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
  39 unsigned long page_mask;
  40 unsigned long page_size;
  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 nr_process = 0;
  47 int nr_thread = 0;
  48 int shm_id = 0;
  49 int temp_stall_ts;
  50
  51 static struct fio_sem *startup_sem;
  52 static volatile int fio_abort;
  53 static int exit_value;
  54
  55 struct io_log *agg_io_log[2];
  56
  57 #define TERMINATE_ALL           (-1)
  58 #define JOB_START_TIMEOUT       (5 * 1000)
  59
  60 static inline void td_set_runstate(struct thread_data *td, int runstate)
  61 {
  62         td->runstate = runstate;
  63 }
  64
  65 static void terminate_threads(int group_id)
  66 {
  67         struct thread_data *td;
  68         int i;
  69
  70         for_each_td(td, i) {
  71                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  72                         /*
  73                          * if the thread is running, just let it exit
  74                          */
  75                         if (td->runstate < TD_RUNNING)
  76                                 kill(td->pid, SIGQUIT);
  77                         td->terminate = 1;
  78                         td->o.start_delay = 0;
  79                 }
  80         }
  81 }
  82
  83 static void sig_handler(int sig)
  84 {
  85         switch (sig) {
  86                 case SIGALRM:
  87                         update_io_ticks();
  88                         disk_util_timer_arm();
  89                         print_thread_status();
  90                         break;
  91                 default:
  92                         printf("\nfio: terminating on signal %d\n", sig);
  93                         fflush(stdout);
  94                         terminate_threads(TERMINATE_ALL);
  95                         break;
  96         }
  97 }
  98
  99 /*
 100  * Check if we are above the minimum rate given.
 101  */
 102 static int check_min_rate(struct thread_data *td, struct timeval *now)
 103 {
 104         unsigned long long bytes = 0;
 105         unsigned long iops = 0;
 106         unsigned long spent;
 107         unsigned long rate;
 108
 109         /*
 110          * No minimum rate set, always ok
 111          */
 112         if (!td->o.ratemin && !td->o.rate_iops_min)
 113                 return 0;
 114
 115         /*
 116          * allow a 2 second settle period in the beginning
 117          */
 118         if (mtime_since(&td->start, now) < 2000)
 119                 return 0;
 120
 121         if (td_read(td)) {
 122                 iops += td->io_blocks[DDIR_READ];
 123                 bytes += td->this_io_bytes[DDIR_READ];
 124         }
 125         if (td_write(td)) {
 126                 iops += td->io_blocks[DDIR_WRITE];
 127                 bytes += td->this_io_bytes[DDIR_WRITE];
 128         }
 129
 130         /*
 131          * if rate blocks is set, sample is running
 132          */
 133         if (td->rate_bytes || td->rate_blocks) {
 134                 spent = mtime_since(&td->lastrate, now);
 135                 if (spent < td->o.ratecycle)
 136                         return 0;
 137
 138                 if (td->o.rate) {
 139                         /*
 140                          * check bandwidth specified rate
 141                          */
 142                         if (bytes < td->rate_bytes) {
 143                                 log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
 144                                 return 1;
 145                         } else {
 146                                 rate = (bytes - td->rate_bytes) / spent;
 147                                 if (rate < td->o.ratemin || bytes < td->rate_bytes) {
 148                                         log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
 149                                         return 1;
 150                                 }
 151                         }
 152                 } else {
 153                         /*
 154                          * checks iops specified rate
 155                          */
 156                         if (iops < td->o.rate_iops) {
 157                                 log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
 158                                 return 1;
 159                         } else {
 160                                 rate = (iops - td->rate_blocks) / spent;
 161                                 if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
 162                                         log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
 163                                 }
 164                         }
 165                 }
 166         }
 167
 168         td->rate_bytes = bytes;
 169         td->rate_blocks = iops;
 170         memcpy(&td->lastrate, now, sizeof(*now));
 171         return 0;
 172 }
 173
 174 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 175 {
 176         if (!td->o.timeout)
 177                 return 0;
 178         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 179                 return 1;
 180
 181         return 0;
 182 }
 183
 184 /*
 185  * When job exits, we can cancel the in-flight IO if we are using async
 186  * io. Attempt to do so.
 187  */
 188 static void cleanup_pending_aio(struct thread_data *td)
 189 {
 190         struct list_head *entry, *n;
 191         struct io_u *io_u;
 192         int r;
 193
 194         /*
 195          * get immediately available events, if any
 196          */
 197         r = io_u_queued_complete(td, 0);
 198         if (r < 0)
 199                 return;
 200
 201         /*
 202          * now cancel remaining active events
 203          */
 204         if (td->io_ops->cancel) {
 205                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 206                         io_u = list_entry(entry, struct io_u, list);
 207
 208                         /*
 209                          * if the io_u isn't in flight, then that generally
 210                          * means someone leaked an io_u. complain but fix
 211                          * it up, so we don't stall here.
 212                          */
 213                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 214                                 log_err("fio: non-busy IO on busy list\n");
 215                                 put_io_u(td, io_u);
 216                         } else {
 217                                 r = td->io_ops->cancel(td, io_u);
 218                                 if (!r)
 219                                         put_io_u(td, io_u);
 220                         }
 221                 }
 222         }
 223
 224         if (td->cur_depth)
 225                 r = io_u_queued_complete(td, td->cur_depth);
 226 }
 227
 228 /*
 229  * Helper to handle the final sync of a file. Works just like the normal
 230  * io path, just does everything sync.
 231  */
 232 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 233 {
 234         struct io_u *io_u = __get_io_u(td);
 235         int ret;
 236
 237         if (!io_u)
 238                 return 1;
 239
 240         io_u->ddir = DDIR_SYNC;
 241         io_u->file = f;
 242
 243         if (td_io_prep(td, io_u)) {
 244                 put_io_u(td, io_u);
 245                 return 1;
 246         }
 247
 248 requeue:
 249         ret = td_io_queue(td, io_u);
 250         if (ret < 0) {
 251                 td_verror(td, io_u->error, "td_io_queue");
 252                 put_io_u(td, io_u);
 253                 return 1;
 254         } else if (ret == FIO_Q_QUEUED) {
 255                 if (io_u_queued_complete(td, 1) < 0)
 256                         return 1;
 257         } else if (ret == FIO_Q_COMPLETED) {
 258                 if (io_u->error) {
 259                         td_verror(td, io_u->error, "td_io_queue");
 260                         return 1;
 261                 }
 262
 263                 if (io_u_sync_complete(td, io_u) < 0)
 264                         return 1;
 265         } else if (ret == FIO_Q_BUSY) {
 266                 if (td_io_commit(td))
 267                         return 1;
 268                 goto requeue;
 269         }
 270
 271         return 0;
 272 }
 273
 274 /*
 275  * The main verify engine. Runs over the writes we previously submitted,
 276  * reads the blocks back in, and checks the crc/md5 of the data.
 277  */
 278 static void do_verify(struct thread_data *td)
 279 {
 280         struct fio_file *f;
 281         struct io_u *io_u;
 282         int ret, min_events;
 283         unsigned int i;
 284
 285         /*
 286          * sync io first and invalidate cache, to make sure we really
 287          * read from disk.
 288          */
 289         for_each_file(td, f, i) {
 290                 if (!(f->flags & FIO_FILE_OPEN))
 291                         continue;
 292                 if (fio_io_sync(td, f))
 293                         break;
 294                 if (file_invalidate_cache(td, f))
 295                         break;
 296         }
 297
 298         if (td->error)
 299                 return;
 300
 301         td_set_runstate(td, TD_VERIFYING);
 302
 303         io_u = NULL;
 304         while (!td->terminate) {
 305                 int ret2;
 306
 307                 io_u = __get_io_u(td);
 308                 if (!io_u)
 309                         break;
 310
 311                 if (runtime_exceeded(td, &io_u->start_time)) {
 312                         put_io_u(td, io_u);
 313                         break;
 314                 }
 315
 316                 if (get_next_verify(td, io_u)) {
 317                         put_io_u(td, io_u);
 318                         break;
 319                 }
 320
 321                 if (td_io_prep(td, io_u)) {
 322                         put_io_u(td, io_u);
 323                         break;
 324                 }
 325
 326                 io_u->end_io = verify_io_u;
 327
 328                 ret = td_io_queue(td, io_u);
 329                 switch (ret) {
 330                 case FIO_Q_COMPLETED:
 331                         if (io_u->error)
 332                                 ret = -io_u->error;
 333                         else if (io_u->resid) {
 334                                 int bytes = io_u->xfer_buflen - io_u->resid;
 335
 336                                 /*
 337                                  * zero read, fail
 338                                  */
 339                                 if (!bytes) {
 340                                         td_verror(td, ENODATA, "full resid");
 341                                         put_io_u(td, io_u);
 342                                         break;
 343                                 }
 344
 345                                 io_u->xfer_buflen = io_u->resid;
 346                                 io_u->xfer_buf += bytes;
 347                                 io_u->offset += bytes;
 348
 349                                 if (io_u->offset == io_u->file->real_file_size)
 350                                         goto sync_done;
 351
 352                                 requeue_io_u(td, &io_u);
 353                         } else {
 354 sync_done:
 355                                 ret = io_u_sync_complete(td, io_u);
 356                                 if (ret < 0)
 357                                         break;
 358                         }
 359                         continue;
 360                 case FIO_Q_QUEUED:
 361                         break;
 362                 case FIO_Q_BUSY:
 363                         requeue_io_u(td, &io_u);
 364                         ret2 = td_io_commit(td);
 365                         if (ret2 < 0)
 366                                 ret = ret2;
 367                         break;
 368                 default:
 369                         assert(ret < 0);
 370                         td_verror(td, -ret, "td_io_queue");
 371                         break;
 372                 }
 373
 374                 if (ret < 0 || td->error)
 375                         break;
 376
 377                 /*
 378                  * if we can queue more, do so. but check if there are
 379                  * completed io_u's first.
 380                  */
 381                 min_events = 0;
 382                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 383                         min_events = 1;
 384
 385                         if (td->cur_depth > td->o.iodepth_low)
 386                                 min_events = td->cur_depth - td->o.iodepth_low;
 387                 }
 388
 389                 /*
 390                  * Reap required number of io units, if any, and do the
 391                  * verification on them through the callback handler
 392                  */
 393                 if (io_u_queued_complete(td, min_events) < 0)
 394                         break;
 395         }
 396
 397         if (!td->error) {
 398                 min_events = td->cur_depth;
 399
 400                 if (min_events)
 401                         ret = io_u_queued_complete(td, min_events);
 402         } else
 403                 cleanup_pending_aio(td);
 404
 405         td_set_runstate(td, TD_RUNNING);
 406 }
 407
 408 /*
 409  * Main IO worker function. It retrieves io_u's to process and queues
 410  * and reaps them, checking for rate and errors along the way.
 411  */
 412 static void do_io(struct thread_data *td)
 413 {
 414         struct timeval s;
 415         unsigned long usec;
 416         unsigned int i;
 417         int ret = 0;
 418
 419         td_set_runstate(td, TD_RUNNING);
 420
 421         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 422                 struct timeval comp_time;
 423                 long bytes_done = 0;
 424                 int min_evts = 0;
 425                 struct io_u *io_u;
 426                 int ret2;
 427
 428                 if (td->terminate)
 429                         break;
 430
 431                 io_u = get_io_u(td);
 432                 if (!io_u)
 433                         break;
 434
 435                 memcpy(&s, &io_u->start_time, sizeof(s));
 436
 437                 if (runtime_exceeded(td, &s)) {
 438                         put_io_u(td, io_u);
 439                         break;
 440                 }
 441
 442                 ret = td_io_queue(td, io_u);
 443                 switch (ret) {
 444                 case FIO_Q_COMPLETED:
 445                         if (io_u->error)
 446                                 ret = -io_u->error;
 447                         else if (io_u->resid) {
 448                                 int bytes = io_u->xfer_buflen - io_u->resid;
 449
 450                                 /*
 451                                  * zero read, fail
 452                                  */
 453                                 if (!bytes) {
 454                                         td_verror(td, ENODATA, "full resid");
 455                                         put_io_u(td, io_u);
 456                                         break;
 457                                 }
 458
 459                                 io_u->xfer_buflen = io_u->resid;
 460                                 io_u->xfer_buf += bytes;
 461                                 io_u->offset += bytes;
 462
 463                                 if (io_u->offset == io_u->file->real_file_size)
 464                                         goto sync_done;
 465
 466                                 requeue_io_u(td, &io_u);
 467                         } else {
 468 sync_done:
 469                                 fio_gettime(&comp_time, NULL);
 470                                 bytes_done = io_u_sync_complete(td, io_u);
 471                                 if (bytes_done < 0)
 472                                         ret = bytes_done;
 473                         }
 474                         break;
 475                 case FIO_Q_QUEUED:
 476                         /*
 477                          * if the engine doesn't have a commit hook,
 478                          * the io_u is really queued. if it does have such
 479                          * a hook, it has to call io_u_queued() itself.
 480                          */
 481                         if (td->io_ops->commit == NULL)
 482                                 io_u_queued(td, io_u);
 483                         break;
 484                 case FIO_Q_BUSY:
 485                         requeue_io_u(td, &io_u);
 486                         ret2 = td_io_commit(td);
 487                         if (ret2 < 0)
 488                                 ret = ret2;
 489                         break;
 490                 default:
 491                         assert(ret < 0);
 492                         put_io_u(td, io_u);
 493                         break;
 494                 }
 495
 496                 if (ret < 0 || td->error)
 497                         break;
 498
 499                 /*
 500                  * See if we need to complete some commands
 501                  */
 502                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 503                         min_evts = 0;
 504                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 505                                 min_evts = 1;
 506
 507                                 if (td->cur_depth > td->o.iodepth_low)
 508                                         min_evts = td->cur_depth - td->o.iodepth_low;
 509                         }
 510
 511                         fio_gettime(&comp_time, NULL);
 512                         bytes_done = io_u_queued_complete(td, min_evts);
 513                         if (bytes_done < 0)
 514                                 break;
 515                 }
 516
 517                 if (!bytes_done)
 518                         continue;
 519
 520                 /*
 521                  * the rate is batched for now, it should work for batches
 522                  * of completions except the very first one which may look
 523                  * a little bursty
 524                  */
 525                 usec = utime_since(&s, &comp_time);
 526
 527                 rate_throttle(td, usec, bytes_done);
 528
 529                 if (check_min_rate(td, &comp_time)) {
 530                         if (exitall_on_terminate)
 531                                 terminate_threads(td->groupid);
 532                         td_verror(td, ENODATA, "check_min_rate");
 533                         break;
 534                 }
 535
 536                 if (td->o.thinktime) {
 537                         unsigned long long b;
 538
 539                         b = td->io_blocks[0] + td->io_blocks[1];
 540                         if (!(b % td->o.thinktime_blocks)) {
 541                                 int left;
 542
 543                                 if (td->o.thinktime_spin)
 544                                         __usec_sleep(td->o.thinktime_spin);
 545
 546                                 left = td->o.thinktime - td->o.thinktime_spin;
 547                                 if (left)
 548                                         usec_sleep(td, left);
 549                         }
 550                 }
 551         }
 552
 553         if (!td->error) {
 554                 struct fio_file *f;
 555
 556                 i = td->cur_depth;
 557                 if (i)
 558                         ret = io_u_queued_complete(td, i);
 559
 560                 if (should_fsync(td) && td->o.end_fsync) {
 561                         td_set_runstate(td, TD_FSYNCING);
 562
 563                         for_each_file(td, f, i) {
 564                                 if (!(f->flags & FIO_FILE_OPEN))
 565                                         continue;
 566                                 fio_io_sync(td, f);
 567                         }
 568                 }
 569         } else
 570                 cleanup_pending_aio(td);
 571 }
 572
 573 static void cleanup_io_u(struct thread_data *td)
 574 {
 575         struct list_head *entry, *n;
 576         struct io_u *io_u;
 577
 578         list_for_each_safe(entry, n, &td->io_u_freelist) {
 579                 io_u = list_entry(entry, struct io_u, list);
 580
 581                 list_del(&io_u->list);
 582                 free(io_u);
 583         }
 584
 585         free_io_mem(td);
 586 }
 587
 588 /*
 589  * "randomly" fill the buffer contents
 590  */
 591 static void fill_rand_buf(struct io_u *io_u, int max_bs)
 592 {
 593         int *ptr = io_u->buf;
 594
 595         while ((void *) ptr - io_u->buf < max_bs) {
 596                 *ptr = rand() * 0x9e370001;
 597                 ptr++;
 598         }
 599 }
 600
 601 static int init_io_u(struct thread_data *td)
 602 {
 603         unsigned long long buf_size;
 604         struct io_u *io_u;
 605         unsigned int max_bs;
 606         int i, max_units;
 607         char *p;
 608
 609         if (td->io_ops->flags & FIO_SYNCIO)
 610                 max_units = 1;
 611         else
 612                 max_units = td->o.iodepth;
 613
 614         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 615         buf_size = (unsigned long long) max_bs * (unsigned long long) max_units;
 616         buf_size += page_mask;
 617         if (buf_size != (size_t) buf_size) {
 618                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 619                 return 1;
 620         }
 621
 622         td->orig_buffer_size = buf_size;
 623
 624         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
 625                 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
 626         else if (td->orig_buffer_size & page_mask)
 627                 td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
 628
 629         if (allocate_io_mem(td))
 630                 return 1;
 631
 632         p = ALIGN(td->orig_buffer);
 633         for (i = 0; i < max_units; i++) {
 634                 io_u = malloc(sizeof(*io_u));
 635                 memset(io_u, 0, sizeof(*io_u));
 636                 INIT_LIST_HEAD(&io_u->list);
 637
 638                 io_u->buf = p + max_bs * i;
 639                 if (td_write(td) || td_rw(td))
 640                         fill_rand_buf(io_u, max_bs);
 641
 642                 io_u->index = i;
 643                 io_u->flags = IO_U_F_FREE;
 644                 list_add(&io_u->list, &td->io_u_freelist);
 645         }
 646
 647         io_u_init_timeout();
 648
 649         return 0;
 650 }
 651
 652 static int switch_ioscheduler(struct thread_data *td)
 653 {
 654         char tmp[256], tmp2[128];
 655         FILE *f;
 656         int ret;
 657
 658         if (td->io_ops->flags & FIO_DISKLESSIO)
 659                 return 0;
 660
 661         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 662
 663         f = fopen(tmp, "r+");
 664         if (!f) {
 665                 if (errno == ENOENT) {
 666                         log_err("fio: os or kernel doesn't support IO scheduler switching\n");
 667                         return 0;
 668                 }
 669                 td_verror(td, errno, "fopen iosched");
 670                 return 1;
 671         }
 672
 673         /*
 674          * Set io scheduler.
 675          */
 676         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 677         if (ferror(f) || ret != 1) {
 678                 td_verror(td, errno, "fwrite");
 679                 fclose(f);
 680                 return 1;
 681         }
 682
 683         rewind(f);
 684
 685         /*
 686          * Read back and check that the selected scheduler is now the default.
 687          */
 688         ret = fread(tmp, 1, sizeof(tmp), f);
 689         if (ferror(f) || ret < 0) {
 690                 td_verror(td, errno, "fread");
 691                 fclose(f);
 692                 return 1;
 693         }
 694
 695         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 696         if (!strstr(tmp, tmp2)) {
 697                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 698                 td_verror(td, EINVAL, "iosched_switch");
 699                 fclose(f);
 700                 return 1;
 701         }
 702
 703         fclose(f);
 704         return 0;
 705 }
 706
 707 static int clear_io_state(struct thread_data *td)
 708 {
 709         struct fio_file *f;
 710         unsigned int i;
 711         int ret;
 712
 713         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 714         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 715         td->zone_bytes = 0;
 716         td->rate_bytes = 0;
 717         td->rate_blocks = 0;
 718         td->rw_end_set[0] = td->rw_end_set[1] = 0;
 719
 720         td->last_was_sync = 0;
 721
 722         for_each_file(td, f, i)
 723                 td_io_close_file(td, f);
 724
 725         ret = 0;
 726         for_each_file(td, f, i) {
 727                 ret = td_io_open_file(td, f);
 728                 if (ret)
 729                         break;
 730         }
 731
 732         return ret;
 733 }
 734
 735 /*
 736  * Entry point for the thread based jobs. The process based jobs end up
 737  * here as well, after a little setup.
 738  */
 739 static void *thread_main(void *data)
 740 {
 741         unsigned long long runtime[2];
 742         struct thread_data *td = data;
 743         unsigned long elapsed;
 744         int clear_state;
 745
 746         if (!td->o.use_thread)
 747                 setsid();
 748
 749         td->pid = getpid();
 750
 751         INIT_LIST_HEAD(&td->io_u_freelist);
 752         INIT_LIST_HEAD(&td->io_u_busylist);
 753         INIT_LIST_HEAD(&td->io_u_requeues);
 754         INIT_LIST_HEAD(&td->io_log_list);
 755         INIT_LIST_HEAD(&td->io_hist_list);
 756         td->io_hist_tree = RB_ROOT;
 757
 758         if (init_io_u(td))
 759                 goto err_sem;
 760
 761         if (fio_setaffinity(td) == -1) {
 762                 td_verror(td, errno, "cpu_set_affinity");
 763                 goto err_sem;
 764         }
 765
 766         if (init_iolog(td))
 767                 goto err_sem;
 768
 769         if (td->ioprio) {
 770                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 771                         td_verror(td, errno, "ioprio_set");
 772                         goto err_sem;
 773                 }
 774         }
 775
 776         if (nice(td->o.nice) == -1) {
 777                 td_verror(td, errno, "nice");
 778                 goto err_sem;
 779         }
 780
 781         if (td->o.ioscheduler && switch_ioscheduler(td))
 782                 goto err_sem;
 783
 784         td_set_runstate(td, TD_INITIALIZED);
 785         fio_sem_up(startup_sem);
 786         fio_sem_down(td->mutex);
 787
 788         /*
 789          * the ->mutex semaphore is now no longer used, close it to avoid
 790          * eating a file descriptor
 791          */
 792         fio_sem_remove(td->mutex);
 793
 794         if (!td->o.create_serialize && setup_files(td))
 795                 goto err;
 796
 797         if (td_io_init(td))
 798                 goto err;
 799
 800         if (open_files(td))
 801                 goto err;
 802
 803         if (init_random_map(td))
 804                 goto err;
 805
 806         if (td->o.exec_prerun) {
 807                 if (system(td->o.exec_prerun) < 0)
 808                         goto err;
 809         }
 810
 811         fio_gettime(&td->epoch, NULL);
 812         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 813         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 814
 815         runtime[0] = runtime[1] = 0;
 816         clear_state = 0;
 817         while (td->o.loops--) {
 818                 fio_gettime(&td->start, NULL);
 819                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 820
 821                 if (td->o.ratemin)
 822                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 823
 824                 if (clear_state && clear_io_state(td))
 825                         break;
 826
 827                 prune_io_piece_log(td);
 828
 829                 do_io(td);
 830
 831                 clear_state = 1;
 832
 833                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
 834                         if (td->rw_end_set[DDIR_READ])
 835                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
 836                         else
 837                                 elapsed = utime_since_now(&td->start);
 838
 839                         runtime[DDIR_READ] += elapsed;
 840                 }
 841                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
 842                         if (td->rw_end_set[DDIR_WRITE])
 843                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
 844                         else
 845                                 elapsed = utime_since_now(&td->start);
 846
 847                         runtime[DDIR_WRITE] += elapsed;
 848                 }
 849
 850                 if (td->error || td->terminate)
 851                         break;
 852
 853                 if (td->o.verify == VERIFY_NONE)
 854                         continue;
 855
 856                 if (clear_io_state(td))
 857                         break;
 858
 859                 fio_gettime(&td->start, NULL);
 860
 861                 do_verify(td);
 862
 863                 runtime[DDIR_READ] += utime_since_now(&td->start);
 864
 865                 if (td->error || td->terminate)
 866                         break;
 867         }
 868
 869         update_rusage_stat(td);
 870         td->ts.runtime[0] = runtime[0] / 1000;
 871         td->ts.runtime[1] = runtime[1] / 1000;
 872         td->ts.total_run_time = mtime_since_now(&td->epoch);
 873         td->ts.io_bytes[0] = td->io_bytes[0];
 874         td->ts.io_bytes[1] = td->io_bytes[1];
 875
 876         if (td->ts.bw_log)
 877                 finish_log(td, td->ts.bw_log, "bw");
 878         if (td->ts.slat_log)
 879                 finish_log(td, td->ts.slat_log, "slat");
 880         if (td->ts.clat_log)
 881                 finish_log(td, td->ts.clat_log, "clat");
 882         if (td->o.write_iolog_file)
 883                 write_iolog_close(td);
 884         if (td->o.exec_postrun) {
 885                 if (system(td->o.exec_postrun) < 0)
 886                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
 887         }
 888
 889         if (exitall_on_terminate)
 890                 terminate_threads(td->groupid);
 891
 892 err:
 893         if (td->error)
 894                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 895         close_files(td);
 896         close_ioengine(td);
 897         cleanup_io_u(td);
 898         options_mem_free(td);
 899         td_set_runstate(td, TD_EXITED);
 900         return (void *) (unsigned long) td->error;
 901 err_sem:
 902         fio_sem_up(startup_sem);
 903         goto err;
 904 }
 905
 906 /*
 907  * We cannot pass the td data into a forked process, so attach the td and
 908  * pass it to the thread worker.
 909  */
 910 static int fork_main(int shmid, int offset)
 911 {
 912         struct thread_data *td;
 913         void *data, *ret;
 914
 915         data = shmat(shmid, NULL, 0);
 916         if (data == (void *) -1) {
 917                 int __err = errno;
 918
 919                 perror("shmat");
 920                 return __err;
 921         }
 922
 923         td = data + offset * sizeof(struct thread_data);
 924         ret = thread_main(td);
 925         shmdt(data);
 926         return (int) (unsigned long) ret;
 927 }
 928
 929 /*
 930  * Run over the job map and reap the threads that have exited, if any.
 931  */
 932 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 933 {
 934         struct thread_data *td;
 935         int i, cputhreads, pending, status, ret;
 936
 937         /*
 938          * reap exited threads (TD_EXITED -> TD_REAPED)
 939          */
 940         pending = cputhreads = 0;
 941         for_each_td(td, i) {
 942                 int flags = 0;
 943
 944                 /*
 945                  * ->io_ops is NULL for a thread that has closed its
 946                  * io engine
 947                  */
 948                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
 949                         cputhreads++;
 950
 951                 if (!td->pid || td->runstate == TD_REAPED)
 952                         continue;
 953                 if (td->o.use_thread) {
 954                         if (td->runstate == TD_EXITED) {
 955                                 td_set_runstate(td, TD_REAPED);
 956                                 goto reaped;
 957                         }
 958                         continue;
 959                 }
 960
 961                 flags = WNOHANG;
 962                 if (td->runstate == TD_EXITED)
 963                         flags = 0;
 964
 965                 /*
 966                  * check if someone quit or got killed in an unusual way
 967                  */
 968                 ret = waitpid(td->pid, &status, flags);
 969                 if (ret < 0) {
 970                         if (errno == ECHILD) {
 971                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
 972                                 td_set_runstate(td, TD_REAPED);
 973                                 goto reaped;
 974                         }
 975                         perror("waitpid");
 976                 } else if (ret == td->pid) {
 977                         if (WIFSIGNALED(status)) {
 978                                 int sig = WTERMSIG(status);
 979
 980                                 if (sig != SIGQUIT)
 981                                         log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
 982                                 td_set_runstate(td, TD_REAPED);
 983                                 goto reaped;
 984                         }
 985                         if (WIFEXITED(status)) {
 986                                 if (WEXITSTATUS(status) && !td->error)
 987                                         td->error = WEXITSTATUS(status);
 988
 989                                 td_set_runstate(td, TD_REAPED);
 990                                 goto reaped;
 991                         }
 992                 }
 993
 994                 /*
 995                  * thread is not dead, continue
 996                  */
 997                 pending++;
 998                 continue;
 999 reaped:
1000                 if (td->o.use_thread) {
1001                         long ret;
1002
1003                         if (pthread_join(td->thread, (void *) &ret))
1004                                 perror("pthread_join");
1005                 }
1006
1007                 (*nr_running)--;
1008                 (*m_rate) -= td->o.ratemin;
1009                 (*t_rate) -= td->o.rate;
1010                 pending--;
1011
1012                 if (td->error)
1013                         exit_value++;
1014         }
1015
1016         if (*nr_running == cputhreads && !pending)
1017                 terminate_threads(TERMINATE_ALL);
1018 }
1019
1020 /*
1021  * Main function for kicking off and reaping jobs, as needed.
1022  */
1023 static void run_threads(void)
1024 {
1025         struct thread_data *td;
1026         unsigned long spent;
1027         int i, todo, nr_running, m_rate, t_rate, nr_started;
1028
1029         if (fio_pin_memory())
1030                 return;
1031
1032         if (!terse_output) {
1033                 printf("Starting ");
1034                 if (nr_thread)
1035                         printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1036                 if (nr_process) {
1037                         if (nr_thread)
1038                                 printf(" and ");
1039                         printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1040                 }
1041                 printf("\n");
1042                 fflush(stdout);
1043         }
1044
1045         signal(SIGINT, sig_handler);
1046         signal(SIGALRM, sig_handler);
1047
1048         todo = thread_number;
1049         nr_running = 0;
1050         nr_started = 0;
1051         m_rate = t_rate = 0;
1052
1053         for_each_td(td, i) {
1054                 print_status_init(td->thread_number - 1);
1055
1056                 if (!td->o.create_serialize) {
1057                         init_disk_util(td);
1058                         continue;
1059                 }
1060
1061                 /*
1062                  * do file setup here so it happens sequentially,
1063                  * we don't want X number of threads getting their
1064                  * client data interspersed on disk
1065                  */
1066                 if (setup_files(td)) {
1067                         exit_value++;
1068                         if (td->error)
1069                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1070                         td_set_runstate(td, TD_REAPED);
1071                         todo--;
1072                 }
1073
1074                 init_disk_util(td);
1075         }
1076
1077         set_genesis_time();
1078
1079         while (todo) {
1080                 struct thread_data *map[MAX_JOBS];
1081                 struct timeval this_start;
1082                 int this_jobs = 0, left;
1083
1084                 /*
1085                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1086                  */
1087                 for_each_td(td, i) {
1088                         if (td->runstate != TD_NOT_CREATED)
1089                                 continue;
1090
1091                         /*
1092                          * never got a chance to start, killed by other
1093                          * thread for some reason
1094                          */
1095                         if (td->terminate) {
1096                                 todo--;
1097                                 continue;
1098                         }
1099
1100                         if (td->o.start_delay) {
1101                                 spent = mtime_since_genesis();
1102
1103                                 if (td->o.start_delay * 1000 > spent)
1104                                         continue;
1105                         }
1106
1107                         if (td->o.stonewall && (nr_started || nr_running))
1108                                 break;
1109
1110                         /*
1111                          * Set state to created. Thread will transition
1112                          * to TD_INITIALIZED when it's done setting up.
1113                          */
1114                         td_set_runstate(td, TD_CREATED);
1115                         map[this_jobs++] = td;
1116                         nr_started++;
1117
1118                         if (td->o.use_thread) {
1119                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1120                                         perror("thread_create");
1121                                         nr_started--;
1122                                         break;
1123                                 }
1124                         } else {
1125                                 if (!fork()) {
1126                                         int ret = fork_main(shm_id, i);
1127
1128                                         exit(ret);
1129                                 }
1130                         }
1131                         fio_sem_down(startup_sem);
1132                 }
1133
1134                 /*
1135                  * Wait for the started threads to transition to
1136                  * TD_INITIALIZED.
1137                  */
1138                 fio_gettime(&this_start, NULL);
1139                 left = this_jobs;
1140                 while (left && !fio_abort) {
1141                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1142                                 break;
1143
1144                         usleep(100000);
1145
1146                         for (i = 0; i < this_jobs; i++) {
1147                                 td = map[i];
1148                                 if (!td)
1149                                         continue;
1150                                 if (td->runstate == TD_INITIALIZED) {
1151                                         map[i] = NULL;
1152                                         left--;
1153                                 } else if (td->runstate >= TD_EXITED) {
1154                                         map[i] = NULL;
1155                                         left--;
1156                                         todo--;
1157                                         nr_running++; /* work-around... */
1158                                 }
1159                         }
1160                 }
1161
1162                 if (left) {
1163                         log_err("fio: %d jobs failed to start\n", left);
1164                         for (i = 0; i < this_jobs; i++) {
1165                                 td = map[i];
1166                                 if (!td)
1167                                         continue;
1168                                 kill(td->pid, SIGTERM);
1169                         }
1170                         break;
1171                 }
1172
1173                 /*
1174                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1175                  */
1176                 for_each_td(td, i) {
1177                         if (td->runstate != TD_INITIALIZED)
1178                                 continue;
1179
1180                         td_set_runstate(td, TD_RUNNING);
1181                         nr_running++;
1182                         nr_started--;
1183                         m_rate += td->o.ratemin;
1184                         t_rate += td->o.rate;
1185                         todo--;
1186                         fio_sem_up(td->mutex);
1187                 }
1188
1189                 reap_threads(&nr_running, &t_rate, &m_rate);
1190
1191                 if (todo)
1192                         usleep(100000);
1193         }
1194
1195         while (nr_running) {
1196                 reap_threads(&nr_running, &t_rate, &m_rate);
1197                 usleep(10000);
1198         }
1199
1200         update_io_ticks();
1201         fio_unpin_memory();
1202 }
1203
1204 int main(int argc, char *argv[])
1205 {
1206         long ps;
1207
1208         /*
1209          * We need locale for number printing, if it isn't set then just
1210          * go with the US format.
1211          */
1212         if (!getenv("LC_NUMERIC"))
1213                 setlocale(LC_NUMERIC, "en_US");
1214
1215         if (parse_options(argc, argv))
1216                 return 1;
1217
1218         if (!thread_number)
1219                 return 0;
1220
1221         ps = sysconf(_SC_PAGESIZE);
1222         if (ps < 0) {
1223                 log_err("Failed to get page size\n");
1224                 return 1;
1225         }
1226
1227         page_size = ps;
1228         page_mask = ps - 1;
1229
1230         if (write_bw_log) {
1231                 setup_log(&agg_io_log[DDIR_READ]);
1232                 setup_log(&agg_io_log[DDIR_WRITE]);
1233         }
1234
1235         startup_sem = fio_sem_init(0);
1236
1237         set_genesis_time();
1238
1239         disk_util_timer_arm();
1240
1241         run_threads();
1242
1243         if (!fio_abort) {
1244                 show_run_stats();
1245                 if (write_bw_log) {
1246                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1247                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1248                 }
1249         }
1250
1251         fio_sem_remove(startup_sem);
1252         return exit_value;
1253 }