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-core.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                         cond_resched();
 232                 }
 233                 dm_stat_free(&s->rcu_head);
 234         }
 235         free_percpu(stats->last);
 236 }
 237
 238 static int dm_stats_create(struct dm_stats *stats, sector_t start, sector_t end,
 239                            sector_t step, unsigned stat_flags,
 240                            unsigned n_histogram_entries,
 241                            unsigned long long *histogram_boundaries,
 242                            const char *program_id, const char *aux_data,
 243                            void (*suspend_callback)(struct mapped_device *),
 244                            void (*resume_callback)(struct mapped_device *),
 245                            struct mapped_device *md)
 246 {
 247         struct list_head *l;
 248         struct dm_stat *s, *tmp_s;
 249         sector_t n_entries;
 250         size_t ni;
 251         size_t shared_alloc_size;
 252         size_t percpu_alloc_size;
 253         size_t histogram_alloc_size;
 254         struct dm_stat_percpu *p;
 255         int cpu;
 256         int ret_id;
 257         int r;
 258
 259         if (end < start || !step)
 260                 return -EINVAL;
 261
 262         n_entries = end - start;
 263         if (dm_sector_div64(n_entries, step))
 264                 n_entries++;
 265
 266         if (n_entries != (size_t)n_entries || !(size_t)(n_entries + 1))
 267                 return -EOVERFLOW;
 268
 269         shared_alloc_size = sizeof(struct dm_stat) + (size_t)n_entries * sizeof(struct dm_stat_shared);
 270         if ((shared_alloc_size - sizeof(struct dm_stat)) / sizeof(struct dm_stat_shared) != n_entries)
 271                 return -EOVERFLOW;
 272
 273         percpu_alloc_size = (size_t)n_entries * sizeof(struct dm_stat_percpu);
 274         if (percpu_alloc_size / sizeof(struct dm_stat_percpu) != n_entries)
 275                 return -EOVERFLOW;
 276
 277         histogram_alloc_size = (n_histogram_entries + 1) * (size_t)n_entries * sizeof(unsigned long long);
 278         if (histogram_alloc_size / (n_histogram_entries + 1) != (size_t)n_entries * sizeof(unsigned long long))
 279                 return -EOVERFLOW;
 280
 281         if (!check_shared_memory(shared_alloc_size + histogram_alloc_size +
 282                                  num_possible_cpus() * (percpu_alloc_size + histogram_alloc_size)))
 283                 return -ENOMEM;
 284
 285         s = dm_kvzalloc(shared_alloc_size, NUMA_NO_NODE);
 286         if (!s)
 287                 return -ENOMEM;
 288
 289         s->stat_flags = stat_flags;
 290         s->n_entries = n_entries;
 291         s->start = start;
 292         s->end = end;
 293         s->step = step;
 294         s->shared_alloc_size = shared_alloc_size;
 295         s->percpu_alloc_size = percpu_alloc_size;
 296         s->histogram_alloc_size = histogram_alloc_size;
 297
 298         s->n_histogram_entries = n_histogram_entries;
 299         s->histogram_boundaries = kmemdup(histogram_boundaries,
 300                                           s->n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
 301         if (!s->histogram_boundaries) {
 302                 r = -ENOMEM;
 303                 goto out;
 304         }
 305
 306         s->program_id = kstrdup(program_id, GFP_KERNEL);
 307         if (!s->program_id) {
 308                 r = -ENOMEM;
 309                 goto out;
 310         }
 311         s->aux_data = kstrdup(aux_data, GFP_KERNEL);
 312         if (!s->aux_data) {
 313                 r = -ENOMEM;
 314                 goto out;
 315         }
 316
 317         for (ni = 0; ni < n_entries; ni++) {
 318                 atomic_set(&s->stat_shared[ni].in_flight[READ], 0);
 319                 atomic_set(&s->stat_shared[ni].in_flight[WRITE], 0);
 320                 cond_resched();
 321         }
 322
 323         if (s->n_histogram_entries) {
 324                 unsigned long long *hi;
 325                 hi = dm_kvzalloc(s->histogram_alloc_size, NUMA_NO_NODE);
 326                 if (!hi) {
 327                         r = -ENOMEM;
 328                         goto out;
 329                 }
 330                 for (ni = 0; ni < n_entries; ni++) {
 331                         s->stat_shared[ni].tmp.histogram = hi;
 332                         hi += s->n_histogram_entries + 1;
 333                         cond_resched();
 334                 }
 335         }
 336
 337         for_each_possible_cpu(cpu) {
 338                 p = dm_kvzalloc(percpu_alloc_size, cpu_to_node(cpu));
 339                 if (!p) {
 340                         r = -ENOMEM;
 341                         goto out;
 342                 }
 343                 s->stat_percpu[cpu] = p;
 344                 if (s->n_histogram_entries) {
 345                         unsigned long long *hi;
 346                         hi = dm_kvzalloc(s->histogram_alloc_size, cpu_to_node(cpu));
 347                         if (!hi) {
 348                                 r = -ENOMEM;
 349                                 goto out;
 350                         }
 351                         for (ni = 0; ni < n_entries; ni++) {
 352                                 p[ni].histogram = hi;
 353                                 hi += s->n_histogram_entries + 1;
 354                                 cond_resched();
 355                         }
 356                 }
 357         }
 358
 359         /*
 360          * Suspend/resume to make sure there is no i/o in flight,
 361          * so that newly created statistics will be exact.
 362          *
 363          * (note: we couldn't suspend earlier because we must not
 364          * allocate memory while suspended)
 365          */
 366         suspend_callback(md);
 367
 368         mutex_lock(&stats->mutex);
 369         s->id = 0;
 370         list_for_each(l, &stats->list) {
 371                 tmp_s = container_of(l, struct dm_stat, list_entry);
 372                 if (WARN_ON(tmp_s->id < s->id)) {
 373                         r = -EINVAL;
 374                         goto out_unlock_resume;
 375                 }
 376                 if (tmp_s->id > s->id)
 377                         break;
 378                 if (unlikely(s->id == INT_MAX)) {
 379                         r = -ENFILE;
 380                         goto out_unlock_resume;
 381                 }
 382                 s->id++;
 383         }
 384         ret_id = s->id;
 385         list_add_tail_rcu(&s->list_entry, l);
 386         mutex_unlock(&stats->mutex);
 387
 388         resume_callback(md);
 389
 390         return ret_id;
 391
 392 out_unlock_resume:
 393         mutex_unlock(&stats->mutex);
 394         resume_callback(md);
 395 out:
 396         dm_stat_free(&s->rcu_head);
 397         return r;
 398 }
 399
 400 static struct dm_stat *__dm_stats_find(struct dm_stats *stats, int id)
 401 {
 402         struct dm_stat *s;
 403
 404         list_for_each_entry(s, &stats->list, list_entry) {
 405                 if (s->id > id)
 406                         break;
 407                 if (s->id == id)
 408                         return s;
 409         }
 410
 411         return NULL;
 412 }
 413
 414 static int dm_stats_delete(struct dm_stats *stats, int id)
 415 {
 416         struct dm_stat *s;
 417         int cpu;
 418
 419         mutex_lock(&stats->mutex);
 420
 421         s = __dm_stats_find(stats, id);
 422         if (!s) {
 423                 mutex_unlock(&stats->mutex);
 424                 return -ENOENT;
 425         }
 426
 427         list_del_rcu(&s->list_entry);
 428         mutex_unlock(&stats->mutex);
 429
 430         /*
 431          * vfree can't be called from RCU callback
 432          */
 433         for_each_possible_cpu(cpu)
 434                 if (is_vmalloc_addr(s->stat_percpu) ||
 435                     is_vmalloc_addr(s->stat_percpu[cpu][0].histogram))
 436                         goto do_sync_free;
 437         if (is_vmalloc_addr(s) ||
 438             is_vmalloc_addr(s->stat_shared[0].tmp.histogram)) {
 439 do_sync_free:
 440                 synchronize_rcu_expedited();
 441                 dm_stat_free(&s->rcu_head);
 442         } else {
 443                 ACCESS_ONCE(dm_stat_need_rcu_barrier) = 1;
 444                 call_rcu(&s->rcu_head, dm_stat_free);
 445         }
 446         return 0;
 447 }
 448
 449 static int dm_stats_list(struct dm_stats *stats, const char *program,
 450                          char *result, unsigned maxlen)
 451 {
 452         struct dm_stat *s;
 453         sector_t len;
 454         unsigned sz = 0;
 455
 456         /*
 457          * Output format:
 458          *   <region_id>: <start_sector>+<length> <step> <program_id> <aux_data>
 459          */
 460
 461         mutex_lock(&stats->mutex);
 462         list_for_each_entry(s, &stats->list, list_entry) {
 463                 if (!program || !strcmp(program, s->program_id)) {
 464                         len = s->end - s->start;
 465                         DMEMIT("%d: %llu+%llu %llu %s %s", s->id,
 466                                 (unsigned long long)s->start,
 467                                 (unsigned long long)len,
 468                                 (unsigned long long)s->step,
 469                                 s->program_id,
 470                                 s->aux_data);
 471                         if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
 472                                 DMEMIT(" precise_timestamps");
 473                         if (s->n_histogram_entries) {
 474                                 unsigned i;
 475                                 DMEMIT(" histogram:");
 476                                 for (i = 0; i < s->n_histogram_entries; i++) {
 477                                         if (i)
 478                                                 DMEMIT(",");
 479                                         DMEMIT("%llu", s->histogram_boundaries[i]);
 480                                 }
 481                         }
 482                         DMEMIT("\n");
 483                 }
 484                 cond_resched();
 485         }
 486         mutex_unlock(&stats->mutex);
 487
 488         return 1;
 489 }
 490
 491 static void dm_stat_round(struct dm_stat *s, struct dm_stat_shared *shared,
 492                           struct dm_stat_percpu *p)
 493 {
 494         /*
 495          * This is racy, but so is part_round_stats_single.
 496          */
 497         unsigned long long now, difference;
 498         unsigned in_flight_read, in_flight_write;
 499
 500         if (likely(!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)))
 501                 now = jiffies;
 502         else
 503                 now = ktime_to_ns(ktime_get());
 504
 505         difference = now - shared->stamp;
 506         if (!difference)
 507                 return;
 508
 509         in_flight_read = (unsigned)atomic_read(&shared->in_flight[READ]);
 510         in_flight_write = (unsigned)atomic_read(&shared->in_flight[WRITE]);
 511         if (in_flight_read)
 512                 p->io_ticks[READ] += difference;
 513         if (in_flight_write)
 514                 p->io_ticks[WRITE] += difference;
 515         if (in_flight_read + in_flight_write) {
 516                 p->io_ticks_total += difference;
 517                 p->time_in_queue += (in_flight_read + in_flight_write) * difference;
 518         }
 519         shared->stamp = now;
 520 }
 521
 522 static void dm_stat_for_entry(struct dm_stat *s, size_t entry,
 523                               int idx, sector_t len,
 524                               struct dm_stats_aux *stats_aux, bool end,
 525                               unsigned long duration_jiffies)
 526 {
 527         struct dm_stat_shared *shared = &s->stat_shared[entry];
 528         struct dm_stat_percpu *p;
 529
 530         /*
 531          * For strict correctness we should use local_irq_save/restore
 532          * instead of preempt_disable/enable.
 533          *
 534          * preempt_disable/enable is racy if the driver finishes bios
 535          * from non-interrupt context as well as from interrupt context
 536          * or from more different interrupts.
 537          *
 538          * On 64-bit architectures the race only results in not counting some
 539          * events, so it is acceptable.  On 32-bit architectures the race could
 540          * cause the counter going off by 2^32, so we need to do proper locking
 541          * there.
 542          *
 543          * part_stat_lock()/part_stat_unlock() have this race too.
 544          */
 545 #if BITS_PER_LONG == 32
 546         unsigned long flags;
 547         local_irq_save(flags);
 548 #else
 549         preempt_disable();
 550 #endif
 551         p = &s->stat_percpu[smp_processor_id()][entry];
 552
 553         if (!end) {
 554                 dm_stat_round(s, shared, p);
 555                 atomic_inc(&shared->in_flight[idx]);
 556         } else {
 557                 unsigned long long duration;
 558                 dm_stat_round(s, shared, p);
 559                 atomic_dec(&shared->in_flight[idx]);
 560                 p->sectors[idx] += len;
 561                 p->ios[idx] += 1;
 562                 p->merges[idx] += stats_aux->merged;
 563                 if (!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)) {
 564                         p->ticks[idx] += duration_jiffies;
 565                         duration = jiffies_to_msecs(duration_jiffies);
 566                 } else {
 567                         p->ticks[idx] += stats_aux->duration_ns;
 568                         duration = stats_aux->duration_ns;
 569                 }
 570                 if (s->n_histogram_entries) {
 571                         unsigned lo = 0, hi = s->n_histogram_entries + 1;
 572                         while (lo + 1 < hi) {
 573                                 unsigned mid = (lo + hi) / 2;
 574                                 if (s->histogram_boundaries[mid - 1] > duration) {
 575                                         hi = mid;
 576                                 } else {
 577                                         lo = mid;
 578                                 }
 579
 580                         }
 581                         p->histogram[lo]++;
 582                 }
 583         }
 584
 585 #if BITS_PER_LONG == 32
 586         local_irq_restore(flags);
 587 #else
 588         preempt_enable();
 589 #endif
 590 }
 591
 592 static void __dm_stat_bio(struct dm_stat *s, int bi_rw,
 593                           sector_t bi_sector, sector_t end_sector,
 594                           bool end, unsigned long duration_jiffies,
 595                           struct dm_stats_aux *stats_aux)
 596 {
 597         sector_t rel_sector, offset, todo, fragment_len;
 598         size_t entry;
 599
 600         if (end_sector <= s->start || bi_sector >= s->end)
 601                 return;
 602         if (unlikely(bi_sector < s->start)) {
 603                 rel_sector = 0;
 604                 todo = end_sector - s->start;
 605         } else {
 606                 rel_sector = bi_sector - s->start;
 607                 todo = end_sector - bi_sector;
 608         }
 609         if (unlikely(end_sector > s->end))
 610                 todo -= (end_sector - s->end);
 611
 612         offset = dm_sector_div64(rel_sector, s->step);
 613         entry = rel_sector;
 614         do {
 615                 if (WARN_ON_ONCE(entry >= s->n_entries)) {
 616                         DMCRIT("Invalid area access in region id %d", s->id);
 617                         return;
 618                 }
 619                 fragment_len = todo;
 620                 if (fragment_len > s->step - offset)
 621                         fragment_len = s->step - offset;
 622                 dm_stat_for_entry(s, entry, bi_rw, fragment_len,
 623                                   stats_aux, end, duration_jiffies);
 624                 todo -= fragment_len;
 625                 entry++;
 626                 offset = 0;
 627         } while (unlikely(todo != 0));
 628 }
 629
 630 void dm_stats_account_io(struct dm_stats *stats, unsigned long bi_rw,
 631                          sector_t bi_sector, unsigned bi_sectors, bool end,
 632                          unsigned long duration_jiffies,
 633                          struct dm_stats_aux *stats_aux)
 634 {
 635         struct dm_stat *s;
 636         sector_t end_sector;
 637         struct dm_stats_last_position *last;
 638         bool got_precise_time;
 639
 640         if (unlikely(!bi_sectors))
 641                 return;
 642
 643         end_sector = bi_sector + bi_sectors;
 644
 645         if (!end) {
 646                 /*
 647                  * A race condition can at worst result in the merged flag being
 648                  * misrepresented, so we don't have to disable preemption here.
 649                  */
 650                 last = raw_cpu_ptr(stats->last);
 651                 stats_aux->merged =
 652                         (bi_sector == (ACCESS_ONCE(last->last_sector) &&
 653                                        ((bi_rw == WRITE) ==
 654                                         (ACCESS_ONCE(last->last_rw) == WRITE))
 655                                        ));
 656                 ACCESS_ONCE(last->last_sector) = end_sector;
 657                 ACCESS_ONCE(last->last_rw) = bi_rw;
 658         }
 659
 660         rcu_read_lock();
 661
 662         got_precise_time = false;
 663         list_for_each_entry_rcu(s, &stats->list, list_entry) {
 664                 if (s->stat_flags & STAT_PRECISE_TIMESTAMPS && !got_precise_time) {
 665                         if (!end)
 666                                 stats_aux->duration_ns = ktime_to_ns(ktime_get());
 667                         else
 668                                 stats_aux->duration_ns = ktime_to_ns(ktime_get()) - stats_aux->duration_ns;
 669                         got_precise_time = true;
 670                 }
 671                 __dm_stat_bio(s, bi_rw, bi_sector, end_sector, end, duration_jiffies, stats_aux);
 672         }
 673
 674         rcu_read_unlock();
 675 }
 676
 677 static void __dm_stat_init_temporary_percpu_totals(struct dm_stat_shared *shared,
 678                                                    struct dm_stat *s, size_t x)
 679 {
 680         int cpu;
 681         struct dm_stat_percpu *p;
 682
 683         local_irq_disable();
 684         p = &s->stat_percpu[smp_processor_id()][x];
 685         dm_stat_round(s, shared, p);
 686         local_irq_enable();
 687
 688         shared->tmp.sectors[READ] = 0;
 689         shared->tmp.sectors[WRITE] = 0;
 690         shared->tmp.ios[READ] = 0;
 691         shared->tmp.ios[WRITE] = 0;
 692         shared->tmp.merges[READ] = 0;
 693         shared->tmp.merges[WRITE] = 0;
 694         shared->tmp.ticks[READ] = 0;
 695         shared->tmp.ticks[WRITE] = 0;
 696         shared->tmp.io_ticks[READ] = 0;
 697         shared->tmp.io_ticks[WRITE] = 0;
 698         shared->tmp.io_ticks_total = 0;
 699         shared->tmp.time_in_queue = 0;
 700
 701         if (s->n_histogram_entries)
 702                 memset(shared->tmp.histogram, 0, (s->n_histogram_entries + 1) * sizeof(unsigned long long));
 703
 704         for_each_possible_cpu(cpu) {
 705                 p = &s->stat_percpu[cpu][x];
 706                 shared->tmp.sectors[READ] += ACCESS_ONCE(p->sectors[READ]);
 707                 shared->tmp.sectors[WRITE] += ACCESS_ONCE(p->sectors[WRITE]);
 708                 shared->tmp.ios[READ] += ACCESS_ONCE(p->ios[READ]);
 709                 shared->tmp.ios[WRITE] += ACCESS_ONCE(p->ios[WRITE]);
 710                 shared->tmp.merges[READ] += ACCESS_ONCE(p->merges[READ]);
 711                 shared->tmp.merges[WRITE] += ACCESS_ONCE(p->merges[WRITE]);
 712                 shared->tmp.ticks[READ] += ACCESS_ONCE(p->ticks[READ]);
 713                 shared->tmp.ticks[WRITE] += ACCESS_ONCE(p->ticks[WRITE]);
 714                 shared->tmp.io_ticks[READ] += ACCESS_ONCE(p->io_ticks[READ]);
 715                 shared->tmp.io_ticks[WRITE] += ACCESS_ONCE(p->io_ticks[WRITE]);
 716                 shared->tmp.io_ticks_total += ACCESS_ONCE(p->io_ticks_total);
 717                 shared->tmp.time_in_queue += ACCESS_ONCE(p->time_in_queue);
 718                 if (s->n_histogram_entries) {
 719                         unsigned i;
 720                         for (i = 0; i < s->n_histogram_entries + 1; i++)
 721                                 shared->tmp.histogram[i] += ACCESS_ONCE(p->histogram[i]);
 722                 }
 723         }
 724 }
 725
 726 static void __dm_stat_clear(struct dm_stat *s, size_t idx_start, size_t idx_end,
 727                             bool init_tmp_percpu_totals)
 728 {
 729         size_t x;
 730         struct dm_stat_shared *shared;
 731         struct dm_stat_percpu *p;
 732
 733         for (x = idx_start; x < idx_end; x++) {
 734                 shared = &s->stat_shared[x];
 735                 if (init_tmp_percpu_totals)
 736                         __dm_stat_init_temporary_percpu_totals(shared, s, x);
 737                 local_irq_disable();
 738                 p = &s->stat_percpu[smp_processor_id()][x];
 739                 p->sectors[READ] -= shared->tmp.sectors[READ];
 740                 p->sectors[WRITE] -= shared->tmp.sectors[WRITE];
 741                 p->ios[READ] -= shared->tmp.ios[READ];
 742                 p->ios[WRITE] -= shared->tmp.ios[WRITE];
 743                 p->merges[READ] -= shared->tmp.merges[READ];
 744                 p->merges[WRITE] -= shared->tmp.merges[WRITE];
 745                 p->ticks[READ] -= shared->tmp.ticks[READ];
 746                 p->ticks[WRITE] -= shared->tmp.ticks[WRITE];
 747                 p->io_ticks[READ] -= shared->tmp.io_ticks[READ];
 748                 p->io_ticks[WRITE] -= shared->tmp.io_ticks[WRITE];
 749                 p->io_ticks_total -= shared->tmp.io_ticks_total;
 750                 p->time_in_queue -= shared->tmp.time_in_queue;
 751                 local_irq_enable();
 752                 if (s->n_histogram_entries) {
 753                         unsigned i;
 754                         for (i = 0; i < s->n_histogram_entries + 1; i++) {
 755                                 local_irq_disable();
 756                                 p = &s->stat_percpu[smp_processor_id()][x];
 757                                 p->histogram[i] -= shared->tmp.histogram[i];
 758                                 local_irq_enable();
 759                         }
 760                 }
 761                 cond_resched();
 762         }
 763 }
 764
 765 static int dm_stats_clear(struct dm_stats *stats, int id)
 766 {
 767         struct dm_stat *s;
 768
 769         mutex_lock(&stats->mutex);
 770
 771         s = __dm_stats_find(stats, id);
 772         if (!s) {
 773                 mutex_unlock(&stats->mutex);
 774                 return -ENOENT;
 775         }
 776
 777         __dm_stat_clear(s, 0, s->n_entries, true);
 778
 779         mutex_unlock(&stats->mutex);
 780
 781         return 1;
 782 }
 783
 784 /*
 785  * This is like jiffies_to_msec, but works for 64-bit values.
 786  */
 787 static unsigned long long dm_jiffies_to_msec64(struct dm_stat *s, unsigned long long j)
 788 {
 789         unsigned long long result;
 790         unsigned mult;
 791
 792         if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
 793                 return j;
 794
 795         result = 0;
 796         if (j)
 797                 result = jiffies_to_msecs(j & 0x3fffff);
 798         if (j >= 1 << 22) {
 799                 mult = jiffies_to_msecs(1 << 22);
 800                 result += (unsigned long long)mult * (unsigned long long)jiffies_to_msecs((j >> 22) & 0x3fffff);
 801         }
 802         if (j >= 1ULL << 44)
 803                 result += (unsigned long long)mult * (unsigned long long)mult * (unsigned long long)jiffies_to_msecs(j >> 44);
 804
 805         return result;
 806 }
 807
 808 static int dm_stats_print(struct dm_stats *stats, int id,
 809                           size_t idx_start, size_t idx_len,
 810                           bool clear, char *result, unsigned maxlen)
 811 {
 812         unsigned sz = 0;
 813         struct dm_stat *s;
 814         size_t x;
 815         sector_t start, end, step;
 816         size_t idx_end;
 817         struct dm_stat_shared *shared;
 818
 819         /*
 820          * Output format:
 821          *   <start_sector>+<length> counters
 822          */
 823
 824         mutex_lock(&stats->mutex);
 825
 826         s = __dm_stats_find(stats, id);
 827         if (!s) {
 828                 mutex_unlock(&stats->mutex);
 829                 return -ENOENT;
 830         }
 831
 832         idx_end = idx_start + idx_len;
 833         if (idx_end < idx_start ||
 834             idx_end > s->n_entries)
 835                 idx_end = s->n_entries;
 836
 837         if (idx_start > idx_end)
 838                 idx_start = idx_end;
 839
 840         step = s->step;
 841         start = s->start + (step * idx_start);
 842
 843         for (x = idx_start; x < idx_end; x++, start = end) {
 844                 shared = &s->stat_shared[x];
 845                 end = start + step;
 846                 if (unlikely(end > s->end))
 847                         end = s->end;
 848
 849                 __dm_stat_init_temporary_percpu_totals(shared, s, x);
 850
 851                 DMEMIT("%llu+%llu %llu %llu %llu %llu %llu %llu %llu %llu %d %llu %llu %llu %llu",
 852                        (unsigned long long)start,
 853                        (unsigned long long)step,
 854                        shared->tmp.ios[READ],
 855                        shared->tmp.merges[READ],
 856                        shared->tmp.sectors[READ],
 857                        dm_jiffies_to_msec64(s, shared->tmp.ticks[READ]),
 858                        shared->tmp.ios[WRITE],
 859                        shared->tmp.merges[WRITE],
 860                        shared->tmp.sectors[WRITE],
 861                        dm_jiffies_to_msec64(s, shared->tmp.ticks[WRITE]),
 862                        dm_stat_in_flight(shared),
 863                        dm_jiffies_to_msec64(s, shared->tmp.io_ticks_total),
 864                        dm_jiffies_to_msec64(s, shared->tmp.time_in_queue),
 865                        dm_jiffies_to_msec64(s, shared->tmp.io_ticks[READ]),
 866                        dm_jiffies_to_msec64(s, shared->tmp.io_ticks[WRITE]));
 867                 if (s->n_histogram_entries) {
 868                         unsigned i;
 869                         for (i = 0; i < s->n_histogram_entries + 1; i++) {
 870                                 DMEMIT("%s%llu", !i ? " " : ":", shared->tmp.histogram[i]);
 871                         }
 872                 }
 873                 DMEMIT("\n");
 874
 875                 if (unlikely(sz + 1 >= maxlen))
 876                         goto buffer_overflow;
 877
 878                 cond_resched();
 879         }
 880
 881         if (clear)
 882                 __dm_stat_clear(s, idx_start, idx_end, false);
 883
 884 buffer_overflow:
 885         mutex_unlock(&stats->mutex);
 886
 887         return 1;
 888 }
 889
 890 static int dm_stats_set_aux(struct dm_stats *stats, int id, const char *aux_data)
 891 {
 892         struct dm_stat *s;
 893         const char *new_aux_data;
 894
 895         mutex_lock(&stats->mutex);
 896
 897         s = __dm_stats_find(stats, id);
 898         if (!s) {
 899                 mutex_unlock(&stats->mutex);
 900                 return -ENOENT;
 901         }
 902
 903         new_aux_data = kstrdup(aux_data, GFP_KERNEL);
 904         if (!new_aux_data) {
 905                 mutex_unlock(&stats->mutex);
 906                 return -ENOMEM;
 907         }
 908
 909         kfree(s->aux_data);
 910         s->aux_data = new_aux_data;
 911
 912         mutex_unlock(&stats->mutex);
 913
 914         return 0;
 915 }
 916
 917 static int parse_histogram(const char *h, unsigned *n_histogram_entries,
 918                            unsigned long long **histogram_boundaries)
 919 {
 920         const char *q;
 921         unsigned n;
 922         unsigned long long last;
 923
 924         *n_histogram_entries = 1;
 925         for (q = h; *q; q++)
 926                 if (*q == ',')
 927                         (*n_histogram_entries)++;
 928
 929         *histogram_boundaries = kmalloc(*n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
 930         if (!*histogram_boundaries)
 931                 return -ENOMEM;
 932
 933         n = 0;
 934         last = 0;
 935         while (1) {
 936                 unsigned long long hi;
 937                 int s;
 938                 char ch;
 939                 s = sscanf(h, "%llu%c", &hi, &ch);
 940                 if (!s || (s == 2 && ch != ','))
 941                         return -EINVAL;
 942                 if (hi <= last)
 943                         return -EINVAL;
 944                 last = hi;
 945                 (*histogram_boundaries)[n] = hi;
 946                 if (s == 1)
 947                         return 0;
 948                 h = strchr(h, ',') + 1;
 949                 n++;
 950         }
 951 }
 952
 953 static int message_stats_create(struct mapped_device *md,
 954                                 unsigned argc, char **argv,
 955                                 char *result, unsigned maxlen)
 956 {
 957         int r;
 958         int id;
 959         char dummy;
 960         unsigned long long start, end, len, step;
 961         unsigned divisor;
 962         const char *program_id, *aux_data;
 963         unsigned stat_flags = 0;
 964
 965         unsigned n_histogram_entries = 0;
 966         unsigned long long *histogram_boundaries = NULL;
 967
 968         struct dm_arg_set as, as_backup;
 969         const char *a;
 970         unsigned feature_args;
 971
 972         /*
 973          * Input format:
 974          *   <range> <step> [<extra_parameters> <parameters>] [<program_id> [<aux_data>]]
 975          */
 976
 977         if (argc < 3)
 978                 goto ret_einval;
 979
 980         as.argc = argc;
 981         as.argv = argv;
 982         dm_consume_args(&as, 1);
 983
 984         a = dm_shift_arg(&as);
 985         if (!strcmp(a, "-")) {
 986                 start = 0;
 987                 len = dm_get_size(md);
 988                 if (!len)
 989                         len = 1;
 990         } else if (sscanf(a, "%llu+%llu%c", &start, &len, &dummy) != 2 ||
 991                    start != (sector_t)start || len != (sector_t)len)
 992                 goto ret_einval;
 993
 994         end = start + len;
 995         if (start >= end)
 996                 goto ret_einval;
 997
 998         a = dm_shift_arg(&as);
 999         if (sscanf(a, "/%u%c", &divisor, &dummy) == 1) {
1000                 if (!divisor)
1001                         return -EINVAL;
1002                 step = end - start;
1003                 if (do_div(step, divisor))
1004                         step++;
1005                 if (!step)
1006                         step = 1;
1007         } else if (sscanf(a, "%llu%c", &step, &dummy) != 1 ||
1008                    step != (sector_t)step || !step)
1009                 goto ret_einval;
1010
1011         as_backup = as;
1012         a = dm_shift_arg(&as);
1013         if (a && sscanf(a, "%u%c", &feature_args, &dummy) == 1) {
1014                 while (feature_args--) {
1015                         a = dm_shift_arg(&as);
1016                         if (!a)
1017                                 goto ret_einval;
1018                         if (!strcasecmp(a, "precise_timestamps"))
1019                                 stat_flags |= STAT_PRECISE_TIMESTAMPS;
1020                         else if (!strncasecmp(a, "histogram:", 10)) {
1021                                 if (n_histogram_entries)
1022                                         goto ret_einval;
1023                                 if ((r = parse_histogram(a + 10, &n_histogram_entries, &histogram_boundaries)))
1024                                         goto ret;
1025                         } else
1026                                 goto ret_einval;
1027                 }
1028         } else {
1029                 as = as_backup;
1030         }
1031
1032         program_id = "-";
1033         aux_data = "-";
1034
1035         a = dm_shift_arg(&as);
1036         if (a)
1037                 program_id = a;
1038
1039         a = dm_shift_arg(&as);
1040         if (a)
1041                 aux_data = a;
1042
1043         if (as.argc)
1044                 goto ret_einval;
1045
1046         /*
1047          * If a buffer overflow happens after we created the region,
1048          * it's too late (the userspace would retry with a larger
1049          * buffer, but the region id that caused the overflow is already
1050          * leaked).  So we must detect buffer overflow in advance.
1051          */
1052         snprintf(result, maxlen, "%d", INT_MAX);
1053         if (dm_message_test_buffer_overflow(result, maxlen)) {
1054                 r = 1;
1055                 goto ret;
1056         }
1057
1058         id = dm_stats_create(dm_get_stats(md), start, end, step, stat_flags,
1059                              n_histogram_entries, histogram_boundaries, program_id, aux_data,
1060                              dm_internal_suspend_fast, dm_internal_resume_fast, md);
1061         if (id < 0) {
1062                 r = id;
1063                 goto ret;
1064         }
1065
1066         snprintf(result, maxlen, "%d", id);
1067
1068         r = 1;
1069         goto ret;
1070
1071 ret_einval:
1072         r = -EINVAL;
1073 ret:
1074         kfree(histogram_boundaries);
1075         return r;
1076 }
1077
1078 static int message_stats_delete(struct mapped_device *md,
1079                                 unsigned argc, char **argv)
1080 {
1081         int id;
1082         char dummy;
1083
1084         if (argc != 2)
1085                 return -EINVAL;
1086
1087         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1088                 return -EINVAL;
1089
1090         return dm_stats_delete(dm_get_stats(md), id);
1091 }
1092
1093 static int message_stats_clear(struct mapped_device *md,
1094                                unsigned argc, char **argv)
1095 {
1096         int id;
1097         char dummy;
1098
1099         if (argc != 2)
1100                 return -EINVAL;
1101
1102         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1103                 return -EINVAL;
1104
1105         return dm_stats_clear(dm_get_stats(md), id);
1106 }
1107
1108 static int message_stats_list(struct mapped_device *md,
1109                               unsigned argc, char **argv,
1110                               char *result, unsigned maxlen)
1111 {
1112         int r;
1113         const char *program = NULL;
1114
1115         if (argc < 1 || argc > 2)
1116                 return -EINVAL;
1117
1118         if (argc > 1) {
1119                 program = kstrdup(argv[1], GFP_KERNEL);
1120                 if (!program)
1121                         return -ENOMEM;
1122         }
1123
1124         r = dm_stats_list(dm_get_stats(md), program, result, maxlen);
1125
1126         kfree(program);
1127
1128         return r;
1129 }
1130
1131 static int message_stats_print(struct mapped_device *md,
1132                                unsigned argc, char **argv, bool clear,
1133                                char *result, unsigned maxlen)
1134 {
1135         int id;
1136         char dummy;
1137         unsigned long idx_start = 0, idx_len = ULONG_MAX;
1138
1139         if (argc != 2 && argc != 4)
1140                 return -EINVAL;
1141
1142         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1143                 return -EINVAL;
1144
1145         if (argc > 3) {
1146                 if (strcmp(argv[2], "-") &&
1147                     sscanf(argv[2], "%lu%c", &idx_start, &dummy) != 1)
1148                         return -EINVAL;
1149                 if (strcmp(argv[3], "-") &&
1150                     sscanf(argv[3], "%lu%c", &idx_len, &dummy) != 1)
1151                         return -EINVAL;
1152         }
1153
1154         return dm_stats_print(dm_get_stats(md), id, idx_start, idx_len, clear,
1155                               result, maxlen);
1156 }
1157
1158 static int message_stats_set_aux(struct mapped_device *md,
1159                                  unsigned argc, char **argv)
1160 {
1161         int id;
1162         char dummy;
1163
1164         if (argc != 3)
1165                 return -EINVAL;
1166
1167         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1168                 return -EINVAL;
1169
1170         return dm_stats_set_aux(dm_get_stats(md), id, argv[2]);
1171 }
1172
1173 int dm_stats_message(struct mapped_device *md, unsigned argc, char **argv,
1174                      char *result, unsigned maxlen)
1175 {
1176         int r;
1177
1178         /* All messages here must start with '@' */
1179         if (!strcasecmp(argv[0], "@stats_create"))
1180                 r = message_stats_create(md, argc, argv, result, maxlen);
1181         else if (!strcasecmp(argv[0], "@stats_delete"))
1182                 r = message_stats_delete(md, argc, argv);
1183         else if (!strcasecmp(argv[0], "@stats_clear"))
1184                 r = message_stats_clear(md, argc, argv);
1185         else if (!strcasecmp(argv[0], "@stats_list"))
1186                 r = message_stats_list(md, argc, argv, result, maxlen);
1187         else if (!strcasecmp(argv[0], "@stats_print"))
1188                 r = message_stats_print(md, argc, argv, false, result, maxlen);
1189         else if (!strcasecmp(argv[0], "@stats_print_clear"))
1190                 r = message_stats_print(md, argc, argv, true, result, maxlen);
1191         else if (!strcasecmp(argv[0], "@stats_set_aux"))
1192                 r = message_stats_set_aux(md, argc, argv);
1193         else
1194                 return 2; /* this wasn't a stats message */
1195
1196         if (r == -EINVAL)
1197                 DMWARN("Invalid parameters for message %s", argv[0]);
1198
1199         return r;
1200 }
1201
1202 int __init dm_statistics_init(void)
1203 {
1204         shared_memory_amount = 0;
1205         dm_stat_need_rcu_barrier = 0;
1206         return 0;
1207 }
1208
1209 void dm_statistics_exit(void)
1210 {
1211         if (dm_stat_need_rcu_barrier)
1212                 rcu_barrier();
1213         if (WARN_ON(shared_memory_amount))
1214                 DMCRIT("shared_memory_amount leaked: %lu", shared_memory_amount);
1215 }
1216
1217 module_param_named(stats_current_allocated_bytes, shared_memory_amount, ulong, S_IRUGO);
1218 MODULE_PARM_DESC(stats_current_allocated_bytes, "Memory currently used by statistics");