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