drivers/md/dm-stats.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/errno.h>
   3 #include <linux/numa.h>
   4 #include <linux/slab.h>
   5 #include <linux/rculist.h>
   6 #include <linux/threads.h>
   7 #include <linux/preempt.h>
   8 #include <linux/irqflags.h>
   9 #include <linux/vmalloc.h>
  10 #include <linux/mm.h>
  11 #include <linux/module.h>
  12 #include <linux/device-mapper.h>
  13
  14 #include "dm-core.h"
  15 #include "dm-stats.h"
  16
  17 #define DM_MSG_PREFIX "stats"
  18
  19 static int dm_stat_need_rcu_barrier;
  20
  21 /*
  22  * Using 64-bit values to avoid overflow (which is a
  23  * problem that block/genhd.c's IO accounting has).
  24  */
  25 struct dm_stat_percpu {
  26         unsigned long long sectors[2];
  27         unsigned long long ios[2];
  28         unsigned long long merges[2];
  29         unsigned long long ticks[2];
  30         unsigned long long io_ticks[2];
  31         unsigned long long io_ticks_total;
  32         unsigned long long time_in_queue;
  33         unsigned long long *histogram;
  34 };
  35
  36 struct dm_stat_shared {
  37         atomic_t in_flight[2];
  38         unsigned long long stamp;
  39         struct dm_stat_percpu tmp;
  40 };
  41
  42 struct dm_stat {
  43         struct list_head list_entry;
  44         int id;
  45         unsigned stat_flags;
  46         size_t n_entries;
  47         sector_t start;
  48         sector_t end;
  49         sector_t step;
  50         unsigned n_histogram_entries;
  51         unsigned long long *histogram_boundaries;
  52         const char *program_id;
  53         const char *aux_data;
  54         struct rcu_head rcu_head;
  55         size_t shared_alloc_size;
  56         size_t percpu_alloc_size;
  57         size_t histogram_alloc_size;
  58         struct dm_stat_percpu *stat_percpu[NR_CPUS];
  59         struct dm_stat_shared stat_shared[0];
  60 };
  61
  62 #define STAT_PRECISE_TIMESTAMPS         1
  63
  64 struct dm_stats_last_position {
  65         sector_t last_sector;
  66         unsigned last_rw;
  67 };
  68
  69 /*
  70  * A typo on the command line could possibly make the kernel run out of memory
  71  * and crash. To prevent the crash we account all used memory. We fail if we
  72  * exhaust 1/4 of all memory or 1/2 of vmalloc space.
  73  */
  74 #define DM_STATS_MEMORY_FACTOR          4
  75 #define DM_STATS_VMALLOC_FACTOR         2
  76
  77 static DEFINE_SPINLOCK(shared_memory_lock);
  78
  79 static unsigned long shared_memory_amount;
  80
  81 static bool __check_shared_memory(size_t alloc_size)
  82 {
  83         size_t a;
  84
  85         a = shared_memory_amount + alloc_size;
  86         if (a < shared_memory_amount)
  87                 return false;
  88         if (a >> PAGE_SHIFT > totalram_pages / DM_STATS_MEMORY_FACTOR)
  89                 return false;
  90 #ifdef CONFIG_MMU
  91         if (a > (VMALLOC_END - VMALLOC_START) / DM_STATS_VMALLOC_FACTOR)
  92                 return false;
  93 #endif
  94         return true;
  95 }
  96
  97 static bool check_shared_memory(size_t alloc_size)
  98 {
  99         bool ret;
 100
 101         spin_lock_irq(&shared_memory_lock);
 102
 103         ret = __check_shared_memory(alloc_size);
 104
 105         spin_unlock_irq(&shared_memory_lock);
 106
 107         return ret;
 108 }
 109
 110 static bool claim_shared_memory(size_t alloc_size)
 111 {
 112         spin_lock_irq(&shared_memory_lock);
 113
 114         if (!__check_shared_memory(alloc_size)) {
 115                 spin_unlock_irq(&shared_memory_lock);
 116                 return false;
 117         }
 118
 119         shared_memory_amount += alloc_size;
 120
 121         spin_unlock_irq(&shared_memory_lock);
 122
 123         return true;
 124 }
 125
 126 static void free_shared_memory(size_t alloc_size)
 127 {
 128         unsigned long flags;
 129
 130         spin_lock_irqsave(&shared_memory_lock, flags);
 131
 132         if (WARN_ON_ONCE(shared_memory_amount < alloc_size)) {
 133                 spin_unlock_irqrestore(&shared_memory_lock, flags);
 134                 DMCRIT("Memory usage accounting bug.");
 135                 return;
 136         }
 137
 138         shared_memory_amount -= alloc_size;
 139
 140         spin_unlock_irqrestore(&shared_memory_lock, flags);
 141 }
 142
 143 static void *dm_kvzalloc(size_t alloc_size, int node)
 144 {
 145         void *p;
 146
 147         if (!claim_shared_memory(alloc_size))
 148                 return NULL;
 149
 150         p = kvzalloc_node(alloc_size, GFP_KERNEL | __GFP_NOMEMALLOC, node);
 151         if (p)
 152                 return p;
 153
 154         free_shared_memory(alloc_size);
 155
 156         return NULL;
 157 }
 158
 159 static void dm_kvfree(void *ptr, size_t alloc_size)
 160 {
 161         if (!ptr)
 162                 return;
 163
 164         free_shared_memory(alloc_size);
 165
 166         kvfree(ptr);
 167 }
 168
 169 static void dm_stat_free(struct rcu_head *head)
 170 {
 171         int cpu;
 172         struct dm_stat *s = container_of(head, struct dm_stat, rcu_head);
 173
 174         kfree(s->histogram_boundaries);
 175         kfree(s->program_id);
 176         kfree(s->aux_data);
 177         for_each_possible_cpu(cpu) {
 178                 dm_kvfree(s->stat_percpu[cpu][0].histogram, s->histogram_alloc_size);
 179                 dm_kvfree(s->stat_percpu[cpu], s->percpu_alloc_size);
 180         }
 181         dm_kvfree(s->stat_shared[0].tmp.histogram, s->histogram_alloc_size);
 182         dm_kvfree(s, s->shared_alloc_size);
 183 }
 184
 185 static int dm_stat_in_flight(struct dm_stat_shared *shared)
 186 {
 187         return atomic_read(&shared->in_flight[READ]) +
 188                atomic_read(&shared->in_flight[WRITE]);
 189 }
 190
 191 void dm_stats_init(struct dm_stats *stats)
 192 {
 193         int cpu;
 194         struct dm_stats_last_position *last;
 195
 196         mutex_init(&stats->mutex);
 197         INIT_LIST_HEAD(&stats->list);
 198         stats->last = alloc_percpu(struct dm_stats_last_position);
 199         for_each_possible_cpu(cpu) {
 200                 last = per_cpu_ptr(stats->last, cpu);
 201                 last->last_sector = (sector_t)ULLONG_MAX;
 202                 last->last_rw = UINT_MAX;
 203         }
 204 }
 205
 206 void dm_stats_cleanup(struct dm_stats *stats)
 207 {
 208         size_t ni;
 209         struct dm_stat *s;
 210         struct dm_stat_shared *shared;
 211
 212         while (!list_empty(&stats->list)) {
 213                 s = container_of(stats->list.next, struct dm_stat, list_entry);
 214                 list_del(&s->list_entry);
 215                 for (ni = 0; ni < s->n_entries; ni++) {
 216                         shared = &s->stat_shared[ni];
 217                         if (WARN_ON(dm_stat_in_flight(shared))) {
 218                                 DMCRIT("leaked in-flight counter at index %lu "
 219                                        "(start %llu, end %llu, step %llu): reads %d, writes %d",
 220                                        (unsigned long)ni,
 221                                        (unsigned long long)s->start,
 222                                        (unsigned long long)s->end,
 223                                        (unsigned long long)s->step,
 224                                        atomic_read(&shared->in_flight[READ]),
 225                                        atomic_read(&shared->in_flight[WRITE]));
 226                         }
 227                         cond_resched();
 228                 }
 229                 dm_stat_free(&s->rcu_head);
 230         }
 231         free_percpu(stats->last);
 232 }
 233
 234 static int dm_stats_create(struct dm_stats *stats, sector_t start, sector_t end,
 235                            sector_t step, unsigned stat_flags,
 236                            unsigned n_histogram_entries,
 237                            unsigned long long *histogram_boundaries,
 238                            const char *program_id, const char *aux_data,
 239                            void (*suspend_callback)(struct mapped_device *),
 240                            void (*resume_callback)(struct mapped_device *),
 241                            struct mapped_device *md)
 242 {
 243         struct list_head *l;
 244         struct dm_stat *s, *tmp_s;
 245         sector_t n_entries;
 246         size_t ni;
 247         size_t shared_alloc_size;
 248         size_t percpu_alloc_size;
 249         size_t histogram_alloc_size;
 250         struct dm_stat_percpu *p;
 251         int cpu;
 252         int ret_id;
 253         int r;
 254
 255         if (end < start || !step)
 256                 return -EINVAL;
 257
 258         n_entries = end - start;
 259         if (dm_sector_div64(n_entries, step))
 260                 n_entries++;
 261
 262         if (n_entries != (size_t)n_entries || !(size_t)(n_entries + 1))
 263                 return -EOVERFLOW;
 264
 265         shared_alloc_size = sizeof(struct dm_stat) + (size_t)n_entries * sizeof(struct dm_stat_shared);
 266         if ((shared_alloc_size - sizeof(struct dm_stat)) / sizeof(struct dm_stat_shared) != n_entries)
 267                 return -EOVERFLOW;
 268
 269         percpu_alloc_size = (size_t)n_entries * sizeof(struct dm_stat_percpu);
 270         if (percpu_alloc_size / sizeof(struct dm_stat_percpu) != n_entries)
 271                 return -EOVERFLOW;
 272
 273         histogram_alloc_size = (n_histogram_entries + 1) * (size_t)n_entries * sizeof(unsigned long long);
 274         if (histogram_alloc_size / (n_histogram_entries + 1) != (size_t)n_entries * sizeof(unsigned long long))
 275                 return -EOVERFLOW;
 276
 277         if (!check_shared_memory(shared_alloc_size + histogram_alloc_size +
 278                                  num_possible_cpus() * (percpu_alloc_size + histogram_alloc_size)))
 279                 return -ENOMEM;
 280
 281         s = dm_kvzalloc(shared_alloc_size, NUMA_NO_NODE);
 282         if (!s)
 283                 return -ENOMEM;
 284
 285         s->stat_flags = stat_flags;
 286         s->n_entries = n_entries;
 287         s->start = start;
 288         s->end = end;
 289         s->step = step;
 290         s->shared_alloc_size = shared_alloc_size;
 291         s->percpu_alloc_size = percpu_alloc_size;
 292         s->histogram_alloc_size = histogram_alloc_size;
 293
 294         s->n_histogram_entries = n_histogram_entries;
 295         s->histogram_boundaries = kmemdup(histogram_boundaries,
 296                                           s->n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
 297         if (!s->histogram_boundaries) {
 298                 r = -ENOMEM;
 299                 goto out;
 300         }
 301
 302         s->program_id = kstrdup(program_id, GFP_KERNEL);
 303         if (!s->program_id) {
 304                 r = -ENOMEM;
 305                 goto out;
 306         }
 307         s->aux_data = kstrdup(aux_data, GFP_KERNEL);
 308         if (!s->aux_data) {
 309                 r = -ENOMEM;
 310                 goto out;
 311         }
 312
 313         for (ni = 0; ni < n_entries; ni++) {
 314                 atomic_set(&s->stat_shared[ni].in_flight[READ], 0);
 315                 atomic_set(&s->stat_shared[ni].in_flight[WRITE], 0);
 316                 cond_resched();
 317         }
 318
 319         if (s->n_histogram_entries) {
 320                 unsigned long long *hi;
 321                 hi = dm_kvzalloc(s->histogram_alloc_size, NUMA_NO_NODE);
 322                 if (!hi) {
 323                         r = -ENOMEM;
 324                         goto out;
 325                 }
 326                 for (ni = 0; ni < n_entries; ni++) {
 327                         s->stat_shared[ni].tmp.histogram = hi;
 328                         hi += s->n_histogram_entries + 1;
 329                         cond_resched();
 330                 }
 331         }
 332
 333         for_each_possible_cpu(cpu) {
 334                 p = dm_kvzalloc(percpu_alloc_size, cpu_to_node(cpu));
 335                 if (!p) {
 336                         r = -ENOMEM;
 337                         goto out;
 338                 }
 339                 s->stat_percpu[cpu] = p;
 340                 if (s->n_histogram_entries) {
 341                         unsigned long long *hi;
 342                         hi = dm_kvzalloc(s->histogram_alloc_size, cpu_to_node(cpu));
 343                         if (!hi) {
 344                                 r = -ENOMEM;
 345                                 goto out;
 346                         }
 347                         for (ni = 0; ni < n_entries; ni++) {
 348                                 p[ni].histogram = hi;
 349                                 hi += s->n_histogram_entries + 1;
 350                                 cond_resched();
 351                         }
 352                 }
 353         }
 354
 355         /*
 356          * Suspend/resume to make sure there is no i/o in flight,
 357          * so that newly created statistics will be exact.
 358          *
 359          * (note: we couldn't suspend earlier because we must not
 360          * allocate memory while suspended)
 361          */
 362         suspend_callback(md);
 363
 364         mutex_lock(&stats->mutex);
 365         s->id = 0;
 366         list_for_each(l, &stats->list) {
 367                 tmp_s = container_of(l, struct dm_stat, list_entry);
 368                 if (WARN_ON(tmp_s->id < s->id)) {
 369                         r = -EINVAL;
 370                         goto out_unlock_resume;
 371                 }
 372                 if (tmp_s->id > s->id)
 373                         break;
 374                 if (unlikely(s->id == INT_MAX)) {
 375                         r = -ENFILE;
 376                         goto out_unlock_resume;
 377                 }
 378                 s->id++;
 379         }
 380         ret_id = s->id;
 381         list_add_tail_rcu(&s->list_entry, l);
 382         mutex_unlock(&stats->mutex);
 383
 384         resume_callback(md);
 385
 386         return ret_id;
 387
 388 out_unlock_resume:
 389         mutex_unlock(&stats->mutex);
 390         resume_callback(md);
 391 out:
 392         dm_stat_free(&s->rcu_head);
 393         return r;
 394 }
 395
 396 static struct dm_stat *__dm_stats_find(struct dm_stats *stats, int id)
 397 {
 398         struct dm_stat *s;
 399
 400         list_for_each_entry(s, &stats->list, list_entry) {
 401                 if (s->id > id)
 402                         break;
 403                 if (s->id == id)
 404                         return s;
 405         }
 406
 407         return NULL;
 408 }
 409
 410 static int dm_stats_delete(struct dm_stats *stats, int id)
 411 {
 412         struct dm_stat *s;
 413         int cpu;
 414
 415         mutex_lock(&stats->mutex);
 416
 417         s = __dm_stats_find(stats, id);
 418         if (!s) {
 419                 mutex_unlock(&stats->mutex);
 420                 return -ENOENT;
 421         }
 422
 423         list_del_rcu(&s->list_entry);
 424         mutex_unlock(&stats->mutex);
 425
 426         /*
 427          * vfree can't be called from RCU callback
 428          */
 429         for_each_possible_cpu(cpu)
 430                 if (is_vmalloc_addr(s->stat_percpu) ||
 431                     is_vmalloc_addr(s->stat_percpu[cpu][0].histogram))
 432                         goto do_sync_free;
 433         if (is_vmalloc_addr(s) ||
 434             is_vmalloc_addr(s->stat_shared[0].tmp.histogram)) {
 435 do_sync_free:
 436                 synchronize_rcu_expedited();
 437                 dm_stat_free(&s->rcu_head);
 438         } else {
 439                 ACCESS_ONCE(dm_stat_need_rcu_barrier) = 1;
 440                 call_rcu(&s->rcu_head, dm_stat_free);
 441         }
 442         return 0;
 443 }
 444
 445 static int dm_stats_list(struct dm_stats *stats, const char *program,
 446                          char *result, unsigned maxlen)
 447 {
 448         struct dm_stat *s;
 449         sector_t len;
 450         unsigned sz = 0;
 451
 452         /*
 453          * Output format:
 454          *   <region_id>: <start_sector>+<length> <step> <program_id> <aux_data>
 455          */
 456
 457         mutex_lock(&stats->mutex);
 458         list_for_each_entry(s, &stats->list, list_entry) {
 459                 if (!program || !strcmp(program, s->program_id)) {
 460                         len = s->end - s->start;
 461                         DMEMIT("%d: %llu+%llu %llu %s %s", s->id,
 462                                 (unsigned long long)s->start,
 463                                 (unsigned long long)len,
 464                                 (unsigned long long)s->step,
 465                                 s->program_id,
 466                                 s->aux_data);
 467                         if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
 468                                 DMEMIT(" precise_timestamps");
 469                         if (s->n_histogram_entries) {
 470                                 unsigned i;
 471                                 DMEMIT(" histogram:");
 472                                 for (i = 0; i < s->n_histogram_entries; i++) {
 473                                         if (i)
 474                                                 DMEMIT(",");
 475                                         DMEMIT("%llu", s->histogram_boundaries[i]);
 476                                 }
 477                         }
 478                         DMEMIT("\n");
 479                 }
 480                 cond_resched();
 481         }
 482         mutex_unlock(&stats->mutex);
 483
 484         return 1;
 485 }
 486
 487 static void dm_stat_round(struct dm_stat *s, struct dm_stat_shared *shared,
 488                           struct dm_stat_percpu *p)
 489 {
 490         /*
 491          * This is racy, but so is part_round_stats_single.
 492          */
 493         unsigned long long now, difference;
 494         unsigned in_flight_read, in_flight_write;
 495
 496         if (likely(!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)))
 497                 now = jiffies;
 498         else
 499                 now = ktime_to_ns(ktime_get());
 500
 501         difference = now - shared->stamp;
 502         if (!difference)
 503                 return;
 504
 505         in_flight_read = (unsigned)atomic_read(&shared->in_flight[READ]);
 506         in_flight_write = (unsigned)atomic_read(&shared->in_flight[WRITE]);
 507         if (in_flight_read)
 508                 p->io_ticks[READ] += difference;
 509         if (in_flight_write)
 510                 p->io_ticks[WRITE] += difference;
 511         if (in_flight_read + in_flight_write) {
 512                 p->io_ticks_total += difference;
 513                 p->time_in_queue += (in_flight_read + in_flight_write) * difference;
 514         }
 515         shared->stamp = now;
 516 }
 517
 518 static void dm_stat_for_entry(struct dm_stat *s, size_t entry,
 519                               int idx, sector_t len,
 520                               struct dm_stats_aux *stats_aux, bool end,
 521                               unsigned long duration_jiffies)
 522 {
 523         struct dm_stat_shared *shared = &s->stat_shared[entry];
 524         struct dm_stat_percpu *p;
 525
 526         /*
 527          * For strict correctness we should use local_irq_save/restore
 528          * instead of preempt_disable/enable.
 529          *
 530          * preempt_disable/enable is racy if the driver finishes bios
 531          * from non-interrupt context as well as from interrupt context
 532          * or from more different interrupts.
 533          *
 534          * On 64-bit architectures the race only results in not counting some
 535          * events, so it is acceptable.  On 32-bit architectures the race could
 536          * cause the counter going off by 2^32, so we need to do proper locking
 537          * there.
 538          *
 539          * part_stat_lock()/part_stat_unlock() have this race too.
 540          */
 541 #if BITS_PER_LONG == 32
 542         unsigned long flags;
 543         local_irq_save(flags);
 544 #else
 545         preempt_disable();
 546 #endif
 547         p = &s->stat_percpu[smp_processor_id()][entry];
 548
 549         if (!end) {
 550                 dm_stat_round(s, shared, p);
 551                 atomic_inc(&shared->in_flight[idx]);
 552         } else {
 553                 unsigned long long duration;
 554                 dm_stat_round(s, shared, p);
 555                 atomic_dec(&shared->in_flight[idx]);
 556                 p->sectors[idx] += len;
 557                 p->ios[idx] += 1;
 558                 p->merges[idx] += stats_aux->merged;
 559                 if (!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)) {
 560                         p->ticks[idx] += duration_jiffies;
 561                         duration = jiffies_to_msecs(duration_jiffies);
 562                 } else {
 563                         p->ticks[idx] += stats_aux->duration_ns;
 564                         duration = stats_aux->duration_ns;
 565                 }
 566                 if (s->n_histogram_entries) {
 567                         unsigned lo = 0, hi = s->n_histogram_entries + 1;
 568                         while (lo + 1 < hi) {
 569                                 unsigned mid = (lo + hi) / 2;
 570                                 if (s->histogram_boundaries[mid - 1] > duration) {
 571                                         hi = mid;
 572                                 } else {
 573                                         lo = mid;
 574                                 }
 575
 576                         }
 577                         p->histogram[lo]++;
 578                 }
 579         }
 580
 581 #if BITS_PER_LONG == 32
 582         local_irq_restore(flags);
 583 #else
 584         preempt_enable();
 585 #endif
 586 }
 587
 588 static void __dm_stat_bio(struct dm_stat *s, int bi_rw,
 589                           sector_t bi_sector, sector_t end_sector,
 590                           bool end, unsigned long duration_jiffies,
 591                           struct dm_stats_aux *stats_aux)
 592 {
 593         sector_t rel_sector, offset, todo, fragment_len;
 594         size_t entry;
 595
 596         if (end_sector <= s->start || bi_sector >= s->end)
 597                 return;
 598         if (unlikely(bi_sector < s->start)) {
 599                 rel_sector = 0;
 600                 todo = end_sector - s->start;
 601         } else {
 602                 rel_sector = bi_sector - s->start;
 603                 todo = end_sector - bi_sector;
 604         }
 605         if (unlikely(end_sector > s->end))
 606                 todo -= (end_sector - s->end);
 607
 608         offset = dm_sector_div64(rel_sector, s->step);
 609         entry = rel_sector;
 610         do {
 611                 if (WARN_ON_ONCE(entry >= s->n_entries)) {
 612                         DMCRIT("Invalid area access in region id %d", s->id);
 613                         return;
 614                 }
 615                 fragment_len = todo;
 616                 if (fragment_len > s->step - offset)
 617                         fragment_len = s->step - offset;
 618                 dm_stat_for_entry(s, entry, bi_rw, fragment_len,
 619                                   stats_aux, end, duration_jiffies);
 620                 todo -= fragment_len;
 621                 entry++;
 622                 offset = 0;
 623         } while (unlikely(todo != 0));
 624 }
 625
 626 void dm_stats_account_io(struct dm_stats *stats, unsigned long bi_rw,
 627                          sector_t bi_sector, unsigned bi_sectors, bool end,
 628                          unsigned long duration_jiffies,
 629                          struct dm_stats_aux *stats_aux)
 630 {
 631         struct dm_stat *s;
 632         sector_t end_sector;
 633         struct dm_stats_last_position *last;
 634         bool got_precise_time;
 635
 636         if (unlikely(!bi_sectors))
 637                 return;
 638
 639         end_sector = bi_sector + bi_sectors;
 640
 641         if (!end) {
 642                 /*
 643                  * A race condition can at worst result in the merged flag being
 644                  * misrepresented, so we don't have to disable preemption here.
 645                  */
 646                 last = raw_cpu_ptr(stats->last);
 647                 stats_aux->merged =
 648                         (bi_sector == (ACCESS_ONCE(last->last_sector) &&
 649                                        ((bi_rw == WRITE) ==
 650                                         (ACCESS_ONCE(last->last_rw) == WRITE))
 651                                        ));
 652                 ACCESS_ONCE(last->last_sector) = end_sector;
 653                 ACCESS_ONCE(last->last_rw) = bi_rw;
 654         }
 655
 656         rcu_read_lock();
 657
 658         got_precise_time = false;
 659         list_for_each_entry_rcu(s, &stats->list, list_entry) {
 660                 if (s->stat_flags & STAT_PRECISE_TIMESTAMPS && !got_precise_time) {
 661                         if (!end)
 662                                 stats_aux->duration_ns = ktime_to_ns(ktime_get());
 663                         else
 664                                 stats_aux->duration_ns = ktime_to_ns(ktime_get()) - stats_aux->duration_ns;
 665                         got_precise_time = true;
 666                 }
 667                 __dm_stat_bio(s, bi_rw, bi_sector, end_sector, end, duration_jiffies, stats_aux);
 668         }
 669
 670         rcu_read_unlock();
 671 }
 672
 673 static void __dm_stat_init_temporary_percpu_totals(struct dm_stat_shared *shared,
 674                                                    struct dm_stat *s, size_t x)
 675 {
 676         int cpu;
 677         struct dm_stat_percpu *p;
 678
 679         local_irq_disable();
 680         p = &s->stat_percpu[smp_processor_id()][x];
 681         dm_stat_round(s, shared, p);
 682         local_irq_enable();
 683
 684         shared->tmp.sectors[READ] = 0;
 685         shared->tmp.sectors[WRITE] = 0;
 686         shared->tmp.ios[READ] = 0;
 687         shared->tmp.ios[WRITE] = 0;
 688         shared->tmp.merges[READ] = 0;
 689         shared->tmp.merges[WRITE] = 0;
 690         shared->tmp.ticks[READ] = 0;
 691         shared->tmp.ticks[WRITE] = 0;
 692         shared->tmp.io_ticks[READ] = 0;
 693         shared->tmp.io_ticks[WRITE] = 0;
 694         shared->tmp.io_ticks_total = 0;
 695         shared->tmp.time_in_queue = 0;
 696
 697         if (s->n_histogram_entries)
 698                 memset(shared->tmp.histogram, 0, (s->n_histogram_entries + 1) * sizeof(unsigned long long));
 699
 700         for_each_possible_cpu(cpu) {
 701                 p = &s->stat_percpu[cpu][x];
 702                 shared->tmp.sectors[READ] += ACCESS_ONCE(p->sectors[READ]);
 703                 shared->tmp.sectors[WRITE] += ACCESS_ONCE(p->sectors[WRITE]);
 704                 shared->tmp.ios[READ] += ACCESS_ONCE(p->ios[READ]);
 705                 shared->tmp.ios[WRITE] += ACCESS_ONCE(p->ios[WRITE]);
 706                 shared->tmp.merges[READ] += ACCESS_ONCE(p->merges[READ]);
 707                 shared->tmp.merges[WRITE] += ACCESS_ONCE(p->merges[WRITE]);
 708                 shared->tmp.ticks[READ] += ACCESS_ONCE(p->ticks[READ]);
 709                 shared->tmp.ticks[WRITE] += ACCESS_ONCE(p->ticks[WRITE]);
 710                 shared->tmp.io_ticks[READ] += ACCESS_ONCE(p->io_ticks[READ]);
 711                 shared->tmp.io_ticks[WRITE] += ACCESS_ONCE(p->io_ticks[WRITE]);
 712                 shared->tmp.io_ticks_total += ACCESS_ONCE(p->io_ticks_total);
 713                 shared->tmp.time_in_queue += ACCESS_ONCE(p->time_in_queue);
 714                 if (s->n_histogram_entries) {
 715                         unsigned i;
 716                         for (i = 0; i < s->n_histogram_entries + 1; i++)
 717                                 shared->tmp.histogram[i] += ACCESS_ONCE(p->histogram[i]);
 718                 }
 719         }
 720 }
 721
 722 static void __dm_stat_clear(struct dm_stat *s, size_t idx_start, size_t idx_end,
 723                             bool init_tmp_percpu_totals)
 724 {
 725         size_t x;
 726         struct dm_stat_shared *shared;
 727         struct dm_stat_percpu *p;
 728
 729         for (x = idx_start; x < idx_end; x++) {
 730                 shared = &s->stat_shared[x];
 731                 if (init_tmp_percpu_totals)
 732                         __dm_stat_init_temporary_percpu_totals(shared, s, x);
 733                 local_irq_disable();
 734                 p = &s->stat_percpu[smp_processor_id()][x];
 735                 p->sectors[READ] -= shared->tmp.sectors[READ];
 736                 p->sectors[WRITE] -= shared->tmp.sectors[WRITE];
 737                 p->ios[READ] -= shared->tmp.ios[READ];
 738                 p->ios[WRITE] -= shared->tmp.ios[WRITE];
 739                 p->merges[READ] -= shared->tmp.merges[READ];
 740                 p->merges[WRITE] -= shared->tmp.merges[WRITE];
 741                 p->ticks[READ] -= shared->tmp.ticks[READ];
 742                 p->ticks[WRITE] -= shared->tmp.ticks[WRITE];
 743                 p->io_ticks[READ] -= shared->tmp.io_ticks[READ];
 744                 p->io_ticks[WRITE] -= shared->tmp.io_ticks[WRITE];
 745                 p->io_ticks_total -= shared->tmp.io_ticks_total;
 746                 p->time_in_queue -= shared->tmp.time_in_queue;
 747                 local_irq_enable();
 748                 if (s->n_histogram_entries) {
 749                         unsigned i;
 750                         for (i = 0; i < s->n_histogram_entries + 1; i++) {
 751                                 local_irq_disable();
 752                                 p = &s->stat_percpu[smp_processor_id()][x];
 753                                 p->histogram[i] -= shared->tmp.histogram[i];
 754                                 local_irq_enable();
 755                         }
 756                 }
 757                 cond_resched();
 758         }
 759 }
 760
 761 static int dm_stats_clear(struct dm_stats *stats, int id)
 762 {
 763         struct dm_stat *s;
 764
 765         mutex_lock(&stats->mutex);
 766
 767         s = __dm_stats_find(stats, id);
 768         if (!s) {
 769                 mutex_unlock(&stats->mutex);
 770                 return -ENOENT;
 771         }
 772
 773         __dm_stat_clear(s, 0, s->n_entries, true);
 774
 775         mutex_unlock(&stats->mutex);
 776
 777         return 1;
 778 }
 779
 780 /*
 781  * This is like jiffies_to_msec, but works for 64-bit values.
 782  */
 783 static unsigned long long dm_jiffies_to_msec64(struct dm_stat *s, unsigned long long j)
 784 {
 785         unsigned long long result;
 786         unsigned mult;
 787
 788         if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
 789                 return j;
 790
 791         result = 0;
 792         if (j)
 793                 result = jiffies_to_msecs(j & 0x3fffff);
 794         if (j >= 1 << 22) {
 795                 mult = jiffies_to_msecs(1 << 22);
 796                 result += (unsigned long long)mult * (unsigned long long)jiffies_to_msecs((j >> 22) & 0x3fffff);
 797         }
 798         if (j >= 1ULL << 44)
 799                 result += (unsigned long long)mult * (unsigned long long)mult * (unsigned long long)jiffies_to_msecs(j >> 44);
 800
 801         return result;
 802 }
 803
 804 static int dm_stats_print(struct dm_stats *stats, int id,
 805                           size_t idx_start, size_t idx_len,
 806                           bool clear, char *result, unsigned maxlen)
 807 {
 808         unsigned sz = 0;
 809         struct dm_stat *s;
 810         size_t x;
 811         sector_t start, end, step;
 812         size_t idx_end;
 813         struct dm_stat_shared *shared;
 814
 815         /*
 816          * Output format:
 817          *   <start_sector>+<length> counters
 818          */
 819
 820         mutex_lock(&stats->mutex);
 821
 822         s = __dm_stats_find(stats, id);
 823         if (!s) {
 824                 mutex_unlock(&stats->mutex);
 825                 return -ENOENT;
 826         }
 827
 828         idx_end = idx_start + idx_len;
 829         if (idx_end < idx_start ||
 830             idx_end > s->n_entries)
 831                 idx_end = s->n_entries;
 832
 833         if (idx_start > idx_end)
 834                 idx_start = idx_end;
 835
 836         step = s->step;
 837         start = s->start + (step * idx_start);
 838
 839         for (x = idx_start; x < idx_end; x++, start = end) {
 840                 shared = &s->stat_shared[x];
 841                 end = start + step;
 842                 if (unlikely(end > s->end))
 843                         end = s->end;
 844
 845                 __dm_stat_init_temporary_percpu_totals(shared, s, x);
 846
 847                 DMEMIT("%llu+%llu %llu %llu %llu %llu %llu %llu %llu %llu %d %llu %llu %llu %llu",
 848                        (unsigned long long)start,
 849                        (unsigned long long)step,
 850                        shared->tmp.ios[READ],
 851                        shared->tmp.merges[READ],
 852                        shared->tmp.sectors[READ],
 853                        dm_jiffies_to_msec64(s, shared->tmp.ticks[READ]),
 854                        shared->tmp.ios[WRITE],
 855                        shared->tmp.merges[WRITE],
 856                        shared->tmp.sectors[WRITE],
 857                        dm_jiffies_to_msec64(s, shared->tmp.ticks[WRITE]),
 858                        dm_stat_in_flight(shared),
 859                        dm_jiffies_to_msec64(s, shared->tmp.io_ticks_total),
 860                        dm_jiffies_to_msec64(s, shared->tmp.time_in_queue),
 861                        dm_jiffies_to_msec64(s, shared->tmp.io_ticks[READ]),
 862                        dm_jiffies_to_msec64(s, shared->tmp.io_ticks[WRITE]));
 863                 if (s->n_histogram_entries) {
 864                         unsigned i;
 865                         for (i = 0; i < s->n_histogram_entries + 1; i++) {
 866                                 DMEMIT("%s%llu", !i ? " " : ":", shared->tmp.histogram[i]);
 867                         }
 868                 }
 869                 DMEMIT("\n");
 870
 871                 if (unlikely(sz + 1 >= maxlen))
 872                         goto buffer_overflow;
 873
 874                 cond_resched();
 875         }
 876
 877         if (clear)
 878                 __dm_stat_clear(s, idx_start, idx_end, false);
 879
 880 buffer_overflow:
 881         mutex_unlock(&stats->mutex);
 882
 883         return 1;
 884 }
 885
 886 static int dm_stats_set_aux(struct dm_stats *stats, int id, const char *aux_data)
 887 {
 888         struct dm_stat *s;
 889         const char *new_aux_data;
 890
 891         mutex_lock(&stats->mutex);
 892
 893         s = __dm_stats_find(stats, id);
 894         if (!s) {
 895                 mutex_unlock(&stats->mutex);
 896                 return -ENOENT;
 897         }
 898
 899         new_aux_data = kstrdup(aux_data, GFP_KERNEL);
 900         if (!new_aux_data) {
 901                 mutex_unlock(&stats->mutex);
 902                 return -ENOMEM;
 903         }
 904
 905         kfree(s->aux_data);
 906         s->aux_data = new_aux_data;
 907
 908         mutex_unlock(&stats->mutex);
 909
 910         return 0;
 911 }
 912
 913 static int parse_histogram(const char *h, unsigned *n_histogram_entries,
 914                            unsigned long long **histogram_boundaries)
 915 {
 916         const char *q;
 917         unsigned n;
 918         unsigned long long last;
 919
 920         *n_histogram_entries = 1;
 921         for (q = h; *q; q++)
 922                 if (*q == ',')
 923                         (*n_histogram_entries)++;
 924
 925         *histogram_boundaries = kmalloc(*n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
 926         if (!*histogram_boundaries)
 927                 return -ENOMEM;
 928
 929         n = 0;
 930         last = 0;
 931         while (1) {
 932                 unsigned long long hi;
 933                 int s;
 934                 char ch;
 935                 s = sscanf(h, "%llu%c", &hi, &ch);
 936                 if (!s || (s == 2 && ch != ','))
 937                         return -EINVAL;
 938                 if (hi <= last)
 939                         return -EINVAL;
 940                 last = hi;
 941                 (*histogram_boundaries)[n] = hi;
 942                 if (s == 1)
 943                         return 0;
 944                 h = strchr(h, ',') + 1;
 945                 n++;
 946         }
 947 }
 948
 949 static int message_stats_create(struct mapped_device *md,
 950                                 unsigned argc, char **argv,
 951                                 char *result, unsigned maxlen)
 952 {
 953         int r;
 954         int id;
 955         char dummy;
 956         unsigned long long start, end, len, step;
 957         unsigned divisor;
 958         const char *program_id, *aux_data;
 959         unsigned stat_flags = 0;
 960
 961         unsigned n_histogram_entries = 0;
 962         unsigned long long *histogram_boundaries = NULL;
 963
 964         struct dm_arg_set as, as_backup;
 965         const char *a;
 966         unsigned feature_args;
 967
 968         /*
 969          * Input format:
 970          *   <range> <step> [<extra_parameters> <parameters>] [<program_id> [<aux_data>]]
 971          */
 972
 973         if (argc < 3)
 974                 goto ret_einval;
 975
 976         as.argc = argc;
 977         as.argv = argv;
 978         dm_consume_args(&as, 1);
 979
 980         a = dm_shift_arg(&as);
 981         if (!strcmp(a, "-")) {
 982                 start = 0;
 983                 len = dm_get_size(md);
 984                 if (!len)
 985                         len = 1;
 986         } else if (sscanf(a, "%llu+%llu%c", &start, &len, &dummy) != 2 ||
 987                    start != (sector_t)start || len != (sector_t)len)
 988                 goto ret_einval;
 989
 990         end = start + len;
 991         if (start >= end)
 992                 goto ret_einval;
 993
 994         a = dm_shift_arg(&as);
 995         if (sscanf(a, "/%u%c", &divisor, &dummy) == 1) {
 996                 if (!divisor)
 997                         return -EINVAL;
 998                 step = end - start;
 999                 if (do_div(step, divisor))
1000                         step++;
1001                 if (!step)
1002                         step = 1;
1003         } else if (sscanf(a, "%llu%c", &step, &dummy) != 1 ||
1004                    step != (sector_t)step || !step)
1005                 goto ret_einval;
1006
1007         as_backup = as;
1008         a = dm_shift_arg(&as);
1009         if (a && sscanf(a, "%u%c", &feature_args, &dummy) == 1) {
1010                 while (feature_args--) {
1011                         a = dm_shift_arg(&as);
1012                         if (!a)
1013                                 goto ret_einval;
1014                         if (!strcasecmp(a, "precise_timestamps"))
1015                                 stat_flags |= STAT_PRECISE_TIMESTAMPS;
1016                         else if (!strncasecmp(a, "histogram:", 10)) {
1017                                 if (n_histogram_entries)
1018                                         goto ret_einval;
1019                                 if ((r = parse_histogram(a + 10, &n_histogram_entries, &histogram_boundaries)))
1020                                         goto ret;
1021                         } else
1022                                 goto ret_einval;
1023                 }
1024         } else {
1025                 as = as_backup;
1026         }
1027
1028         program_id = "-";
1029         aux_data = "-";
1030
1031         a = dm_shift_arg(&as);
1032         if (a)
1033                 program_id = a;
1034
1035         a = dm_shift_arg(&as);
1036         if (a)
1037                 aux_data = a;
1038
1039         if (as.argc)
1040                 goto ret_einval;
1041
1042         /*
1043          * If a buffer overflow happens after we created the region,
1044          * it's too late (the userspace would retry with a larger
1045          * buffer, but the region id that caused the overflow is already
1046          * leaked).  So we must detect buffer overflow in advance.
1047          */
1048         snprintf(result, maxlen, "%d", INT_MAX);
1049         if (dm_message_test_buffer_overflow(result, maxlen)) {
1050                 r = 1;
1051                 goto ret;
1052         }
1053
1054         id = dm_stats_create(dm_get_stats(md), start, end, step, stat_flags,
1055                              n_histogram_entries, histogram_boundaries, program_id, aux_data,
1056                              dm_internal_suspend_fast, dm_internal_resume_fast, md);
1057         if (id < 0) {
1058                 r = id;
1059                 goto ret;
1060         }
1061
1062         snprintf(result, maxlen, "%d", id);
1063
1064         r = 1;
1065         goto ret;
1066
1067 ret_einval:
1068         r = -EINVAL;
1069 ret:
1070         kfree(histogram_boundaries);
1071         return r;
1072 }
1073
1074 static int message_stats_delete(struct mapped_device *md,
1075                                 unsigned argc, char **argv)
1076 {
1077         int id;
1078         char dummy;
1079
1080         if (argc != 2)
1081                 return -EINVAL;
1082
1083         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1084                 return -EINVAL;
1085
1086         return dm_stats_delete(dm_get_stats(md), id);
1087 }
1088
1089 static int message_stats_clear(struct mapped_device *md,
1090                                unsigned argc, char **argv)
1091 {
1092         int id;
1093         char dummy;
1094
1095         if (argc != 2)
1096                 return -EINVAL;
1097
1098         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1099                 return -EINVAL;
1100
1101         return dm_stats_clear(dm_get_stats(md), id);
1102 }
1103
1104 static int message_stats_list(struct mapped_device *md,
1105                               unsigned argc, char **argv,
1106                               char *result, unsigned maxlen)
1107 {
1108         int r;
1109         const char *program = NULL;
1110
1111         if (argc < 1 || argc > 2)
1112                 return -EINVAL;
1113
1114         if (argc > 1) {
1115                 program = kstrdup(argv[1], GFP_KERNEL);
1116                 if (!program)
1117                         return -ENOMEM;
1118         }
1119
1120         r = dm_stats_list(dm_get_stats(md), program, result, maxlen);
1121
1122         kfree(program);
1123
1124         return r;
1125 }
1126
1127 static int message_stats_print(struct mapped_device *md,
1128                                unsigned argc, char **argv, bool clear,
1129                                char *result, unsigned maxlen)
1130 {
1131         int id;
1132         char dummy;
1133         unsigned long idx_start = 0, idx_len = ULONG_MAX;
1134
1135         if (argc != 2 && argc != 4)
1136                 return -EINVAL;
1137
1138         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1139                 return -EINVAL;
1140
1141         if (argc > 3) {
1142                 if (strcmp(argv[2], "-") &&
1143                     sscanf(argv[2], "%lu%c", &idx_start, &dummy) != 1)
1144                         return -EINVAL;
1145                 if (strcmp(argv[3], "-") &&
1146                     sscanf(argv[3], "%lu%c", &idx_len, &dummy) != 1)
1147                         return -EINVAL;
1148         }
1149
1150         return dm_stats_print(dm_get_stats(md), id, idx_start, idx_len, clear,
1151                               result, maxlen);
1152 }
1153
1154 static int message_stats_set_aux(struct mapped_device *md,
1155                                  unsigned argc, char **argv)
1156 {
1157         int id;
1158         char dummy;
1159
1160         if (argc != 3)
1161                 return -EINVAL;
1162
1163         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1164                 return -EINVAL;
1165
1166         return dm_stats_set_aux(dm_get_stats(md), id, argv[2]);
1167 }
1168
1169 int dm_stats_message(struct mapped_device *md, unsigned argc, char **argv,
1170                      char *result, unsigned maxlen)
1171 {
1172         int r;
1173
1174         /* All messages here must start with '@' */
1175         if (!strcasecmp(argv[0], "@stats_create"))
1176                 r = message_stats_create(md, argc, argv, result, maxlen);
1177         else if (!strcasecmp(argv[0], "@stats_delete"))
1178                 r = message_stats_delete(md, argc, argv);
1179         else if (!strcasecmp(argv[0], "@stats_clear"))
1180                 r = message_stats_clear(md, argc, argv);
1181         else if (!strcasecmp(argv[0], "@stats_list"))
1182                 r = message_stats_list(md, argc, argv, result, maxlen);
1183         else if (!strcasecmp(argv[0], "@stats_print"))
1184                 r = message_stats_print(md, argc, argv, false, result, maxlen);
1185         else if (!strcasecmp(argv[0], "@stats_print_clear"))
1186                 r = message_stats_print(md, argc, argv, true, result, maxlen);
1187         else if (!strcasecmp(argv[0], "@stats_set_aux"))
1188                 r = message_stats_set_aux(md, argc, argv);
1189         else
1190                 return 2; /* this wasn't a stats message */
1191
1192         if (r == -EINVAL)
1193                 DMWARN("Invalid parameters for message %s", argv[0]);
1194
1195         return r;
1196 }
1197
1198 int __init dm_statistics_init(void)
1199 {
1200         shared_memory_amount = 0;
1201         dm_stat_need_rcu_barrier = 0;
1202         return 0;
1203 }
1204
1205 void dm_statistics_exit(void)
1206 {
1207         if (dm_stat_need_rcu_barrier)
1208                 rcu_barrier();
1209         if (WARN_ON(shared_memory_amount))
1210                 DMCRIT("shared_memory_amount leaked: %lu", shared_memory_amount);
1211 }
1212
1213 module_param_named(stats_current_allocated_bytes, shared_memory_amount, ulong, S_IRUGO);
1214 MODULE_PARM_DESC(stats_current_allocated_bytes, "Memory currently used by statistics");