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 unsigned long done_secs = 0;
  53
  54 static struct fio_mutex *startup_mutex;
  55 static volatile int fio_abort;
  56 static int exit_value;
  57
  58 struct io_log *agg_io_log[2];
  59
  60 #define TERMINATE_ALL           (-1)
  61 #define JOB_START_TIMEOUT       (5 * 1000)
  62
  63 static inline void td_set_runstate(struct thread_data *td, int runstate)
  64 {
  65         if (td->runstate == runstate)
  66                 return;
  67
  68         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", td->pid, td->runstate,
  69                                                                 runstate);
  70         td->runstate = runstate;
  71 }
  72
  73 static void terminate_threads(int group_id)
  74 {
  75         struct thread_data *td;
  76         int i;
  77
  78         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
  79
  80         for_each_td(td, i) {
  81                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  82                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
  83                                                         td->o.name, td->pid);
  84                         td->terminate = 1;
  85                         td->o.start_delay = 0;
  86
  87                         /*
  88                          * if the thread is running, just let it exit
  89                          */
  90                         if (td->runstate < TD_RUNNING)
  91                                 kill(td->pid, SIGQUIT);
  92                         else {
  93                                 struct ioengine_ops *ops = td->io_ops;
  94
  95                                 if (ops && (ops->flags & FIO_SIGQUIT))
  96                                         kill(td->pid, SIGQUIT);
  97                         }
  98                 }
  99         }
 100 }
 101
 102 static void sig_handler(int sig)
 103 {
 104         switch (sig) {
 105         case SIGALRM:
 106                 update_io_ticks();
 107                 disk_util_timer_arm();
 108                 print_thread_status();
 109                 break;
 110         default:
 111                 printf("\nfio: terminating on signal %d\n", sig);
 112                 fflush(stdout);
 113                 terminate_threads(TERMINATE_ALL);
 114                 break;
 115         }
 116 }
 117
 118 /*
 119  * Check if we are above the minimum rate given.
 120  */
 121 static int check_min_rate(struct thread_data *td, struct timeval *now)
 122 {
 123         unsigned long long bytes = 0;
 124         unsigned long iops = 0;
 125         unsigned long spent;
 126         unsigned long rate;
 127
 128         /*
 129          * No minimum rate set, always ok
 130          */
 131         if (!td->o.ratemin && !td->o.rate_iops_min)
 132                 return 0;
 133
 134         /*
 135          * allow a 2 second settle period in the beginning
 136          */
 137         if (mtime_since(&td->start, now) < 2000)
 138                 return 0;
 139
 140         if (td_read(td)) {
 141                 iops += td->io_blocks[DDIR_READ];
 142                 bytes += td->this_io_bytes[DDIR_READ];
 143         }
 144         if (td_write(td)) {
 145                 iops += td->io_blocks[DDIR_WRITE];
 146                 bytes += td->this_io_bytes[DDIR_WRITE];
 147         }
 148
 149         /*
 150          * if rate blocks is set, sample is running
 151          */
 152         if (td->rate_bytes || td->rate_blocks) {
 153                 spent = mtime_since(&td->lastrate, now);
 154                 if (spent < td->o.ratecycle)
 155                         return 0;
 156
 157                 if (td->o.rate) {
 158                         /*
 159                          * check bandwidth specified rate
 160                          */
 161                         if (bytes < td->rate_bytes) {
 162                                 log_err("%s: min rate %u not met\n", td->o.name,
 163                                                                 td->o.ratemin);
 164                                 return 1;
 165                         } else {
 166                                 rate = (bytes - td->rate_bytes) / spent;
 167                                 if (rate < td->o.ratemin ||
 168                                     bytes < td->rate_bytes) {
 169                                         log_err("%s: min rate %u not met, got"
 170                                                 " %luKiB/sec\n", td->o.name,
 171                                                         td->o.ratemin, rate);
 172                                         return 1;
 173                                 }
 174                         }
 175                 } else {
 176                         /*
 177                          * checks iops specified rate
 178                          */
 179                         if (iops < td->o.rate_iops) {
 180                                 log_err("%s: min iops rate %u not met\n",
 181                                                 td->o.name, td->o.rate_iops);
 182                                 return 1;
 183                         } else {
 184                                 rate = (iops - td->rate_blocks) / spent;
 185                                 if (rate < td->o.rate_iops_min ||
 186                                     iops < td->rate_blocks) {
 187                                         log_err("%s: min iops rate %u not met,"
 188                                                 " got %lu\n", td->o.name,
 189                                                         td->o.rate_iops_min,
 190                                                         rate);
 191                                 }
 192                         }
 193                 }
 194         }
 195
 196         td->rate_bytes = bytes;
 197         td->rate_blocks = iops;
 198         memcpy(&td->lastrate, now, sizeof(*now));
 199         return 0;
 200 }
 201
 202 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 203 {
 204         if (!td->o.timeout)
 205                 return 0;
 206         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 207                 return 1;
 208
 209         return 0;
 210 }
 211
 212 /*
 213  * When job exits, we can cancel the in-flight IO if we are using async
 214  * io. Attempt to do so.
 215  */
 216 static void cleanup_pending_aio(struct thread_data *td)
 217 {
 218         struct list_head *entry, *n;
 219         struct io_u *io_u;
 220         int r;
 221
 222         /*
 223          * get immediately available events, if any
 224          */
 225         r = io_u_queued_complete(td, 0);
 226         if (r < 0)
 227                 return;
 228
 229         /*
 230          * now cancel remaining active events
 231          */
 232         if (td->io_ops->cancel) {
 233                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 234                         io_u = list_entry(entry, struct io_u, list);
 235
 236                         /*
 237                          * if the io_u isn't in flight, then that generally
 238                          * means someone leaked an io_u. complain but fix
 239                          * it up, so we don't stall here.
 240                          */
 241                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 242                                 log_err("fio: non-busy IO on busy list\n");
 243                                 put_io_u(td, io_u);
 244                         } else {
 245                                 r = td->io_ops->cancel(td, io_u);
 246                                 if (!r)
 247                                         put_io_u(td, io_u);
 248                         }
 249                 }
 250         }
 251
 252         if (td->cur_depth)
 253                 r = io_u_queued_complete(td, td->cur_depth);
 254 }
 255
 256 /*
 257  * Helper to handle the final sync of a file. Works just like the normal
 258  * io path, just does everything sync.
 259  */
 260 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 261 {
 262         struct io_u *io_u = __get_io_u(td);
 263         int ret;
 264
 265         if (!io_u)
 266                 return 1;
 267
 268         io_u->ddir = DDIR_SYNC;
 269         io_u->file = f;
 270
 271         if (td_io_prep(td, io_u)) {
 272                 put_io_u(td, io_u);
 273                 return 1;
 274         }
 275
 276 requeue:
 277         ret = td_io_queue(td, io_u);
 278         if (ret < 0) {
 279                 td_verror(td, io_u->error, "td_io_queue");
 280                 put_io_u(td, io_u);
 281                 return 1;
 282         } else if (ret == FIO_Q_QUEUED) {
 283                 if (io_u_queued_complete(td, 1) < 0)
 284                         return 1;
 285         } else if (ret == FIO_Q_COMPLETED) {
 286                 if (io_u->error) {
 287                         td_verror(td, io_u->error, "td_io_queue");
 288                         return 1;
 289                 }
 290
 291                 if (io_u_sync_complete(td, io_u) < 0)
 292                         return 1;
 293         } else if (ret == FIO_Q_BUSY) {
 294                 if (td_io_commit(td))
 295                         return 1;
 296                 goto requeue;
 297         }
 298
 299         return 0;
 300 }
 301
 302 /*
 303  * The main verify engine. Runs over the writes we previously submitted,
 304  * reads the blocks back in, and checks the crc/md5 of the data.
 305  */
 306 static void do_verify(struct thread_data *td)
 307 {
 308         struct fio_file *f;
 309         struct io_u *io_u;
 310         int ret, min_events;
 311         unsigned int i;
 312
 313         /*
 314          * sync io first and invalidate cache, to make sure we really
 315          * read from disk.
 316          */
 317         for_each_file(td, f, i) {
 318                 if (!(f->flags & FIO_FILE_OPEN))
 319                         continue;
 320                 if (fio_io_sync(td, f))
 321                         break;
 322                 if (file_invalidate_cache(td, f))
 323                         break;
 324         }
 325
 326         if (td->error)
 327                 return;
 328
 329         td_set_runstate(td, TD_VERIFYING);
 330
 331         io_u = NULL;
 332         while (!td->terminate) {
 333                 int ret2;
 334
 335                 io_u = __get_io_u(td);
 336                 if (!io_u)
 337                         break;
 338
 339                 if (runtime_exceeded(td, &io_u->start_time)) {
 340                         put_io_u(td, io_u);
 341                         td->terminate = 1;
 342                         break;
 343                 }
 344
 345                 if (get_next_verify(td, io_u)) {
 346                         put_io_u(td, io_u);
 347                         break;
 348                 }
 349
 350                 if (td_io_prep(td, io_u)) {
 351                         put_io_u(td, io_u);
 352                         break;
 353                 }
 354
 355                 io_u->end_io = verify_io_u;
 356
 357                 ret = td_io_queue(td, io_u);
 358                 switch (ret) {
 359                 case FIO_Q_COMPLETED:
 360                         if (io_u->error)
 361                                 ret = -io_u->error;
 362                         else if (io_u->resid) {
 363                                 int bytes = io_u->xfer_buflen - io_u->resid;
 364                                 struct fio_file *f = io_u->file;
 365
 366                                 /*
 367                                  * zero read, fail
 368                                  */
 369                                 if (!bytes) {
 370                                         td_verror(td, ENODATA, "full resid");
 371                                         put_io_u(td, io_u);
 372                                         break;
 373                                 }
 374
 375                                 io_u->xfer_buflen = io_u->resid;
 376                                 io_u->xfer_buf += bytes;
 377                                 io_u->offset += bytes;
 378
 379                                 td->ts.short_io_u[io_u->ddir]++;
 380
 381                                 if (io_u->offset == f->real_file_size)
 382                                         goto sync_done;
 383
 384                                 requeue_io_u(td, &io_u);
 385                         } else {
 386 sync_done:
 387                                 ret = io_u_sync_complete(td, io_u);
 388                                 if (ret < 0)
 389                                         break;
 390                         }
 391                         continue;
 392                 case FIO_Q_QUEUED:
 393                         break;
 394                 case FIO_Q_BUSY:
 395                         requeue_io_u(td, &io_u);
 396                         ret2 = td_io_commit(td);
 397                         if (ret2 < 0)
 398                                 ret = ret2;
 399                         break;
 400                 default:
 401                         assert(ret < 0);
 402                         td_verror(td, -ret, "td_io_queue");
 403                         break;
 404                 }
 405
 406                 if (ret < 0 || td->error)
 407                         break;
 408
 409                 /*
 410                  * if we can queue more, do so. but check if there are
 411                  * completed io_u's first.
 412                  */
 413                 min_events = 0;
 414                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 415                         min_events = 1;
 416
 417                         if (td->cur_depth > td->o.iodepth_low)
 418                                 min_events = td->cur_depth - td->o.iodepth_low;
 419                 }
 420
 421                 /*
 422                  * Reap required number of io units, if any, and do the
 423                  * verification on them through the callback handler
 424                  */
 425                 if (io_u_queued_complete(td, min_events) < 0)
 426                         break;
 427         }
 428
 429         if (!td->error) {
 430                 min_events = td->cur_depth;
 431
 432                 if (min_events)
 433                         ret = io_u_queued_complete(td, min_events);
 434         } else
 435                 cleanup_pending_aio(td);
 436
 437         td_set_runstate(td, TD_RUNNING);
 438 }
 439
 440 /*
 441  * Main IO worker function. It retrieves io_u's to process and queues
 442  * and reaps them, checking for rate and errors along the way.
 443  */
 444 static void do_io(struct thread_data *td)
 445 {
 446         struct timeval s;
 447         unsigned long usec;
 448         unsigned int i;
 449         int ret = 0;
 450
 451         td_set_runstate(td, TD_RUNNING);
 452
 453         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 454                 struct timeval comp_time;
 455                 long bytes_done = 0;
 456                 int min_evts = 0;
 457                 struct io_u *io_u;
 458                 int ret2;
 459
 460                 if (td->terminate)
 461                         break;
 462
 463                 io_u = get_io_u(td);
 464                 if (!io_u)
 465                         break;
 466
 467                 memcpy(&s, &io_u->start_time, sizeof(s));
 468
 469                 if (runtime_exceeded(td, &s)) {
 470                         put_io_u(td, io_u);
 471                         td->terminate = 1;
 472                         break;
 473                 }
 474
 475                 /*
 476                  * Add verification end_io handler, if asked to verify
 477                  * a previously written file.
 478                  */
 479                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
 480                         io_u->end_io = verify_io_u;
 481                         td_set_runstate(td, TD_VERIFYING);
 482                 } else
 483                         td_set_runstate(td, TD_RUNNING);
 484
 485                 ret = td_io_queue(td, io_u);
 486                 switch (ret) {
 487                 case FIO_Q_COMPLETED:
 488                         if (io_u->error)
 489                                 ret = -io_u->error;
 490                         else if (io_u->resid) {
 491                                 int bytes = io_u->xfer_buflen - io_u->resid;
 492                                 struct fio_file *f = io_u->file;
 493
 494                                 /*
 495                                  * zero read, fail
 496                                  */
 497                                 if (!bytes) {
 498                                         td_verror(td, ENODATA, "full resid");
 499                                         put_io_u(td, io_u);
 500                                         break;
 501                                 }
 502
 503                                 io_u->xfer_buflen = io_u->resid;
 504                                 io_u->xfer_buf += bytes;
 505                                 io_u->offset += bytes;
 506
 507                                 td->ts.short_io_u[io_u->ddir]++;
 508
 509                                 if (io_u->offset == f->real_file_size)
 510                                         goto sync_done;
 511
 512                                 requeue_io_u(td, &io_u);
 513                         } else {
 514 sync_done:
 515                                 fio_gettime(&comp_time, NULL);
 516                                 bytes_done = io_u_sync_complete(td, io_u);
 517                                 if (bytes_done < 0)
 518                                         ret = bytes_done;
 519                         }
 520                         break;
 521                 case FIO_Q_QUEUED:
 522                         /*
 523                          * if the engine doesn't have a commit hook,
 524                          * the io_u is really queued. if it does have such
 525                          * a hook, it has to call io_u_queued() itself.
 526                          */
 527                         if (td->io_ops->commit == NULL)
 528                                 io_u_queued(td, io_u);
 529                         break;
 530                 case FIO_Q_BUSY:
 531                         requeue_io_u(td, &io_u);
 532                         ret2 = td_io_commit(td);
 533                         if (ret2 < 0)
 534                                 ret = ret2;
 535                         break;
 536                 default:
 537                         assert(ret < 0);
 538                         put_io_u(td, io_u);
 539                         break;
 540                 }
 541
 542                 if (ret < 0 || td->error)
 543                         break;
 544
 545                 /*
 546                  * See if we need to complete some commands
 547                  */
 548                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 549                         min_evts = 0;
 550                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 551                                 min_evts = 1;
 552
 553                                 if (td->cur_depth > td->o.iodepth_low)
 554                                         min_evts = td->cur_depth
 555                                                         - td->o.iodepth_low;
 556                         }
 557
 558                         fio_gettime(&comp_time, NULL);
 559                         bytes_done = io_u_queued_complete(td, min_evts);
 560                         if (bytes_done < 0)
 561                                 break;
 562                 }
 563
 564                 if (!bytes_done)
 565                         continue;
 566
 567                 /*
 568                  * the rate is batched for now, it should work for batches
 569                  * of completions except the very first one which may look
 570                  * a little bursty
 571                  */
 572                 usec = utime_since(&s, &comp_time);
 573
 574                 rate_throttle(td, usec, bytes_done);
 575
 576                 if (check_min_rate(td, &comp_time)) {
 577                         if (exitall_on_terminate)
 578                                 terminate_threads(td->groupid);
 579                         td_verror(td, ENODATA, "check_min_rate");
 580                         break;
 581                 }
 582
 583                 if (td->o.thinktime) {
 584                         unsigned long long b;
 585
 586                         b = td->io_blocks[0] + td->io_blocks[1];
 587                         if (!(b % td->o.thinktime_blocks)) {
 588                                 int left;
 589
 590                                 if (td->o.thinktime_spin)
 591                                         __usec_sleep(td->o.thinktime_spin);
 592
 593                                 left = td->o.thinktime - td->o.thinktime_spin;
 594                                 if (left)
 595                                         usec_sleep(td, left);
 596                         }
 597                 }
 598         }
 599
 600         if (td->o.fill_device && td->error == ENOSPC) {
 601                 td->error = 0;
 602                 td->terminate = 1;
 603         }
 604         if (!td->error) {
 605                 struct fio_file *f;
 606
 607                 i = td->cur_depth;
 608                 if (i)
 609                         ret = io_u_queued_complete(td, i);
 610
 611                 if (should_fsync(td) && td->o.end_fsync) {
 612                         td_set_runstate(td, TD_FSYNCING);
 613
 614                         for_each_file(td, f, i) {
 615                                 if (!(f->flags & FIO_FILE_OPEN))
 616                                         continue;
 617                                 fio_io_sync(td, f);
 618                         }
 619                 }
 620         } else
 621                 cleanup_pending_aio(td);
 622
 623         /*
 624          * stop job if we failed doing any IO
 625          */
 626         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 627                 td->done = 1;
 628 }
 629
 630 static void cleanup_io_u(struct thread_data *td)
 631 {
 632         struct list_head *entry, *n;
 633         struct io_u *io_u;
 634
 635         list_for_each_safe(entry, n, &td->io_u_freelist) {
 636                 io_u = list_entry(entry, struct io_u, list);
 637
 638                 list_del(&io_u->list);
 639                 free(io_u);
 640         }
 641
 642         free_io_mem(td);
 643 }
 644
 645 /*
 646  * "randomly" fill the buffer contents
 647  */
 648 static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
 649 {
 650         long *ptr = io_u->buf;
 651
 652         if (!td->o.zero_buffers) {
 653                 while ((void *) ptr - io_u->buf < max_bs) {
 654                         *ptr = rand() * GOLDEN_RATIO_PRIME;
 655                         ptr++;
 656                 }
 657         } else
 658                 memset(ptr, 0, max_bs);
 659 }
 660
 661 static int init_io_u(struct thread_data *td)
 662 {
 663         struct io_u *io_u;
 664         unsigned int max_bs;
 665         int i, max_units;
 666         char *p;
 667
 668         max_units = td->o.iodepth;
 669         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 670         td->orig_buffer_size = (unsigned long long) max_bs
 671                                         * (unsigned long long) max_units;
 672
 673         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 674                 unsigned long bs;
 675
 676                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 677                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 678         }
 679
 680         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 681                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 682                 return 1;
 683         }
 684
 685         if (allocate_io_mem(td))
 686                 return 1;
 687
 688         if (td->o.odirect)
 689                 p = ALIGN(td->orig_buffer);
 690         else
 691                 p = td->orig_buffer;
 692
 693         for (i = 0; i < max_units; i++) {
 694                 if (td->terminate)
 695                         return 1;
 696                 io_u = malloc(sizeof(*io_u));
 697                 memset(io_u, 0, sizeof(*io_u));
 698                 INIT_LIST_HEAD(&io_u->list);
 699
 700                 if (!(td->io_ops->flags & FIO_NOIO)) {
 701                         io_u->buf = p + max_bs * i;
 702
 703                         if (td_write(td))
 704                                 fill_io_buf(td, io_u, max_bs);
 705                 }
 706
 707                 io_u->index = i;
 708                 io_u->flags = IO_U_F_FREE;
 709                 list_add(&io_u->list, &td->io_u_freelist);
 710         }
 711
 712         io_u_init_timeout();
 713
 714         return 0;
 715 }
 716
 717 static int switch_ioscheduler(struct thread_data *td)
 718 {
 719         char tmp[256], tmp2[128];
 720         FILE *f;
 721         int ret;
 722
 723         if (td->io_ops->flags & FIO_DISKLESSIO)
 724                 return 0;
 725
 726         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 727
 728         f = fopen(tmp, "r+");
 729         if (!f) {
 730                 if (errno == ENOENT) {
 731                         log_err("fio: os or kernel doesn't support IO scheduler"
 732                                 " switching\n");
 733                         return 0;
 734                 }
 735                 td_verror(td, errno, "fopen iosched");
 736                 return 1;
 737         }
 738
 739         /*
 740          * Set io scheduler.
 741          */
 742         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 743         if (ferror(f) || ret != 1) {
 744                 td_verror(td, errno, "fwrite");
 745                 fclose(f);
 746                 return 1;
 747         }
 748
 749         rewind(f);
 750
 751         /*
 752          * Read back and check that the selected scheduler is now the default.
 753          */
 754         ret = fread(tmp, 1, sizeof(tmp), f);
 755         if (ferror(f) || ret < 0) {
 756                 td_verror(td, errno, "fread");
 757                 fclose(f);
 758                 return 1;
 759         }
 760
 761         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 762         if (!strstr(tmp, tmp2)) {
 763                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 764                 td_verror(td, EINVAL, "iosched_switch");
 765                 fclose(f);
 766                 return 1;
 767         }
 768
 769         fclose(f);
 770         return 0;
 771 }
 772
 773 static int keep_running(struct thread_data *td)
 774 {
 775         unsigned long long io_done;
 776
 777         if (td->done)
 778                 return 0;
 779         if (td->o.time_based)
 780                 return 1;
 781         if (td->o.loops) {
 782                 td->o.loops--;
 783                 return 1;
 784         }
 785
 786         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 787                         + td->io_skip_bytes;
 788         if (io_done < td->o.size)
 789                 return 1;
 790
 791         return 0;
 792 }
 793
 794 static int clear_io_state(struct thread_data *td)
 795 {
 796         struct fio_file *f;
 797         unsigned int i;
 798         int ret;
 799
 800         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 801         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 802         td->zone_bytes = 0;
 803         td->rate_bytes = 0;
 804         td->rate_blocks = 0;
 805         td->rw_end_set[0] = td->rw_end_set[1] = 0;
 806
 807         td->last_was_sync = 0;
 808
 809         /*
 810          * reset file done count if we are to start over
 811          */
 812         if (td->o.time_based || td->o.loops)
 813                 td->nr_done_files = 0;
 814
 815         close_files(td);
 816
 817         ret = 0;
 818         for_each_file(td, f, i) {
 819                 f->flags &= ~FIO_FILE_DONE;
 820                 ret = td_io_open_file(td, f);
 821                 if (ret)
 822                         break;
 823         }
 824
 825         return ret;
 826 }
 827
 828 /*
 829  * Entry point for the thread based jobs. The process based jobs end up
 830  * here as well, after a little setup.
 831  */
 832 static void *thread_main(void *data)
 833 {
 834         unsigned long long runtime[2], elapsed;
 835         struct thread_data *td = data;
 836         int clear_state;
 837
 838         if (!td->o.use_thread)
 839                 setsid();
 840
 841         td->pid = getpid();
 842
 843         dprint(FD_PROCESS, "jobs pid=%d started\n", td->pid);
 844
 845         INIT_LIST_HEAD(&td->io_u_freelist);
 846         INIT_LIST_HEAD(&td->io_u_busylist);
 847         INIT_LIST_HEAD(&td->io_u_requeues);
 848         INIT_LIST_HEAD(&td->io_log_list);
 849         INIT_LIST_HEAD(&td->io_hist_list);
 850         td->io_hist_tree = RB_ROOT;
 851
 852         td_set_runstate(td, TD_INITIALIZED);
 853         fio_mutex_up(startup_mutex);
 854         fio_mutex_down(td->mutex);
 855
 856         /*
 857          * the ->mutex mutex is now no longer used, close it to avoid
 858          * eating a file descriptor
 859          */
 860         fio_mutex_remove(td->mutex);
 861
 862         /*
 863          * May alter parameters that init_io_u() will use, so we need to
 864          * do this first.
 865          */
 866         if (init_iolog(td))
 867                 goto err;
 868
 869         if (init_io_u(td))
 870                 goto err;
 871
 872         if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
 873                 td_verror(td, errno, "cpu_set_affinity");
 874                 goto err;
 875         }
 876
 877         if (td->ioprio_set) {
 878                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 879                         td_verror(td, errno, "ioprio_set");
 880                         goto err;
 881                 }
 882         }
 883
 884         if (nice(td->o.nice) == -1) {
 885                 td_verror(td, errno, "nice");
 886                 goto err;
 887         }
 888
 889         if (td->o.ioscheduler && switch_ioscheduler(td))
 890                 goto err;
 891
 892         if (!td->o.create_serialize && setup_files(td))
 893                 goto err;
 894
 895         if (td_io_init(td))
 896                 goto err;
 897
 898         if (open_files(td))
 899                 goto err;
 900
 901         if (init_random_map(td))
 902                 goto err;
 903
 904         if (td->o.exec_prerun) {
 905                 if (system(td->o.exec_prerun) < 0)
 906                         goto err;
 907         }
 908
 909         fio_gettime(&td->epoch, NULL);
 910         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 911         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 912
 913         runtime[0] = runtime[1] = 0;
 914         clear_state = 0;
 915         while (keep_running(td)) {
 916                 fio_gettime(&td->start, NULL);
 917                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 918
 919                 if (td->o.ratemin)
 920                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
 921                                                         sizeof(td->lastrate));
 922
 923                 if (clear_state && clear_io_state(td))
 924                         break;
 925
 926                 prune_io_piece_log(td);
 927
 928                 do_io(td);
 929
 930                 clear_state = 1;
 931
 932                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
 933                         if (td->rw_end_set[DDIR_READ])
 934                                 elapsed = utime_since(&td->start,
 935                                                       &td->rw_end[DDIR_READ]);
 936                         else
 937                                 elapsed = utime_since_now(&td->start);
 938
 939                         runtime[DDIR_READ] += elapsed;
 940                 }
 941                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
 942                         if (td->rw_end_set[DDIR_WRITE])
 943                                 elapsed = utime_since(&td->start,
 944                                                       &td->rw_end[DDIR_WRITE]);
 945                         else
 946                                 elapsed = utime_since_now(&td->start);
 947
 948                         runtime[DDIR_WRITE] += elapsed;
 949                 }
 950
 951                 if (td->error || td->terminate)
 952                         break;
 953
 954                 if (!td->o.do_verify ||
 955                     td->o.verify == VERIFY_NONE ||
 956                     (td->io_ops->flags & FIO_UNIDIR))
 957                         continue;
 958
 959                 if (clear_io_state(td))
 960                         break;
 961
 962                 fio_gettime(&td->start, NULL);
 963
 964                 do_verify(td);
 965
 966                 runtime[DDIR_READ] += utime_since_now(&td->start);
 967
 968                 if (td->error || td->terminate)
 969                         break;
 970         }
 971
 972         update_rusage_stat(td);
 973         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
 974         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
 975         td->ts.total_run_time = mtime_since_now(&td->epoch);
 976         td->ts.io_bytes[0] = td->io_bytes[0];
 977         td->ts.io_bytes[1] = td->io_bytes[1];
 978
 979         if (td->ts.bw_log)
 980                 finish_log(td, td->ts.bw_log, "bw");
 981         if (td->ts.slat_log)
 982                 finish_log(td, td->ts.slat_log, "slat");
 983         if (td->ts.clat_log)
 984                 finish_log(td, td->ts.clat_log, "clat");
 985         if (td->o.exec_postrun) {
 986                 if (system(td->o.exec_postrun) < 0)
 987                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
 988         }
 989
 990         if (exitall_on_terminate)
 991                 terminate_threads(td->groupid);
 992
 993 err:
 994         if (td->error)
 995                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error,
 996                                                         td->verror);
 997         close_and_free_files(td);
 998         close_ioengine(td);
 999         cleanup_io_u(td);
1000
1001         /*
1002          * do this very late, it will log file closing as well
1003          */
1004         if (td->o.write_iolog_file)
1005                 write_iolog_close(td);
1006
1007         options_mem_free(td);
1008         td_set_runstate(td, TD_EXITED);
1009         return (void *) (unsigned long) td->error;
1010 }
1011
1012 /*
1013  * We cannot pass the td data into a forked process, so attach the td and
1014  * pass it to the thread worker.
1015  */
1016 static int fork_main(int shmid, int offset)
1017 {
1018         struct thread_data *td;
1019         void *data, *ret;
1020
1021         data = shmat(shmid, NULL, 0);
1022         if (data == (void *) -1) {
1023                 int __err = errno;
1024
1025                 perror("shmat");
1026                 return __err;
1027         }
1028
1029         td = data + offset * sizeof(struct thread_data);
1030         ret = thread_main(td);
1031         shmdt(data);
1032         return (int) (unsigned long) ret;
1033 }
1034
1035 /*
1036  * Run over the job map and reap the threads that have exited, if any.
1037  */
1038 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1039 {
1040         struct thread_data *td;
1041         int i, cputhreads, realthreads, pending, status, ret;
1042
1043         /*
1044          * reap exited threads (TD_EXITED -> TD_REAPED)
1045          */
1046         realthreads = pending = cputhreads = 0;
1047         for_each_td(td, i) {
1048                 int flags = 0;
1049
1050                 /*
1051                  * ->io_ops is NULL for a thread that has closed its
1052                  * io engine
1053                  */
1054                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1055                         cputhreads++;
1056                 else
1057                         realthreads++;
1058
1059                 if (!td->pid) {
1060                         pending++;
1061                         continue;
1062                 }
1063                 if (td->runstate == TD_REAPED)
1064                         continue;
1065                 if (td->o.use_thread) {
1066                         if (td->runstate == TD_EXITED) {
1067                                 td_set_runstate(td, TD_REAPED);
1068                                 goto reaped;
1069                         }
1070                         continue;
1071                 }
1072
1073                 flags = WNOHANG;
1074                 if (td->runstate == TD_EXITED)
1075                         flags = 0;
1076
1077                 /*
1078                  * check if someone quit or got killed in an unusual way
1079                  */
1080                 ret = waitpid(td->pid, &status, flags);
1081                 if (ret < 0) {
1082                         if (errno == ECHILD) {
1083                                 log_err("fio: pid=%d disappeared %d\n", td->pid,
1084                                                                 td->runstate);
1085                                 td_set_runstate(td, TD_REAPED);
1086                                 goto reaped;
1087                         }
1088                         perror("waitpid");
1089                 } else if (ret == td->pid) {
1090                         if (WIFSIGNALED(status)) {
1091                                 int sig = WTERMSIG(status);
1092
1093                                 if (sig != SIGQUIT)
1094                                         log_err("fio: pid=%d, got signal=%d\n",
1095                                                                 td->pid, sig);
1096                                 td_set_runstate(td, TD_REAPED);
1097                                 goto reaped;
1098                         }
1099                         if (WIFEXITED(status)) {
1100                                 if (WEXITSTATUS(status) && !td->error)
1101                                         td->error = WEXITSTATUS(status);
1102
1103                                 td_set_runstate(td, TD_REAPED);
1104                                 goto reaped;
1105                         }
1106                 }
1107
1108                 /*
1109                  * thread is not dead, continue
1110                  */
1111                 pending++;
1112                 continue;
1113 reaped:
1114                 (*nr_running)--;
1115                 (*m_rate) -= td->o.ratemin;
1116                 (*t_rate) -= td->o.rate;
1117                 if (!td->pid)
1118                         pending--;
1119
1120                 if (td->error)
1121                         exit_value++;
1122
1123                 done_secs += mtime_since_now(&td->epoch) / 1000;
1124         }
1125
1126         if (*nr_running == cputhreads && !pending && realthreads)
1127                 terminate_threads(TERMINATE_ALL);
1128 }
1129
1130 /*
1131  * Main function for kicking off and reaping jobs, as needed.
1132  */
1133 static void run_threads(void)
1134 {
1135         struct thread_data *td;
1136         unsigned long spent;
1137         int i, todo, nr_running, m_rate, t_rate, nr_started;
1138
1139         if (fio_pin_memory())
1140                 return;
1141
1142         if (!terse_output) {
1143                 printf("Starting ");
1144                 if (nr_thread)
1145                         printf("%d thread%s", nr_thread,
1146                                                 nr_thread > 1 ? "s" : "");
1147                 if (nr_process) {
1148                         if (nr_thread)
1149                                 printf(" and ");
1150                         printf("%d process%s", nr_process,
1151                                                 nr_process > 1 ? "es" : "");
1152                 }
1153                 printf("\n");
1154                 fflush(stdout);
1155         }
1156
1157         signal(SIGINT, sig_handler);
1158         signal(SIGALRM, sig_handler);
1159
1160         todo = thread_number;
1161         nr_running = 0;
1162         nr_started = 0;
1163         m_rate = t_rate = 0;
1164
1165         for_each_td(td, i) {
1166                 print_status_init(td->thread_number - 1);
1167
1168                 if (!td->o.create_serialize) {
1169                         init_disk_util(td);
1170                         continue;
1171                 }
1172
1173                 /*
1174                  * do file setup here so it happens sequentially,
1175                  * we don't want X number of threads getting their
1176                  * client data interspersed on disk
1177                  */
1178                 if (setup_files(td)) {
1179                         exit_value++;
1180                         if (td->error)
1181                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid,
1182                                                         td->error, td->verror);
1183                         td_set_runstate(td, TD_REAPED);
1184                         todo--;
1185                 } else {
1186                         struct fio_file *f;
1187                         unsigned int i;
1188
1189                         /*
1190                          * for sharing to work, each job must always open
1191                          * its own files. so close them, if we opened them
1192                          * for creation
1193                          */
1194                         for_each_file(td, f, i)
1195                                 td_io_close_file(td, f);
1196                 }
1197
1198                 init_disk_util(td);
1199         }
1200
1201         set_genesis_time();
1202
1203         while (todo) {
1204                 struct thread_data *map[MAX_JOBS];
1205                 struct timeval this_start;
1206                 int this_jobs = 0, left;
1207
1208                 /*
1209                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1210                  */
1211                 for_each_td(td, i) {
1212                         if (td->runstate != TD_NOT_CREATED)
1213                                 continue;
1214
1215                         /*
1216                          * never got a chance to start, killed by other
1217                          * thread for some reason
1218                          */
1219                         if (td->terminate) {
1220                                 todo--;
1221                                 continue;
1222                         }
1223
1224                         if (td->o.start_delay) {
1225                                 spent = mtime_since_genesis();
1226
1227                                 if (td->o.start_delay * 1000 > spent)
1228                                         continue;
1229                         }
1230
1231                         if (td->o.stonewall && (nr_started || nr_running)) {
1232                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1233                                                         td->o.name);
1234                                 break;
1235                         }
1236
1237                         /*
1238                          * Set state to created. Thread will transition
1239                          * to TD_INITIALIZED when it's done setting up.
1240                          */
1241                         td_set_runstate(td, TD_CREATED);
1242                         map[this_jobs++] = td;
1243                         nr_started++;
1244
1245                         if (td->o.use_thread) {
1246                                 dprint(FD_PROCESS, "will pthread_create\n");
1247                                 if (pthread_create(&td->thread, NULL,
1248                                                    thread_main, td)) {
1249                                         perror("pthread_create");
1250                                         nr_started--;
1251                                         break;
1252                                 }
1253                                 if (pthread_detach(td->thread) < 0)
1254                                         perror("pthread_detach");
1255                         } else {
1256                                 dprint(FD_PROCESS, "will fork\n");
1257                                 if (!fork()) {
1258                                         int ret = fork_main(shm_id, i);
1259
1260                                         exit(ret);
1261                                 }
1262                         }
1263                         fio_mutex_down(startup_mutex);
1264                 }
1265
1266                 /*
1267                  * Wait for the started threads to transition to
1268                  * TD_INITIALIZED.
1269                  */
1270                 fio_gettime(&this_start, NULL);
1271                 left = this_jobs;
1272                 while (left && !fio_abort) {
1273                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1274                                 break;
1275
1276                         usleep(100000);
1277
1278                         for (i = 0; i < this_jobs; i++) {
1279                                 td = map[i];
1280                                 if (!td)
1281                                         continue;
1282                                 if (td->runstate == TD_INITIALIZED) {
1283                                         map[i] = NULL;
1284                                         left--;
1285                                 } else if (td->runstate >= TD_EXITED) {
1286                                         map[i] = NULL;
1287                                         left--;
1288                                         todo--;
1289                                         nr_running++; /* work-around... */
1290                                 }
1291                         }
1292                 }
1293
1294                 if (left) {
1295                         log_err("fio: %d jobs failed to start\n", left);
1296                         for (i = 0; i < this_jobs; i++) {
1297                                 td = map[i];
1298                                 if (!td)
1299                                         continue;
1300                                 kill(td->pid, SIGTERM);
1301                         }
1302                         break;
1303                 }
1304
1305                 /*
1306                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1307                  */
1308                 for_each_td(td, i) {
1309                         if (td->runstate != TD_INITIALIZED)
1310                                 continue;
1311
1312                         td_set_runstate(td, TD_RUNNING);
1313                         nr_running++;
1314                         nr_started--;
1315                         m_rate += td->o.ratemin;
1316                         t_rate += td->o.rate;
1317                         todo--;
1318                         fio_mutex_up(td->mutex);
1319                 }
1320
1321                 reap_threads(&nr_running, &t_rate, &m_rate);
1322
1323                 if (todo)
1324                         usleep(100000);
1325         }
1326
1327         while (nr_running) {
1328                 reap_threads(&nr_running, &t_rate, &m_rate);
1329                 usleep(10000);
1330         }
1331
1332         update_io_ticks();
1333         fio_unpin_memory();
1334 }
1335
1336 int main(int argc, char *argv[])
1337 {
1338         long ps;
1339
1340         sinit();
1341
1342         /*
1343          * We need locale for number printing, if it isn't set then just
1344          * go with the US format.
1345          */
1346         if (!getenv("LC_NUMERIC"))
1347                 setlocale(LC_NUMERIC, "en_US");
1348
1349         if (parse_options(argc, argv))
1350                 return 1;
1351
1352         if (!thread_number)
1353                 return 0;
1354
1355         ps = sysconf(_SC_PAGESIZE);
1356         if (ps < 0) {
1357                 log_err("Failed to get page size\n");
1358                 return 1;
1359         }
1360
1361         page_size = ps;
1362         page_mask = ps - 1;
1363
1364         if (write_bw_log) {
1365                 setup_log(&agg_io_log[DDIR_READ]);
1366                 setup_log(&agg_io_log[DDIR_WRITE]);
1367         }
1368
1369         startup_mutex = fio_mutex_init(0);
1370
1371         set_genesis_time();
1372
1373         disk_util_timer_arm();
1374
1375         run_threads();
1376
1377         if (!fio_abort) {
1378                 show_run_stats();
1379                 if (write_bw_log) {
1380                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1381                         __finish_log(agg_io_log[DDIR_WRITE],
1382                                         "agg-write_bw.log");
1383                 }
1384         }
1385
1386         fio_mutex_remove(startup_mutex);
1387         return exit_value;
1388 }