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 int dm_stats_init(struct dm_stats *stats)
 192 {
 193         int cpu;
 194         struct dm_stats_last_position *last;
 195
 196         mutex_init(&stats->mutex);
 197         INIT_LIST_HEAD(&stats->list);
 198         stats->last = alloc_percpu(struct dm_stats_last_position);
 199         if (!stats->last)
 200                 return -ENOMEM;
 201
 202         for_each_possible_cpu(cpu) {
 203                 last = per_cpu_ptr(stats->last, cpu);
 204                 last->last_sector = (sector_t)ULLONG_MAX;
 205                 last->last_rw = UINT_MAX;
 206         }
 207
 208         return 0;
 209 }
 210
 211 void dm_stats_cleanup(struct dm_stats *stats)
 212 {
 213         size_t ni;
 214         struct dm_stat *s;
 215         struct dm_stat_shared *shared;
 216
 217         while (!list_empty(&stats->list)) {
 218                 s = container_of(stats->list.next, struct dm_stat, list_entry);
 219                 list_del(&s->list_entry);
 220                 for (ni = 0; ni < s->n_entries; ni++) {
 221                         shared = &s->stat_shared[ni];
 222                         if (WARN_ON(dm_stat_in_flight(shared))) {
 223                                 DMCRIT("leaked in-flight counter at index %lu "
 224                                        "(start %llu, end %llu, step %llu): reads %d, writes %d",
 225                                        (unsigned long)ni,
 226                                        (unsigned long long)s->start,
 227                                        (unsigned long long)s->end,
 228                                        (unsigned long long)s->step,
 229                                        atomic_read(&shared->in_flight[READ]),
 230                                        atomic_read(&shared->in_flight[WRITE]));
 231                         }
 232                         cond_resched();
 233                 }
 234                 dm_stat_free(&s->rcu_head);
 235         }
 236         free_percpu(stats->last);
 237 }
 238
 239 static int dm_stats_create(struct dm_stats *stats, sector_t start, sector_t end,
 240                            sector_t step, unsigned stat_flags,
 241                            unsigned n_histogram_entries,
 242                            unsigned long long *histogram_boundaries,
 243                            const char *program_id, const char *aux_data,
 244                            void (*suspend_callback)(struct mapped_device *),
 245                            void (*resume_callback)(struct mapped_device *),
 246                            struct mapped_device *md)
 247 {
 248         struct list_head *l;
 249         struct dm_stat *s, *tmp_s;
 250         sector_t n_entries;
 251         size_t ni;
 252         size_t shared_alloc_size;
 253         size_t percpu_alloc_size;
 254         size_t histogram_alloc_size;
 255         struct dm_stat_percpu *p;
 256         int cpu;
 257         int ret_id;
 258         int r;
 259
 260         if (end < start || !step)
 261                 return -EINVAL;
 262
 263         n_entries = end - start;
 264         if (dm_sector_div64(n_entries, step))
 265                 n_entries++;
 266
 267         if (n_entries != (size_t)n_entries || !(size_t)(n_entries + 1))
 268                 return -EOVERFLOW;
 269
 270         shared_alloc_size = sizeof(struct dm_stat) + (size_t)n_entries * sizeof(struct dm_stat_shared);
 271         if ((shared_alloc_size - sizeof(struct dm_stat)) / sizeof(struct dm_stat_shared) != n_entries)
 272                 return -EOVERFLOW;
 273
 274         percpu_alloc_size = (size_t)n_entries * sizeof(struct dm_stat_percpu);
 275         if (percpu_alloc_size / sizeof(struct dm_stat_percpu) != n_entries)
 276                 return -EOVERFLOW;
 277
 278         histogram_alloc_size = (n_histogram_entries + 1) * (size_t)n_entries * sizeof(unsigned long long);
 279         if (histogram_alloc_size / (n_histogram_entries + 1) != (size_t)n_entries * sizeof(unsigned long long))
 280                 return -EOVERFLOW;
 281
 282         if (!check_shared_memory(shared_alloc_size + histogram_alloc_size +
 283                                  num_possible_cpus() * (percpu_alloc_size + histogram_alloc_size)))
 284                 return -ENOMEM;
 285
 286         s = dm_kvzalloc(shared_alloc_size, NUMA_NO_NODE);
 287         if (!s)
 288                 return -ENOMEM;
 289
 290         s->stat_flags = stat_flags;
 291         s->n_entries = n_entries;
 292         s->start = start;
 293         s->end = end;
 294         s->step = step;
 295         s->shared_alloc_size = shared_alloc_size;
 296         s->percpu_alloc_size = percpu_alloc_size;
 297         s->histogram_alloc_size = histogram_alloc_size;
 298
 299         s->n_histogram_entries = n_histogram_entries;
 300         s->histogram_boundaries = kmemdup(histogram_boundaries,
 301                                           s->n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
 302         if (!s->histogram_boundaries) {
 303                 r = -ENOMEM;
 304                 goto out;
 305         }
 306
 307         s->program_id = kstrdup(program_id, GFP_KERNEL);
 308         if (!s->program_id) {
 309                 r = -ENOMEM;
 310                 goto out;
 311         }
 312         s->aux_data = kstrdup(aux_data, GFP_KERNEL);
 313         if (!s->aux_data) {
 314                 r = -ENOMEM;
 315                 goto out;
 316         }
 317
 318         for (ni = 0; ni < n_entries; ni++) {
 319                 atomic_set(&s->stat_shared[ni].in_flight[READ], 0);
 320                 atomic_set(&s->stat_shared[ni].in_flight[WRITE], 0);
 321                 cond_resched();
 322         }
 323
 324         if (s->n_histogram_entries) {
 325                 unsigned long long *hi;
 326                 hi = dm_kvzalloc(s->histogram_alloc_size, NUMA_NO_NODE);
 327                 if (!hi) {
 328                         r = -ENOMEM;
 329                         goto out;
 330                 }
 331                 for (ni = 0; ni < n_entries; ni++) {
 332                         s->stat_shared[ni].tmp.histogram = hi;
 333                         hi += s->n_histogram_entries + 1;
 334                         cond_resched();
 335                 }
 336         }
 337
 338         for_each_possible_cpu(cpu) {
 339                 p = dm_kvzalloc(percpu_alloc_size, cpu_to_node(cpu));
 340                 if (!p) {
 341                         r = -ENOMEM;
 342                         goto out;
 343                 }
 344                 s->stat_percpu[cpu] = p;
 345                 if (s->n_histogram_entries) {
 346                         unsigned long long *hi;
 347                         hi = dm_kvzalloc(s->histogram_alloc_size, cpu_to_node(cpu));
 348                         if (!hi) {
 349                                 r = -ENOMEM;
 350                                 goto out;
 351                         }
 352                         for (ni = 0; ni < n_entries; ni++) {
 353                                 p[ni].histogram = hi;
 354                                 hi += s->n_histogram_entries + 1;
 355                                 cond_resched();
 356                         }
 357                 }
 358         }
 359
 360         /*
 361          * Suspend/resume to make sure there is no i/o in flight,
 362          * so that newly created statistics will be exact.
 363          *
 364          * (note: we couldn't suspend earlier because we must not
 365          * allocate memory while suspended)
 366          */
 367         suspend_callback(md);
 368
 369         mutex_lock(&stats->mutex);
 370         s->id = 0;
 371         list_for_each(l, &stats->list) {
 372                 tmp_s = container_of(l, struct dm_stat, list_entry);
 373                 if (WARN_ON(tmp_s->id < s->id)) {
 374                         r = -EINVAL;
 375                         goto out_unlock_resume;
 376                 }
 377                 if (tmp_s->id > s->id)
 378                         break;
 379                 if (unlikely(s->id == INT_MAX)) {
 380                         r = -ENFILE;
 381                         goto out_unlock_resume;
 382                 }
 383                 s->id++;
 384         }
 385         ret_id = s->id;
 386         list_add_tail_rcu(&s->list_entry, l);
 387         mutex_unlock(&stats->mutex);
 388
 389         resume_callback(md);
 390
 391         return ret_id;
 392
 393 out_unlock_resume:
 394         mutex_unlock(&stats->mutex);
 395         resume_callback(md);
 396 out:
 397         dm_stat_free(&s->rcu_head);
 398         return r;
 399 }
 400
 401 static struct dm_stat *__dm_stats_find(struct dm_stats *stats, int id)
 402 {
 403         struct dm_stat *s;
 404
 405         list_for_each_entry(s, &stats->list, list_entry) {
 406                 if (s->id > id)
 407                         break;
 408                 if (s->id == id)
 409                         return s;
 410         }
 411
 412         return NULL;
 413 }
 414
 415 static int dm_stats_delete(struct dm_stats *stats, int id)
 416 {
 417         struct dm_stat *s;
 418         int cpu;
 419
 420         mutex_lock(&stats->mutex);
 421
 422         s = __dm_stats_find(stats, id);
 423         if (!s) {
 424                 mutex_unlock(&stats->mutex);
 425                 return -ENOENT;
 426         }
 427
 428         list_del_rcu(&s->list_entry);
 429         mutex_unlock(&stats->mutex);
 430
 431         /*
 432          * vfree can't be called from RCU callback
 433          */
 434         for_each_possible_cpu(cpu)
 435                 if (is_vmalloc_addr(s->stat_percpu) ||
 436                     is_vmalloc_addr(s->stat_percpu[cpu][0].histogram))
 437                         goto do_sync_free;
 438         if (is_vmalloc_addr(s) ||
 439             is_vmalloc_addr(s->stat_shared[0].tmp.histogram)) {
 440 do_sync_free:
 441                 synchronize_rcu_expedited();
 442                 dm_stat_free(&s->rcu_head);
 443         } else {
 444                 ACCESS_ONCE(dm_stat_need_rcu_barrier) = 1;
 445                 call_rcu(&s->rcu_head, dm_stat_free);
 446         }
 447         return 0;
 448 }
 449
 450 static int dm_stats_list(struct dm_stats *stats, const char *program,
 451                          char *result, unsigned maxlen)
 452 {
 453         struct dm_stat *s;
 454         sector_t len;
 455         unsigned sz = 0;
 456
 457         /*
 458          * Output format:
 459          *   <region_id>: <start_sector>+<length> <step> <program_id> <aux_data>
 460          */
 461
 462         mutex_lock(&stats->mutex);
 463         list_for_each_entry(s, &stats->list, list_entry) {
 464                 if (!program || !strcmp(program, s->program_id)) {
 465                         len = s->end - s->start;
 466                         DMEMIT("%d: %llu+%llu %llu %s %s", s->id,
 467                                 (unsigned long long)s->start,
 468                                 (unsigned long long)len,
 469                                 (unsigned long long)s->step,
 470                                 s->program_id,
 471                                 s->aux_data);
 472                         if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
 473                                 DMEMIT(" precise_timestamps");
 474                         if (s->n_histogram_entries) {
 475                                 unsigned i;
 476                                 DMEMIT(" histogram:");
 477                                 for (i = 0; i < s->n_histogram_entries; i++) {
 478                                         if (i)
 479                                                 DMEMIT(",");
 480                                         DMEMIT("%llu", s->histogram_boundaries[i]);
 481                                 }
 482                         }
 483                         DMEMIT("\n");
 484                 }
 485                 cond_resched();
 486         }
 487         mutex_unlock(&stats->mutex);
 488
 489         return 1;
 490 }
 491
 492 static void dm_stat_round(struct dm_stat *s, struct dm_stat_shared *shared,
 493                           struct dm_stat_percpu *p)
 494 {
 495         /*
 496          * This is racy, but so is part_round_stats_single.
 497          */
 498         unsigned long long now, difference;
 499         unsigned in_flight_read, in_flight_write;
 500
 501         if (likely(!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)))
 502                 now = jiffies;
 503         else
 504                 now = ktime_to_ns(ktime_get());
 505
 506         difference = now - shared->stamp;
 507         if (!difference)
 508                 return;
 509
 510         in_flight_read = (unsigned)atomic_read(&shared->in_flight[READ]);
 511         in_flight_write = (unsigned)atomic_read(&shared->in_flight[WRITE]);
 512         if (in_flight_read)
 513                 p->io_ticks[READ] += difference;
 514         if (in_flight_write)
 515                 p->io_ticks[WRITE] += difference;
 516         if (in_flight_read + in_flight_write) {
 517                 p->io_ticks_total += difference;
 518                 p->time_in_queue += (in_flight_read + in_flight_write) * difference;
 519         }
 520         shared->stamp = now;
 521 }
 522
 523 static void dm_stat_for_entry(struct dm_stat *s, size_t entry,
 524                               int idx, sector_t len,
 525                               struct dm_stats_aux *stats_aux, bool end,
 526                               unsigned long duration_jiffies)
 527 {
 528         struct dm_stat_shared *shared = &s->stat_shared[entry];
 529         struct dm_stat_percpu *p;
 530
 531         /*
 532          * For strict correctness we should use local_irq_save/restore
 533          * instead of preempt_disable/enable.
 534          *
 535          * preempt_disable/enable is racy if the driver finishes bios
 536          * from non-interrupt context as well as from interrupt context
 537          * or from more different interrupts.
 538          *
 539          * On 64-bit architectures the race only results in not counting some
 540          * events, so it is acceptable.  On 32-bit architectures the race could
 541          * cause the counter going off by 2^32, so we need to do proper locking
 542          * there.
 543          *
 544          * part_stat_lock()/part_stat_unlock() have this race too.
 545          */
 546 #if BITS_PER_LONG == 32
 547         unsigned long flags;
 548         local_irq_save(flags);
 549 #else
 550         preempt_disable();
 551 #endif
 552         p = &s->stat_percpu[smp_processor_id()][entry];
 553
 554         if (!end) {
 555                 dm_stat_round(s, shared, p);
 556                 atomic_inc(&shared->in_flight[idx]);
 557         } else {
 558                 unsigned long long duration;
 559                 dm_stat_round(s, shared, p);
 560                 atomic_dec(&shared->in_flight[idx]);
 561                 p->sectors[idx] += len;
 562                 p->ios[idx] += 1;
 563                 p->merges[idx] += stats_aux->merged;
 564                 if (!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)) {
 565                         p->ticks[idx] += duration_jiffies;
 566                         duration = jiffies_to_msecs(duration_jiffies);
 567                 } else {
 568                         p->ticks[idx] += stats_aux->duration_ns;
 569                         duration = stats_aux->duration_ns;
 570                 }
 571                 if (s->n_histogram_entries) {
 572                         unsigned lo = 0, hi = s->n_histogram_entries + 1;
 573                         while (lo + 1 < hi) {
 574                                 unsigned mid = (lo + hi) / 2;
 575                                 if (s->histogram_boundaries[mid - 1] > duration) {
 576                                         hi = mid;
 577                                 } else {
 578                                         lo = mid;
 579                                 }
 580
 581                         }
 582                         p->histogram[lo]++;
 583                 }
 584         }
 585
 586 #if BITS_PER_LONG == 32
 587         local_irq_restore(flags);
 588 #else
 589         preempt_enable();
 590 #endif
 591 }
 592
 593 static void __dm_stat_bio(struct dm_stat *s, int bi_rw,
 594                           sector_t bi_sector, sector_t end_sector,
 595                           bool end, unsigned long duration_jiffies,
 596                           struct dm_stats_aux *stats_aux)
 597 {
 598         sector_t rel_sector, offset, todo, fragment_len;
 599         size_t entry;
 600
 601         if (end_sector <= s->start || bi_sector >= s->end)
 602                 return;
 603         if (unlikely(bi_sector < s->start)) {
 604                 rel_sector = 0;
 605                 todo = end_sector - s->start;
 606         } else {
 607                 rel_sector = bi_sector - s->start;
 608                 todo = end_sector - bi_sector;
 609         }
 610         if (unlikely(end_sector > s->end))
 611                 todo -= (end_sector - s->end);
 612
 613         offset = dm_sector_div64(rel_sector, s->step);
 614         entry = rel_sector;
 615         do {
 616                 if (WARN_ON_ONCE(entry >= s->n_entries)) {
 617                         DMCRIT("Invalid area access in region id %d", s->id);
 618                         return;
 619                 }
 620                 fragment_len = todo;
 621                 if (fragment_len > s->step - offset)
 622                         fragment_len = s->step - offset;
 623                 dm_stat_for_entry(s, entry, bi_rw, fragment_len,
 624                                   stats_aux, end, duration_jiffies);
 625                 todo -= fragment_len;
 626                 entry++;
 627                 offset = 0;
 628         } while (unlikely(todo != 0));
 629 }
 630
 631 void dm_stats_account_io(struct dm_stats *stats, unsigned long bi_rw,
 632                          sector_t bi_sector, unsigned bi_sectors, bool end,
 633                          unsigned long duration_jiffies,
 634                          struct dm_stats_aux *stats_aux)
 635 {
 636         struct dm_stat *s;
 637         sector_t end_sector;
 638         struct dm_stats_last_position *last;
 639         bool got_precise_time;
 640
 641         if (unlikely(!bi_sectors))
 642                 return;
 643
 644         end_sector = bi_sector + bi_sectors;
 645
 646         if (!end) {
 647                 /*
 648                  * A race condition can at worst result in the merged flag being
 649                  * misrepresented, so we don't have to disable preemption here.
 650                  */
 651                 last = raw_cpu_ptr(stats->last);
 652                 stats_aux->merged =
 653                         (bi_sector == (ACCESS_ONCE(last->last_sector) &&
 654                                        ((bi_rw == WRITE) ==
 655                                         (ACCESS_ONCE(last->last_rw) == WRITE))
 656                                        ));
 657                 ACCESS_ONCE(last->last_sector) = end_sector;
 658                 ACCESS_ONCE(last->last_rw) = bi_rw;
 659         }
 660
 661         rcu_read_lock();
 662
 663         got_precise_time = false;
 664         list_for_each_entry_rcu(s, &stats->list, list_entry) {
 665                 if (s->stat_flags & STAT_PRECISE_TIMESTAMPS && !got_precise_time) {
 666                         if (!end)
 667                                 stats_aux->duration_ns = ktime_to_ns(ktime_get());
 668                         else
 669                                 stats_aux->duration_ns = ktime_to_ns(ktime_get()) - stats_aux->duration_ns;
 670                         got_precise_time = true;
 671                 }
 672                 __dm_stat_bio(s, bi_rw, bi_sector, end_sector, end, duration_jiffies, stats_aux);
 673         }
 674
 675         rcu_read_unlock();
 676 }
 677
 678 static void __dm_stat_init_temporary_percpu_totals(struct dm_stat_shared *shared,
 679                                                    struct dm_stat *s, size_t x)
 680 {
 681         int cpu;
 682         struct dm_stat_percpu *p;
 683
 684         local_irq_disable();
 685         p = &s->stat_percpu[smp_processor_id()][x];
 686         dm_stat_round(s, shared, p);
 687         local_irq_enable();
 688
 689         shared->tmp.sectors[READ] = 0;
 690         shared->tmp.sectors[WRITE] = 0;
 691         shared->tmp.ios[READ] = 0;
 692         shared->tmp.ios[WRITE] = 0;
 693         shared->tmp.merges[READ] = 0;
 694         shared->tmp.merges[WRITE] = 0;
 695         shared->tmp.ticks[READ] = 0;
 696         shared->tmp.ticks[WRITE] = 0;
 697         shared->tmp.io_ticks[READ] = 0;
 698         shared->tmp.io_ticks[WRITE] = 0;
 699         shared->tmp.io_ticks_total = 0;
 700         shared->tmp.time_in_queue = 0;
 701
 702         if (s->n_histogram_entries)
 703                 memset(shared->tmp.histogram, 0, (s->n_histogram_entries + 1) * sizeof(unsigned long long));
 704
 705         for_each_possible_cpu(cpu) {
 706                 p = &s->stat_percpu[cpu][x];
 707                 shared->tmp.sectors[READ] += ACCESS_ONCE(p->sectors[READ]);
 708                 shared->tmp.sectors[WRITE] += ACCESS_ONCE(p->sectors[WRITE]);
 709                 shared->tmp.ios[READ] += ACCESS_ONCE(p->ios[READ]);
 710                 shared->tmp.ios[WRITE] += ACCESS_ONCE(p->ios[WRITE]);
 711                 shared->tmp.merges[READ] += ACCESS_ONCE(p->merges[READ]);
 712                 shared->tmp.merges[WRITE] += ACCESS_ONCE(p->merges[WRITE]);
 713                 shared->tmp.ticks[READ] += ACCESS_ONCE(p->ticks[READ]);
 714                 shared->tmp.ticks[WRITE] += ACCESS_ONCE(p->ticks[WRITE]);
 715                 shared->tmp.io_ticks[READ] += ACCESS_ONCE(p->io_ticks[READ]);
 716                 shared->tmp.io_ticks[WRITE] += ACCESS_ONCE(p->io_ticks[WRITE]);
 717                 shared->tmp.io_ticks_total += ACCESS_ONCE(p->io_ticks_total);
 718                 shared->tmp.time_in_queue += ACCESS_ONCE(p->time_in_queue);
 719                 if (s->n_histogram_entries) {
 720                         unsigned i;
 721                         for (i = 0; i < s->n_histogram_entries + 1; i++)
 722                                 shared->tmp.histogram[i] += ACCESS_ONCE(p->histogram[i]);
 723                 }
 724         }
 725 }
 726
 727 static void __dm_stat_clear(struct dm_stat *s, size_t idx_start, size_t idx_end,
 728                             bool init_tmp_percpu_totals)
 729 {
 730         size_t x;
 731         struct dm_stat_shared *shared;
 732         struct dm_stat_percpu *p;
 733
 734         for (x = idx_start; x < idx_end; x++) {
 735                 shared = &s->stat_shared[x];
 736                 if (init_tmp_percpu_totals)
 737                         __dm_stat_init_temporary_percpu_totals(shared, s, x);
 738                 local_irq_disable();
 739                 p = &s->stat_percpu[smp_processor_id()][x];
 740                 p->sectors[READ] -= shared->tmp.sectors[READ];
 741                 p->sectors[WRITE] -= shared->tmp.sectors[WRITE];
 742                 p->ios[READ] -= shared->tmp.ios[READ];
 743                 p->ios[WRITE] -= shared->tmp.ios[WRITE];
 744                 p->merges[READ] -= shared->tmp.merges[READ];
 745                 p->merges[WRITE] -= shared->tmp.merges[WRITE];
 746                 p->ticks[READ] -= shared->tmp.ticks[READ];
 747                 p->ticks[WRITE] -= shared->tmp.ticks[WRITE];
 748                 p->io_ticks[READ] -= shared->tmp.io_ticks[READ];
 749                 p->io_ticks[WRITE] -= shared->tmp.io_ticks[WRITE];
 750                 p->io_ticks_total -= shared->tmp.io_ticks_total;
 751                 p->time_in_queue -= shared->tmp.time_in_queue;
 752                 local_irq_enable();
 753                 if (s->n_histogram_entries) {
 754                         unsigned i;
 755                         for (i = 0; i < s->n_histogram_entries + 1; i++) {
 756                                 local_irq_disable();
 757                                 p = &s->stat_percpu[smp_processor_id()][x];
 758                                 p->histogram[i] -= shared->tmp.histogram[i];
 759                                 local_irq_enable();
 760                         }
 761                 }
 762                 cond_resched();
 763         }
 764 }
 765
 766 static int dm_stats_clear(struct dm_stats *stats, int id)
 767 {
 768         struct dm_stat *s;
 769
 770         mutex_lock(&stats->mutex);
 771
 772         s = __dm_stats_find(stats, id);
 773         if (!s) {
 774                 mutex_unlock(&stats->mutex);
 775                 return -ENOENT;
 776         }
 777
 778         __dm_stat_clear(s, 0, s->n_entries, true);
 779
 780         mutex_unlock(&stats->mutex);
 781
 782         return 1;
 783 }
 784
 785 /*
 786  * This is like jiffies_to_msec, but works for 64-bit values.
 787  */
 788 static unsigned long long dm_jiffies_to_msec64(struct dm_stat *s, unsigned long long j)
 789 {
 790         unsigned long long result;
 791         unsigned mult;
 792
 793         if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
 794                 return j;
 795
 796         result = 0;
 797         if (j)
 798                 result = jiffies_to_msecs(j & 0x3fffff);
 799         if (j >= 1 << 22) {
 800                 mult = jiffies_to_msecs(1 << 22);
 801                 result += (unsigned long long)mult * (unsigned long long)jiffies_to_msecs((j >> 22) & 0x3fffff);
 802         }
 803         if (j >= 1ULL << 44)
 804                 result += (unsigned long long)mult * (unsigned long long)mult * (unsigned long long)jiffies_to_msecs(j >> 44);
 805
 806         return result;
 807 }
 808
 809 static int dm_stats_print(struct dm_stats *stats, int id,
 810                           size_t idx_start, size_t idx_len,
 811                           bool clear, char *result, unsigned maxlen)
 812 {
 813         unsigned sz = 0;
 814         struct dm_stat *s;
 815         size_t x;
 816         sector_t start, end, step;
 817         size_t idx_end;
 818         struct dm_stat_shared *shared;
 819
 820         /*
 821          * Output format:
 822          *   <start_sector>+<length> counters
 823          */
 824
 825         mutex_lock(&stats->mutex);
 826
 827         s = __dm_stats_find(stats, id);
 828         if (!s) {
 829                 mutex_unlock(&stats->mutex);
 830                 return -ENOENT;
 831         }
 832
 833         idx_end = idx_start + idx_len;
 834         if (idx_end < idx_start ||
 835             idx_end > s->n_entries)
 836                 idx_end = s->n_entries;
 837
 838         if (idx_start > idx_end)
 839                 idx_start = idx_end;
 840
 841         step = s->step;
 842         start = s->start + (step * idx_start);
 843
 844         for (x = idx_start; x < idx_end; x++, start = end) {
 845                 shared = &s->stat_shared[x];
 846                 end = start + step;
 847                 if (unlikely(end > s->end))
 848                         end = s->end;
 849
 850                 __dm_stat_init_temporary_percpu_totals(shared, s, x);
 851
 852                 DMEMIT("%llu+%llu %llu %llu %llu %llu %llu %llu %llu %llu %d %llu %llu %llu %llu",
 853                        (unsigned long long)start,
 854                        (unsigned long long)step,
 855                        shared->tmp.ios[READ],
 856                        shared->tmp.merges[READ],
 857                        shared->tmp.sectors[READ],
 858                        dm_jiffies_to_msec64(s, shared->tmp.ticks[READ]),
 859                        shared->tmp.ios[WRITE],
 860                        shared->tmp.merges[WRITE],
 861                        shared->tmp.sectors[WRITE],
 862                        dm_jiffies_to_msec64(s, shared->tmp.ticks[WRITE]),
 863                        dm_stat_in_flight(shared),
 864                        dm_jiffies_to_msec64(s, shared->tmp.io_ticks_total),
 865                        dm_jiffies_to_msec64(s, shared->tmp.time_in_queue),
 866                        dm_jiffies_to_msec64(s, shared->tmp.io_ticks[READ]),
 867                        dm_jiffies_to_msec64(s, shared->tmp.io_ticks[WRITE]));
 868                 if (s->n_histogram_entries) {
 869                         unsigned i;
 870                         for (i = 0; i < s->n_histogram_entries + 1; i++) {
 871                                 DMEMIT("%s%llu", !i ? " " : ":", shared->tmp.histogram[i]);
 872                         }
 873                 }
 874                 DMEMIT("\n");
 875
 876                 if (unlikely(sz + 1 >= maxlen))
 877                         goto buffer_overflow;
 878
 879                 cond_resched();
 880         }
 881
 882         if (clear)
 883                 __dm_stat_clear(s, idx_start, idx_end, false);
 884
 885 buffer_overflow:
 886         mutex_unlock(&stats->mutex);
 887
 888         return 1;
 889 }
 890
 891 static int dm_stats_set_aux(struct dm_stats *stats, int id, const char *aux_data)
 892 {
 893         struct dm_stat *s;
 894         const char *new_aux_data;
 895
 896         mutex_lock(&stats->mutex);
 897
 898         s = __dm_stats_find(stats, id);
 899         if (!s) {
 900                 mutex_unlock(&stats->mutex);
 901                 return -ENOENT;
 902         }
 903
 904         new_aux_data = kstrdup(aux_data, GFP_KERNEL);
 905         if (!new_aux_data) {
 906                 mutex_unlock(&stats->mutex);
 907                 return -ENOMEM;
 908         }
 909
 910         kfree(s->aux_data);
 911         s->aux_data = new_aux_data;
 912
 913         mutex_unlock(&stats->mutex);
 914
 915         return 0;
 916 }
 917
 918 static int parse_histogram(const char *h, unsigned *n_histogram_entries,
 919                            unsigned long long **histogram_boundaries)
 920 {
 921         const char *q;
 922         unsigned n;
 923         unsigned long long last;
 924
 925         *n_histogram_entries = 1;
 926         for (q = h; *q; q++)
 927                 if (*q == ',')
 928                         (*n_histogram_entries)++;
 929
 930         *histogram_boundaries = kmalloc(*n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
 931         if (!*histogram_boundaries)
 932                 return -ENOMEM;
 933
 934         n = 0;
 935         last = 0;
 936         while (1) {
 937                 unsigned long long hi;
 938                 int s;
 939                 char ch;
 940                 s = sscanf(h, "%llu%c", &hi, &ch);
 941                 if (!s || (s == 2 && ch != ','))
 942                         return -EINVAL;
 943                 if (hi <= last)
 944                         return -EINVAL;
 945                 last = hi;
 946                 (*histogram_boundaries)[n] = hi;
 947                 if (s == 1)
 948                         return 0;
 949                 h = strchr(h, ',') + 1;
 950                 n++;
 951         }
 952 }
 953
 954 static int message_stats_create(struct mapped_device *md,
 955                                 unsigned argc, char **argv,
 956                                 char *result, unsigned maxlen)
 957 {
 958         int r;
 959         int id;
 960         char dummy;
 961         unsigned long long start, end, len, step;
 962         unsigned divisor;
 963         const char *program_id, *aux_data;
 964         unsigned stat_flags = 0;
 965
 966         unsigned n_histogram_entries = 0;
 967         unsigned long long *histogram_boundaries = NULL;
 968
 969         struct dm_arg_set as, as_backup;
 970         const char *a;
 971         unsigned feature_args;
 972
 973         /*
 974          * Input format:
 975          *   <range> <step> [<extra_parameters> <parameters>] [<program_id> [<aux_data>]]
 976          */
 977
 978         if (argc < 3)
 979                 goto ret_einval;
 980
 981         as.argc = argc;
 982         as.argv = argv;
 983         dm_consume_args(&as, 1);
 984
 985         a = dm_shift_arg(&as);
 986         if (!strcmp(a, "-")) {
 987                 start = 0;
 988                 len = dm_get_size(md);
 989                 if (!len)
 990                         len = 1;
 991         } else if (sscanf(a, "%llu+%llu%c", &start, &len, &dummy) != 2 ||
 992                    start != (sector_t)start || len != (sector_t)len)
 993                 goto ret_einval;
 994
 995         end = start + len;
 996         if (start >= end)
 997                 goto ret_einval;
 998
 999         a = dm_shift_arg(&as);
1000         if (sscanf(a, "/%u%c", &divisor, &dummy) == 1) {
1001                 if (!divisor)
1002                         return -EINVAL;
1003                 step = end - start;
1004                 if (do_div(step, divisor))
1005                         step++;
1006                 if (!step)
1007                         step = 1;
1008         } else if (sscanf(a, "%llu%c", &step, &dummy) != 1 ||
1009                    step != (sector_t)step || !step)
1010                 goto ret_einval;
1011
1012         as_backup = as;
1013         a = dm_shift_arg(&as);
1014         if (a && sscanf(a, "%u%c", &feature_args, &dummy) == 1) {
1015                 while (feature_args--) {
1016                         a = dm_shift_arg(&as);
1017                         if (!a)
1018                                 goto ret_einval;
1019                         if (!strcasecmp(a, "precise_timestamps"))
1020                                 stat_flags |= STAT_PRECISE_TIMESTAMPS;
1021                         else if (!strncasecmp(a, "histogram:", 10)) {
1022                                 if (n_histogram_entries)
1023                                         goto ret_einval;
1024                                 if ((r = parse_histogram(a + 10, &n_histogram_entries, &histogram_boundaries)))
1025                                         goto ret;
1026                         } else
1027                                 goto ret_einval;
1028                 }
1029         } else {
1030                 as = as_backup;
1031         }
1032
1033         program_id = "-";
1034         aux_data = "-";
1035
1036         a = dm_shift_arg(&as);
1037         if (a)
1038                 program_id = a;
1039
1040         a = dm_shift_arg(&as);
1041         if (a)
1042                 aux_data = a;
1043
1044         if (as.argc)
1045                 goto ret_einval;
1046
1047         /*
1048          * If a buffer overflow happens after we created the region,
1049          * it's too late (the userspace would retry with a larger
1050          * buffer, but the region id that caused the overflow is already
1051          * leaked).  So we must detect buffer overflow in advance.
1052          */
1053         snprintf(result, maxlen, "%d", INT_MAX);
1054         if (dm_message_test_buffer_overflow(result, maxlen)) {
1055                 r = 1;
1056                 goto ret;
1057         }
1058
1059         id = dm_stats_create(dm_get_stats(md), start, end, step, stat_flags,
1060                              n_histogram_entries, histogram_boundaries, program_id, aux_data,
1061                              dm_internal_suspend_fast, dm_internal_resume_fast, md);
1062         if (id < 0) {
1063                 r = id;
1064                 goto ret;
1065         }
1066
1067         snprintf(result, maxlen, "%d", id);
1068
1069         r = 1;
1070         goto ret;
1071
1072 ret_einval:
1073         r = -EINVAL;
1074 ret:
1075         kfree(histogram_boundaries);
1076         return r;
1077 }
1078
1079 static int message_stats_delete(struct mapped_device *md,
1080                                 unsigned argc, char **argv)
1081 {
1082         int id;
1083         char dummy;
1084
1085         if (argc != 2)
1086                 return -EINVAL;
1087
1088         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1089                 return -EINVAL;
1090
1091         return dm_stats_delete(dm_get_stats(md), id);
1092 }
1093
1094 static int message_stats_clear(struct mapped_device *md,
1095                                unsigned argc, char **argv)
1096 {
1097         int id;
1098         char dummy;
1099
1100         if (argc != 2)
1101                 return -EINVAL;
1102
1103         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1104                 return -EINVAL;
1105
1106         return dm_stats_clear(dm_get_stats(md), id);
1107 }
1108
1109 static int message_stats_list(struct mapped_device *md,
1110                               unsigned argc, char **argv,
1111                               char *result, unsigned maxlen)
1112 {
1113         int r;
1114         const char *program = NULL;
1115
1116         if (argc < 1 || argc > 2)
1117                 return -EINVAL;
1118
1119         if (argc > 1) {
1120                 program = kstrdup(argv[1], GFP_KERNEL);
1121                 if (!program)
1122                         return -ENOMEM;
1123         }
1124
1125         r = dm_stats_list(dm_get_stats(md), program, result, maxlen);
1126
1127         kfree(program);
1128
1129         return r;
1130 }
1131
1132 static int message_stats_print(struct mapped_device *md,
1133                                unsigned argc, char **argv, bool clear,
1134                                char *result, unsigned maxlen)
1135 {
1136         int id;
1137         char dummy;
1138         unsigned long idx_start = 0, idx_len = ULONG_MAX;
1139
1140         if (argc != 2 && argc != 4)
1141                 return -EINVAL;
1142
1143         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1144                 return -EINVAL;
1145
1146         if (argc > 3) {
1147                 if (strcmp(argv[2], "-") &&
1148                     sscanf(argv[2], "%lu%c", &idx_start, &dummy) != 1)
1149                         return -EINVAL;
1150                 if (strcmp(argv[3], "-") &&
1151                     sscanf(argv[3], "%lu%c", &idx_len, &dummy) != 1)
1152                         return -EINVAL;
1153         }
1154
1155         return dm_stats_print(dm_get_stats(md), id, idx_start, idx_len, clear,
1156                               result, maxlen);
1157 }
1158
1159 static int message_stats_set_aux(struct mapped_device *md,
1160                                  unsigned argc, char **argv)
1161 {
1162         int id;
1163         char dummy;
1164
1165         if (argc != 3)
1166                 return -EINVAL;
1167
1168         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1169                 return -EINVAL;
1170
1171         return dm_stats_set_aux(dm_get_stats(md), id, argv[2]);
1172 }
1173
1174 int dm_stats_message(struct mapped_device *md, unsigned argc, char **argv,
1175                      char *result, unsigned maxlen)
1176 {
1177         int r;
1178
1179         /* All messages here must start with '@' */
1180         if (!strcasecmp(argv[0], "@stats_create"))
1181                 r = message_stats_create(md, argc, argv, result, maxlen);
1182         else if (!strcasecmp(argv[0], "@stats_delete"))
1183                 r = message_stats_delete(md, argc, argv);
1184         else if (!strcasecmp(argv[0], "@stats_clear"))
1185                 r = message_stats_clear(md, argc, argv);
1186         else if (!strcasecmp(argv[0], "@stats_list"))
1187                 r = message_stats_list(md, argc, argv, result, maxlen);
1188         else if (!strcasecmp(argv[0], "@stats_print"))
1189                 r = message_stats_print(md, argc, argv, false, result, maxlen);
1190         else if (!strcasecmp(argv[0], "@stats_print_clear"))
1191                 r = message_stats_print(md, argc, argv, true, result, maxlen);
1192         else if (!strcasecmp(argv[0], "@stats_set_aux"))
1193                 r = message_stats_set_aux(md, argc, argv);
1194         else
1195                 return 2; /* this wasn't a stats message */
1196
1197         if (r == -EINVAL)
1198                 DMWARN("Invalid parameters for message %s", argv[0]);
1199
1200         return r;
1201 }
1202
1203 int __init dm_statistics_init(void)
1204 {
1205         shared_memory_amount = 0;
1206         dm_stat_need_rcu_barrier = 0;
1207         return 0;
1208 }
1209
1210 void dm_statistics_exit(void)
1211 {
1212         if (dm_stat_need_rcu_barrier)
1213                 rcu_barrier();
1214         if (WARN_ON(shared_memory_amount))
1215                 DMCRIT("shared_memory_amount leaked: %lu", shared_memory_amount);
1216 }
1217
1218 module_param_named(stats_current_allocated_bytes, shared_memory_amount, ulong, S_IRUGO);
1219 MODULE_PARM_DESC(stats_current_allocated_bytes, "Memory currently used by statistics");