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                         break;
 312                 }
 313
 314                 if (get_next_verify(td, io_u)) {
 315                         put_io_u(td, io_u);
 316                         break;
 317                 }
 318
 319                 if (td_io_prep(td, io_u)) {
 320                         put_io_u(td, io_u);
 321                         break;
 322                 }
 323
 324                 io_u->end_io = verify_io_u;
 325
 326                 ret = td_io_queue(td, io_u);
 327                 switch (ret) {
 328                 case FIO_Q_COMPLETED:
 329                         if (io_u->error)
 330                                 ret = -io_u->error;
 331                         else if (io_u->resid) {
 332                                 int bytes = io_u->xfer_buflen - io_u->resid;
 333                                 struct fio_file *f = io_u->file;
 334
 335                                 /*
 336                                  * zero read, fail
 337                                  */
 338                                 if (!bytes) {
 339                                         td_verror(td, ENODATA, "full resid");
 340                                         put_io_u(td, io_u);
 341                                         break;
 342                                 }
 343
 344                                 io_u->xfer_buflen = io_u->resid;
 345                                 io_u->xfer_buf += bytes;
 346                                 io_u->offset += bytes;
 347                                 f->last_completed_pos = io_u->offset;
 348
 349                                 td->ts.short_io_u[io_u->ddir]++;
 350
 351                                 if (io_u->offset == f->real_file_size)
 352                                         goto sync_done;
 353
 354                                 requeue_io_u(td, &io_u);
 355                         } else {
 356 sync_done:
 357                                 ret = io_u_sync_complete(td, io_u);
 358                                 if (ret < 0)
 359                                         break;
 360                         }
 361                         continue;
 362                 case FIO_Q_QUEUED:
 363                         break;
 364                 case FIO_Q_BUSY:
 365                         requeue_io_u(td, &io_u);
 366                         ret2 = td_io_commit(td);
 367                         if (ret2 < 0)
 368                                 ret = ret2;
 369                         break;
 370                 default:
 371                         assert(ret < 0);
 372                         td_verror(td, -ret, "td_io_queue");
 373                         break;
 374                 }
 375
 376                 if (ret < 0 || td->error)
 377                         break;
 378
 379                 /*
 380                  * if we can queue more, do so. but check if there are
 381                  * completed io_u's first.
 382                  */
 383                 min_events = 0;
 384                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 385                         min_events = 1;
 386
 387                         if (td->cur_depth > td->o.iodepth_low)
 388                                 min_events = td->cur_depth - td->o.iodepth_low;
 389                 }
 390
 391                 /*
 392                  * Reap required number of io units, if any, and do the
 393                  * verification on them through the callback handler
 394                  */
 395                 if (io_u_queued_complete(td, min_events) < 0)
 396                         break;
 397         }
 398
 399         if (!td->error) {
 400                 min_events = td->cur_depth;
 401
 402                 if (min_events)
 403                         ret = io_u_queued_complete(td, min_events);
 404         } else
 405                 cleanup_pending_aio(td);
 406
 407         td_set_runstate(td, TD_RUNNING);
 408 }
 409
 410 /*
 411  * Main IO worker function. It retrieves io_u's to process and queues
 412  * and reaps them, checking for rate and errors along the way.
 413  */
 414 static void do_io(struct thread_data *td)
 415 {
 416         struct timeval s;
 417         unsigned long usec;
 418         unsigned int i;
 419         int ret = 0;
 420
 421         td_set_runstate(td, TD_RUNNING);
 422
 423         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 424                 struct timeval comp_time;
 425                 long bytes_done = 0;
 426                 int min_evts = 0;
 427                 struct io_u *io_u;
 428                 int ret2;
 429
 430                 if (td->terminate)
 431                         break;
 432
 433                 io_u = get_io_u(td);
 434                 if (!io_u)
 435                         break;
 436
 437                 memcpy(&s, &io_u->start_time, sizeof(s));
 438
 439                 if (runtime_exceeded(td, &s)) {
 440                         put_io_u(td, io_u);
 441                         break;
 442                 }
 443
 444                 ret = td_io_queue(td, io_u);
 445                 switch (ret) {
 446                 case FIO_Q_COMPLETED:
 447                         if (io_u->error)
 448                                 ret = -io_u->error;
 449                         else if (io_u->resid) {
 450                                 int bytes = io_u->xfer_buflen - io_u->resid;
 451                                 struct fio_file *f = io_u->file;
 452
 453                                 /*
 454                                  * zero read, fail
 455                                  */
 456                                 if (!bytes) {
 457                                         td_verror(td, ENODATA, "full resid");
 458                                         put_io_u(td, io_u);
 459                                         break;
 460                                 }
 461
 462                                 io_u->xfer_buflen = io_u->resid;
 463                                 io_u->xfer_buf += bytes;
 464                                 io_u->offset += bytes;
 465                                 f->last_completed_pos = io_u->offset;
 466
 467                                 td->ts.short_io_u[io_u->ddir]++;
 468
 469                                 if (io_u->offset == f->real_file_size)
 470                                         goto sync_done;
 471
 472                                 requeue_io_u(td, &io_u);
 473                         } else {
 474 sync_done:
 475                                 fio_gettime(&comp_time, NULL);
 476                                 bytes_done = io_u_sync_complete(td, io_u);
 477                                 if (bytes_done < 0)
 478                                         ret = bytes_done;
 479                         }
 480                         break;
 481                 case FIO_Q_QUEUED:
 482                         /*
 483                          * if the engine doesn't have a commit hook,
 484                          * the io_u is really queued. if it does have such
 485                          * a hook, it has to call io_u_queued() itself.
 486                          */
 487                         if (td->io_ops->commit == NULL)
 488                                 io_u_queued(td, io_u);
 489                         break;
 490                 case FIO_Q_BUSY:
 491                         requeue_io_u(td, &io_u);
 492                         ret2 = td_io_commit(td);
 493                         if (ret2 < 0)
 494                                 ret = ret2;
 495                         break;
 496                 default:
 497                         assert(ret < 0);
 498                         put_io_u(td, io_u);
 499                         break;
 500                 }
 501
 502                 if (ret < 0 || td->error)
 503                         break;
 504
 505                 /*
 506                  * See if we need to complete some commands
 507                  */
 508                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 509                         min_evts = 0;
 510                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 511                                 min_evts = 1;
 512
 513                                 if (td->cur_depth > td->o.iodepth_low)
 514                                         min_evts = td->cur_depth - td->o.iodepth_low;
 515                         }
 516
 517                         fio_gettime(&comp_time, NULL);
 518                         bytes_done = io_u_queued_complete(td, min_evts);
 519                         if (bytes_done < 0)
 520                                 break;
 521                 }
 522
 523                 if (!bytes_done)
 524                         continue;
 525
 526                 /*
 527                  * the rate is batched for now, it should work for batches
 528                  * of completions except the very first one which may look
 529                  * a little bursty
 530                  */
 531                 usec = utime_since(&s, &comp_time);
 532
 533                 rate_throttle(td, usec, bytes_done);
 534
 535                 if (check_min_rate(td, &comp_time)) {
 536                         if (exitall_on_terminate)
 537                                 terminate_threads(td->groupid);
 538                         td_verror(td, ENODATA, "check_min_rate");
 539                         break;
 540                 }
 541
 542                 if (td->o.thinktime) {
 543                         unsigned long long b;
 544
 545                         b = td->io_blocks[0] + td->io_blocks[1];
 546                         if (!(b % td->o.thinktime_blocks)) {
 547                                 int left;
 548
 549                                 if (td->o.thinktime_spin)
 550                                         __usec_sleep(td->o.thinktime_spin);
 551
 552                                 left = td->o.thinktime - td->o.thinktime_spin;
 553                                 if (left)
 554                                         usec_sleep(td, left);
 555                         }
 556                 }
 557         }
 558
 559         if (!td->error) {
 560                 struct fio_file *f;
 561
 562                 i = td->cur_depth;
 563                 if (i)
 564                         ret = io_u_queued_complete(td, i);
 565
 566                 if (should_fsync(td) && td->o.end_fsync) {
 567                         td_set_runstate(td, TD_FSYNCING);
 568
 569                         for_each_file(td, f, i) {
 570                                 if (!(f->flags & FIO_FILE_OPEN))
 571                                         continue;
 572                                 fio_io_sync(td, f);
 573                         }
 574                 }
 575         } else
 576                 cleanup_pending_aio(td);
 577 }
 578
 579 static void cleanup_io_u(struct thread_data *td)
 580 {
 581         struct list_head *entry, *n;
 582         struct io_u *io_u;
 583
 584         list_for_each_safe(entry, n, &td->io_u_freelist) {
 585                 io_u = list_entry(entry, struct io_u, list);
 586
 587                 list_del(&io_u->list);
 588                 free(io_u);
 589         }
 590
 591         free_io_mem(td);
 592 }
 593
 594 /*
 595  * "randomly" fill the buffer contents
 596  */
 597 static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
 598 {
 599         long *ptr = io_u->buf;
 600
 601         if (!td->o.zero_buffers) {
 602                 while ((void *) ptr - io_u->buf < max_bs) {
 603                         *ptr = rand() * GOLDEN_RATIO_PRIME;
 604                         ptr++;
 605                 }
 606         } else
 607                 memset(ptr, 0, max_bs);
 608 }
 609
 610 static int init_io_u(struct thread_data *td)
 611 {
 612         unsigned long long buf_size;
 613         struct io_u *io_u;
 614         unsigned int max_bs;
 615         int i, max_units;
 616         char *p;
 617
 618         if (td->io_ops->flags & FIO_SYNCIO)
 619                 max_units = 1;
 620         else
 621                 max_units = td->o.iodepth;
 622
 623         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 624         buf_size = (unsigned long long) max_bs * (unsigned long long) max_units;
 625         buf_size += page_mask;
 626         if (buf_size != (size_t) buf_size) {
 627                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 628                 return 1;
 629         }
 630
 631         td->orig_buffer_size = buf_size;
 632
 633         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
 634                 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
 635         else if (td->orig_buffer_size & page_mask)
 636                 td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
 637
 638         if (allocate_io_mem(td))
 639                 return 1;
 640
 641         p = ALIGN(td->orig_buffer);
 642         for (i = 0; i < max_units; i++) {
 643                 io_u = malloc(sizeof(*io_u));
 644                 memset(io_u, 0, sizeof(*io_u));
 645                 INIT_LIST_HEAD(&io_u->list);
 646
 647                 io_u->buf = p + max_bs * i;
 648
 649                 if (td_write(td))
 650                         fill_io_buf(td, io_u, max_bs);
 651
 652                 io_u->index = i;
 653                 io_u->flags = IO_U_F_FREE;
 654                 list_add(&io_u->list, &td->io_u_freelist);
 655         }
 656
 657         io_u_init_timeout();
 658
 659         return 0;
 660 }
 661
 662 static int switch_ioscheduler(struct thread_data *td)
 663 {
 664         char tmp[256], tmp2[128];
 665         FILE *f;
 666         int ret;
 667
 668         if (td->io_ops->flags & FIO_DISKLESSIO)
 669                 return 0;
 670
 671         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 672
 673         f = fopen(tmp, "r+");
 674         if (!f) {
 675                 if (errno == ENOENT) {
 676                         log_err("fio: os or kernel doesn't support IO scheduler switching\n");
 677                         return 0;
 678                 }
 679                 td_verror(td, errno, "fopen iosched");
 680                 return 1;
 681         }
 682
 683         /*
 684          * Set io scheduler.
 685          */
 686         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 687         if (ferror(f) || ret != 1) {
 688                 td_verror(td, errno, "fwrite");
 689                 fclose(f);
 690                 return 1;
 691         }
 692
 693         rewind(f);
 694
 695         /*
 696          * Read back and check that the selected scheduler is now the default.
 697          */
 698         ret = fread(tmp, 1, sizeof(tmp), f);
 699         if (ferror(f) || ret < 0) {
 700                 td_verror(td, errno, "fread");
 701                 fclose(f);
 702                 return 1;
 703         }
 704
 705         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 706         if (!strstr(tmp, tmp2)) {
 707                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 708                 td_verror(td, EINVAL, "iosched_switch");
 709                 fclose(f);
 710                 return 1;
 711         }
 712
 713         fclose(f);
 714         return 0;
 715 }
 716
 717 static int clear_io_state(struct thread_data *td)
 718 {
 719         struct fio_file *f;
 720         unsigned int i;
 721         int ret;
 722
 723         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 724         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 725         td->zone_bytes = 0;
 726         td->rate_bytes = 0;
 727         td->rate_blocks = 0;
 728         td->rw_end_set[0] = td->rw_end_set[1] = 0;
 729
 730         td->last_was_sync = 0;
 731
 732         for_each_file(td, f, i)
 733                 td_io_close_file(td, f);
 734
 735         ret = 0;
 736         for_each_file(td, f, i) {
 737                 ret = td_io_open_file(td, f);
 738                 if (ret)
 739                         break;
 740         }
 741
 742         return ret;
 743 }
 744
 745 /*
 746  * Entry point for the thread based jobs. The process based jobs end up
 747  * here as well, after a little setup.
 748  */
 749 static void *thread_main(void *data)
 750 {
 751         unsigned long long runtime[2];
 752         struct thread_data *td = data;
 753         unsigned long elapsed;
 754         struct timeval t;
 755         int clear_state;
 756
 757         if (!td->o.use_thread)
 758                 setsid();
 759
 760         td->pid = getpid();
 761
 762         INIT_LIST_HEAD(&td->io_u_freelist);
 763         INIT_LIST_HEAD(&td->io_u_busylist);
 764         INIT_LIST_HEAD(&td->io_u_requeues);
 765         INIT_LIST_HEAD(&td->io_log_list);
 766         INIT_LIST_HEAD(&td->io_hist_list);
 767         td->io_hist_tree = RB_ROOT;
 768
 769         if (init_io_u(td))
 770                 goto err_sem;
 771
 772         if (fio_setaffinity(td) == -1) {
 773                 td_verror(td, errno, "cpu_set_affinity");
 774                 goto err_sem;
 775         }
 776
 777         if (init_iolog(td))
 778                 goto err_sem;
 779
 780         if (td->ioprio) {
 781                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 782                         td_verror(td, errno, "ioprio_set");
 783                         goto err_sem;
 784                 }
 785         }
 786
 787         if (nice(td->o.nice) == -1) {
 788                 td_verror(td, errno, "nice");
 789                 goto err_sem;
 790         }
 791
 792         if (td->o.ioscheduler && switch_ioscheduler(td))
 793                 goto err_sem;
 794
 795         td_set_runstate(td, TD_INITIALIZED);
 796         fio_sem_up(startup_sem);
 797         fio_sem_down(td->mutex);
 798
 799         /*
 800          * the ->mutex semaphore is now no longer used, close it to avoid
 801          * eating a file descriptor
 802          */
 803         fio_sem_remove(td->mutex);
 804
 805         if (!td->o.create_serialize && setup_files(td))
 806                 goto err;
 807
 808         if (td_io_init(td))
 809                 goto err;
 810
 811         if (open_files(td))
 812                 goto err;
 813
 814         if (init_random_map(td))
 815                 goto err;
 816
 817         if (td->o.exec_prerun) {
 818                 if (system(td->o.exec_prerun) < 0)
 819                         goto err;
 820         }
 821
 822         fio_gettime(&td->epoch, NULL);
 823         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 824         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 825
 826         runtime[0] = runtime[1] = 0;
 827         clear_state = 0;
 828         while (td->o.time_based || td->o.loops--) {
 829                 fio_gettime(&td->start, NULL);
 830                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 831
 832                 if (td->o.ratemin)
 833                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 834
 835                 if (clear_state && clear_io_state(td))
 836                         break;
 837
 838                 prune_io_piece_log(td);
 839
 840                 do_io(td);
 841
 842                 clear_state = 1;
 843
 844                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
 845                         if (td->rw_end_set[DDIR_READ])
 846                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
 847                         else
 848                                 elapsed = utime_since_now(&td->start);
 849
 850                         runtime[DDIR_READ] += elapsed;
 851                 }
 852                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
 853                         if (td->rw_end_set[DDIR_WRITE])
 854                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
 855                         else
 856                                 elapsed = utime_since_now(&td->start);
 857
 858                         runtime[DDIR_WRITE] += elapsed;
 859                 }
 860
 861                 if (td->error || td->terminate)
 862                         break;
 863
 864                 fio_gettime(&t, NULL);
 865                 if (runtime_exceeded(td, &t))
 866                         break;
 867
 868                 if (td->o.verify == VERIFY_NONE)
 869                         continue;
 870
 871                 if (clear_io_state(td))
 872                         break;
 873
 874                 fio_gettime(&td->start, NULL);
 875
 876                 do_verify(td);
 877
 878                 runtime[DDIR_READ] += utime_since_now(&td->start);
 879
 880                 if (td->error || td->terminate)
 881                         break;
 882         }
 883
 884         update_rusage_stat(td);
 885         td->ts.runtime[0] = runtime[0] / 1000;
 886         td->ts.runtime[1] = runtime[1] / 1000;
 887         td->ts.total_run_time = mtime_since_now(&td->epoch);
 888         td->ts.io_bytes[0] = td->io_bytes[0];
 889         td->ts.io_bytes[1] = td->io_bytes[1];
 890
 891         if (td->ts.bw_log)
 892                 finish_log(td, td->ts.bw_log, "bw");
 893         if (td->ts.slat_log)
 894                 finish_log(td, td->ts.slat_log, "slat");
 895         if (td->ts.clat_log)
 896                 finish_log(td, td->ts.clat_log, "clat");
 897         if (td->o.write_iolog_file)
 898                 write_iolog_close(td);
 899         if (td->o.exec_postrun) {
 900                 if (system(td->o.exec_postrun) < 0)
 901                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
 902         }
 903
 904         if (exitall_on_terminate)
 905                 terminate_threads(td->groupid);
 906
 907 err:
 908         if (td->error)
 909                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
 910         close_files(td);
 911         close_ioengine(td);
 912         cleanup_io_u(td);
 913         options_mem_free(td);
 914         td_set_runstate(td, TD_EXITED);
 915         return (void *) (unsigned long) td->error;
 916 err_sem:
 917         fio_sem_up(startup_sem);
 918         goto err;
 919 }
 920
 921 /*
 922  * We cannot pass the td data into a forked process, so attach the td and
 923  * pass it to the thread worker.
 924  */
 925 static int fork_main(int shmid, int offset)
 926 {
 927         struct thread_data *td;
 928         void *data, *ret;
 929
 930         data = shmat(shmid, NULL, 0);
 931         if (data == (void *) -1) {
 932                 int __err = errno;
 933
 934                 perror("shmat");
 935                 return __err;
 936         }
 937
 938         td = data + offset * sizeof(struct thread_data);
 939         ret = thread_main(td);
 940         shmdt(data);
 941         return (int) (unsigned long) ret;
 942 }
 943
 944 /*
 945  * Run over the job map and reap the threads that have exited, if any.
 946  */
 947 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 948 {
 949         struct thread_data *td;
 950         int i, cputhreads, pending, status, ret;
 951
 952         /*
 953          * reap exited threads (TD_EXITED -> TD_REAPED)
 954          */
 955         pending = cputhreads = 0;
 956         for_each_td(td, i) {
 957                 int flags = 0;
 958
 959                 /*
 960                  * ->io_ops is NULL for a thread that has closed its
 961                  * io engine
 962                  */
 963                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
 964                         cputhreads++;
 965
 966                 if (!td->pid || td->runstate == TD_REAPED)
 967                         continue;
 968                 if (td->o.use_thread) {
 969                         if (td->runstate == TD_EXITED) {
 970                                 td_set_runstate(td, TD_REAPED);
 971                                 goto reaped;
 972                         }
 973                         continue;
 974                 }
 975
 976                 flags = WNOHANG;
 977                 if (td->runstate == TD_EXITED)
 978                         flags = 0;
 979
 980                 /*
 981                  * check if someone quit or got killed in an unusual way
 982                  */
 983                 ret = waitpid(td->pid, &status, flags);
 984                 if (ret < 0) {
 985                         if (errno == ECHILD) {
 986                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
 987                                 td_set_runstate(td, TD_REAPED);
 988                                 goto reaped;
 989                         }
 990                         perror("waitpid");
 991                 } else if (ret == td->pid) {
 992                         if (WIFSIGNALED(status)) {
 993                                 int sig = WTERMSIG(status);
 994
 995                                 if (sig != SIGQUIT)
 996                                         log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
 997                                 td_set_runstate(td, TD_REAPED);
 998                                 goto reaped;
 999                         }
1000                         if (WIFEXITED(status)) {
1001                                 if (WEXITSTATUS(status) && !td->error)
1002                                         td->error = WEXITSTATUS(status);
1003
1004                                 td_set_runstate(td, TD_REAPED);
1005                                 goto reaped;
1006                         }
1007                 }
1008
1009                 /*
1010                  * thread is not dead, continue
1011                  */
1012                 pending++;
1013                 continue;
1014 reaped:
1015                 if (td->o.use_thread) {
1016                         long ret;
1017
1018                         if (pthread_join(td->thread, (void *) &ret))
1019                                 perror("pthread_join");
1020                 }
1021
1022                 (*nr_running)--;
1023                 (*m_rate) -= td->o.ratemin;
1024                 (*t_rate) -= td->o.rate;
1025                 pending--;
1026
1027                 if (td->error)
1028                         exit_value++;
1029         }
1030
1031         if (*nr_running == cputhreads && !pending)
1032                 terminate_threads(TERMINATE_ALL);
1033 }
1034
1035 /*
1036  * Main function for kicking off and reaping jobs, as needed.
1037  */
1038 static void run_threads(void)
1039 {
1040         struct thread_data *td;
1041         unsigned long spent;
1042         int i, todo, nr_running, m_rate, t_rate, nr_started;
1043
1044         if (fio_pin_memory())
1045                 return;
1046
1047         if (!terse_output) {
1048                 printf("Starting ");
1049                 if (nr_thread)
1050                         printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1051                 if (nr_process) {
1052                         if (nr_thread)
1053                                 printf(" and ");
1054                         printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1055                 }
1056                 printf("\n");
1057                 fflush(stdout);
1058         }
1059
1060         signal(SIGINT, sig_handler);
1061         signal(SIGALRM, sig_handler);
1062
1063         todo = thread_number;
1064         nr_running = 0;
1065         nr_started = 0;
1066         m_rate = t_rate = 0;
1067
1068         for_each_td(td, i) {
1069                 print_status_init(td->thread_number - 1);
1070
1071                 if (!td->o.create_serialize) {
1072                         init_disk_util(td);
1073                         continue;
1074                 }
1075
1076                 /*
1077                  * do file setup here so it happens sequentially,
1078                  * we don't want X number of threads getting their
1079                  * client data interspersed on disk
1080                  */
1081                 if (setup_files(td)) {
1082                         exit_value++;
1083                         if (td->error)
1084                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1085                         td_set_runstate(td, TD_REAPED);
1086                         todo--;
1087                 }
1088
1089                 init_disk_util(td);
1090         }
1091
1092         set_genesis_time();
1093
1094         while (todo) {
1095                 struct thread_data *map[MAX_JOBS];
1096                 struct timeval this_start;
1097                 int this_jobs = 0, left;
1098
1099                 /*
1100                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1101                  */
1102                 for_each_td(td, i) {
1103                         if (td->runstate != TD_NOT_CREATED)
1104                                 continue;
1105
1106                         /*
1107                          * never got a chance to start, killed by other
1108                          * thread for some reason
1109                          */
1110                         if (td->terminate) {
1111                                 todo--;
1112                                 continue;
1113                         }
1114
1115                         if (td->o.start_delay) {
1116                                 spent = mtime_since_genesis();
1117
1118                                 if (td->o.start_delay * 1000 > spent)
1119                                         continue;
1120                         }
1121
1122                         if (td->o.stonewall && (nr_started || nr_running))
1123                                 break;
1124
1125                         /*
1126                          * Set state to created. Thread will transition
1127                          * to TD_INITIALIZED when it's done setting up.
1128                          */
1129                         td_set_runstate(td, TD_CREATED);
1130                         map[this_jobs++] = td;
1131                         nr_started++;
1132
1133                         if (td->o.use_thread) {
1134                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1135                                         perror("thread_create");
1136                                         nr_started--;
1137                                         break;
1138                                 }
1139                         } else {
1140                                 if (!fork()) {
1141                                         int ret = fork_main(shm_id, i);
1142
1143                                         exit(ret);
1144                                 }
1145                         }
1146                         fio_sem_down(startup_sem);
1147                 }
1148
1149                 /*
1150                  * Wait for the started threads to transition to
1151                  * TD_INITIALIZED.
1152                  */
1153                 fio_gettime(&this_start, NULL);
1154                 left = this_jobs;
1155                 while (left && !fio_abort) {
1156                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1157                                 break;
1158
1159                         usleep(100000);
1160
1161                         for (i = 0; i < this_jobs; i++) {
1162                                 td = map[i];
1163                                 if (!td)
1164                                         continue;
1165                                 if (td->runstate == TD_INITIALIZED) {
1166                                         map[i] = NULL;
1167                                         left--;
1168                                 } else if (td->runstate >= TD_EXITED) {
1169                                         map[i] = NULL;
1170                                         left--;
1171                                         todo--;
1172                                         nr_running++; /* work-around... */
1173                                 }
1174                         }
1175                 }
1176
1177                 if (left) {
1178                         log_err("fio: %d jobs failed to start\n", left);
1179                         for (i = 0; i < this_jobs; i++) {
1180                                 td = map[i];
1181                                 if (!td)
1182                                         continue;
1183                                 kill(td->pid, SIGTERM);
1184                         }
1185                         break;
1186                 }
1187
1188                 /*
1189                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1190                  */
1191                 for_each_td(td, i) {
1192                         if (td->runstate != TD_INITIALIZED)
1193                                 continue;
1194
1195                         td_set_runstate(td, TD_RUNNING);
1196                         nr_running++;
1197                         nr_started--;
1198                         m_rate += td->o.ratemin;
1199                         t_rate += td->o.rate;
1200                         todo--;
1201                         fio_sem_up(td->mutex);
1202                 }
1203
1204                 reap_threads(&nr_running, &t_rate, &m_rate);
1205
1206                 if (todo)
1207                         usleep(100000);
1208         }
1209
1210         while (nr_running) {
1211                 reap_threads(&nr_running, &t_rate, &m_rate);
1212                 usleep(10000);
1213         }
1214
1215         update_io_ticks();
1216         fio_unpin_memory();
1217 }
1218
1219 int main(int argc, char *argv[])
1220 {
1221         long ps;
1222
1223         /*
1224          * We need locale for number printing, if it isn't set then just
1225          * go with the US format.
1226          */
1227         if (!getenv("LC_NUMERIC"))
1228                 setlocale(LC_NUMERIC, "en_US");
1229
1230         if (parse_options(argc, argv))
1231                 return 1;
1232
1233         if (!thread_number)
1234                 return 0;
1235
1236         ps = sysconf(_SC_PAGESIZE);
1237         if (ps < 0) {
1238                 log_err("Failed to get page size\n");
1239                 return 1;
1240         }
1241
1242         page_size = ps;
1243         page_mask = ps - 1;
1244
1245         if (write_bw_log) {
1246                 setup_log(&agg_io_log[DDIR_READ]);
1247                 setup_log(&agg_io_log[DDIR_WRITE]);
1248         }
1249
1250         startup_sem = fio_sem_init(0);
1251
1252         set_genesis_time();
1253
1254         disk_util_timer_arm();
1255
1256         run_threads();
1257
1258         if (!fio_abort) {
1259                 show_run_stats();
1260                 if (write_bw_log) {
1261                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1262                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1263                 }
1264         }
1265
1266         fio_sem_remove(startup_sem);
1267         return exit_value;
1268 }