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