors[READ],
 712                        dm_jiffies_to_msec64(shared->tmp.ticks[READ]),
 713                        shared->tmp.ios[WRITE],
 714                        shared->tmp.merges[WRITE],
 715                        shared->tmp.sectors[WRITE],
 716                        dm_jiffies_to_msec64(shared->tmp.ticks[WRITE]),
 717                        dm_stat_in_flight(shared),
 718                        dm_jiffies_to_msec64(shared->tmp.io_ticks_total),
 719                        dm_jiffies_to_msec64(shared->tmp.time_in_queue),
 720                        dm_jiffies_to_msec64(shared->tmp.io_ticks[READ]),
 721                        dm_jiffies_to_msec64(shared->tmp.io_ticks[WRITE]));
 722
 723                 if (unlikely(sz + 1 >= maxlen))
 724                         goto buffer_overflow;
 725         }
 726
 727         if (clear)
 728                 __dm_stat_clear(s, idx_start, idx_end, false);
 729
 730 buffer_overflow:
 731         mutex_unlock(&stats->mutex);
 732
 733         return 1;
 734 }
 735
 736 static int dm_stats_set_aux(struct dm_stats *stats, int id, const char *aux_data)
 737 {
 738         struct dm_stat *s;
 739         const char *new_aux_data;
 740
 741         mutex_lock(&stats->mutex);
 742
 743         s = __dm_stats_find(stats, id);
 744         if (!s) {
 745                 mutex_unlock(&stats->mutex);
 746                 return -ENOENT;
 747         }
 748
 749         new_aux_data = kstrdup(aux_data, GFP_KERNEL);
 750         if (!new_aux_data) {
 751                 mutex_unlock(&stats->mutex);
 752                 return -ENOMEM;
 753         }
 754
 755         kfree(s->aux_data);
 756         s->aux_data = new_aux_data;
 757
 758         mutex_unlock(&stats->mutex);
 759
 760         return 0;
 761 }
 762
 763 static int message_stats_create(struct mapped_device *md,
 764                                 unsigned argc, char **argv,
 765                                 char *result, unsigned maxlen)
 766 {
 767         int id;
 768         char dummy;
 769         unsigned long long start, end, len, step;
 770         unsigned divisor;
 771         const char *program_id, *aux_data;
 772
 773         /*
 774          * Input format:
 775          *   <range> <step> [<program_id> [<aux_data>]]
 776          */
 777
 778         if (argc < 3 || argc > 5)
 779                 return -EINVAL;
 780
 781         if (!strcmp(argv[1], "-")) {
 782                 start = 0;
 783                 len = dm_get_size(md);
 784                 if (!len)
 785                         len = 1;
 786         } else if (sscanf(argv[1], "%llu+%llu%c", &start, &len, &dummy) != 2 ||
 787                    start != (sector_t)start || len != (sector_t)len)
 788                 return -EINVAL;
 789
 790         end = start + len;
 791         if (start >= end)
 792                 return -EINVAL;
 793
 794         if (sscanf(argv[2], "/%u%c", &divisor, &dummy) == 1) {
 795                 if (!divisor)
 796                         return -EINVAL;
 797                 step = end - start;
 798                 if (do_div(step, divisor))
 799                         step++;
 800                 if (!step)
 801                         step = 1;
 802         } else if (sscanf(argv[2], "%llu%c", &step, &dummy) != 1 ||
 803                    step != (sector_t)step || !step)
 804                 return -EINVAL;
 805
 806         program_id = "-";
 807         aux_data = "-";
 808
 809         if (argc > 3)
 810                 program_id = argv[3];
 811
 812         if (argc > 4)
 813                 aux_data = argv[4];
 814
 815         /*
 816          * If a buffer overflow happens after we created the region,
 817          * it's too late (the userspace would retry with a larger
 818          * buffer, but the region id that caused the overflow is already
 819          * leaked).  So we must detect buffer overflow in advance.
 820          */
 821         snprintf(result, maxlen, "%d", INT_MAX);
 822         if (dm_message_test_buffer_overflow(result, maxlen))
 823                 return 1;
 824
 825         id = dm_stats_create(dm_get_stats(md), start, end, step, program_id, aux_data,
 826                              dm_internal_suspend_fast, dm_internal_resume_fast, md);
 827         if (id < 0)
 828                 return id;
 829
 830         snprintf(result, maxlen, "%d", id);
 831
 832         return 1;
 833 }
 834
 835 static int message_stats_delete(struct mapped_device *md,
 836                                 unsigned argc, char **argv)
 837 {
 838         int id;
 839         char dummy;
 840
 841         if (argc != 2)
 842                 return -EINVAL;
 843
 844         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
 845                 return -EINVAL;
 846
 847         return dm_stats_delete(dm_get_stats(md), id);
 848 }
 849
 850 static int message_stats_clear(struct mapped_device *md,
 851                                unsigned argc, char **argv)
 852 {
 853         int id;
 854         char dummy;
 855
 856         if (argc != 2)
 857                 return -EINVAL;
 858
 859         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
 860                 return -EINVAL;
 861
 862         return dm_stats_clear(dm_get_stats(md), id);
 863 }
 864
 865 static int message_stats_list(struct mapped_device *md,
 866                               unsigned argc, char **argv,
 867                               char *result, unsigned maxlen)
 868 {
 869         int r;
 870         const char *program = NULL;
 871
 872         if (argc < 1 || argc > 2)
 873                 return -EINVAL;
 874
 875         if (argc > 1) {
 876                 program = kstrdup(argv[1], GFP_KERNEL);
 877                 if (!program)
 878                         return -ENOMEM;
 879         }
 880
 881         r = dm_stats_list(dm_get_stats(md), program, result, maxlen);
 882
 883         kfree(program);
 884
 885         return r;
 886 }
 887
 888 static int message_stats_print(struct mapped_device *md,
 889                                unsigned argc, char **argv, bool clear,
 890                                char *result, unsigned maxlen)
 891 {
 892         int id;
 893         char dummy;
 894         unsigned long idx_start = 0, idx_len = ULONG_MAX;
 895
 896         if (argc != 2 && argc != 4)
 897                 return -EINVAL;
 898
 899         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
 900                 return -EINVAL;
 901
 902         if (argc > 3) {
 903                 if (strcmp(argv[2], "-") &&
 904                     sscanf(argv[2], "%lu%c", &idx_start, &dummy) != 1)
 905                         return -EINVAL;
 906                 if (strcmp(argv[3], "-") &&
 907                     sscanf(argv[3], "%lu%c", &idx_len, &dummy) != 1)
 908                         return -EINVAL;
 909         }
 910
 911         return dm_stats_print(dm_get_stats(md), id, idx_start, idx_len, clear,
 912                               result, maxlen);
 913 }
 914
 915 static int message_stats_set_aux(struct mapped_device *md,
 916                                  unsigned argc, char **argv)
 917 {
 918         int id;
 919         char dummy;
 920
 921         if (argc != 3)
 922                 return -EINVAL;
 923
 924         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
 925                 return -EINVAL;
 926
 927         return dm_stats_set_aux(dm_get_stats(md), id, argv[2]);
 928 }
 929
 930 int dm_stats_message(struct mapped_device *md, unsigned argc, char **argv,
 931                      char *result, unsigned maxlen)
 932 {
 933         int r;
 934
 935         if (dm_request_based(md)) {
 936                 DMWARN("Statistics are only supported for bio-based devices");
 937                 return -EOPNOTSUPP;
 938         }
 939
 940         /* All messages here must start with '@' */
 941         if (!strcasecmp(argv[0], "@stats_create"))
 942                 r = message_stats_create(md, argc, argv, result, maxlen);
 943         else if (!strcasecmp(argv[0], "@stats_delete"))
 944                 r = message_stats_delete(md, argc, argv);
 945         else if (!strcasecmp(argv[0], "@stats_clear"))
 946                 r = message_stats_clear(md, argc, argv);
 947         else if (!strcasecmp(argv[0], "@stats_list"))
 948                 r = message_stats_list(md, argc, argv, result, maxlen);
 949         else if (!strcasecmp(argv[0], "@stats_print"))
 950                 r = message_stats_print(md, argc, argv, false, result, maxlen);
 951         else if (!strcasecmp(argv[0], "@stats_print_clear"))
 952                 r = message_stats_print(md, argc, argv, true, result, maxlen);
 953         else if (!strcasecmp(argv[0], "@stats_set_aux"))
 954                 r = message_stats_set_aux(md, argc, argv);
 955         else
 956                 return 2; /* this wasn't a stats message */
 957
 958         if (r == -EINVAL)
 959                 DMWARN("Invalid parameters for message %s", argv[0]);
 960
 961         return r;
 962 }
 963
 964 int __init dm_statistics_init(void)
 965 {
 966         shared_memory_amount = 0;
 967         dm_stat_need_rcu_barrier = 0;
 968         return 0;
 969 }
 970
 971 void dm_statistics_exit(void)
 972 {
 973         if (dm_stat_need_rcu_barrier)
 974                 rcu_barrier();
 975         if (WARN_ON(shared_memory_amount))
 976                 DMCRIT("shared_memory_amount leaked: %lu", shared_memory_amount);
 977 }
 978
 979 module_param_named(stats_current_allocated_bytes, shared_memory_amount, ulong, S_IRUGO);
 980 MODULE_PARM_DESC(stats_current_allocated_bytes, "Memory currently used by statistics");