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 static void sig_alrm(int sig)
 111 {
 112         if (threads) {
 113                 update_io_ticks();
 114                 print_thread_status();
 115                 status_timer_arm();
 116         }
 117 }
 118
 119 static void sig_int(int sig)
 120 {
 121         if (threads) {
 122                 printf("\nfio: terminating on signal %d\n", sig);
 123                 fflush(stdout);
 124                 terminate_threads(TERMINATE_ALL);
 125         }
 126 }
 127
 128 /*
 129  * We need to rearm on BSD/solaris. Switch this to sigaction in the future...
 130  */
 131 static void set_sig_handlers(void)
 132 {
 133         struct sigaction act;
 134
 135         memset(&act, 0, sizeof(act));
 136         act.sa_handler = sig_alrm;
 137         act.sa_flags = SA_RESTART;
 138         sigaction(SIGALRM, &act, NULL);
 139
 140         memset(&act, 0, sizeof(act));
 141         act.sa_handler = sig_int;
 142         act.sa_flags = SA_RESTART;
 143         sigaction(SIGINT, &act, NULL);
 144 }
 145
 146 /*
 147  * Check if we are above the minimum rate given.
 148  */
 149 static int check_min_rate(struct thread_data *td, struct timeval *now)
 150 {
 151         unsigned long long bytes = 0;
 152         unsigned long iops = 0;
 153         unsigned long spent;
 154         unsigned long rate;
 155
 156         /*
 157          * No minimum rate set, always ok
 158          */
 159         if (!td->o.ratemin && !td->o.rate_iops_min)
 160                 return 0;
 161
 162         /*
 163          * allow a 2 second settle period in the beginning
 164          */
 165         if (mtime_since(&td->start, now) < 2000)
 166                 return 0;
 167
 168         if (td_read(td)) {
 169                 iops += td->io_blocks[DDIR_READ];
 170                 bytes += td->this_io_bytes[DDIR_READ];
 171         }
 172         if (td_write(td)) {
 173                 iops += td->io_blocks[DDIR_WRITE];
 174                 bytes += td->this_io_bytes[DDIR_WRITE];
 175         }
 176
 177         /*
 178          * if rate blocks is set, sample is running
 179          */
 180         if (td->rate_bytes || td->rate_blocks) {
 181                 spent = mtime_since(&td->lastrate, now);
 182                 if (spent < td->o.ratecycle)
 183                         return 0;
 184
 185                 if (td->o.rate) {
 186                         /*
 187                          * check bandwidth specified rate
 188                          */
 189                         if (bytes < td->rate_bytes) {
 190                                 log_err("%s: min rate %u not met\n", td->o.name,
 191                                                                 td->o.ratemin);
 192                                 return 1;
 193                         } else {
 194                                 rate = (bytes - td->rate_bytes) / spent;
 195                                 if (rate < td->o.ratemin ||
 196                                     bytes < td->rate_bytes) {
 197                                         log_err("%s: min rate %u not met, got"
 198                                                 " %luKiB/sec\n", td->o.name,
 199                                                         td->o.ratemin, rate);
 200                                         return 1;
 201                                 }
 202                         }
 203                 } else {
 204                         /*
 205                          * checks iops specified rate
 206                          */
 207                         if (iops < td->o.rate_iops) {
 208                                 log_err("%s: min iops rate %u not met\n",
 209                                                 td->o.name, td->o.rate_iops);
 210                                 return 1;
 211                         } else {
 212                                 rate = (iops - td->rate_blocks) / spent;
 213                                 if (rate < td->o.rate_iops_min ||
 214                                     iops < td->rate_blocks) {
 215                                         log_err("%s: min iops rate %u not met,"
 216                                                 " got %lu\n", td->o.name,
 217                                                         td->o.rate_iops_min,
 218                                                         rate);
 219                                 }
 220                         }
 221                 }
 222         }
 223
 224         td->rate_bytes = bytes;
 225         td->rate_blocks = iops;
 226         memcpy(&td->lastrate, now, sizeof(*now));
 227         return 0;
 228 }
 229
 230 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 231 {
 232         if (!td->o.timeout)
 233                 return 0;
 234         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 235                 return 1;
 236
 237         return 0;
 238 }
 239
 240 /*
 241  * When job exits, we can cancel the in-flight IO if we are using async
 242  * io. Attempt to do so.
 243  */
 244 static void cleanup_pending_aio(struct thread_data *td)
 245 {
 246         struct list_head *entry, *n;
 247         struct io_u *io_u;
 248         int r;
 249
 250         /*
 251          * get immediately available events, if any
 252          */
 253         r = io_u_queued_complete(td, 0);
 254         if (r < 0)
 255                 return;
 256
 257         /*
 258          * now cancel remaining active events
 259          */
 260         if (td->io_ops->cancel) {
 261                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 262                         io_u = list_entry(entry, struct io_u, list);
 263
 264                         /*
 265                          * if the io_u isn't in flight, then that generally
 266                          * means someone leaked an io_u. complain but fix
 267                          * it up, so we don't stall here.
 268                          */
 269                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 270                                 log_err("fio: non-busy IO on busy list\n");
 271                                 put_io_u(td, io_u);
 272                         } else {
 273                                 r = td->io_ops->cancel(td, io_u);
 274                                 if (!r)
 275                                         put_io_u(td, io_u);
 276                         }
 277                 }
 278         }
 279
 280         if (td->cur_depth)
 281                 r = io_u_queued_complete(td, td->cur_depth);
 282 }
 283
 284 /*
 285  * Helper to handle the final sync of a file. Works just like the normal
 286  * io path, just does everything sync.
 287  */
 288 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 289 {
 290         struct io_u *io_u = __get_io_u(td);
 291         int ret;
 292
 293         if (!io_u)
 294                 return 1;
 295
 296         io_u->ddir = DDIR_SYNC;
 297         io_u->file = f;
 298
 299         if (td_io_prep(td, io_u)) {
 300                 put_io_u(td, io_u);
 301                 return 1;
 302         }
 303
 304 requeue:
 305         ret = td_io_queue(td, io_u);
 306         if (ret < 0) {
 307                 td_verror(td, io_u->error, "td_io_queue");
 308                 put_io_u(td, io_u);
 309                 return 1;
 310         } else if (ret == FIO_Q_QUEUED) {
 311                 if (io_u_queued_complete(td, 1) < 0)
 312                         return 1;
 313         } else if (ret == FIO_Q_COMPLETED) {
 314                 if (io_u->error) {
 315                         td_verror(td, io_u->error, "td_io_queue");
 316                         return 1;
 317                 }
 318
 319                 if (io_u_sync_complete(td, io_u) < 0)
 320                         return 1;
 321         } else if (ret == FIO_Q_BUSY) {
 322                 if (td_io_commit(td))
 323                         return 1;
 324                 goto requeue;
 325         }
 326
 327         return 0;
 328 }
 329
 330 /*
 331  * The main verify engine. Runs over the writes we previously submitted,
 332  * reads the blocks back in, and checks the crc/md5 of the data.
 333  */
 334 static void do_verify(struct thread_data *td)
 335 {
 336         struct fio_file *f;
 337         struct io_u *io_u;
 338         int ret, min_events;
 339         unsigned int i;
 340
 341         /*
 342          * sync io first and invalidate cache, to make sure we really
 343          * read from disk.
 344          */
 345         for_each_file(td, f, i) {
 346                 if (!(f->flags & FIO_FILE_OPEN))
 347                         continue;
 348                 if (fio_io_sync(td, f))
 349                         break;
 350                 if (file_invalidate_cache(td, f))
 351                         break;
 352         }
 353
 354         if (td->error)
 355                 return;
 356
 357         td_set_runstate(td, TD_VERIFYING);
 358
 359         io_u = NULL;
 360         while (!td->terminate) {
 361                 int ret2;
 362
 363                 io_u = __get_io_u(td);
 364                 if (!io_u)
 365                         break;
 366
 367                 if (runtime_exceeded(td, &io_u->start_time)) {
 368                         put_io_u(td, io_u);
 369                         td->terminate = 1;
 370                         break;
 371                 }
 372
 373                 if (get_next_verify(td, io_u)) {
 374                         put_io_u(td, io_u);
 375                         break;
 376                 }
 377
 378                 if (td_io_prep(td, io_u)) {
 379                         put_io_u(td, io_u);
 380                         break;
 381                 }
 382
 383                 io_u->end_io = verify_io_u;
 384
 385                 ret = td_io_queue(td, io_u);
 386                 switch (ret) {
 387                 case FIO_Q_COMPLETED:
 388                         if (io_u->error)
 389                                 ret = -io_u->error;
 390                         else if (io_u->resid) {
 391                                 int bytes = io_u->xfer_buflen - io_u->resid;
 392                                 struct fio_file *f = io_u->file;
 393
 394                                 /*
 395                                  * zero read, fail
 396                                  */
 397                                 if (!bytes) {
 398                                         td_verror(td, ENODATA, "full resid");
 399                                         put_io_u(td, io_u);
 400                                         break;
 401                                 }
 402
 403                                 io_u->xfer_buflen = io_u->resid;
 404                                 io_u->xfer_buf += bytes;
 405                                 io_u->offset += bytes;
 406
 407                                 td->ts.short_io_u[io_u->ddir]++;
 408
 409                                 if (io_u->offset == f->real_file_size)
 410                                         goto sync_done;
 411
 412                                 requeue_io_u(td, &io_u);
 413                         } else {
 414 sync_done:
 415                                 ret = io_u_sync_complete(td, io_u);
 416                                 if (ret < 0)
 417                                         break;
 418                         }
 419                         continue;
 420                 case FIO_Q_QUEUED:
 421                         break;
 422                 case FIO_Q_BUSY:
 423                         requeue_io_u(td, &io_u);
 424                         ret2 = td_io_commit(td);
 425                         if (ret2 < 0)
 426                                 ret = ret2;
 427                         break;
 428                 default:
 429                         assert(ret < 0);
 430                         td_verror(td, -ret, "td_io_queue");
 431                         break;
 432                 }
 433
 434                 if (ret < 0 || td->error)
 435                         break;
 436
 437                 /*
 438                  * if we can queue more, do so. but check if there are
 439                  * completed io_u's first.
 440                  */
 441                 min_events = 0;
 442                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 443                         if (td->cur_depth >= td->o.iodepth_low)
 444                                 min_events = td->cur_depth - td->o.iodepth_low;
 445                         if (!min_events)
 446                                 min_events = 1;
 447                 }
 448
 449                 /*
 450                  * Reap required number of io units, if any, and do the
 451                  * verification on them through the callback handler
 452                  */
 453                 if (io_u_queued_complete(td, min_events) < 0)
 454                         break;
 455         }
 456
 457         if (!td->error) {
 458                 min_events = td->cur_depth;
 459
 460                 if (min_events)
 461                         ret = io_u_queued_complete(td, min_events);
 462         } else
 463                 cleanup_pending_aio(td);
 464
 465         td_set_runstate(td, TD_RUNNING);
 466 }
 467
 468 /*
 469  * Main IO worker function. It retrieves io_u's to process and queues
 470  * and reaps them, checking for rate and errors along the way.
 471  */
 472 static void do_io(struct thread_data *td)
 473 {
 474         struct timeval s;
 475         unsigned long usec;
 476         unsigned int i;
 477         int ret = 0;
 478
 479         td_set_runstate(td, TD_RUNNING);
 480
 481         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 482                 struct timeval comp_time;
 483                 long bytes_done = 0;
 484                 int min_evts = 0;
 485                 struct io_u *io_u;
 486                 int ret2;
 487
 488                 if (td->terminate)
 489                         break;
 490
 491                 io_u = get_io_u(td);
 492                 if (!io_u)
 493                         break;
 494
 495                 memcpy(&s, &io_u->start_time, sizeof(s));
 496
 497                 if (runtime_exceeded(td, &s)) {
 498                         put_io_u(td, io_u);
 499                         td->terminate = 1;
 500                         break;
 501                 }
 502
 503                 /*
 504                  * Add verification end_io handler, if asked to verify
 505                  * a previously written file.
 506                  */
 507                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
 508                         io_u->end_io = verify_io_u;
 509                         td_set_runstate(td, TD_VERIFYING);
 510                 } else
 511                         td_set_runstate(td, TD_RUNNING);
 512
 513                 ret = td_io_queue(td, io_u);
 514                 switch (ret) {
 515                 case FIO_Q_COMPLETED:
 516                         if (io_u->error)
 517                                 ret = -io_u->error;
 518                         else if (io_u->resid) {
 519                                 int bytes = io_u->xfer_buflen - io_u->resid;
 520                                 struct fio_file *f = io_u->file;
 521
 522                                 /*
 523                                  * zero read, fail
 524                                  */
 525                                 if (!bytes) {
 526                                         td_verror(td, ENODATA, "full resid");
 527                                         put_io_u(td, io_u);
 528                                         break;
 529                                 }
 530
 531                                 io_u->xfer_buflen = io_u->resid;
 532                                 io_u->xfer_buf += bytes;
 533                                 io_u->offset += bytes;
 534
 535                                 td->ts.short_io_u[io_u->ddir]++;
 536
 537                                 if (io_u->offset == f->real_file_size)
 538                                         goto sync_done;
 539
 540                                 requeue_io_u(td, &io_u);
 541                         } else {
 542 sync_done:
 543                                 fio_gettime(&comp_time, NULL);
 544                                 bytes_done = io_u_sync_complete(td, io_u);
 545                                 if (bytes_done < 0)
 546                                         ret = bytes_done;
 547                         }
 548                         break;
 549                 case FIO_Q_QUEUED:
 550                         /*
 551                          * if the engine doesn't have a commit hook,
 552                          * the io_u is really queued. if it does have such
 553                          * a hook, it has to call io_u_queued() itself.
 554                          */
 555                         if (td->io_ops->commit == NULL)
 556                                 io_u_queued(td, io_u);
 557                         break;
 558                 case FIO_Q_BUSY:
 559                         requeue_io_u(td, &io_u);
 560                         ret2 = td_io_commit(td);
 561                         if (ret2 < 0)
 562                                 ret = ret2;
 563                         break;
 564                 default:
 565                         assert(ret < 0);
 566                         put_io_u(td, io_u);
 567                         break;
 568                 }
 569
 570                 if (ret < 0 || td->error)
 571                         break;
 572
 573                 /*
 574                  * See if we need to complete some commands
 575                  */
 576                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 577                         min_evts = 0;
 578                         if (td->cur_depth >= td->o.iodepth_low)
 579                                 min_evts = td->cur_depth - td->o.iodepth_low;
 580                         if (!min_evts)
 581                                 min_evts = 1;
 582                         fio_gettime(&comp_time, NULL);
 583                         bytes_done = io_u_queued_complete(td, min_evts);
 584                         if (bytes_done < 0)
 585                                 break;
 586                 }
 587
 588                 if (!bytes_done)
 589                         continue;
 590
 591                 /*
 592                  * the rate is batched for now, it should work for batches
 593                  * of completions except the very first one which may look
 594                  * a little bursty
 595                  */
 596                 usec = utime_since(&s, &comp_time);
 597
 598                 rate_throttle(td, usec, bytes_done);
 599
 600                 if (check_min_rate(td, &comp_time)) {
 601                         if (exitall_on_terminate)
 602                                 terminate_threads(td->groupid);
 603                         td_verror(td, ENODATA, "check_min_rate");
 604                         break;
 605                 }
 606
 607                 if (td->o.thinktime) {
 608                         unsigned long long b;
 609
 610                         b = td->io_blocks[0] + td->io_blocks[1];
 611                         if (!(b % td->o.thinktime_blocks)) {
 612                                 int left;
 613
 614                                 if (td->o.thinktime_spin)
 615                                         __usec_sleep(td->o.thinktime_spin);
 616
 617                                 left = td->o.thinktime - td->o.thinktime_spin;
 618                                 if (left)
 619                                         usec_sleep(td, left);
 620                         }
 621                 }
 622         }
 623
 624         if (td->o.fill_device && td->error == ENOSPC) {
 625                 td->error = 0;
 626                 td->terminate = 1;
 627         }
 628         if (!td->error) {
 629                 struct fio_file *f;
 630
 631                 i = td->cur_depth;
 632                 if (i)
 633                         ret = io_u_queued_complete(td, i);
 634
 635                 if (should_fsync(td) && td->o.end_fsync) {
 636                         td_set_runstate(td, TD_FSYNCING);
 637
 638                         for_each_file(td, f, i) {
 639                                 if (!(f->flags & FIO_FILE_OPEN))
 640                                         continue;
 641                                 fio_io_sync(td, f);
 642                         }
 643                 }
 644         } else
 645                 cleanup_pending_aio(td);
 646
 647         /*
 648          * stop job if we failed doing any IO
 649          */
 650         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 651                 td->done = 1;
 652 }
 653
 654 static void cleanup_io_u(struct thread_data *td)
 655 {
 656         struct list_head *entry, *n;
 657         struct io_u *io_u;
 658
 659         list_for_each_safe(entry, n, &td->io_u_freelist) {
 660                 io_u = list_entry(entry, struct io_u, list);
 661
 662                 list_del(&io_u->list);
 663                 free(io_u);
 664         }
 665
 666         free_io_mem(td);
 667 }
 668
 669 static int init_io_u(struct thread_data *td)
 670 {
 671         struct io_u *io_u;
 672         unsigned int max_bs;
 673         int i, max_units;
 674         char *p;
 675
 676         max_units = td->o.iodepth;
 677         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 678         td->orig_buffer_size = (unsigned long long) max_bs
 679                                         * (unsigned long long) max_units;
 680
 681         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 682                 unsigned long bs;
 683
 684                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 685                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 686         }
 687
 688         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 689                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 690                 return 1;
 691         }
 692
 693         if (allocate_io_mem(td))
 694                 return 1;
 695
 696         if (td->o.odirect)
 697                 p = ALIGN(td->orig_buffer);
 698         else
 699                 p = td->orig_buffer;
 700
 701         for (i = 0; i < max_units; i++) {
 702                 if (td->terminate)
 703                         return 1;
 704                 io_u = malloc(sizeof(*io_u));
 705                 memset(io_u, 0, sizeof(*io_u));
 706                 INIT_LIST_HEAD(&io_u->list);
 707
 708                 if (!(td->io_ops->flags & FIO_NOIO)) {
 709                         io_u->buf = p + max_bs * i;
 710
 711                         if (td_write(td) && !td->o.refill_buffers)
 712                                 io_u_fill_buffer(td, io_u, max_bs);
 713                 }
 714
 715                 io_u->index = i;
 716                 io_u->flags = IO_U_F_FREE;
 717                 list_add(&io_u->list, &td->io_u_freelist);
 718         }
 719
 720         io_u_init_timeout();
 721
 722         return 0;
 723 }
 724
 725 static int switch_ioscheduler(struct thread_data *td)
 726 {
 727         char tmp[256], tmp2[128];
 728         FILE *f;
 729         int ret;
 730
 731         if (td->io_ops->flags & FIO_DISKLESSIO)
 732                 return 0;
 733
 734         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 735
 736         f = fopen(tmp, "r+");
 737         if (!f) {
 738                 if (errno == ENOENT) {
 739                         log_err("fio: os or kernel doesn't support IO scheduler"
 740                                 " switching\n");
 741                         return 0;
 742                 }
 743                 td_verror(td, errno, "fopen iosched");
 744                 return 1;
 745         }
 746
 747         /*
 748          * Set io scheduler.
 749          */
 750         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 751         if (ferror(f) || ret != 1) {
 752                 td_verror(td, errno, "fwrite");
 753                 fclose(f);
 754                 return 1;
 755         }
 756
 757         rewind(f);
 758
 759         /*
 760          * Read back and check that the selected scheduler is now the default.
 761          */
 762         ret = fread(tmp, 1, sizeof(tmp), f);
 763         if (ferror(f) || ret < 0) {
 764                 td_verror(td, errno, "fread");
 765                 fclose(f);
 766                 return 1;
 767         }
 768
 769         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 770         if (!strstr(tmp, tmp2)) {
 771                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 772                 td_verror(td, EINVAL, "iosched_switch");
 773                 fclose(f);
 774                 return 1;
 775         }
 776
 777         fclose(f);
 778         return 0;
 779 }
 780
 781 static int keep_running(struct thread_data *td)
 782 {
 783         unsigned long long io_done;
 784
 785         if (td->done)
 786                 return 0;
 787         if (td->o.time_based)
 788                 return 1;
 789         if (td->o.loops) {
 790                 td->o.loops--;
 791                 return 1;
 792         }
 793
 794         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 795                         + td->io_skip_bytes;
 796         if (io_done < td->o.size)
 797                 return 1;
 798
 799         return 0;
 800 }
 801
 802 static int clear_io_state(struct thread_data *td)
 803 {
 804         struct fio_file *f;
 805         unsigned int i;
 806         int ret;
 807
 808         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 809         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 810         td->zone_bytes = 0;
 811         td->rate_bytes = 0;
 812         td->rate_blocks = 0;
 813         td->rw_end_set[0] = td->rw_end_set[1] = 0;
 814
 815         td->last_was_sync = 0;
 816
 817         /*
 818          * reset file done count if we are to start over
 819          */
 820         if (td->o.time_based || td->o.loops)
 821                 td->nr_done_files = 0;
 822
 823         close_files(td);
 824
 825         ret = 0;
 826         for_each_file(td, f, i) {
 827                 f->flags &= ~FIO_FILE_DONE;
 828                 ret = td_io_open_file(td, f);
 829                 if (ret)
 830                         break;
 831         }
 832
 833         return ret;
 834 }
 835
 836 /*
 837  * Entry point for the thread based jobs. The process based jobs end up
 838  * here as well, after a little setup.
 839  */
 840 static void *thread_main(void *data)
 841 {
 842         unsigned long long runtime[2], elapsed;
 843         struct thread_data *td = data;
 844         int clear_state;
 845
 846         if (!td->o.use_thread)
 847                 setsid();
 848
 849         td->pid = getpid();
 850
 851         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 852
 853         INIT_LIST_HEAD(&td->io_u_freelist);
 854         INIT_LIST_HEAD(&td->io_u_busylist);
 855         INIT_LIST_HEAD(&td->io_u_requeues);
 856         INIT_LIST_HEAD(&td->io_log_list);
 857         INIT_LIST_HEAD(&td->io_hist_list);
 858         td->io_hist_tree = RB_ROOT;
 859
 860         td_set_runstate(td, TD_INITIALIZED);
 861         fio_mutex_up(startup_mutex);
 862         fio_mutex_down(td->mutex);
 863
 864         /*
 865          * the ->mutex mutex is now no longer used, close it to avoid
 866          * eating a file descriptor
 867          */
 868         fio_mutex_remove(td->mutex);
 869
 870         /*
 871          * May alter parameters that init_io_u() will use, so we need to
 872          * do this first.
 873          */
 874         if (init_iolog(td))
 875                 goto err;
 876
 877         if (init_io_u(td))
 878                 goto err;
 879
 880         if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
 881                 td_verror(td, errno, "cpu_set_affinity");
 882                 goto err;
 883         }
 884
 885         if (td->ioprio_set) {
 886                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 887                         td_verror(td, errno, "ioprio_set");
 888                         goto err;
 889                 }
 890         }
 891
 892         if (nice(td->o.nice) == -1) {
 893                 td_verror(td, errno, "nice");
 894                 goto err;
 895         }
 896
 897         if (td->o.ioscheduler && switch_ioscheduler(td))
 898                 goto err;
 899
 900         if (!td->o.create_serialize && setup_files(td))
 901                 goto err;
 902
 903         if (td_io_init(td))
 904                 goto err;
 905
 906         if (open_files(td))
 907                 goto err;
 908
 909         if (init_random_map(td))
 910                 goto err;
 911
 912         if (td->o.exec_prerun) {
 913                 if (system(td->o.exec_prerun) < 0)
 914                         goto err;
 915         }
 916
 917         fio_gettime(&td->epoch, NULL);
 918         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 919         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 920
 921         runtime[0] = runtime[1] = 0;
 922         clear_state = 0;
 923         while (keep_running(td)) {
 924                 fio_gettime(&td->start, NULL);
 925                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 926
 927                 if (td->o.ratemin)
 928                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
 929                                                         sizeof(td->lastrate));
 930
 931                 if (clear_state && clear_io_state(td))
 932                         break;
 933
 934                 prune_io_piece_log(td);
 935
 936                 do_io(td);
 937
 938                 clear_state = 1;
 939
 940                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
 941                         if (td->rw_end_set[DDIR_READ])
 942                                 elapsed = utime_since(&td->start,
 943                                                       &td->rw_end[DDIR_READ]);
 944                         else
 945                                 elapsed = utime_since_now(&td->start);
 946
 947                         runtime[DDIR_READ] += elapsed;
 948                 }
 949                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
 950                         if (td->rw_end_set[DDIR_WRITE])
 951                                 elapsed = utime_since(&td->start,
 952                                                       &td->rw_end[DDIR_WRITE]);
 953                         else
 954                                 elapsed = utime_since_now(&td->start);
 955
 956                         runtime[DDIR_WRITE] += elapsed;
 957                 }
 958
 959                 if (td->error || td->terminate)
 960                         break;
 961
 962                 if (!td->o.do_verify ||
 963                     td->o.verify == VERIFY_NONE ||
 964                     (td->io_ops->flags & FIO_UNIDIR))
 965                         continue;
 966
 967                 if (clear_io_state(td))
 968                         break;
 969
 970                 fio_gettime(&td->start, NULL);
 971
 972                 do_verify(td);
 973
 974                 runtime[DDIR_READ] += utime_since_now(&td->start);
 975
 976                 if (td->error || td->terminate)
 977                         break;
 978         }
 979
 980         update_rusage_stat(td);
 981         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
 982         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
 983         td->ts.total_run_time = mtime_since_now(&td->epoch);
 984         td->ts.io_bytes[0] = td->io_bytes[0];
 985         td->ts.io_bytes[1] = td->io_bytes[1];
 986
 987         if (td->ts.bw_log)
 988                 finish_log(td, td->ts.bw_log, "bw");
 989         if (td->ts.slat_log)
 990                 finish_log(td, td->ts.slat_log, "slat");
 991         if (td->ts.clat_log)
 992                 finish_log(td, td->ts.clat_log, "clat");
 993         if (td->o.exec_postrun) {
 994                 if (system(td->o.exec_postrun) < 0)
 995                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
 996         }
 997
 998         if (exitall_on_terminate)
 999                 terminate_threads(td->groupid);
1000
1001 err:
1002         if (td->error)
1003                 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1004                                                         td->verror);
1005         close_and_free_files(td);
1006         close_ioengine(td);
1007         cleanup_io_u(td);
1008
1009         /*
1010          * do this very late, it will log file closing as well
1011          */
1012         if (td->o.write_iolog_file)
1013                 write_iolog_close(td);
1014
1015         options_mem_free(td);
1016         td_set_runstate(td, TD_EXITED);
1017         return (void *) (unsigned long) td->error;
1018 }
1019
1020 /*
1021  * We cannot pass the td data into a forked process, so attach the td and
1022  * pass it to the thread worker.
1023  */
1024 static int fork_main(int shmid, int offset)
1025 {
1026         struct thread_data *td;
1027         void *data, *ret;
1028
1029         data = shmat(shmid, NULL, 0);
1030         if (data == (void *) -1) {
1031                 int __err = errno;
1032
1033                 perror("shmat");
1034                 return __err;
1035         }
1036
1037         td = data + offset * sizeof(struct thread_data);
1038         ret = thread_main(td);
1039         shmdt(data);
1040         return (int) (unsigned long) ret;
1041 }
1042
1043 /*
1044  * Run over the job map and reap the threads that have exited, if any.
1045  */
1046 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1047 {
1048         struct thread_data *td;
1049         int i, cputhreads, realthreads, pending, status, ret;
1050
1051         /*
1052          * reap exited threads (TD_EXITED -> TD_REAPED)
1053          */
1054         realthreads = pending = cputhreads = 0;
1055         for_each_td(td, i) {
1056                 int flags = 0;
1057
1058                 /*
1059                  * ->io_ops is NULL for a thread that has closed its
1060                  * io engine
1061                  */
1062                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1063                         cputhreads++;
1064                 else
1065                         realthreads++;
1066
1067                 if (!td->pid) {
1068                         pending++;
1069                         continue;
1070                 }
1071                 if (td->runstate == TD_REAPED)
1072                         continue;
1073                 if (td->o.use_thread) {
1074                         if (td->runstate == TD_EXITED) {
1075                                 td_set_runstate(td, TD_REAPED);
1076                                 goto reaped;
1077                         }
1078                         continue;
1079                 }
1080
1081                 flags = WNOHANG;
1082                 if (td->runstate == TD_EXITED)
1083                         flags = 0;
1084
1085                 /*
1086                  * check if someone quit or got killed in an unusual way
1087                  */
1088                 ret = waitpid(td->pid, &status, flags);
1089                 if (ret < 0) {
1090                         if (errno == ECHILD) {
1091                                 log_err("fio: pid=%d disappeared %d\n",
1092                                                 (int) td->pid, td->runstate);
1093                                 td_set_runstate(td, TD_REAPED);
1094                                 goto reaped;
1095                         }
1096                         perror("waitpid");
1097                 } else if (ret == td->pid) {
1098                         if (WIFSIGNALED(status)) {
1099                                 int sig = WTERMSIG(status);
1100
1101                                 if (sig != SIGQUIT)
1102                                         log_err("fio: pid=%d, got signal=%d\n",
1103                                                         (int) td->pid, sig);
1104                                 td_set_runstate(td, TD_REAPED);
1105                                 goto reaped;
1106                         }
1107                         if (WIFEXITED(status)) {
1108                                 if (WEXITSTATUS(status) && !td->error)
1109                                         td->error = WEXITSTATUS(status);
1110
1111                                 td_set_runstate(td, TD_REAPED);
1112                                 goto reaped;
1113                         }
1114                 }
1115
1116                 /*
1117                  * thread is not dead, continue
1118                  */
1119                 pending++;
1120                 continue;
1121 reaped:
1122                 (*nr_running)--;
1123                 (*m_rate) -= td->o.ratemin;
1124                 (*t_rate) -= td->o.rate;
1125                 if (!td->pid)
1126                         pending--;
1127
1128                 if (td->error)
1129                         exit_value++;
1130
1131                 done_secs += mtime_since_now(&td->epoch) / 1000;
1132         }
1133
1134         if (*nr_running == cputhreads && !pending && realthreads)
1135                 terminate_threads(TERMINATE_ALL);
1136 }
1137
1138 /*
1139  * Main function for kicking off and reaping jobs, as needed.
1140  */
1141 static void run_threads(void)
1142 {
1143         struct thread_data *td;
1144         unsigned long spent;
1145         int i, todo, nr_running, m_rate, t_rate, nr_started;
1146
1147         if (fio_pin_memory())
1148                 return;
1149
1150         if (!terse_output) {
1151                 printf("Starting ");
1152                 if (nr_thread)
1153                         printf("%d thread%s", nr_thread,
1154                                                 nr_thread > 1 ? "s" : "");
1155                 if (nr_process) {
1156                         if (nr_thread)
1157                                 printf(" and ");
1158                         printf("%d process%s", nr_process,
1159                                                 nr_process > 1 ? "es" : "");
1160                 }
1161                 printf("\n");
1162                 fflush(stdout);
1163         }
1164
1165         set_sig_handlers();
1166
1167         todo = thread_number;
1168         nr_running = 0;
1169         nr_started = 0;
1170         m_rate = t_rate = 0;
1171
1172         for_each_td(td, i) {
1173                 print_status_init(td->thread_number - 1);
1174
1175                 if (!td->o.create_serialize) {
1176                         init_disk_util(td);
1177                         continue;
1178                 }
1179
1180                 /*
1181                  * do file setup here so it happens sequentially,
1182                  * we don't want X number of threads getting their
1183                  * client data interspersed on disk
1184                  */
1185                 if (setup_files(td)) {
1186                         exit_value++;
1187                         if (td->error)
1188                                 log_err("fio: pid=%d, err=%d/%s\n",
1189                                         (int) td->pid, td->error, td->verror);
1190                         td_set_runstate(td, TD_REAPED);
1191                         todo--;
1192                 } else {
1193                         struct fio_file *f;
1194                         unsigned int i;
1195
1196                         /*
1197                          * for sharing to work, each job must always open
1198                          * its own files. so close them, if we opened them
1199                          * for creation
1200                          */
1201                         for_each_file(td, f, i)
1202                                 td_io_close_file(td, f);
1203                 }
1204
1205                 init_disk_util(td);
1206         }
1207
1208         set_genesis_time();
1209
1210         while (todo) {
1211                 struct thread_data *map[MAX_JOBS];
1212                 struct timeval this_start;
1213                 int this_jobs = 0, left;
1214
1215                 /*
1216                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1217                  */
1218                 for_each_td(td, i) {
1219                         if (td->runstate != TD_NOT_CREATED)
1220                                 continue;
1221
1222                         /*
1223                          * never got a chance to start, killed by other
1224                          * thread for some reason
1225                          */
1226                         if (td->terminate) {
1227                                 todo--;
1228                                 continue;
1229                         }
1230
1231                         if (td->o.start_delay) {
1232                                 spent = mtime_since_genesis();
1233
1234                                 if (td->o.start_delay * 1000 > spent)
1235                                         continue;
1236                         }
1237
1238                         if (td->o.stonewall && (nr_started || nr_running)) {
1239                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1240                                                         td->o.name);
1241                                 break;
1242                         }
1243
1244                         /*
1245                          * Set state to created. Thread will transition
1246                          * to TD_INITIALIZED when it's done setting up.
1247                          */
1248                         td_set_runstate(td, TD_CREATED);
1249                         map[this_jobs++] = td;
1250                         nr_started++;
1251
1252                         if (td->o.use_thread) {
1253                                 dprint(FD_PROCESS, "will pthread_create\n");
1254                                 if (pthread_create(&td->thread, NULL,
1255                                                    thread_main, td)) {
1256                                         perror("pthread_create");
1257                                         nr_started--;
1258                                         break;
1259                                 }
1260                                 if (pthread_detach(td->thread) < 0)
1261                                         perror("pthread_detach");
1262                         } else {
1263                                 pid_t pid;
1264                                 dprint(FD_PROCESS, "will fork\n");
1265                                 pid = fork();
1266                                 if (!pid) {
1267                                         int ret = fork_main(shm_id, i);
1268
1269                                         _exit(ret);
1270                                 } else if (i == fio_debug_jobno)
1271                                         *fio_debug_jobp = pid;
1272                         }
1273                         fio_mutex_down(startup_mutex);
1274                 }
1275
1276                 /*
1277                  * Wait for the started threads to transition to
1278                  * TD_INITIALIZED.
1279                  */
1280                 fio_gettime(&this_start, NULL);
1281                 left = this_jobs;
1282                 while (left && !fio_abort) {
1283                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1284                                 break;
1285
1286                         usleep(100000);
1287
1288                         for (i = 0; i < this_jobs; i++) {
1289                                 td = map[i];
1290                                 if (!td)
1291                                         continue;
1292                                 if (td->runstate == TD_INITIALIZED) {
1293                                         map[i] = NULL;
1294                                         left--;
1295                                 } else if (td->runstate >= TD_EXITED) {
1296                                         map[i] = NULL;
1297                                         left--;
1298                                         todo--;
1299                                         nr_running++; /* work-around... */
1300                                 }
1301                         }
1302                 }
1303
1304                 if (left) {
1305                         log_err("fio: %d jobs failed to start\n", left);
1306                         for (i = 0; i < this_jobs; i++) {
1307                                 td = map[i];
1308                                 if (!td)
1309                                         continue;
1310                                 kill(td->pid, SIGTERM);
1311                         }
1312                         break;
1313                 }
1314
1315                 /*
1316                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1317                  */
1318                 for_each_td(td, i) {
1319                         if (td->runstate != TD_INITIALIZED)
1320                                 continue;
1321
1322                         td_set_runstate(td, TD_RUNNING);
1323                         nr_running++;
1324                         nr_started--;
1325                         m_rate += td->o.ratemin;
1326                         t_rate += td->o.rate;
1327                         todo--;
1328                         fio_mutex_up(td->mutex);
1329                 }
1330
1331                 reap_threads(&nr_running, &t_rate, &m_rate);
1332
1333                 if (todo)
1334                         usleep(100000);
1335         }
1336
1337         while (nr_running) {
1338                 reap_threads(&nr_running, &t_rate, &m_rate);
1339                 usleep(10000);
1340         }
1341
1342         update_io_ticks();
1343         fio_unpin_memory();
1344 }
1345
1346 int main(int argc, char *argv[])
1347 {
1348         long ps;
1349
1350         sinit();
1351
1352         /*
1353          * We need locale for number printing, if it isn't set then just
1354          * go with the US format.
1355          */
1356         if (!getenv("LC_NUMERIC"))
1357                 setlocale(LC_NUMERIC, "en_US");
1358
1359         if (parse_options(argc, argv))
1360                 return 1;
1361
1362         if (!thread_number)
1363                 return 0;
1364
1365         ps = sysconf(_SC_PAGESIZE);
1366         if (ps < 0) {
1367                 log_err("Failed to get page size\n");
1368                 return 1;
1369         }
1370
1371         page_size = ps;
1372         page_mask = ps - 1;
1373
1374         if (write_bw_log) {
1375                 setup_log(&agg_io_log[DDIR_READ]);
1376                 setup_log(&agg_io_log[DDIR_WRITE]);
1377         }
1378
1379         startup_mutex = fio_mutex_init(0);
1380
1381         set_genesis_time();
1382
1383         status_timer_arm();
1384
1385         run_threads();
1386
1387         if (!fio_abort) {
1388                 show_run_stats();
1389                 if (write_bw_log) {
1390                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1391                         __finish_log(agg_io_log[DDIR_WRITE],
1392                                         "agg-write_bw.log");
1393                 }
1394         }
1395
1396         fio_mutex_remove(startup_mutex);
1397         return exit_value;
1398 }