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 "hash.h"
  39
  40 unsigned long page_mask;
  41 unsigned long page_size;
  42 #define ALIGN(buf)      \
  43         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  44
  45 int groupid = 0;
  46 int thread_number = 0;
  47 int nr_process = 0;
  48 int nr_thread = 0;
  49 int shm_id = 0;
  50 int temp_stall_ts;
  51
  52 static struct fio_sem *startup_sem;
  53 static volatile int fio_abort;
  54 static int exit_value;
  55
  56 struct io_log *agg_io_log[2];
  57
  58 #define TERMINATE_ALL           (-1)
  59 #define JOB_START_TIMEOUT       (5 * 1000)
  60
  61 static inline void td_set_runstate(struct thread_data *td, int runstate)
  62 {
  63         td->runstate = runstate;
  64 }
  65
  66 static void terminate_threads(int group_id)
  67 {
  68         struct thread_data *td;
  69         int i;
  70
  71         for_each_td(td, i) {
  72                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  73                         if (td->runstate <= TD_RUNNING)
  74                                 kill(td->pid, SIGQUIT);
  75                         td->terminate = 1;
  76                         td->o.start_delay = 0;
  77                 }
  78         }
  79 }
  80
  81 static void sig_handler(int sig)
  82 {
  83         switch (sig) {
  84                 case SIGALRM:
  85                         update_io_ticks();
  86                         disk_util_timer_arm();
  87                         print_thread_status();
  88                         break;
  89                 default:
  90                         printf("\nfio: terminating on signal %d\n", sig);
  91                         fflush(stdout);
  92                         terminate_threads(TERMINATE_ALL);
  93                         break;
  94         }
  95 }
  96
  97 /*
  98  * Check if we are above the minimum rate given.
  99  */
 100 static int check_min_rate(struct thread_data *td, struct timeval *now)
 101 {
 102         unsigned long long bytes = 0;
 103         unsigned long iops = 0;
 104         unsigned long spent;
 105         unsigned long rate;
 106
 107         /*
 108          * No minimum rate set, always ok
 109          */
 110         if (!td->o.ratemin && !td->o.rate_iops_min)
 111                 return 0;
 112
 113         /*
 114          * allow a 2 second settle period in the beginning
 115          */
 116         if (mtime_since(&td->start, now) < 2000)
 117                 return 0;
 118
 119         if (td_read(td)) {
 120                 iops += td->io_blocks[DDIR_READ];
 121                 bytes += td->this_io_bytes[DDIR_READ];
 122         }
 123         if (td_write(td)) {
 124                 iops += td->io_blocks[DDIR_WRITE];
 125                 bytes += td->this_io_bytes[DDIR_WRITE];
 126         }
 127
 128         /*
 129          * if rate blocks is set, sample is running
 130          */
 131         if (td->rate_bytes || td->rate_blocks) {
 132                 spent = mtime_since(&td->lastrate, now);
 133                 if (spent < td->o.ratecycle)
 134                         return 0;
 135
 136                 if (td->o.rate) {
 137                         /*
 138                          * check bandwidth specified rate
 139                          */
 140                         if (bytes < td->rate_bytes) {
 141                                 log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
 142                                 return 1;
 143                         } else {
 144                                 rate = (bytes - td->rate_bytes) / spent;
 145                                 if (rate < td->o.ratemin || bytes < td->rate_bytes) {
 146                                         log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
 147                                         return 1;
 148                                 }
 149                         }
 150                 } else {
 151                         /*
 152                          * checks iops specified rate
 153                          */
 154                         if (iops < td->o.rate_iops) {
 155                                 log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
 156                                 return 1;
 157                         } else {
 158                                 rate = (iops - td->rate_blocks) / spent;
 159                                 if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
 160                                         log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
 161                                 }
 162                         }
 163                 }
 164         }
 165
 166         td->rate_bytes = bytes;
 167         td->rate_blocks = iops;
 168         memcpy(&td->lastrate, now, sizeof(*now));
 169         return 0;
 170 }
 171
 172 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 173 {
 174         if (!td->o.timeout)
 175                 return 0;
 176         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 177                 return 1;
 178
 179         return 0;
 180 }
 181
 182 /*
 183  * When job exits, we can cancel the in-flight IO if we are using async
 184  * io. Attempt to do so.
 185  */
 186 static void cleanup_pending_aio(struct thread_data *td)
 187 {
 188         struct list_head *entry, *n;
 189         struct io_u *io_u;
 190         int r;
 191
 192         /*
 193          * get immediately available events, if any
 194          */
 195         r = io_u_queued_complete(td, 0);
 196         if (r < 0)
 197                 return;
 198
 199         /*
 200          * now cancel remaining active events
 201          */
 202         if (td->io_ops->cancel) {
 203                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 204                         io_u = list_entry(entry, struct io_u, list);
 205
 206                         /*
 207                          * if the io_u isn't in flight, then that generally
 208                          * means someone leaked an io_u. complain but fix
 209                          * it up, so we don't stall here.
 210                          */
 211                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 212                                 log_err("fio: non-busy IO on busy list\n");
 213                                 put_io_u(td, io_u);
 214                         } else {
 215                                 r = td->io_ops->cancel(td, io_u);
 216                                 if (!r)
 217                                         put_io_u(td, io_u);
 218                         }
 219                 }
 220         }
 221
 222         if (td->cur_depth)
 223                 r = io_u_queued_complete(td, td->cur_depth);
 224 }
 225
 226 /*
 227  * Helper to handle the final sync of a file. Works just like the normal
 228  * io path, just does everything sync.
 229  */
 230 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 231 {
 232         struct io_u *io_u = __get_io_u(td);
 233         int ret;
 234
 235         if (!io_u)
 236                 return 1;
 237
 238         io_u->ddir = DDIR_SYNC;
 239         io_u->file = f;
 240
 241         if (td_io_prep(td, io_u)) {
 242                 put_io_u(td, io_u);
 243                 return 1;
 244         }
 245
 246 requeue:
 247         ret = td_io_queue(td, io_u);
 248         if (ret < 0) {
 249                 td_verror(td, io_u->error, "td_io_queue");
 250                 put_io_u(td, io_u);
 251                 return 1;
 252         } else if (ret == FIO_Q_QUEUED) {
 253                 if (io_u_queued_complete(td, 1) < 0)
 254                         return 1;
 255         } else if (ret == FIO_Q_COMPLETED) {
 256                 if (io_u->error) {
 257                         td_verror(td, io_u->error, "td_io_queue");
 258                         return 1;
 259                 }
 260
 261                 if (io_u_sync_complete(td, io_u) < 0)
 262                         return 1;
 263         } else if (ret == FIO_Q_BUSY) {
 264                 if (td_io_commit(td))
 265                         return 1;
 266                 goto requeue;
 267         }
 268
 269         return 0;
 270 }
 271
 272 /*
 273  * The main verify engine. Runs over the writes we previously submitted,
 274  * reads the blocks back in, and checks the crc/md5 of the data.
 275  */
 276 static void do_verify(struct thread_data *td)
 277 {
 278         struct fio_file *f;
 279         struct io_u *io_u;
 280         int ret, min_events;
 281         unsigned int i;
 282
 283         /*
 284          * sync io first and invalidate cache, to make sure we really
 285          * read from disk.
 286          */
 287         for_each_file(td, f, i) {
 288                 if (!(f->flags & FIO_FILE_OPEN))
 289                         continue;
 290                 if (fio_io_sync(td, f))
 291                         break;
 292                 if (file_invalidate_cache(td, f))
 293                         break;
 294         }
 295
 296         if (td->error)
 297                 return;
 298
 299         td_set_runstate(td, TD_VERIFYING);
 300
 301         io_u = NULL;
 302         while (!td->terminate) {
 303                 int ret2;
 304
 305                 io_u = __get_io_u(td);
 306                 if (!io_u)
 307                         break;
 308
 309                 if (runtime_exceeded(td, &io_u->start_time)) {
 310                         put_io_u(td, io_u);
 311                         td->terminate = 1;
 312                         break;
 313                 }
 314
 315                 if (get_next_verify(td, io_u)) {
 316                         put_io_u(td, io_u);
 317                         break;
 318                 }
 319
 320                 if (td_io_prep(td, io_u)) {
 321                         put_io_u(td, io_u);
 322                         break;
 323                 }
 324
 325                 io_u->end_io = verify_io_u;
 326
 327                 ret = td_io_queue(td, io_u);
 328                 switch (ret) {
 329                 case FIO_Q_COMPLETED:
 330                         if (io_u->error)
 331                                 ret = -io_u->error;
 332                         else if (io_u->resid) {
 333                                 int bytes = io_u->xfer_buflen - io_u->resid;
 334                                 struct fio_file *f = io_u->file;
 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                                 f->last_completed_pos = io_u->offset;
 349
 350                                 td->ts.short_io_u[io_u->ddir]++;
 351
 352                                 if (io_u->offset == f->real_file_size)
 353                                         goto sync_done;
 354
 355                                 requeue_io_u(td, &io_u);
 356                         } else {
 357 sync_done:
 358                                 ret = io_u_sync_complete(td, io_u);
 359                                 if (ret < 0)
 360                                         break;
 361                         }
 362                         continue;
 363                 case FIO_Q_QUEUED:
 364                         break;
 365                 case FIO_Q_BUSY:
 366                         requeue_io_u(td, &io_u);
 367                         ret2 = td_io_commit(td);
 368                         if (ret2 < 0)
 369                                 ret = ret2;
 370                         break;
 371                 default:
 372                         assert(ret < 0);
 373                         td_verror(td, -ret, "td_io_queue");
 374                         break;
 375                 }
 376
 377                 if (ret < 0 || td->error)
 378                         break;
 379
 380                 /*
 381                  * if we can queue more, do so. but check if there are
 382                  * completed io_u's first.
 383                  */
 384                 min_events = 0;
 385                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 386                         min_events = 1;
 387
 388                         if (td->cur_depth > td->o.iodepth_low)
 389                                 min_events = td->cur_depth - td->o.iodepth_low;
 390                 }
 391
 392                 /*
 393                  * Reap required number of io units, if any, and do the
 394                  * verification on them through the callback handler
 395                  */
 396                 if (io_u_queued_complete(td, min_events) < 0)
 397                         break;
 398         }
 399
 400         if (!td->error) {
 401                 min_events = td->cur_depth;
 402
 403                 if (min_events)
 404                         ret = io_u_queued_complete(td, min_events);
 405         } else
 406                 cleanup_pending_aio(td);
 407
 408         td_set_runstate(td, TD_RUNNING);
 409 }
 410
 411 /*
 412  * Main IO worker function. It retrieves io_u's to process and queues
 413  * and reaps them, checking for rate and errors along the way.
 414  */
 415 static void do_io(struct thread_data *td)
 416 {
 417         struct timeval s;
 418         unsigned long usec;
 419         unsigned int i;
 420         int ret = 0;
 421
 422         td_set_runstate(td, TD_RUNNING);
 423
 424         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 425                 struct timeval comp_time;
 426                 long bytes_done = 0;
 427                 int min_evts = 0;
 428                 struct io_u *io_u;
 429                 int ret2;
 430
 431                 if (td->terminate)
 432                         break;
 433
 434                 io_u = get_io_u(td);
 435                 if (!io_u)
 436                         break;
 437
 438                 memcpy(&s, &io_u->start_time, sizeof(s));
 439
 440                 if (runtime_exceeded(td, &s)) {
 441                         put_io_u(td, io_u);
 442                         td->terminate = 1;
 443                         break;
 444                 }
 445
 446                 ret = td_io_queue(td, io_u);
 447                 switch (ret) {
 448                 case FIO_Q_COMPLETED:
 449                         if (io_u->error)
 450                                 ret = -io_u->error;
 451                         else if (io_u->resid) {
 452                                 int bytes = io_u->xfer_buflen - io_u->resid;
 453                                 struct fio_file *f = io_u->file;
 454
 455                                 /*
 456                                  * zero read, fail
 457                                  */
 458                                 if (!bytes) {
 459                                         td_verror(td, ENODATA, "full resid");
 460                                         put_io_u(td, io_u);
 461                                         break;
 462                                 }
 463
 464                                 io_u->xfer_buflen = io_u->resid;
 465                                 io_u->xfer_buf += bytes;
 466                                 io_u->offset += bytes;
 467                                 f->last_completed_pos = io_u->offset;
 468
 469                                 td->ts.short_io_u[io_u->ddir]++;
 470
 471                                 if (io_u->offset == f->real_file_size)
 472                                         goto sync_done;
 473
 474                                 requeue_io_u(td, &io_u);
 475                         } else {
 476 sync_done:
 477                                 fio_gettime(&comp_time, NULL);
 478                                 bytes_done = io_u_sync_complete(td, io_u);
 479                                 if (bytes_done < 0)
 480                                         ret = bytes_done;
 481                         }
 482                         break;
 483                 case FIO_Q_QUEUED:
 484                         /*
 485                          * if the engine doesn't have a commit hook,
 486                          * the io_u is really queued. if it does have such
 487                          * a hook, it has to call io_u_queued() itself.
 488                          */
 489                         if (td->io_ops->commit == NULL)
 490                                 io_u_queued(td, io_u);
 491                         break;
 492                 case FIO_Q_BUSY:
 493                         requeue_io_u(td, &io_u);
 494                         ret2 = td_io_commit(td);
 495                         if (ret2 < 0)
 496                                 ret = ret2;
 497                         break;
 498                 default:
 499                         assert(ret < 0);
 500                         put_io_u(td, io_u);
 501                         break;
 502                 }
 503
 504                 if (ret < 0 || td->error)
 505                         break;
 506
 507                 /*
 508                  * See if we need to complete some commands
 509                  */
 510                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 511                         min_evts = 0;
 512                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 513                                 min_evts = 1;
 514
 515                                 if (td->cur_depth > td->o.iodepth_low)
 516                                         min_evts = td->cur_depth - td->o.iodepth_low;
 517                         }
 518
 519                         fio_gettime(&comp_time, NULL);
 520                         bytes_done = io_u_queued_complete(td, min_evts);
 521                         if (bytes_done < 0)
 522                                 break;
 523                 }
 524
 525                 if (!bytes_done)
 526                         continue;
 527
 528                 /*
 529                  * the rate is batched for now, it should work for batches
 530                  * of completions except the very first one which may look
 531                  * a little bursty
 532                  */
 533                 usec = utime_since(&s, &comp_time);
 534
 535                 rate_throttle(td, usec, bytes_done);
 536
 537                 if (check_min_rate(td, &comp_time)) {
 538                         if (exitall_on_terminate)
 539                                 terminate_threads(td->groupid);
 540                         td_verror(td, ENODATA, "check_min_rate");
 541                         break;
 542                 }
 543
 544                 if (td->o.thinktime) {
 545                         unsigned long long b;
 546
 547                         b = td->io_blocks[0] + td->io_blocks[1];
 548                         if (!(b % td->o.thinktime_blocks)) {
 549                                 int left;
 550
 551                                 if (td->o.thinktime_spin)
 552                                         __usec_sleep(td->o.thinktime_spin);
 553
 554                                 left = td->o.thinktime - td->o.thinktime_spin;
 555                                 if (left)
 556                                         usec_sleep(td, left);
 557                         }
 558                 }
 559         }
 560
 561         if (!td->error) {
 562                 struct fio_file *f;
 563
 564                 i = td->cur_depth;
 565                 if (i)
 566                         ret = io_u_queued_complete(td, i);
 567
 568                 if (should_fsync(td) && td->o.end_fsync) {
 569                         td_set_runstate(td, TD_FSYNCING);
 570
 571                         for_each_file(td, f, i) {
 572                                 if (!(f->flags & FIO_FILE_OPEN))
 573                                         continue;
 574                                 fio_io_sync(td, f);
 575                         }
 576                 }
 577         } else
 578                 cleanup_pending_aio(td);
 579 }
 580
 581 static void cleanup_io_u(struct thread_data *td)
 582 {
 583         struct list_head *entry, *n;
 584         struct io_u *io_u;
 585
 586         list_for_each_safe(entry, n, &td->io_u_freelist) {
 587                 io_u = list_entry(entry, struct io_u, list);
 588
 589                 list_del(&io_u->list);
 590                 free(io_u);
 591         }
 592
 593         free_io_mem(td);
 594 }
 595
 596 /*
 597  * "randomly" fill the buffer contents
 598  */
 599 static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
 600 {
 601         long *ptr = io_u->buf;
 602
 603         if (!td->o.zero_buffers) {
 604                 while ((void *) ptr - io_u->buf < max_bs) {
 605                         *ptr = rand() * GOLDEN_RATIO_PRIME;
 606                         ptr++;
 607                 }
 608         } else
 609                 memset(ptr, 0, max_bs);
 610 }
 611
 612 static int init_io_u(struct thread_data *td)
 613 {
 614         unsigned long long buf_size;
 615         struct io_u *io_u;
 616         unsigned int max_bs;
 617         int i, max_units;
 618         char *p;
 619
 620         if (td->io_ops->flags & FIO_SYNCIO)
 621                 max_units = 1;
 622         else
 623                 max_units = td->o.iodepth;
 624
 625         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 626         buf_size = (unsigned long long) max_bs * (unsigned long long) max_units;
 627         buf_size += page_mask;
 628         if (buf_size != (size_t) buf_size) {
 629                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 630                 return 1;
 631         }
 632
 633         td->orig_buffer_size = buf_size;
 634
 635         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
 636                 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
 637         else if (td->orig_buffer_size & page_mask)
 638                 td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
 639
 640         if (allocate_io_mem(td))
 641                 return 1;
 642
 643         p = ALIGN(td->orig_buffer);
 644         for (i = 0; i < max_units; i++) {
 645                 io_u = malloc(sizeof(*io_u));
 646                 memset(io_u, 0, sizeof(*io_u));
 647                 INIT_LIST_HEAD(&io_u->list);
 648
 649                 io_u->buf = p + max_bs * i;
 650
 651                 if (td_write(td))
 652                         fill_io_buf(td, io_u, max_bs);
 653
 654                 io_u->index = i;
 655                 io_u->flags = IO_U_F_FREE;
 656                 list_add(&io_u->list, &td->io_u_freelist);
 657         }
 658
 659         io_u_init_timeout();
 660
 661         return 0;
 662 }
 663
 664 static int switch_ioscheduler(struct thread_data *td)
 665 {
 666         char tmp[256], tmp2[128];
 667         FILE *f;
 668         int ret;
 669
 670         if (td->io_ops->flags & FIO_DISKLESSIO)
 671                 return 0;
 672
 673         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 674
 675         f = fopen(tmp, "r+");
 676         if (!f) {
 677                 if (errno == ENOENT) {
 678                         log_err("fio: os or kernel doesn't support IO scheduler switching\n");
 679                         return 0;
 680                 }
 681                 td_verror(td, errno, "fopen iosched");
 682                 return 1;
 683         }
 684
 685         /*
 686          * Set io scheduler.
 687          */
 688         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 689         if (ferror(f) || ret != 1) {
 690                 td_verror(td, errno, "fwrite");
 691                 fclose(f);
 692                 return 1;
 693         }
 694
 695         rewind(f);
 696
 697         /*
 698          * Read back and check that the selected scheduler is now the default.
 699          */
 700         ret = fread(tmp, 1, sizeof(tmp), f);
 701         if (ferror(f) || ret < 0) {
 702                 td_verror(td, errno, "fread");
 703                 fclose(f);
 704                 return 1;
 705         }
 706
 707         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 708         if (!strstr(tmp, tmp2)) {
 709                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 710                 td_verror(td, EINVAL, "iosched_switch");
 711                 fclose(f);
 712                 return 1;
 713         }
 714
 715         fclose(f);
 716         return 0;
 717 }
 718
 719 static int clear_io_state(struct thread_data *td)
 720 {
 721         struct fio_file *f;
 722         unsigned int i;
 723         int ret;
 724
 725         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 726         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 727         td->zone_bytes = 0;
 728         td->rate_bytes = 0;
 729         td->rate_blocks = 0;
 730         td->rw_end_set[0] = td->rw_end_set[1] = 0;
 731
 732         td->last_was_sync = 0;
 733
 734         for_each_file(td, f, i)
 735                 td_io_close_file(td, f);
 736
 737         ret = 0;
 738         for_each_file(td, f, i) {
 739                 ret = td_io_open_file(td, f);
 740                 if (ret)
 741                         break;
 742         }
 743
 744         return ret;
 745 }
 746
 747 /*
 748  * Entry point for the thread based jobs. The process based jobs end up
 749  * here as well, after a little setup.
 750  */
 751 static void *thread_main(void *data)
 752 {
 753         unsigned long long runtime[2];
 754         struct thread_data *td = data;
 755         unsigned long elapsed;
 756         int clear_state;
 757
 758         if (!td->o.use_thread)
 759                 setsid();
 760
 761         td->pid = getpid();
 762
 763         INIT_LIST_HEAD(&td->io_u_freelist);
 764         INIT_LIST_HEAD(&td->io_u_busylist);
 765         INIT_LIST_HEAD(&td->io_u_requeues);
 766         INIT_LIST_HEAD(&td->io_log_list);
 767         INIT_LIST_HEAD(&td->io_hist_list);
 768         td->io_hist_tree = RB_ROOT;
 769
 770         if (init_io_u(td))
 771                 goto err_sem;
 772
 773         if (fio_setaffinity(td) == -1) {
 774                 td_verror(td, errno, "cpu_set_affinity");
 775                 goto err_sem;
 776         }
 777
 778         if (init_iolog(td))
 779                 goto err_sem;
 780
 781         if (td->ioprio) {
 782                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 783                         td_verror(td, errno, "ioprio_set");
 784                         goto err_sem;
 785                 }
 786         }
 787
 788         if (nice(td->o.nice) == -1) {
 789                 td_verror(td, errno, "nice");
 790                 goto err_sem;
 791         }
 792
 793         if (td->o.ioscheduler && switch_ioscheduler(td))
 794                 goto err_sem;
 795
 796         td_set_runstate(td, TD_INITIALIZED);
 797         fio_sem_up(startup_sem);
 798         fio_sem_down(td->mutex);
 799
 800         /*
 801          * the ->mutex semaphore is now no longer used, close it to avoid
 802          * eating a file descriptor
 803          */
 804         fio_sem_remove(td->mutex);
 805
 806         if (!td->o.create_serialize && setup_files(td))
 807                 goto err;
 808
 809         if (td_io_init(td))
 810                 goto err;
 811
 812         if (open_files(td))
 813                 goto err;
 814
 815         if (init_random_map(td))
 816                 goto err;
 817
 818         if (td->o.exec_prerun) {
 819                 if (system(td->o.exec_prerun) < 0)
 820                         goto err;
 821         }
 822
 823         fio_gettime(&td->epoch, NULL);
 824         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 825         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 826
 827         runtime[0] = runtime[1] = 0;
 828         clear_state = 0;
 829         while (td->o.time_based || td->o.loops--) {
 830                 fio_gettime(&td->start, NULL);
 831                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 832
 833                 if (td->o.ratemin)
 834                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 835
 836                 if (clear_state && clear_io_state(td))
 837                         break;
 838
 839                 prune_io_piece_log(td);
 840
 841                 do_io(td);
 842
 843                 clear_state = 1;
 844
 845                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
 846                         if (td->rw_end_set[DDIR_READ])
 847                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
 848                         else
 849                                 elapsed = utime_since_now(&td->start);
 850
 851                         runtime[DDIR_READ] += elapsed;
 852                 }
 853                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
 854                         if (td->rw_end_set[DDIR_WRITE])
 855                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
 856                         else
 857                                 elapsed = utime_since_now(&td->start);
 858
 859                         runtime[DDIR_WRITE] += elapsed;
 860                 }
 861
 862                 if (td->error || td->terminate)
 863                         break;
 864
 865                 if (td->o.verify == VERIFY_NONE)
 866                         continue;
 867
 868                 if (clear_io_state(td))
 869                         break;
 870
 871                 fio_gettime(&td->start, NULL);
 872
 873                 do_verify(td);
 874
 875                 runtime[DDIR_READ] += utime_since_now(&td->start);
 876
 877                 if (td->error || td->terminate)
 878                         break;
 879         }
 880
 881         update_rusage_stat(td);
 882         td->ts.runtime[0] = runtime[0] / 1000;
 883         td->ts.runtime[1] = runtime[1] / 1000;
 884         td->ts.total_run_time = mtime_since_now(&td->epoch);
 885         td->ts.io_bytes[0] = td->io_bytes[0];
 886         td->ts.io_bytes[1] = td->io_bytes[1];
 887
 888         if (td->ts.bw_log)
 889                 finish_log(td, td->ts.bw_log, "bw");
 890         if (td->ts.slat_log)
 891                 finish_log(td, td->ts.slat_log, "slat");
 892         if (td->ts.clat_log)
 893                 finish_log(td, td->ts.clat_log, "clat");
 894         if (td->o.write_iolog_file)
 895                 write_iolog_close(td);
 896         if (td->o.exec_postrun) {
 897                 if (system(td->o.exec_postrun) < 0)
 898                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
 899         }
 900
 901         if (exitall_on_terminate)
 902                 terminate_threads(td->groupid);
 903
 904 err:
 905         if (td->error)
 906                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 907         close_files(td);
 908         close_ioengine(td);
 909         cleanup_io_u(td);
 910         options_mem_free(td);
 911         td_set_runstate(td, TD_EXITED);
 912         return (void *) (unsigned long) td->error;
 913 err_sem:
 914         fio_sem_up(startup_sem);
 915         goto err;
 916 }
 917
 918 /*
 919  * We cannot pass the td data into a forked process, so attach the td and
 920  * pass it to the thread worker.
 921  */
 922 static int fork_main(int shmid, int offset)
 923 {
 924         struct thread_data *td;
 925         void *data, *ret;
 926
 927         data = shmat(shmid, NULL, 0);
 928         if (data == (void *) -1) {
 929                 int __err = errno;
 930
 931                 perror("shmat");
 932                 return __err;
 933         }
 934
 935         td = data + offset * sizeof(struct thread_data);
 936         ret = thread_main(td);
 937         shmdt(data);
 938         return (int) (unsigned long) ret;
 939 }
 940
 941 /*
 942  * Run over the job map and reap the threads that have exited, if any.
 943  */
 944 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 945 {
 946         struct thread_data *td;
 947         int i, cputhreads, pending, status, ret;
 948
 949         /*
 950          * reap exited threads (TD_EXITED -> TD_REAPED)
 951          */
 952         pending = cputhreads = 0;
 953         for_each_td(td, i) {
 954                 int flags = 0;
 955
 956                 /*
 957                  * ->io_ops is NULL for a thread that has closed its
 958                  * io engine
 959                  */
 960                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
 961                         cputhreads++;
 962
 963                 if (!td->pid || td->runstate == TD_REAPED)
 964                         continue;
 965                 if (td->o.use_thread) {
 966                         if (td->runstate == TD_EXITED) {
 967                                 td_set_runstate(td, TD_REAPED);
 968                                 goto reaped;
 969                         }
 970                         continue;
 971                 }
 972
 973                 flags = WNOHANG;
 974                 if (td->runstate == TD_EXITED)
 975                         flags = 0;
 976
 977                 /*
 978                  * check if someone quit or got killed in an unusual way
 979                  */
 980                 ret = waitpid(td->pid, &status, flags);
 981                 if (ret < 0) {
 982                         if (errno == ECHILD) {
 983                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
 984                                 td_set_runstate(td, TD_REAPED);
 985                                 goto reaped;
 986                         }
 987                         perror("waitpid");
 988                 } else if (ret == td->pid) {
 989                         if (WIFSIGNALED(status)) {
 990                                 int sig = WTERMSIG(status);
 991
 992                                 if (sig != SIGQUIT)
 993                                         log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
 994                                 td_set_runstate(td, TD_REAPED);
 995                                 goto reaped;
 996                         }
 997                         if (WIFEXITED(status)) {
 998                                 if (WEXITSTATUS(status) && !td->error)
 999                                         td->error = WEXITSTATUS(status);
1000
1001                                 td_set_runstate(td, TD_REAPED);
1002                                 goto reaped;
1003                         }
1004                 }
1005
1006                 /*
1007                  * thread is not dead, continue
1008                  */
1009                 pending++;
1010                 continue;
1011 reaped:
1012                 if (td->o.use_thread) {
1013                         long ret;
1014
1015                         if (pthread_join(td->thread, (void *) &ret))
1016                                 perror("pthread_join");
1017                 }
1018
1019                 (*nr_running)--;
1020                 (*m_rate) -= td->o.ratemin;
1021                 (*t_rate) -= td->o.rate;
1022                 pending--;
1023
1024                 if (td->error)
1025                         exit_value++;
1026         }
1027
1028         if (*nr_running == cputhreads && !pending)
1029                 terminate_threads(TERMINATE_ALL);
1030 }
1031
1032 /*
1033  * Main function for kicking off and reaping jobs, as needed.
1034  */
1035 static void run_threads(void)
1036 {
1037         struct thread_data *td;
1038         unsigned long spent;
1039         int i, todo, nr_running, m_rate, t_rate, nr_started;
1040
1041         if (fio_pin_memory())
1042                 return;
1043
1044         if (!terse_output) {
1045                 printf("Starting ");
1046                 if (nr_thread)
1047                         printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1048                 if (nr_process) {
1049                         if (nr_thread)
1050                                 printf(" and ");
1051                         printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1052                 }
1053                 printf("\n");
1054                 fflush(stdout);
1055         }
1056
1057         signal(SIGINT, sig_handler);
1058         signal(SIGALRM, sig_handler);
1059
1060         todo = thread_number;
1061         nr_running = 0;
1062         nr_started = 0;
1063         m_rate = t_rate = 0;
1064
1065         for_each_td(td, i) {
1066                 print_status_init(td->thread_number - 1);
1067
1068                 if (!td->o.create_serialize) {
1069                         init_disk_util(td);
1070                         continue;
1071                 }
1072
1073                 /*
1074                  * do file setup here so it happens sequentially,
1075                  * we don't want X number of threads getting their
1076                  * client data interspersed on disk
1077                  */
1078                 if (setup_files(td)) {
1079                         exit_value++;
1080                         if (td->error)
1081                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1082                         td_set_runstate(td, TD_REAPED);
1083                         todo--;
1084                 }
1085
1086                 init_disk_util(td);
1087         }
1088
1089         set_genesis_time();
1090
1091         while (todo) {
1092                 struct thread_data *map[MAX_JOBS];
1093                 struct timeval this_start;
1094                 int this_jobs = 0, left;
1095
1096                 /*
1097                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1098                  */
1099                 for_each_td(td, i) {
1100                         if (td->runstate != TD_NOT_CREATED)
1101                                 continue;
1102
1103                         /*
1104                          * never got a chance to start, killed by other
1105                          * thread for some reason
1106                          */
1107                         if (td->terminate) {
1108                                 todo--;
1109                                 continue;
1110                         }
1111
1112                         if (td->o.start_delay) {
1113                                 spent = mtime_since_genesis();
1114
1115                                 if (td->o.start_delay * 1000 > spent)
1116                                         continue;
1117                         }
1118
1119                         if (td->o.stonewall && (nr_started || nr_running))
1120                                 break;
1121
1122                         /*
1123                          * Set state to created. Thread will transition
1124                          * to TD_INITIALIZED when it's done setting up.
1125                          */
1126                         td_set_runstate(td, TD_CREATED);
1127                         map[this_jobs++] = td;
1128                         nr_started++;
1129
1130                         if (td->o.use_thread) {
1131                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1132                                         perror("thread_create");
1133                                         nr_started--;
1134                                         break;
1135                                 }
1136                         } else {
1137                                 if (!fork()) {
1138                                         int ret = fork_main(shm_id, i);
1139
1140                                         exit(ret);
1141                                 }
1142                         }
1143                         fio_sem_down(startup_sem);
1144                 }
1145
1146                 /*
1147                  * Wait for the started threads to transition to
1148                  * TD_INITIALIZED.
1149                  */
1150                 fio_gettime(&this_start, NULL);
1151                 left = this_jobs;
1152                 while (left && !fio_abort) {
1153                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1154                                 break;
1155
1156                         usleep(100000);
1157
1158                         for (i = 0; i < this_jobs; i++) {
1159                                 td = map[i];
1160                                 if (!td)
1161                                         continue;
1162                                 if (td->runstate == TD_INITIALIZED) {
1163                                         map[i] = NULL;
1164                                         left--;
1165                                 } else if (td->runstate >= TD_EXITED) {
1166                                         map[i] = NULL;
1167                                         left--;
1168                                         todo--;
1169                                         nr_running++; /* work-around... */
1170                                 }
1171                         }
1172                 }
1173
1174                 if (left) {
1175                         log_err("fio: %d jobs failed to start\n", left);
1176                         for (i = 0; i < this_jobs; i++) {
1177                                 td = map[i];
1178                                 if (!td)
1179                                         continue;
1180                                 kill(td->pid, SIGTERM);
1181                         }
1182                         break;
1183                 }
1184
1185                 /*
1186                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1187                  */
1188                 for_each_td(td, i) {
1189                         if (td->runstate != TD_INITIALIZED)
1190                                 continue;
1191
1192                         td_set_runstate(td, TD_RUNNING);
1193                         nr_running++;
1194                         nr_started--;
1195                         m_rate += td->o.ratemin;
1196                         t_rate += td->o.rate;
1197                         todo--;
1198                         fio_sem_up(td->mutex);
1199                 }
1200
1201                 reap_threads(&nr_running, &t_rate, &m_rate);
1202
1203                 if (todo)
1204                         usleep(100000);
1205         }
1206
1207         while (nr_running) {
1208                 reap_threads(&nr_running, &t_rate, &m_rate);
1209                 usleep(10000);
1210         }
1211
1212         update_io_ticks();
1213         fio_unpin_memory();
1214 }
1215
1216 int main(int argc, char *argv[])
1217 {
1218         long ps;
1219
1220         /*
1221          * We need locale for number printing, if it isn't set then just
1222          * go with the US format.
1223          */
1224         if (!getenv("LC_NUMERIC"))
1225                 setlocale(LC_NUMERIC, "en_US");
1226
1227         if (parse_options(argc, argv))
1228                 return 1;
1229
1230         if (!thread_number)
1231                 return 0;
1232
1233         ps = sysconf(_SC_PAGESIZE);
1234         if (ps < 0) {
1235                 log_err("Failed to get page size\n");
1236                 return 1;
1237         }
1238
1239         page_size = ps;
1240         page_mask = ps - 1;
1241
1242         if (write_bw_log) {
1243                 setup_log(&agg_io_log[DDIR_READ]);
1244                 setup_log(&agg_io_log[DDIR_WRITE]);
1245         }
1246
1247         startup_sem = fio_sem_init(0);
1248
1249         set_genesis_time();
1250
1251         disk_util_timer_arm();
1252
1253         run_threads();
1254
1255         if (!fio_abort) {
1256                 show_run_stats();
1257                 if (write_bw_log) {
1258                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1259                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1260                 }
1261         }
1262
1263         fio_sem_remove(startup_sem);
1264         return exit_value;
1265 }