diskutil.c

   1 #include <stdio.h>
   2 #include <string.h>
   3 #include <sys/time.h>
   4 #include <sys/types.h>
   5 #include <sys/stat.h>
   6 #include <dirent.h>
   7 #include <libgen.h>
   8 #include <math.h>
   9
  10 #include "fio.h"
  11 #include "smalloc.h"
  12 #include "diskutil.h"
  13
  14 static int last_majdev, last_mindev;
  15 static struct disk_util *last_du;
  16
  17 static struct fio_mutex *disk_util_mutex;
  18 static int disk_util_exit;
  19
  20 FLIST_HEAD(disk_list);
  21
  22 static struct disk_util *__init_per_file_disk_util(struct thread_data *td,
  23                 int majdev, int mindev, char *path);
  24
  25 static void disk_util_free(struct disk_util *du)
  26 {
  27         if (du == last_du)
  28                 last_du = NULL;
  29
  30         while (!flist_empty(&du->slaves)) {
  31                 struct disk_util *slave;
  32
  33                 slave = flist_entry(du->slaves.next, struct disk_util, slavelist);
  34                 flist_del(&slave->slavelist);
  35                 slave->users--;
  36         }
  37
  38         fio_mutex_remove(du->lock);
  39         sfree(du);
  40 }
  41
  42 static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
  43 {
  44         unsigned in_flight;
  45         unsigned long long sectors[2];
  46         char line[256];
  47         FILE *f;
  48         char *p;
  49         int ret;
  50
  51         dprint(FD_DISKUTIL, "open stat file: %s\n", du->path);
  52
  53         f = fopen(du->path, "r");
  54         if (!f)
  55                 return 1;
  56
  57         p = fgets(line, sizeof(line), f);
  58         if (!p) {
  59                 fclose(f);
  60                 return 1;
  61         }
  62
  63         dprint(FD_DISKUTIL, "%s: %s", du->path, p);
  64
  65         ret = sscanf(p, "%u %u %llu %u %u %u %llu %u %u %u %u\n", &dus->ios[0],
  66                                         &dus->merges[0], &sectors[0],
  67                                         &dus->ticks[0], &dus->ios[1],
  68                                         &dus->merges[1], &sectors[1],
  69                                         &dus->ticks[1], &in_flight,
  70                                         &dus->io_ticks, &dus->time_in_queue);
  71         fclose(f);
  72         dprint(FD_DISKUTIL, "%s: stat read ok? %d\n", du->path, ret == 1);
  73         dus->sectors[0] = sectors[0];
  74         dus->sectors[1] = sectors[1];
  75         return ret != 11;
  76 }
  77
  78 static void update_io_tick_disk(struct disk_util *du)
  79 {
  80         struct disk_util_stat __dus, *dus, *ldus;
  81         struct timeval t;
  82
  83         if (!du->users)
  84                 return;
  85         if (get_io_ticks(du, &__dus))
  86                 return;
  87
  88         dus = &du->dus;
  89         ldus = &du->last_dus;
  90
  91         dus->sectors[0] += (__dus.sectors[0] - ldus->sectors[0]);
  92         dus->sectors[1] += (__dus.sectors[1] - ldus->sectors[1]);
  93         dus->ios[0] += (__dus.ios[0] - ldus->ios[0]);
  94         dus->ios[1] += (__dus.ios[1] - ldus->ios[1]);
  95         dus->merges[0] += (__dus.merges[0] - ldus->merges[0]);
  96         dus->merges[1] += (__dus.merges[1] - ldus->merges[1]);
  97         dus->ticks[0] += (__dus.ticks[0] - ldus->ticks[0]);
  98         dus->ticks[1] += (__dus.ticks[1] - ldus->ticks[1]);
  99         dus->io_ticks += (__dus.io_ticks - ldus->io_ticks);
 100         dus->time_in_queue += (__dus.time_in_queue - ldus->time_in_queue);
 101
 102         fio_gettime(&t, NULL);
 103         dus->msec += mtime_since(&du->time, &t);
 104         memcpy(&du->time, &t, sizeof(t));
 105         memcpy(ldus, &__dus, sizeof(__dus));
 106 }
 107
 108 int update_io_ticks(void)
 109 {
 110         struct flist_head *entry;
 111         struct disk_util *du;
 112         int ret = 0;
 113
 114         dprint(FD_DISKUTIL, "update io ticks\n");
 115
 116         fio_mutex_down(disk_util_mutex);
 117
 118         if (!disk_util_exit) {
 119                 flist_for_each(entry, &disk_list) {
 120                         du = flist_entry(entry, struct disk_util, list);
 121                         update_io_tick_disk(du);
 122                 }
 123         } else
 124                 ret = 1;
 125
 126         fio_mutex_up(disk_util_mutex);
 127         return ret;
 128 }
 129
 130 static struct disk_util *disk_util_exists(int major, int minor)
 131 {
 132         struct flist_head *entry;
 133         struct disk_util *du;
 134
 135         flist_for_each(entry, &disk_list) {
 136                 du = flist_entry(entry, struct disk_util, list);
 137
 138                 if (major == du->major && minor == du->minor)
 139                         return du;
 140         }
 141
 142         return NULL;
 143 }
 144
 145 static int get_device_numbers(char *file_name, int *maj, int *min)
 146 {
 147         struct stat st;
 148         int majdev, mindev;
 149         char tempname[PATH_MAX], *p;
 150
 151         if (!lstat(file_name, &st)) {
 152                 if (S_ISBLK(st.st_mode)) {
 153                         majdev = major(st.st_rdev);
 154                         mindev = minor(st.st_rdev);
 155                 } else if (S_ISCHR(st.st_mode)) {
 156                         majdev = major(st.st_rdev);
 157                         mindev = minor(st.st_rdev);
 158                         if (fio_lookup_raw(st.st_rdev, &majdev, &mindev))
 159                                 return -1;
 160                 } else if (S_ISFIFO(st.st_mode))
 161                         return -1;
 162                 else {
 163                         majdev = major(st.st_dev);
 164                         mindev = minor(st.st_dev);
 165                 }
 166         } else {
 167                 /*
 168                  * must be a file, open "." in that path
 169                  */
 170                 strncpy(tempname, file_name, PATH_MAX - 1);
 171                 p = dirname(tempname);
 172                 if (stat(p, &st)) {
 173                         perror("disk util stat");
 174                         return -1;
 175                 }
 176
 177                 majdev = major(st.st_dev);
 178                 mindev = minor(st.st_dev);
 179         }
 180
 181         *min = mindev;
 182         *maj = majdev;
 183
 184         return 0;
 185 }
 186
 187 static int read_block_dev_entry(char *path, int *maj, int *min)
 188 {
 189         char line[256], *p;
 190         FILE *f;
 191
 192         f = fopen(path, "r");
 193         if (!f) {
 194                 perror("open path");
 195                 return 1;
 196         }
 197
 198         p = fgets(line, sizeof(line), f);
 199         fclose(f);
 200
 201         if (!p)
 202                 return 1;
 203
 204         if (sscanf(p, "%u:%u", maj, min) != 2)
 205                 return 1;
 206
 207         return 0;
 208 }
 209
 210 static void find_add_disk_slaves(struct thread_data *td, char *path,
 211                                  struct disk_util *masterdu)
 212 {
 213         DIR *dirhandle = NULL;
 214         struct dirent *dirent = NULL;
 215         char slavesdir[PATH_MAX], temppath[PATH_MAX], slavepath[PATH_MAX];
 216         struct disk_util *slavedu = NULL;
 217         int majdev, mindev;
 218         ssize_t linklen;
 219
 220         sprintf(slavesdir, "%s/%s", path, "slaves");
 221         dirhandle = opendir(slavesdir);
 222         if (!dirhandle)
 223                 return;
 224
 225         while ((dirent = readdir(dirhandle)) != NULL) {
 226                 if (!strcmp(dirent->d_name, ".") ||
 227                     !strcmp(dirent->d_name, ".."))
 228                         continue;
 229
 230                 sprintf(temppath, "%s%s%s", slavesdir, FIO_OS_PATH_SEPARATOR, dirent->d_name);
 231                 /* Can we always assume that the slaves device entries
 232                  * are links to the real directories for the slave
 233                  * devices?
 234                  */
 235                 linklen = readlink(temppath, slavepath, PATH_MAX - 0);
 236                 if (linklen  < 0) {
 237                         perror("readlink() for slave device.");
 238                         return;
 239                 }
 240                 slavepath[linklen] = '\0';
 241
 242                 sprintf(temppath, "%s/%s/dev", slavesdir, slavepath);
 243                 if (read_block_dev_entry(temppath, &majdev, &mindev)) {
 244                         perror("Error getting slave device numbers.");
 245                         return;
 246                 }
 247
 248                 /*
 249                  * See if this maj,min already exists
 250                  */
 251                 slavedu = disk_util_exists(majdev, mindev);
 252                 if (slavedu)
 253                         continue;
 254
 255                 sprintf(temppath, "%s%s%s", slavesdir, FIO_OS_PATH_SEPARATOR, slavepath);
 256                 __init_per_file_disk_util(td, majdev, mindev, temppath);
 257                 slavedu = disk_util_exists(majdev, mindev);
 258
 259                 /* Should probably use an assert here. slavedu should
 260                  * always be present at this point. */
 261                 if (slavedu) {
 262                         slavedu->users++;
 263                         flist_add_tail(&slavedu->slavelist, &masterdu->slaves);
 264                 }
 265         }
 266
 267         closedir(dirhandle);
 268 }
 269
 270 static struct disk_util *disk_util_add(struct thread_data *td, int majdev,
 271                                        int mindev, char *path)
 272 {
 273         struct disk_util *du, *__du;
 274         struct flist_head *entry;
 275
 276         dprint(FD_DISKUTIL, "add maj/min %d/%d: %s\n", majdev, mindev, path);
 277
 278         du = smalloc(sizeof(*du));
 279         memset(du, 0, sizeof(*du));
 280         INIT_FLIST_HEAD(&du->list);
 281         sprintf(du->path, "%s/stat", path);
 282         strncpy((char *) du->dus.name, basename(path), FIO_DU_NAME_SZ);
 283         du->sysfs_root = path;
 284         du->major = majdev;
 285         du->minor = mindev;
 286         INIT_FLIST_HEAD(&du->slavelist);
 287         INIT_FLIST_HEAD(&du->slaves);
 288         du->lock = fio_mutex_init(1);
 289         du->users = 0;
 290
 291         fio_mutex_down(disk_util_mutex);
 292
 293         flist_for_each(entry, &disk_list) {
 294                 __du = flist_entry(entry, struct disk_util, list);
 295
 296                 dprint(FD_DISKUTIL, "found %s in list\n", __du->dus.name);
 297
 298                 if (!strcmp((char *) du->dus.name, (char *) __du->dus.name)) {
 299                         disk_util_free(du);
 300                         fio_mutex_up(disk_util_mutex);
 301                         return __du;
 302                 }
 303         }
 304
 305         dprint(FD_DISKUTIL, "add %s to list\n", du->dus.name);
 306
 307         fio_gettime(&du->time, NULL);
 308         get_io_ticks(du, &du->last_dus);
 309
 310         flist_add_tail(&du->list, &disk_list);
 311         find_add_disk_slaves(td, path, du);
 312         fio_mutex_up(disk_util_mutex);
 313         return du;
 314 }
 315
 316 static int check_dev_match(int majdev, int mindev, char *path)
 317 {
 318         int major, minor;
 319
 320         if (read_block_dev_entry(path, &major, &minor))
 321                 return 1;
 322
 323         if (majdev == major && mindev == minor)
 324                 return 0;
 325
 326         return 1;
 327 }
 328
 329 static int find_block_dir(int majdev, int mindev, char *path, int link_ok)
 330 {
 331         struct dirent *dir;
 332         struct stat st;
 333         int found = 0;
 334         DIR *D;
 335
 336         D = opendir(path);
 337         if (!D)
 338                 return 0;
 339
 340         while ((dir = readdir(D)) != NULL) {
 341                 char full_path[256];
 342
 343                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
 344                         continue;
 345
 346                 sprintf(full_path, "%s%s%s", path, FIO_OS_PATH_SEPARATOR, dir->d_name);
 347
 348                 if (!strcmp(dir->d_name, "dev")) {
 349                         if (!check_dev_match(majdev, mindev, full_path)) {
 350                                 found = 1;
 351                                 break;
 352                         }
 353                 }
 354
 355                 if (link_ok) {
 356                         if (stat(full_path, &st) == -1) {
 357                                 perror("stat");
 358                                 break;
 359                         }
 360                 } else {
 361                         if (lstat(full_path, &st) == -1) {
 362                                 perror("stat");
 363                                 break;
 364                         }
 365                 }
 366
 367                 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
 368                         continue;
 369
 370                 found = find_block_dir(majdev, mindev, full_path, 0);
 371                 if (found) {
 372                         strcpy(path, full_path);
 373                         break;
 374                 }
 375         }
 376
 377         closedir(D);
 378         return found;
 379 }
 380
 381 static struct disk_util *__init_per_file_disk_util(struct thread_data *td,
 382                                                    int majdev, int mindev,
 383                                                    char *path)
 384 {
 385         struct stat st;
 386         char tmp[PATH_MAX];
 387         char *p;
 388
 389         /*
 390          * If there's a ../queue/ directory there, we are inside a partition.
 391          * Check if that is the case and jump back. For loop/md/dm etc we
 392          * are already in the right spot.
 393          */
 394         sprintf(tmp, "%s/../queue", path);
 395         if (!stat(tmp, &st)) {
 396                 p = dirname(path);
 397                 sprintf(tmp, "%s/queue", p);
 398                 if (stat(tmp, &st)) {
 399                         log_err("unknown sysfs layout\n");
 400                         return NULL;
 401                 }
 402                 strncpy(tmp, p, PATH_MAX - 1);
 403                 sprintf(path, "%s", tmp);
 404         }
 405
 406         if (td->o.ioscheduler && !td->sysfs_root)
 407                 td->sysfs_root = strdup(path);
 408
 409         return disk_util_add(td, majdev, mindev, path);
 410 }
 411
 412 static struct disk_util *init_per_file_disk_util(struct thread_data *td,
 413                                                  char *filename)
 414 {
 415
 416         char foo[PATH_MAX];
 417         struct disk_util *du;
 418         int mindev, majdev;
 419
 420         if (get_device_numbers(filename, &majdev, &mindev))
 421                 return NULL;
 422
 423         dprint(FD_DISKUTIL, "%s belongs to maj/min %d/%d\n", filename, majdev,
 424                         mindev);
 425
 426         du = disk_util_exists(majdev, mindev);
 427         if (du) {
 428                 if (td->o.ioscheduler && !td->sysfs_root)
 429                         td->sysfs_root = strdup(du->sysfs_root);
 430
 431                 return du;
 432         }
 433
 434         /*
 435          * for an fs without a device, we will repeatedly stat through
 436          * sysfs which can take oodles of time for thousands of files. so
 437          * cache the last lookup and compare with that before going through
 438          * everything again.
 439          */
 440         if (mindev == last_mindev && majdev == last_majdev)
 441                 return last_du;
 442
 443         last_mindev = mindev;
 444         last_majdev = majdev;
 445
 446         sprintf(foo, "/sys/block");
 447         if (!find_block_dir(majdev, mindev, foo, 1))
 448                 return NULL;
 449
 450         return __init_per_file_disk_util(td, majdev, mindev, foo);
 451 }
 452
 453 static struct disk_util *__init_disk_util(struct thread_data *td,
 454                                           struct fio_file *f)
 455 {
 456         return init_per_file_disk_util(td, f->file_name);
 457 }
 458
 459 void init_disk_util(struct thread_data *td)
 460 {
 461         struct fio_file *f;
 462         unsigned int i;
 463
 464         if (!td->o.do_disk_util ||
 465             (td->io_ops->flags & (FIO_DISKLESSIO | FIO_NODISKUTIL)))
 466                 return;
 467
 468         for_each_file(td, f, i)
 469                 f->du = __init_disk_util(td, f);
 470 }
 471
 472 static void show_agg_stats(struct disk_util_agg *agg, int terse)
 473 {
 474         if (!agg->slavecount)
 475                 return;
 476
 477         if (!terse) {
 478                 log_info(", aggrios=%u/%u, aggrmerge=%u/%u, aggrticks=%u/%u,"
 479                                 " aggrin_queue=%u, aggrutil=%3.2f%%",
 480                                 agg->ios[0] / agg->slavecount,
 481                                 agg->ios[1] / agg->slavecount,
 482                                 agg->merges[0] / agg->slavecount,
 483                                 agg->merges[1] / agg->slavecount,
 484                                 agg->ticks[0] / agg->slavecount,
 485                                 agg->ticks[1] / agg->slavecount,
 486                                 agg->time_in_queue / agg->slavecount,
 487                                 agg->max_util.u.f);
 488         } else {
 489                 log_info(";slaves;%u;%u;%u;%u;%u;%u;%u;%3.2f%%",
 490                                 agg->ios[0] / agg->slavecount,
 491                                 agg->ios[1] / agg->slavecount,
 492                                 agg->merges[0] / agg->slavecount,
 493                                 agg->merges[1] / agg->slavecount,
 494                                 agg->ticks[0] / agg->slavecount,
 495                                 agg->ticks[1] / agg->slavecount,
 496                                 agg->time_in_queue / agg->slavecount,
 497                                 agg->max_util.u.f);
 498         }
 499 }
 500
 501 static void aggregate_slaves_stats(struct disk_util *masterdu)
 502 {
 503         struct disk_util_agg *agg = &masterdu->agg;
 504         struct disk_util_stat *dus;
 505         struct flist_head *entry;
 506         struct disk_util *slavedu;
 507         double util;
 508
 509         flist_for_each(entry, &masterdu->slaves) {
 510                 slavedu = flist_entry(entry, struct disk_util, slavelist);
 511                 dus = &slavedu->dus;
 512                 agg->ios[0] += dus->ios[0];
 513                 agg->ios[1] += dus->ios[1];
 514                 agg->merges[0] += dus->merges[0];
 515                 agg->merges[1] += dus->merges[1];
 516                 agg->sectors[0] += dus->sectors[0];
 517                 agg->sectors[1] += dus->sectors[1];
 518                 agg->ticks[0] += dus->ticks[0];
 519                 agg->ticks[1] += dus->ticks[1];
 520                 agg->time_in_queue += dus->time_in_queue;
 521                 agg->slavecount++;
 522
 523                 util = (double) (100 * dus->io_ticks / (double) slavedu->dus.msec);
 524                 /* System utilization is the utilization of the
 525                  * component with the highest utilization.
 526                  */
 527                 if (util > agg->max_util.u.f)
 528                         agg->max_util.u.f = util;
 529
 530         }
 531
 532         if (agg->max_util.u.f > 100.0)
 533                 agg->max_util.u.f = 100.0;
 534 }
 535
 536 void free_disk_util(void)
 537 {
 538         struct disk_util *du;
 539
 540         disk_util_exit = 1;
 541         wait_for_disk_thread_exit();
 542
 543         fio_mutex_down(disk_util_mutex);
 544
 545         while (!flist_empty(&disk_list)) {
 546                 du = flist_entry(disk_list.next, struct disk_util, list);
 547                 flist_del(&du->list);
 548                 disk_util_free(du);
 549         }
 550
 551         last_majdev = last_mindev = -1;
 552         fio_mutex_up(disk_util_mutex);
 553         fio_mutex_remove(disk_util_mutex);
 554 }
 555
 556 void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
 557                      int terse)
 558 {
 559         double util = 0;
 560
 561         if (dus->msec)
 562                 util = (double) 100 * dus->io_ticks / (double) dus->msec;
 563         if (util > 100.0)
 564                 util = 100.0;
 565
 566         if (!terse) {
 567                 if (agg->slavecount)
 568                         log_info("  ");
 569
 570                 log_info("  %s: ios=%u/%u, merge=%u/%u, ticks=%u/%u, "
 571                          "in_queue=%u, util=%3.2f%%", dus->name,
 572                                         dus->ios[0], dus->ios[1],
 573                                         dus->merges[0], dus->merges[1],
 574                                         dus->ticks[0], dus->ticks[1],
 575                                         dus->time_in_queue, util);
 576         } else {
 577                 log_info(";%s;%u;%u;%u;%u;%u;%u;%u;%3.2f%%",
 578                                         dus->name, dus->ios[0], dus->ios[1],
 579                                         dus->merges[0], dus->merges[1],
 580                                         dus->ticks[0], dus->ticks[1],
 581                                         dus->time_in_queue, util);
 582         }
 583
 584         /*
 585          * If the device has slaves, aggregate the stats for
 586          * those slave devices also.
 587          */
 588         show_agg_stats(agg, terse);
 589
 590         if (!terse)
 591                 log_info("\n");
 592 }
 593
 594 void show_disk_util(int terse)
 595 {
 596         struct flist_head *entry;
 597         struct disk_util *du;
 598
 599         fio_mutex_down(disk_util_mutex);
 600
 601         if (flist_empty(&disk_list)) {
 602                 fio_mutex_up(disk_util_mutex);
 603                 return;
 604         }
 605
 606         if (!terse)
 607                 log_info("\nDisk stats (read/write):\n");
 608
 609         flist_for_each(entry, &disk_list) {
 610                 du = flist_entry(entry, struct disk_util, list);
 611
 612                 aggregate_slaves_stats(du);
 613                 print_disk_util(&du->dus, &du->agg, terse);
 614         }
 615
 616         fio_mutex_up(disk_util_mutex);
 617 }
 618
 619 void setup_disk_util(void)
 620 {
 621         disk_util_mutex = fio_mutex_init(1);
 622 }