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