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)
  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                         kill(td->pid, SIGQUIT);
  71                         td->terminate = 1;
  72                         td->start_delay = 0;
  73                 }
  74         }
  75 }
  76
  77 static void sig_handler(int sig)
  78 {
  79         switch (sig) {
  80                 case SIGALRM:
  81                         update_io_ticks();
  82                         disk_util_timer_arm();
  83                         print_thread_status();
  84                         break;
  85                 default:
  86                         printf("\nfio: terminating on signal %d\n", sig);
  87                         fflush(stdout);
  88                         terminate_threads(TERMINATE_ALL);
  89                         break;
  90         }
  91 }
  92
  93 /*
  94  * Check if we are above the minimum rate given.
  95  */
  96 static int check_min_rate(struct thread_data *td, struct timeval *now)
  97 {
  98         unsigned long long bytes = 0;
  99         unsigned long spent;
 100         unsigned long rate;
 101
 102         /*
 103          * No minimum rate set, always ok
 104          */
 105         if (!td->ratemin)
 106                 return 0;
 107
 108         /*
 109          * allow a 2 second settle period in the beginning
 110          */
 111         if (mtime_since(&td->start, now) < 2000)
 112                 return 0;
 113
 114         if (td_read(td))
 115                 bytes += td->this_io_bytes[DDIR_READ];
 116         if (td_write(td))
 117                 bytes += td->this_io_bytes[DDIR_WRITE];
 118
 119         /*
 120          * if rate blocks is set, sample is running
 121          */
 122         if (td->rate_bytes) {
 123                 spent = mtime_since(&td->lastrate, now);
 124                 if (spent < td->ratecycle)
 125                         return 0;
 126
 127                 if (bytes < td->rate_bytes) {
 128                         fprintf(f_out, "%s: min rate %u not met\n", td->name, td->ratemin);
 129                         return 1;
 130                 } else {
 131                         rate = (bytes - td->rate_bytes) / spent;
 132                         if (rate < td->ratemin || bytes < td->rate_bytes) {
 133                                 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
 134                                 return 1;
 135                         }
 136                 }
 137         }
 138
 139         td->rate_bytes = bytes;
 140         memcpy(&td->lastrate, now, sizeof(*now));
 141         return 0;
 142 }
 143
 144 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 145 {
 146         if (!td->timeout)
 147                 return 0;
 148         if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
 149                 return 1;
 150
 151         return 0;
 152 }
 153
 154 /*
 155  * When job exits, we can cancel the in-flight IO if we are using async
 156  * io. Attempt to do so.
 157  */
 158 static void cleanup_pending_aio(struct thread_data *td)
 159 {
 160         struct list_head *entry, *n;
 161         struct io_u *io_u;
 162         int r;
 163
 164         /*
 165          * get immediately available events, if any
 166          */
 167         r = io_u_queued_complete(td, 0);
 168         if (r < 0)
 169                 return;
 170
 171         /*
 172          * now cancel remaining active events
 173          */
 174         if (td->io_ops->cancel) {
 175                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 176                         io_u = list_entry(entry, struct io_u, list);
 177
 178                         /*
 179                          * if the io_u isn't in flight, then that generally
 180                          * means someone leaked an io_u. complain but fix
 181                          * it up, so we don't stall here.
 182                          */
 183                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 184                                 log_err("fio: non-busy IO on busy list\n");
 185                                 put_io_u(td, io_u);
 186                         } else {
 187                                 r = td->io_ops->cancel(td, io_u);
 188                                 if (!r)
 189                                         put_io_u(td, io_u);
 190                         }
 191                 }
 192         }
 193
 194         if (td->cur_depth)
 195                 r = io_u_queued_complete(td, td->cur_depth);
 196 }
 197
 198 /*
 199  * Helper to handle the final sync of a file. Works just like the normal
 200  * io path, just does everything sync.
 201  */
 202 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 203 {
 204         struct io_u *io_u = __get_io_u(td);
 205         int ret;
 206
 207         if (!io_u)
 208                 return 1;
 209
 210         io_u->ddir = DDIR_SYNC;
 211         io_u->file = f;
 212
 213         if (td_io_prep(td, io_u)) {
 214                 put_io_u(td, io_u);
 215                 return 1;
 216         }
 217
 218 requeue:
 219         ret = td_io_queue(td, io_u);
 220         if (ret < 0) {
 221                 td_verror(td, io_u->error, "td_io_queue");
 222                 put_io_u(td, io_u);
 223                 return 1;
 224         } else if (ret == FIO_Q_QUEUED) {
 225                 if (io_u_queued_complete(td, 1) < 0)
 226                         return 1;
 227         } else if (ret == FIO_Q_COMPLETED) {
 228                 if (io_u->error) {
 229                         td_verror(td, io_u->error, "td_io_queue");
 230                         return 1;
 231                 }
 232
 233                 if (io_u_sync_complete(td, io_u) < 0)
 234                         return 1;
 235         } else if (ret == FIO_Q_BUSY) {
 236                 if (td_io_commit(td))
 237                         return 1;
 238                 goto requeue;
 239         }
 240
 241         return 0;
 242 }
 243
 244 /*
 245  * The main verify engine. Runs over the writes we previously submitted,
 246  * reads the blocks back in, and checks the crc/md5 of the data.
 247  */
 248 static void do_verify(struct thread_data *td)
 249 {
 250         struct fio_file *f;
 251         struct io_u *io_u;
 252         int ret, i, min_events;
 253
 254         /*
 255          * sync io first and invalidate cache, to make sure we really
 256          * read from disk.
 257          */
 258         for_each_file(td, f, i) {
 259                 if (fio_io_sync(td, f))
 260                         break;
 261                 if (file_invalidate_cache(td, f))
 262                         break;
 263         }
 264
 265         if (td->error)
 266                 return;
 267
 268         td_set_runstate(td, TD_VERIFYING);
 269
 270         io_u = NULL;
 271         while (!td->terminate) {
 272                 int ret2;
 273
 274                 io_u = __get_io_u(td);
 275                 if (!io_u)
 276                         break;
 277
 278                 if (runtime_exceeded(td, &io_u->start_time)) {
 279                         put_io_u(td, io_u);
 280                         break;
 281                 }
 282
 283                 if (get_next_verify(td, io_u)) {
 284                         put_io_u(td, io_u);
 285                         break;
 286                 }
 287
 288                 if (td_io_prep(td, io_u)) {
 289                         put_io_u(td, io_u);
 290                         break;
 291                 }
 292
 293                 io_u->end_io = verify_io_u;
 294
 295                 ret = td_io_queue(td, io_u);
 296                 switch (ret) {
 297                 case FIO_Q_COMPLETED:
 298                         if (io_u->error)
 299                                 ret = -io_u->error;
 300                         else if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
 301                                 int bytes = io_u->xfer_buflen - io_u->resid;
 302
 303                                 io_u->xfer_buflen = io_u->resid;
 304                                 io_u->xfer_buf += bytes;
 305                                 requeue_io_u(td, &io_u);
 306                         } else {
 307                                 ret = io_u_sync_complete(td, io_u);
 308                                 if (ret < 0)
 309                                         break;
 310                         }
 311                         continue;
 312                 case FIO_Q_QUEUED:
 313                         break;
 314                 case FIO_Q_BUSY:
 315                         requeue_io_u(td, &io_u);
 316                         ret2 = td_io_commit(td);
 317                         if (ret2 < 0)
 318                                 ret = ret2;
 319                         break;
 320                 default:
 321                         assert(ret < 0);
 322                         td_verror(td, -ret, "td_io_queue");
 323                         break;
 324                 }
 325
 326                 if (ret < 0 || td->error)
 327                         break;
 328
 329                 /*
 330                  * if we can queue more, do so. but check if there are
 331                  * completed io_u's first.
 332                  */
 333                 min_events = 0;
 334                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 335                         min_events = 1;
 336
 337                         if (td->cur_depth > td->iodepth_low)
 338                                 min_events = td->cur_depth - td->iodepth_low;
 339                 }
 340
 341                 /*
 342                  * Reap required number of io units, if any, and do the
 343                  * verification on them through the callback handler
 344                  */
 345                 if (io_u_queued_complete(td, min_events) < 0)
 346                         break;
 347         }
 348
 349         if (!td->error) {
 350                 min_events = td->cur_depth;
 351
 352                 if (min_events)
 353                         ret = io_u_queued_complete(td, min_events);
 354         } else
 355                 cleanup_pending_aio(td);
 356
 357         td_set_runstate(td, TD_RUNNING);
 358 }
 359
 360 /*
 361  * Not really an io thread, all it does is burn CPU cycles in the specified
 362  * manner.
 363  */
 364 static void do_cpuio(struct thread_data *td)
 365 {
 366         struct timeval e;
 367         int split = 100 / td->cpuload;
 368         int i = 0;
 369
 370         while (!td->terminate) {
 371                 fio_gettime(&e, NULL);
 372
 373                 if (runtime_exceeded(td, &e))
 374                         break;
 375
 376                 if (!(i % split))
 377                         __usec_sleep(10000);
 378                 else
 379                         usec_sleep(td, 10000);
 380
 381                 i++;
 382         }
 383 }
 384
 385 /*
 386  * Main IO worker function. It retrieves io_u's to process and queues
 387  * and reaps them, checking for rate and errors along the way.
 388  */
 389 static void do_io(struct thread_data *td)
 390 {
 391         struct timeval s;
 392         unsigned long usec;
 393         int i, ret = 0;
 394
 395         td_set_runstate(td, TD_RUNNING);
 396
 397         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
 398                 struct timeval comp_time;
 399                 long bytes_done = 0;
 400                 int min_evts = 0;
 401                 struct io_u *io_u;
 402                 int ret2;
 403
 404                 if (td->terminate)
 405                         break;
 406
 407                 io_u = get_io_u(td);
 408                 if (!io_u)
 409                         break;
 410
 411                 memcpy(&s, &io_u->start_time, sizeof(s));
 412
 413                 if (runtime_exceeded(td, &s)) {
 414                         put_io_u(td, io_u);
 415                         break;
 416                 }
 417
 418                 ret = td_io_queue(td, io_u);
 419                 switch (ret) {
 420                 case FIO_Q_COMPLETED:
 421                         if (io_u->error)
 422                                 ret = -io_u->error;
 423                         else if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
 424                                 int bytes = io_u->xfer_buflen - io_u->resid;
 425
 426                                 io_u->xfer_buflen = io_u->resid;
 427                                 io_u->xfer_buf += bytes;
 428                                 requeue_io_u(td, &io_u);
 429                         } else {
 430                                 fio_gettime(&comp_time, NULL);
 431                                 bytes_done = io_u_sync_complete(td, io_u);
 432                                 if (bytes_done < 0)
 433                                         ret = bytes_done;
 434                         }
 435                         break;
 436                 case FIO_Q_QUEUED:
 437                         /*
 438                          * if the engine doesn't have a commit hook,
 439                          * the io_u is really queued. if it does have such
 440                          * a hook, it has to call io_u_queued() itself.
 441                          */
 442                         if (td->io_ops->commit == NULL)
 443                                 io_u_queued(td, io_u);
 444                         break;
 445                 case FIO_Q_BUSY:
 446                         requeue_io_u(td, &io_u);
 447                         ret2 = td_io_commit(td);
 448                         if (ret2 < 0)
 449                                 ret = ret2;
 450                         break;
 451                 default:
 452                         assert(ret < 0);
 453                         put_io_u(td, io_u);
 454                         break;
 455                 }
 456
 457                 if (ret < 0 || td->error)
 458                         break;
 459
 460                 /*
 461                  * See if we need to complete some commands
 462                  */
 463                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 464                         min_evts = 0;
 465                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 466                                 min_evts = 1;
 467
 468                                 if (td->cur_depth > td->iodepth_low)
 469                                         min_evts = td->cur_depth - td->iodepth_low;
 470                         }
 471
 472                         fio_gettime(&comp_time, NULL);
 473                         bytes_done = io_u_queued_complete(td, min_evts);
 474                         if (bytes_done < 0)
 475                                 break;
 476                 }
 477
 478                 if (!bytes_done)
 479                         continue;
 480
 481                 /*
 482                  * the rate is batched for now, it should work for batches
 483                  * of completions except the very first one which may look
 484                  * a little bursty
 485                  */
 486                 usec = utime_since(&s, &comp_time);
 487
 488                 rate_throttle(td, usec, bytes_done);
 489
 490                 if (check_min_rate(td, &comp_time)) {
 491                         if (exitall_on_terminate)
 492                                 terminate_threads(td->groupid);
 493                         td_verror(td, ENODATA, "check_min_rate");
 494                         break;
 495                 }
 496
 497                 if (td->thinktime) {
 498                         unsigned long long b;
 499
 500                         b = td->io_blocks[0] + td->io_blocks[1];
 501                         if (!(b % td->thinktime_blocks)) {
 502                                 int left;
 503
 504                                 if (td->thinktime_spin)
 505                                         __usec_sleep(td->thinktime_spin);
 506
 507                                 left = td->thinktime - td->thinktime_spin;
 508                                 if (left)
 509                                         usec_sleep(td, left);
 510                         }
 511                 }
 512         }
 513
 514         if (!td->error) {
 515                 struct fio_file *f;
 516
 517                 i = td->cur_depth;
 518                 if (i)
 519                         ret = io_u_queued_complete(td, i);
 520
 521                 if (should_fsync(td) && td->end_fsync) {
 522                         td_set_runstate(td, TD_FSYNCING);
 523                         for_each_file(td, f, i)
 524                                 fio_io_sync(td, f);
 525                 }
 526         } else
 527                 cleanup_pending_aio(td);
 528 }
 529
 530 static void cleanup_io_u(struct thread_data *td)
 531 {
 532         struct list_head *entry, *n;
 533         struct io_u *io_u;
 534
 535         list_for_each_safe(entry, n, &td->io_u_freelist) {
 536                 io_u = list_entry(entry, struct io_u, list);
 537
 538                 list_del(&io_u->list);
 539                 free(io_u);
 540         }
 541
 542         free_io_mem(td);
 543 }
 544
 545 /*
 546  * "randomly" fill the buffer contents
 547  */
 548 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 549 {
 550         int *ptr = io_u->buf;
 551
 552         while ((void *) ptr - io_u->buf < max_bs) {
 553                 *ptr = rand() * 0x9e370001;
 554                 ptr++;
 555         }
 556 }
 557
 558 static int init_io_u(struct thread_data *td)
 559 {
 560         struct io_u *io_u;
 561         unsigned int max_bs;
 562         int i, max_units;
 563         char *p;
 564
 565         if (td->io_ops->flags & FIO_CPUIO)
 566                 return 0;
 567
 568         if (td->io_ops->flags & FIO_SYNCIO)
 569                 max_units = 1;
 570         else
 571                 max_units = td->iodepth;
 572
 573         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
 574         td->orig_buffer_size = max_bs * max_units;
 575
 576         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
 577                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
 578         else
 579                 td->orig_buffer_size += page_mask;
 580
 581         if (allocate_io_mem(td))
 582                 return 1;
 583
 584         p = ALIGN(td->orig_buffer);
 585         for (i = 0; i < max_units; i++) {
 586                 io_u = malloc(sizeof(*io_u));
 587                 memset(io_u, 0, sizeof(*io_u));
 588                 INIT_LIST_HEAD(&io_u->list);
 589
 590                 io_u->buf = p + max_bs * i;
 591                 if (td_write(td) || td_rw(td))
 592                         fill_rand_buf(io_u, max_bs);
 593
 594                 io_u->index = i;
 595                 io_u->flags = IO_U_F_FREE;
 596                 list_add(&io_u->list, &td->io_u_freelist);
 597         }
 598
 599         io_u_init_timeout();
 600
 601         return 0;
 602 }
 603
 604 static int switch_ioscheduler(struct thread_data *td)
 605 {
 606         char tmp[256], tmp2[128];
 607         FILE *f;
 608         int ret;
 609
 610         if (td->io_ops->flags & FIO_CPUIO)
 611                 return 0;
 612
 613         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 614
 615         f = fopen(tmp, "r+");
 616         if (!f) {
 617                 td_verror(td, errno, "fopen");
 618                 return 1;
 619         }
 620
 621         /*
 622          * Set io scheduler.
 623          */
 624         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
 625         if (ferror(f) || ret != 1) {
 626                 td_verror(td, errno, "fwrite");
 627                 fclose(f);
 628                 return 1;
 629         }
 630
 631         rewind(f);
 632
 633         /*
 634          * Read back and check that the selected scheduler is now the default.
 635          */
 636         ret = fread(tmp, 1, sizeof(tmp), f);
 637         if (ferror(f) || ret < 0) {
 638                 td_verror(td, errno, "fread");
 639                 fclose(f);
 640                 return 1;
 641         }
 642
 643         sprintf(tmp2, "[%s]", td->ioscheduler);
 644         if (!strstr(tmp, tmp2)) {
 645                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
 646                 td_verror(td, EINVAL, "iosched_switch");
 647                 fclose(f);
 648                 return 1;
 649         }
 650
 651         fclose(f);
 652         return 0;
 653 }
 654
 655 static void clear_io_state(struct thread_data *td)
 656 {
 657         struct fio_file *f;
 658         int i;
 659
 660         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 661         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 662         td->zone_bytes = 0;
 663
 664         td->last_was_sync = 0;
 665
 666         for_each_file(td, f, i) {
 667                 f->last_completed_pos = 0;
 668
 669                 f->last_pos = 0;
 670                 if (td->io_ops->flags & FIO_SYNCIO)
 671                         lseek(f->fd, SEEK_SET, 0);
 672
 673                 if (f->file_map)
 674                         memset(f->file_map, 0, f->num_maps * sizeof(long));
 675         }
 676 }
 677
 678 /*
 679  * Entry point for the thread based jobs. The process based jobs end up
 680  * here as well, after a little setup.
 681  */
 682 static void *thread_main(void *data)
 683 {
 684         unsigned long long runtime[2];
 685         struct thread_data *td = data;
 686
 687         if (!td->use_thread)
 688                 setsid();
 689
 690         td->pid = getpid();
 691
 692         INIT_LIST_HEAD(&td->io_u_freelist);
 693         INIT_LIST_HEAD(&td->io_u_busylist);
 694         INIT_LIST_HEAD(&td->io_u_requeues);
 695         INIT_LIST_HEAD(&td->io_hist_list);
 696         INIT_LIST_HEAD(&td->io_log_list);
 697
 698         if (init_io_u(td))
 699                 goto err;
 700
 701         if (fio_setaffinity(td) == -1) {
 702                 td_verror(td, errno, "cpu_set_affinity");
 703                 goto err;
 704         }
 705
 706         if (init_iolog(td))
 707                 goto err;
 708
 709         if (td->ioprio) {
 710                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 711                         td_verror(td, errno, "ioprio_set");
 712                         goto err;
 713                 }
 714         }
 715
 716         if (nice(td->nice) == -1) {
 717                 td_verror(td, errno, "nice");
 718                 goto err;
 719         }
 720
 721         if (init_random_state(td))
 722                 goto err;
 723
 724         if (td->ioscheduler && switch_ioscheduler(td))
 725                 goto err;
 726
 727         td_set_runstate(td, TD_INITIALIZED);
 728         fio_sem_up(&startup_sem);
 729         fio_sem_down(&td->mutex);
 730
 731         if (!td->create_serialize && setup_files(td))
 732                 goto err;
 733         if (open_files(td))
 734                 goto err;
 735
 736         /*
 737          * Do this late, as some IO engines would like to have the
 738          * files setup prior to initializing structures.
 739          */
 740         if (td_io_init(td))
 741                 goto err;
 742
 743         if (td->exec_prerun) {
 744                 if (system(td->exec_prerun) < 0)
 745                         goto err;
 746         }
 747
 748         fio_gettime(&td->epoch, NULL);
 749         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 750         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 751
 752         runtime[0] = runtime[1] = 0;
 753         while (td->loops--) {
 754                 fio_gettime(&td->start, NULL);
 755                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 756
 757                 if (td->ratemin)
 758                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 759
 760                 clear_io_state(td);
 761                 prune_io_piece_log(td);
 762
 763                 if (td->io_ops->flags & FIO_CPUIO)
 764                         do_cpuio(td);
 765                 else
 766                         do_io(td);
 767
 768                 if (td_read(td) && td->io_bytes[DDIR_READ])
 769                         runtime[DDIR_READ] += utime_since_now(&td->start);
 770                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
 771                         runtime[DDIR_WRITE] += utime_since_now(&td->start);
 772
 773                 if (td->error || td->terminate)
 774                         break;
 775
 776                 if (td->verify == VERIFY_NONE)
 777                         continue;
 778
 779                 clear_io_state(td);
 780                 fio_gettime(&td->start, NULL);
 781
 782                 do_verify(td);
 783
 784                 runtime[DDIR_READ] += utime_since_now(&td->start);
 785
 786                 if (td->error || td->terminate)
 787                         break;
 788         }
 789
 790         update_rusage_stat(td);
 791         fio_gettime(&td->end_time, NULL);
 792         td->runtime[0] = runtime[0] / 1000;
 793         td->runtime[1] = runtime[1] / 1000;
 794
 795         if (td->ts.bw_log)
 796                 finish_log(td, td->ts.bw_log, "bw");
 797         if (td->ts.slat_log)
 798                 finish_log(td, td->ts.slat_log, "slat");
 799         if (td->ts.clat_log)
 800                 finish_log(td, td->ts.clat_log, "clat");
 801         if (td->write_iolog_file)
 802                 write_iolog_close(td);
 803         if (td->exec_postrun) {
 804                 if (system(td->exec_postrun) < 0)
 805                         log_err("fio: postrun %s failed\n", td->exec_postrun);
 806         }
 807
 808         if (exitall_on_terminate)
 809                 terminate_threads(td->groupid);
 810
 811 err:
 812         if (td->error)
 813                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 814         close_files(td);
 815         close_ioengine(td);
 816         cleanup_io_u(td);
 817         td_set_runstate(td, TD_EXITED);
 818         return (void *) (unsigned long) td->error;
 819 }
 820
 821 /*
 822  * We cannot pass the td data into a forked process, so attach the td and
 823  * pass it to the thread worker.
 824  */
 825 static int fork_main(int shmid, int offset)
 826 {
 827         struct thread_data *td;
 828         void *data, *ret;
 829
 830         data = shmat(shmid, NULL, 0);
 831         if (data == (void *) -1) {
 832                 int __err = errno;
 833
 834                 perror("shmat");
 835                 return __err;
 836         }
 837
 838         td = data + offset * sizeof(struct thread_data);
 839         ret = thread_main(td);
 840         shmdt(data);
 841         return (int) (unsigned long) ret;
 842 }
 843
 844 /*
 845  * Run over the job map and reap the threads that have exited, if any.
 846  */
 847 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 848 {
 849         struct thread_data *td;
 850         int i, cputhreads, pending, status, ret;
 851
 852         /*
 853          * reap exited threads (TD_EXITED -> TD_REAPED)
 854          */
 855         pending = cputhreads = 0;
 856         for_each_td(td, i) {
 857                 int flags = 0;
 858
 859                 /*
 860                  * ->io_ops is NULL for a thread that has closed its
 861                  * io engine
 862                  */
 863                 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
 864                         cputhreads++;
 865
 866                 if (!td->pid || td->runstate == TD_REAPED)
 867                         continue;
 868                 if (td->use_thread) {
 869                         if (td->runstate == TD_EXITED) {
 870                                 td_set_runstate(td, TD_REAPED);
 871                                 goto reaped;
 872                         }
 873                         continue;
 874                 }
 875
 876                 flags = WNOHANG;
 877                 if (td->runstate == TD_EXITED)
 878                         flags = 0;
 879
 880                 /*
 881                  * check if someone quit or got killed in an unusual way
 882                  */
 883                 ret = waitpid(td->pid, &status, flags);
 884                 if (ret < 0) {
 885                         if (errno == ECHILD) {
 886                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
 887                                 td_set_runstate(td, TD_REAPED);
 888                                 goto reaped;
 889                         }
 890                         perror("waitpid");
 891                 } else if (ret == td->pid) {
 892                         if (WIFSIGNALED(status)) {
 893                                 int sig = WTERMSIG(status);
 894
 895                                 if (sig != SIGQUIT)
 896                                         log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
 897                                 td_set_runstate(td, TD_REAPED);
 898                                 goto reaped;
 899                         }
 900                         if (WIFEXITED(status)) {
 901                                 if (WEXITSTATUS(status) && !td->error)
 902                                         td->error = WEXITSTATUS(status);
 903
 904                                 td_set_runstate(td, TD_REAPED);
 905                                 goto reaped;
 906                         }
 907                 }
 908
 909                 /*
 910                  * thread is not dead, continue
 911                  */
 912                 continue;
 913 reaped:
 914                 if (td->use_thread) {
 915                         long ret;
 916
 917                         if (pthread_join(td->thread, (void *) &ret))
 918                                 perror("pthread_join");
 919                 }
 920
 921                 (*nr_running)--;
 922                 (*m_rate) -= td->ratemin;
 923                 (*t_rate) -= td->rate;
 924
 925                 if (td->error)
 926                         exit_value++;
 927         }
 928
 929         if (*nr_running == cputhreads && !pending)
 930                 terminate_threads(TERMINATE_ALL);
 931 }
 932
 933 /*
 934  * Main function for kicking off and reaping jobs, as needed.
 935  */
 936 static void run_threads(void)
 937 {
 938         struct thread_data *td;
 939         unsigned long spent;
 940         int i, todo, nr_running, m_rate, t_rate, nr_started;
 941
 942         if (fio_pin_memory())
 943                 return;
 944
 945         if (!terse_output) {
 946                 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
 947                 fflush(stdout);
 948         }
 949
 950         signal(SIGINT, sig_handler);
 951         signal(SIGALRM, sig_handler);
 952
 953         todo = thread_number;
 954         nr_running = 0;
 955         nr_started = 0;
 956         m_rate = t_rate = 0;
 957
 958         for_each_td(td, i) {
 959                 print_status_init(td->thread_number - 1);
 960
 961                 if (!td->create_serialize) {
 962                         init_disk_util(td);
 963                         continue;
 964                 }
 965
 966                 /*
 967                  * do file setup here so it happens sequentially,
 968                  * we don't want X number of threads getting their
 969                  * client data interspersed on disk
 970                  */
 971                 if (setup_files(td)) {
 972                         exit_value++;
 973                         if (td->error)
 974                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 975                         td_set_runstate(td, TD_REAPED);
 976                         todo--;
 977                 }
 978
 979                 init_disk_util(td);
 980         }
 981
 982         set_genesis_time();
 983
 984         while (todo) {
 985                 struct thread_data *map[MAX_JOBS];
 986                 struct timeval this_start;
 987                 int this_jobs = 0, left;
 988
 989                 /*
 990                  * create threads (TD_NOT_CREATED -> TD_CREATED)
 991                  */
 992                 for_each_td(td, i) {
 993                         if (td->runstate != TD_NOT_CREATED)
 994                                 continue;
 995
 996                         /*
 997                          * never got a chance to start, killed by other
 998                          * thread for some reason
 999                          */
1000                         if (td->terminate) {
1001                                 todo--;
1002                                 continue;
1003                         }
1004
1005                         if (td->start_delay) {
1006                                 spent = mtime_since_genesis();
1007
1008                                 if (td->start_delay * 1000 > spent)
1009                                         continue;
1010                         }
1011
1012                         if (td->stonewall && (nr_started || nr_running))
1013                                 break;
1014
1015                         /*
1016                          * Set state to created. Thread will transition
1017                          * to TD_INITIALIZED when it's done setting up.
1018                          */
1019                         td_set_runstate(td, TD_CREATED);
1020                         map[this_jobs++] = td;
1021                         fio_sem_init(&startup_sem, 1);
1022                         nr_started++;
1023
1024                         if (td->use_thread) {
1025                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1026                                         perror("thread_create");
1027                                         nr_started--;
1028                                 }
1029                         } else {
1030                                 if (fork())
1031                                         fio_sem_down(&startup_sem);
1032                                 else {
1033                                         int ret = fork_main(shm_id, i);
1034
1035                                         exit(ret);
1036                                 }
1037                         }
1038                 }
1039
1040                 /*
1041                  * Wait for the started threads to transition to
1042                  * TD_INITIALIZED.
1043                  */
1044                 fio_gettime(&this_start, NULL);
1045                 left = this_jobs;
1046                 while (left && !fio_abort) {
1047                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1048                                 break;
1049
1050                         usleep(100000);
1051
1052                         for (i = 0; i < this_jobs; i++) {
1053                                 td = map[i];
1054                                 if (!td)
1055                                         continue;
1056                                 if (td->runstate == TD_INITIALIZED) {
1057                                         map[i] = NULL;
1058                                         left--;
1059                                 } else if (td->runstate >= TD_EXITED) {
1060                                         map[i] = NULL;
1061                                         left--;
1062                                         todo--;
1063                                         nr_running++; /* work-around... */
1064                                 }
1065                         }
1066                 }
1067
1068                 if (left) {
1069                         log_err("fio: %d jobs failed to start\n", left);
1070                         for (i = 0; i < this_jobs; i++) {
1071                                 td = map[i];
1072                                 if (!td)
1073                                         continue;
1074                                 kill(td->pid, SIGTERM);
1075                         }
1076                         break;
1077                 }
1078
1079                 /*
1080                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1081                  */
1082                 for_each_td(td, i) {
1083                         if (td->runstate != TD_INITIALIZED)
1084                                 continue;
1085
1086                         td_set_runstate(td, TD_RUNNING);
1087                         nr_running++;
1088                         nr_started--;
1089                         m_rate += td->ratemin;
1090                         t_rate += td->rate;
1091                         todo--;
1092                         fio_sem_up(&td->mutex);
1093                 }
1094
1095                 reap_threads(&nr_running, &t_rate, &m_rate);
1096
1097                 if (todo)
1098                         usleep(100000);
1099         }
1100
1101         while (nr_running) {
1102                 reap_threads(&nr_running, &t_rate, &m_rate);
1103                 usleep(10000);
1104         }
1105
1106         update_io_ticks();
1107         fio_unpin_memory();
1108 }
1109
1110 int main(int argc, char *argv[])
1111 {
1112         long ps;
1113
1114         /*
1115          * We need locale for number printing, if it isn't set then just
1116          * go with the US format.
1117          */
1118         if (!getenv("LC_NUMERIC"))
1119                 setlocale(LC_NUMERIC, "en_US");
1120
1121         if (parse_options(argc, argv))
1122                 return 1;
1123
1124         if (!thread_number) {
1125                 log_err("Nothing to do\n");
1126                 return 1;
1127         }
1128
1129         ps = sysconf(_SC_PAGESIZE);
1130         if (ps < 0) {
1131                 log_err("Failed to get page size\n");
1132                 return 1;
1133         }
1134
1135         page_mask = ps - 1;
1136
1137         if (write_bw_log) {
1138                 setup_log(&agg_io_log[DDIR_READ]);
1139                 setup_log(&agg_io_log[DDIR_WRITE]);
1140         }
1141
1142         set_genesis_time();
1143
1144         disk_util_timer_arm();
1145
1146         run_threads();
1147
1148         if (!fio_abort) {
1149                 show_run_stats();
1150                 if (write_bw_log) {
1151                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1152                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1153                 }
1154         }
1155
1156         return exit_value;
1157 }