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