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         fio_mutex_down(disk_util_mutex);
 136
 137         flist_for_each(entry, &disk_list) {
 138                 du = flist_entry(entry, struct disk_util, list);
 139
 140                 if (major == du->major && minor == du->minor) {
 141                         fio_mutex_up(disk_util_mutex);
 142                         return du;
 143                 }
 144         }
 145
 146         fio_mutex_up(disk_util_mutex);
 147         return NULL;
 148 }
 149
 150 static int get_device_numbers(char *file_name, int *maj, int *min)
 151 {
 152         struct stat st;
 153         int majdev, mindev;
 154         char tempname[PATH_MAX], *p;
 155
 156         if (!lstat(file_name, &st)) {
 157                 if (S_ISBLK(st.st_mode)) {
 158                         majdev = major(st.st_rdev);
 159                         mindev = minor(st.st_rdev);
 160                 } else if (S_ISCHR(st.st_mode)) {
 161                         majdev = major(st.st_rdev);
 162                         mindev = minor(st.st_rdev);
 163                         if (fio_lookup_raw(st.st_rdev, &majdev, &mindev))
 164                                 return -1;
 165                 } else if (S_ISFIFO(st.st_mode))
 166                         return -1;
 167                 else {
 168                         majdev = major(st.st_dev);
 169                         mindev = minor(st.st_dev);
 170                 }
 171         } else {
 172                 /*
 173                  * must be a file, open "." in that path
 174                  */
 175                 strncpy(tempname, file_name, PATH_MAX - 1);
 176                 p = dirname(tempname);
 177                 if (stat(p, &st)) {
 178                         perror("disk util stat");
 179                         return -1;
 180                 }
 181
 182                 majdev = major(st.st_dev);
 183                 mindev = minor(st.st_dev);
 184         }
 185
 186         *min = mindev;
 187         *maj = majdev;
 188
 189         return 0;
 190 }
 191
 192 static int read_block_dev_entry(char *path, int *maj, int *min)
 193 {
 194         char line[256], *p;
 195         FILE *f;
 196
 197         f = fopen(path, "r");
 198         if (!f) {
 199                 perror("open path");
 200                 return 1;
 201         }
 202
 203         p = fgets(line, sizeof(line), f);
 204         fclose(f);
 205
 206         if (!p)
 207                 return 1;
 208
 209         if (sscanf(p, "%u:%u", maj, min) != 2)
 210                 return 1;
 211
 212         return 0;
 213 }
 214
 215 static void find_add_disk_slaves(struct thread_data *td, char *path,
 216                                  struct disk_util *masterdu)
 217 {
 218         DIR *dirhandle = NULL;
 219         struct dirent *dirent = NULL;
 220         char slavesdir[PATH_MAX], temppath[PATH_MAX], slavepath[PATH_MAX];
 221         struct disk_util *slavedu = NULL;
 222         int majdev, mindev;
 223         ssize_t linklen;
 224
 225         sprintf(slavesdir, "%s/%s", path, "slaves");
 226         dirhandle = opendir(slavesdir);
 227         if (!dirhandle)
 228                 return;
 229
 230         while ((dirent = readdir(dirhandle)) != NULL) {
 231                 if (!strcmp(dirent->d_name, ".") ||
 232                     !strcmp(dirent->d_name, ".."))
 233                         continue;
 234
 235                 sprintf(temppath, "%s%s%s", slavesdir, FIO_OS_PATH_SEPARATOR, dirent->d_name);
 236                 /* Can we always assume that the slaves device entries
 237                  * are links to the real directories for the slave
 238                  * devices?
 239                  */
 240                 linklen = readlink(temppath, slavepath, PATH_MAX - 0);
 241                 if (linklen  < 0) {
 242                         perror("readlink() for slave device.");
 243                         return;
 244                 }
 245                 slavepath[linklen] = '\0';
 246
 247                 sprintf(temppath, "%s/%s/dev", slavesdir, slavepath);
 248                 if (read_block_dev_entry(temppath, &majdev, &mindev)) {
 249                         perror("Error getting slave device numbers.");
 250                         return;
 251                 }
 252
 253                 /*
 254                  * See if this maj,min already exists
 255                  */
 256                 slavedu = disk_util_exists(majdev, mindev);
 257                 if (slavedu)
 258                         continue;
 259
 260                 sprintf(temppath, "%s%s%s", slavesdir, FIO_OS_PATH_SEPARATOR, slavepath);
 261                 __init_per_file_disk_util(td, majdev, mindev, temppath);
 262                 slavedu = disk_util_exists(majdev, mindev);
 263
 264                 /* Should probably use an assert here. slavedu should
 265                  * always be present at this point. */
 266                 if (slavedu) {
 267                         slavedu->users++;
 268                         flist_add_tail(&slavedu->slavelist, &masterdu->slaves);
 269                 }
 270         }
 271
 272         closedir(dirhandle);
 273 }
 274
 275 static struct disk_util *disk_util_add(struct thread_data *td, int majdev,
 276                                        int mindev, char *path)
 277 {
 278         struct disk_util *du, *__du;
 279         struct flist_head *entry;
 280
 281         dprint(FD_DISKUTIL, "add maj/min %d/%d: %s\n", majdev, mindev, path);
 282
 283         du = smalloc(sizeof(*du));
 284         memset(du, 0, sizeof(*du));
 285         INIT_FLIST_HEAD(&du->list);
 286         sprintf(du->path, "%s/stat", path);
 287         strncpy((char *) du->dus.name, basename(path), FIO_DU_NAME_SZ);
 288         du->sysfs_root = path;
 289         du->major = majdev;
 290         du->minor = mindev;
 291         INIT_FLIST_HEAD(&du->slavelist);
 292         INIT_FLIST_HEAD(&du->slaves);
 293         du->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
 294         du->users = 0;
 295
 296         fio_mutex_down(disk_util_mutex);
 297
 298         flist_for_each(entry, &disk_list) {
 299                 __du = flist_entry(entry, struct disk_util, list);
 300
 301                 dprint(FD_DISKUTIL, "found %s in list\n", __du->dus.name);
 302
 303                 if (!strcmp((char *) du->dus.name, (char *) __du->dus.name)) {
 304                         disk_util_free(du);
 305                         fio_mutex_up(disk_util_mutex);
 306                         return __du;
 307                 }
 308         }
 309
 310         dprint(FD_DISKUTIL, "add %s to list\n", du->dus.name);
 311
 312         fio_gettime(&du->time, NULL);
 313         get_io_ticks(du, &du->last_dus);
 314
 315         flist_add_tail(&du->list, &disk_list);
 316         fio_mutex_up(disk_util_mutex);
 317
 318         find_add_disk_slaves(td, path, du);
 319         return du;
 320 }
 321
 322 static int check_dev_match(int majdev, int mindev, char *path)
 323 {
 324         int major, minor;
 325
 326         if (read_block_dev_entry(path, &major, &minor))
 327                 return 1;
 328
 329         if (majdev == major && mindev == minor)
 330                 return 0;
 331
 332         return 1;
 333 }
 334
 335 static int find_block_dir(int majdev, int mindev, char *path, int link_ok)
 336 {
 337         struct dirent *dir;
 338         struct stat st;
 339         int found = 0;
 340         DIR *D;
 341
 342         D = opendir(path);
 343         if (!D)
 344                 return 0;
 345
 346         while ((dir = readdir(D)) != NULL) {
 347                 char full_path[256];
 348
 349                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
 350                         continue;
 351
 352                 sprintf(full_path, "%s%s%s", path, FIO_OS_PATH_SEPARATOR, dir->d_name);
 353
 354                 if (!strcmp(dir->d_name, "dev")) {
 355                         if (!check_dev_match(majdev, mindev, full_path)) {
 356                                 found = 1;
 357                                 break;
 358                         }
 359                 }
 360
 361                 if (link_ok) {
 362                         if (stat(full_path, &st) == -1) {
 363                                 perror("stat");
 364                                 break;
 365                         }
 366                 } else {
 367                         if (lstat(full_path, &st) == -1) {
 368                                 perror("stat");
 369                                 break;
 370                         }
 371                 }
 372
 373                 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
 374                         continue;
 375
 376                 found = find_block_dir(majdev, mindev, full_path, 0);
 377                 if (found) {
 378                         strcpy(path, full_path);
 379                         break;
 380                 }
 381         }
 382
 383         closedir(D);
 384         return found;
 385 }
 386
 387 static struct disk_util *__init_per_file_disk_util(struct thread_data *td,
 388                                                    int majdev, int mindev,
 389                                                    char *path)
 390 {
 391         struct stat st;
 392         char tmp[PATH_MAX];
 393         char *p;
 394
 395         /*
 396          * If there's a ../queue/ directory there, we are inside a partition.
 397          * Check if that is the case and jump back. For loop/md/dm etc we
 398          * are already in the right spot.
 399          */
 400         sprintf(tmp, "%s/../queue", path);
 401         if (!stat(tmp, &st)) {
 402                 p = dirname(path);
 403                 sprintf(tmp, "%s/queue", p);
 404                 if (stat(tmp, &st)) {
 405                         log_err("unknown sysfs layout\n");
 406                         return NULL;
 407                 }
 408                 strncpy(tmp, p, PATH_MAX - 1);
 409                 sprintf(path, "%s", tmp);
 410         }
 411
 412         if (td->o.ioscheduler && !td->sysfs_root)
 413                 td->sysfs_root = strdup(path);
 414
 415         return disk_util_add(td, majdev, mindev, path);
 416 }
 417
 418 static struct disk_util *init_per_file_disk_util(struct thread_data *td,
 419                                                  char *filename)
 420 {
 421
 422         char foo[PATH_MAX];
 423         struct disk_util *du;
 424         int mindev, majdev;
 425
 426         if (get_device_numbers(filename, &majdev, &mindev))
 427                 return NULL;
 428
 429         dprint(FD_DISKUTIL, "%s belongs to maj/min %d/%d\n", filename, majdev,
 430                         mindev);
 431
 432         du = disk_util_exists(majdev, mindev);
 433         if (du) {
 434                 if (td->o.ioscheduler && !td->sysfs_root)
 435                         td->sysfs_root = strdup(du->sysfs_root);
 436
 437                 return du;
 438         }
 439
 440         /*
 441          * for an fs without a device, we will repeatedly stat through
 442          * sysfs which can take oodles of time for thousands of files. so
 443          * cache the last lookup and compare with that before going through
 444          * everything again.
 445          */
 446         if (mindev == last_mindev && majdev == last_majdev)
 447                 return last_du;
 448
 449         last_mindev = mindev;
 450         last_majdev = majdev;
 451
 452         sprintf(foo, "/sys/block");
 453         if (!find_block_dir(majdev, mindev, foo, 1))
 454                 return NULL;
 455
 456         return __init_per_file_disk_util(td, majdev, mindev, foo);
 457 }
 458
 459 static struct disk_util *__init_disk_util(struct thread_data *td,
 460                                           struct fio_file *f)
 461 {
 462         return init_per_file_disk_util(td, f->file_name);
 463 }
 464
 465 void init_disk_util(struct thread_data *td)
 466 {
 467         struct fio_file *f;
 468         unsigned int i;
 469
 470         if (!td->o.do_disk_util ||
 471             (td->io_ops->flags & (FIO_DISKLESSIO | FIO_NODISKUTIL)))
 472                 return;
 473
 474         for_each_file(td, f, i)
 475                 f->du = __init_disk_util(td, f);
 476 }
 477
 478 static void show_agg_stats(struct disk_util_agg *agg, int terse)
 479 {
 480         if (!agg->slavecount)
 481                 return;
 482
 483         if (!terse) {
 484                 log_info(", aggrios=%u/%u, aggrmerge=%u/%u, aggrticks=%u/%u,"
 485                                 " aggrin_queue=%u, aggrutil=%3.2f%%",
 486                                 agg->ios[0] / agg->slavecount,
 487                                 agg->ios[1] / agg->slavecount,
 488                                 agg->merges[0] / agg->slavecount,
 489                                 agg->merges[1] / agg->slavecount,
 490                                 agg->ticks[0] / agg->slavecount,
 491                                 agg->ticks[1] / agg->slavecount,
 492                                 agg->time_in_queue / agg->slavecount,
 493                                 agg->max_util.u.f);
 494         } else {
 495                 log_info(";slaves;%u;%u;%u;%u;%u;%u;%u;%3.2f%%",
 496                                 agg->ios[0] / agg->slavecount,
 497                                 agg->ios[1] / agg->slavecount,
 498                                 agg->merges[0] / agg->slavecount,
 499                                 agg->merges[1] / agg->slavecount,
 500                                 agg->ticks[0] / agg->slavecount,
 501                                 agg->ticks[1] / agg->slavecount,
 502                                 agg->time_in_queue / agg->slavecount,
 503                                 agg->max_util.u.f);
 504         }
 505 }
 506
 507 static void aggregate_slaves_stats(struct disk_util *masterdu)
 508 {
 509         struct disk_util_agg *agg = &masterdu->agg;
 510         struct disk_util_stat *dus;
 511         struct flist_head *entry;
 512         struct disk_util *slavedu;
 513         double util;
 514
 515         flist_for_each(entry, &masterdu->slaves) {
 516                 slavedu = flist_entry(entry, struct disk_util, slavelist);
 517                 dus = &slavedu->dus;
 518                 agg->ios[0] += dus->ios[0];
 519                 agg->ios[1] += dus->ios[1];
 520                 agg->merges[0] += dus->merges[0];
 521                 agg->merges[1] += dus->merges[1];
 522                 agg->sectors[0] += dus->sectors[0];
 523                 agg->sectors[1] += dus->sectors[1];
 524                 agg->ticks[0] += dus->ticks[0];
 525                 agg->ticks[1] += dus->ticks[1];
 526                 agg->time_in_queue += dus->time_in_queue;
 527                 agg->slavecount++;
 528
 529                 util = (double) (100 * dus->io_ticks / (double) slavedu->dus.msec);
 530                 /* System utilization is the utilization of the
 531                  * component with the highest utilization.
 532                  */
 533                 if (util > agg->max_util.u.f)
 534                         agg->max_util.u.f = util;
 535
 536         }
 537
 538         if (agg->max_util.u.f > 100.0)
 539                 agg->max_util.u.f = 100.0;
 540 }
 541
 542 void free_disk_util(void)
 543 {
 544         struct disk_util *du;
 545
 546         disk_util_exit = 1;
 547         wait_for_disk_thread_exit();
 548
 549         fio_mutex_down(disk_util_mutex);
 550
 551         while (!flist_empty(&disk_list)) {
 552                 du = flist_entry(disk_list.next, struct disk_util, list);
 553                 flist_del(&du->list);
 554                 disk_util_free(du);
 555         }
 556
 557         last_majdev = last_mindev = -1;
 558         fio_mutex_up(disk_util_mutex);
 559         fio_mutex_remove(disk_util_mutex);
 560 }
 561
 562 void print_disk_util(struct disk_util_stat *dus, struct disk_util_agg *agg,
 563                      int terse)
 564 {
 565         double util = 0;
 566
 567         if (dus->msec)
 568                 util = (double) 100 * dus->io_ticks / (double) dus->msec;
 569         if (util > 100.0)
 570                 util = 100.0;
 571
 572         if (!terse) {
 573                 if (agg->slavecount)
 574                         log_info("  ");
 575
 576                 log_info("  %s: ios=%u/%u, merge=%u/%u, ticks=%u/%u, "
 577                          "in_queue=%u, util=%3.2f%%", dus->name,
 578                                         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         } else {
 583                 log_info(";%s;%u;%u;%u;%u;%u;%u;%u;%3.2f%%",
 584                                         dus->name, dus->ios[0], dus->ios[1],
 585                                         dus->merges[0], dus->merges[1],
 586                                         dus->ticks[0], dus->ticks[1],
 587                                         dus->time_in_queue, util);
 588         }
 589
 590         /*
 591          * If the device has slaves, aggregate the stats for
 592          * those slave devices also.
 593          */
 594         show_agg_stats(agg, terse);
 595
 596         if (!terse)
 597                 log_info("\n");
 598 }
 599
 600 void show_disk_util(int terse)
 601 {
 602         struct flist_head *entry;
 603         struct disk_util *du;
 604
 605         fio_mutex_down(disk_util_mutex);
 606
 607         if (flist_empty(&disk_list)) {
 608                 fio_mutex_up(disk_util_mutex);
 609                 return;
 610         }
 611
 612         if (!terse)
 613                 log_info("\nDisk stats (read/write):\n");
 614
 615         flist_for_each(entry, &disk_list) {
 616                 du = flist_entry(entry, struct disk_util, list);
 617
 618                 aggregate_slaves_stats(du);
 619                 print_disk_util(&du->dus, &du->agg, terse);
 620         }
 621
 622         fio_mutex_up(disk_util_mutex);
 623 }
 624
 625 void setup_disk_util(void)
 626 {
 627         disk_util_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
 628 }