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 #include "smalloc.h"
  40
  41 unsigned long page_mask;
  42 unsigned long page_size;
  43 #define ALIGN(buf)      \
  44         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  45
  46 int groupid = 0;
  47 int thread_number = 0;
  48 int nr_process = 0;
  49 int nr_thread = 0;
  50 int shm_id = 0;
  51 int temp_stall_ts;
  52
  53 static struct fio_mutex *startup_mutex;
  54 static volatile int fio_abort;
  55 static int exit_value;
  56
  57 struct io_log *agg_io_log[2];
  58
  59 #define TERMINATE_ALL           (-1)
  60 #define JOB_START_TIMEOUT       (5 * 1000)
  61
  62 static inline void td_set_runstate(struct thread_data *td, int runstate)
  63 {
  64         if (td->runstate == runstate)
  65                 return;
  66
  67         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", td->pid, td->runstate,
  68                                                                 runstate);
  69         td->runstate = runstate;
  70 }
  71
  72 static void terminate_threads(int group_id)
  73 {
  74         struct thread_data *td;
  75         int i;
  76
  77         for_each_td(td, i) {
  78                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  79                         dprint(FD_PROCESS, "setting terminate on %d\n",
  80                                                                 td->pid);
  81                         td->terminate = 1;
  82                         td->o.start_delay = 0;
  83
  84                         /*
  85                          * if the thread is running, just let it exit
  86                          */
  87                         if (td->runstate < TD_RUNNING)
  88                                 kill(td->pid, SIGQUIT);
  89                         else {
  90                                 struct ioengine_ops *ops = td->io_ops;
  91
  92                                 if (ops && (ops->flags & FIO_SIGQUIT))
  93                                         kill(td->pid, SIGQUIT);
  94                         }
  95                 }
  96         }
  97 }
  98
  99 static void sig_handler(int sig)
 100 {
 101         switch (sig) {
 102         case SIGALRM:
 103                 update_io_ticks();
 104                 disk_util_timer_arm();
 105                 print_thread_status();
 106                 break;
 107         default:
 108                 printf("\nfio: terminating on signal %d\n", sig);
 109                 fflush(stdout);
 110                 terminate_threads(TERMINATE_ALL);
 111                 break;
 112         }
 113 }
 114
 115 /*
 116  * Check if we are above the minimum rate given.
 117  */
 118 static int check_min_rate(struct thread_data *td, struct timeval *now)
 119 {
 120         unsigned long long bytes = 0;
 121         unsigned long iops = 0;
 122         unsigned long spent;
 123         unsigned long rate;
 124
 125         /*
 126          * No minimum rate set, always ok
 127          */
 128         if (!td->o.ratemin && !td->o.rate_iops_min)
 129                 return 0;
 130
 131         /*
 132          * allow a 2 second settle period in the beginning
 133          */
 134         if (mtime_since(&td->start, now) < 2000)
 135                 return 0;
 136
 137         if (td_read(td)) {
 138                 iops += td->io_blocks[DDIR_READ];
 139                 bytes += td->this_io_bytes[DDIR_READ];
 140         }
 141         if (td_write(td)) {
 142                 iops += td->io_blocks[DDIR_WRITE];
 143                 bytes += td->this_io_bytes[DDIR_WRITE];
 144         }
 145
 146         /*
 147          * if rate blocks is set, sample is running
 148          */
 149         if (td->rate_bytes || td->rate_blocks) {
 150                 spent = mtime_since(&td->lastrate, now);
 151                 if (spent < td->o.ratecycle)
 152                         return 0;
 153
 154                 if (td->o.rate) {
 155                         /*
 156                          * check bandwidth specified rate
 157                          */
 158                         if (bytes < td->rate_bytes) {
 159                                 log_err("%s: min rate %u not met\n", td->o.name,
 160                                                                 td->o.ratemin);
 161                                 return 1;
 162                         } else {
 163                                 rate = (bytes - td->rate_bytes) / spent;
 164                                 if (rate < td->o.ratemin ||
 165                                     bytes < td->rate_bytes) {
 166                                         log_err("%s: min rate %u not met, got"
 167                                                 " %luKiB/sec\n", td->o.name,
 168                                                         td->o.ratemin, rate);
 169                                         return 1;
 170                                 }
 171                         }
 172                 } else {
 173                         /*
 174                          * checks iops specified rate
 175                          */
 176                         if (iops < td->o.rate_iops) {
 177                                 log_err("%s: min iops rate %u not met\n",
 178                                                 td->o.name, td->o.rate_iops);
 179                                 return 1;
 180                         } else {
 181                                 rate = (iops - td->rate_blocks) / spent;
 182                                 if (rate < td->o.rate_iops_min ||
 183                                     iops < td->rate_blocks) {
 184                                         log_err("%s: min iops rate %u not met,"
 185                                                 " got %lu\n", td->o.name,
 186                                                         td->o.rate_iops_min,
 187                                                         rate);
 188                                 }
 189                         }
 190                 }
 191         }
 192
 193         td->rate_bytes = bytes;
 194         td->rate_blocks = iops;
 195         memcpy(&td->lastrate, now, sizeof(*now));
 196         return 0;
 197 }
 198
 199 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 200 {
 201         if (!td->o.timeout)
 202                 return 0;
 203         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 204                 return 1;
 205
 206         return 0;
 207 }
 208
 209 /*
 210  * When job exits, we can cancel the in-flight IO if we are using async
 211  * io. Attempt to do so.
 212  */
 213 static void cleanup_pending_aio(struct thread_data *td)
 214 {
 215         struct list_head *entry, *n;
 216         struct io_u *io_u;
 217         int r;
 218
 219         /*
 220          * get immediately available events, if any
 221          */
 222         r = io_u_queued_complete(td, 0);
 223         if (r < 0)
 224                 return;
 225
 226         /*
 227          * now cancel remaining active events
 228          */
 229         if (td->io_ops->cancel) {
 230                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 231                         io_u = list_entry(entry, struct io_u, list);
 232
 233                         /*
 234                          * if the io_u isn't in flight, then that generally
 235                          * means someone leaked an io_u. complain but fix
 236                          * it up, so we don't stall here.
 237                          */
 238                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 239                                 log_err("fio: non-busy IO on busy list\n");
 240                                 put_io_u(td, io_u);
 241                         } else {
 242                                 r = td->io_ops->cancel(td, io_u);
 243                                 if (!r)
 244                                         put_io_u(td, io_u);
 245                         }
 246                 }
 247         }
 248
 249         if (td->cur_depth)
 250                 r = io_u_queued_complete(td, td->cur_depth);
 251 }
 252
 253 /*
 254  * Helper to handle the final sync of a file. Works just like the normal
 255  * io path, just does everything sync.
 256  */
 257 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 258 {
 259         struct io_u *io_u = __get_io_u(td);
 260         int ret;
 261
 262         if (!io_u)
 263                 return 1;
 264
 265         io_u->ddir = DDIR_SYNC;
 266         io_u->file = f;
 267
 268         if (td_io_prep(td, io_u)) {
 269                 put_io_u(td, io_u);
 270                 return 1;
 271         }
 272
 273 requeue:
 274         ret = td_io_queue(td, io_u);
 275         if (ret < 0) {
 276                 td_verror(td, io_u->error, "td_io_queue");
 277                 put_io_u(td, io_u);
 278                 return 1;
 279         } else if (ret == FIO_Q_QUEUED) {
 280                 if (io_u_queued_complete(td, 1) < 0)
 281                         return 1;
 282         } else if (ret == FIO_Q_COMPLETED) {
 283                 if (io_u->error) {
 284                         td_verror(td, io_u->error, "td_io_queue");
 285                         return 1;
 286                 }
 287
 288                 if (io_u_sync_complete(td, io_u) < 0)
 289                         return 1;
 290         } else if (ret == FIO_Q_BUSY) {
 291                 if (td_io_commit(td))
 292                         return 1;
 293                 goto requeue;
 294         }
 295
 296         return 0;
 297 }
 298
 299 /*
 300  * The main verify engine. Runs over the writes we previously submitted,
 301  * reads the blocks back in, and checks the crc/md5 of the data.
 302  */
 303 static void do_verify(struct thread_data *td)
 304 {
 305         struct fio_file *f;
 306         struct io_u *io_u;
 307         int ret, min_events;
 308         unsigned int i;
 309
 310         /*
 311          * sync io first and invalidate cache, to make sure we really
 312          * read from disk.
 313          */
 314         for_each_file(td, f, i) {
 315                 if (!(f->flags & FIO_FILE_OPEN))
 316                         continue;
 317                 if (fio_io_sync(td, f))
 318                         break;
 319                 if (file_invalidate_cache(td, f))
 320                         break;
 321         }
 322
 323         if (td->error)
 324                 return;
 325
 326         td_set_runstate(td, TD_VERIFYING);
 327
 328         io_u = NULL;
 329         while (!td->terminate) {
 330                 int ret2;
 331
 332                 io_u = __get_io_u(td);
 333                 if (!io_u)
 334                         break;
 335
 336                 if (runtime_exceeded(td, &io_u->start_time)) {
 337                         put_io_u(td, io_u);
 338                         td->terminate = 1;
 339                         break;
 340                 }
 341
 342                 if (get_next_verify(td, io_u)) {
 343                         put_io_u(td, io_u);
 344                         break;
 345                 }
 346
 347                 if (td_io_prep(td, io_u)) {
 348                         put_io_u(td, io_u);
 349                         break;
 350                 }
 351
 352                 io_u->end_io = verify_io_u;
 353
 354                 ret = td_io_queue(td, io_u);
 355                 switch (ret) {
 356                 case FIO_Q_COMPLETED:
 357                         if (io_u->error)
 358                                 ret = -io_u->error;
 359                         else if (io_u->resid) {
 360                                 int bytes = io_u->xfer_buflen - io_u->resid;
 361                                 struct fio_file *f = io_u->file;
 362
 363                                 /*
 364                                  * zero read, fail
 365                                  */
 366                                 if (!bytes) {
 367                                         td_verror(td, ENODATA, "full resid");
 368                                         put_io_u(td, io_u);
 369                                         break;
 370                                 }
 371
 372                                 io_u->xfer_buflen = io_u->resid;
 373                                 io_u->xfer_buf += bytes;
 374                                 io_u->offset += bytes;
 375
 376                                 td->ts.short_io_u[io_u->ddir]++;
 377
 378                                 if (io_u->offset == f->real_file_size)
 379                                         goto sync_done;
 380
 381                                 requeue_io_u(td, &io_u);
 382                         } else {
 383 sync_done:
 384                                 ret = io_u_sync_complete(td, io_u);
 385                                 if (ret < 0)
 386                                         break;
 387                         }
 388                         continue;
 389                 case FIO_Q_QUEUED:
 390                         break;
 391                 case FIO_Q_BUSY:
 392                         requeue_io_u(td, &io_u);
 393                         ret2 = td_io_commit(td);
 394                         if (ret2 < 0)
 395                                 ret = ret2;
 396                         break;
 397                 default:
 398                         assert(ret < 0);
 399                         td_verror(td, -ret, "td_io_queue");
 400                         break;
 401                 }
 402
 403                 if (ret < 0 || td->error)
 404                         break;
 405
 406                 /*
 407                  * if we can queue more, do so. but check if there are
 408                  * completed io_u's first.
 409                  */
 410                 min_events = 0;
 411                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 412                         min_events = 1;
 413
 414                         if (td->cur_depth > td->o.iodepth_low)
 415                                 min_events = td->cur_depth - td->o.iodepth_low;
 416                 }
 417
 418                 /*
 419                  * Reap required number of io units, if any, and do the
 420                  * verification on them through the callback handler
 421                  */
 422                 if (io_u_queued_complete(td, min_events) < 0)
 423                         break;
 424         }
 425
 426         if (!td->error) {
 427                 min_events = td->cur_depth;
 428
 429                 if (min_events)
 430                         ret = io_u_queued_complete(td, min_events);
 431         } else
 432                 cleanup_pending_aio(td);
 433
 434         td_set_runstate(td, TD_RUNNING);
 435 }
 436
 437 /*
 438  * Main IO worker function. It retrieves io_u's to process and queues
 439  * and reaps them, checking for rate and errors along the way.
 440  */
 441 static void do_io(struct thread_data *td)
 442 {
 443         struct timeval s;
 444         unsigned long usec;
 445         unsigned int i;
 446         int ret = 0;
 447
 448         td_set_runstate(td, TD_RUNNING);
 449
 450         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 451                 struct timeval comp_time;
 452                 long bytes_done = 0;
 453                 int min_evts = 0;
 454                 struct io_u *io_u;
 455                 int ret2;
 456
 457                 if (td->terminate)
 458                         break;
 459
 460                 io_u = get_io_u(td);
 461                 if (!io_u)
 462                         break;
 463
 464                 memcpy(&s, &io_u->start_time, sizeof(s));
 465
 466                 if (runtime_exceeded(td, &s)) {
 467                         put_io_u(td, io_u);
 468                         td->terminate = 1;
 469                         break;
 470                 }
 471
 472                 /*
 473                  * Add verification end_io handler, if asked to verify
 474                  * a previously written file.
 475                  */
 476                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
 477                         io_u->end_io = verify_io_u;
 478                         td_set_runstate(td, TD_VERIFYING);
 479                 } else
 480                         td_set_runstate(td, TD_RUNNING);
 481
 482                 ret = td_io_queue(td, io_u);
 483                 switch (ret) {
 484                 case FIO_Q_COMPLETED:
 485                         if (io_u->error)
 486                                 ret = -io_u->error;
 487                         else if (io_u->resid) {
 488                                 int bytes = io_u->xfer_buflen - io_u->resid;
 489                                 struct fio_file *f = io_u->file;
 490
 491                                 /*
 492                                  * zero read, fail
 493                                  */
 494                                 if (!bytes) {
 495                                         td_verror(td, ENODATA, "full resid");
 496                                         put_io_u(td, io_u);
 497                                         break;
 498                                 }
 499
 500                                 io_u->xfer_buflen = io_u->resid;
 501                                 io_u->xfer_buf += bytes;
 502                                 io_u->offset += bytes;
 503
 504                                 td->ts.short_io_u[io_u->ddir]++;
 505
 506                                 if (io_u->offset == f->real_file_size)
 507                                         goto sync_done;
 508
 509                                 requeue_io_u(td, &io_u);
 510                         } else {
 511 sync_done:
 512                                 fio_gettime(&comp_time, NULL);
 513                                 bytes_done = io_u_sync_complete(td, io_u);
 514                                 if (bytes_done < 0)
 515                                         ret = bytes_done;
 516                         }
 517                         break;
 518                 case FIO_Q_QUEUED:
 519                         /*
 520                          * if the engine doesn't have a commit hook,
 521                          * the io_u is really queued. if it does have such
 522                          * a hook, it has to call io_u_queued() itself.
 523                          */
 524                         if (td->io_ops->commit == NULL)
 525                                 io_u_queued(td, io_u);
 526                         break;
 527                 case FIO_Q_BUSY:
 528                         requeue_io_u(td, &io_u);
 529                         ret2 = td_io_commit(td);
 530                         if (ret2 < 0)
 531                                 ret = ret2;
 532                         break;
 533                 default:
 534                         assert(ret < 0);
 535                         put_io_u(td, io_u);
 536                         break;
 537                 }
 538
 539                 if (ret < 0 || td->error)
 540                         break;
 541
 542                 /*
 543                  * See if we need to complete some commands
 544                  */
 545                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 546                         min_evts = 0;
 547                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 548                                 min_evts = 1;
 549
 550                                 if (td->cur_depth > td->o.iodepth_low)
 551                                         min_evts = td->cur_depth
 552                                                         - td->o.iodepth_low;
 553                         }
 554
 555                         fio_gettime(&comp_time, NULL);
 556                         bytes_done = io_u_queued_complete(td, min_evts);
 557                         if (bytes_done < 0)
 558                                 break;
 559                 }
 560
 561                 if (!bytes_done)
 562                         continue;
 563
 564                 /*
 565                  * the rate is batched for now, it should work for batches
 566                  * of completions except the very first one which may look
 567                  * a little bursty
 568                  */
 569                 usec = utime_since(&s, &comp_time);
 570
 571                 rate_throttle(td, usec, bytes_done);
 572
 573                 if (check_min_rate(td, &comp_time)) {
 574                         if (exitall_on_terminate)
 575                                 terminate_threads(td->groupid);
 576                         td_verror(td, ENODATA, "check_min_rate");
 577                         break;
 578                 }
 579
 580                 if (td->o.thinktime) {
 581                         unsigned long long b;
 582
 583                         b = td->io_blocks[0] + td->io_blocks[1];
 584                         if (!(b % td->o.thinktime_blocks)) {
 585                                 int left;
 586
 587                                 if (td->o.thinktime_spin)
 588                                         __usec_sleep(td->o.thinktime_spin);
 589
 590                                 left = td->o.thinktime - td->o.thinktime_spin;
 591                                 if (left)
 592                                         usec_sleep(td, left);
 593                         }
 594                 }
 595         }
 596
 597         if (td->o.fill_device && td->error == ENOSPC) {
 598                 td->error = 0;
 599                 td->terminate = 1;
 600         }
 601         if (!td->error) {
 602                 struct fio_file *f;
 603
 604                 i = td->cur_depth;
 605                 if (i)
 606                         ret = io_u_queued_complete(td, i);
 607
 608                 if (should_fsync(td) && td->o.end_fsync) {
 609                         td_set_runstate(td, TD_FSYNCING);
 610
 611                         for_each_file(td, f, i) {
 612                                 if (!(f->flags & FIO_FILE_OPEN))
 613                                         continue;
 614                                 fio_io_sync(td, f);
 615                         }
 616                 }
 617         } else
 618                 cleanup_pending_aio(td);
 619
 620         /*
 621          * stop job if we failed doing any IO
 622          */
 623         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 624                 td->done = 1;
 625 }
 626
 627 static void cleanup_io_u(struct thread_data *td)
 628 {
 629         struct list_head *entry, *n;
 630         struct io_u *io_u;
 631
 632         list_for_each_safe(entry, n, &td->io_u_freelist) {
 633                 io_u = list_entry(entry, struct io_u, list);
 634
 635                 list_del(&io_u->list);
 636                 free(io_u);
 637         }
 638
 639         free_io_mem(td);
 640 }
 641
 642 /*
 643  * "randomly" fill the buffer contents
 644  */
 645 static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
 646 {
 647         long *ptr = io_u->buf;
 648
 649         if (!td->o.zero_buffers) {
 650                 while ((void *) ptr - io_u->buf < max_bs) {
 651                         *ptr = rand() * GOLDEN_RATIO_PRIME;
 652                         ptr++;
 653                 }
 654         } else
 655                 memset(ptr, 0, max_bs);
 656 }
 657
 658 static int init_io_u(struct thread_data *td)
 659 {
 660         struct io_u *io_u;
 661         unsigned int max_bs;
 662         int i, max_units;
 663         char *p;
 664
 665         max_units = td->o.iodepth;
 666         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 667         td->orig_buffer_size = (unsigned long long) max_bs
 668                                         * (unsigned long long) max_units;
 669
 670         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 671                 unsigned long bs;
 672
 673                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 674                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 675         }
 676
 677         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 678                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 679                 return 1;
 680         }
 681
 682         if (allocate_io_mem(td))
 683                 return 1;
 684
 685         if (td->o.odirect)
 686                 p = ALIGN(td->orig_buffer);
 687         else
 688                 p = td->orig_buffer;
 689
 690         for (i = 0; i < max_units; i++) {
 691                 if (td->terminate)
 692                         return 1;
 693                 io_u = malloc(sizeof(*io_u));
 694                 memset(io_u, 0, sizeof(*io_u));
 695                 INIT_LIST_HEAD(&io_u->list);
 696
 697                 if (!(td->io_ops->flags & FIO_NOIO)) {
 698                         io_u->buf = p + max_bs * i;
 699
 700                         if (td_write(td))
 701                                 fill_io_buf(td, io_u, max_bs);
 702                 }
 703
 704                 io_u->index = i;
 705                 io_u->flags = IO_U_F_FREE;
 706                 list_add(&io_u->list, &td->io_u_freelist);
 707         }
 708
 709         io_u_init_timeout();
 710
 711         return 0;
 712 }
 713
 714 static int switch_ioscheduler(struct thread_data *td)
 715 {
 716         char tmp[256], tmp2[128];
 717         FILE *f;
 718         int ret;
 719
 720         if (td->io_ops->flags & FIO_DISKLESSIO)
 721                 return 0;
 722
 723         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 724
 725         f = fopen(tmp, "r+");
 726         if (!f) {
 727                 if (errno == ENOENT) {
 728                         log_err("fio: os or kernel doesn't support IO scheduler"
 729                                 " switching\n");
 730                         return 0;
 731                 }
 732                 td_verror(td, errno, "fopen iosched");
 733                 return 1;
 734         }
 735
 736         /*
 737          * Set io scheduler.
 738          */
 739         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 740         if (ferror(f) || ret != 1) {
 741                 td_verror(td, errno, "fwrite");
 742                 fclose(f);
 743                 return 1;
 744         }
 745
 746         rewind(f);
 747
 748         /*
 749          * Read back and check that the selected scheduler is now the default.
 750          */
 751         ret = fread(tmp, 1, sizeof(tmp), f);
 752         if (ferror(f) || ret < 0) {
 753                 td_verror(td, errno, "fread");
 754                 fclose(f);
 755                 return 1;
 756         }
 757
 758         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 759         if (!strstr(tmp, tmp2)) {
 760                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 761                 td_verror(td, EINVAL, "iosched_switch");
 762                 fclose(f);
 763                 return 1;
 764         }
 765
 766         fclose(f);
 767         return 0;
 768 }
 769
 770 static int keep_running(struct thread_data *td)
 771 {
 772         unsigned long long io_done;
 773
 774         if (td->done)
 775                 return 0;
 776         if (td->o.time_based)
 777                 return 1;
 778         if (td->o.loops) {
 779                 td->o.loops--;
 780                 return 1;
 781         }
 782
 783         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 784                         + td->io_skip_bytes;
 785         if (io_done < td->o.size)
 786                 return 1;
 787
 788         return 0;
 789 }
 790
 791 static int clear_io_state(struct thread_data *td)
 792 {
 793         struct fio_file *f;
 794         unsigned int i;
 795         int ret;
 796
 797         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 798         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 799         td->zone_bytes = 0;
 800         td->rate_bytes = 0;
 801         td->rate_blocks = 0;
 802         td->rw_end_set[0] = td->rw_end_set[1] = 0;
 803
 804         td->last_was_sync = 0;
 805
 806         /*
 807          * reset file done count if we are to start over
 808          */
 809         if (td->o.time_based || td->o.loops)
 810                 td->nr_done_files = 0;
 811
 812         close_files(td);
 813
 814         ret = 0;
 815         for_each_file(td, f, i) {
 816                 f->flags &= ~FIO_FILE_DONE;
 817                 ret = td_io_open_file(td, f);
 818                 if (ret)
 819                         break;
 820         }
 821
 822         return ret;
 823 }
 824
 825 /*
 826  * Entry point for the thread based jobs. The process based jobs end up
 827  * here as well, after a little setup.
 828  */
 829 static void *thread_main(void *data)
 830 {
 831         unsigned long long runtime[2], elapsed;
 832         struct thread_data *td = data;
 833         int clear_state;
 834
 835         if (!td->o.use_thread)
 836                 setsid();
 837
 838         td->pid = getpid();
 839
 840         dprint(FD_PROCESS, "jobs pid=%d started\n", td->pid);
 841
 842         INIT_LIST_HEAD(&td->io_u_freelist);
 843         INIT_LIST_HEAD(&td->io_u_busylist);
 844         INIT_LIST_HEAD(&td->io_u_requeues);
 845         INIT_LIST_HEAD(&td->io_log_list);
 846         INIT_LIST_HEAD(&td->io_hist_list);
 847         td->io_hist_tree = RB_ROOT;
 848
 849         td_set_runstate(td, TD_INITIALIZED);
 850         fio_mutex_up(startup_mutex);
 851         fio_mutex_down(td->mutex);
 852
 853         /*
 854          * the ->mutex mutex is now no longer used, close it to avoid
 855          * eating a file descriptor
 856          */
 857         fio_mutex_remove(td->mutex);
 858
 859         /*
 860          * May alter parameters that init_io_u() will use, so we need to
 861          * do this first.
 862          */
 863         if (init_iolog(td))
 864                 goto err;
 865
 866         if (init_io_u(td))
 867                 goto err;
 868
 869         if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
 870                 td_verror(td, errno, "cpu_set_affinity");
 871                 goto err;
 872         }
 873
 874         if (td->ioprio_set) {
 875                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 876                         td_verror(td, errno, "ioprio_set");
 877                         goto err;
 878                 }
 879         }
 880
 881         if (nice(td->o.nice) == -1) {
 882                 td_verror(td, errno, "nice");
 883                 goto err;
 884         }
 885
 886         if (td->o.ioscheduler && switch_ioscheduler(td))
 887                 goto err;
 888
 889         if (!td->o.create_serialize && setup_files(td))
 890                 goto err;
 891
 892         if (td_io_init(td))
 893                 goto err;
 894
 895         if (open_files(td))
 896                 goto err;
 897
 898         if (init_random_map(td))
 899                 goto err;
 900
 901         if (td->o.exec_prerun) {
 902                 if (system(td->o.exec_prerun) < 0)
 903                         goto err;
 904         }
 905
 906         fio_gettime(&td->epoch, NULL);
 907         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 908         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 909
 910         runtime[0] = runtime[1] = 0;
 911         clear_state = 0;
 912         while (keep_running(td)) {
 913                 fio_gettime(&td->start, NULL);
 914                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 915
 916                 if (td->o.ratemin)
 917                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
 918                                                         sizeof(td->lastrate));
 919
 920                 if (clear_state && clear_io_state(td))
 921                         break;
 922
 923                 prune_io_piece_log(td);
 924
 925                 do_io(td);
 926
 927                 clear_state = 1;
 928
 929                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
 930                         if (td->rw_end_set[DDIR_READ])
 931                                 elapsed = utime_since(&td->start,
 932                                                       &td->rw_end[DDIR_READ]);
 933                         else
 934                                 elapsed = utime_since_now(&td->start);
 935
 936                         runtime[DDIR_READ] += elapsed;
 937                 }
 938                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
 939                         if (td->rw_end_set[DDIR_WRITE])
 940                                 elapsed = utime_since(&td->start,
 941                                                       &td->rw_end[DDIR_WRITE]);
 942                         else
 943                                 elapsed = utime_since_now(&td->start);
 944
 945                         runtime[DDIR_WRITE] += elapsed;
 946                 }
 947
 948                 if (td->error || td->terminate)
 949                         break;
 950
 951                 if (!td->o.do_verify ||
 952                     td->o.verify == VERIFY_NONE ||
 953                     (td->io_ops->flags & FIO_UNIDIR))
 954                         continue;
 955
 956                 if (clear_io_state(td))
 957                         break;
 958
 959                 fio_gettime(&td->start, NULL);
 960
 961                 do_verify(td);
 962
 963                 runtime[DDIR_READ] += utime_since_now(&td->start);
 964
 965                 if (td->error || td->terminate)
 966                         break;
 967         }
 968
 969         update_rusage_stat(td);
 970         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
 971         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
 972         td->ts.total_run_time = mtime_since_now(&td->epoch);
 973         td->ts.io_bytes[0] = td->io_bytes[0];
 974         td->ts.io_bytes[1] = td->io_bytes[1];
 975
 976         if (td->ts.bw_log)
 977                 finish_log(td, td->ts.bw_log, "bw");
 978         if (td->ts.slat_log)
 979                 finish_log(td, td->ts.slat_log, "slat");
 980         if (td->ts.clat_log)
 981                 finish_log(td, td->ts.clat_log, "clat");
 982         if (td->o.exec_postrun) {
 983                 if (system(td->o.exec_postrun) < 0)
 984                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
 985         }
 986
 987         if (exitall_on_terminate)
 988                 terminate_threads(td->groupid);
 989
 990 err:
 991         if (td->error)
 992                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error,
 993                                                         td->verror);
 994         close_and_free_files(td);
 995         close_ioengine(td);
 996         cleanup_io_u(td);
 997
 998         /*
 999          * do this very late, it will log file closing as well
1000          */
1001         if (td->o.write_iolog_file)
1002                 write_iolog_close(td);
1003
1004         options_mem_free(td);
1005         td_set_runstate(td, TD_EXITED);
1006         return (void *) (unsigned long) td->error;
1007 }
1008
1009 /*
1010  * We cannot pass the td data into a forked process, so attach the td and
1011  * pass it to the thread worker.
1012  */
1013 static int fork_main(int shmid, int offset)
1014 {
1015         struct thread_data *td;
1016         void *data, *ret;
1017
1018         data = shmat(shmid, NULL, 0);
1019         if (data == (void *) -1) {
1020                 int __err = errno;
1021
1022                 perror("shmat");
1023                 return __err;
1024         }
1025
1026         td = data + offset * sizeof(struct thread_data);
1027         ret = thread_main(td);
1028         shmdt(data);
1029         return (int) (unsigned long) ret;
1030 }
1031
1032 /*
1033  * Run over the job map and reap the threads that have exited, if any.
1034  */
1035 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1036 {
1037         struct thread_data *td;
1038         int i, cputhreads, realthreads, pending, status, ret;
1039
1040         /*
1041          * reap exited threads (TD_EXITED -> TD_REAPED)
1042          */
1043         realthreads = pending = cputhreads = 0;
1044         for_each_td(td, i) {
1045                 int flags = 0;
1046
1047                 /*
1048                  * ->io_ops is NULL for a thread that has closed its
1049                  * io engine
1050                  */
1051                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1052                         cputhreads++;
1053                 else
1054                         realthreads++;
1055
1056                 if (!td->pid) {
1057                         pending++;
1058                         continue;
1059                 }
1060                 if (td->runstate == TD_REAPED)
1061                         continue;
1062                 if (td->o.use_thread) {
1063                         if (td->runstate == TD_EXITED) {
1064                                 td_set_runstate(td, TD_REAPED);
1065                                 goto reaped;
1066                         }
1067                         continue;
1068                 }
1069
1070                 flags = WNOHANG;
1071                 if (td->runstate == TD_EXITED)
1072                         flags = 0;
1073
1074                 /*
1075                  * check if someone quit or got killed in an unusual way
1076                  */
1077                 ret = waitpid(td->pid, &status, flags);
1078                 if (ret < 0) {
1079                         if (errno == ECHILD) {
1080                                 log_err("fio: pid=%d disappeared %d\n", td->pid,
1081                                                                 td->runstate);
1082                                 td_set_runstate(td, TD_REAPED);
1083                                 goto reaped;
1084                         }
1085                         perror("waitpid");
1086                 } else if (ret == td->pid) {
1087                         if (WIFSIGNALED(status)) {
1088                                 int sig = WTERMSIG(status);
1089
1090                                 if (sig != SIGQUIT)
1091                                         log_err("fio: pid=%d, got signal=%d\n",
1092                                                                 td->pid, sig);
1093                                 td_set_runstate(td, TD_REAPED);
1094                                 goto reaped;
1095                         }
1096                         if (WIFEXITED(status)) {
1097                                 if (WEXITSTATUS(status) && !td->error)
1098                                         td->error = WEXITSTATUS(status);
1099
1100                                 td_set_runstate(td, TD_REAPED);
1101                                 goto reaped;
1102                         }
1103                 }
1104
1105                 /*
1106                  * thread is not dead, continue
1107                  */
1108                 pending++;
1109                 continue;
1110 reaped:
1111                 if (td->o.use_thread) {
1112                         long ret;
1113
1114                         dprint(FD_PROCESS, "joining tread %d\n", td->pid);
1115                         if (pthread_join(td->thread, (void *) &ret)) {
1116                                 dprint(FD_PROCESS, "join failed %ld\n", ret);
1117                                 perror("pthread_join");
1118                         }
1119                 }
1120
1121                 (*nr_running)--;
1122                 (*m_rate) -= td->o.ratemin;
1123                 (*t_rate) -= td->o.rate;
1124                 pending--;
1125
1126                 if (td->error)
1127                         exit_value++;
1128         }
1129
1130         if (*nr_running == cputhreads && !pending && realthreads)
1131                 terminate_threads(TERMINATE_ALL);
1132 }
1133
1134 /*
1135  * Main function for kicking off and reaping jobs, as needed.
1136  */
1137 static void run_threads(void)
1138 {
1139         struct thread_data *td;
1140         unsigned long spent;
1141         int i, todo, nr_running, m_rate, t_rate, nr_started;
1142
1143         if (fio_pin_memory())
1144                 return;
1145
1146         if (!terse_output) {
1147                 printf("Starting ");
1148                 if (nr_thread)
1149                         printf("%d thread%s", nr_thread,
1150                                                 nr_thread > 1 ? "s" : "");
1151                 if (nr_process) {
1152                         if (nr_thread)
1153                                 printf(" and ");
1154                         printf("%d process%s", nr_process,
1155                                                 nr_process > 1 ? "es" : "");
1156                 }
1157                 printf("\n");
1158                 fflush(stdout);
1159         }
1160
1161         signal(SIGINT, sig_handler);
1162         signal(SIGALRM, sig_handler);
1163
1164         todo = thread_number;
1165         nr_running = 0;
1166         nr_started = 0;
1167         m_rate = t_rate = 0;
1168
1169         for_each_td(td, i) {
1170                 print_status_init(td->thread_number - 1);
1171
1172                 if (!td->o.create_serialize) {
1173                         init_disk_util(td);
1174                         continue;
1175                 }
1176
1177                 /*
1178                  * do file setup here so it happens sequentially,
1179                  * we don't want X number of threads getting their
1180                  * client data interspersed on disk
1181                  */
1182                 if (setup_files(td)) {
1183                         exit_value++;
1184                         if (td->error)
1185                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid,
1186                                                         td->error, td->verror);
1187                         td_set_runstate(td, TD_REAPED);
1188                         todo--;
1189                 } else {
1190                         struct fio_file *f;
1191                         unsigned int i;
1192
1193                         /*
1194                          * for sharing to work, each job must always open
1195                          * its own files. so close them, if we opened them
1196                          * for creation
1197                          */
1198                         for_each_file(td, f, i)
1199                                 td_io_close_file(td, f);
1200                 }
1201
1202                 init_disk_util(td);
1203         }
1204
1205         set_genesis_time();
1206
1207         while (todo) {
1208                 struct thread_data *map[MAX_JOBS];
1209                 struct timeval this_start;
1210                 int this_jobs = 0, left;
1211
1212                 /*
1213                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1214                  */
1215                 for_each_td(td, i) {
1216                         if (td->runstate != TD_NOT_CREATED)
1217                                 continue;
1218
1219                         /*
1220                          * never got a chance to start, killed by other
1221                          * thread for some reason
1222                          */
1223                         if (td->terminate) {
1224                                 todo--;
1225                                 continue;
1226                         }
1227
1228                         if (td->o.start_delay) {
1229                                 spent = mtime_since_genesis();
1230
1231                                 if (td->o.start_delay * 1000 > spent)
1232                                         continue;
1233                         }
1234
1235                         if (td->o.stonewall && (nr_started || nr_running))
1236                                 break;
1237
1238                         /*
1239                          * Set state to created. Thread will transition
1240                          * to TD_INITIALIZED when it's done setting up.
1241                          */
1242                         td_set_runstate(td, TD_CREATED);
1243                         map[this_jobs++] = td;
1244                         nr_started++;
1245
1246                         if (td->o.use_thread) {
1247                                 dprint(FD_PROCESS, "will pthread_create\n");
1248                                 if (pthread_create(&td->thread, NULL,
1249                                                    thread_main, td)) {
1250                                         perror("thread_create");
1251                                         nr_started--;
1252                                         break;
1253                                 }
1254                         } else {
1255                                 dprint(FD_PROCESS, "will fork\n");
1256                                 if (!fork()) {
1257                                         int ret = fork_main(shm_id, i);
1258
1259                                         exit(ret);
1260                                 }
1261                         }
1262                         fio_mutex_down(startup_mutex);
1263                 }
1264
1265                 /*
1266                  * Wait for the started threads to transition to
1267                  * TD_INITIALIZED.
1268                  */
1269                 fio_gettime(&this_start, NULL);
1270                 left = this_jobs;
1271                 while (left && !fio_abort) {
1272                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1273                                 break;
1274
1275                         usleep(100000);
1276
1277                         for (i = 0; i < this_jobs; i++) {
1278                                 td = map[i];
1279                                 if (!td)
1280                                         continue;
1281                                 if (td->runstate == TD_INITIALIZED) {
1282                                         map[i] = NULL;
1283                                         left--;
1284                                 } else if (td->runstate >= TD_EXITED) {
1285                                         map[i] = NULL;
1286                                         left--;
1287                                         todo--;
1288                                         nr_running++; /* work-around... */
1289                                 }
1290                         }
1291                 }
1292
1293                 if (left) {
1294                         log_err("fio: %d jobs failed to start\n", left);
1295                         for (i = 0; i < this_jobs; i++) {
1296                                 td = map[i];
1297                                 if (!td)
1298                                         continue;
1299                                 kill(td->pid, SIGTERM);
1300                         }
1301                         break;
1302                 }
1303
1304                 /*
1305                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1306                  */
1307                 for_each_td(td, i) {
1308                         if (td->runstate != TD_INITIALIZED)
1309                                 continue;
1310
1311                         td_set_runstate(td, TD_RUNNING);
1312                         nr_running++;
1313                         nr_started--;
1314                         m_rate += td->o.ratemin;
1315                         t_rate += td->o.rate;
1316                         todo--;
1317                         fio_mutex_up(td->mutex);
1318                 }
1319
1320                 reap_threads(&nr_running, &t_rate, &m_rate);
1321
1322                 if (todo)
1323                         usleep(100000);
1324         }
1325
1326         while (nr_running) {
1327                 reap_threads(&nr_running, &t_rate, &m_rate);
1328                 usleep(10000);
1329         }
1330
1331         update_io_ticks();
1332         fio_unpin_memory();
1333 }
1334
1335 int main(int argc, char *argv[])
1336 {
1337         long ps;
1338
1339         sinit();
1340
1341         /*
1342          * We need locale for number printing, if it isn't set then just
1343          * go with the US format.
1344          */
1345         if (!getenv("LC_NUMERIC"))
1346                 setlocale(LC_NUMERIC, "en_US");
1347
1348         if (parse_options(argc, argv))
1349                 return 1;
1350
1351         if (!thread_number)
1352                 return 0;
1353
1354         ps = sysconf(_SC_PAGESIZE);
1355         if (ps < 0) {
1356                 log_err("Failed to get page size\n");
1357                 return 1;
1358         }
1359
1360         page_size = ps;
1361         page_mask = ps - 1;
1362
1363         if (write_bw_log) {
1364                 setup_log(&agg_io_log[DDIR_READ]);
1365                 setup_log(&agg_io_log[DDIR_WRITE]);
1366         }
1367
1368         startup_mutex = fio_mutex_init(0);
1369
1370         set_genesis_time();
1371
1372         disk_util_timer_arm();
1373
1374         run_threads();
1375
1376         if (!fio_abort) {
1377                 show_run_stats();
1378                 if (write_bw_log) {
1379                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1380                         __finish_log(agg_io_log[DDIR_WRITE],
1381                                         "agg-write_bw.log");
1382                 }
1383         }
1384
1385         fio_mutex_remove(startup_mutex);
1386         return exit_value;
1387 }