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 stat_flags;
  46         size_t n_entries;
  47         sector_t start;
  48         sector_t end;
  49         sector_t step;
  50         unsigned 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 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 void 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         for_each_possible_cpu(cpu) {
 201                 last = per_cpu_ptr(stats->last, cpu);
 202                 last->last_sector = (sector_t)ULLONG_MAX;
 203                 last->last_rw = UINT_MAX;
 204         }
 205 }
 206
 207 void dm_stats_cleanup(struct dm_stats *stats)
 208 {
 209         size_t ni;
 210         struct dm_stat *s;
 211         struct dm_stat_shared *shared;
 212
 213         while (!list_empty(&stats->list)) {
 214                 s = container_of(stats->list.next, struct dm_stat, list_entry);
 215                 list_del(&s->list_entry);
 216                 for (ni = 0; ni < s->n_entries; ni++) {
 217                         shared = &s->stat_shared[ni];
 218                         if (WARN_ON(dm_stat_in_flight(shared))) {
 219                                 DMCRIT("leaked in-flight counter at index %lu "
 220                                        "(start %llu, end %llu, step %llu): reads %d, writes %d",
 221                                        (unsigned long)ni,
 222                                        (unsigned long long)s->start,
 223                                        (unsigned long long)s->end,
 224                                        (unsigned long long)s->step,
 225                                        atomic_read(&shared->in_flight[READ]),
 226                                        atomic_read(&shared->in_flight[WRITE]));
 227                         }
 228                         cond_resched();
 229                 }
 230                 dm_stat_free(&s->rcu_head);
 231         }
 232         free_percpu(stats->last);
 233         mutex_destroy(&stats->mutex);
 234 }
 235
 236 static void dm_stats_recalc_precise_timestamps(struct dm_stats *stats)
 237 {
 238         struct list_head *l;
 239         struct dm_stat *tmp_s;
 240         bool precise_timestamps = false;
 241
 242         list_for_each(l, &stats->list) {
 243                 tmp_s = container_of(l, struct dm_stat, list_entry);
 244                 if (tmp_s->stat_flags & STAT_PRECISE_TIMESTAMPS) {
 245                         precise_timestamps = true;
 246                         break;
 247                 }
 248         }
 249         stats->precise_timestamps = precise_timestamps;
 250 }
 251
 252 static int dm_stats_create(struct dm_stats *stats, sector_t start, sector_t end,
 253                            sector_t step, unsigned stat_flags,
 254                            unsigned n_histogram_entries,
 255                            unsigned long long *histogram_boundaries,
 256                            const char *program_id, const char *aux_data,
 257                            void (*suspend_callback)(struct mapped_device *),
 258                            void (*resume_callback)(struct mapped_device *),
 259                            struct mapped_device *md)
 260 {
 261         struct list_head *l;
 262         struct dm_stat *s, *tmp_s;
 263         sector_t n_entries;
 264         size_t ni;
 265         size_t shared_alloc_size;
 266         size_t percpu_alloc_size;
 267         size_t histogram_alloc_size;
 268         struct dm_stat_percpu *p;
 269         int cpu;
 270         int ret_id;
 271         int r;
 272
 273         if (end < start || !step)
 274                 return -EINVAL;
 275
 276         n_entries = end - start;
 277         if (dm_sector_div64(n_entries, step))
 278                 n_entries++;
 279
 280         if (n_entries != (size_t)n_entries || !(size_t)(n_entries + 1))
 281                 return -EOVERFLOW;
 282
 283         shared_alloc_size = struct_size(s, stat_shared, n_entries);
 284         if ((shared_alloc_size - sizeof(struct dm_stat)) / sizeof(struct dm_stat_shared) != n_entries)
 285                 return -EOVERFLOW;
 286
 287         percpu_alloc_size = (size_t)n_entries * sizeof(struct dm_stat_percpu);
 288         if (percpu_alloc_size / sizeof(struct dm_stat_percpu) != n_entries)
 289                 return -EOVERFLOW;
 290
 291         histogram_alloc_size = (n_histogram_entries + 1) * (size_t)n_entries * sizeof(unsigned long long);
 292         if (histogram_alloc_size / (n_histogram_entries + 1) != (size_t)n_entries * sizeof(unsigned long long))
 293                 return -EOVERFLOW;
 294
 295         if (!check_shared_memory(shared_alloc_size + histogram_alloc_size +
 296                                  num_possible_cpus() * (percpu_alloc_size + histogram_alloc_size)))
 297                 return -ENOMEM;
 298
 299         s = dm_kvzalloc(shared_alloc_size, NUMA_NO_NODE);
 300         if (!s)
 301                 return -ENOMEM;
 302
 303         s->stat_flags = stat_flags;
 304         s->n_entries = n_entries;
 305         s->start = start;
 306         s->end = end;
 307         s->step = step;
 308         s->shared_alloc_size = shared_alloc_size;
 309         s->percpu_alloc_size = percpu_alloc_size;
 310         s->histogram_alloc_size = histogram_alloc_size;
 311
 312         s->n_histogram_entries = n_histogram_entries;
 313         s->histogram_boundaries = kmemdup(histogram_boundaries,
 314                                           s->n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
 315         if (!s->histogram_boundaries) {
 316                 r = -ENOMEM;
 317                 goto out;
 318         }
 319
 320         s->program_id = kstrdup(program_id, GFP_KERNEL);
 321         if (!s->program_id) {
 322                 r = -ENOMEM;
 323                 goto out;
 324         }
 325         s->aux_data = kstrdup(aux_data, GFP_KERNEL);
 326         if (!s->aux_data) {
 327                 r = -ENOMEM;
 328                 goto out;
 329         }
 330
 331         for (ni = 0; ni < n_entries; ni++) {
 332                 atomic_set(&s->stat_shared[ni].in_flight[READ], 0);
 333                 atomic_set(&s->stat_shared[ni].in_flight[WRITE], 0);
 334                 cond_resched();
 335         }
 336
 337         if (s->n_histogram_entries) {
 338                 unsigned long long *hi;
 339                 hi = dm_kvzalloc(s->histogram_alloc_size, NUMA_NO_NODE);
 340                 if (!hi) {
 341                         r = -ENOMEM;
 342                         goto out;
 343                 }
 344                 for (ni = 0; ni < n_entries; ni++) {
 345                         s->stat_shared[ni].tmp.histogram = hi;
 346                         hi += s->n_histogram_entries + 1;
 347                         cond_resched();
 348                 }
 349         }
 350
 351         for_each_possible_cpu(cpu) {
 352                 p = dm_kvzalloc(percpu_alloc_size, cpu_to_node(cpu));
 353                 if (!p) {
 354                         r = -ENOMEM;
 355                         goto out;
 356                 }
 357                 s->stat_percpu[cpu] = p;
 358                 if (s->n_histogram_entries) {
 359                         unsigned long long *hi;
 360                         hi = dm_kvzalloc(s->histogram_alloc_size, cpu_to_node(cpu));
 361                         if (!hi) {
 362                                 r = -ENOMEM;
 363                                 goto out;
 364                         }
 365                         for (ni = 0; ni < n_entries; ni++) {
 366                                 p[ni].histogram = hi;
 367                                 hi += s->n_histogram_entries + 1;
 368                                 cond_resched();
 369                         }
 370                 }
 371         }
 372
 373         /*
 374          * Suspend/resume to make sure there is no i/o in flight,
 375          * so that newly created statistics will be exact.
 376          *
 377          * (note: we couldn't suspend earlier because we must not
 378          * allocate memory while suspended)
 379          */
 380         suspend_callback(md);
 381
 382         mutex_lock(&stats->mutex);
 383         s->id = 0;
 384         list_for_each(l, &stats->list) {
 385                 tmp_s = container_of(l, struct dm_stat, list_entry);
 386                 if (WARN_ON(tmp_s->id < s->id)) {
 387                         r = -EINVAL;
 388                         goto out_unlock_resume;
 389                 }
 390                 if (tmp_s->id > s->id)
 391                         break;
 392                 if (unlikely(s->id == INT_MAX)) {
 393                         r = -ENFILE;
 394                         goto out_unlock_resume;
 395                 }
 396                 s->id++;
 397         }
 398         ret_id = s->id;
 399         list_add_tail_rcu(&s->list_entry, l);
 400
 401         dm_stats_recalc_precise_timestamps(stats);
 402
 403         mutex_unlock(&stats->mutex);
 404
 405         resume_callback(md);
 406
 407         return ret_id;
 408
 409 out_unlock_resume:
 410         mutex_unlock(&stats->mutex);
 411         resume_callback(md);
 412 out:
 413         dm_stat_free(&s->rcu_head);
 414         return r;
 415 }
 416
 417 static struct dm_stat *__dm_stats_find(struct dm_stats *stats, int id)
 418 {
 419         struct dm_stat *s;
 420
 421         list_for_each_entry(s, &stats->list, list_entry) {
 422                 if (s->id > id)
 423                         break;
 424                 if (s->id == id)
 425                         return s;
 426         }
 427
 428         return NULL;
 429 }
 430
 431 static int dm_stats_delete(struct dm_stats *stats, int id)
 432 {
 433         struct dm_stat *s;
 434         int cpu;
 435
 436         mutex_lock(&stats->mutex);
 437
 438         s = __dm_stats_find(stats, id);
 439         if (!s) {
 440                 mutex_unlock(&stats->mutex);
 441                 return -ENOENT;
 442         }
 443
 444         list_del_rcu(&s->list_entry);
 445
 446         dm_stats_recalc_precise_timestamps(stats);
 447
 448         mutex_unlock(&stats->mutex);
 449
 450         /*
 451          * vfree can't be called from RCU callback
 452          */
 453         for_each_possible_cpu(cpu)
 454                 if (is_vmalloc_addr(s->stat_percpu) ||
 455                     is_vmalloc_addr(s->stat_percpu[cpu][0].histogram))
 456                         goto do_sync_free;
 457         if (is_vmalloc_addr(s) ||
 458             is_vmalloc_addr(s->stat_shared[0].tmp.histogram)) {
 459 do_sync_free:
 460                 synchronize_rcu_expedited();
 461                 dm_stat_free(&s->rcu_head);
 462         } else {
 463                 WRITE_ONCE(dm_stat_need_rcu_barrier, 1);
 464                 call_rcu(&s->rcu_head, dm_stat_free);
 465         }
 466         return 0;
 467 }
 468
 469 static int dm_stats_list(struct dm_stats *stats, const char *program,
 470                          char *result, unsigned maxlen)
 471 {
 472         struct dm_stat *s;
 473         sector_t len;
 474         unsigned sz = 0;
 475
 476         /*
 477          * Output format:
 478          *   <region_id>: <start_sector>+<length> <step> <program_id> <aux_data>
 479          */
 480
 481         mutex_lock(&stats->mutex);
 482         list_for_each_entry(s, &stats->list, list_entry) {
 483                 if (!program || !strcmp(program, s->program_id)) {
 484                         len = s->end - s->start;
 485                         DMEMIT("%d: %llu+%llu %llu %s %s", s->id,
 486                                 (unsigned long long)s->start,
 487                                 (unsigned long long)len,
 488                                 (unsigned long long)s->step,
 489                                 s->program_id,
 490                                 s->aux_data);
 491                         if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
 492                                 DMEMIT(" precise_timestamps");
 493                         if (s->n_histogram_entries) {
 494                                 unsigned i;
 495                                 DMEMIT(" histogram:");
 496                                 for (i = 0; i < s->n_histogram_entries; i++) {
 497                                         if (i)
 498                                                 DMEMIT(",");
 499                                         DMEMIT("%llu", s->histogram_boundaries[i]);
 500                                 }
 501                         }
 502                         DMEMIT("\n");
 503                 }
 504                 cond_resched();
 505         }
 506         mutex_unlock(&stats->mutex);
 507
 508         return 1;
 509 }
 510
 511 static void dm_stat_round(struct dm_stat *s, struct dm_stat_shared *shared,
 512                           struct dm_stat_percpu *p)
 513 {
 514         /*
 515          * This is racy, but so is part_round_stats_single.
 516          */
 517         unsigned long long now, difference;
 518         unsigned in_flight_read, in_flight_write;
 519
 520         if (likely(!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)))
 521                 now = jiffies;
 522         else
 523                 now = ktime_to_ns(ktime_get());
 524
 525         difference = now - shared->stamp;
 526         if (!difference)
 527                 return;
 528
 529         in_flight_read = (unsigned)atomic_read(&shared->in_flight[READ]);
 530         in_flight_write = (unsigned)atomic_read(&shared->in_flight[WRITE]);
 531         if (in_flight_read)
 532                 p->io_ticks[READ] += difference;
 533         if (in_flight_write)
 534                 p->io_ticks[WRITE] += difference;
 535         if (in_flight_read + in_flight_write) {
 536                 p->io_ticks_total += difference;
 537                 p->time_in_queue += (in_flight_read + in_flight_write) * difference;
 538         }
 539         shared->stamp = now;
 540 }
 541
 542 static void dm_stat_for_entry(struct dm_stat *s, size_t entry,
 543                               int idx, sector_t len,
 544                               struct dm_stats_aux *stats_aux, bool end,
 545                               unsigned long duration_jiffies)
 546 {
 547         struct dm_stat_shared *shared = &s->stat_shared[entry];
 548         struct dm_stat_percpu *p;
 549
 550         /*
 551          * For strict correctness we should use local_irq_save/restore
 552          * instead of preempt_disable/enable.
 553          *
 554          * preempt_disable/enable is racy if the driver finishes bios
 555          * from non-interrupt context as well as from interrupt context
 556          * or from more different interrupts.
 557          *
 558          * On 64-bit architectures the race only results in not counting some
 559          * events, so it is acceptable.  On 32-bit architectures the race could
 560          * cause the counter going off by 2^32, so we need to do proper locking
 561          * there.
 562          *
 563          * part_stat_lock()/part_stat_unlock() have this race too.
 564          */
 565 #if BITS_PER_LONG == 32
 566         unsigned long flags;
 567         local_irq_save(flags);
 568 #else
 569         preempt_disable();
 570 #endif
 571         p = &s->stat_percpu[smp_processor_id()][entry];
 572
 573         if (!end) {
 574                 dm_stat_round(s, shared, p);
 575                 atomic_inc(&shared->in_flight[idx]);
 576         } else {
 577                 unsigned long long duration;
 578                 dm_stat_round(s, shared, p);
 579                 atomic_dec(&shared->in_flight[idx]);
 580                 p->sectors[idx] += len;
 581                 p->ios[idx] += 1;
 582                 p->merges[idx] += stats_aux->merged;
 583                 if (!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)) {
 584                         p->ticks[idx] += duration_jiffies;
 585                         duration = jiffies_to_msecs(duration_jiffies);
 586                 } else {
 587                         p->ticks[idx] += stats_aux->duration_ns;
 588                         duration = stats_aux->duration_ns;
 589                 }
 590                 if (s->n_histogram_entries) {
 591                         unsigned lo = 0, hi = s->n_histogram_entries + 1;
 592                         while (lo + 1 < hi) {
 593                                 unsigned mid = (lo + hi) / 2;
 594                                 if (s->histogram_boundaries[mid - 1] > duration) {
 595                                         hi = mid;
 596                                 } else {
 597                                         lo = mid;
 598                                 }
 599
 600                         }
 601                         p->histogram[lo]++;
 602                 }
 603         }
 604
 605 #if BITS_PER_LONG == 32
 606         local_irq_restore(flags);
 607 #else
 608         preempt_enable();
 609 #endif
 610 }
 611
 612 static void __dm_stat_bio(struct dm_stat *s, int bi_rw,
 613                           sector_t bi_sector, sector_t end_sector,
 614                           bool end, unsigned long duration_jiffies,
 615                           struct dm_stats_aux *stats_aux)
 616 {
 617         sector_t rel_sector, offset, todo, fragment_len;
 618         size_t entry;
 619
 620         if (end_sector <= s->start || bi_sector >= s->end)
 621                 return;
 622         if (unlikely(bi_sector < s->start)) {
 623                 rel_sector = 0;
 624                 todo = end_sector - s->start;
 625         } else {
 626                 rel_sector = bi_sector - s->start;
 627                 todo = end_sector - bi_sector;
 628         }
 629         if (unlikely(end_sector > s->end))
 630                 todo -= (end_sector - s->end);
 631
 632         offset = dm_sector_div64(rel_sector, s->step);
 633         entry = rel_sector;
 634         do {
 635                 if (WARN_ON_ONCE(entry >= s->n_entries)) {
 636                         DMCRIT("Invalid area access in region id %d", s->id);
 637                         return;
 638                 }
 639                 fragment_len = todo;
 640                 if (fragment_len > s->step - offset)
 641                         fragment_len = s->step - offset;
 642                 dm_stat_for_entry(s, entry, bi_rw, fragment_len,
 643                                   stats_aux, end, duration_jiffies);
 644                 todo -= fragment_len;
 645                 entry++;
 646                 offset = 0;
 647         } while (unlikely(todo != 0));
 648 }
 649
 650 void dm_stats_account_io(struct dm_stats *stats, unsigned long bi_rw,
 651                          sector_t bi_sector, unsigned bi_sectors, bool end,
 652                          unsigned long start_time,
 653                          struct dm_stats_aux *stats_aux)
 654 {
 655         struct dm_stat *s;
 656         sector_t end_sector;
 657         struct dm_stats_last_position *last;
 658         bool got_precise_time;
 659         unsigned long duration_jiffies = 0;
 660
 661         if (unlikely(!bi_sectors))
 662                 return;
 663
 664         end_sector = bi_sector + bi_sectors;
 665
 666         if (!end) {
 667                 /*
 668                  * A race condition can at worst result in the merged flag being
 669                  * misrepresented, so we don't have to disable preemption here.
 670                  */
 671                 last = raw_cpu_ptr(stats->last);
 672                 stats_aux->merged =
 673                         (bi_sector == (READ_ONCE(last->last_sector) &&
 674                                        ((bi_rw == WRITE) ==
 675                                         (READ_ONCE(last->last_rw) == WRITE))
 676                                        ));
 677                 WRITE_ONCE(last->last_sector, end_sector);
 678                 WRITE_ONCE(last->last_rw, bi_rw);
 679         } else
 680                 duration_jiffies = jiffies - start_time;
 681
 682         rcu_read_lock();
 683
 684         got_precise_time = false;
 685         list_for_each_entry_rcu(s, &stats->list, list_entry) {
 686                 if (s->stat_flags & STAT_PRECISE_TIMESTAMPS && !got_precise_time) {
 687                         /* start (!end) duration_ns is set by DM core's alloc_io() */
 688                         if (end)
 689                                 stats_aux->duration_ns = ktime_to_ns(ktime_get()) - stats_aux->duration_ns;
 690                         got_precise_time = true;
 691                 }
 692                 __dm_stat_bio(s, bi_rw, bi_sector, end_sector, end, duration_jiffies, stats_aux);
 693         }
 694
 695         rcu_read_unlock();
 696 }
 697
 698 static void __dm_stat_init_temporary_percpu_totals(struct dm_stat_shared *shared,
 699                                                    struct dm_stat *s, size_t x)
 700 {
 701         int cpu;
 702         struct dm_stat_percpu *p;
 703
 704         local_irq_disable();
 705         p = &s->stat_percpu[smp_processor_id()][x];
 706         dm_stat_round(s, shared, p);
 707         local_irq_enable();
 708
 709         shared->tmp.sectors[READ] = 0;
 710         shared->tmp.sectors[WRITE] = 0;
 711         shared->tmp.ios[READ] = 0;
 712         shared->tmp.ios[WRITE] = 0;
 713         shared->tmp.merges[READ] = 0;
 714         shared->tmp.merges[WRITE] = 0;
 715         shared->tmp.ticks[READ] = 0;
 716         shared->tmp.ticks[WRITE] = 0;
 717         shared->tmp.io_ticks[READ] = 0;
 718         shared->tmp.io_ticks[WRITE] = 0;
 719         shared->tmp.io_ticks_total = 0;
 720         shared->tmp.time_in_queue = 0;
 721
 722         if (s->n_histogram_entries)
 723                 memset(shared->tmp.histogram, 0, (s->n_histogram_entries + 1) * sizeof(unsigned long long));
 724
 725         for_each_possible_cpu(cpu) {
 726                 p = &s->stat_percpu[cpu][x];
 727                 shared->tmp.sectors[READ] += READ_ONCE(p->sectors[READ]);
 728                 shared->tmp.sectors[WRITE] += READ_ONCE(p->sectors[WRITE]);
 729                 shared->tmp.ios[READ] += READ_ONCE(p->ios[READ]);
 730                 shared->tmp.ios[WRITE] += READ_ONCE(p->ios[WRITE]);
 731                 shared->tmp.merges[READ] += READ_ONCE(p->merges[READ]);
 732                 shared->tmp.merges[WRITE] += READ_ONCE(p->merges[WRITE]);
 733                 shared->tmp.ticks[READ] += READ_ONCE(p->ticks[READ]);
 734                 shared->tmp.ticks[WRITE] += READ_ONCE(p->ticks[WRITE]);
 735                 shared->tmp.io_ticks[READ] += READ_ONCE(p->io_ticks[READ]);
 736                 shared->tmp.io_ticks[WRITE] += READ_ONCE(p->io_ticks[WRITE]);
 737                 shared->tmp.io_ticks_total += READ_ONCE(p->io_ticks_total);
 738                 shared->tmp.time_in_queue += READ_ONCE(p->time_in_queue);
 739                 if (s->n_histogram_entries) {
 740                         unsigned i;
 741                         for (i = 0; i < s->n_histogram_entries + 1; i++)
 742                                 shared->tmp.histogram[i] += READ_ONCE(p->histogram[i]);
 743                 }
 744         }
 745 }
 746
 747 static void __dm_stat_clear(struct dm_stat *s, size_t idx_start, size_t idx_end,
 748                             bool init_tmp_percpu_totals)
 749 {
 750         size_t x;
 751         struct dm_stat_shared *shared;
 752         struct dm_stat_percpu *p;
 753
 754         for (x = idx_start; x < idx_end; x++) {
 755                 shared = &s->stat_shared[x];
 756                 if (init_tmp_percpu_totals)
 757                         __dm_stat_init_temporary_percpu_totals(shared, s, x);
 758                 local_irq_disable();
 759                 p = &s->stat_percpu[smp_processor_id()][x];
 760                 p->sectors[READ] -= shared->tmp.sectors[READ];
 761                 p->sectors[WRITE] -= shared->tmp.sectors[WRITE];
 762                 p->ios[READ] -= shared->tmp.ios[READ];
 763                 p->ios[WRITE] -= shared->tmp.ios[WRITE];
 764                 p->merges[READ] -= shared->tmp.merges[READ];
 765                 p->merges[WRITE] -= shared->tmp.merges[WRITE];
 766                 p->ticks[READ] -= shared->tmp.ticks[READ];
 767                 p->ticks[WRITE] -= shared->tmp.ticks[WRITE];
 768                 p->io_ticks[READ] -= shared->tmp.io_ticks[READ];
 769                 p->io_ticks[WRITE] -= shared->tmp.io_ticks[WRITE];
 770                 p->io_ticks_total -= shared->tmp.io_ticks_total;
 771                 p->time_in_queue -= shared->tmp.time_in_queue;
 772                 local_irq_enable();
 773                 if (s->n_histogram_entries) {
 774                         unsigned i;
 775                         for (i = 0; i < s->n_histogram_entries + 1; i++) {
 776                                 local_irq_disable();
 777                                 p = &s->stat_percpu[smp_processor_id()][x];
 778                                 p->histogram[i] -= shared->tmp.histogram[i];
 779                                 local_irq_enable();
 780                         }
 781                 }
 782                 cond_resched();
 783         }
 784 }
 785
 786 static int dm_stats_clear(struct dm_stats *stats, int id)
 787 {
 788         struct dm_stat *s;
 789
 790         mutex_lock(&stats->mutex);
 791
 792         s = __dm_stats_find(stats, id);
 793         if (!s) {
 794                 mutex_unlock(&stats->mutex);
 795                 return -ENOENT;
 796         }
 797
 798         __dm_stat_clear(s, 0, s->n_entries, true);
 799
 800         mutex_unlock(&stats->mutex);
 801
 802         return 1;
 803 }
 804
 805 /*
 806  * This is like jiffies_to_msec, but works for 64-bit values.
 807  */
 808 static unsigned long long dm_jiffies_to_msec64(struct dm_stat *s, unsigned long long j)
 809 {
 810         unsigned long long result;
 811         unsigned mult;
 812
 813         if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
 814                 return j;
 815
 816         result = 0;
 817         if (j)
 818                 result = jiffies_to_msecs(j & 0x3fffff);
 819         if (j >= 1 << 22) {
 820                 mult = jiffies_to_msecs(1 << 22);
 821                 result += (unsigned long long)mult * (unsigned long long)jiffies_to_msecs((j >> 22) & 0x3fffff);
 822         }
 823         if (j >= 1ULL << 44)
 824                 result += (unsigned long long)mult * (unsigned long long)mult * (unsigned long long)jiffies_to_msecs(j >> 44);
 825
 826         return result;
 827 }
 828
 829 static int dm_stats_print(struct dm_stats *stats, int id,
 830                           size_t idx_start, size_t idx_len,
 831                           bool clear, char *result, unsigned maxlen)
 832 {
 833         unsigned sz = 0;
 834         struct dm_stat *s;
 835         size_t x;
 836         sector_t start, end, step;
 837         size_t idx_end;
 838         struct dm_stat_shared *shared;
 839
 840         /*
 841          * Output format:
 842          *   <start_sector>+<length> counters
 843          */
 844
 845         mutex_lock(&stats->mutex);
 846
 847         s = __dm_stats_find(stats, id);
 848         if (!s) {
 849                 mutex_unlock(&stats->mutex);
 850                 return -ENOENT;
 851         }
 852
 853         idx_end = idx_start + idx_len;
 854         if (idx_end < idx_start ||
 855             idx_end > s->n_entries)
 856                 idx_end = s->n_entries;
 857
 858         if (idx_start > idx_end)
 859                 idx_start = idx_end;
 860
 861         step = s->step;
 862         start = s->start + (step * idx_start);
 863
 864         for (x = idx_start; x < idx_end; x++, start = end) {
 865                 shared = &s->stat_shared[x];
 866                 end = start + step;
 867                 if (unlikely(end > s->end))
 868                         end = s->end;
 869
 870                 __dm_stat_init_temporary_percpu_totals(shared, s, x);
 871
 872                 DMEMIT("%llu+%llu %llu %llu %llu %llu %llu %llu %llu %llu %d %llu %llu %llu %llu",
 873                        (unsigned long long)start,
 874                        (unsigned long long)step,
 875                        shared->tmp.ios[READ],
 876                        shared->tmp.merges[READ],
 877                        shared->tmp.sectors[READ],
 878                        dm_jiffies_to_msec64(s, shared->tmp.ticks[READ]),
 879                        shared->tmp.ios[WRITE],
 880                        shared->tmp.merges[WRITE],
 881                        shared->tmp.sectors[WRITE],
 882                        dm_jiffies_to_msec64(s, shared->tmp.ticks[WRITE]),
 883                        dm_stat_in_flight(shared),
 884                        dm_jiffies_to_msec64(s, shared->tmp.io_ticks_total),
 885                        dm_jiffies_to_msec64(s, shared->tmp.time_in_queue),
 886                        dm_jiffies_to_msec64(s, shared->tmp.io_ticks[READ]),
 887                        dm_jiffies_to_msec64(s, shared->tmp.io_ticks[WRITE]));
 888                 if (s->n_histogram_entries) {
 889                         unsigned i;
 890                         for (i = 0; i < s->n_histogram_entries + 1; i++) {
 891                                 DMEMIT("%s%llu", !i ? " " : ":", shared->tmp.histogram[i]);
 892                         }
 893                 }
 894                 DMEMIT("\n");
 895
 896                 if (unlikely(sz + 1 >= maxlen))
 897                         goto buffer_overflow;
 898
 899                 cond_resched();
 900         }
 901
 902         if (clear)
 903                 __dm_stat_clear(s, idx_start, idx_end, false);
 904
 905 buffer_overflow:
 906         mutex_unlock(&stats->mutex);
 907
 908         return 1;
 909 }
 910
 911 static int dm_stats_set_aux(struct dm_stats *stats, int id, const char *aux_data)
 912 {
 913         struct dm_stat *s;
 914         const char *new_aux_data;
 915
 916         mutex_lock(&stats->mutex);
 917
 918         s = __dm_stats_find(stats, id);
 919         if (!s) {
 920                 mutex_unlock(&stats->mutex);
 921                 return -ENOENT;
 922         }
 923
 924         new_aux_data = kstrdup(aux_data, GFP_KERNEL);
 925         if (!new_aux_data) {
 926                 mutex_unlock(&stats->mutex);
 927                 return -ENOMEM;
 928         }
 929
 930         kfree(s->aux_data);
 931         s->aux_data = new_aux_data;
 932
 933         mutex_unlock(&stats->mutex);
 934
 935         return 0;
 936 }
 937
 938 static int parse_histogram(const char *h, unsigned *n_histogram_entries,
 939                            unsigned long long **histogram_boundaries)
 940 {
 941         const char *q;
 942         unsigned n;
 943         unsigned long long last;
 944
 945         *n_histogram_entries = 1;
 946         for (q = h; *q; q++)
 947                 if (*q == ',')
 948                         (*n_histogram_entries)++;
 949
 950         *histogram_boundaries = kmalloc_array(*n_histogram_entries,
 951                                               sizeof(unsigned long long),
 952                                               GFP_KERNEL);
 953         if (!*histogram_boundaries)
 954                 return -ENOMEM;
 955
 956         n = 0;
 957         last = 0;
 958         while (1) {
 959                 unsigned long long hi;
 960                 int s;
 961                 char ch;
 962                 s = sscanf(h, "%llu%c", &hi, &ch);
 963                 if (!s || (s == 2 && ch != ','))
 964                         return -EINVAL;
 965                 if (hi <= last)
 966                         return -EINVAL;
 967                 last = hi;
 968                 (*histogram_boundaries)[n] = hi;
 969                 if (s == 1)
 970                         return 0;
 971                 h = strchr(h, ',') + 1;
 972                 n++;
 973         }
 974 }
 975
 976 static int message_stats_create(struct mapped_device *md,
 977                                 unsigned argc, char **argv,
 978                                 char *result, unsigned maxlen)
 979 {
 980         int r;
 981         int id;
 982         char dummy;
 983         unsigned long long start, end, len, step;
 984         unsigned divisor;
 985         const char *program_id, *aux_data;
 986         unsigned stat_flags = 0;
 987
 988         unsigned n_histogram_entries = 0;
 989         unsigned long long *histogram_boundaries = NULL;
 990
 991         struct dm_arg_set as, as_backup;
 992         const char *a;
 993         unsigned feature_args;
 994
 995         /*
 996          * Input format:
 997          *   <range> <step> [<extra_parameters> <parameters>] [<program_id> [<aux_data>]]
 998          */
 999
1000         if (argc < 3)
1001                 goto ret_einval;
1002
1003         as.argc = argc;
1004         as.argv = argv;
1005         dm_consume_args(&as, 1);
1006
1007         a = dm_shift_arg(&as);
1008         if (!strcmp(a, "-")) {
1009                 start = 0;
1010                 len = dm_get_size(md);
1011                 if (!len)
1012                         len = 1;
1013         } else if (sscanf(a, "%llu+%llu%c", &start, &len, &dummy) != 2 ||
1014                    start != (sector_t)start || len != (sector_t)len)
1015                 goto ret_einval;
1016
1017         end = start + len;
1018         if (start >= end)
1019                 goto ret_einval;
1020
1021         a = dm_shift_arg(&as);
1022         if (sscanf(a, "/%u%c", &divisor, &dummy) == 1) {
1023                 if (!divisor)
1024                         return -EINVAL;
1025                 step = end - start;
1026                 if (do_div(step, divisor))
1027                         step++;
1028                 if (!step)
1029                         step = 1;
1030         } else if (sscanf(a, "%llu%c", &step, &dummy) != 1 ||
1031                    step != (sector_t)step || !step)
1032                 goto ret_einval;
1033
1034         as_backup = as;
1035         a = dm_shift_arg(&as);
1036         if (a && sscanf(a, "%u%c", &feature_args, &dummy) == 1) {
1037                 while (feature_args--) {
1038                         a = dm_shift_arg(&as);
1039                         if (!a)
1040                                 goto ret_einval;
1041                         if (!strcasecmp(a, "precise_timestamps"))
1042                                 stat_flags |= STAT_PRECISE_TIMESTAMPS;
1043                         else if (!strncasecmp(a, "histogram:", 10)) {
1044                                 if (n_histogram_entries)
1045                                         goto ret_einval;
1046                                 if ((r = parse_histogram(a + 10, &n_histogram_entries, &histogram_boundaries)))
1047                                         goto ret;
1048                         } else
1049                                 goto ret_einval;
1050                 }
1051         } else {
1052                 as = as_backup;
1053         }
1054
1055         program_id = "-";
1056         aux_data = "-";
1057
1058         a = dm_shift_arg(&as);
1059         if (a)
1060                 program_id = a;
1061
1062         a = dm_shift_arg(&as);
1063         if (a)
1064                 aux_data = a;
1065
1066         if (as.argc)
1067                 goto ret_einval;
1068
1069         /*
1070          * If a buffer overflow happens after we created the region,
1071          * it's too late (the userspace would retry with a larger
1072          * buffer, but the region id that caused the overflow is already
1073          * leaked).  So we must detect buffer overflow in advance.
1074          */
1075         snprintf(result, maxlen, "%d", INT_MAX);
1076         if (dm_message_test_buffer_overflow(result, maxlen)) {
1077                 r = 1;
1078                 goto ret;
1079         }
1080
1081         id = dm_stats_create(dm_get_stats(md), start, end, step, stat_flags,
1082                              n_histogram_entries, histogram_boundaries, program_id, aux_data,
1083                              dm_internal_suspend_fast, dm_internal_resume_fast, md);
1084         if (id < 0) {
1085                 r = id;
1086                 goto ret;
1087         }
1088
1089         snprintf(result, maxlen, "%d", id);
1090
1091         r = 1;
1092         goto ret;
1093
1094 ret_einval:
1095         r = -EINVAL;
1096 ret:
1097         kfree(histogram_boundaries);
1098         return r;
1099 }
1100
1101 static int message_stats_delete(struct mapped_device *md,
1102                                 unsigned argc, char **argv)
1103 {
1104         int id;
1105         char dummy;
1106
1107         if (argc != 2)
1108                 return -EINVAL;
1109
1110         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1111                 return -EINVAL;
1112
1113         return dm_stats_delete(dm_get_stats(md), id);
1114 }
1115
1116 static int message_stats_clear(struct mapped_device *md,
1117                                unsigned argc, char **argv)
1118 {
1119         int id;
1120         char dummy;
1121
1122         if (argc != 2)
1123                 return -EINVAL;
1124
1125         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1126                 return -EINVAL;
1127
1128         return dm_stats_clear(dm_get_stats(md), id);
1129 }
1130
1131 static int message_stats_list(struct mapped_device *md,
1132                               unsigned argc, char **argv,
1133                               char *result, unsigned maxlen)
1134 {
1135         int r;
1136         const char *program = NULL;
1137
1138         if (argc < 1 || argc > 2)
1139                 return -EINVAL;
1140
1141         if (argc > 1) {
1142                 program = kstrdup(argv[1], GFP_KERNEL);
1143                 if (!program)
1144                         return -ENOMEM;
1145         }
1146
1147         r = dm_stats_list(dm_get_stats(md), program, result, maxlen);
1148
1149         kfree(program);
1150
1151         return r;
1152 }
1153
1154 static int message_stats_print(struct mapped_device *md,
1155                                unsigned argc, char **argv, bool clear,
1156                                char *result, unsigned maxlen)
1157 {
1158         int id;
1159         char dummy;
1160         unsigned long idx_start = 0, idx_len = ULONG_MAX;
1161
1162         if (argc != 2 && argc != 4)
1163                 return -EINVAL;
1164
1165         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1166                 return -EINVAL;
1167
1168         if (argc > 3) {
1169                 if (strcmp(argv[2], "-") &&
1170                     sscanf(argv[2], "%lu%c", &idx_start, &dummy) != 1)
1171                         return -EINVAL;
1172                 if (strcmp(argv[3], "-") &&
1173                     sscanf(argv[3], "%lu%c", &idx_len, &dummy) != 1)
1174                         return -EINVAL;
1175         }
1176
1177         return dm_stats_print(dm_get_stats(md), id, idx_start, idx_len, clear,
1178                               result, maxlen);
1179 }
1180
1181 static int message_stats_set_aux(struct mapped_device *md,
1182                                  unsigned argc, char **argv)
1183 {
1184         int id;
1185         char dummy;
1186
1187         if (argc != 3)
1188                 return -EINVAL;
1189
1190         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1191                 return -EINVAL;
1192
1193         return dm_stats_set_aux(dm_get_stats(md), id, argv[2]);
1194 }
1195
1196 int dm_stats_message(struct mapped_device *md, unsigned argc, char **argv,
1197                      char *result, unsigned maxlen)
1198 {
1199         int r;
1200
1201         /* All messages here must start with '@' */
1202         if (!strcasecmp(argv[0], "@stats_create"))
1203                 r = message_stats_create(md, argc, argv, result, maxlen);
1204         else if (!strcasecmp(argv[0], "@stats_delete"))
1205                 r = message_stats_delete(md, argc, argv);
1206         else if (!strcasecmp(argv[0], "@stats_clear"))
1207                 r = message_stats_clear(md, argc, argv);
1208         else if (!strcasecmp(argv[0], "@stats_list"))
1209                 r = message_stats_list(md, argc, argv, result, maxlen);
1210         else if (!strcasecmp(argv[0], "@stats_print"))
1211                 r = message_stats_print(md, argc, argv, false, result, maxlen);
1212         else if (!strcasecmp(argv[0], "@stats_print_clear"))
1213                 r = message_stats_print(md, argc, argv, true, result, maxlen);
1214         else if (!strcasecmp(argv[0], "@stats_set_aux"))
1215                 r = message_stats_set_aux(md, argc, argv);
1216         else
1217                 return 2; /* this wasn't a stats message */
1218
1219         if (r == -EINVAL)
1220                 DMWARN("Invalid parameters for message %s", argv[0]);
1221
1222         return r;
1223 }
1224
1225 int __init dm_statistics_init(void)
1226 {
1227         shared_memory_amount = 0;
1228         dm_stat_need_rcu_barrier = 0;
1229         return 0;
1230 }
1231
1232 void dm_statistics_exit(void)
1233 {
1234         if (dm_stat_need_rcu_barrier)
1235                 rcu_barrier();
1236         if (WARN_ON(shared_memory_amount))
1237                 DMCRIT("shared_memory_amount leaked: %lu", shared_memory_amount);
1238 }
1239
1240 module_param_named(stats_current_allocated_bytes, shared_memory_amount, ulong, S_IRUGO);
1241 MODULE_PARM_DESC(stats_current_allocated_bytes, "Memory currently used by statistics");