fio.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
   6  *
   7  * The license below covers all files distributed with fio unless otherwise
   8  * noted in the file itself.
   9  *
  10  *  This program is free software; you can redistribute it and/or modify
  11  *  it under the terms of the GNU General Public License version 2 as
  12  *  published by the Free Software Foundation.
  13  *
  14  *  This program is distributed in the hope that it will be useful,
  15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *  GNU General Public License for more details.
  18  *
  19  *  You should have received a copy of the GNU General Public License
  20  *  along with this program; if not, write to the Free Software
  21  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <fcntl.h>
  26 #include <string.h>
  27 #include <signal.h>
  28 #include <time.h>
  29 #include <locale.h>
  30 #include <assert.h>
  31 #include <sys/stat.h>
  32 #include <sys/wait.h>
  33 #include <sys/ipc.h>
  34 #include <sys/shm.h>
  35 #include <sys/mman.h>
  36
  37 #include "fio.h"
  38 #include "os.h"
  39
  40 static unsigned long page_mask;
  41 #define ALIGN(buf)      \
  42         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  43
  44 int groupid = 0;
  45 int thread_number = 0;
  46 int shm_id = 0;
  47 int temp_stall_ts;
  48
  49 static volatile int startup_sem;
  50 static volatile int fio_abort;
  51 static int exit_value;
  52
  53 struct io_log *agg_io_log[2];
  54
  55 #define TERMINATE_ALL           (-1)
  56 #define JOB_START_TIMEOUT       (5 * 1000)
  57
  58 static inline void td_set_runstate(struct thread_data *td, int runstate)
  59 {
  60         td->runstate = runstate;
  61 }
  62
  63 static void terminate_threads(int group_id, int forced_kill)
  64 {
  65         struct thread_data *td;
  66         int i;
  67
  68         for_each_td(td, i) {
  69                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  70                         td->terminate = 1;
  71                         td->start_delay = 0;
  72                         if (forced_kill)
  73                                 td_set_runstate(td, TD_EXITED);
  74                 }
  75         }
  76 }
  77
  78 static void sig_handler(int sig)
  79 {
  80         switch (sig) {
  81                 case SIGALRM:
  82                         update_io_ticks();
  83                         disk_util_timer_arm();
  84                         print_thread_status();
  85                         break;
  86                 default:
  87                         printf("\nfio: terminating on signal %d\n", sig);
  88                         fflush(stdout);
  89                         terminate_threads(TERMINATE_ALL, 0);
  90                         break;
  91         }
  92 }
  93
  94 /*
  95  * Check if we are above the minimum rate given.
  96  */
  97 static int check_min_rate(struct thread_data *td, struct timeval *now)
  98 {
  99         unsigned long long bytes = 0;
 100         unsigned long spent;
 101         unsigned long rate;
 102
 103         /*
 104          * allow a 2 second settle period in the beginning
 105          */
 106         if (mtime_since(&td->start, now) < 2000)
 107                 return 0;
 108
 109         if (td_read(td))
 110                 bytes += td->this_io_bytes[DDIR_READ];
 111         if (td_write(td))
 112                 bytes += td->this_io_bytes[DDIR_WRITE];
 113
 114         /*
 115          * if rate blocks is set, sample is running
 116          */
 117         if (td->rate_bytes) {
 118                 spent = mtime_since(&td->lastrate, now);
 119                 if (spent < td->ratecycle)
 120                         return 0;
 121
 122                 if (bytes < td->rate_bytes) {
 123                         fprintf(f_out, "%s: min rate %u not met\n", td->name, td->ratemin);
 124                         return 1;
 125                 } else {
 126                         rate = (bytes - td->rate_bytes) / spent;
 127                         if (rate < td->ratemin || bytes < td->rate_bytes) {
 128                                 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
 129                                 return 1;
 130                         }
 131                 }
 132         }
 133
 134         td->rate_bytes = bytes;
 135         memcpy(&td->lastrate, now, sizeof(*now));
 136         return 0;
 137 }
 138
 139 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 140 {
 141         if (!td->timeout)
 142                 return 0;
 143         if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
 144                 return 1;
 145
 146         return 0;
 147 }
 148
 149 /*
 150  * When job exits, we can cancel the in-flight IO if we are using async
 151  * io. Attempt to do so.
 152  */
 153 static void cleanup_pending_aio(struct thread_data *td)
 154 {
 155         struct list_head *entry, *n;
 156         struct io_u *io_u;
 157         int r;
 158
 159         /*
 160          * get immediately available events, if any
 161          */
 162         r = io_u_queued_complete(td, 0);
 163         if (r < 0)
 164                 return;
 165
 166         /*
 167          * now cancel remaining active events
 168          */
 169         if (td->io_ops->cancel) {
 170                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 171                         io_u = list_entry(entry, struct io_u, list);
 172
 173                         /*
 174                          * if the io_u isn't in flight, then that generally
 175                          * means someone leaked an io_u. complain but fix
 176                          * it up, so we don't stall here.
 177                          */
 178                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 179                                 log_err("fio: non-busy IO on busy list\n");
 180                                 put_io_u(td, io_u);
 181                         } else {
 182                                 r = td->io_ops->cancel(td, io_u);
 183                                 if (!r)
 184                                         put_io_u(td, io_u);
 185                         }
 186                 }
 187         }
 188
 189         if (td->cur_depth)
 190                 r = io_u_queued_complete(td, td->cur_depth);
 191 }
 192
 193 /*
 194  * Helper to handle the final sync of a file. Works just like the normal
 195  * io path, just does everything sync.
 196  */
 197 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 198 {
 199         struct io_u *io_u = __get_io_u(td);
 200         int ret;
 201
 202         if (!io_u)
 203                 return 1;
 204
 205         io_u->ddir = DDIR_SYNC;
 206         io_u->file = f;
 207
 208         if (td_io_prep(td, io_u)) {
 209                 put_io_u(td, io_u);
 210                 return 1;
 211         }
 212
 213 requeue:
 214         ret = td_io_queue(td, io_u);
 215         if (ret < 0) {
 216                 td_verror(td, io_u->error, "td_io_queue");
 217                 put_io_u(td, io_u);
 218                 return 1;
 219         } else if (ret == FIO_Q_QUEUED) {
 220                 if (io_u_queued_complete(td, 1) < 0)
 221                         return 1;
 222         } else if (ret == FIO_Q_COMPLETED) {
 223                 if (io_u->error) {
 224                         td_verror(td, io_u->error, "td_io_queue");
 225                         return 1;
 226                 }
 227
 228                 if (io_u_sync_complete(td, io_u) < 0)
 229                         return 1;
 230         } else if (ret == FIO_Q_BUSY) {
 231                 if (td_io_commit(td))
 232                         return 1;
 233                 goto requeue;
 234         }
 235
 236         return 0;
 237 }
 238
 239 /*
 240  * The main verify engine. Runs over the writes we previusly submitted,
 241  * reads the blocks back in, and checks the crc/md5 of the data.
 242  */
 243 static void do_verify(struct thread_data *td)
 244 {
 245         struct fio_file *f;
 246         struct io_u *io_u;
 247         int ret, i, min_events;
 248
 249         /*
 250          * sync io first and invalidate cache, to make sure we really
 251          * read from disk.
 252          */
 253         for_each_file(td, f, i) {
 254                 if (fio_io_sync(td, f))
 255                         break;
 256                 if (file_invalidate_cache(td, f))
 257                         break;
 258         }
 259
 260         if (td->error)
 261                 return;
 262
 263         td_set_runstate(td, TD_VERIFYING);
 264
 265         io_u = NULL;
 266         while (!td->terminate) {
 267                 io_u = __get_io_u(td);
 268                 if (!io_u)
 269                         break;
 270
 271                 if (runtime_exceeded(td, &io_u->start_time)) {
 272                         put_io_u(td, io_u);
 273                         break;
 274                 }
 275
 276                 if (get_next_verify(td, io_u)) {
 277                         put_io_u(td, io_u);
 278                         break;
 279                 }
 280
 281                 if (td_io_prep(td, io_u)) {
 282                         put_io_u(td, io_u);
 283                         break;
 284                 }
 285
 286                 io_u->end_io = verify_io_u;
 287 requeue:
 288                 ret = td_io_queue(td, io_u);
 289
 290                 switch (ret) {
 291                 case FIO_Q_COMPLETED:
 292                         if (io_u->error)
 293                                 ret = -io_u->error;
 294                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
 295                                 int bytes = io_u->xfer_buflen - io_u->resid;
 296
 297                                 io_u->xfer_buflen = io_u->resid;
 298                                 io_u->xfer_buf += bytes;
 299                                 goto requeue;
 300                         }
 301                         ret = io_u_sync_complete(td, io_u);
 302                         if (ret < 0)
 303                                 break;
 304                         continue;
 305                 case FIO_Q_QUEUED:
 306                         break;
 307                 case FIO_Q_BUSY:
 308                         requeue_io_u(td, &io_u);
 309                         ret = td_io_commit(td);
 310                         break;
 311                 default:
 312                         assert(ret < 0);
 313                         td_verror(td, -ret, "td_io_queue");
 314                         break;
 315                 }
 316
 317                 if (ret < 0 || td->error)
 318                         break;
 319
 320                 /*
 321                  * if we can queue more, do so. but check if there are
 322                  * completed io_u's first.
 323                  */
 324                 min_events = 0;
 325                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 326                         min_events = 1;
 327
 328                         if (td->cur_depth > td->iodepth_low)
 329                                 min_events = td->cur_depth - td->iodepth_low;
 330                 }
 331
 332                 /*
 333                  * Reap required number of io units, if any, and do the
 334                  * verification on them through the callback handler
 335                  */
 336                 if (io_u_queued_complete(td, min_events) < 0)
 337                         break;
 338         }
 339
 340         if (td->cur_depth)
 341                 cleanup_pending_aio(td);
 342
 343         td_set_runstate(td, TD_RUNNING);
 344 }
 345
 346 /*
 347  * Not really an io thread, all it does is burn CPU cycles in the specified
 348  * manner.
 349  */
 350 static void do_cpuio(struct thread_data *td)
 351 {
 352         struct timeval e;
 353         int split = 100 / td->cpuload;
 354         int i = 0;
 355
 356         while (!td->terminate) {
 357                 fio_gettime(&e, NULL);
 358
 359                 if (runtime_exceeded(td, &e))
 360                         break;
 361
 362                 if (!(i % split))
 363                         __usec_sleep(10000);
 364                 else
 365                         usec_sleep(td, 10000);
 366
 367                 i++;
 368         }
 369 }
 370
 371 /*
 372  * Main IO worker function. It retrieves io_u's to process and queues
 373  * and reaps them, checking for rate and errors along the way.
 374  */
 375 static void do_io(struct thread_data *td)
 376 {
 377         struct timeval s;
 378         unsigned long usec;
 379         int i, ret = 0;
 380
 381         td_set_runstate(td, TD_RUNNING);
 382
 383         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
 384                 struct timeval comp_time;
 385                 long bytes_done = 0;
 386                 int min_evts = 0;
 387                 struct io_u *io_u;
 388
 389                 if (td->terminate)
 390                         break;
 391
 392                 io_u = get_io_u(td);
 393                 if (!io_u)
 394                         break;
 395
 396                 memcpy(&s, &io_u->start_time, sizeof(s));
 397
 398                 if (runtime_exceeded(td, &s)) {
 399                         put_io_u(td, io_u);
 400                         break;
 401                 }
 402 requeue:
 403                 ret = td_io_queue(td, io_u);
 404
 405                 switch (ret) {
 406                 case FIO_Q_COMPLETED:
 407                         if (io_u->error) {
 408                                 ret = io_u->error;
 409                                 break;
 410                         }
 411                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
 412                                 int bytes = io_u->xfer_buflen - io_u->resid;
 413
 414                                 io_u->xfer_buflen = io_u->resid;
 415                                 io_u->xfer_buf += bytes;
 416                                 goto requeue;
 417                         }
 418                         fio_gettime(&comp_time, NULL);
 419                         bytes_done = io_u_sync_complete(td, io_u);
 420                         if (bytes_done < 0)
 421                                 ret = bytes_done;
 422                         break;
 423                 case FIO_Q_QUEUED:
 424                         /*
 425                          * if the engine doesn't have a commit hook,
 426                          * the io_u is really queued. if it does have such
 427                          * a hook, it has to call io_u_queued() itself.
 428                          */
 429                         if (td->io_ops->commit == NULL)
 430                                 io_u_queued(td, io_u);
 431                         break;
 432                 case FIO_Q_BUSY:
 433                         requeue_io_u(td, &io_u);
 434                         ret = td_io_commit(td);
 435                         break;
 436                 default:
 437                         assert(ret < 0);
 438                         put_io_u(td, io_u);
 439                         break;
 440                 }
 441
 442                 if (ret < 0 || td->error)
 443                         break;
 444
 445                 /*
 446                  * See if we need to complete some commands
 447                  */
 448                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 449                         min_evts = 0;
 450                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 451                                 min_evts = 1;
 452
 453                                 if (td->cur_depth > td->iodepth_low)
 454                                         min_evts = td->cur_depth - td->iodepth_low;
 455                         }
 456
 457                         fio_gettime(&comp_time, NULL);
 458                         bytes_done = io_u_queued_complete(td, min_evts);
 459                         if (bytes_done < 0)
 460                                 break;
 461                 }
 462
 463                 if (!bytes_done)
 464                         continue;
 465
 466                 /*
 467                  * the rate is batched for now, it should work for batches
 468                  * of completions except the very first one which may look
 469                  * a little bursty
 470                  */
 471                 usec = utime_since(&s, &comp_time);
 472
 473                 rate_throttle(td, usec, bytes_done);
 474
 475                 if (check_min_rate(td, &comp_time)) {
 476                         if (exitall_on_terminate)
 477                                 terminate_threads(td->groupid, 0);
 478                         td_verror(td, ENODATA, "check_min_rate");
 479                         break;
 480                 }
 481
 482                 if (td->thinktime) {
 483                         unsigned long long b;
 484
 485                         b = td->io_blocks[0] + td->io_blocks[1];
 486                         if (!(b % td->thinktime_blocks)) {
 487                                 int left;
 488
 489                                 if (td->thinktime_spin)
 490                                         __usec_sleep(td->thinktime_spin);
 491
 492                                 left = td->thinktime - td->thinktime_spin;
 493                                 if (left)
 494                                         usec_sleep(td, left);
 495                         }
 496                 }
 497         }
 498
 499         if (!td->error) {
 500                 struct fio_file *f;
 501
 502                 if (td->cur_depth)
 503                         cleanup_pending_aio(td);
 504
 505                 if (should_fsync(td) && td->end_fsync) {
 506                         td_set_runstate(td, TD_FSYNCING);
 507                         for_each_file(td, f, i)
 508                                 fio_io_sync(td, f);
 509                 }
 510         }
 511 }
 512
 513 static void cleanup_io_u(struct thread_data *td)
 514 {
 515         struct list_head *entry, *n;
 516         struct io_u *io_u;
 517
 518         list_for_each_safe(entry, n, &td->io_u_freelist) {
 519                 io_u = list_entry(entry, struct io_u, list);
 520
 521                 list_del(&io_u->list);
 522                 free(io_u);
 523         }
 524
 525         free_io_mem(td);
 526 }
 527
 528 /*
 529  * "randomly" fill the buffer contents
 530  */
 531 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 532 {
 533         int *ptr = io_u->buf;
 534
 535         while ((void *) ptr - io_u->buf < max_bs) {
 536                 *ptr = rand() * 0x9e370001;
 537                 ptr++;
 538         }
 539 }
 540
 541 static int init_io_u(struct thread_data *td)
 542 {
 543         struct io_u *io_u;
 544         unsigned int max_bs;
 545         int i, max_units;
 546         char *p;
 547
 548         if (td->io_ops->flags & FIO_CPUIO)
 549                 return 0;
 550
 551         if (td->io_ops->flags & FIO_SYNCIO)
 552                 max_units = 1;
 553         else
 554                 max_units = td->iodepth;
 555
 556         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
 557         td->orig_buffer_size = max_bs * max_units;
 558
 559         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
 560                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
 561         else
 562                 td->orig_buffer_size += page_mask;
 563
 564         if (allocate_io_mem(td))
 565                 return 1;
 566
 567         p = ALIGN(td->orig_buffer);
 568         for (i = 0; i < max_units; i++) {
 569                 io_u = malloc(sizeof(*io_u));
 570                 memset(io_u, 0, sizeof(*io_u));
 571                 INIT_LIST_HEAD(&io_u->list);
 572
 573                 io_u->buf = p + max_bs * i;
 574                 if (td_write(td) || td_rw(td))
 575                         fill_rand_buf(io_u, max_bs);
 576
 577                 io_u->index = i;
 578                 io_u->flags = IO_U_F_FREE;
 579                 list_add(&io_u->list, &td->io_u_freelist);
 580         }
 581
 582         io_u_init_timeout();
 583
 584         return 0;
 585 }
 586
 587 static int switch_ioscheduler(struct thread_data *td)
 588 {
 589         char tmp[256], tmp2[128];
 590         FILE *f;
 591         int ret;
 592
 593         if (td->io_ops->flags & FIO_CPUIO)
 594                 return 0;
 595
 596         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 597
 598         f = fopen(tmp, "r+");
 599         if (!f) {
 600                 td_verror(td, errno, "fopen");
 601                 return 1;
 602         }
 603
 604         /*
 605          * Set io scheduler.
 606          */
 607         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
 608         if (ferror(f) || ret != 1) {
 609                 td_verror(td, errno, "fwrite");
 610                 fclose(f);
 611                 return 1;
 612         }
 613
 614         rewind(f);
 615
 616         /*
 617          * Read back and check that the selected scheduler is now the default.
 618          */
 619         ret = fread(tmp, 1, sizeof(tmp), f);
 620         if (ferror(f) || ret < 0) {
 621                 td_verror(td, errno, "fread");
 622                 fclose(f);
 623                 return 1;
 624         }
 625
 626         sprintf(tmp2, "[%s]", td->ioscheduler);
 627         if (!strstr(tmp, tmp2)) {
 628                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
 629                 td_verror(td, EINVAL, "iosched_switch");
 630                 fclose(f);
 631                 return 1;
 632         }
 633
 634         fclose(f);
 635         return 0;
 636 }
 637
 638 static void clear_io_state(struct thread_data *td)
 639 {
 640         struct fio_file *f;
 641         int i;
 642
 643         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 644         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 645         td->zone_bytes = 0;
 646
 647         td->last_was_sync = 0;
 648
 649         for_each_file(td, f, i) {
 650                 f->last_completed_pos = 0;
 651
 652                 f->last_pos = 0;
 653                 if (td->io_ops->flags & FIO_SYNCIO)
 654                         lseek(f->fd, SEEK_SET, 0);
 655
 656                 if (f->file_map)
 657                         memset(f->file_map, 0, f->num_maps * sizeof(long));
 658         }
 659 }
 660
 661 /*
 662  * Entry point for the thread based jobs. The process based jobs end up
 663  * here as well, after a little setup.
 664  */
 665 static void *thread_main(void *data)
 666 {
 667         unsigned long long runtime[2];
 668         struct thread_data *td = data;
 669
 670         if (!td->use_thread)
 671                 setsid();
 672
 673         td->pid = getpid();
 674
 675         INIT_LIST_HEAD(&td->io_u_freelist);
 676         INIT_LIST_HEAD(&td->io_u_busylist);
 677         INIT_LIST_HEAD(&td->io_u_requeues);
 678         INIT_LIST_HEAD(&td->io_hist_list);
 679         INIT_LIST_HEAD(&td->io_log_list);
 680
 681         if (init_io_u(td))
 682                 goto err;
 683
 684         if (fio_setaffinity(td) == -1) {
 685                 td_verror(td, errno, "cpu_set_affinity");
 686                 goto err;
 687         }
 688
 689         if (init_iolog(td))
 690                 goto err;
 691
 692         if (td->ioprio) {
 693                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 694                         td_verror(td, errno, "ioprio_set");
 695                         goto err;
 696                 }
 697         }
 698
 699         if (nice(td->nice) == -1) {
 700                 td_verror(td, errno, "nice");
 701                 goto err;
 702         }
 703
 704         if (init_random_state(td))
 705                 goto err;
 706
 707         if (td->ioscheduler && switch_ioscheduler(td))
 708                 goto err;
 709
 710         td_set_runstate(td, TD_INITIALIZED);
 711         fio_sem_up(&startup_sem);
 712         fio_sem_down(&td->mutex);
 713
 714         if (!td->create_serialize && setup_files(td))
 715                 goto err;
 716         if (open_files(td))
 717                 goto err;
 718
 719         /*
 720          * Do this late, as some IO engines would like to have the
 721          * files setup prior to initializing structures.
 722          */
 723         if (td_io_init(td))
 724                 goto err;
 725
 726         if (td->exec_prerun) {
 727                 if (system(td->exec_prerun) < 0)
 728                         goto err;
 729         }
 730
 731         fio_gettime(&td->epoch, NULL);
 732         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 733         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 734
 735         runtime[0] = runtime[1] = 0;
 736         while (td->loops--) {
 737                 fio_gettime(&td->start, NULL);
 738                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 739
 740                 if (td->ratemin)
 741                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 742
 743                 clear_io_state(td);
 744                 prune_io_piece_log(td);
 745
 746                 if (td->io_ops->flags & FIO_CPUIO)
 747                         do_cpuio(td);
 748                 else
 749                         do_io(td);
 750
 751                 if (td_read(td) && td->io_bytes[DDIR_READ])
 752                         runtime[DDIR_READ] += utime_since_now(&td->start);
 753                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
 754                         runtime[DDIR_WRITE] += utime_since_now(&td->start);
 755
 756                 if (td->error || td->terminate)
 757                         break;
 758
 759                 if (td->verify == VERIFY_NONE)
 760                         continue;
 761
 762                 clear_io_state(td);
 763                 fio_gettime(&td->start, NULL);
 764
 765                 do_verify(td);
 766
 767                 runtime[DDIR_READ] += utime_since_now(&td->start);
 768
 769                 if (td->error || td->terminate)
 770                         break;
 771         }
 772
 773         update_rusage_stat(td);
 774         fio_gettime(&td->end_time, NULL);
 775         td->runtime[0] = runtime[0] / 1000;
 776         td->runtime[1] = runtime[1] / 1000;
 777
 778         if (td->ts.bw_log)
 779                 finish_log(td, td->ts.bw_log, "bw");
 780         if (td->ts.slat_log)
 781                 finish_log(td, td->ts.slat_log, "slat");
 782         if (td->ts.clat_log)
 783                 finish_log(td, td->ts.clat_log, "clat");
 784         if (td->write_iolog_file)
 785                 write_iolog_close(td);
 786         if (td->exec_postrun) {
 787                 if (system(td->exec_postrun) < 0)
 788                         log_err("fio: postrun %s failed\n", td->exec_postrun);
 789         }
 790
 791         if (exitall_on_terminate)
 792                 terminate_threads(td->groupid, 0);
 793
 794 err:
 795         if (td->error)
 796                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 797         close_files(td);
 798         close_ioengine(td);
 799         cleanup_io_u(td);
 800         td_set_runstate(td, TD_EXITED);
 801         return (void *) (unsigned long) td->error;
 802 }
 803
 804 /*
 805  * We cannot pass the td data into a forked process, so attach the td and
 806  * pass it to the thread worker.
 807  */
 808 static int fork_main(int shmid, int offset)
 809 {
 810         struct thread_data *td;
 811         void *data, *ret;
 812
 813         data = shmat(shmid, NULL, 0);
 814         if (data == (void *) -1) {
 815                 int __err = errno;
 816
 817                 perror("shmat");
 818                 return __err;
 819         }
 820
 821         td = data + offset * sizeof(struct thread_data);
 822         ret = thread_main(td);
 823         shmdt(data);
 824         return (int) (unsigned long) ret;
 825 }
 826
 827 /*
 828  * Run over the job map and reap the threads that have exited, if any.
 829  */
 830 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 831 {
 832         struct thread_data *td;
 833         int i, cputhreads, pending, status, ret;
 834
 835         /*
 836          * reap exited threads (TD_EXITED -> TD_REAPED)
 837          */
 838         pending = cputhreads = 0;
 839         for_each_td(td, i) {
 840                 int flags = 0;
 841
 842                 /*
 843                  * ->io_ops is NULL for a thread that has closed its
 844                  * io engine
 845                  */
 846                 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
 847                         cputhreads++;
 848
 849                 if (!td->pid || td->runstate == TD_REAPED)
 850                         continue;
 851                 if (td->use_thread) {
 852                         if (td->runstate == TD_EXITED) {
 853                                 td_set_runstate(td, TD_REAPED);
 854                                 goto reaped;
 855                         }
 856                         continue;
 857                 }
 858
 859                 flags = WNOHANG;
 860                 if (td->runstate == TD_EXITED)
 861                         flags = 0;
 862
 863                 /*
 864                  * check if someone quit or got killed in an unusual way
 865                  */
 866                 ret = waitpid(td->pid, &status, flags);
 867                 if (ret < 0) {
 868                         if (errno == ECHILD) {
 869                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
 870                                 td_set_runstate(td, TD_REAPED);
 871                                 goto reaped;
 872                         }
 873                         perror("waitpid");
 874                 } else if (ret == td->pid) {
 875                         if (WIFSIGNALED(status)) {
 876                                 int sig = WTERMSIG(status);
 877
 878                                 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
 879                                 td_set_runstate(td, TD_REAPED);
 880                                 goto reaped;
 881                         }
 882                         if (WIFEXITED(status)) {
 883                                 if (WEXITSTATUS(status) && !td->error)
 884                                         td->error = WEXITSTATUS(status);
 885
 886                                 td_set_runstate(td, TD_REAPED);
 887                                 goto reaped;
 888                         }
 889                 }
 890
 891                 /*
 892                  * thread is not dead, continue
 893                  */
 894                 continue;
 895 reaped:
 896                 if (td->use_thread) {
 897                         long ret;
 898
 899                         if (pthread_join(td->thread, (void *) &ret))
 900                                 perror("pthread_join");
 901                 }
 902
 903                 (*nr_running)--;
 904                 (*m_rate) -= td->ratemin;
 905                 (*t_rate) -= td->rate;
 906
 907                 if (td->error)
 908                         exit_value++;
 909         }
 910
 911         if (*nr_running == cputhreads && !pending)
 912                 terminate_threads(TERMINATE_ALL, 0);
 913 }
 914
 915 /*
 916  * Main function for kicking off and reaping jobs, as needed.
 917  */
 918 static void run_threads(void)
 919 {
 920         struct thread_data *td;
 921         unsigned long spent;
 922         int i, todo, nr_running, m_rate, t_rate, nr_started;
 923
 924         if (fio_pin_memory())
 925                 return;
 926
 927         if (!terse_output) {
 928                 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
 929                 fflush(stdout);
 930         }
 931
 932         signal(SIGINT, sig_handler);
 933         signal(SIGALRM, sig_handler);
 934
 935         todo = thread_number;
 936         nr_running = 0;
 937         nr_started = 0;
 938         m_rate = t_rate = 0;
 939
 940         for_each_td(td, i) {
 941                 print_status_init(td->thread_number - 1);
 942
 943                 if (!td->create_serialize) {
 944                         init_disk_util(td);
 945                         continue;
 946                 }
 947
 948                 /*
 949                  * do file setup here so it happens sequentially,
 950                  * we don't want X number of threads getting their
 951                  * client data interspersed on disk
 952                  */
 953                 if (setup_files(td)) {
 954                         exit_value++;
 955                         if (td->error)
 956                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 957                         td_set_runstate(td, TD_REAPED);
 958                         todo--;
 959                 }
 960
 961                 init_disk_util(td);
 962         }
 963
 964         set_genesis_time();
 965
 966         while (todo) {
 967                 struct thread_data *map[MAX_JOBS];
 968                 struct timeval this_start;
 969                 int this_jobs = 0, left;
 970
 971                 /*
 972                  * create threads (TD_NOT_CREATED -> TD_CREATED)
 973                  */
 974                 for_each_td(td, i) {
 975                         if (td->runstate != TD_NOT_CREATED)
 976                                 continue;
 977
 978                         /*
 979                          * never got a chance to start, killed by other
 980                          * thread for some reason
 981                          */
 982                         if (td->terminate) {
 983                                 todo--;
 984                                 continue;
 985                         }
 986
 987                         if (td->start_delay) {
 988                                 spent = mtime_since_genesis();
 989
 990                                 if (td->start_delay * 1000 > spent)
 991                                         continue;
 992                         }
 993
 994                         if (td->stonewall && (nr_started || nr_running))
 995                                 break;
 996
 997                         /*
 998                          * Set state to created. Thread will transition
 999                          * to TD_INITIALIZED when it's done setting up.
1000                          */
1001                         td_set_runstate(td, TD_CREATED);
1002                         map[this_jobs++] = td;
1003                         fio_sem_init(&startup_sem, 1);
1004                         nr_started++;
1005
1006                         if (td->use_thread) {
1007                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1008                                         perror("thread_create");
1009                                         nr_started--;
1010                                 }
1011                         } else {
1012                                 if (fork())
1013                                         fio_sem_down(&startup_sem);
1014                                 else {
1015                                         int ret = fork_main(shm_id, i);
1016
1017                                         exit(ret);
1018                                 }
1019                         }
1020                 }
1021
1022                 /*
1023                  * Wait for the started threads to transition to
1024                  * TD_INITIALIZED.
1025                  */
1026                 fio_gettime(&this_start, NULL);
1027                 left = this_jobs;
1028                 while (left && !fio_abort) {
1029                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1030                                 break;
1031
1032                         usleep(100000);
1033
1034                         for (i = 0; i < this_jobs; i++) {
1035                                 td = map[i];
1036                                 if (!td)
1037                                         continue;
1038                                 if (td->runstate == TD_INITIALIZED) {
1039                                         map[i] = NULL;
1040                                         left--;
1041                                 } else if (td->runstate >= TD_EXITED) {
1042                                         map[i] = NULL;
1043                                         left--;
1044                                         todo--;
1045                                         nr_running++; /* work-around... */
1046                                 }
1047                         }
1048                 }
1049
1050                 if (left) {
1051                         log_err("fio: %d jobs failed to start\n", left);
1052                         for (i = 0; i < this_jobs; i++) {
1053                                 td = map[i];
1054                                 if (!td)
1055                                         continue;
1056                                 kill(td->pid, SIGTERM);
1057                         }
1058                         break;
1059                 }
1060
1061                 /*
1062                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1063                  */
1064                 for_each_td(td, i) {
1065                         if (td->runstate != TD_INITIALIZED)
1066                                 continue;
1067
1068                         td_set_runstate(td, TD_RUNNING);
1069                         nr_running++;
1070                         nr_started--;
1071                         m_rate += td->ratemin;
1072                         t_rate += td->rate;
1073                         todo--;
1074                         fio_sem_up(&td->mutex);
1075                 }
1076
1077                 reap_threads(&nr_running, &t_rate, &m_rate);
1078
1079                 if (todo)
1080                         usleep(100000);
1081         }
1082
1083         while (nr_running) {
1084                 reap_threads(&nr_running, &t_rate, &m_rate);
1085                 usleep(10000);
1086         }
1087
1088         update_io_ticks();
1089         fio_unpin_memory();
1090 }
1091
1092 int main(int argc, char *argv[])
1093 {
1094         long ps;
1095
1096         /*
1097          * We need locale for number printing, if it isn't set then just
1098          * go with the US format.
1099          */
1100         if (!getenv("LC_NUMERIC"))
1101                 setlocale(LC_NUMERIC, "en_US");
1102
1103         if (parse_options(argc, argv))
1104                 return 1;
1105
1106         if (!thread_number) {
1107                 log_err("Nothing to do\n");
1108                 return 1;
1109         }
1110
1111         ps = sysconf(_SC_PAGESIZE);
1112         if (ps < 0) {
1113                 log_err("Failed to get page size\n");
1114                 return 1;
1115         }
1116
1117         page_mask = ps - 1;
1118
1119         if (write_bw_log) {
1120                 setup_log(&agg_io_log[DDIR_READ]);
1121                 setup_log(&agg_io_log[DDIR_WRITE]);
1122         }
1123
1124         set_genesis_time();
1125
1126         disk_util_timer_arm();
1127
1128         run_threads();
1129
1130         if (!fio_abort) {
1131                 show_run_stats();
1132                 if (write_bw_log) {
1133                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1134                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1135                 }
1136         }
1137
1138         return exit_value;
1139 }