drivers/md/dm-stats.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/errno.h>
   3 #include <linux/numa.h>
   4 #include <linux/slab.h>
   5 #include <linux/rculist.h>
   6 #include <linux/threads.h>
   7 #include <linux/preempt.h>
   8 #include <linux/irqflags.h>
   9 #include <linux/vmalloc.h>
  10 #include <linux/mm.h>
  11 #include <linux/module.h>
  12 #include <linux/device-mapper.h>
  13
  14 #include "dm-core.h"
  15 #include "dm-stats.h"
  16
  17 #define DM_MSG_PREFIX "stats"
  18
  19 static int dm_stat_need_rcu_barrier;
  20
  21 /*
  22  * Using 64-bit values to avoid overflow (which is a
  23  * problem that block/genhd.c's IO accounting has).
  24  */
  25 struct dm_stat_percpu {
  26         unsigned long long sectors[2];
  27         unsigned long long ios[2];
  28         unsigned long long merges[2];
  29         unsigned long long ticks[2];
  30         unsigned long long io_ticks[2];
  31         unsigned long long io_ticks_total;
  32         unsigned long long time_in_queue;
  33         unsigned long long *histogram;
  34 };
  35
  36 struct dm_stat_shared {
  37         atomic_t in_flight[2];
  38         unsigned long long stamp;
  39         struct dm_stat_percpu tmp;
  40 };
  41
  42 struct dm_stat {
  43         struct list_head list_entry;
  44         int id;
  45         unsigned int stat_flags;
  46         size_t n_entries;
  47         sector_t start;
  48         sector_t end;
  49         sector_t step;
  50         unsigned int n_histogram_entries;
  51         unsigned long long *histogram_boundaries;
  52         const char *program_id;
  53         const char *aux_data;
  54         struct rcu_head rcu_head;
  55         size_t shared_alloc_size;
  56         size_t percpu_alloc_size;
  57         size_t histogram_alloc_size;
  58         struct dm_stat_percpu *stat_percpu[NR_CPUS];
  59         struct dm_stat_shared stat_shared[];
  60 };
  61
  62 #define STAT_PRECISE_TIMESTAMPS         1
  63
  64 struct dm_stats_last_position {
  65         sector_t last_sector;
  66         unsigned int last_rw;
  67 };
  68
  69 /*
  70  * A typo on the command line could possibly make the kernel run out of memory
  71  * and crash. To prevent the crash we account all used memory. We fail if we
  72  * exhaust 1/4 of all memory or 1/2 of vmalloc space.
  73  */
  74 #define DM_STATS_MEMORY_FACTOR          4
  75 #define DM_STATS_VMALLOC_FACTOR         2
  76
  77 static DEFINE_SPINLOCK(shared_memory_lock);
  78
  79 static unsigned long shared_memory_amount;
  80
  81 static bool __check_shared_memory(size_t alloc_size)
  82 {
  83         size_t a;
  84
  85         a = shared_memory_amount + alloc_size;
  86         if (a < shared_memory_amount)
  87                 return false;
  88         if (a >> PAGE_SHIFT > totalram_pages() / DM_STATS_MEMORY_FACTOR)
  89                 return false;
  90 #ifdef CONFIG_MMU
  91         if (a > (VMALLOC_END - VMALLOC_START) / DM_STATS_VMALLOC_FACTOR)
  92                 return false;
  93 #endif
  94         return true;
  95 }
  96
  97 static bool check_shared_memory(size_t alloc_size)
  98 {
  99         bool ret;
 100
 101         spin_lock_irq(&shared_memory_lock);
 102
 103         ret = __check_shared_memory(alloc_size);
 104
 105         spin_unlock_irq(&shared_memory_lock);
 106
 107         return ret;
 108 }
 109
 110 static bool claim_shared_memory(size_t alloc_size)
 111 {
 112         spin_lock_irq(&shared_memory_lock);
 113
 114         if (!__check_shared_memory(alloc_size)) {
 115                 spin_unlock_irq(&shared_memory_lock);
 116                 return false;
 117         }
 118
 119         shared_memory_amount += alloc_size;
 120
 121         spin_unlock_irq(&shared_memory_lock);
 122
 123         return true;
 124 }
 125
 126 static void free_shared_memory(size_t alloc_size)
 127 {
 128         unsigned long flags;
 129
 130         spin_lock_irqsave(&shared_memory_lock, flags);
 131
 132         if (WARN_ON_ONCE(shared_memory_amount < alloc_size)) {
 133                 spin_unlock_irqrestore(&shared_memory_lock, flags);
 134                 DMCRIT("Memory usage accounting bug.");
 135                 return;
 136         }
 137
 138         shared_memory_amount -= alloc_size;
 139
 140         spin_unlock_irqrestore(&shared_memory_lock, flags);
 141 }
 142
 143 static void *dm_kvzalloc(size_t alloc_size, int node)
 144 {
 145         void *p;
 146
 147         if (!claim_shared_memory(alloc_size))
 148                 return NULL;
 149
 150         p = kvzalloc_node(alloc_size, GFP_KERNEL | __GFP_NOMEMALLOC, node);
 151         if (p)
 152                 return p;
 153
 154         free_shared_memory(alloc_size);
 155
 156         return NULL;
 157 }
 158
 159 static void dm_kvfree(void *ptr, size_t alloc_size)
 160 {
 161         if (!ptr)
 162                 return;
 163
 164         free_shared_memory(alloc_size);
 165
 166         kvfree(ptr);
 167 }
 168
 169 static void dm_stat_free(struct rcu_head *head)
 170 {
 171         int cpu;
 172         struct dm_stat *s = container_of(head, struct dm_stat, rcu_head);
 173
 174         kfree(s->histogram_boundaries);
 175         kfree(s->program_id);
 176         kfree(s->aux_data);
 177         for_each_possible_cpu(cpu) {
 178                 dm_kvfree(s->stat_percpu[cpu][0].histogram, s->histogram_alloc_size);
 179                 dm_kvfree(s->stat_percpu[cpu], s->percpu_alloc_size);
 180         }
 181         dm_kvfree(s->stat_shared[0].tmp.histogram, s->histogram_alloc_size);
 182         dm_kvfree(s, s->shared_alloc_size);
 183 }
 184
 185 static int dm_stat_in_flight(struct dm_stat_shared *shared)
 186 {
 187         return atomic_read(&shared->in_flight[READ]) +
 188                atomic_read(&shared->in_flight[WRITE]);
 189 }
 190
 191 int dm_stats_init(struct dm_stats *stats)
 192 {
 193         int cpu;
 194         struct dm_stats_last_position *last;
 195
 196         mutex_init(&stats->mutex);
 197         INIT_LIST_HEAD(&stats->list);
 198         stats->precise_timestamps = false;
 199         stats->last = alloc_percpu(struct dm_stats_last_position);
 200         if (!stats->last)
 201                 return -ENOMEM;
 202
 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         return 0;
 210 }
 211
 212 void dm_stats_cleanup(struct dm_stats *stats)
 213 {
 214         size_t ni;
 215         struct dm_stat *s;
 216         struct dm_stat_shared *shared;
 217
 218         while (!list_empty(&stats->list)) {
 219                 s = container_of(stats->list.next, struct dm_stat, list_entry);
 220                 list_del(&s->list_entry);
 221                 for (ni = 0; ni < s->n_entries; ni++) {
 222                         shared = &s->stat_shared[ni];
 223                         if (WARN_ON(dm_stat_in_flight(shared))) {
 224                                 DMCRIT("leaked in-flight counter at index %lu "
 225                                        "(start %llu, end %llu, step %llu): reads %d, writes %d",
 226                                        (unsigned long)ni,
 227                                        (unsigned long long)s->start,
 228                                        (unsigned long long)s->end,
 229                                        (unsigned long long)s->step,
 230                                        atomic_read(&shared->in_flight[READ]),
 231                                        atomic_read(&shared->in_flight[WRITE]));
 232                         }
 233                         cond_resched();
 234                 }
 235                 dm_stat_free(&s->rcu_head);
 236         }
 237         free_percpu(stats->last);
 238         mutex_destroy(&stats->mutex);
 239 }
 240
 241 static void dm_stats_recalc_precise_timestamps(struct dm_stats *stats)
 242 {
 243         struct list_head *l;
 244         struct dm_stat *tmp_s;
 245         bool precise_timestamps = false;
 246
 247         list_for_each(l, &stats->list) {
 248                 tmp_s = container_of(l, struct dm_stat, list_entry);
 249                 if (tmp_s->stat_flags & STAT_PRECISE_TIMESTAMPS) {
 250                         precise_timestamps = true;
 251                         break;
 252                 }
 253         }
 254         stats->precise_timestamps = precise_timestamps;
 255 }
 256
 257 static int dm_stats_create(struct dm_stats *stats, sector_t start, sector_t end,
 258                            sector_t step, unsigned int stat_flags,
 259                            unsigned int n_histogram_entries,
 260                            unsigned long long *histogram_boundaries,
 261                            const char *program_id, const char *aux_data,
 262                            void (*suspend_callback)(struct mapped_device *),
 263                            void (*resume_callback)(struct mapped_device *),
 264                            struct mapped_device *md)
 265 {
 266         struct list_head *l;
 267         struct dm_stat *s, *tmp_s;
 268         sector_t n_entries;
 269         size_t ni;
 270         size_t shared_alloc_size;
 271         size_t percpu_alloc_size;
 272         size_t histogram_alloc_size;
 273         struct dm_stat_percpu *p;
 274         int cpu;
 275         int ret_id;
 276         int r;
 277
 278         if (end < start || !step)
 279                 return -EINVAL;
 280
 281         n_entries = end - start;
 282         if (dm_sector_div64(n_entries, step))
 283                 n_entries++;
 284
 285         if (n_entries != (size_t)n_entries || !(size_t)(n_entries + 1))
 286                 return -EOVERFLOW;
 287
 288         shared_alloc_size = struct_size(s, stat_shared, n_entries);
 289         if ((shared_alloc_size - sizeof(struct dm_stat)) / sizeof(struct dm_stat_shared) != n_entries)
 290                 return -EOVERFLOW;
 291
 292         percpu_alloc_size = (size_t)n_entries * sizeof(struct dm_stat_percpu);
 293         if (percpu_alloc_size / sizeof(struct dm_stat_percpu) != n_entries)
 294                 return -EOVERFLOW;
 295
 296         histogram_alloc_size = (n_histogram_entries + 1) * (size_t)n_entries * sizeof(unsigned long long);
 297         if (histogram_alloc_size / (n_histogram_entries + 1) != (size_t)n_entries * sizeof(unsigned long long))
 298                 return -EOVERFLOW;
 299
 300         if (!check_shared_memory(shared_alloc_size + histogram_alloc_size +
 301                                  num_possible_cpus() * (percpu_alloc_size + histogram_alloc_size)))
 302                 return -ENOMEM;
 303
 304         s = dm_kvzalloc(shared_alloc_size, NUMA_NO_NODE);
 305         if (!s)
 306                 return -ENOMEM;
 307
 308         s->stat_flags = stat_flags;
 309         s->n_entries = n_entries;
 310         s->start = start;
 311         s->end = end;
 312         s->step = step;
 313         s->shared_alloc_size = shared_alloc_size;
 314         s->percpu_alloc_size = percpu_alloc_size;
 315         s->histogram_alloc_size = histogram_alloc_size;
 316
 317         s->n_histogram_entries = n_histogram_entries;
 318         s->histogram_boundaries = kmemdup(histogram_boundaries,
 319                                           s->n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
 320         if (!s->histogram_boundaries) {
 321                 r = -ENOMEM;
 322                 goto out;
 323         }
 324
 325         s->program_id = kstrdup(program_id, GFP_KERNEL);
 326         if (!s->program_id) {
 327                 r = -ENOMEM;
 328                 goto out;
 329         }
 330         s->aux_data = kstrdup(aux_data, GFP_KERNEL);
 331         if (!s->aux_data) {
 332                 r = -ENOMEM;
 333                 goto out;
 334         }
 335
 336         for (ni = 0; ni < n_entries; ni++) {
 337                 atomic_set(&s->stat_shared[ni].in_flight[READ], 0);
 338                 atomic_set(&s->stat_shared[ni].in_flight[WRITE], 0);
 339                 cond_resched();
 340         }
 341
 342         if (s->n_histogram_entries) {
 343                 unsigned long long *hi;
 344                 hi = dm_kvzalloc(s->histogram_alloc_size, NUMA_NO_NODE);
 345                 if (!hi) {
 346                         r = -ENOMEM;
 347                         goto out;
 348                 }
 349                 for (ni = 0; ni < n_entries; ni++) {
 350                         s->stat_shared[ni].tmp.histogram = hi;
 351                         hi += s->n_histogram_entries + 1;
 352                         cond_resched();
 353                 }
 354         }
 355
 356         for_each_possible_cpu(cpu) {
 357                 p = dm_kvzalloc(percpu_alloc_size, cpu_to_node(cpu));
 358                 if (!p) {
 359                         r = -ENOMEM;
 360                         goto out;
 361                 }
 362                 s->stat_percpu[cpu] = p;
 363                 if (s->n_histogram_entries) {
 364                         unsigned long long *hi;
 365                         hi = dm_kvzalloc(s->histogram_alloc_size, cpu_to_node(cpu));
 366                         if (!hi) {
 367                                 r = -ENOMEM;
 368                                 goto out;
 369                         }
 370                         for (ni = 0; ni < n_entries; ni++) {
 371                                 p[ni].histogram = hi;
 372                                 hi += s->n_histogram_entries + 1;
 373                                 cond_resched();
 374                         }
 375                 }
 376         }
 377
 378         /*
 379          * Suspend/resume to make sure there is no i/o in flight,
 380          * so that newly created statistics will be exact.
 381          *
 382          * (note: we couldn't suspend earlier because we must not
 383          * allocate memory while suspended)
 384          */
 385         suspend_callback(md);
 386
 387         mutex_lock(&stats->mutex);
 388         s->id = 0;
 389         list_for_each(l, &stats->list) {
 390                 tmp_s = container_of(l, struct dm_stat, list_entry);
 391                 if (WARN_ON(tmp_s->id < s->id)) {
 392                         r = -EINVAL;
 393                         goto out_unlock_resume;
 394                 }
 395                 if (tmp_s->id > s->id)
 396                         break;
 397                 if (unlikely(s->id == INT_MAX)) {
 398                         r = -ENFILE;
 399                         goto out_unlock_resume;
 400                 }
 401                 s->id++;
 402         }
 403         ret_id = s->id;
 404         list_add_tail_rcu(&s->list_entry, l);
 405
 406         dm_stats_recalc_precise_timestamps(stats);
 407
 408         if (!static_key_enabled(&stats_enabled.key))
 409                 static_branch_enable(&stats_enabled);
 410
 411         mutex_unlock(&stats->mutex);
 412
 413         resume_callback(md);
 414
 415         return ret_id;
 416
 417 out_unlock_resume:
 418         mutex_unlock(&stats->mutex);
 419         resume_callback(md);
 420 out:
 421         dm_stat_free(&s->rcu_head);
 422         return r;
 423 }
 424
 425 static struct dm_stat *__dm_stats_find(struct dm_stats *stats, int id)
 426 {
 427         struct dm_stat *s;
 428
 429         list_for_each_entry(s, &stats->list, list_entry) {
 430                 if (s->id > id)
 431                         break;
 432                 if (s->id == id)
 433                         return s;
 434         }
 435
 436         return NULL;
 437 }
 438
 439 static int dm_stats_delete(struct dm_stats *stats, int id)
 440 {
 441         struct dm_stat *s;
 442         int cpu;
 443
 444         mutex_lock(&stats->mutex);
 445
 446         s = __dm_stats_find(stats, id);
 447         if (!s) {
 448                 mutex_unlock(&stats->mutex);
 449                 return -ENOENT;
 450         }
 451
 452         list_del_rcu(&s->list_entry);
 453
 454         dm_stats_recalc_precise_timestamps(stats);
 455
 456         mutex_unlock(&stats->mutex);
 457
 458         /*
 459          * vfree can't be called from RCU callback
 460          */
 461         for_each_possible_cpu(cpu)
 462                 if (is_vmalloc_addr(s->stat_percpu) ||
 463                     is_vmalloc_addr(s->stat_percpu[cpu][0].histogram))
 464                         goto do_sync_free;
 465         if (is_vmalloc_addr(s) ||
 466             is_vmalloc_addr(s->stat_shared[0].tmp.histogram)) {
 467 do_sync_free:
 468                 synchronize_rcu_expedited();
 469                 dm_stat_free(&s->rcu_head);
 470         } else {
 471                 WRITE_ONCE(dm_stat_need_rcu_barrier, 1);
 472                 call_rcu(&s->rcu_head, dm_stat_free);
 473         }
 474         return 0;
 475 }
 476
 477 static int dm_stats_list(struct dm_stats *stats, const char *program,
 478                          char *result, unsigned int maxlen)
 479 {
 480         struct dm_stat *s;
 481         sector_t len;
 482         unsigned int sz = 0;
 483
 484         /*
 485          * Output format:
 486          *   <region_id>: <start_sector>+<length> <step> <program_id> <aux_data>
 487          */
 488
 489         mutex_lock(&stats->mutex);
 490         list_for_each_entry(s, &stats->list, list_entry) {
 491                 if (!program || !strcmp(program, s->program_id)) {
 492                         len = s->end - s->start;
 493                         DMEMIT("%d: %llu+%llu %llu %s %s", s->id,
 494                                 (unsigned long long)s->start,
 495                                 (unsigned long long)len,
 496                                 (unsigned long long)s->step,
 497                                 s->program_id,
 498                                 s->aux_data);
 499                         if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
 500                                 DMEMIT(" precise_timestamps");
 501                         if (s->n_histogram_entries) {
 502                                 unsigned int i;
 503                                 DMEMIT(" histogram:");
 504                                 for (i = 0; i < s->n_histogram_entries; i++) {
 505                                         if (i)
 506                                                 DMEMIT(",");
 507                                         DMEMIT("%llu", s->histogram_boundaries[i]);
 508                                 }
 509                         }
 510                         DMEMIT("\n");
 511                 }
 512                 cond_resched();
 513         }
 514         mutex_unlock(&stats->mutex);
 515
 516         return 1;
 517 }
 518
 519 static void dm_stat_round(struct dm_stat *s, struct dm_stat_shared *shared,
 520                           struct dm_stat_percpu *p)
 521 {
 522         /*
 523          * This is racy, but so is part_round_stats_single.
 524          */
 525         unsigned long long now, difference;
 526         unsigned int in_flight_read, in_flight_write;
 527
 528         if (likely(!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)))
 529                 now = jiffies;
 530         else
 531                 now = ktime_to_ns(ktime_get());
 532
 533         difference = now - shared->stamp;
 534         if (!difference)
 535                 return;
 536
 537         in_flight_read = (unsigned int)atomic_read(&shared->in_flight[READ]);
 538         in_flight_write = (unsigned int)atomic_read(&shared->in_flight[WRITE]);
 539         if (in_flight_read)
 540                 p->io_ticks[READ] += difference;
 541         if (in_flight_write)
 542                 p->io_ticks[WRITE] += difference;
 543         if (in_flight_read + in_flight_write) {
 544                 p->io_ticks_total += difference;
 545                 p->time_in_queue += (in_flight_read + in_flight_write) * difference;
 546         }
 547         shared->stamp = now;
 548 }
 549
 550 static void dm_stat_for_entry(struct dm_stat *s, size_t entry,
 551                               int idx, sector_t len,
 552                               struct dm_stats_aux *stats_aux, bool end,
 553                               unsigned long duration_jiffies)
 554 {
 555         struct dm_stat_shared *shared = &s->stat_shared[entry];
 556         struct dm_stat_percpu *p;
 557
 558         /*
 559          * For strict correctness we should use local_irq_save/restore
 560          * instead of preempt_disable/enable.
 561          *
 562          * preempt_disable/enable is racy if the driver finishes bios
 563          * from non-interrupt context as well as from interrupt context
 564          * or from more different interrupts.
 565          *
 566          * On 64-bit architectures the race only results in not counting some
 567          * events, so it is acceptable.  On 32-bit architectures the race could
 568          * cause the counter going off by 2^32, so we need to do proper locking
 569          * there.
 570          *
 571          * part_stat_lock()/part_stat_unlock() have this race too.
 572          */
 573 #if BITS_PER_LONG == 32
 574         unsigned long flags;
 575         local_irq_save(flags);
 576 #else
 577         preempt_disable();
 578 #endif
 579         p = &s->stat_percpu[smp_processor_id()][entry];
 580
 581         if (!end) {
 582                 dm_stat_round(s, shared, p);
 583                 atomic_inc(&shared->in_flight[idx]);
 584         } else {
 585                 unsigned long long duration;
 586                 dm_stat_round(s, shared, p);
 587                 atomic_dec(&shared->in_flight[idx]);
 588                 p->sectors[idx] += len;
 589                 p->ios[idx] += 1;
 590                 p->merges[idx] += stats_aux->merged;
 591                 if (!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)) {
 592                         p->ticks[idx] += duration_jiffies;
 593                         duration = jiffies_to_msecs(duration_jiffies);
 594                 } else {
 595                         p->ticks[idx] += stats_aux->duration_ns;
 596                         duration = stats_aux->duration_ns;
 597                 }
 598                 if (s->n_histogram_entries) {
 599                         unsigned int lo = 0, hi = s->n_histogram_entries + 1;
 600                         while (lo + 1 < hi) {
 601                                 unsigned int mid = (lo + hi) / 2;
 602                                 if (s->histogram_boundaries[mid - 1] > duration) {
 603                                         hi = mid;
 604                                 } else {
 605                                         lo = mid;
 606                                 }
 607
 608                         }
 609                         p->histogram[lo]++;
 610                 }
 611         }
 612
 613 #if BITS_PER_LONG == 32
 614         local_irq_restore(flags);
 615 #else
 616         preempt_enable();
 617 #endif
 618 }
 619
 620 static void __dm_stat_bio(struct dm_stat *s, int bi_rw,
 621                           sector_t bi_sector, sector_t end_sector,
 622                           bool end, unsigned long duration_jiffies,
 623                           struct dm_stats_aux *stats_aux)
 624 {
 625         sector_t rel_sector, offset, todo, fragment_len;
 626         size_t entry;
 627
 628         if (end_sector <= s->start || bi_sector >= s->end)
 629                 return;
 630         if (unlikely(bi_sector < s->start)) {
 631                 rel_sector = 0;
 632                 todo = end_sector - s->start;
 633         } else {
 634                 rel_sector = bi_sector - s->start;
 635                 todo = end_sector - bi_sector;
 636         }
 637         if (unlikely(end_sector > s->end))
 638                 todo -= (end_sector - s->end);
 639
 640         offset = dm_sector_div64(rel_sector, s->step);
 641         entry = rel_sector;
 642         do {
 643                 if (WARN_ON_ONCE(entry >= s->n_entries)) {
 644                         DMCRIT("Invalid area access in region id %d", s->id);
 645                         return;
 646                 }
 647                 fragment_len = todo;
 648                 if (fragment_len > s->step - offset)
 649                         fragment_len = s->step - offset;
 650                 dm_stat_for_entry(s, entry, bi_rw, fragment_len,
 651                                   stats_aux, end, duration_jiffies);
 652                 todo -= fragment_len;
 653                 entry++;
 654                 offset = 0;
 655         } while (unlikely(todo != 0));
 656 }
 657
 658 void dm_stats_account_io(struct dm_stats *stats, unsigned long bi_rw,
 659                          sector_t bi_sector, unsigned int bi_sectors, bool end,
 660                          unsigned long start_time,
 661                          struct dm_stats_aux *stats_aux)
 662 {
 663         struct dm_stat *s;
 664         sector_t end_sector;
 665         struct dm_stats_last_position *last;
 666         bool got_precise_time;
 667         unsigned long duration_jiffies = 0;
 668
 669         if (unlikely(!bi_sectors))
 670                 return;
 671
 672         end_sector = bi_sector + bi_sectors;
 673
 674         if (!end) {
 675                 /*
 676                  * A race condition can at worst result in the merged flag being
 677                  * misrepresented, so we don't have to disable preemption here.
 678                  */
 679                 last = raw_cpu_ptr(stats->last);
 680                 stats_aux->merged =
 681                         (bi_sector == (READ_ONCE(last->last_sector) &&
 682                                        ((bi_rw == WRITE) ==
 683                                         (READ_ONCE(last->last_rw) == WRITE))
 684                                        ));
 685                 WRITE_ONCE(last->last_sector, end_sector);
 686                 WRITE_ONCE(last->last_rw, bi_rw);
 687         } else
 688                 duration_jiffies = jiffies - start_time;
 689
 690         rcu_read_lock();
 691
 692         got_precise_time = false;
 693         list_for_each_entry_rcu(s, &stats->list, list_entry) {
 694                 if (s->stat_flags & STAT_PRECISE_TIMESTAMPS && !got_precise_time) {
 695                         /* start (!end) duration_ns is set by DM core's alloc_io() */
 696                         if (end)
 697                                 stats_aux->duration_ns = ktime_to_ns(ktime_get()) - stats_aux->duration_ns;
 698                         got_precise_time = true;
 699                 }
 700                 __dm_stat_bio(s, bi_rw, bi_sector, end_sector, end, duration_jiffies, stats_aux);
 701         }
 702
 703         rcu_read_unlock();
 704 }
 705
 706 static void __dm_stat_init_temporary_percpu_totals(struct dm_stat_shared *shared,
 707                                                    struct dm_stat *s, size_t x)
 708 {
 709         int cpu;
 710         struct dm_stat_percpu *p;
 711
 712         local_irq_disable();
 713         p = &s->stat_percpu[smp_processor_id()][x];
 714         dm_stat_round(s, shared, p);
 715         local_irq_enable();
 716
 717         shared->tmp.sectors[READ] = 0;
 718         shared->tmp.sectors[WRITE] = 0;
 719         shared->tmp.ios[READ] = 0;
 720         shared->tmp.ios[WRITE] = 0;
 721         shared->tmp.merges[READ] = 0;
 722         shared->tmp.merges[WRITE] = 0;
 723         shared->tmp.ticks[READ] = 0;
 724         shared->tmp.ticks[WRITE] = 0;
 725         shared->tmp.io_ticks[READ] = 0;
 726         shared->tmp.io_ticks[WRITE] = 0;
 727         shared->tmp.io_ticks_total = 0;
 728         shared->tmp.time_in_queue = 0;
 729
 730         if (s->n_histogram_entries)
 731                 memset(shared->tmp.histogram, 0, (s->n_histogram_entries + 1) * sizeof(unsigned long long));
 732
 733         for_each_possible_cpu(cpu) {
 734                 p = &s->stat_percpu[cpu][x];
 735                 shared->tmp.sectors[READ] += READ_ONCE(p->sectors[READ]);
 736                 shared->tmp.sectors[WRITE] += READ_ONCE(p->sectors[WRITE]);
 737                 shared->tmp.ios[READ] += READ_ONCE(p->ios[READ]);
 738                 shared->tmp.ios[WRITE] += READ_ONCE(p->ios[WRITE]);
 739                 shared->tmp.merges[READ] += READ_ONCE(p->merges[READ]);
 740                 shared->tmp.merges[WRITE] += READ_ONCE(p->merges[WRITE]);
 741                 shared->tmp.ticks[READ] += READ_ONCE(p->ticks[READ]);
 742                 shared->tmp.ticks[WRITE] += READ_ONCE(p->ticks[WRITE]);
 743                 shared->tmp.io_ticks[READ] += READ_ONCE(p->io_ticks[READ]);
 744                 shared->tmp.io_ticks[WRITE] += READ_ONCE(p->io_ticks[WRITE]);
 745                 shared->tmp.io_ticks_total += READ_ONCE(p->io_ticks_total);
 746                 shared->tmp.time_in_queue += READ_ONCE(p->time_in_queue);
 747                 if (s->n_histogram_entries) {
 748                         unsigned int i;
 749                         for (i = 0; i < s->n_histogram_entries + 1; i++)
 750                                 shared->tmp.histogram[i] += READ_ONCE(p->histogram[i]);
 751                 }
 752         }
 753 }
 754
 755 static void __dm_stat_clear(struct dm_stat *s, size_t idx_start, size_t idx_end,
 756                             bool init_tmp_percpu_totals)
 757 {
 758         size_t x;
 759         struct dm_stat_shared *shared;
 760         struct dm_stat_percpu *p;
 761
 762         for (x = idx_start; x < idx_end; x++) {
 763                 shared = &s->stat_shared[x];
 764                 if (init_tmp_percpu_totals)
 765                         __dm_stat_init_temporary_percpu_totals(shared, s, x);
 766                 local_irq_disable();
 767                 p = &s->stat_percpu[smp_processor_id()][x];
 768                 p->sectors[READ] -= shared->tmp.sectors[READ];
 769                 p->sectors[WRITE] -= shared->tmp.sectors[WRITE];
 770                 p->ios[READ] -= shared->tmp.ios[READ];
 771                 p->ios[WRITE] -= shared->tmp.ios[WRITE];
 772                 p->merges[READ] -= shared->tmp.merges[READ];
 773                 p->merges[WRITE] -= shared->tmp.merges[WRITE];
 774                 p->ticks[READ] -= shared->tmp.ticks[READ];
 775                 p->ticks[WRITE] -= shared->tmp.ticks[WRITE];
 776                 p->io_ticks[READ] -= shared->tmp.io_ticks[READ];
 777                 p->io_ticks[WRITE] -= shared->tmp.io_ticks[WRITE];
 778                 p->io_ticks_total -= shared->tmp.io_ticks_total;
 779                 p->time_in_queue -= shared->tmp.time_in_queue;
 780                 local_irq_enable();
 781                 if (s->n_histogram_entries) {
 782                         unsigned int i;
 783                         for (i = 0; i < s->n_histogram_entries + 1; i++) {
 784                                 local_irq_disable();
 785                                 p = &s->stat_percpu[smp_processor_id()][x];
 786                                 p->histogram[i] -= shared->tmp.histogram[i];
 787                                 local_irq_enable();
 788                         }
 789                 }
 790                 cond_resched();
 791         }
 792 }
 793
 794 static int dm_stats_clear(struct dm_stats *stats, int id)
 795 {
 796         struct dm_stat *s;
 797
 798         mutex_lock(&stats->mutex);
 799
 800         s = __dm_stats_find(stats, id);
 801         if (!s) {
 802                 mutex_unlock(&stats->mutex);
 803                 return -ENOENT;
 804         }
 805
 806         __dm_stat_clear(s, 0, s->n_entries, true);
 807
 808         mutex_unlock(&stats->mutex);
 809
 810         return 1;
 811 }
 812
 813 /*
 814  * This is like jiffies_to_msec, but works for 64-bit values.
 815  */
 816 static unsigned long long dm_jiffies_to_msec64(struct dm_stat *s, unsigned long long j)
 817 {
 818         unsigned long long result;
 819         unsigned int mult;
 820
 821         if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
 822                 return j;
 823
 824         result = 0;
 825         if (j)
 826                 result = jiffies_to_msecs(j & 0x3fffff);
 827         if (j >= 1 << 22) {
 828                 mult = jiffies_to_msecs(1 << 22);
 829                 result += (unsigned long long)mult * (unsigned long long)jiffies_to_msecs((j >> 22) & 0x3fffff);
 830         }
 831         if (j >= 1ULL << 44)
 832                 result += (unsigned long long)mult * (unsigned long long)mult * (unsigned long long)jiffies_to_msecs(j >> 44);
 833
 834         return result;
 835 }
 836
 837 static int dm_stats_print(struct dm_stats *stats, int id,
 838                           size_t idx_start, size_t idx_len,
 839                           bool clear, char *result, unsigned int maxlen)
 840 {
 841         unsigned int sz = 0;
 842         struct dm_stat *s;
 843         size_t x;
 844         sector_t start, end, step;
 845         size_t idx_end;
 846         struct dm_stat_shared *shared;
 847
 848         /*
 849          * Output format:
 850          *   <start_sector>+<length> counters
 851          */
 852
 853         mutex_lock(&stats->mutex);
 854
 855         s = __dm_stats_find(stats, id);
 856         if (!s) {
 857                 mutex_unlock(&stats->mutex);
 858                 return -ENOENT;
 859         }
 860
 861         idx_end = idx_start + idx_len;
 862         if (idx_end < idx_start ||
 863             idx_end > s->n_entries)
 864                 idx_end = s->n_entries;
 865
 866         if (idx_start > idx_end)
 867                 idx_start = idx_end;
 868
 869         step = s->step;
 870         start = s->start + (step * idx_start);
 871
 872         for (x = idx_start; x < idx_end; x++, start = end) {
 873                 shared = &s->stat_shared[x];
 874                 end = start + step;
 875                 if (unlikely(end > s->end))
 876                         end = s->end;
 877
 878                 __dm_stat_init_temporary_percpu_totals(shared, s, x);
 879
 880                 DMEMIT("%llu+%llu %llu %llu %llu %llu %llu %llu %llu %llu %d %llu %llu %llu %llu",
 881                        (unsigned long long)start,
 882                        (unsigned long long)step,
 883                        shared->tmp.ios[READ],
 884                        shared->tmp.merges[READ],
 885                        shared->tmp.sectors[READ],
 886                        dm_jiffies_to_msec64(s, shared->tmp.ticks[READ]),
 887                        shared->tmp.ios[WRITE],
 888                        shared->tmp.merges[WRITE],
 889                        shared->tmp.sectors[WRITE],
 890                        dm_jiffies_to_msec64(s, shared->tmp.ticks[WRITE]),
 891                        dm_stat_in_flight(shared),
 892                        dm_jiffies_to_msec64(s, shared->tmp.io_ticks_total),
 893                        dm_jiffies_to_msec64(s, shared->tmp.time_in_queue),
 894                        dm_jiffies_to_msec64(s, shared->tmp.io_ticks[READ]),
 895                        dm_jiffies_to_msec64(s, shared->tmp.io_ticks[WRITE]));
 896                 if (s->n_histogram_entries) {
 897                         unsigned int i;
 898                         for (i = 0; i < s->n_histogram_entries + 1; i++) {
 899                                 DMEMIT("%s%llu", !i ? " " : ":", shared->tmp.histogram[i]);
 900                         }
 901                 }
 902                 DMEMIT("\n");
 903
 904                 if (unlikely(sz + 1 >= maxlen))
 905                         goto buffer_overflow;
 906
 907                 cond_resched();
 908         }
 909
 910         if (clear)
 911                 __dm_stat_clear(s, idx_start, idx_end, false);
 912
 913 buffer_overflow:
 914         mutex_unlock(&stats->mutex);
 915
 916         return 1;
 917 }
 918
 919 static int dm_stats_set_aux(struct dm_stats *stats, int id, const char *aux_data)
 920 {
 921         struct dm_stat *s;
 922         const char *new_aux_data;
 923
 924         mutex_lock(&stats->mutex);
 925
 926         s = __dm_stats_find(stats, id);
 927         if (!s) {
 928                 mutex_unlock(&stats->mutex);
 929                 return -ENOENT;
 930         }
 931
 932         new_aux_data = kstrdup(aux_data, GFP_KERNEL);
 933         if (!new_aux_data) {
 934                 mutex_unlock(&stats->mutex);
 935                 return -ENOMEM;
 936         }
 937
 938         kfree(s->aux_data);
 939         s->aux_data = new_aux_data;
 940
 941         mutex_unlock(&stats->mutex);
 942
 943         return 0;
 944 }
 945
 946 static int parse_histogram(const char *h, unsigned int *n_histogram_entries,
 947                            unsigned long long **histogram_boundaries)
 948 {
 949         const char *q;
 950         unsigned int n;
 951         unsigned long long last;
 952
 953         *n_histogram_entries = 1;
 954         for (q = h; *q; q++)
 955                 if (*q == ',')
 956                         (*n_histogram_entries)++;
 957
 958         *histogram_boundaries = kmalloc_array(*n_histogram_entries,
 959                                               sizeof(unsigned long long),
 960                                               GFP_KERNEL);
 961         if (!*histogram_boundaries)
 962                 return -ENOMEM;
 963
 964         n = 0;
 965         last = 0;
 966         while (1) {
 967                 unsigned long long hi;
 968                 int s;
 969                 char ch;
 970                 s = sscanf(h, "%llu%c", &hi, &ch);
 971                 if (!s || (s == 2 && ch != ','))
 972                         return -EINVAL;
 973                 if (hi <= last)
 974                         return -EINVAL;
 975                 last = hi;
 976                 (*histogram_boundaries)[n] = hi;
 977                 if (s == 1)
 978                         return 0;
 979                 h = strchr(h, ',') + 1;
 980                 n++;
 981         }
 982 }
 983
 984 static int message_stats_create(struct mapped_device *md,
 985                                 unsigned int argc, char **argv,
 986                                 char *result, unsigned int maxlen)
 987 {
 988         int r;
 989         int id;
 990         char dummy;
 991         unsigned long long start, end, len, step;
 992         unsigned int divisor;
 993         const char *program_id, *aux_data;
 994         unsigned int stat_flags = 0;
 995
 996         unsigned int n_histogram_entries = 0;
 997         unsigned long long *histogram_boundaries = NULL;
 998
 999         struct dm_arg_set as, as_backup;
1000         const char *a;
1001         unsigned int feature_args;
1002
1003         /*
1004          * Input format:
1005          *   <range> <step> [<extra_parameters> <parameters>] [<program_id> [<aux_data>]]
1006          */
1007
1008         if (argc < 3)
1009                 goto ret_einval;
1010
1011         as.argc = argc;
1012         as.argv = argv;
1013         dm_consume_args(&as, 1);
1014
1015         a = dm_shift_arg(&as);
1016         if (!strcmp(a, "-")) {
1017                 start = 0;
1018                 len = dm_get_size(md);
1019                 if (!len)
1020                         len = 1;
1021         } else if (sscanf(a, "%llu+%llu%c", &start, &len, &dummy) != 2 ||
1022                    start != (sector_t)start || len != (sector_t)len)
1023                 goto ret_einval;
1024
1025         end = start + len;
1026         if (start >= end)
1027                 goto ret_einval;
1028
1029         a = dm_shift_arg(&as);
1030         if (sscanf(a, "/%u%c", &divisor, &dummy) == 1) {
1031                 if (!divisor)
1032                         return -EINVAL;
1033                 step = end - start;
1034                 if (do_div(step, divisor))
1035                         step++;
1036                 if (!step)
1037                         step = 1;
1038         } else if (sscanf(a, "%llu%c", &step, &dummy) != 1 ||
1039                    step != (sector_t)step || !step)
1040                 goto ret_einval;
1041
1042         as_backup = as;
1043         a = dm_shift_arg(&as);
1044         if (a && sscanf(a, "%u%c", &feature_args, &dummy) == 1) {
1045                 while (feature_args--) {
1046                         a = dm_shift_arg(&as);
1047                         if (!a)
1048                                 goto ret_einval;
1049                         if (!strcasecmp(a, "precise_timestamps"))
1050                                 stat_flags |= STAT_PRECISE_TIMESTAMPS;
1051                         else if (!strncasecmp(a, "histogram:", 10)) {
1052                                 if (n_histogram_entries)
1053                                         goto ret_einval;
1054                                 if ((r = parse_histogram(a + 10, &n_histogram_entries, &histogram_boundaries)))
1055                                         goto ret;
1056                         } else
1057                                 goto ret_einval;
1058                 }
1059         } else {
1060                 as = as_backup;
1061         }
1062
1063         program_id = "-";
1064         aux_data = "-";
1065
1066         a = dm_shift_arg(&as);
1067         if (a)
1068                 program_id = a;
1069
1070         a = dm_shift_arg(&as);
1071         if (a)
1072                 aux_data = a;
1073
1074         if (as.argc)
1075                 goto ret_einval;
1076
1077         /*
1078          * If a buffer overflow happens after we created the region,
1079          * it's too late (the userspace would retry with a larger
1080          * buffer, but the region id that caused the overflow is already
1081          * leaked).  So we must detect buffer overflow in advance.
1082          */
1083         snprintf(result, maxlen, "%d", INT_MAX);
1084         if (dm_message_test_buffer_overflow(result, maxlen)) {
1085                 r = 1;
1086                 goto ret;
1087         }
1088
1089         id = dm_stats_create(dm_get_stats(md), start, end, step, stat_flags,
1090                              n_histogram_entries, histogram_boundaries, program_id, aux_data,
1091                              dm_internal_suspend_fast, dm_internal_resume_fast, md);
1092         if (id < 0) {
1093                 r = id;
1094                 goto ret;
1095         }
1096
1097         snprintf(result, maxlen, "%d", id);
1098
1099         r = 1;
1100         goto ret;
1101
1102 ret_einval:
1103         r = -EINVAL;
1104 ret:
1105         kfree(histogram_boundaries);
1106         return r;
1107 }
1108
1109 static int message_stats_delete(struct mapped_device *md,
1110                                 unsigned int argc, char **argv)
1111 {
1112         int id;
1113         char dummy;
1114
1115         if (argc != 2)
1116                 return -EINVAL;
1117
1118         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1119                 return -EINVAL;
1120
1121         return dm_stats_delete(dm_get_stats(md), id);
1122 }
1123
1124 static int message_stats_clear(struct mapped_device *md,
1125                                unsigned int argc, char **argv)
1126 {
1127         int id;
1128         char dummy;
1129
1130         if (argc != 2)
1131                 return -EINVAL;
1132
1133         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1134                 return -EINVAL;
1135
1136         return dm_stats_clear(dm_get_stats(md), id);
1137 }
1138
1139 static int message_stats_list(struct mapped_device *md,
1140                               unsigned int argc, char **argv,
1141                               char *result, unsigned int maxlen)
1142 {
1143         int r;
1144         const char *program = NULL;
1145
1146         if (argc < 1 || argc > 2)
1147                 return -EINVAL;
1148
1149         if (argc > 1) {
1150                 program = kstrdup(argv[1], GFP_KERNEL);
1151                 if (!program)
1152                         return -ENOMEM;
1153         }
1154
1155         r = dm_stats_list(dm_get_stats(md), program, result, maxlen);
1156
1157         kfree(program);
1158
1159         return r;
1160 }
1161
1162 static int message_stats_print(struct mapped_device *md,
1163                                unsigned int argc, char **argv, bool clear,
1164                                char *result, unsigned int maxlen)
1165 {
1166         int id;
1167         char dummy;
1168         unsigned long idx_start = 0, idx_len = ULONG_MAX;
1169
1170         if (argc != 2 && argc != 4)
1171                 return -EINVAL;
1172
1173         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1174                 return -EINVAL;
1175
1176         if (argc > 3) {
1177                 if (strcmp(argv[2], "-") &&
1178                     sscanf(argv[2], "%lu%c", &idx_start, &dummy) != 1)
1179                         return -EINVAL;
1180                 if (strcmp(argv[3], "-") &&
1181                     sscanf(argv[3], "%lu%c", &idx_len, &dummy) != 1)
1182                         return -EINVAL;
1183         }
1184
1185         return dm_stats_print(dm_get_stats(md), id, idx_start, idx_len, clear,
1186                               result, maxlen);
1187 }
1188
1189 static int message_stats_set_aux(struct mapped_device *md,
1190                                  unsigned int argc, char **argv)
1191 {
1192         int id;
1193         char dummy;
1194
1195         if (argc != 3)
1196                 return -EINVAL;
1197
1198         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1199                 return -EINVAL;
1200
1201         return dm_stats_set_aux(dm_get_stats(md), id, argv[2]);
1202 }
1203
1204 int dm_stats_message(struct mapped_device *md, unsigned int argc, char **argv,
1205                      char *result, unsigned int maxlen)
1206 {
1207         int r;
1208
1209         /* All messages here must start with '@' */
1210         if (!strcasecmp(argv[0], "@stats_create"))
1211                 r = message_stats_create(md, argc, argv, result, maxlen);
1212         else if (!strcasecmp(argv[0], "@stats_delete"))
1213                 r = message_stats_delete(md, argc, argv);
1214         else if (!strcasecmp(argv[0], "@stats_clear"))
1215                 r = message_stats_clear(md, argc, argv);
1216         else if (!strcasecmp(argv[0], "@stats_list"))
1217                 r = message_stats_list(md, argc, argv, result, maxlen);
1218         else if (!strcasecmp(argv[0], "@stats_print"))
1219                 r = message_stats_print(md, argc, argv, false, result, maxlen);
1220         else if (!strcasecmp(argv[0], "@stats_print_clear"))
1221                 r = message_stats_print(md, argc, argv, true, result, maxlen);
1222         else if (!strcasecmp(argv[0], "@stats_set_aux"))
1223                 r = message_stats_set_aux(md, argc, argv);
1224         else
1225                 return 2; /* this wasn't a stats message */
1226
1227         if (r == -EINVAL)
1228                 DMCRIT("Invalid parameters for message %s", argv[0]);
1229
1230         return r;
1231 }
1232
1233 int __init dm_statistics_init(void)
1234 {
1235         shared_memory_amount = 0;
1236         dm_stat_need_rcu_barrier = 0;
1237         return 0;
1238 }
1239
1240 void dm_statistics_exit(void)
1241 {
1242         if (dm_stat_need_rcu_barrier)
1243                 rcu_barrier();
1244         if (WARN_ON(shared_memory_amount))
1245                 DMCRIT("shared_memory_amount leaked: %lu", shared_memory_amount);
1246 }
1247
1248 module_param_named(stats_current_allocated_bytes, shared_memory_amount, ulong, S_IRUGO);
1249 MODULE_PARM_DESC(stats_current_allocated_bytes, "Memory currently used by statistics");