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
  19 FLIST_HEAD(disk_list);
  20
  21 static struct disk_util *__init_per_file_disk_util(struct thread_data *td,
  22                 int majdev, int mindev, char *path);
  23
  24 static void disk_util_free(struct disk_util *du)
  25 {
  26         if (du == last_du)
  27                 last_du = NULL;
  28
  29         while (!flist_empty(&du->slaves)) {
  30                 struct disk_util *slave;
  31
  32                 slave = flist_entry(du->slaves.next, struct disk_util, slavelist);
  33                 flist_del(&slave->slavelist);
  34                 slave->users--;
  35         }
  36
  37         fio_mutex_remove(du->lock);
  38         sfree(du);
  39 }
  40
  41 static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
  42 {
  43         unsigned in_flight;
  44         unsigned long long sectors[2];
  45         char line[256];
  46         FILE *f;
  47         char *p;
  48         int ret;
  49
  50         dprint(FD_DISKUTIL, "open stat file: %s\n", du->path);
  51
  52         f = fopen(du->path, "r");
  53         if (!f)
  54                 return 1;
  55
  56         p = fgets(line, sizeof(line), f);
  57         if (!p) {
  58                 fclose(f);
  59                 return 1;
  60         }
  61
  62         dprint(FD_DISKUTIL, "%s: %s", du->path, p);
  63
  64         ret = sscanf(p, "%u %u %llu %u %u %u %llu %u %u %u %u\n", &dus->ios[0],
  65                                         &dus->merges[0], &sectors[0],
  66                                         &dus->ticks[0], &dus->ios[1],
  67                                         &dus->merges[1], &sectors[1],
  68                                         &dus->ticks[1], &in_flight,
  69                                         &dus->io_ticks, &dus->time_in_queue);
  70         fclose(f);
  71         dprint(FD_DISKUTIL, "%s: stat read ok? %d\n", du->path, ret == 1);
  72         dus->sectors[0] = sectors[0];
  73         dus->sectors[1] = sectors[1];
  74         return ret != 11;
  75 }
  76
  77 static void update_io_tick_disk(struct disk_util *du)
  78 {
  79         struct disk_util_stat __dus, *dus, *ldus;
  80         struct timeval t;
  81
  82         if (!du->users)
  83                 return;
  84         if (get_io_ticks(du, &__dus))
  85                 return;
  86
  87         dus = &du->dus;
  88         ldus = &du->last_dus;
  89
  90         dus->sectors[0] += (__dus.sectors[0] - ldus->sectors[0]);
  91         dus->sectors[1] += (__dus.sectors[1] - ldus->sectors[1]);
  92         dus->ios[0] += (__dus.ios[0] - ldus->ios[0]);
  93         dus->ios[1] += (__dus.ios[1] - ldus->ios[1]);
  94         dus->merges[0] += (__dus.merges[0] - ldus->merges[0]);
  95         dus->merges[1] += (__dus.merges[1] - ldus->merges[1]);
  96         dus->ticks[0] += (__dus.ticks[0] - ldus->ticks[0]);
  97         dus->ticks[1] += (__dus.ticks[1] - ldus->ticks[1]);
  98         dus->io_ticks += (__dus.io_ticks - ldus->io_ticks);
  99         dus->time_in_queue += (__dus.time_in_queue - ldus->time_in_queue);
 100
 101         fio_gettime(&t, NULL);
 102         dus->msec += mtime_since(&du->time, &t);
 103         memcpy(&du->time, &t, sizeof(t));
 104         memcpy(ldus, &__dus, sizeof(__dus));
 105 }
 106
 107 int update_io_ticks(void)
 108 {
 109         struct flist_head *entry;
 110         struct disk_util *du;
 111         int ret = 0;
 112
 113         dprint(FD_DISKUTIL, "update io ticks\n");
 114
 115         fio_mutex_down(disk_util_mutex);
 116
 117         if (!disk_util_exit) {
 118                 flist_for_each(entry, &disk_list) {
 119                         du = flist_entry(entry, struct disk_util, list);
 120                         update_io_tick_disk(du);
 121                 }
 122         } else
 123                 ret = 1;
 124
 125         fio_mutex_up(disk_util_mutex);
 126         return ret;
 127 }
 128
 129 static struct disk_util *disk_util_exists(int major, int minor)
 130 {
 131         struct flist_head *entry;
 132         struct disk_util *du;
 133
 134         fio_mutex_down(disk_util_mutex);
 135
 136         flist_for_each(entry, &disk_list) {
 137                 du = flist_entry(entry, struct disk_util, list);
 138
 139                 if (major == du->major && minor == du->minor) {
 140                         fio_mutex_up(disk_util_mutex);
 141                         return du;
 142                 }
 143         }
 144
 145         fio_mutex_up(disk_util_mutex);
 146         return NULL;
 147 }
 148
 149 static int get_device_numbers(char *file_name, int *maj, int *min)
 150 {
 151         struct stat st;
 152         int majdev, mindev;
 153         char tempname[PATH_MAX], *p;
 154
 155         if (!lstat(file_name, &st)) {
 156                 if (S_ISBLK(st.st_mode)) {
 157                         majdev = major(st.st_rdev);
 158                         mindev = minor(st.st_rdev);
 159                 } else if (S_ISCHR(st.st_mode)) {
 160                         majdev = major(st.st_rdev);
 161                         mindev = minor(st.st_rdev);
 162                         if (fio_lookup_raw(st.st_rdev, &majdev, &mindev))
 163                                 return -1;
 164                 } else if (S_ISFIFO(st.st_mode))
 165                         return -1;
 166                 else {
 167                         majdev = major(st.st_dev);
 168                         mindev = minor(st.st_dev);
 169                 }
 170         } else {
 171                 /*
 172                  * must be a file, open "." in that path
 173                  */
 174                 strncpy(tempname, file_name, PATH_MAX - 1);
 175                 p = dirname(tempname);
 176                 if (stat(p, &st)) {
 177                         perror("disk util stat");
 178                         return -1;
 179                 }
 180
 181                 majdev = major(st.st_dev);
 182                 mindev = minor(st.st_dev);
 183         }
 184
 185         *min = mindev;
 186         *maj = majdev;
 187
 188         return 0;
 189 }
 190
 191 static int read_block_dev_entry(char *path, int *maj, int *min)
 192 {
 193         char line[256], *p;
 194         FILE *f;
 195
 196         f = fopen(path, "r");
 197         if (!f) {
 198                 perror("open path");
 199                 return 1;
 200         }
 201
 202         p = fgets(line, sizeof(line), f);
 203         fclose(f);
 204
 205         if (!p)
 206                 return 1;
 207
 208         if (sscanf(p, "%u:%u", maj, min) != 2)
 209                 return 1;
 210
 211         return 0;
 212 }
 213
 214 static void find_add_disk_slaves(struct thread_data *td, char *path,
 215                                  struct disk_util *masterdu)
 216 {
 217         DIR *dirhandle = NULL;
 218         struct dirent *dirent = NULL;
 219         char slavesdir[PATH_MAX], temppath[PATH_MAX], slavepath[PATH_MAX];
 220         struct disk_util *slavedu = NULL;
 221         int majdev, mindev;
 222         ssize_t linklen;
 223
 224         sprintf(slavesdir, "%s/%s", path, "slaves");
 225         dirhandle = opendir(slavesdir);
 226         if (!dirhandle)
 227                 return;
 228
 229         while ((dirent = readdir(dirhandle)) != NULL) {
 230                 if (!strcmp(dirent->d_name, ".") ||
 231                     !strcmp(dirent->d_name, ".."))
 232                         continue;
 233
 234                 sprintf(temppath, "%s%s%s", slavesdir, FIO_OS_PATH_SEPARATOR, dirent->d_name);
 235                 /* Can we always assume that the slaves device entries
 236                  * are links to the real directories for the slave
 237                  * devices?
 238                  */
 239                 linklen = readlink(temppath, slavepath, PATH_MAX - 0);
 240                 if (linklen  < 0) {
 241                         perror("readlink() for slave device.");
 242                         return;
 243                 }
 244                 slavepath[linklen] = '\0';
 245
 246                 sprintf(temppath, "%s/%s/dev", slavesdir, slavepath);
 247                 if (read_block_dev_entry(temppath, &majdev, &mindev)) {
 248                         perror("Error getting slave device numbers.");
 249                         return;
 250                 }
 251
 252                 /*
 253                  * See if this maj,min already exists
 254                  */
 255                 slavedu = disk_util_exists(majdev, mindev);
 256                 if (slavedu)
 257                         continue;
 258
 259                 sprintf(temppath, "%s%s%s", slavesdir, FIO_OS_PATH_SEPARATOR, slavepath);
 260                 __init_per_file_disk_util(td, majdev, mindev, temppath);
 261                 slavedu = disk_util_exists(majdev, mindev);
 262
 263                 /* Should probably use an assert here. slavedu should
 264                  * always be present at this point. */
 265                 if (slavedu) {
 266                         slavedu->users++;
 267                         flist_add_tail(&slavedu->slavelist, &masterdu->slaves);
 268                 }
 269         }
 270
 271         closedir(dirhandle);
 272 }
 273
 274 static struct disk_util *disk_util_add(struct thread_data *td, int majdev,
 275                                        int mindev, char *path)
 276 {
 277         struct disk_util *du, *__du;
 278         struct flist_head *entry;
 279
 280         dprint(FD_DISKUTIL, "add maj/min %d/%d: %s\n", majdev, mindev, path);
 281
 282         du = smalloc(sizeof(*du));
 283         memset(du, 0, sizeof(*du));
 284         INIT_FLIST_HEAD(&du->list);
 285         sprintf(du->path, "%s/stat", path);
 286         strncpy((char *) du->dus.name, basename(path), FIO_DU_NAME_SZ);
 287         du->sysfs_root = path;
 288         du->major = majdev;
 289         du->minor = mindev;
 290         INIT_FLIST_HEAD(&du->slavelist);
 291         INIT_FLIST_HEAD(&du->slaves);
 292         du->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
 293         du->users = 0;
 294
 295         fio_mutex_down(disk_util_mutex);
 296
 297         flist_for_each(entry, &disk_list) {
 298                 __du = flist_entry(entry, struct disk_util, list);
 299
 300                 dprint(FD_DISKUTIL, "found %s in list\n", __du->dus.name);
 301
 302                 if (!strcmp((char *) du->dus.name, (char *) __du->dus.name)) {
 303                         disk_util_free(du);
 304                         fio_mutex_up(disk_util_mutex);
 305                         return __du;
 306                 }
 307         }
 308
 309         dprint(FD_DISKUTIL, "add %s to list\n", du->dus.name);
 310
 311         fio_gettime(&du->time, NULL);
 312         get_io_ticks(du, &du->last_dus);
 313
 314         flist_add_tail(&du->list, &disk_list);
 315         fio_mutex_up(disk_util_mutex);
 316
 317         find_add_disk_slaves(td, path, du);
 318         return du;
 319 }
 320
 321 static int check_dev_match(int majdev, int mindev, char *path)
 322 {
 323         int major, minor;
 324
 325         if (read_block_dev_entry(path, &major, &minor))
 326                 return 1;
 327
 328         if (majdev == major && mindev == minor)
 329                 return 0;
 330
 331         return 1;
 332 }
 333
 334 static int find_block_dir(int majdev, int mindev, char *path, int link_ok)
 335 {
 336         struct dirent *dir;
 337         struct stat st;
 338         int found = 0;
 339         DIR *D;
 340
 341         D = opendir(path);
 342         if (!D)
 343                 return 0;
 344
 345         while ((dir = readdir(D)) != NULL) {
 346                 char full_path[256];
 347
 348                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
 349                         continue;
 350
 351                 sprintf(full_path, "%s%s%s", path, FIO_OS_PATH_SEPARATOR, dir->d_name);
 352
 353                 if (!strcmp(dir->d_name, "dev")) {
 354                         if (!check_dev_match(majdev, mindev, full_path)) {
 355                                 found = 1;
 356                                 break;
 357                         }
 358                 }
 359
 360                 if (link_ok) {
 361                         if (stat(full_path, &st) == -1) {
 362                                 perror("stat");
 363                                 break;
 364                         }
 365                 } else {
 366                         if (lstat(full_path, &st) == -1) {
 367                                 perror("stat");
 368                                 break;
 369                         }
 370                 }
 371
 372                 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
 373                         continue;
 374
 375                 found = find_block_dir(majdev, mindev, full_path, 0);
 376                 if (found) {
 377                         strcpy(path, full_path);
 378                         break;
 379                 }
 380         }
 381
 382         closedir(D);
 383         return found;
 384 }
 385
 386 static struct disk_util *__init_per_file_disk_util(struct thread_data *td,
 387                                                    int majdev, int mindev,
 388                                                    char *path)
 389 {
 390         struct stat st;
 391         char tmp[PATH_MAX];
 392         char *p;
 393
 394         /*
 395          * If there's a ../queue/ directory there, we are inside a partition.
 396          * Check if that is the case and jump back. For loop/md/dm etc we
 397          * are already in the right spot.
 398          */
 399         sprintf(tmp, "%s/../queue", path);
 400         if (!stat(tmp, &st)) {
 401                 p = dirname(path);
 402                 sprintf(tmp, "%s/queue", p);
 403                 if (stat(tmp, &st)) {
 404                         log_err("unknown sysfs layout\n");
 405                         return NULL;
 406                 }
 407                 strncpy(tmp, p, PATH_MAX - 1);
 408                 sprintf(path, "%s", tmp);
 409         }
 410
 411         if (td->o.ioscheduler && !td->sysfs_root)
 412                 td->sysfs_root = strdup(path);
 413
 414         return disk_util_add(td, majdev, mindev, path);
 415 }
 416
 417 static struct disk_util *init_per_file_disk_util(struct thread_data *td,
 418                                                  char *filename)
 419 {
 420
 421         char foo[PATH_MAX];
 422         struct disk_util *du;
 423         int mindev, majdev;
 424
 425         if (get_device_numbers(filename, &majdev, &mindev))
 426                 return NULL;
 427
 428         dprint(FD_DISKUTIL, "%s belongs to maj/min %d/%d\n", filename, majdev,
 429                         mindev);
 430
 431         du = disk_util_exists(majdev, mindev);
 432         if (du) {
 433                 if (td->o.ioscheduler && !td->sysfs_root)
 434                         td->sysfs_root = strdup(du->sysfs_root);
 435
 436                 return du;
 437         }
 438
 439         /*
 440          * for an fs without a device, we will repeatedly stat through
 441          * sysfs which can take oodles of time for thousands of files. so
 442          * cache the last lookup and compare with that before going through
 443          * everything again.
 444          */
 445         if (mindev == last_mindev && majdev == last_majdev)
 446                 return last_du;
 447
 448         last_mindev = mindev;
 449         last_majdev = majdev;
 450
 451         sprintf(foo, "/sys/block");
 452         if (!find_block_dir(majdev, mindev, foo, 1))
 453                 return NULL;
 454
 455         return __init_per_file_disk_util(td, majdev, mindev, foo);
 456 }
 457
 458 static struct disk_util *__init_disk_util(struct thread_data *td,
 459                                           struct fio_file *f)
 460 {
 461         return init_per_file_disk_util(td, f->file_name);
 462 }
 463
 464 void init_disk_util(struct thread_data *td)
 465 {
 466         struct fio_file *f;
 467         unsigned int i;
 468
 469         if (!td->o.do_disk_util ||
 470             (td->io_ops->flags & (FIO_DISKLESSIO | FIO_NODISKUTIL)))
 471                 return;
 472
 473         for_each_file(td, f, i)
 474                 f->du = __init_disk_util(td, f);
 475 }
 476
 477 static void show_agg_stats(struct disk_util_agg *agg, int terse)
 478 {
 479         if (!agg->slavecount)
 480                 return;
 481
 482         if (!terse) {
 483                 log_info(", aggrios=%u/%u, aggrmerge=%u/%u, aggrticks=%u/%u,"
 484                                 " aggrin_queue=%u, aggrutil=%3.2f%%",
 485                                 agg->ios[0] / agg->slavecount,
 486                                 agg->ios[1] / agg->slavecount,
 487                                 agg->merges[0] / agg->slavecount,
 488                                 agg->merges[1] / agg->slavecount,
 489                                 agg->ticks[0] / agg->slavecount,
 490                                 agg->ticks[1] / agg->slavecount,
 491                                 agg->time_in_queue / agg->slavecount,
 492                                 agg->max_util.u.f);
 493         } else {
 494                 log_info(";slaves;%u;%u;%u;%u;%u;%u;%u;%3.2f%%",
 495                                 agg->ios[0] / agg->slavecount,
 496                                 agg->ios[1] / agg->slavecount,
 497                                 agg->merges[0] / agg->slavecount,
 498                                 agg->merges[1] / agg->slavecount,
 499                                 agg->ticks[0] / agg->slavecount,
 500                                 agg->ticks[1] / agg->slavecount,
 501                                 agg->time_in_queue / agg->slavecount,
 502                                 agg->max_util.u.f);
 503         }
 504 }
 505
 506 static void aggregate_slaves_stats(struct disk_util *masterdu)
 507 {
 508         struct disk_util_agg *agg = &masterdu->agg;
 509         struct disk_util_stat *dus;
 510         struct flist_head *entry;
 511         struct disk_util *slavedu;
 512         double util;
 513
 514         flist_for_each(entry, &masterdu->slaves) {
 515                 slavedu = flist_entry(entry, struct disk_util, slavelist);
 516                 dus = &slavedu->dus;
 517                 agg->ios[0] += dus->ios[0];
 518                 agg->ios[1] += dus->ios[1];
 519                 agg->merges[0] += dus->merges[0];
 520                 agg->merges[1] += dus->merges[1];
 521                 agg->sectors[0] += dus->sectors[0];
 522                 agg->sectors[1] += dus->sectors[1];
 523                 agg->ticks[0] += dus->ticks[0];
 524                 agg->ticks[1] += dus->ticks[1];
 525                 agg->time_in_queue += dus->time_in_queue;
 526                 agg->slavecount++;
 527
 528                 util = (double) (100 * dus->io_ticks / (double) slavedu->dus.msec);
 529                 /* System utilization is the utilization of the
 530                  * component with the highest utilization.
 531                  */
 532                 if (util > agg->max_util.u.f)
 533                         agg->max_util.u.f = util;
 534
 535         }
 536
 537         if (agg->max_util.u.f > 100.0)
 538                 agg->max_util.u.f = 100.0;
 539 }
 540
 541 void disk_util_prune_entries(void)
 542 {
 543         fio_mutex_down(disk_util_mutex);
 544
 545         while (!flist_empty(&disk_list)) {
 546                 struct disk_util *du;
 547
 548                 du = flist_entry(disk_list.next, struct disk_util, list);
 549                 flist_del(&du->list);
 550                 disk_util_free(du);
 551         }
 552
 553         last_majdev = last_mindev = -1;
 554         fio_mutex_up(disk_util_mutex);
 555         fio_mutex_remove(disk_util_mutex);
 556 }
 557
 558 void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
 559                      int terse)
 560 {
 561         double util = 0;
 562
 563         if (dus->msec)
 564                 util = (double) 100 * dus->io_ticks / (double) dus->msec;
 565         if (util > 100.0)
 566                 util = 100.0;
 567
 568         if (!terse) {
 569                 if (agg->slavecount)
 570                         log_info("  ");
 571
 572                 log_info("  %s: ios=%u/%u, merge=%u/%u, ticks=%u/%u, "
 573                          "in_queue=%u, util=%3.2f%%", dus->name,
 574                                         dus->ios[0], dus->ios[1],
 575                                         dus->merges[0], dus->merges[1],
 576                                         dus->ticks[0], dus->ticks[1],
 577                                         dus->time_in_queue, util);
 578         } else {
 579                 log_info(";%s;%u;%u;%u;%u;%u;%u;%u;%3.2f%%",
 580                                         dus->name, dus->ios[0], dus->ios[1],
 581                                         dus->merges[0], dus->merges[1],
 582                                         dus->ticks[0], dus->ticks[1],
 583                                         dus->time_in_queue, util);
 584         }
 585
 586         /*
 587          * If the device has slaves, aggregate the stats for
 588          * those slave devices also.
 589          */
 590         show_agg_stats(agg, terse);
 591
 592         if (!terse)
 593                 log_info("\n");
 594 }
 595
 596 static void print_disk_util_json(struct disk_util *du, struct json_array *array)
 597 {
 598         double util = 0;
 599         struct disk_util_stat *dus = &du->dus;
 600         struct disk_util_agg *agg = &du->agg;
 601         struct json_object *obj;
 602
 603         obj = json_create_object();
 604         json_array_add_value_object(array, obj);
 605
 606         if (dus->msec)
 607                 util = (double) 100 * dus->io_ticks / (double) dus->msec;
 608         if (util > 100.0)
 609                 util = 100.0;
 610
 611
 612         json_object_add_value_string(obj, "name", dus->name);
 613         json_object_add_value_int(obj, "read_ios", dus->ios[0]);
 614         json_object_add_value_int(obj, "write_ios", dus->ios[1]);
 615         json_object_add_value_int(obj, "read_merges", dus->merges[0]);
 616         json_object_add_value_int(obj, "write_merges", dus->merges[1]);
 617         json_object_add_value_int(obj, "read_ticks", dus->ticks[0]);
 618         json_object_add_value_int(obj, "write_ticks", dus->ticks[1]);
 619         json_object_add_value_int(obj, "in_queue", dus->time_in_queue);
 620         json_object_add_value_float(obj, "util", util);
 621
 622         /*
 623          * If the device has slaves, aggregate the stats for
 624          * those slave devices also.
 625          */
 626         if (!agg->slavecount)
 627                 return;
 628         json_object_add_value_int(obj, "aggr_read_ios",
 629                                 agg->ios[0] / agg->slavecount);
 630         json_object_add_value_int(obj, "aggr_write_ios",
 631                                 agg->ios[1] / agg->slavecount);
 632         json_object_add_value_int(obj, "aggr_read_merges",
 633                                 agg->merges[0] / agg->slavecount);
 634         json_object_add_value_int(obj, "aggr_write_merge",
 635                                 agg->merges[1] / agg->slavecount);
 636         json_object_add_value_int(obj, "aggr_read_ticks",
 637                                 agg->ticks[0] / agg->slavecount);
 638         json_object_add_value_int(obj, "aggr_write_ticks",
 639                                 agg->ticks[1] / agg->slavecount);
 640         json_object_add_value_int(obj, "aggr_in_queue",
 641                                 agg->time_in_queue / agg->slavecount);
 642         json_object_add_value_float(obj, "aggr_util", agg->max_util.u.f);
 643 }
 644
 645 void show_disk_util(int terse, struct json_object *parent)
 646 {
 647         struct flist_head *entry;
 648         struct disk_util *du;
 649         struct json_array *array = NULL;
 650
 651         fio_mutex_down(disk_util_mutex);
 652
 653         if (flist_empty(&disk_list)) {
 654                 fio_mutex_up(disk_util_mutex);
 655                 return;
 656         }
 657
 658         if (!terse)
 659                 log_info("\nDisk stats (read/write):\n");
 660
 661         if (output_format == FIO_OUTPUT_JSON) {
 662                 array = json_create_array();
 663                 json_object_add_value_array(parent, "disk_util", array);
 664         }
 665
 666         flist_for_each(entry, &disk_list) {
 667                 du = flist_entry(entry, struct disk_util, list);
 668
 669                 aggregate_slaves_stats(du);
 670                 if (output_format == FIO_OUTPUT_JSON)
 671                         print_disk_util_json(du, array);
 672                 else
 673                         print_disk_util(&du->dus, &du->agg, terse);
 674         }
 675
 676         fio_mutex_up(disk_util_mutex);
 677 }
 678
 679 void setup_disk_util(void)
 680 {
 681         disk_util_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
 682 }