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, NULL);
 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, NULL);
 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, NULL) < 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, NULL) < 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 requeue:
 286                 ret = td_io_queue(td, io_u);
 287
 288                 switch (ret) {
 289                 case FIO_Q_COMPLETED:
 290                         if (io_u->error)
 291                                 ret = -io_u->error;
 292                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
 293                                 int bytes = io_u->xfer_buflen - io_u->resid;
 294
 295                                 io_u->xfer_buflen = io_u->resid;
 296                                 io_u->xfer_buf += bytes;
 297                                 goto requeue;
 298                         }
 299                         ret = io_u_sync_complete(td, io_u, verify_io_u);
 300                         if (ret < 0)
 301                                 break;
 302                         continue;
 303                 case FIO_Q_QUEUED:
 304                         break;
 305                 case FIO_Q_BUSY:
 306                         requeue_io_u(td, &io_u);
 307                         ret = td_io_commit(td);
 308                         break;
 309                 default:
 310                         assert(ret < 0);
 311                         td_verror(td, -ret, "td_io_queue");
 312                         break;
 313                 }
 314
 315                 if (ret < 0 || td->error)
 316                         break;
 317
 318                 /*
 319                  * if we can queue more, do so. but check if there are
 320                  * completed io_u's first.
 321                  */
 322                 min_events = 0;
 323                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 324                         min_events = 1;
 325
 326                         if (td->cur_depth > td->iodepth_low)
 327                                 min_events = td->cur_depth - td->iodepth_low;
 328                 }
 329
 330                 /*
 331                  * Reap required number of io units, if any, and do the
 332                  * verification on them through the callback handler
 333                  */
 334                 if (io_u_queued_complete(td, min_events, verify_io_u) < 0)
 335                         break;
 336         }
 337
 338         if (td->cur_depth)
 339                 cleanup_pending_aio(td);
 340
 341         td_set_runstate(td, TD_RUNNING);
 342 }
 343
 344 /*
 345  * Not really an io thread, all it does is burn CPU cycles in the specified
 346  * manner.
 347  */
 348 static void do_cpuio(struct thread_data *td)
 349 {
 350         struct timeval e;
 351         int split = 100 / td->cpuload;
 352         int i = 0;
 353
 354         while (!td->terminate) {
 355                 fio_gettime(&e, NULL);
 356
 357                 if (runtime_exceeded(td, &e))
 358                         break;
 359
 360                 if (!(i % split))
 361                         __usec_sleep(10000);
 362                 else
 363                         usec_sleep(td, 10000);
 364
 365                 i++;
 366         }
 367 }
 368
 369 /*
 370  * Main IO worker function. It retrieves io_u's to process and queues
 371  * and reaps them, checking for rate and errors along the way.
 372  */
 373 static void do_io(struct thread_data *td)
 374 {
 375         struct timeval s;
 376         unsigned long usec;
 377         int i, ret = 0;
 378
 379         td_set_runstate(td, TD_RUNNING);
 380
 381         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
 382                 struct timeval comp_time;
 383                 long bytes_done = 0;
 384                 int min_evts = 0;
 385                 struct io_u *io_u;
 386
 387                 if (td->terminate)
 388                         break;
 389
 390                 io_u = get_io_u(td);
 391                 if (!io_u)
 392                         break;
 393
 394                 memcpy(&s, &io_u->start_time, sizeof(s));
 395
 396                 if (runtime_exceeded(td, &s)) {
 397                         put_io_u(td, io_u);
 398                         break;
 399                 }
 400 requeue:
 401                 ret = td_io_queue(td, io_u);
 402
 403                 switch (ret) {
 404                 case FIO_Q_COMPLETED:
 405                         if (io_u->error) {
 406                                 ret = io_u->error;
 407                                 break;
 408                         }
 409                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
 410                                 int bytes = io_u->xfer_buflen - io_u->resid;
 411
 412                                 io_u->xfer_buflen = io_u->resid;
 413                                 io_u->xfer_buf += bytes;
 414                                 goto requeue;
 415                         }
 416                         fio_gettime(&comp_time, NULL);
 417                         bytes_done = io_u_sync_complete(td, io_u, NULL);
 418                         if (bytes_done < 0)
 419                                 ret = bytes_done;
 420                         break;
 421                 case FIO_Q_QUEUED:
 422                         /*
 423                          * if the engine doesn't have a commit hook,
 424                          * the io_u is really queued. if it does have such
 425                          * a hook, it has to call io_u_queued() itself.
 426                          */
 427                         if (td->io_ops->commit == NULL)
 428                                 io_u_queued(td, io_u);
 429                         break;
 430                 case FIO_Q_BUSY:
 431                         requeue_io_u(td, &io_u);
 432                         ret = td_io_commit(td);
 433                         break;
 434                 default:
 435                         assert(ret < 0);
 436                         put_io_u(td, io_u);
 437                         break;
 438                 }
 439
 440                 if (ret < 0 || td->error)
 441                         break;
 442
 443                 /*
 444                  * See if we need to complete some commands
 445                  */
 446                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 447                         min_evts = 0;
 448                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 449                                 min_evts = 1;
 450
 451                                 if (td->cur_depth > td->iodepth_low)
 452                                         min_evts = td->cur_depth - td->iodepth_low;
 453                         }
 454
 455                         fio_gettime(&comp_time, NULL);
 456                         bytes_done = io_u_queued_complete(td, min_evts, NULL);
 457                         if (bytes_done < 0)
 458                                 break;
 459                 }
 460
 461                 if (!bytes_done)
 462                         continue;
 463
 464                 /*
 465                  * the rate is batched for now, it should work for batches
 466                  * of completions except the very first one which may look
 467                  * a little bursty
 468                  */
 469                 usec = utime_since(&s, &comp_time);
 470
 471                 rate_throttle(td, usec, bytes_done);
 472
 473                 if (check_min_rate(td, &comp_time)) {
 474                         if (exitall_on_terminate)
 475                                 terminate_threads(td->groupid, 0);
 476                         td_verror(td, ENODATA, "check_min_rate");
 477                         break;
 478                 }
 479
 480                 if (td->thinktime) {
 481                         unsigned long long b;
 482
 483                         b = td->io_blocks[0] + td->io_blocks[1];
 484                         if (!(b % td->thinktime_blocks)) {
 485                                 int left;
 486
 487                                 if (td->thinktime_spin)
 488                                         __usec_sleep(td->thinktime_spin);
 489
 490                                 left = td->thinktime - td->thinktime_spin;
 491                                 if (left)
 492                                         usec_sleep(td, left);
 493                         }
 494                 }
 495         }
 496
 497         if (!td->error) {
 498                 struct fio_file *f;
 499
 500                 if (td->cur_depth)
 501                         cleanup_pending_aio(td);
 502
 503                 if (should_fsync(td) && td->end_fsync) {
 504                         td_set_runstate(td, TD_FSYNCING);
 505                         for_each_file(td, f, i)
 506                                 fio_io_sync(td, f);
 507                 }
 508         }
 509 }
 510
 511 static void cleanup_io_u(struct thread_data *td)
 512 {
 513         struct list_head *entry, *n;
 514         struct io_u *io_u;
 515
 516         list_for_each_safe(entry, n, &td->io_u_freelist) {
 517                 io_u = list_entry(entry, struct io_u, list);
 518
 519                 list_del(&io_u->list);
 520                 free(io_u);
 521         }
 522
 523         free_io_mem(td);
 524 }
 525
 526 /*
 527  * "randomly" fill the buffer contents
 528  */
 529 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 530 {
 531         int *ptr = io_u->buf;
 532
 533         while ((void *) ptr - io_u->buf < max_bs) {
 534                 *ptr = rand() * 0x9e370001;
 535                 ptr++;
 536         }
 537 }
 538
 539 static int init_io_u(struct thread_data *td)
 540 {
 541         struct io_u *io_u;
 542         unsigned int max_bs;
 543         int i, max_units;
 544         char *p;
 545
 546         if (td->io_ops->flags & FIO_CPUIO)
 547                 return 0;
 548
 549         if (td->io_ops->flags & FIO_SYNCIO)
 550                 max_units = 1;
 551         else
 552                 max_units = td->iodepth;
 553
 554         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
 555         td->orig_buffer_size = max_bs * max_units;
 556
 557         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
 558                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
 559         else
 560                 td->orig_buffer_size += page_mask;
 561
 562         if (allocate_io_mem(td))
 563                 return 1;
 564
 565         p = ALIGN(td->orig_buffer);
 566         for (i = 0; i < max_units; i++) {
 567                 io_u = malloc(sizeof(*io_u));
 568                 memset(io_u, 0, sizeof(*io_u));
 569                 INIT_LIST_HEAD(&io_u->list);
 570
 571                 io_u->buf = p + max_bs * i;
 572                 if (td_write(td) || td_rw(td))
 573                         fill_rand_buf(io_u, max_bs);
 574
 575                 io_u->index = i;
 576                 io_u->flags = IO_U_F_FREE;
 577                 list_add(&io_u->list, &td->io_u_freelist);
 578         }
 579
 580         io_u_init_timeout();
 581
 582         return 0;
 583 }
 584
 585 static int switch_ioscheduler(struct thread_data *td)
 586 {
 587         char tmp[256], tmp2[128];
 588         FILE *f;
 589         int ret;
 590
 591         if (td->io_ops->flags & FIO_CPUIO)
 592                 return 0;
 593
 594         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 595
 596         f = fopen(tmp, "r+");
 597         if (!f) {
 598                 td_verror(td, errno, "fopen");
 599                 return 1;
 600         }
 601
 602         /*
 603          * Set io scheduler.
 604          */
 605         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
 606         if (ferror(f) || ret != 1) {
 607                 td_verror(td, errno, "fwrite");
 608                 fclose(f);
 609                 return 1;
 610         }
 611
 612         rewind(f);
 613
 614         /*
 615          * Read back and check that the selected scheduler is now the default.
 616          */
 617         ret = fread(tmp, 1, sizeof(tmp), f);
 618         if (ferror(f) || ret < 0) {
 619                 td_verror(td, errno, "fread");
 620                 fclose(f);
 621                 return 1;
 622         }
 623
 624         sprintf(tmp2, "[%s]", td->ioscheduler);
 625         if (!strstr(tmp, tmp2)) {
 626                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
 627                 td_verror(td, EINVAL, "iosched_switch");
 628                 fclose(f);
 629                 return 1;
 630         }
 631
 632         fclose(f);
 633         return 0;
 634 }
 635
 636 static void clear_io_state(struct thread_data *td)
 637 {
 638         struct fio_file *f;
 639         int i;
 640
 641         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 642         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 643         td->zone_bytes = 0;
 644
 645         td->last_was_sync = 0;
 646
 647         for_each_file(td, f, i) {
 648                 f->last_completed_pos = 0;
 649
 650                 f->last_pos = 0;
 651                 if (td->io_ops->flags & FIO_SYNCIO)
 652                         lseek(f->fd, SEEK_SET, 0);
 653
 654                 if (f->file_map)
 655                         memset(f->file_map, 0, f->num_maps * sizeof(long));
 656         }
 657 }
 658
 659 /*
 660  * Entry point for the thread based jobs. The process based jobs end up
 661  * here as well, after a little setup.
 662  */
 663 static void *thread_main(void *data)
 664 {
 665         unsigned long long runtime[2];
 666         struct thread_data *td = data;
 667
 668         if (!td->use_thread)
 669                 setsid();
 670
 671         td->pid = getpid();
 672
 673         INIT_LIST_HEAD(&td->io_u_freelist);
 674         INIT_LIST_HEAD(&td->io_u_busylist);
 675         INIT_LIST_HEAD(&td->io_u_requeues);
 676         INIT_LIST_HEAD(&td->io_hist_list);
 677         INIT_LIST_HEAD(&td->io_log_list);
 678
 679         if (init_io_u(td))
 680                 goto err;
 681
 682         if (fio_setaffinity(td) == -1) {
 683                 td_verror(td, errno, "cpu_set_affinity");
 684                 goto err;
 685         }
 686
 687         if (init_iolog(td))
 688                 goto err;
 689
 690         if (td->ioprio) {
 691                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 692                         td_verror(td, errno, "ioprio_set");
 693                         goto err;
 694                 }
 695         }
 696
 697         if (nice(td->nice) == -1) {
 698                 td_verror(td, errno, "nice");
 699                 goto err;
 700         }
 701
 702         if (init_random_state(td))
 703                 goto err;
 704
 705         if (td->ioscheduler && switch_ioscheduler(td))
 706                 goto err;
 707
 708         td_set_runstate(td, TD_INITIALIZED);
 709         fio_sem_up(&startup_sem);
 710         fio_sem_down(&td->mutex);
 711
 712         if (!td->create_serialize && setup_files(td))
 713                 goto err;
 714         if (open_files(td))
 715                 goto err;
 716
 717         /*
 718          * Do this late, as some IO engines would like to have the
 719          * files setup prior to initializing structures.
 720          */
 721         if (td_io_init(td))
 722                 goto err;
 723
 724         if (td->exec_prerun) {
 725                 if (system(td->exec_prerun) < 0)
 726                         goto err;
 727         }
 728
 729         fio_gettime(&td->epoch, NULL);
 730         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 731         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 732
 733         runtime[0] = runtime[1] = 0;
 734         while (td->loops--) {
 735                 fio_gettime(&td->start, NULL);
 736                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 737
 738                 if (td->ratemin)
 739                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 740
 741                 clear_io_state(td);
 742                 prune_io_piece_log(td);
 743
 744                 if (td->io_ops->flags & FIO_CPUIO)
 745                         do_cpuio(td);
 746                 else
 747                         do_io(td);
 748
 749                 if (td_read(td) && td->io_bytes[DDIR_READ])
 750                         runtime[DDIR_READ] += utime_since_now(&td->start);
 751                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
 752                         runtime[DDIR_WRITE] += utime_since_now(&td->start);
 753
 754                 if (td->error || td->terminate)
 755                         break;
 756
 757                 if (td->verify == VERIFY_NONE)
 758                         continue;
 759
 760                 clear_io_state(td);
 761                 fio_gettime(&td->start, NULL);
 762
 763                 do_verify(td);
 764
 765                 runtime[DDIR_READ] += utime_since_now(&td->start);
 766
 767                 if (td->error || td->terminate)
 768                         break;
 769         }
 770
 771         update_rusage_stat(td);
 772         fio_gettime(&td->end_time, NULL);
 773         td->runtime[0] = runtime[0] / 1000;
 774         td->runtime[1] = runtime[1] / 1000;
 775
 776         if (td->ts.bw_log)
 777                 finish_log(td, td->ts.bw_log, "bw");
 778         if (td->ts.slat_log)
 779                 finish_log(td, td->ts.slat_log, "slat");
 780         if (td->ts.clat_log)
 781                 finish_log(td, td->ts.clat_log, "clat");
 782         if (td->write_iolog_file)
 783                 write_iolog_close(td);
 784         if (td->exec_postrun) {
 785                 if (system(td->exec_postrun) < 0)
 786                         log_err("fio: postrun %s failed\n", td->exec_postrun);
 787         }
 788
 789         if (exitall_on_terminate)
 790                 terminate_threads(td->groupid, 0);
 791
 792 err:
 793         if (td->error)
 794                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 795         close_files(td);
 796         close_ioengine(td);
 797         cleanup_io_u(td);
 798         td_set_runstate(td, TD_EXITED);
 799         return (void *) (unsigned long) td->error;
 800 }
 801
 802 /*
 803  * We cannot pass the td data into a forked process, so attach the td and
 804  * pass it to the thread worker.
 805  */
 806 static int fork_main(int shmid, int offset)
 807 {
 808         struct thread_data *td;
 809         void *data, *ret;
 810
 811         data = shmat(shmid, NULL, 0);
 812         if (data == (void *) -1) {
 813                 int __err = errno;
 814
 815                 perror("shmat");
 816                 return __err;
 817         }
 818
 819         td = data + offset * sizeof(struct thread_data);
 820         ret = thread_main(td);
 821         shmdt(data);
 822         return (int) (unsigned long) ret;
 823 }
 824
 825 /*
 826  * Run over the job map and reap the threads that have exited, if any.
 827  */
 828 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 829 {
 830         struct thread_data *td;
 831         int i, cputhreads, pending, status, ret;
 832
 833         /*
 834          * reap exited threads (TD_EXITED -> TD_REAPED)
 835          */
 836         pending = cputhreads = 0;
 837         for_each_td(td, i) {
 838                 int flags = 0;
 839
 840                 /*
 841                  * ->io_ops is NULL for a thread that has closed its
 842                  * io engine
 843                  */
 844                 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
 845                         cputhreads++;
 846
 847                 if (!td->pid || td->runstate == TD_REAPED)
 848                         continue;
 849                 if (td->use_thread) {
 850                         if (td->runstate == TD_EXITED) {
 851                                 td_set_runstate(td, TD_REAPED);
 852                                 goto reaped;
 853                         }
 854                         continue;
 855                 }
 856
 857                 flags = WNOHANG;
 858                 if (td->runstate == TD_EXITED)
 859                         flags = 0;
 860
 861                 /*
 862                  * check if someone quit or got killed in an unusual way
 863                  */
 864                 ret = waitpid(td->pid, &status, flags);
 865                 if (ret < 0) {
 866                         if (errno == ECHILD) {
 867                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
 868                                 td_set_runstate(td, TD_REAPED);
 869                                 goto reaped;
 870                         }
 871                         perror("waitpid");
 872                 } else if (ret == td->pid) {
 873                         if (WIFSIGNALED(status)) {
 874                                 int sig = WTERMSIG(status);
 875
 876                                 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
 877                                 td_set_runstate(td, TD_REAPED);
 878                                 goto reaped;
 879                         }
 880                         if (WIFEXITED(status)) {
 881                                 if (WEXITSTATUS(status) && !td->error)
 882                                         td->error = WEXITSTATUS(status);
 883
 884                                 td_set_runstate(td, TD_REAPED);
 885                                 goto reaped;
 886                         }
 887                 }
 888
 889                 /*
 890                  * thread is not dead, continue
 891                  */
 892                 continue;
 893 reaped:
 894                 if (td->use_thread) {
 895                         long ret;
 896
 897                         if (pthread_join(td->thread, (void *) &ret))
 898                                 perror("pthread_join");
 899                 }
 900
 901                 (*nr_running)--;
 902                 (*m_rate) -= td->ratemin;
 903                 (*t_rate) -= td->rate;
 904
 905                 if (td->error)
 906                         exit_value++;
 907         }
 908
 909         if (*nr_running == cputhreads && !pending)
 910                 terminate_threads(TERMINATE_ALL, 0);
 911 }
 912
 913 /*
 914  * Main function for kicking off and reaping jobs, as needed.
 915  */
 916 static void run_threads(void)
 917 {
 918         struct thread_data *td;
 919         unsigned long spent;
 920         int i, todo, nr_running, m_rate, t_rate, nr_started;
 921
 922         if (fio_pin_memory())
 923                 return;
 924
 925         if (!terse_output) {
 926                 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
 927                 fflush(stdout);
 928         }
 929
 930         signal(SIGINT, sig_handler);
 931         signal(SIGALRM, sig_handler);
 932
 933         todo = thread_number;
 934         nr_running = 0;
 935         nr_started = 0;
 936         m_rate = t_rate = 0;
 937
 938         for_each_td(td, i) {
 939                 print_status_init(td->thread_number - 1);
 940
 941                 if (!td->create_serialize) {
 942                         init_disk_util(td);
 943                         continue;
 944                 }
 945
 946                 /*
 947                  * do file setup here so it happens sequentially,
 948                  * we don't want X number of threads getting their
 949                  * client data interspersed on disk
 950                  */
 951                 if (setup_files(td)) {
 952                         exit_value++;
 953                         if (td->error)
 954                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 955                         td_set_runstate(td, TD_REAPED);
 956                         todo--;
 957                 }
 958
 959                 init_disk_util(td);
 960         }
 961
 962         set_genesis_time();
 963
 964         while (todo) {
 965                 struct thread_data *map[MAX_JOBS];
 966                 struct timeval this_start;
 967                 int this_jobs = 0, left;
 968
 969                 /*
 970                  * create threads (TD_NOT_CREATED -> TD_CREATED)
 971                  */
 972                 for_each_td(td, i) {
 973                         if (td->runstate != TD_NOT_CREATED)
 974                                 continue;
 975
 976                         /*
 977                          * never got a chance to start, killed by other
 978                          * thread for some reason
 979                          */
 980                         if (td->terminate) {
 981                                 todo--;
 982                                 continue;
 983                         }
 984
 985                         if (td->start_delay) {
 986                                 spent = mtime_since_genesis();
 987
 988                                 if (td->start_delay * 1000 > spent)
 989                                         continue;
 990                         }
 991
 992                         if (td->stonewall && (nr_started || nr_running))
 993                                 break;
 994
 995                         /*
 996                          * Set state to created. Thread will transition
 997                          * to TD_INITIALIZED when it's done setting up.
 998                          */
 999                         td_set_runstate(td, TD_CREATED);
1000                         map[this_jobs++] = td;
1001                         fio_sem_init(&startup_sem, 1);
1002                         nr_started++;
1003
1004                         if (td->use_thread) {
1005                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1006                                         perror("thread_create");
1007                                         nr_started--;
1008                                 }
1009                         } else {
1010                                 if (fork())
1011                                         fio_sem_down(&startup_sem);
1012                                 else {
1013                                         int ret = fork_main(shm_id, i);
1014
1015                                         exit(ret);
1016                                 }
1017                         }
1018                 }
1019
1020                 /*
1021                  * Wait for the started threads to transition to
1022                  * TD_INITIALIZED.
1023                  */
1024                 fio_gettime(&this_start, NULL);
1025                 left = this_jobs;
1026                 while (left && !fio_abort) {
1027                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1028                                 break;
1029
1030                         usleep(100000);
1031
1032                         for (i = 0; i < this_jobs; i++) {
1033                                 td = map[i];
1034                                 if (!td)
1035                                         continue;
1036                                 if (td->runstate == TD_INITIALIZED) {
1037                                         map[i] = NULL;
1038                                         left--;
1039                                 } else if (td->runstate >= TD_EXITED) {
1040                                         map[i] = NULL;
1041                                         left--;
1042                                         todo--;
1043                                         nr_running++; /* work-around... */
1044                                 }
1045                         }
1046                 }
1047
1048                 if (left) {
1049                         log_err("fio: %d jobs failed to start\n", left);
1050                         for (i = 0; i < this_jobs; i++) {
1051                                 td = map[i];
1052                                 if (!td)
1053                                         continue;
1054                                 kill(td->pid, SIGTERM);
1055                         }
1056                         break;
1057                 }
1058
1059                 /*
1060                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1061                  */
1062                 for_each_td(td, i) {
1063                         if (td->runstate != TD_INITIALIZED)
1064                                 continue;
1065
1066                         td_set_runstate(td, TD_RUNNING);
1067                         nr_running++;
1068                         nr_started--;
1069                         m_rate += td->ratemin;
1070                         t_rate += td->rate;
1071                         todo--;
1072                         fio_sem_up(&td->mutex);
1073                 }
1074
1075                 reap_threads(&nr_running, &t_rate, &m_rate);
1076
1077                 if (todo)
1078                         usleep(100000);
1079         }
1080
1081         while (nr_running) {
1082                 reap_threads(&nr_running, &t_rate, &m_rate);
1083                 usleep(10000);
1084         }
1085
1086         update_io_ticks();
1087         fio_unpin_memory();
1088 }
1089
1090 int main(int argc, char *argv[])
1091 {
1092         long ps;
1093
1094         /*
1095          * We need locale for number printing, if it isn't set then just
1096          * go with the US format.
1097          */
1098         if (!getenv("LC_NUMERIC"))
1099                 setlocale(LC_NUMERIC, "en_US");
1100
1101         if (parse_options(argc, argv))
1102                 return 1;
1103
1104         if (!thread_number) {
1105                 log_err("Nothing to do\n");
1106                 return 1;
1107         }
1108
1109         ps = sysconf(_SC_PAGESIZE);
1110         if (ps < 0) {
1111                 log_err("Failed to get page size\n");
1112                 return 1;
1113         }
1114
1115         page_mask = ps - 1;
1116
1117         if (write_bw_log) {
1118                 setup_log(&agg_io_log[DDIR_READ]);
1119                 setup_log(&agg_io_log[DDIR_WRITE]);
1120         }
1121
1122         set_genesis_time();
1123
1124         disk_util_timer_arm();
1125
1126         run_threads();
1127
1128         if (!fio_abort) {
1129                 show_run_stats();
1130                 if (write_bw_log) {
1131                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1132                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1133                 }
1134         }
1135
1136         return exit_value;
1137 }