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 static struct itimerval itimer;
  58
  59 struct io_log *agg_io_log[2];
  60
  61 #define TERMINATE_ALL           (-1)
  62 #define JOB_START_TIMEOUT       (5 * 1000)
  63
  64 static inline void td_set_runstate(struct thread_data *td, int runstate)
  65 {
  66         if (td->runstate == runstate)
  67                 return;
  68
  69         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
  70                                                 td->runstate, runstate);
  71         td->runstate = runstate;
  72 }
  73
  74 static void terminate_threads(int group_id)
  75 {
  76         struct thread_data *td;
  77         int i;
  78
  79         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
  80
  81         for_each_td(td, i) {
  82                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  83                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
  84                                                 td->o.name, (int) td->pid);
  85                         td->terminate = 1;
  86                         td->o.start_delay = 0;
  87
  88                         /*
  89                          * if the thread is running, just let it exit
  90                          */
  91                         if (td->runstate < TD_RUNNING)
  92                                 kill(td->pid, SIGQUIT);
  93                         else {
  94                                 struct ioengine_ops *ops = td->io_ops;
  95
  96                                 if (ops && (ops->flags & FIO_SIGQUIT))
  97                                         kill(td->pid, SIGQUIT);
  98                         }
  99                 }
 100         }
 101 }
 102
 103 static void status_timer_arm(void)
 104 {
 105         itimer.it_value.tv_sec = 0;
 106         itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
 107         setitimer(ITIMER_REAL, &itimer, NULL);
 108 }
 109
 110 /*
 111  * We need to rearm on BSD/solaris. Switch this to sigaction in the future...
 112  */
 113 static void set_sig_handlers(void (*sighandler)(int))
 114 {
 115         signal(SIGINT, sighandler);
 116         signal(SIGALRM, sighandler);
 117 }
 118
 119 static void sig_handler(int sig)
 120 {
 121         set_sig_handlers(sig_handler);
 122
 123         if (!threads)
 124                 return;
 125
 126         switch (sig) {
 127         case SIGALRM:
 128                 update_io_ticks();
 129                 print_thread_status();
 130                 status_timer_arm();
 131                 break;
 132         default:
 133                 printf("\nfio: terminating on signal %d\n", sig);
 134                 fflush(stdout);
 135                 terminate_threads(TERMINATE_ALL);
 136                 break;
 137         }
 138 }
 139
 140 /*
 141  * Check if we are above the minimum rate given.
 142  */
 143 static int check_min_rate(struct thread_data *td, struct timeval *now)
 144 {
 145         unsigned long long bytes = 0;
 146         unsigned long iops = 0;
 147         unsigned long spent;
 148         unsigned long rate;
 149
 150         /*
 151          * No minimum rate set, always ok
 152          */
 153         if (!td->o.ratemin && !td->o.rate_iops_min)
 154                 return 0;
 155
 156         /*
 157          * allow a 2 second settle period in the beginning
 158          */
 159         if (mtime_since(&td->start, now) < 2000)
 160                 return 0;
 161
 162         if (td_read(td)) {
 163                 iops += td->io_blocks[DDIR_READ];
 164                 bytes += td->this_io_bytes[DDIR_READ];
 165         }
 166         if (td_write(td)) {
 167                 iops += td->io_blocks[DDIR_WRITE];
 168                 bytes += td->this_io_bytes[DDIR_WRITE];
 169         }
 170
 171         /*
 172          * if rate blocks is set, sample is running
 173          */
 174         if (td->rate_bytes || td->rate_blocks) {
 175                 spent = mtime_since(&td->lastrate, now);
 176                 if (spent < td->o.ratecycle)
 177                         return 0;
 178
 179                 if (td->o.rate) {
 180                         /*
 181                          * check bandwidth specified rate
 182                          */
 183                         if (bytes < td->rate_bytes) {
 184                                 log_err("%s: min rate %u not met\n", td->o.name,
 185                                                                 td->o.ratemin);
 186                                 return 1;
 187                         } else {
 188                                 rate = (bytes - td->rate_bytes) / spent;
 189                                 if (rate < td->o.ratemin ||
 190                                     bytes < td->rate_bytes) {
 191                                         log_err("%s: min rate %u not met, got"
 192                                                 " %luKiB/sec\n", td->o.name,
 193                                                         td->o.ratemin, rate);
 194                                         return 1;
 195                                 }
 196                         }
 197                 } else {
 198                         /*
 199                          * checks iops specified rate
 200                          */
 201                         if (iops < td->o.rate_iops) {
 202                                 log_err("%s: min iops rate %u not met\n",
 203                                                 td->o.name, td->o.rate_iops);
 204                                 return 1;
 205                         } else {
 206                                 rate = (iops - td->rate_blocks) / spent;
 207                                 if (rate < td->o.rate_iops_min ||
 208                                     iops < td->rate_blocks) {
 209                                         log_err("%s: min iops rate %u not met,"
 210                                                 " got %lu\n", td->o.name,
 211                                                         td->o.rate_iops_min,
 212                                                         rate);
 213                                 }
 214                         }
 215                 }
 216         }
 217
 218         td->rate_bytes = bytes;
 219         td->rate_blocks = iops;
 220         memcpy(&td->lastrate, now, sizeof(*now));
 221         return 0;
 222 }
 223
 224 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 225 {
 226         if (!td->o.timeout)
 227                 return 0;
 228         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 229                 return 1;
 230
 231         return 0;
 232 }
 233
 234 /*
 235  * When job exits, we can cancel the in-flight IO if we are using async
 236  * io. Attempt to do so.
 237  */
 238 static void cleanup_pending_aio(struct thread_data *td)
 239 {
 240         struct list_head *entry, *n;
 241         struct io_u *io_u;
 242         int r;
 243
 244         /*
 245          * get immediately available events, if any
 246          */
 247         r = io_u_queued_complete(td, 0);
 248         if (r < 0)
 249                 return;
 250
 251         /*
 252          * now cancel remaining active events
 253          */
 254         if (td->io_ops->cancel) {
 255                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 256                         io_u = list_entry(entry, struct io_u, list);
 257
 258                         /*
 259                          * if the io_u isn't in flight, then that generally
 260                          * means someone leaked an io_u. complain but fix
 261                          * it up, so we don't stall here.
 262                          */
 263                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 264                                 log_err("fio: non-busy IO on busy list\n");
 265                                 put_io_u(td, io_u);
 266                         } else {
 267                                 r = td->io_ops->cancel(td, io_u);
 268                                 if (!r)
 269                                         put_io_u(td, io_u);
 270                         }
 271                 }
 272         }
 273
 274         if (td->cur_depth)
 275                 r = io_u_queued_complete(td, td->cur_depth);
 276 }
 277
 278 /*
 279  * Helper to handle the final sync of a file. Works just like the normal
 280  * io path, just does everything sync.
 281  */
 282 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 283 {
 284         struct io_u *io_u = __get_io_u(td);
 285         int ret;
 286
 287         if (!io_u)
 288                 return 1;
 289
 290         io_u->ddir = DDIR_SYNC;
 291         io_u->file = f;
 292
 293         if (td_io_prep(td, io_u)) {
 294                 put_io_u(td, io_u);
 295                 return 1;
 296         }
 297
 298 requeue:
 299         ret = td_io_queue(td, io_u);
 300         if (ret < 0) {
 301                 td_verror(td, io_u->error, "td_io_queue");
 302                 put_io_u(td, io_u);
 303                 return 1;
 304         } else if (ret == FIO_Q_QUEUED) {
 305                 if (io_u_queued_complete(td, 1) < 0)
 306                         return 1;
 307         } else if (ret == FIO_Q_COMPLETED) {
 308                 if (io_u->error) {
 309                         td_verror(td, io_u->error, "td_io_queue");
 310                         return 1;
 311                 }
 312
 313                 if (io_u_sync_complete(td, io_u) < 0)
 314                         return 1;
 315         } else if (ret == FIO_Q_BUSY) {
 316                 if (td_io_commit(td))
 317                         return 1;
 318                 goto requeue;
 319         }
 320
 321         return 0;
 322 }
 323
 324 /*
 325  * The main verify engine. Runs over the writes we previously submitted,
 326  * reads the blocks back in, and checks the crc/md5 of the data.
 327  */
 328 static void do_verify(struct thread_data *td)
 329 {
 330         struct fio_file *f;
 331         struct io_u *io_u;
 332         int ret, min_events;
 333         unsigned int i;
 334
 335         /*
 336          * sync io first and invalidate cache, to make sure we really
 337          * read from disk.
 338          */
 339         for_each_file(td, f, i) {
 340                 if (!(f->flags & FIO_FILE_OPEN))
 341                         continue;
 342                 if (fio_io_sync(td, f))
 343                         break;
 344                 if (file_invalidate_cache(td, f))
 345                         break;
 346         }
 347
 348         if (td->error)
 349                 return;
 350
 351         td_set_runstate(td, TD_VERIFYING);
 352
 353         io_u = NULL;
 354         while (!td->terminate) {
 355                 int ret2;
 356
 357                 io_u = __get_io_u(td);
 358                 if (!io_u)
 359                         break;
 360
 361                 if (runtime_exceeded(td, &io_u->start_time)) {
 362                         put_io_u(td, io_u);
 363                         td->terminate = 1;
 364                         break;
 365                 }
 366
 367                 if (get_next_verify(td, io_u)) {
 368                         put_io_u(td, io_u);
 369                         break;
 370                 }
 371
 372                 if (td_io_prep(td, io_u)) {
 373                         put_io_u(td, io_u);
 374                         break;
 375                 }
 376
 377                 io_u->end_io = verify_io_u;
 378
 379                 ret = td_io_queue(td, io_u);
 380                 switch (ret) {
 381                 case FIO_Q_COMPLETED:
 382                         if (io_u->error)
 383                                 ret = -io_u->error;
 384                         else if (io_u->resid) {
 385                                 int bytes = io_u->xfer_buflen - io_u->resid;
 386                                 struct fio_file *f = io_u->file;
 387
 388                                 /*
 389                                  * zero read, fail
 390                                  */
 391                                 if (!bytes) {
 392                                         td_verror(td, ENODATA, "full resid");
 393                                         put_io_u(td, io_u);
 394                                         break;
 395                                 }
 396
 397                                 io_u->xfer_buflen = io_u->resid;
 398                                 io_u->xfer_buf += bytes;
 399                                 io_u->offset += bytes;
 400
 401                                 td->ts.short_io_u[io_u->ddir]++;
 402
 403                                 if (io_u->offset == f->real_file_size)
 404                                         goto sync_done;
 405
 406                                 requeue_io_u(td, &io_u);
 407                         } else {
 408 sync_done:
 409                                 ret = io_u_sync_complete(td, io_u);
 410                                 if (ret < 0)
 411                                         break;
 412                         }
 413                         continue;
 414                 case FIO_Q_QUEUED:
 415                         break;
 416                 case FIO_Q_BUSY:
 417                         requeue_io_u(td, &io_u);
 418                         ret2 = td_io_commit(td);
 419                         if (ret2 < 0)
 420                                 ret = ret2;
 421                         break;
 422                 default:
 423                         assert(ret < 0);
 424                         td_verror(td, -ret, "td_io_queue");
 425                         break;
 426                 }
 427
 428                 if (ret < 0 || td->error)
 429                         break;
 430
 431                 /*
 432                  * if we can queue more, do so. but check if there are
 433                  * completed io_u's first.
 434                  */
 435                 min_events = 0;
 436                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 437                         if (td->cur_depth >= td->o.iodepth_low)
 438                                 min_events = td->cur_depth - td->o.iodepth_low;
 439                         if (!min_events)
 440                                 min_events = 1;
 441                 }
 442
 443                 /*
 444                  * Reap required number of io units, if any, and do the
 445                  * verification on them through the callback handler
 446                  */
 447                 if (io_u_queued_complete(td, min_events) < 0)
 448                         break;
 449         }
 450
 451         if (!td->error) {
 452                 min_events = td->cur_depth;
 453
 454                 if (min_events)
 455                         ret = io_u_queued_complete(td, min_events);
 456         } else
 457                 cleanup_pending_aio(td);
 458
 459         td_set_runstate(td, TD_RUNNING);
 460 }
 461
 462 /*
 463  * Main IO worker function. It retrieves io_u's to process and queues
 464  * and reaps them, checking for rate and errors along the way.
 465  */
 466 static void do_io(struct thread_data *td)
 467 {
 468         struct timeval s;
 469         unsigned long usec;
 470         unsigned int i;
 471         int ret = 0;
 472
 473         td_set_runstate(td, TD_RUNNING);
 474
 475         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 476                 struct timeval comp_time;
 477                 long bytes_done = 0;
 478                 int min_evts = 0;
 479                 struct io_u *io_u;
 480                 int ret2;
 481
 482                 if (td->terminate)
 483                         break;
 484
 485                 io_u = get_io_u(td);
 486                 if (!io_u)
 487                         break;
 488
 489                 memcpy(&s, &io_u->start_time, sizeof(s));
 490
 491                 if (runtime_exceeded(td, &s)) {
 492                         put_io_u(td, io_u);
 493                         td->terminate = 1;
 494                         break;
 495                 }
 496
 497                 /*
 498                  * Add verification end_io handler, if asked to verify
 499                  * a previously written file.
 500                  */
 501                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
 502                         io_u->end_io = verify_io_u;
 503                         td_set_runstate(td, TD_VERIFYING);
 504                 } else
 505                         td_set_runstate(td, TD_RUNNING);
 506
 507                 ret = td_io_queue(td, io_u);
 508                 switch (ret) {
 509                 case FIO_Q_COMPLETED:
 510                         if (io_u->error)
 511                                 ret = -io_u->error;
 512                         else if (io_u->resid) {
 513                                 int bytes = io_u->xfer_buflen - io_u->resid;
 514                                 struct fio_file *f = io_u->file;
 515
 516                                 /*
 517                                  * zero read, fail
 518                                  */
 519                                 if (!bytes) {
 520                                         td_verror(td, ENODATA, "full resid");
 521                                         put_io_u(td, io_u);
 522                                         break;
 523                                 }
 524
 525                                 io_u->xfer_buflen = io_u->resid;
 526                                 io_u->xfer_buf += bytes;
 527                                 io_u->offset += bytes;
 528
 529                                 td->ts.short_io_u[io_u->ddir]++;
 530
 531                                 if (io_u->offset == f->real_file_size)
 532                                         goto sync_done;
 533
 534                                 requeue_io_u(td, &io_u);
 535                         } else {
 536 sync_done:
 537                                 fio_gettime(&comp_time, NULL);
 538                                 bytes_done = io_u_sync_complete(td, io_u);
 539                                 if (bytes_done < 0)
 540                                         ret = bytes_done;
 541                         }
 542                         break;
 543                 case FIO_Q_QUEUED:
 544                         /*
 545                          * if the engine doesn't have a commit hook,
 546                          * the io_u is really queued. if it does have such
 547                          * a hook, it has to call io_u_queued() itself.
 548                          */
 549                         if (td->io_ops->commit == NULL)
 550                                 io_u_queued(td, io_u);
 551                         break;
 552                 case FIO_Q_BUSY:
 553                         requeue_io_u(td, &io_u);
 554                         ret2 = td_io_commit(td);
 555                         if (ret2 < 0)
 556                                 ret = ret2;
 557                         break;
 558                 default:
 559                         assert(ret < 0);
 560                         put_io_u(td, io_u);
 561                         break;
 562                 }
 563
 564                 if (ret < 0 || td->error)
 565                         break;
 566
 567                 /*
 568                  * See if we need to complete some commands
 569                  */
 570                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 571                         min_evts = 0;
 572                         if (td->cur_depth >= td->o.iodepth_low)
 573                                 min_evts = td->cur_depth - td->o.iodepth_low;
 574                         if (!min_evts)
 575                                 min_evts = 1;
 576                         fio_gettime(&comp_time, NULL);
 577                         bytes_done = io_u_queued_complete(td, min_evts);
 578                         if (bytes_done < 0)
 579                                 break;
 580                 }
 581
 582                 if (!bytes_done)
 583                         continue;
 584
 585                 /*
 586                  * the rate is batched for now, it should work for batches
 587                  * of completions except the very first one which may look
 588                  * a little bursty
 589                  */
 590                 usec = utime_since(&s, &comp_time);
 591
 592                 rate_throttle(td, usec, bytes_done);
 593
 594                 if (check_min_rate(td, &comp_time)) {
 595                         if (exitall_on_terminate)
 596                                 terminate_threads(td->groupid);
 597                         td_verror(td, ENODATA, "check_min_rate");
 598                         break;
 599                 }
 600
 601                 if (td->o.thinktime) {
 602                         unsigned long long b;
 603
 604                         b = td->io_blocks[0] + td->io_blocks[1];
 605                         if (!(b % td->o.thinktime_blocks)) {
 606                                 int left;
 607
 608                                 if (td->o.thinktime_spin)
 609                                         __usec_sleep(td->o.thinktime_spin);
 610
 611                                 left = td->o.thinktime - td->o.thinktime_spin;
 612                                 if (left)
 613                                         usec_sleep(td, left);
 614                         }
 615                 }
 616         }
 617
 618         if (td->o.fill_device && td->error == ENOSPC) {
 619                 td->error = 0;
 620                 td->terminate = 1;
 621         }
 622         if (!td->error) {
 623                 struct fio_file *f;
 624
 625                 i = td->cur_depth;
 626                 if (i)
 627                         ret = io_u_queued_complete(td, i);
 628
 629                 if (should_fsync(td) && td->o.end_fsync) {
 630                         td_set_runstate(td, TD_FSYNCING);
 631
 632                         for_each_file(td, f, i) {
 633                                 if (!(f->flags & FIO_FILE_OPEN))
 634                                         continue;
 635                                 fio_io_sync(td, f);
 636                         }
 637                 }
 638         } else
 639                 cleanup_pending_aio(td);
 640
 641         /*
 642          * stop job if we failed doing any IO
 643          */
 644         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 645                 td->done = 1;
 646 }
 647
 648 static void cleanup_io_u(struct thread_data *td)
 649 {
 650         struct list_head *entry, *n;
 651         struct io_u *io_u;
 652
 653         list_for_each_safe(entry, n, &td->io_u_freelist) {
 654                 io_u = list_entry(entry, struct io_u, list);
 655
 656                 list_del(&io_u->list);
 657                 free(io_u);
 658         }
 659
 660         free_io_mem(td);
 661 }
 662
 663 static int init_io_u(struct thread_data *td)
 664 {
 665         struct io_u *io_u;
 666         unsigned int max_bs;
 667         int i, max_units;
 668         char *p;
 669
 670         max_units = td->o.iodepth;
 671         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 672         td->orig_buffer_size = (unsigned long long) max_bs
 673                                         * (unsigned long long) max_units;
 674
 675         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 676                 unsigned long bs;
 677
 678                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 679                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 680         }
 681
 682         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 683                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 684                 return 1;
 685         }
 686
 687         if (allocate_io_mem(td))
 688                 return 1;
 689
 690         if (td->o.odirect)
 691                 p = ALIGN(td->orig_buffer);
 692         else
 693                 p = td->orig_buffer;
 694
 695         for (i = 0; i < max_units; i++) {
 696                 if (td->terminate)
 697                         return 1;
 698                 io_u = malloc(sizeof(*io_u));
 699                 memset(io_u, 0, sizeof(*io_u));
 700                 INIT_LIST_HEAD(&io_u->list);
 701
 702                 if (!(td->io_ops->flags & FIO_NOIO)) {
 703                         io_u->buf = p + max_bs * i;
 704
 705                         if (td_write(td) && !td->o.refill_buffers)
 706                                 io_u_fill_buffer(td, io_u, max_bs);
 707                 }
 708
 709                 io_u->index = i;
 710                 io_u->flags = IO_U_F_FREE;
 711                 list_add(&io_u->list, &td->io_u_freelist);
 712         }
 713
 714         io_u_init_timeout();
 715
 716         return 0;
 717 }
 718
 719 static int switch_ioscheduler(struct thread_data *td)
 720 {
 721         char tmp[256], tmp2[128];
 722         FILE *f;
 723         int ret;
 724
 725         if (td->io_ops->flags & FIO_DISKLESSIO)
 726                 return 0;
 727
 728         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 729
 730         f = fopen(tmp, "r+");
 731         if (!f) {
 732                 if (errno == ENOENT) {
 733                         log_err("fio: os or kernel doesn't support IO scheduler"
 734                                 " switching\n");
 735                         return 0;
 736                 }
 737                 td_verror(td, errno, "fopen iosched");
 738                 return 1;
 739         }
 740
 741         /*
 742          * Set io scheduler.
 743          */
 744         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 745         if (ferror(f) || ret != 1) {
 746                 td_verror(td, errno, "fwrite");
 747                 fclose(f);
 748                 return 1;
 749         }
 750
 751         rewind(f);
 752
 753         /*
 754          * Read back and check that the selected scheduler is now the default.
 755          */
 756         ret = fread(tmp, 1, sizeof(tmp), f);
 757         if (ferror(f) || ret < 0) {
 758                 td_verror(td, errno, "fread");
 759                 fclose(f);
 760                 return 1;
 761         }
 762
 763         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 764         if (!strstr(tmp, tmp2)) {
 765                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 766                 td_verror(td, EINVAL, "iosched_switch");
 767                 fclose(f);
 768                 return 1;
 769         }
 770
 771         fclose(f);
 772         return 0;
 773 }
 774
 775 static int keep_running(struct thread_data *td)
 776 {
 777         unsigned long long io_done;
 778
 779         if (td->done)
 780                 return 0;
 781         if (td->o.time_based)
 782                 return 1;
 783         if (td->o.loops) {
 784                 td->o.loops--;
 785                 return 1;
 786         }
 787
 788         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 789                         + td->io_skip_bytes;
 790         if (io_done < td->o.size)
 791                 return 1;
 792
 793         return 0;
 794 }
 795
 796 static int clear_io_state(struct thread_data *td)
 797 {
 798         struct fio_file *f;
 799         unsigned int i;
 800         int ret;
 801
 802         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 803         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 804         td->zone_bytes = 0;
 805         td->rate_bytes = 0;
 806         td->rate_blocks = 0;
 807         td->rw_end_set[0] = td->rw_end_set[1] = 0;
 808
 809         td->last_was_sync = 0;
 810
 811         /*
 812          * reset file done count if we are to start over
 813          */
 814         if (td->o.time_based || td->o.loops)
 815                 td->nr_done_files = 0;
 816
 817         close_files(td);
 818
 819         ret = 0;
 820         for_each_file(td, f, i) {
 821                 f->flags &= ~FIO_FILE_DONE;
 822                 ret = td_io_open_file(td, f);
 823                 if (ret)
 824                         break;
 825         }
 826
 827         return ret;
 828 }
 829
 830 /*
 831  * Entry point for the thread based jobs. The process based jobs end up
 832  * here as well, after a little setup.
 833  */
 834 static void *thread_main(void *data)
 835 {
 836         unsigned long long runtime[2], elapsed;
 837         struct thread_data *td = data;
 838         int clear_state;
 839
 840         if (!td->o.use_thread)
 841                 setsid();
 842
 843         td->pid = getpid();
 844
 845         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 846
 847         INIT_LIST_HEAD(&td->io_u_freelist);
 848         INIT_LIST_HEAD(&td->io_u_busylist);
 849         INIT_LIST_HEAD(&td->io_u_requeues);
 850         INIT_LIST_HEAD(&td->io_log_list);
 851         INIT_LIST_HEAD(&td->io_hist_list);
 852         td->io_hist_tree = RB_ROOT;
 853
 854         td_set_runstate(td, TD_INITIALIZED);
 855         fio_mutex_up(startup_mutex);
 856         fio_mutex_down(td->mutex);
 857
 858         /*
 859          * the ->mutex mutex is now no longer used, close it to avoid
 860          * eating a file descriptor
 861          */
 862         fio_mutex_remove(td->mutex);
 863
 864         /*
 865          * May alter parameters that init_io_u() will use, so we need to
 866          * do this first.
 867          */
 868         if (init_iolog(td))
 869                 goto err;
 870
 871         if (init_io_u(td))
 872                 goto err;
 873
 874         if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
 875                 td_verror(td, errno, "cpu_set_affinity");
 876                 goto err;
 877         }
 878
 879         if (td->ioprio_set) {
 880                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 881                         td_verror(td, errno, "ioprio_set");
 882                         goto err;
 883                 }
 884         }
 885
 886         if (nice(td->o.nice) == -1) {
 887                 td_verror(td, errno, "nice");
 888                 goto err;
 889         }
 890
 891         if (td->o.ioscheduler && switch_ioscheduler(td))
 892                 goto err;
 893
 894         if (!td->o.create_serialize && setup_files(td))
 895                 goto err;
 896
 897         if (td_io_init(td))
 898                 goto err;
 899
 900         if (open_files(td))
 901                 goto err;
 902
 903         if (init_random_map(td))
 904                 goto err;
 905
 906         if (td->o.exec_prerun) {
 907                 if (system(td->o.exec_prerun) < 0)
 908                         goto err;
 909         }
 910
 911         fio_gettime(&td->epoch, NULL);
 912         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 913         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 914
 915         runtime[0] = runtime[1] = 0;
 916         clear_state = 0;
 917         while (keep_running(td)) {
 918                 fio_gettime(&td->start, NULL);
 919                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 920
 921                 if (td->o.ratemin)
 922                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
 923                                                         sizeof(td->lastrate));
 924
 925                 if (clear_state && clear_io_state(td))
 926                         break;
 927
 928                 prune_io_piece_log(td);
 929
 930                 do_io(td);
 931
 932                 clear_state = 1;
 933
 934                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
 935                         if (td->rw_end_set[DDIR_READ])
 936                                 elapsed = utime_since(&td->start,
 937                                                       &td->rw_end[DDIR_READ]);
 938                         else
 939                                 elapsed = utime_since_now(&td->start);
 940
 941                         runtime[DDIR_READ] += elapsed;
 942                 }
 943                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
 944                         if (td->rw_end_set[DDIR_WRITE])
 945                                 elapsed = utime_since(&td->start,
 946                                                       &td->rw_end[DDIR_WRITE]);
 947                         else
 948                                 elapsed = utime_since_now(&td->start);
 949
 950                         runtime[DDIR_WRITE] += elapsed;
 951                 }
 952
 953                 if (td->error || td->terminate)
 954                         break;
 955
 956                 if (!td->o.do_verify ||
 957                     td->o.verify == VERIFY_NONE ||
 958                     (td->io_ops->flags & FIO_UNIDIR))
 959                         continue;
 960
 961                 if (clear_io_state(td))
 962                         break;
 963
 964                 fio_gettime(&td->start, NULL);
 965
 966                 do_verify(td);
 967
 968                 runtime[DDIR_READ] += utime_since_now(&td->start);
 969
 970                 if (td->error || td->terminate)
 971                         break;
 972         }
 973
 974         update_rusage_stat(td);
 975         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
 976         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
 977         td->ts.total_run_time = mtime_since_now(&td->epoch);
 978         td->ts.io_bytes[0] = td->io_bytes[0];
 979         td->ts.io_bytes[1] = td->io_bytes[1];
 980
 981         if (td->ts.bw_log)
 982                 finish_log(td, td->ts.bw_log, "bw");
 983         if (td->ts.slat_log)
 984                 finish_log(td, td->ts.slat_log, "slat");
 985         if (td->ts.clat_log)
 986                 finish_log(td, td->ts.clat_log, "clat");
 987         if (td->o.exec_postrun) {
 988                 if (system(td->o.exec_postrun) < 0)
 989                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
 990         }
 991
 992         if (exitall_on_terminate)
 993                 terminate_threads(td->groupid);
 994
 995 err:
 996         if (td->error)
 997                 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
 998                                                         td->verror);
 999         close_and_free_files(td);
1000         close_ioengine(td);
1001         cleanup_io_u(td);
1002
1003         /*
1004          * do this very late, it will log file closing as well
1005          */
1006         if (td->o.write_iolog_file)
1007                 write_iolog_close(td);
1008
1009         options_mem_free(td);
1010         td_set_runstate(td, TD_EXITED);
1011         return (void *) (unsigned long) td->error;
1012 }
1013
1014 /*
1015  * We cannot pass the td data into a forked process, so attach the td and
1016  * pass it to the thread worker.
1017  */
1018 static int fork_main(int shmid, int offset)
1019 {
1020         struct thread_data *td;
1021         void *data, *ret;
1022
1023         data = shmat(shmid, NULL, 0);
1024         if (data == (void *) -1) {
1025                 int __err = errno;
1026
1027                 perror("shmat");
1028                 return __err;
1029         }
1030
1031         td = data + offset * sizeof(struct thread_data);
1032         ret = thread_main(td);
1033         shmdt(data);
1034         return (int) (unsigned long) ret;
1035 }
1036
1037 /*
1038  * Run over the job map and reap the threads that have exited, if any.
1039  */
1040 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1041 {
1042         struct thread_data *td;
1043         int i, cputhreads, realthreads, pending, status, ret;
1044
1045         /*
1046          * reap exited threads (TD_EXITED -> TD_REAPED)
1047          */
1048         realthreads = pending = cputhreads = 0;
1049         for_each_td(td, i) {
1050                 int flags = 0;
1051
1052                 /*
1053                  * ->io_ops is NULL for a thread that has closed its
1054                  * io engine
1055                  */
1056                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1057                         cputhreads++;
1058                 else
1059                         realthreads++;
1060
1061                 if (!td->pid) {
1062                         pending++;
1063                         continue;
1064                 }
1065                 if (td->runstate == TD_REAPED)
1066                         continue;
1067                 if (td->o.use_thread) {
1068                         if (td->runstate == TD_EXITED) {
1069                                 td_set_runstate(td, TD_REAPED);
1070                                 goto reaped;
1071                         }
1072                         continue;
1073                 }
1074
1075                 flags = WNOHANG;
1076                 if (td->runstate == TD_EXITED)
1077                         flags = 0;
1078
1079                 /*
1080                  * check if someone quit or got killed in an unusual way
1081                  */
1082                 ret = waitpid(td->pid, &status, flags);
1083                 if (ret < 0) {
1084                         if (errno == ECHILD) {
1085                                 log_err("fio: pid=%d disappeared %d\n",
1086                                                 (int) td->pid, td->runstate);
1087                                 td_set_runstate(td, TD_REAPED);
1088                                 goto reaped;
1089                         }
1090                         perror("waitpid");
1091                 } else if (ret == td->pid) {
1092                         if (WIFSIGNALED(status)) {
1093                                 int sig = WTERMSIG(status);
1094
1095                                 if (sig != SIGQUIT)
1096                                         log_err("fio: pid=%d, got signal=%d\n",
1097                                                         (int) td->pid, sig);
1098                                 td_set_runstate(td, TD_REAPED);
1099                                 goto reaped;
1100                         }
1101                         if (WIFEXITED(status)) {
1102                                 if (WEXITSTATUS(status) && !td->error)
1103                                         td->error = WEXITSTATUS(status);
1104
1105                                 td_set_runstate(td, TD_REAPED);
1106                                 goto reaped;
1107                         }
1108                 }
1109
1110                 /*
1111                  * thread is not dead, continue
1112                  */
1113                 pending++;
1114                 continue;
1115 reaped:
1116                 (*nr_running)--;
1117                 (*m_rate) -= td->o.ratemin;
1118                 (*t_rate) -= td->o.rate;
1119                 if (!td->pid)
1120                         pending--;
1121
1122                 if (td->error)
1123                         exit_value++;
1124
1125                 done_secs += mtime_since_now(&td->epoch) / 1000;
1126         }
1127
1128         if (*nr_running == cputhreads && !pending && realthreads)
1129                 terminate_threads(TERMINATE_ALL);
1130 }
1131
1132 /*
1133  * Main function for kicking off and reaping jobs, as needed.
1134  */
1135 static void run_threads(void)
1136 {
1137         struct thread_data *td;
1138         unsigned long spent;
1139         int i, todo, nr_running, m_rate, t_rate, nr_started;
1140
1141         if (fio_pin_memory())
1142                 return;
1143
1144         if (!terse_output) {
1145                 printf("Starting ");
1146                 if (nr_thread)
1147                         printf("%d thread%s", nr_thread,
1148                                                 nr_thread > 1 ? "s" : "");
1149                 if (nr_process) {
1150                         if (nr_thread)
1151                                 printf(" and ");
1152                         printf("%d process%s", nr_process,
1153                                                 nr_process > 1 ? "es" : "");
1154                 }
1155                 printf("\n");
1156                 fflush(stdout);
1157         }
1158
1159         set_sig_handlers(sig_handler);
1160
1161         todo = thread_number;
1162         nr_running = 0;
1163         nr_started = 0;
1164         m_rate = t_rate = 0;
1165
1166         for_each_td(td, i) {
1167                 print_status_init(td->thread_number - 1);
1168
1169                 if (!td->o.create_serialize) {
1170                         init_disk_util(td);
1171                         continue;
1172                 }
1173
1174                 /*
1175                  * do file setup here so it happens sequentially,
1176                  * we don't want X number of threads getting their
1177                  * client data interspersed on disk
1178                  */
1179                 if (setup_files(td)) {
1180                         exit_value++;
1181                         if (td->error)
1182                                 log_err("fio: pid=%d, err=%d/%s\n",
1183                                         (int) td->pid, td->error, td->verror);
1184                         td_set_runstate(td, TD_REAPED);
1185                         todo--;
1186                 } else {
1187                         struct fio_file *f;
1188                         unsigned int i;
1189
1190                         /*
1191                          * for sharing to work, each job must always open
1192                          * its own files. so close them, if we opened them
1193                          * for creation
1194                          */
1195                         for_each_file(td, f, i)
1196                                 td_io_close_file(td, f);
1197                 }
1198
1199                 init_disk_util(td);
1200         }
1201
1202         set_genesis_time();
1203
1204         while (todo) {
1205                 struct thread_data *map[MAX_JOBS];
1206                 struct timeval this_start;
1207                 int this_jobs = 0, left;
1208
1209                 /*
1210                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1211                  */
1212                 for_each_td(td, i) {
1213                         if (td->runstate != TD_NOT_CREATED)
1214                                 continue;
1215
1216                         /*
1217                          * never got a chance to start, killed by other
1218                          * thread for some reason
1219                          */
1220                         if (td->terminate) {
1221                                 todo--;
1222                                 continue;
1223                         }
1224
1225                         if (td->o.start_delay) {
1226                                 spent = mtime_since_genesis();
1227
1228                                 if (td->o.start_delay * 1000 > spent)
1229                                         continue;
1230                         }
1231
1232                         if (td->o.stonewall && (nr_started || nr_running)) {
1233                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1234                                                         td->o.name);
1235                                 break;
1236                         }
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("pthread_create");
1251                                         nr_started--;
1252                                         break;
1253                                 }
1254                                 if (pthread_detach(td->thread) < 0)
1255                                         perror("pthread_detach");
1256                         } else {
1257                                 pid_t pid;
1258                                 dprint(FD_PROCESS, "will fork\n");
1259                                 pid = fork();
1260                                 if (!pid) {
1261                                         int ret = fork_main(shm_id, i);
1262
1263                                         _exit(ret);
1264                                 } else if (i == fio_debug_jobno)
1265                                         *fio_debug_jobp = pid;
1266                         }
1267                         fio_mutex_down(startup_mutex);
1268                 }
1269
1270                 /*
1271                  * Wait for the started threads to transition to
1272                  * TD_INITIALIZED.
1273                  */
1274                 fio_gettime(&this_start, NULL);
1275                 left = this_jobs;
1276                 while (left && !fio_abort) {
1277                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1278                                 break;
1279
1280                         usleep(100000);
1281
1282                         for (i = 0; i < this_jobs; i++) {
1283                                 td = map[i];
1284                                 if (!td)
1285                                         continue;
1286                                 if (td->runstate == TD_INITIALIZED) {
1287                                         map[i] = NULL;
1288                                         left--;
1289                                 } else if (td->runstate >= TD_EXITED) {
1290                                         map[i] = NULL;
1291                                         left--;
1292                                         todo--;
1293                                         nr_running++; /* work-around... */
1294                                 }
1295                         }
1296                 }
1297
1298                 if (left) {
1299                         log_err("fio: %d jobs failed to start\n", left);
1300                         for (i = 0; i < this_jobs; i++) {
1301                                 td = map[i];
1302                                 if (!td)
1303                                         continue;
1304                                 kill(td->pid, SIGTERM);
1305                         }
1306                         break;
1307                 }
1308
1309                 /*
1310                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1311                  */
1312                 for_each_td(td, i) {
1313                         if (td->runstate != TD_INITIALIZED)
1314                                 continue;
1315
1316                         td_set_runstate(td, TD_RUNNING);
1317                         nr_running++;
1318                         nr_started--;
1319                         m_rate += td->o.ratemin;
1320                         t_rate += td->o.rate;
1321                         todo--;
1322                         fio_mutex_up(td->mutex);
1323                 }
1324
1325                 reap_threads(&nr_running, &t_rate, &m_rate);
1326
1327                 if (todo)
1328                         usleep(100000);
1329         }
1330
1331         while (nr_running) {
1332                 reap_threads(&nr_running, &t_rate, &m_rate);
1333                 usleep(10000);
1334         }
1335
1336         update_io_ticks();
1337         fio_unpin_memory();
1338 }
1339
1340 int main(int argc, char *argv[])
1341 {
1342         long ps;
1343
1344         sinit();
1345
1346         /*
1347          * We need locale for number printing, if it isn't set then just
1348          * go with the US format.
1349          */
1350         if (!getenv("LC_NUMERIC"))
1351                 setlocale(LC_NUMERIC, "en_US");
1352
1353         if (parse_options(argc, argv))
1354                 return 1;
1355
1356         if (!thread_number)
1357                 return 0;
1358
1359         ps = sysconf(_SC_PAGESIZE);
1360         if (ps < 0) {
1361                 log_err("Failed to get page size\n");
1362                 return 1;
1363         }
1364
1365         page_size = ps;
1366         page_mask = ps - 1;
1367
1368         if (write_bw_log) {
1369                 setup_log(&agg_io_log[DDIR_READ]);
1370                 setup_log(&agg_io_log[DDIR_WRITE]);
1371         }
1372
1373         startup_mutex = fio_mutex_init(0);
1374
1375         set_genesis_time();
1376
1377         status_timer_arm();
1378
1379         run_threads();
1380
1381         if (!fio_abort) {
1382                 show_run_stats();
1383                 if (write_bw_log) {
1384                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1385                         __finish_log(agg_io_log[DDIR_WRITE],
1386                                         "agg-write_bw.log");
1387                 }
1388         }
1389
1390         fio_mutex_remove(startup_mutex);
1391         return exit_value;
1392 }