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