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         unsigned long long *histogram;
  33 };
  34
  35 struct dm_stat_shared {
  36         atomic_t in_flight[2];
  37         unsigned long long stamp;
  38         struct dm_stat_percpu tmp;
  39 };
  40
  41 struct dm_stat {
  42         struct list_head list_entry;
  43         int id;
  44         unsigned stat_flags;
  45         size_t n_entries;
  46         sector_t start;
  47         sector_t end;
  48         sector_t step;
  49         unsigned n_histogram_entries;
  50         unsigned long long *histogram_boundaries;
  51         const char *program_id;
  52         const char *aux_data;
  53         struct rcu_head rcu_head;
  54         size_t shared_alloc_size;
  55         size_t percpu_alloc_size;
  56         size_t histogram_alloc_size;
  57         struct dm_stat_percpu *stat_percpu[NR_CPUS];
  58         struct dm_stat_shared stat_shared[0];
  59 };
  60
  61 #define STAT_PRECISE_TIMESTAMPS         1
  62
  63 struct dm_stats_last_position {
  64         sector_t last_sector;
  65         unsigned last_rw;
  66 };
  67
  68 /*
  69  * A typo on the command line could possibly make the kernel run out of memory
  70  * and crash. To prevent the crash we account all used memory. We fail if we
  71  * exhaust 1/4 of all memory or 1/2 of vmalloc space.
  72  */
  73 #define DM_STATS_MEMORY_FACTOR          4
  74 #define DM_STATS_VMALLOC_FACTOR         2
  75
  76 static DEFINE_SPINLOCK(shared_memory_lock);
  77
  78 static unsigned long shared_memory_amount;
  79
  80 static bool __check_shared_memory(size_t alloc_size)
  81 {
  82         size_t a;
  83
  84         a = shared_memory_amount + alloc_size;
  85         if (a < shared_memory_amount)
  86                 return false;
  87         if (a >> PAGE_SHIFT > totalram_pages / DM_STATS_MEMORY_FACTOR)
  88                 return false;
  89 #ifdef CONFIG_MMU
  90         if (a > (VMALLOC_END - VMALLOC_START) / DM_STATS_VMALLOC_FACTOR)
  91                 return false;
  92 #endif
  93         return true;
  94 }
  95
  96 static bool check_shared_memory(size_t alloc_size)
  97 {
  98         bool ret;
  99
 100         spin_lock_irq(&shared_memory_lock);
 101
 102         ret = __check_shared_memory(alloc_size);
 103
 104         spin_unlock_irq(&shared_memory_lock);
 105
 106         return ret;
 107 }
 108
 109 static bool claim_shared_memory(size_t alloc_size)
 110 {
 111         spin_lock_irq(&shared_memory_lock);
 112
 113         if (!__check_shared_memory(alloc_size)) {
 114                 spin_unlock_irq(&shared_memory_lock);
 115                 return false;
 116         }
 117
 118         shared_memory_amount += alloc_size;
 119
 120         spin_unlock_irq(&shared_memory_lock);
 121
 122         return true;
 123 }
 124
 125 static void free_shared_memory(size_t alloc_size)
 126 {
 127         unsigned long flags;
 128
 129         spin_lock_irqsave(&shared_memory_lock, flags);
 130
 131         if (WARN_ON_ONCE(shared_memory_amount < alloc_size)) {
 132                 spin_unlock_irqrestore(&shared_memory_lock, flags);
 133                 DMCRIT("Memory usage accounting bug.");
 134                 return;
 135         }
 136
 137         shared_memory_amount -= alloc_size;
 138
 139         spin_unlock_irqrestore(&shared_memory_lock, flags);
 140 }
 141
 142 static void *dm_kvzalloc(size_t alloc_size, int node)
 143 {
 144         void *p;
 145
 146         if (!claim_shared_memory(alloc_size))
 147                 return NULL;
 148
 149         if (alloc_size <= KMALLOC_MAX_SIZE) {
 150                 p = kzalloc_node(alloc_size, GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN, node);
 151                 if (p)
 152                         return p;
 153         }
 154         p = vzalloc_node(alloc_size, node);
 155         if (p)
 156                 return p;
 157
 158         free_shared_memory(alloc_size);
 159
 160         return NULL;
 161 }
 162
 163 static void dm_kvfree(void *ptr, size_t alloc_size)
 164 {
 165         if (!ptr)
 166                 return;
 167
 168         free_shared_memory(alloc_size);
 169
 170         kvfree(ptr);
 171 }
 172
 173 static void dm_stat_free(struct rcu_head *head)
 174 {
 175         int cpu;
 176         struct dm_stat *s = container_of(head, struct dm_stat, rcu_head);
 177
 178         kfree(s->program_id);
 179         kfree(s->aux_data);
 180         for_each_possible_cpu(cpu) {
 181                 dm_kvfree(s->stat_percpu[cpu][0].histogram, s->histogram_alloc_size);
 182                 dm_kvfree(s->stat_percpu[cpu], s->percpu_alloc_size);
 183         }
 184         dm_kvfree(s->stat_shared[0].tmp.histogram, s->histogram_alloc_size);
 185         dm_kvfree(s, s->shared_alloc_size);
 186 }
 187
 188 static int dm_stat_in_flight(struct dm_stat_shared *shared)
 189 {
 190         return atomic_read(&shared->in_flight[READ]) +
 191                atomic_read(&shared->in_flight[WRITE]);
 192 }
 193
 194 void dm_stats_init(struct dm_stats *stats)
 195 {
 196         int cpu;
 197         struct dm_stats_last_position *last;
 198
 199         mutex_init(&stats->mutex);
 200         INIT_LIST_HEAD(&stats->list);
 201         stats->last = alloc_percpu(struct dm_stats_last_position);
 202         for_each_possible_cpu(cpu) {
 203                 last = per_cpu_ptr(stats->last, cpu);
 204                 last->last_sector = (sector_t)ULLONG_MAX;
 205                 last->last_rw = UINT_MAX;
 206         }
 207 }
 208
 209 void dm_stats_cleanup(struct dm_stats *stats)
 210 {
 211         size_t ni;
 212         struct dm_stat *s;
 213         struct dm_stat_shared *shared;
 214
 215         while (!list_empty(&stats->list)) {
 216                 s = container_of(stats->list.next, struct dm_stat, list_entry);
 217                 list_del(&s->list_entry);
 218                 for (ni = 0; ni < s->n_entries; ni++) {
 219                         shared = &s->stat_shared[ni];
 220                         if (WARN_ON(dm_stat_in_flight(shared))) {
 221                                 DMCRIT("leaked in-flight counter at index %lu "
 222                                        "(start %llu, end %llu, step %llu): reads %d, writes %d",
 223                                        (unsigned long)ni,
 224                                        (unsigned long long)s->start,
 225                                        (unsigned long long)s->end,
 226                                        (unsigned long long)s->step,
 227                                        atomic_read(&shared->in_flight[READ]),
 228                                        atomic_read(&shared->in_flight[WRITE]));
 229                         }
 230                 }
 231                 dm_stat_free(&s->rcu_head);
 232         }
 233         free_percpu(stats->last);
 234 }
 235
 236 static int dm_stats_create(struct dm_stats *stats, sector_t start, sector_t end,
 237                            sector_t step, unsigned stat_flags,
 238                            unsigned n_histogram_entries,
 239                            unsigned long long *histogram_boundaries,
 240                            const char *program_id, const char *aux_data,
 241                            void (*suspend_callback)(struct mapped_device *),
 242                            void (*resume_callback)(struct mapped_device *),
 243                            struct mapped_device *md)
 244 {
 245         struct list_head *l;
 246         struct dm_stat *s, *tmp_s;
 247         sector_t n_entries;
 248         size_t ni;
 249         size_t shared_alloc_size;
 250         size_t percpu_alloc_size;
 251         size_t histogram_alloc_size;
 252         struct dm_stat_percpu *p;
 253         int cpu;
 254         int ret_id;
 255         int r;
 256
 257         if (end < start || !step)
 258                 return -EINVAL;
 259
 260         n_entries = end - start;
 261         if (dm_sector_div64(n_entries, step))
 262                 n_entries++;
 263
 264         if (n_entries != (size_t)n_entries || !(size_t)(n_entries + 1))
 265                 return -EOVERFLOW;
 266
 267         shared_alloc_size = sizeof(struct dm_stat) + (size_t)n_entries * sizeof(struct dm_stat_shared);
 268         if ((shared_alloc_size - sizeof(struct dm_stat)) / sizeof(struct dm_stat_shared) != n_entries)
 269                 return -EOVERFLOW;
 270
 271         percpu_alloc_size = (size_t)n_entries * sizeof(struct dm_stat_percpu);
 272         if (percpu_alloc_size / sizeof(struct dm_stat_percpu) != n_entries)
 273                 return -EOVERFLOW;
 274
 275         histogram_alloc_size = (n_histogram_entries + 1) * (size_t)n_entries * sizeof(unsigned long long);
 276         if (histogram_alloc_size / (n_histogram_entries + 1) != (size_t)n_entries * sizeof(unsigned long long))
 277                 return -EOVERFLOW;
 278
 279         if (!check_shared_memory(shared_alloc_size + histogram_alloc_size +
 280                                  num_possible_cpus() * (percpu_alloc_size + histogram_alloc_size)))
 281                 return -ENOMEM;
 282
 283         s = dm_kvzalloc(shared_alloc_size, NUMA_NO_NODE);
 284         if (!s)
 285                 return -ENOMEM;
 286
 287         s->stat_flags = stat_flags;
 288         s->n_entries = n_entries;
 289         s->start = start;
 290         s->end = end;
 291         s->step = step;
 292         s->shared_alloc_size = shared_alloc_size;
 293         s->percpu_alloc_size = percpu_alloc_size;
 294         s->histogram_alloc_size = histogram_alloc_size;
 295
 296         s->n_histogram_entries = n_histogram_entries;
 297         s->histogram_boundaries = kmemdup(histogram_boundaries,
 298                                           s->n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
 299         if (!s->histogram_boundaries) {
 300                 r = -ENOMEM;
 301                 goto out;
 302         }
 303
 304         s->program_id = kstrdup(program_id, GFP_KERNEL);
 305         if (!s->program_id) {
 306                 r = -ENOMEM;
 307                 goto out;
 308         }
 309         s->aux_data = kstrdup(aux_data, GFP_KERNEL);
 310         if (!s->aux_data) {
 311                 r = -ENOMEM;
 312                 goto out;
 313         }
 314
 315         for (ni = 0; ni < n_entries; ni++) {
 316                 atomic_set(&s->stat_shared[ni].in_flight[READ], 0);
 317                 atomic_set(&s->stat_shared[ni].in_flight[WRITE], 0);
 318         }
 319
 320         if (s->n_histogram_entries) {
 321                 unsigned long long *hi;
 322                 hi = dm_kvzalloc(s->histogram_alloc_size, NUMA_NO_NODE);
 323                 if (!hi) {
 324                         r = -ENOMEM;
 325                         goto out;
 326                 }
 327                 for (ni = 0; ni < n_entries; ni++) {
 328                         s->stat_shared[ni].tmp.histogram = hi;
 329                         hi += s->n_histogram_entries + 1;
 330                 }
 331         }
 332
 333         for_each_possible_cpu(cpu) {
 334                 p = dm_kvzalloc(percpu_alloc_size, cpu_to_node(cpu));
 335                 if (!p) {
 336                         r = -ENOMEM;
 337                         goto out;
 338                 }
 339                 s->stat_percpu[cpu] = p;
 340                 if (s->n_histogram_entries) {
 341                         unsigned long long *hi;
 342                         hi = dm_kvzalloc(s->histogram_alloc_size, cpu_to_node(cpu));
 343                         if (!hi) {
 344                                 r = -ENOMEM;
 345                                 goto out;
 346                         }
 347                         for (ni = 0; ni < n_entries; ni++) {
 348                                 p[ni].histogram = hi;
 349                                 hi += s->n_histogram_entries + 1;
 350                         }
 351                 }
 352         }
 353
 354         /*
 355          * Suspend/resume to make sure there is no i/o in flight,
 356          * so that newly created statistics will be exact.
 357          *
 358          * (note: we couldn't suspend earlier because we must not
 359          * allocate memory while suspended)
 360          */
 361         suspend_callback(md);
 362
 363         mutex_lock(&stats->mutex);
 364         s->id = 0;
 365         list_for_each(l, &stats->list) {
 366                 tmp_s = container_of(l, struct dm_stat, list_entry);
 367                 if (WARN_ON(tmp_s->id < s->id)) {
 368                         r = -EINVAL;
 369                         goto out_unlock_resume;
 370                 }
 371                 if (tmp_s->id > s->id)
 372                         break;
 373                 if (unlikely(s->id == INT_MAX)) {
 374                         r = -ENFILE;
 375                         goto out_unlock_resume;
 376                 }
 377                 s->id++;
 378         }
 379         ret_id = s->id;
 380         list_add_tail_rcu(&s->list_entry, l);
 381         mutex_unlock(&stats->mutex);
 382
 383         resume_callback(md);
 384
 385         return ret_id;
 386
 387 out_unlock_resume:
 388         mutex_unlock(&stats->mutex);
 389         resume_callback(md);
 390 out:
 391         dm_stat_free(&s->rcu_head);
 392         return r;
 393 }
 394
 395 static struct dm_stat *__dm_stats_find(struct dm_stats *stats, int id)
 396 {
 397         struct dm_stat *s;
 398
 399         list_for_each_entry(s, &stats->list, list_entry) {
 400                 if (s->id > id)
 401                         break;
 402                 if (s->id == id)
 403                         return s;
 404         }
 405
 406         return NULL;
 407 }
 408
 409 static int dm_stats_delete(struct dm_stats *stats, int id)
 410 {
 411         struct dm_stat *s;
 412         int cpu;
 413
 414         mutex_lock(&stats->mutex);
 415
 416         s = __dm_stats_find(stats, id);
 417         if (!s) {
 418                 mutex_unlock(&stats->mutex);
 419                 return -ENOENT;
 420         }
 421
 422         list_del_rcu(&s->list_entry);
 423         mutex_unlock(&stats->mutex);
 424
 425         /*
 426          * vfree can't be called from RCU callback
 427          */
 428         for_each_possible_cpu(cpu)
 429                 if (is_vmalloc_addr(s->stat_percpu) ||
 430                     is_vmalloc_addr(s->stat_percpu[cpu][0].histogram))
 431                         goto do_sync_free;
 432         if (is_vmalloc_addr(s) ||
 433             is_vmalloc_addr(s->stat_shared[0].tmp.histogram)) {
 434 do_sync_free:
 435                 synchronize_rcu_expedited();
 436                 dm_stat_free(&s->rcu_head);
 437         } else {
 438                 ACCESS_ONCE(dm_stat_need_rcu_barrier) = 1;
 439                 call_rcu(&s->rcu_head, dm_stat_free);
 440         }
 441         return 0;
 442 }
 443
 444 static int dm_stats_list(struct dm_stats *stats, const char *program,
 445                          char *result, unsigned maxlen)
 446 {
 447         struct dm_stat *s;
 448         sector_t len;
 449         unsigned sz = 0;
 450
 451         /*
 452          * Output format:
 453          *   <region_id>: <start_sector>+<length> <step> <program_id> <aux_data>
 454          */
 455
 456         mutex_lock(&stats->mutex);
 457         list_for_each_entry(s, &stats->list, list_entry) {
 458                 if (!program || !strcmp(program, s->program_id)) {
 459                         len = s->end - s->start;
 460                         DMEMIT("%d: %llu+%llu %llu %s %s", s->id,
 461                                 (unsigned long long)s->start,
 462                                 (unsigned long long)len,
 463                                 (unsigned long long)s->step,
 464                                 s->program_id,
 465                                 s->aux_data);
 466                         if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
 467                                 DMEMIT(" precise_timestamps");
 468                         if (s->n_histogram_entries) {
 469                                 unsigned i;
 470                                 DMEMIT(" histogram:");
 471                                 for (i = 0; i < s->n_histogram_entries; i++) {
 472                                         if (i)
 473                                                 DMEMIT(",");
 474                                         DMEMIT("%llu", s->histogram_boundaries[i]);
 475                                 }
 476                         }
 477                         DMEMIT("\n");
 478                 }
 479         }
 480         mutex_unlock(&stats->mutex);
 481
 482         return 1;
 483 }
 484
 485 static void dm_stat_round(struct dm_stat *s, struct dm_stat_shared *shared,
 486                           struct dm_stat_percpu *p)
 487 {
 488         /*
 489          * This is racy, but so is part_round_stats_single.
 490          */
 491         unsigned long long now, difference;
 492         unsigned in_flight_read, in_flight_write;
 493
 494         if (likely(!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)))
 495                 now = jiffies;
 496         else
 497                 now = ktime_to_ns(ktime_get());
 498
 499         difference = now - shared->stamp;
 500         if (!difference)
 501                 return;
 502
 503         in_flight_read = (unsigned)atomic_read(&shared->in_flight[READ]);
 504         in_flight_write = (unsigned)atomic_read(&shared->in_flight[WRITE]);
 505         if (in_flight_read)
 506                 p->io_ticks[READ] += difference;
 507         if (in_flight_write)
 508                 p->io_ticks[WRITE] += difference;
 509         if (in_flight_read + in_flight_write) {
 510                 p->io_ticks_total += difference;
 511                 p->time_in_queue += (in_flight_read + in_flight_write) * difference;
 512         }
 513         shared->stamp = now;
 514 }
 515
 516 static void dm_stat_for_entry(struct dm_stat *s, size_t entry,
 517                               unsigned long bi_rw, sector_t len,
 518                               struct dm_stats_aux *stats_aux, bool end,
 519                               unsigned long duration_jiffies)
 520 {
 521         unsigned long idx = bi_rw & REQ_WRITE;
 522         struct dm_stat_shared *shared = &s->stat_shared[entry];
 523         struct dm_stat_percpu *p;
 524
 525         /*
 526          * For strict correctness we should use local_irq_save/restore
 527          * instead of preempt_disable/enable.
 528          *
 529          * preempt_disable/enable is racy if the driver finishes bios
 530          * from non-interrupt context as well as from interrupt context
 531          * or from more different interrupts.
 532          *
 533          * On 64-bit architectures the race only results in not counting some
 534          * events, so it is acceptable.  On 32-bit architectures the race could
 535          * cause the counter going off by 2^32, so we need to do proper locking
 536          * there.
 537          *
 538          * part_stat_lock()/part_stat_unlock() have this race too.
 539          */
 540 #if BITS_PER_LONG == 32
 541         unsigned long flags;
 542         local_irq_save(flags);
 543 #else
 544         preempt_disable();
 545 #endif
 546         p = &s->stat_percpu[smp_processor_id()][entry];
 547
 548         if (!end) {
 549                 dm_stat_round(s, shared, p);
 550                 atomic_inc(&shared->in_flight[idx]);
 551         } else {
 552                 unsigned long long duration;
 553                 dm_stat_round(s, shared, p);
 554                 atomic_dec(&shared->in_flight[idx]);
 555                 p->sectors[idx] += len;
 556                 p->ios[idx] += 1;
 557                 p->merges[idx] += stats_aux->merged;
 558                 if (!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)) {
 559                         p->ticks[idx] += duration_jiffies;
 560                         duration = jiffies_to_msecs(duration_jiffies);
 561                 } else {
 562                         p->ticks[idx] += stats_aux->duration_ns;
 563                         duration = stats_aux->duration_ns;
 564                 }
 565                 if (s->n_histogram_entries) {
 566                         unsigned lo = 0, hi = s->n_histogram_entries + 1;
 567                         while (lo + 1 < hi) {
 568                                 unsigned mid = (lo + hi) / 2;
 569                                 if (s->histogram_boundaries[mid - 1] > duration) {
 570                                         hi = mid;
 571                                 } else {
 572                                         lo = mid;
 573                                 }
 574
 575                         }
 576                         p->histogram[lo]++;
 577                 }
 578         }
 579
 580 #if BITS_PER_LONG == 32
 581         local_irq_restore(flags);
 582 #else
 583         preempt_enable();
 584 #endif
 585 }
 586
 587 static void __dm_stat_bio(struct dm_stat *s, unsigned long bi_rw,
 588                           sector_t bi_sector, sector_t end_sector,
 589                           bool end, unsigned long duration_jiffies,
 590                           struct dm_stats_aux *stats_aux)
 591 {
 592         sector_t rel_sector, offset, todo, fragment_len;
 593         size_t entry;
 594
 595         if (end_sector <= s->start || bi_sector >= s->end)
 596                 return;
 597         if (unlikely(bi_sector < s->start)) {
 598                 rel_sector = 0;
 599                 todo = end_sector - s->start;
 600         } else {
 601                 rel_sector = bi_sector - s->start;
 602                 todo = end_sector - bi_sector;
 603         }
 604         if (unlikely(end_sector > s->end))
 605                 todo -= (end_sector - s->end);
 606
 607         offset = dm_sector_div64(rel_sector, s->step);
 608         entry = rel_sector;
 609         do {
 610                 if (WARN_ON_ONCE(entry >= s->n_entries)) {
 611                         DMCRIT("Invalid area access in region id %d", s->id);
 612                         return;
 613                 }
 614                 fragment_len = todo;
 615                 if (fragment_len > s->step - offset)
 616                         fragment_len = s->step - offset;
 617                 dm_stat_for_entry(s, entry, bi_rw, fragment_len,
 618                                   stats_aux, end, duration_jiffies);
 619                 todo -= fragment_len;
 620                 entry++;
 621                 offset = 0;
 622         } while (unlikely(todo != 0));
 623 }
 624
 625 void dm_stats_account_io(struct dm_stats *stats, unsigned long bi_rw,
 626                          sector_t bi_sector, unsigned bi_sectors, bool end,
 627                          unsigned long duration_jiffies,
 628                          struct dm_stats_aux *stats_aux)
 629 {
 630         struct dm_stat *s;
 631         sector_t end_sector;
 632         struct dm_stats_last_position *last;
 633         bool got_precise_time;
 634
 635         if (unlikely(!bi_sectors))
 636                 return;
 637
 638         end_sector = bi_sector + bi_sectors;
 639
 640         if (!end) {
 641                 /*
 642                  * A race condition can at worst result in the merged flag being
 643                  * misrepresented, so we don't have to disable preemption here.
 644                  */
 645                 last = raw_cpu_ptr(stats->last);
 646                 stats_aux->merged =
 647                         (bi_sector == (ACCESS_ONCE(last->last_sector) &&
 648                                        ((bi_rw & (REQ_WRITE | REQ_DISCARD)) ==
 649                                         (ACCESS_ONCE(last->last_rw) & (REQ_WRITE | REQ_DISCARD)))
 650                                        ));
 651                 ACCESS_ONCE(last->last_sector) = end_sector;
 652                 ACCESS_ONCE(last->last_rw) = bi_rw;
 653         }
 654
 655         rcu_read_lock();
 656
 657         got_precise_time = false;
 658         list_for_each_entry_rcu(s, &stats->list, list_entry) {
 659                 if (s->stat_flags & STAT_PRECISE_TIMESTAMPS && !got_precise_time) {
 660                         if (!end)
 661                                 stats_aux->duration_ns = ktime_to_ns(ktime_get());
 662                         else
 663                                 stats_aux->duration_ns = ktime_to_ns(ktime_get()) - stats_aux->duration_ns;
 664                         got_precise_time = true;
 665                 }
 666                 __dm_stat_bio(s, bi_rw, bi_sector, end_sector, end, duration_jiffies, stats_aux);
 667         }
 668
 669         rcu_read_unlock();
 670 }
 671
 672 static void __dm_stat_init_temporary_percpu_totals(struct dm_stat_shared *shared,
 673                                                    struct dm_stat *s, size_t x)
 674 {
 675         int cpu;
 676         struct dm_stat_percpu *p;
 677
 678         local_irq_disable();
 679         p = &s->stat_percpu[smp_processor_id()][x];
 680         dm_stat_round(s, shared, p);
 681         local_irq_enable();
 682
 683         shared->tmp.sectors[READ] = 0;
 684         shared->tmp.sectors[WRITE] = 0;
 685         shared->tmp.ios[READ] = 0;
 686         shared->tmp.ios[WRITE] = 0;
 687         shared->tmp.merges[READ] = 0;
 688         shared->tmp.merges[WRITE] = 0;
 689         shared->tmp.ticks[READ] = 0;
 690         shared->tmp.ticks[WRITE] = 0;
 691         shared->tmp.io_ticks[READ] = 0;
 692         shared->tmp.io_ticks[WRITE] = 0;
 693         shared->tmp.io_ticks_total = 0;
 694         shared->tmp.time_in_queue = 0;
 695
 696         if (s->n_histogram_entries)
 697                 memset(shared->tmp.histogram, 0, (s->n_histogram_entries + 1) * sizeof(unsigned long long));
 698
 699         for_each_possible_cpu(cpu) {
 700                 p = &s->stat_percpu[cpu][x];
 701                 shared->tmp.sectors[READ] += ACCESS_ONCE(p->sectors[READ]);
 702                 shared->tmp.sectors[WRITE] += ACCESS_ONCE(p->sectors[WRITE]);
 703                 shared->tmp.ios[READ] += ACCESS_ONCE(p->ios[READ]);
 704                 shared->tmp.ios[WRITE] += ACCESS_ONCE(p->ios[WRITE]);
 705                 shared->tmp.merges[READ] += ACCESS_ONCE(p->merges[READ]);
 706                 shared->tmp.merges[WRITE] += ACCESS_ONCE(p->merges[WRITE]);
 707                 shared->tmp.ticks[READ] += ACCESS_ONCE(p->ticks[READ]);
 708                 shared->tmp.ticks[WRITE] += ACCESS_ONCE(p->ticks[WRITE]);
 709                 shared->tmp.io_ticks[READ] += ACCESS_ONCE(p->io_ticks[READ]);
 710                 shared->tmp.io_ticks[WRITE] += ACCESS_ONCE(p->io_ticks[WRITE]);
 711                 shared->tmp.io_ticks_total += ACCESS_ONCE(p->io_ticks_total);
 712                 shared->tmp.time_in_queue += ACCESS_ONCE(p->time_in_queue);
 713                 if (s->n_histogram_entries) {
 714                         unsigned i;
 715                         for (i = 0; i < s->n_histogram_entries + 1; i++)
 716                                 shared->tmp.histogram[i] += ACCESS_ONCE(p->histogram[i]);
 717                 }
 718         }
 719 }
 720
 721 static void __dm_stat_clear(struct dm_stat *s, size_t idx_start, size_t idx_end,
 722                             bool init_tmp_percpu_totals)
 723 {
 724         size_t x;
 725         struct dm_stat_shared *shared;
 726         struct dm_stat_percpu *p;
 727
 728         for (x = idx_start; x < idx_end; x++) {
 729                 shared = &s->stat_shared[x];
 730                 if (init_tmp_percpu_totals)
 731                         __dm_stat_init_temporary_percpu_totals(shared, s, x);
 732                 local_irq_disable();
 733                 p = &s->stat_percpu[smp_processor_id()][x];
 734                 p->sectors[READ] -= shared->tmp.sectors[READ];
 735                 p->sectors[WRITE] -= shared->tmp.sectors[WRITE];
 736                 p->ios[READ] -= shared->tmp.ios[READ];
 737                 p->ios[WRITE] -= shared->tmp.ios[WRITE];
 738                 p->merges[READ] -= shared->tmp.merges[READ];
 739                 p->merges[WRITE] -= shared->tmp.merges[WRITE];
 740                 p->ticks[READ] -= shared->tmp.ticks[READ];
 741                 p->ticks[WRITE] -= shared->tmp.ticks[WRITE];
 742                 p->io_ticks[READ] -= shared->tmp.io_ticks[READ];
 743                 p->io_ticks[WRITE] -= shared->tmp.io_ticks[WRITE];
 744                 p->io_ticks_total -= shared->tmp.io_ticks_total;
 745                 p->time_in_queue -= shared->tmp.time_in_queue;
 746                 local_irq_enable();
 747                 if (s->n_histogram_entries) {
 748                         unsigned i;
 749                         for (i = 0; i < s->n_histogram_entries + 1; i++) {
 750                                 local_irq_disable();
 751                                 p = &s->stat_percpu[smp_processor_id()][x];
 752                                 p->histogram[i] -= shared->tmp.histogram[i];
 753                                 local_irq_enable();
 754                         }
 755                 }
 756         }
 757 }
 758
 759 static int dm_stats_clear(struct dm_stats *stats, int id)
 760 {
 761         struct dm_stat *s;
 762
 763         mutex_lock(&stats->mutex);
 764
 765         s = __dm_stats_find(stats, id);
 766         if (!s) {
 767                 mutex_unlock(&stats->mutex);
 768                 return -ENOENT;
 769         }
 770
 771         __dm_stat_clear(s, 0, s->n_entries, true);
 772
 773         mutex_unlock(&stats->mutex);
 774
 775         return 1;
 776 }
 777
 778 /*
 779  * This is like jiffies_to_msec, but works for 64-bit values.
 780  */
 781 static unsigned long long dm_jiffies_to_msec64(struct dm_stat *s, unsigned long long j)
 782 {
 783         unsigned long long result;
 784         unsigned mult;
 785
 786         if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
 787                 return j;
 788
 789         result = 0;
 790         if (j)
 791                 result = jiffies_to_msecs(j & 0x3fffff);
 792         if (j >= 1 << 22) {
 793                 mult = jiffies_to_msecs(1 << 22);
 794                 result += (unsigned long long)mult * (unsigned long long)jiffies_to_msecs((j >> 22) & 0x3fffff);
 795         }
 796         if (j >= 1ULL << 44)
 797                 result += (unsigned long long)mult * (unsigned long long)mult * (unsigned long long)jiffies_to_msecs(j >> 44);
 798
 799         return result;
 800 }
 801
 802 static int dm_stats_print(struct dm_stats *stats, int id,
 803                           size_t idx_start, size_t idx_len,
 804                           bool clear, char *result, unsigned maxlen)
 805 {
 806         unsigned sz = 0;
 807         struct dm_stat *s;
 808         size_t x;
 809         sector_t start, end, step;
 810         size_t idx_end;
 811         struct dm_stat_shared *shared;
 812
 813         /*
 814          * Output format:
 815          *   <start_sector>+<length> counters
 816          */
 817
 818         mutex_lock(&stats->mutex);
 819
 820         s = __dm_stats_find(stats, id);
 821         if (!s) {
 822                 mutex_unlock(&stats->mutex);
 823                 return -ENOENT;
 824         }
 825
 826         idx_end = idx_start + idx_len;
 827         if (idx_end < idx_start ||
 828             idx_end > s->n_entries)
 829                 idx_end = s->n_entries;
 830
 831         if (idx_start > idx_end)
 832                 idx_start = idx_end;
 833
 834         step = s->step;
 835         start = s->start + (step * idx_start);
 836
 837         for (x = idx_start; x < idx_end; x++, start = end) {
 838                 shared = &s->stat_shared[x];
 839                 end = start + step;
 840                 if (unlikely(end > s->end))
 841                         end = s->end;
 842
 843                 __dm_stat_init_temporary_percpu_totals(shared, s, x);
 844
 845                 DMEMIT("%llu+%llu %llu %llu %llu %llu %llu %llu %llu %llu %d %llu %llu %llu %llu",
 846                        (unsigned long long)start,
 847                        (unsigned long long)step,
 848                        shared->tmp.ios[READ],
 849                        shared->tmp.merges[READ],
 850                        shared->tmp.sectors[READ],
 851                        dm_jiffies_to_msec64(s, shared->tmp.ticks[READ]),
 852                        shared->tmp.ios[WRITE],
 853                        shared->tmp.merges[WRITE],
 854                        shared->tmp.sectors[WRITE],
 855                        dm_jiffies_to_msec64(s, shared->tmp.ticks[WRITE]),
 856                        dm_stat_in_flight(shared),
 857                        dm_jiffies_to_msec64(s, shared->tmp.io_ticks_total),
 858                        dm_jiffies_to_msec64(s, shared->tmp.time_in_queue),
 859                        dm_jiffies_to_msec64(s, shared->tmp.io_ticks[READ]),
 860                        dm_jiffies_to_msec64(s, shared->tmp.io_ticks[WRITE]));
 861                 if (s->n_histogram_entries) {
 862                         unsigned i;
 863                         for (i = 0; i < s->n_histogram_entries + 1; i++) {
 864                                 DMEMIT("%s%llu", !i ? " " : ":", shared->tmp.histogram[i]);
 865                         }
 866                 }
 867                 DMEMIT("\n");
 868
 869                 if (unlikely(sz + 1 >= maxlen))
 870                         goto buffer_overflow;
 871         }
 872
 873         if (clear)
 874                 __dm_stat_clear(s, idx_start, idx_end, false);
 875
 876 buffer_overflow:
 877         mutex_unlock(&stats->mutex);
 878
 879         return 1;
 880 }
 881
 882 static int dm_stats_set_aux(struct dm_stats *stats, int id, const char *aux_data)
 883 {
 884         struct dm_stat *s;
 885         const char *new_aux_data;
 886
 887         mutex_lock(&stats->mutex);
 888
 889         s = __dm_stats_find(stats, id);
 890         if (!s) {
 891                 mutex_unlock(&stats->mutex);
 892                 return -ENOENT;
 893         }
 894
 895         new_aux_data = kstrdup(aux_data, GFP_KERNEL);
 896         if (!new_aux_data) {
 897                 mutex_unlock(&stats->mutex);
 898                 return -ENOMEM;
 899         }
 900
 901         kfree(s->aux_data);
 902         s->aux_data = new_aux_data;
 903
 904         mutex_unlock(&stats->mutex);
 905
 906         return 0;
 907 }
 908
 909 static int parse_histogram(const char *h, unsigned *n_histogram_entries,
 910                            unsigned long long **histogram_boundaries)
 911 {
 912         const char *q;
 913         unsigned n;
 914         unsigned long long last;
 915
 916         *n_histogram_entries = 1;
 917         for (q = h; *q; q++)
 918                 if (*q == ',')
 919                         (*n_histogram_entries)++;
 920
 921         *histogram_boundaries = kmalloc(*n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
 922         if (!*histogram_boundaries)
 923                 return -ENOMEM;
 924
 925         n = 0;
 926         last = 0;
 927         while (1) {
 928                 unsigned long long hi;
 929                 int s;
 930                 char ch;
 931                 s = sscanf(h, "%llu%c", &hi, &ch);
 932                 if (!s || (s == 2 && ch != ','))
 933                         return -EINVAL;
 934                 if (hi <= last)
 935                         return -EINVAL;
 936                 last = hi;
 937                 (*histogram_boundaries)[n] = hi;
 938                 if (s == 1)
 939                         return 0;
 940                 h = strchr(h, ',') + 1;
 941                 n++;
 942         }
 943 }
 944
 945 static int message_stats_create(struct mapped_device *md,
 946                                 unsigned argc, char **argv,
 947                                 char *result, unsigned maxlen)
 948 {
 949         int r;
 950         int id;
 951         char dummy;
 952         unsigned long long start, end, len, step;
 953         unsigned divisor;
 954         const char *program_id, *aux_data;
 955         unsigned stat_flags = 0;
 956
 957         unsigned n_histogram_entries = 0;
 958         unsigned long long *histogram_boundaries = NULL;
 959
 960         struct dm_arg_set as, as_backup;
 961         const char *a;
 962         unsigned feature_args;
 963
 964         /*
 965          * Input format:
 966          *   <range> <step> [<extra_parameters> <parameters>] [<program_id> [<aux_data>]]
 967          */
 968
 969         if (argc < 3)
 970                 goto ret_einval;
 971
 972         as.argc = argc;
 973         as.argv = argv;
 974         dm_consume_args(&as, 1);
 975
 976         a = dm_shift_arg(&as);
 977         if (!strcmp(a, "-")) {
 978                 start = 0;
 979                 len = dm_get_size(md);
 980                 if (!len)
 981                         len = 1;
 982         } else if (sscanf(a, "%llu+%llu%c", &start, &len, &dummy) != 2 ||
 983                    start != (sector_t)start || len != (sector_t)len)
 984                 goto ret_einval;
 985
 986         end = start + len;
 987         if (start >= end)
 988                 goto ret_einval;
 989
 990         a = dm_shift_arg(&as);
 991         if (sscanf(a, "/%u%c", &divisor, &dummy) == 1) {
 992                 if (!divisor)
 993                         return -EINVAL;
 994                 step = end - start;
 995                 if (do_div(step, divisor))
 996                         step++;
 997                 if (!step)
 998                         step = 1;
 999         } else if (sscanf(a, "%llu%c", &step, &dummy) != 1 ||
1000                    step != (sector_t)step || !step)
1001                 goto ret_einval;
1002
1003         as_backup = as;
1004         a = dm_shift_arg(&as);
1005         if (a && sscanf(a, "%u%c", &feature_args, &dummy) == 1) {
1006                 while (feature_args--) {
1007                         a = dm_shift_arg(&as);
1008                         if (!a)
1009                                 goto ret_einval;
1010                         if (!strcasecmp(a, "precise_timestamps"))
1011                                 stat_flags |= STAT_PRECISE_TIMESTAMPS;
1012                         else if (!strncasecmp(a, "histogram:", 10)) {
1013                                 if (n_histogram_entries)
1014                                         goto ret_einval;
1015                                 if ((r = parse_histogram(a + 10, &n_histogram_entries, &histogram_boundaries)))
1016                                         goto ret;
1017                         } else
1018                                 goto ret_einval;
1019                 }
1020         } else {
1021                 as = as_backup;
1022         }
1023
1024         program_id = "-";
1025         aux_data = "-";
1026
1027         a = dm_shift_arg(&as);
1028         if (a)
1029                 program_id = a;
1030
1031         a = dm_shift_arg(&as);
1032         if (a)
1033                 aux_data = a;
1034
1035         if (as.argc)
1036                 goto ret_einval;
1037
1038         /*
1039          * If a buffer overflow happens after we created the region,
1040          * it's too late (the userspace would retry with a larger
1041          * buffer, but the region id that caused the overflow is already
1042          * leaked).  So we must detect buffer overflow in advance.
1043          */
1044         snprintf(result, maxlen, "%d", INT_MAX);
1045         if (dm_message_test_buffer_overflow(result, maxlen)) {
1046                 r = 1;
1047                 goto ret;
1048         }
1049
1050         id = dm_stats_create(dm_get_stats(md), start, end, step, stat_flags,
1051                              n_histogram_entries, histogram_boundaries, program_id, aux_data,
1052                              dm_internal_suspend_fast, dm_internal_resume_fast, md);
1053         if (id < 0) {
1054                 r = id;
1055                 goto ret;
1056         }
1057
1058         snprintf(result, maxlen, "%d", id);
1059
1060         r = 1;
1061         goto ret;
1062
1063 ret_einval:
1064         r = -EINVAL;
1065 ret:
1066         kfree(histogram_boundaries);
1067         return r;
1068 }
1069
1070 static int message_stats_delete(struct mapped_device *md,
1071                                 unsigned argc, char **argv)
1072 {
1073         int id;
1074         char dummy;
1075
1076         if (argc != 2)
1077                 return -EINVAL;
1078
1079         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1080                 return -EINVAL;
1081
1082         return dm_stats_delete(dm_get_stats(md), id);
1083 }
1084
1085 static int message_stats_clear(struct mapped_device *md,
1086                                unsigned argc, char **argv)
1087 {
1088         int id;
1089         char dummy;
1090
1091         if (argc != 2)
1092                 return -EINVAL;
1093
1094         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1095                 return -EINVAL;
1096
1097         return dm_stats_clear(dm_get_stats(md), id);
1098 }
1099
1100 static int message_stats_list(struct mapped_device *md,
1101                               unsigned argc, char **argv,
1102                               char *result, unsigned maxlen)
1103 {
1104         int r;
1105         const char *program = NULL;
1106
1107         if (argc < 1 || argc > 2)
1108                 return -EINVAL;
1109
1110         if (argc > 1) {
1111                 program = kstrdup(argv[1], GFP_KERNEL);
1112                 if (!program)
1113                         return -ENOMEM;
1114         }
1115
1116         r = dm_stats_list(dm_get_stats(md), program, result, maxlen);
1117
1118         kfree(program);
1119
1120         return r;
1121 }
1122
1123 static int message_stats_print(struct mapped_device *md,
1124                                unsigned argc, char **argv, bool clear,
1125                                char *result, unsigned maxlen)
1126 {
1127         int id;
1128         char dummy;
1129         unsigned long idx_start = 0, idx_len = ULONG_MAX;
1130
1131         if (argc != 2 && argc != 4)
1132                 return -EINVAL;
1133
1134         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1135                 return -EINVAL;
1136
1137         if (argc > 3) {
1138                 if (strcmp(argv[2], "-") &&
1139                     sscanf(argv[2], "%lu%c", &idx_start, &dummy) != 1)
1140                         return -EINVAL;
1141                 if (strcmp(argv[3], "-") &&
1142                     sscanf(argv[3], "%lu%c", &idx_len, &dummy) != 1)
1143                         return -EINVAL;
1144         }
1145
1146         return dm_stats_print(dm_get_stats(md), id, idx_start, idx_len, clear,
1147                               result, maxlen);
1148 }
1149
1150 static int message_stats_set_aux(struct mapped_device *md,
1151                                  unsigned argc, char **argv)
1152 {
1153         int id;
1154         char dummy;
1155
1156         if (argc != 3)
1157                 return -EINVAL;
1158
1159         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1160                 return -EINVAL;
1161
1162         return dm_stats_set_aux(dm_get_stats(md), id, argv[2]);
1163 }
1164
1165 int dm_stats_message(struct mapped_device *md, unsigned argc, char **argv,
1166                      char *result, unsigned maxlen)
1167 {
1168         int r;
1169
1170         /* All messages here must start with '@' */
1171         if (!strcasecmp(argv[0], "@stats_create"))
1172                 r = message_stats_create(md, argc, argv, result, maxlen);
1173         else if (!strcasecmp(argv[0], "@stats_delete"))
1174                 r = message_stats_delete(md, argc, argv);
1175         else if (!strcasecmp(argv[0], "@stats_clear"))
1176                 r = message_stats_clear(md, argc, argv);
1177         else if (!strcasecmp(argv[0], "@stats_list"))
1178                 r = message_stats_list(md, argc, argv, result, maxlen);
1179         else if (!strcasecmp(argv[0], "@stats_print"))
1180                 r = message_stats_print(md, argc, argv, false, result, maxlen);
1181         else if (!strcasecmp(argv[0], "@stats_print_clear"))
1182                 r = message_stats_print(md, argc, argv, true, result, maxlen);
1183         else if (!strcasecmp(argv[0], "@stats_set_aux"))
1184                 r = message_stats_set_aux(md, argc, argv);
1185         else
1186                 return 2; /* this wasn't a stats message */
1187
1188         if (r == -EINVAL)
1189                 DMWARN("Invalid parameters for message %s", argv[0]);
1190
1191         return r;
1192 }
1193
1194 int __init dm_statistics_init(void)
1195 {
1196         shared_memory_amount = 0;
1197         dm_stat_need_rcu_barrier = 0;
1198         return 0;
1199 }
1200
1201 void dm_statistics_exit(void)
1202 {
1203         if (dm_stat_need_rcu_barrier)
1204                 rcu_barrier();
1205         if (WARN_ON(shared_memory_amount))
1206                 DMCRIT("shared_memory_amount leaked: %lu", shared_memory_amount);
1207 }
1208
1209 module_param_named(stats_current_allocated_bytes, shared_memory_amount, ulong, S_IRUGO);
1210 MODULE_PARM_DESC(stats_current_allocated_bytes, "Memory currently used by statistics");