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->histogram_boundaries);
 179         kfree(s->program_id);
 180         kfree(s->aux_data);
 181         for_each_possible_cpu(cpu) {
 182                 dm_kvfree(s->stat_percpu[cpu][0].histogram, s->histogram_alloc_size);
 183                 dm_kvfree(s->stat_percpu[cpu], s->percpu_alloc_size);
 184         }
 185         dm_kvfree(s->stat_shared[0].tmp.histogram, s->histogram_alloc_size);
 186         dm_kvfree(s, s->shared_alloc_size);
 187 }
 188
 189 static int dm_stat_in_flight(struct dm_stat_shared *shared)
 190 {
 191         return atomic_read(&shared->in_flight[READ]) +
 192                atomic_read(&shared->in_flight[WRITE]);
 193 }
 194
 195 void dm_stats_init(struct dm_stats *stats)
 196 {
 197         int cpu;
 198         struct dm_stats_last_position *last;
 199
 200         mutex_init(&stats->mutex);
 201         INIT_LIST_HEAD(&stats->list);
 202         stats->last = alloc_percpu(struct dm_stats_last_position);
 203         for_each_possible_cpu(cpu) {
 204                 last = per_cpu_ptr(stats->last, cpu);
 205                 last->last_sector = (sector_t)ULLONG_MAX;
 206                 last->last_rw = UINT_MAX;
 207         }
 208 }
 209
 210 void dm_stats_cleanup(struct dm_stats *stats)
 211 {
 212         size_t ni;
 213         struct dm_stat *s;
 214         struct dm_stat_shared *shared;
 215
 216         while (!list_empty(&stats->list)) {
 217                 s = container_of(stats->list.next, struct dm_stat, list_entry);
 218                 list_del(&s->list_entry);
 219                 for (ni = 0; ni < s->n_entries; ni++) {
 220                         shared = &s->stat_shared[ni];
 221                         if (WARN_ON(dm_stat_in_flight(shared))) {
 222                                 DMCRIT("leaked in-flight counter at index %lu "
 223                                        "(start %llu, end %llu, step %llu): reads %d, writes %d",
 224                                        (unsigned long)ni,
 225                                        (unsigned long long)s->start,
 226                                        (unsigned long long)s->end,
 227                                        (unsigned long long)s->step,
 228                                        atomic_read(&shared->in_flight[READ]),
 229                                        atomic_read(&shared->in_flight[WRITE]));
 230                         }
 231                 }
 232                 dm_stat_free(&s->rcu_head);
 233         }
 234         free_percpu(stats->last);
 235 }
 236
 237 static int dm_stats_create(struct dm_stats *stats, sector_t start, sector_t end,
 238                            sector_t step, unsigned stat_flags,
 239                            unsigned n_histogram_entries,
 240                            unsigned long long *histogram_boundaries,
 241                            const char *program_id, const char *aux_data,
 242                            void (*suspend_callback)(struct mapped_device *),
 243                            void (*resume_callback)(struct mapped_device *),
 244                            struct mapped_device *md)
 245 {
 246         struct list_head *l;
 247         struct dm_stat *s, *tmp_s;
 248         sector_t n_entries;
 249         size_t ni;
 250         size_t shared_alloc_size;
 251         size_t percpu_alloc_size;
 252         size_t histogram_alloc_size;
 253         struct dm_stat_percpu *p;
 254         int cpu;
 255         int ret_id;
 256         int r;
 257
 258         if (end < start || !step)
 259                 return -EINVAL;
 260
 261         n_entries = end - start;
 262         if (dm_sector_div64(n_entries, step))
 263                 n_entries++;
 264
 265         if (n_entries != (size_t)n_entries || !(size_t)(n_entries + 1))
 266                 return -EOVERFLOW;
 267
 268         shared_alloc_size = sizeof(struct dm_stat) + (size_t)n_entries * sizeof(struct dm_stat_shared);
 269         if ((shared_alloc_size - sizeof(struct dm_stat)) / sizeof(struct dm_stat_shared) != n_entries)
 270                 return -EOVERFLOW;
 271
 272         percpu_alloc_size = (size_t)n_entries * sizeof(struct dm_stat_percpu);
 273         if (percpu_alloc_size / sizeof(struct dm_stat_percpu) != n_entries)
 274                 return -EOVERFLOW;
 275
 276         histogram_alloc_size = (n_histogram_entries + 1) * (size_t)n_entries * sizeof(unsigned long long);
 277         if (histogram_alloc_size / (n_histogram_entries + 1) != (size_t)n_entries * sizeof(unsigned long long))
 278                 return -EOVERFLOW;
 279
 280         if (!check_shared_memory(shared_alloc_size + histogram_alloc_size +
 281                                  num_possible_cpus() * (percpu_alloc_size + histogram_alloc_size)))
 282                 return -ENOMEM;
 283
 284         s = dm_kvzalloc(shared_alloc_size, NUMA_NO_NODE);
 285         if (!s)
 286                 return -ENOMEM;
 287
 288         s->stat_flags = stat_flags;
 289         s->n_entries = n_entries;
 290         s->start = start;
 291         s->end = end;
 292         s->step = step;
 293         s->shared_alloc_size = shared_alloc_size;
 294         s->percpu_alloc_size = percpu_alloc_size;
 295         s->histogram_alloc_size = histogram_alloc_size;
 296
 297         s->n_histogram_entries = n_histogram_entries;
 298         s->histogram_boundaries = kmemdup(histogram_boundaries,
 299                                           s->n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
 300         if (!s->histogram_boundaries) {
 301                 r = -ENOMEM;
 302                 goto out;
 303         }
 304
 305         s->program_id = kstrdup(program_id, GFP_KERNEL);
 306         if (!s->program_id) {
 307                 r = -ENOMEM;
 308                 goto out;
 309         }
 310         s->aux_data = kstrdup(aux_data, GFP_KERNEL);
 311         if (!s->aux_data) {
 312                 r = -ENOMEM;
 313                 goto out;
 314         }
 315
 316         for (ni = 0; ni < n_entries; ni++) {
 317                 atomic_set(&s->stat_shared[ni].in_flight[READ], 0);
 318                 atomic_set(&s->stat_shared[ni].in_flight[WRITE], 0);
 319         }
 320
 321         if (s->n_histogram_entries) {
 322                 unsigned long long *hi;
 323                 hi = dm_kvzalloc(s->histogram_alloc_size, NUMA_NO_NODE);
 324                 if (!hi) {
 325                         r = -ENOMEM;
 326                         goto out;
 327                 }
 328                 for (ni = 0; ni < n_entries; ni++) {
 329                         s->stat_shared[ni].tmp.histogram = hi;
 330                         hi += s->n_histogram_entries + 1;
 331                 }
 332         }
 333
 334         for_each_possible_cpu(cpu) {
 335                 p = dm_kvzalloc(percpu_alloc_size, cpu_to_node(cpu));
 336                 if (!p) {
 337                         r = -ENOMEM;
 338                         goto out;
 339                 }
 340                 s->stat_percpu[cpu] = p;
 341                 if (s->n_histogram_entries) {
 342                         unsigned long long *hi;
 343                         hi = dm_kvzalloc(s->histogram_alloc_size, cpu_to_node(cpu));
 344                         if (!hi) {
 345                                 r = -ENOMEM;
 346                                 goto out;
 347                         }
 348                         for (ni = 0; ni < n_entries; ni++) {
 349                                 p[ni].histogram = hi;
 350                                 hi += s->n_histogram_entries + 1;
 351                         }
 352                 }
 353         }
 354
 355         /*
 356          * Suspend/resume to make sure there is no i/o in flight,
 357          * so that newly created statistics will be exact.
 358          *
 359          * (note: we couldn't suspend earlier because we must not
 360          * allocate memory while suspended)
 361          */
 362         suspend_callback(md);
 363
 364         mutex_lock(&stats->mutex);
 365         s->id = 0;
 366         list_for_each(l, &stats->list) {
 367                 tmp_s = container_of(l, struct dm_stat, list_entry);
 368                 if (WARN_ON(tmp_s->id < s->id)) {
 369                         r = -EINVAL;
 370                         goto out_unlock_resume;
 371                 }
 372                 if (tmp_s->id > s->id)
 373                         break;
 374                 if (unlikely(s->id == INT_MAX)) {
 375                         r = -ENFILE;
 376                         goto out_unlock_resume;
 377                 }
 378                 s->id++;
 379         }
 380         ret_id = s->id;
 381         list_add_tail_rcu(&s->list_entry, l);
 382         mutex_unlock(&stats->mutex);
 383
 384         resume_callback(md);
 385
 386         return ret_id;
 387
 388 out_unlock_resume:
 389         mutex_unlock(&stats->mutex);
 390         resume_callback(md);
 391 out:
 392         dm_stat_free(&s->rcu_head);
 393         return r;
 394 }
 395
 396 static struct dm_stat *__dm_stats_find(struct dm_stats *stats, int id)
 397 {
 398         struct dm_stat *s;
 399
 400         list_for_each_entry(s, &stats->list, list_entry) {
 401                 if (s->id > id)
 402                         break;
 403                 if (s->id == id)
 404                         return s;
 405         }
 406
 407         return NULL;
 408 }
 409
 410 static int dm_stats_delete(struct dm_stats *stats, int id)
 411 {
 412         struct dm_stat *s;
 413         int cpu;
 414
 415         mutex_lock(&stats->mutex);
 416
 417         s = __dm_stats_find(stats, id);
 418         if (!s) {
 419                 mutex_unlock(&stats->mutex);
 420                 return -ENOENT;
 421         }
 422
 423         list_del_rcu(&s->list_entry);
 424         mutex_unlock(&stats->mutex);
 425
 426         /*
 427          * vfree can't be called from RCU callback
 428          */
 429         for_each_possible_cpu(cpu)
 430                 if (is_vmalloc_addr(s->stat_percpu) ||
 431                     is_vmalloc_addr(s->stat_percpu[cpu][0].histogram))
 432                         goto do_sync_free;
 433         if (is_vmalloc_addr(s) ||
 434             is_vmalloc_addr(s->stat_shared[0].tmp.histogram)) {
 435 do_sync_free:
 436                 synchronize_rcu_expedited();
 437                 dm_stat_free(&s->rcu_head);
 438         } else {
 439                 ACCESS_ONCE(dm_stat_need_rcu_barrier) = 1;
 440                 call_rcu(&s->rcu_head, dm_stat_free);
 441         }
 442         return 0;
 443 }
 444
 445 static int dm_stats_list(struct dm_stats *stats, const char *program,
 446                          char *result, unsigned maxlen)
 447 {
 448         struct dm_stat *s;
 449         sector_t len;
 450         unsigned sz = 0;
 451
 452         /*
 453          * Output format:
 454          *   <region_id>: <start_sector>+<length> <step> <program_id> <aux_data>
 455          */
 456
 457         mutex_lock(&stats->mutex);
 458         list_for_each_entry(s, &stats->list, list_entry) {
 459                 if (!program || !strcmp(program, s->program_id)) {
 460                         len = s->end - s->start;
 461                         DMEMIT("%d: %llu+%llu %llu %s %s", s->id,
 462                                 (unsigned long long)s->start,
 463                                 (unsigned long long)len,
 464                                 (unsigned long long)s->step,
 465                                 s->program_id,
 466                                 s->aux_data);
 467                         if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
 468                                 DMEMIT(" precise_timestamps");
 469                         if (s->n_histogram_entries) {
 470                                 unsigned i;
 471                                 DMEMIT(" histogram:");
 472                                 for (i = 0; i < s->n_histogram_entries; i++) {
 473                                         if (i)
 474                                                 DMEMIT(",");
 475                                         DMEMIT("%llu", s->histogram_boundaries[i]);
 476                                 }
 477                         }
 478                         DMEMIT("\n");
 479                 }
 480         }
 481         mutex_unlock(&stats->mutex);
 482
 483         return 1;
 484 }
 485
 486 static void dm_stat_round(struct dm_stat *s, struct dm_stat_shared *shared,
 487                           struct dm_stat_percpu *p)
 488 {
 489         /*
 490          * This is racy, but so is part_round_stats_single.
 491          */
 492         unsigned long long now, difference;
 493         unsigned in_flight_read, in_flight_write;
 494
 495         if (likely(!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)))
 496                 now = jiffies;
 497         else
 498                 now = ktime_to_ns(ktime_get());
 499
 500         difference = now - shared->stamp;
 501         if (!difference)
 502                 return;
 503
 504         in_flight_read = (unsigned)atomic_read(&shared->in_flight[READ]);
 505         in_flight_write = (unsigned)atomic_read(&shared->in_flight[WRITE]);
 506         if (in_flight_read)
 507                 p->io_ticks[READ] += difference;
 508         if (in_flight_write)
 509                 p->io_ticks[WRITE] += difference;
 510         if (in_flight_read + in_flight_write) {
 511                 p->io_ticks_total += difference;
 512                 p->time_in_queue += (in_flight_read + in_flight_write) * difference;
 513         }
 514         shared->stamp = now;
 515 }
 516
 517 static void dm_stat_for_entry(struct dm_stat *s, size_t entry,
 518                               unsigned long bi_rw, sector_t len,
 519                               struct dm_stats_aux *stats_aux, bool end,
 520                               unsigned long duration_jiffies)
 521 {
 522         unsigned long idx = bi_rw & REQ_WRITE;
 523         struct dm_stat_shared *shared = &s->stat_shared[entry];
 524         struct dm_stat_percpu *p;
 525
 526         /*
 527          * For strict correctness we should use local_irq_save/restore
 528          * instead of preempt_disable/enable.
 529          *
 530          * preempt_disable/enable is racy if the driver finishes bios
 531          * from non-interrupt context as well as from interrupt context
 532          * or from more different interrupts.
 533          *
 534          * On 64-bit architectures the race only results in not counting some
 535          * events, so it is acceptable.  On 32-bit architectures the race could
 536          * cause the counter going off by 2^32, so we need to do proper locking
 537          * there.
 538          *
 539          * part_stat_lock()/part_stat_unlock() have this race too.
 540          */
 541 #if BITS_PER_LONG == 32
 542         unsigned long flags;
 543         local_irq_save(flags);
 544 #else
 545         preempt_disable();
 546 #endif
 547         p = &s->stat_percpu[smp_processor_id()][entry];
 548
 549         if (!end) {
 550                 dm_stat_round(s, shared, p);
 551                 atomic_inc(&shared->in_flight[idx]);
 552         } else {
 553                 unsigned long long duration;
 554                 dm_stat_round(s, shared, p);
 555                 atomic_dec(&shared->in_flight[idx]);
 556                 p->sectors[idx] += len;
 557                 p->ios[idx] += 1;
 558                 p->merges[idx] += stats_aux->merged;
 559                 if (!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)) {
 560                         p->ticks[idx] += duration_jiffies;
 561                         duration = jiffies_to_msecs(duration_jiffies);
 562                 } else {
 563                         p->ticks[idx] += stats_aux->duration_ns;
 564                         duration = stats_aux->duration_ns;
 565                 }
 566                 if (s->n_histogram_entries) {
 567                         unsigned lo = 0, hi = s->n_histogram_entries + 1;
 568                         while (lo + 1 < hi) {
 569                                 unsigned mid = (lo + hi) / 2;
 570                                 if (s->histogram_boundaries[mid - 1] > duration) {
 571                                         hi = mid;
 572                                 } else {
 573                                         lo = mid;
 574                                 }
 575
 576                         }
 577                         p->histogram[lo]++;
 578                 }
 579         }
 580
 581 #if BITS_PER_LONG == 32
 582         local_irq_restore(flags);
 583 #else
 584         preempt_enable();
 585 #endif
 586 }
 587
 588 static void __dm_stat_bio(struct dm_stat *s, unsigned long bi_rw,
 589                           sector_t bi_sector, sector_t end_sector,
 590                           bool end, unsigned long duration_jiffies,
 591                           struct dm_stats_aux *stats_aux)
 592 {
 593         sector_t rel_sector, offset, todo, fragment_len;
 594         size_t entry;
 595
 596         if (end_sector <= s->start || bi_sector >= s->end)
 597                 return;
 598         if (unlikely(bi_sector < s->start)) {
 599                 rel_sector = 0;
 600                 todo = end_sector - s->start;
 601         } else {
 602                 rel_sector = bi_sector - s->start;
 603                 todo = end_sector - bi_sector;
 604         }
 605         if (unlikely(end_sector > s->end))
 606                 todo -= (end_sector - s->end);
 607
 608         offset = dm_sector_div64(rel_sector, s->step);
 609         entry = rel_sector;
 610         do {
 611                 if (WARN_ON_ONCE(entry >= s->n_entries)) {
 612                         DMCRIT("Invalid area access in region id %d", s->id);
 613                         return;
 614                 }
 615                 fragment_len = todo;
 616                 if (fragment_len > s->step - offset)
 617                         fragment_len = s->step - offset;
 618                 dm_stat_for_entry(s, entry, bi_rw, fragment_len,
 619                                   stats_aux, end, duration_jiffies);
 620                 todo -= fragment_len;
 621                 entry++;
 622                 offset = 0;
 623         } while (unlikely(todo != 0));
 624 }
 625
 626 void dm_stats_account_io(struct dm_stats *stats, unsigned long bi_rw,
 627                          sector_t bi_sector, unsigned bi_sectors, bool end,
 628                          unsigned long duration_jiffies,
 629                          struct dm_stats_aux *stats_aux)
 630 {
 631         struct dm_stat *s;
 632         sector_t end_sector;
 633         struct dm_stats_last_position *last;
 634         bool got_precise_time;
 635
 636         if (unlikely(!bi_sectors))
 637                 return;
 638
 639         end_sector = bi_sector + bi_sectors;
 640
 641         if (!end) {
 642                 /*
 643                  * A race condition can at worst result in the merged flag being
 644                  * misrepresented, so we don't have to disable preemption here.
 645                  */
 646                 last = raw_cpu_ptr(stats->last);
 647                 stats_aux->merged =
 648                         (bi_sector == (ACCESS_ONCE(last->last_sector) &&
 649                                        ((bi_rw & (REQ_WRITE | REQ_DISCARD)) ==
 650                                         (ACCESS_ONCE(last->last_rw) & (REQ_WRITE | REQ_DISCARD)))
 651                                        ));
 652                 ACCESS_ONCE(last->last_sector) = end_sector;
 653                 ACCESS_ONCE(last->last_rw) = bi_rw;
 654         }
 655
 656         rcu_read_lock();
 657
 658         got_precise_time = false;
 659         list_for_each_entry_rcu(s, &stats->list, list_entry) {
 660                 if (s->stat_flags & STAT_PRECISE_TIMESTAMPS && !got_precise_time) {
 661                         if (!end)
 662                                 stats_aux->duration_ns = ktime_to_ns(ktime_get());
 663                         else
 664                                 stats_aux->duration_ns = ktime_to_ns(ktime_get()) - stats_aux->duration_ns;
 665                         got_precise_time = true;
 666                 }
 667                 __dm_stat_bio(s, bi_rw, bi_sector, end_sector, end, duration_jiffies, stats_aux);
 668         }
 669
 670         rcu_read_unlock();
 671 }
 672
 673 static void __dm_stat_init_temporary_percpu_totals(struct dm_stat_shared *shared,
 674                                                    struct dm_stat *s, size_t x)
 675 {
 676         int cpu;
 677         struct dm_stat_percpu *p;
 678
 679         local_irq_disable();
 680         p = &s->stat_percpu[smp_processor_id()][x];
 681         dm_stat_round(s, shared, p);
 682         local_irq_enable();
 683
 684         shared->tmp.sectors[READ] = 0;
 685         shared->tmp.sectors[WRITE] = 0;
 686         shared->tmp.ios[READ] = 0;
 687         shared->tmp.ios[WRITE] = 0;
 688         shared->tmp.merges[READ] = 0;
 689         shared->tmp.merges[WRITE] = 0;
 690         shared->tmp.ticks[READ] = 0;
 691         shared->tmp.ticks[WRITE] = 0;
 692         shared->tmp.io_ticks[READ] = 0;
 693         shared->tmp.io_ticks[WRITE] = 0;
 694         shared->tmp.io_ticks_total = 0;
 695         shared->tmp.time_in_queue = 0;
 696
 697         if (s->n_histogram_entries)
 698                 memset(shared->tmp.histogram, 0, (s->n_histogram_entries + 1) * sizeof(unsigned long long));
 699
 700         for_each_possible_cpu(cpu) {
 701                 p = &s->stat_percpu[cpu][x];
 702                 shared->tmp.sectors[READ] += ACCESS_ONCE(p->sectors[READ]);
 703                 shared->tmp.sectors[WRITE] += ACCESS_ONCE(p->sectors[WRITE]);
 704                 shared->tmp.ios[READ] += ACCESS_ONCE(p->ios[READ]);
 705                 shared->tmp.ios[WRITE] += ACCESS_ONCE(p->ios[WRITE]);
 706                 shared->tmp.merges[READ] += ACCESS_ONCE(p->merges[READ]);
 707                 shared->tmp.merges[WRITE] += ACCESS_ONCE(p->merges[WRITE]);
 708                 shared->tmp.ticks[READ] += ACCESS_ONCE(p->ticks[READ]);
 709                 shared->tmp.ticks[WRITE] += ACCESS_ONCE(p->ticks[WRITE]);
 710                 shared->tmp.io_ticks[READ] += ACCESS_ONCE(p->io_ticks[READ]);
 711                 shared->tmp.io_ticks[WRITE] += ACCESS_ONCE(p->io_ticks[WRITE]);
 712                 shared->tmp.io_ticks_total += ACCESS_ONCE(p->io_ticks_total);
 713                 shared->tmp.time_in_queue += ACCESS_ONCE(p->time_in_queue);
 714                 if (s->n_histogram_entries) {
 715                         unsigned i;
 716                         for (i = 0; i < s->n_histogram_entries + 1; i++)
 717                                 shared->tmp.histogram[i] += ACCESS_ONCE(p->histogram[i]);
 718                 }
 719         }
 720 }
 721
 722 static void __dm_stat_clear(struct dm_stat *s, size_t idx_start, size_t idx_end,
 723                             bool init_tmp_percpu_totals)
 724 {
 725         size_t x;
 726         struct dm_stat_shared *shared;
 727         struct dm_stat_percpu *p;
 728
 729         for (x = idx_start; x < idx_end; x++) {
 730                 shared = &s->stat_shared[x];
 731                 if (init_tmp_percpu_totals)
 732                         __dm_stat_init_temporary_percpu_totals(shared, s, x);
 733                 local_irq_disable();
 734                 p = &s->stat_percpu[smp_processor_id()][x];
 735                 p->sectors[READ] -= shared->tmp.sectors[READ];
 736                 p->sectors[WRITE] -= shared->tmp.sectors[WRITE];
 737                 p->ios[READ] -= shared->tmp.ios[READ];
 738                 p->ios[WRITE] -= shared->tmp.ios[WRITE];
 739                 p->merges[READ] -= shared->tmp.merges[READ];
 740                 p->merges[WRITE] -= shared->tmp.merges[WRITE];
 741                 p->ticks[READ] -= shared->tmp.ticks[READ];
 742                 p->ticks[WRITE] -= shared->tmp.ticks[WRITE];
 743                 p->io_ticks[READ] -= shared->tmp.io_ticks[READ];
 744                 p->io_ticks[WRITE] -= shared->tmp.io_ticks[WRITE];
 745                 p->io_ticks_total -= shared->tmp.io_ticks_total;
 746                 p->time_in_queue -= shared->tmp.time_in_queue;
 747                 local_irq_enable();
 748                 if (s->n_histogram_entries) {
 749                         unsigned i;
 750                         for (i = 0; i < s->n_histogram_entries + 1; i++) {
 751                                 local_irq_disable();
 752                                 p = &s->stat_percpu[smp_processor_id()][x];
 753                                 p->histogram[i] -= shared->tmp.histogram[i];
 754                                 local_irq_enable();
 755                         }
 756                 }
 757         }
 758 }
 759
 760 static int dm_stats_clear(struct dm_stats *stats, int id)
 761 {
 762         struct dm_stat *s;
 763
 764         mutex_lock(&stats->mutex);
 765
 766         s = __dm_stats_find(stats, id);
 767         if (!s) {
 768                 mutex_unlock(&stats->mutex);
 769                 return -ENOENT;
 770         }
 771
 772         __dm_stat_clear(s, 0, s->n_entries, true);
 773
 774         mutex_unlock(&stats->mutex);
 775
 776         return 1;
 777 }
 778
 779 /*
 780  * This is like jiffies_to_msec, but works for 64-bit values.
 781  */
 782 static unsigned long long dm_jiffies_to_msec64(struct dm_stat *s, unsigned long long j)
 783 {
 784         unsigned long long result;
 785         unsigned mult;
 786
 787         if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
 788                 return j;
 789
 790         result = 0;
 791         if (j)
 792                 result = jiffies_to_msecs(j & 0x3fffff);
 793         if (j >= 1 << 22) {
 794                 mult = jiffies_to_msecs(1 << 22);
 795                 result += (unsigned long long)mult * (unsigned long long)jiffies_to_msecs((j >> 22) & 0x3fffff);
 796         }
 797         if (j >= 1ULL << 44)
 798                 result += (unsigned long long)mult * (unsigned long long)mult * (unsigned long long)jiffies_to_msecs(j >> 44);
 799
 800         return result;
 801 }
 802
 803 static int dm_stats_print(struct dm_stats *stats, int id,
 804                           size_t idx_start, size_t idx_len,
 805                           bool clear, char *result, unsigned maxlen)
 806 {
 807         unsigned sz = 0;
 808         struct dm_stat *s;
 809         size_t x;
 810         sector_t start, end, step;
 811         size_t idx_end;
 812         struct dm_stat_shared *shared;
 813
 814         /*
 815          * Output format:
 816          *   <start_sector>+<length> counters
 817          */
 818
 819         mutex_lock(&stats->mutex);
 820
 821         s = __dm_stats_find(stats, id);
 822         if (!s) {
 823                 mutex_unlock(&stats->mutex);
 824                 return -ENOENT;
 825         }
 826
 827         idx_end = idx_start + idx_len;
 828         if (idx_end < idx_start ||
 829             idx_end > s->n_entries)
 830                 idx_end = s->n_entries;
 831
 832         if (idx_start > idx_end)
 833                 idx_start = idx_end;
 834
 835         step = s->step;
 836         start = s->start + (step * idx_start);
 837
 838         for (x = idx_start; x < idx_end; x++, start = end) {
 839                 shared = &s->stat_shared[x];
 840                 end = start + step;
 841                 if (unlikely(end > s->end))
 842                         end = s->end;
 843
 844                 __dm_stat_init_temporary_percpu_totals(shared, s, x);
 845
 846                 DMEMIT("%llu+%llu %llu %llu %llu %llu %llu %llu %llu %llu %d %llu %llu %llu %llu",
 847                        (unsigned long long)start,
 848                        (unsigned long long)step,
 849                        shared->tmp.ios[READ],
 850                        shared->tmp.merges[READ],
 851                        shared->tmp.sectors[READ],
 852                        dm_jiffies_to_msec64(s, shared->tmp.ticks[READ]),
 853                        shared->tmp.ios[WRITE],
 854                        shared->tmp.merges[WRITE],
 855                        shared->tmp.sectors[WRITE],
 856                        dm_jiffies_to_msec64(s, shared->tmp.ticks[WRITE]),
 857                        dm_stat_in_flight(shared),
 858                        dm_jiffies_to_msec64(s, shared->tmp.io_ticks_total),
 859                        dm_jiffies_to_msec64(s, shared->tmp.time_in_queue),
 860                        dm_jiffies_to_msec64(s, shared->tmp.io_ticks[READ]),
 861                        dm_jiffies_to_msec64(s, shared->tmp.io_ticks[WRITE]));
 862                 if (s->n_histogram_entries) {
 863                         unsigned i;
 864                         for (i = 0; i < s->n_histogram_entries + 1; i++) {
 865                                 DMEMIT("%s%llu", !i ? " " : ":", shared->tmp.histogram[i]);
 866                         }
 867                 }
 868                 DMEMIT("\n");
 869
 870                 if (unlikely(sz + 1 >= maxlen))
 871                         goto buffer_overflow;
 872         }
 873
 874         if (clear)
 875                 __dm_stat_clear(s, idx_start, idx_end, false);
 876
 877 buffer_overflow:
 878         mutex_unlock(&stats->mutex);
 879
 880         return 1;
 881 }
 882
 883 static int dm_stats_set_aux(struct dm_stats *stats, int id, const char *aux_data)
 884 {
 885         struct dm_stat *s;
 886         const char *new_aux_data;
 887
 888         mutex_lock(&stats->mutex);
 889
 890         s = __dm_stats_find(stats, id);
 891         if (!s) {
 892                 mutex_unlock(&stats->mutex);
 893                 return -ENOENT;
 894         }
 895
 896         new_aux_data = kstrdup(aux_data, GFP_KERNEL);
 897         if (!new_aux_data) {
 898                 mutex_unlock(&stats->mutex);
 899                 return -ENOMEM;
 900         }
 901
 902         kfree(s->aux_data);
 903         s->aux_data = new_aux_data;
 904
 905         mutex_unlock(&stats->mutex);
 906
 907         return 0;
 908 }
 909
 910 static int parse_histogram(const char *h, unsigned *n_histogram_entries,
 911                            unsigned long long **histogram_boundaries)
 912 {
 913         const char *q;
 914         unsigned n;
 915         unsigned long long last;
 916
 917         *n_histogram_entries = 1;
 918         for (q = h; *q; q++)
 919                 if (*q == ',')
 920                         (*n_histogram_entries)++;
 921
 922         *histogram_boundaries = kmalloc(*n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
 923         if (!*histogram_boundaries)
 924                 return -ENOMEM;
 925
 926         n = 0;
 927         last = 0;
 928         while (1) {
 929                 unsigned long long hi;
 930                 int s;
 931                 char ch;
 932                 s = sscanf(h, "%llu%c", &hi, &ch);
 933                 if (!s || (s == 2 && ch != ','))
 934                         return -EINVAL;
 935                 if (hi <= last)
 936                         return -EINVAL;
 937                 last = hi;
 938                 (*histogram_boundaries)[n] = hi;
 939                 if (s == 1)
 940                         return 0;
 941                 h = strchr(h, ',') + 1;
 942                 n++;
 943         }
 944 }
 945
 946 static int message_stats_create(struct mapped_device *md,
 947                                 unsigned argc, char **argv,
 948                                 char *result, unsigned maxlen)
 949 {
 950         int r;
 951         int id;
 952         char dummy;
 953         unsigned long long start, end, len, step;
 954         unsigned divisor;
 955         const char *program_id, *aux_data;
 956         unsigned stat_flags = 0;
 957
 958         unsigned n_histogram_entries = 0;
 959         unsigned long long *histogram_boundaries = NULL;
 960
 961         struct dm_arg_set as, as_backup;
 962         const char *a;
 963         unsigned feature_args;
 964
 965         /*
 966          * Input format:
 967          *   <range> <step> [<extra_parameters> <parameters>] [<program_id> [<aux_data>]]
 968          */
 969
 970         if (argc < 3)
 971                 goto ret_einval;
 972
 973         as.argc = argc;
 974         as.argv = argv;
 975         dm_consume_args(&as, 1);
 976
 977         a = dm_shift_arg(&as);
 978         if (!strcmp(a, "-")) {
 979                 start = 0;
 980                 len = dm_get_size(md);
 981                 if (!len)
 982                         len = 1;
 983         } else if (sscanf(a, "%llu+%llu%c", &start, &len, &dummy) != 2 ||
 984                    start != (sector_t)start || len != (sector_t)len)
 985                 goto ret_einval;
 986
 987         end = start + len;
 988         if (start >= end)
 989                 goto ret_einval;
 990
 991         a = dm_shift_arg(&as);
 992         if (sscanf(a, "/%u%c", &divisor, &dummy) == 1) {
 993                 if (!divisor)
 994                         return -EINVAL;
 995                 step = end - start;
 996                 if (do_div(step, divisor))
 997                         step++;
 998                 if (!step)
 999                         step = 1;
1000         } else if (sscanf(a, "%llu%c", &step, &dummy) != 1 ||
1001                    step != (sector_t)step || !step)
1002                 goto ret_einval;
1003
1004         as_backup = as;
1005         a = dm_shift_arg(&as);
1006         if (a && sscanf(a, "%u%c", &feature_args, &dummy) == 1) {
1007                 while (feature_args--) {
1008                         a = dm_shift_arg(&as);
1009                         if (!a)
1010                                 goto ret_einval;
1011                         if (!strcasecmp(a, "precise_timestamps"))
1012                                 stat_flags |= STAT_PRECISE_TIMESTAMPS;
1013                         else if (!strncasecmp(a, "histogram:", 10)) {
1014                                 if (n_histogram_entries)
1015                                         goto ret_einval;
1016                                 if ((r = parse_histogram(a + 10, &n_histogram_entries, &histogram_boundaries)))
1017                                         goto ret;
1018                         } else
1019                                 goto ret_einval;
1020                 }
1021         } else {
1022                 as = as_backup;
1023         }
1024
1025         program_id = "-";
1026         aux_data = "-";
1027
1028         a = dm_shift_arg(&as);
1029         if (a)
1030                 program_id = a;
1031
1032         a = dm_shift_arg(&as);
1033         if (a)
1034                 aux_data = a;
1035
1036         if (as.argc)
1037                 goto ret_einval;
1038
1039         /*
1040          * If a buffer overflow happens after we created the region,
1041          * it's too late (the userspace would retry with a larger
1042          * buffer, but the region id that caused the overflow is already
1043          * leaked).  So we must detect buffer overflow in advance.
1044          */
1045         snprintf(result, maxlen, "%d", INT_MAX);
1046         if (dm_message_test_buffer_overflow(result, maxlen)) {
1047                 r = 1;
1048                 goto ret;
1049         }
1050
1051         id = dm_stats_create(dm_get_stats(md), start, end, step, stat_flags,
1052                              n_histogram_entries, histogram_boundaries, program_id, aux_data,
1053                              dm_internal_suspend_fast, dm_internal_resume_fast, md);
1054         if (id < 0) {
1055                 r = id;
1056                 goto ret;
1057         }
1058
1059         snprintf(result, maxlen, "%d", id);
1060
1061         r = 1;
1062         goto ret;
1063
1064 ret_einval:
1065         r = -EINVAL;
1066 ret:
1067         kfree(histogram_boundaries);
1068         return r;
1069 }
1070
1071 static int message_stats_delete(struct mapped_device *md,
1072                                 unsigned argc, char **argv)
1073 {
1074         int id;
1075         char dummy;
1076
1077         if (argc != 2)
1078                 return -EINVAL;
1079
1080         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1081                 return -EINVAL;
1082
1083         return dm_stats_delete(dm_get_stats(md), id);
1084 }
1085
1086 static int message_stats_clear(struct mapped_device *md,
1087                                unsigned argc, char **argv)
1088 {
1089         int id;
1090         char dummy;
1091
1092         if (argc != 2)
1093                 return -EINVAL;
1094
1095         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1096                 return -EINVAL;
1097
1098         return dm_stats_clear(dm_get_stats(md), id);
1099 }
1100
1101 static int message_stats_list(struct mapped_device *md,
1102                               unsigned argc, char **argv,
1103                               char *result, unsigned maxlen)
1104 {
1105         int r;
1106         const char *program = NULL;
1107
1108         if (argc < 1 || argc > 2)
1109                 return -EINVAL;
1110
1111         if (argc > 1) {
1112                 program = kstrdup(argv[1], GFP_KERNEL);
1113                 if (!program)
1114                         return -ENOMEM;
1115         }
1116
1117         r = dm_stats_list(dm_get_stats(md), program, result, maxlen);
1118
1119         kfree(program);
1120
1121         return r;
1122 }
1123
1124 static int message_stats_print(struct mapped_device *md,
1125                                unsigned argc, char **argv, bool clear,
1126                                char *result, unsigned maxlen)
1127 {
1128         int id;
1129         char dummy;
1130         unsigned long idx_start = 0, idx_len = ULONG_MAX;
1131
1132         if (argc != 2 && argc != 4)
1133                 return -EINVAL;
1134
1135         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1136                 return -EINVAL;
1137
1138         if (argc > 3) {
1139                 if (strcmp(argv[2], "-") &&
1140                     sscanf(argv[2], "%lu%c", &idx_start, &dummy) != 1)
1141                         return -EINVAL;
1142                 if (strcmp(argv[3], "-") &&
1143                     sscanf(argv[3], "%lu%c", &idx_len, &dummy) != 1)
1144                         return -EINVAL;
1145         }
1146
1147         return dm_stats_print(dm_get_stats(md), id, idx_start, idx_len, clear,
1148                               result, maxlen);
1149 }
1150
1151 static int message_stats_set_aux(struct mapped_device *md,
1152                                  unsigned argc, char **argv)
1153 {
1154         int id;
1155         char dummy;
1156
1157         if (argc != 3)
1158                 return -EINVAL;
1159
1160         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1161                 return -EINVAL;
1162
1163         return dm_stats_set_aux(dm_get_stats(md), id, argv[2]);
1164 }
1165
1166 int dm_stats_message(struct mapped_device *md, unsigned argc, char **argv,
1167                      char *result, unsigned maxlen)
1168 {
1169         int r;
1170
1171         /* All messages here must start with '@' */
1172         if (!strcasecmp(argv[0], "@stats_create"))
1173                 r = message_stats_create(md, argc, argv, result, maxlen);
1174         else if (!strcasecmp(argv[0], "@stats_delete"))
1175                 r = message_stats_delete(md, argc, argv);
1176         else if (!strcasecmp(argv[0], "@stats_clear"))
1177                 r = message_stats_clear(md, argc, argv);
1178         else if (!strcasecmp(argv[0], "@stats_list"))
1179                 r = message_stats_list(md, argc, argv, result, maxlen);
1180         else if (!strcasecmp(argv[0], "@stats_print"))
1181                 r = message_stats_print(md, argc, argv, false, result, maxlen);
1182         else if (!strcasecmp(argv[0], "@stats_print_clear"))
1183                 r = message_stats_print(md, argc, argv, true, result, maxlen);
1184         else if (!strcasecmp(argv[0], "@stats_set_aux"))
1185                 r = message_stats_set_aux(md, argc, argv);
1186         else
1187                 return 2; /* this wasn't a stats message */
1188
1189         if (r == -EINVAL)
1190                 DMWARN("Invalid parameters for message %s", argv[0]);
1191
1192         return r;
1193 }
1194
1195 int __init dm_statistics_init(void)
1196 {
1197         shared_memory_amount = 0;
1198         dm_stat_need_rcu_barrier = 0;
1199         return 0;
1200 }
1201
1202 void dm_statistics_exit(void)
1203 {
1204         if (dm_stat_need_rcu_barrier)
1205                 rcu_barrier();
1206         if (WARN_ON(shared_memory_amount))
1207                 DMCRIT("shared_memory_amount leaked: %lu", shared_memory_amount);
1208 }
1209
1210 module_param_named(stats_current_allocated_bytes, shared_memory_amount, ulong, S_IRUGO);
1211 MODULE_PARM_DESC(stats_current_allocated_bytes, "Memory currently used by statistics");